Browse the corpus

Main Previous national-level exclusive breastfeeding (EBF) prevalence estimates within Africa4,7,8 found substantial heterogeneity between countries, while studies comparing urban and rural locations8,9, and subnational-level estimates in select countries8,10,11, also identified considerable within-country heterogeneity. We found that EBF prevalence and trends varied greatly across the African continent between 2000 and 2017, often irrespective of national or subnational boundaries (Fig. 1a–d). The greatest observable patterns of improvement, where estimated EBF levels had increased from <25% to ≥40% in the modeled period, were along or near the East African Rift, including Sudan, South Sudan, Democratic Republic of the Congo (DRC), Kenya, Tanzania, Zambia and Malawi. Within these countries, an estimated 68 second administrative subdivisions (out of 534) had low EBF prevalence (estimates: <25%) in 2000, which subsequently increased to meet or exceed the World Health Organization’s (WHO’s) Global Nutrition Target (GNT; estimated EBF prevalence of ≥50%) by 2017. The estimated national prevalence nearly doubled in some countries in western (for example, Burkina Faso) and southern (for example, Namibia) sub-Saharan Africa (SSA) between 2000 and 2017. This was achieved by reducing the number of areas with low EBF prevalence. At the same time, estimates at higher spatial resolutions highlight corners of persistent need in countries that made notable national progress towards EBF targets, including in eastern Angola and eastern and coastal areas in South Africa. For example, we estimated a 13.6 percentage-point increase (95% uncertainty interval: 8.3–19.6) in national EBF prevalence in South Africa, from 10.2% (8.1–12.6%) in 2000 to 23.8% (18.5–30.0%) in 2017. Yet, areas with persistently lower levels, such as the City of Johannesburg (4.9% (2.9–7.7%) in 2000; 17.4% (10.4–27.0%) in 2017) and throughout Gauteng province (5.7% (3.5–8.7%) in 2000; 19.4% (12.0–29.0%) in 2017), contribute to South Africa’s relatively low national average.Fig. 1 EBF prevalence (2000–2017) among infants under 6 months and progress towards the 2025 WHO GNT.

of Johannesburg (4.9% (2.9–7.7%) in 2000; 17.4% (10.4–27.0%) in 2017) and throughout Gauteng province (5.7% (3.5–8.7%) in 2000; 19.4% (12.0–29.0%) in 2017), contribute to South Africa’s relatively low national average.Fig. 1 EBF prevalence (2000–2017) among infants under 6 months and progress towards the 2025 WHO GNT. a–c, Prevalence of EBF practices at the 5 km × 5 km resolution in 2000 (a), 2010 (b) and 2017 (c). d, Prevalence of EBF at the first administrative subdivision in 2017. e, Overlapping population-weighted lowest and highest 10% of grid cells and weighted AROC in EBF from 2000–2017. f, Overlapping population-weighted quartiles of EBF and relative 95% uncertainty in 2017. Cut-offs for the quartiles were 25.0% (25th percentile), 38.5% (50th percentile) and 52.3% (75th percentile) for the EBF prevalence axis, and 0.500 (25th percentile), 0.902 (50th percentile) and 0.137 (75th percentile) for the relative uncertainty axis (calculated as the absolute range of the uncertainty intervals divided by the estimate). g, Weighted annualized percentage change in EBF prevalence from 2000–2017. h, Grid cell level prevalence of EBF predicted for 2025, projected from 2017 based on AROC between 2000 and 2017. i, Acceleration in the annualized increase in EBF required to meet WHO GNT by 2025. Dark blue pixels were either non-increasing or must accelerate their rate of increase by more than 400% above 2000–2017 rates during 2017–2025 to achieve the target. White pixels require no increase to meet WHO GNT by 2025. Maps reflect administrative boundaries, land cover, lakes and population; gray-colored grid cells had fewer than ten people per 1 km × 1 km grid cell and were classified as ‘barren or sparsely vegetated’, or were not included in this analysis.

chieve the target. White pixels require no increase to meet WHO GNT by 2025. Maps reflect administrative boundaries, land cover, lakes and population; gray-colored grid cells had fewer than ten people per 1 km × 1 km grid cell and were classified as ‘barren or sparsely vegetated’, or were not included in this analysis. Figure 1e features the best- and worst-performing locales by overlaying the highest- and lowest-prevalence areas (90th and 10th percentiles, respectively) across Africa in 2000 and 2017, and areas with the highest and lowest weighted annualized rates of change (AROC) for the 18-year study period. Burundi and Rwanda—nearly homogeneously—were among the top achievers in Africa in both 2000 and 2017, as were north-western parts of Ethiopia, areas scattered throughout Uganda, and south-western Zambia. Sudan showed some of the highest and most consistent rates of increase within its borders. Areas in southern Côte d’Ivoire, eastern and western Burkina Faso, south-western Niger, southern Nigeria, northern Central African Republic (CAR), northern Angola, southern DRC and central South Africa had among the lowest EBF prevalence in 2000; however, high AROC propelled most of these areas out of the lowest decile by 2017. Conversely, areas in northern Nigeria and throughout Chad were among the lowest-prevalence and lowest-AROC areas, indicating stagnation or reversals in progress. The majority of areas in Gabon and Somalia, as well as a large geographic area in south-eastern Niger and pockets in north-eastern Angola and southern Tunisia, were in the lowest-prevalence decile in 2000 and 2017.

d were among the lowest-prevalence and lowest-AROC areas, indicating stagnation or reversals in progress. The majority of areas in Gabon and Somalia, as well as a large geographic area in south-eastern Niger and pockets in north-eastern Angola and southern Tunisia, were in the lowest-prevalence decile in 2000 and 2017. Our detailed spatial estimates display broad within-country differences throughout Africa that would otherwise be masked by national or less granular subnational estimates. ‘Hot spots’ of low EBF prevalence are highlighted at higher resolutions (Fig. 2). Nationally in 2017, Ethiopia (58.2% (50.4–65.8%)), Tanzania (52.6% (46.0–58.9%)), DRC (45.9% (40.0–52.5%)), Kenya (37.6% (26.8–49.5%)) and Namibia (40.9% (31.6–50.2%)) were at or approaching the 2025 prevalence target (see Fig. 1f for a relative uncertainty map). However, some second administrative subdivisions in south-eastern Ethiopia and Tanzania with slower EBF uptake fell short, and will fail to meet targets based on current trajectories (Supplementary Tables 11 and 12), while local-level areas in north-eastern Namibia and south-western DRC and Kenya were found to have lower prevalence (<25%). Within-country disparities in estimated EBF prevalence were both common and widespread: in 2017, at least a twofold difference in estimated EBF prevalence existed across second administrative subdivisions in 53.1% (26 of 49) of African countries; at least a threefold difference occurred in 14.3% (7 of 49) of countries, and a more than sixfold difference was estimated in Niger and Nigeria.Fig. 2 EBF prevalence in 2017 among infants under 6 months at different levels of spatial resolution. a–d, Prevalence of EBF in 2017 at the national (a), first administrative subdivision (b), second administrative subdivision (c) and 5 km × 5 km grid cell level (d). Maps reflect administrative boundaries, land cover, lakes and population; gray-colored grid cells had fewer than ten people per 1 km × 1 km grid cell and were classified as ‘barren or sparsely vegetated’, or were not included in this analysis.

, second administrative subdivision (c) and 5 km × 5 km grid cell level (d). Maps reflect administrative boundaries, land cover, lakes and population; gray-colored grid cells had fewer than ten people per 1 km × 1 km grid cell and were classified as ‘barren or sparsely vegetated’, or were not included in this analysis. The weighted AROC between 2000 and 2017 (Fig. 1g) and corresponding projected levels of EBF prevalence in 2025 (Fig. 1h) were highly variable across the continent, with declines in EBF prevalence observed in several countries. In some cases—as in Madagascar—these decreases occurred against a background of initially high EBF prevalence in 2000, and a few central areas are nonetheless projected to meet WHO GNT of at least 50% EBF by 2025, assuming that recent trends continue. Although Ethiopia’s north-western areas met WHO GNT by 2017, some of these locations experienced annualized declines and failed to meet the minimum 1.2% relative annual increase recommended for well-performing countries6 (Fig. 3). Following current trajectories, fewer than half of African countries (36.7%; 18 of 49) are projected to meet or exceed WHO GNT by 2025 based on national-level estimates (Supplementary Table 12). Success in meeting WHO GNT in 2025 was predicted for all first administrative subdivisions in just eight countries (Burundi, Guinea-Bissau, Lesotho, Malawi, Rwanda, São Tomé and Príncipe, Sierra Leone and Zambia) and for all second administrative subdivisions in just three countries (Guinea-Bissau, Rwanda, and São Tomé and Príncipe) (Fig. 3 and Supplementary Table 12).Fig. 3 Progress towards WHO GNT 2025 during 2013–2017. a–e, Results are shown for 2013 (a), 2014 (b), 2015 (c), 2016 (d) and 2017 (e). Areas in purple highlight places that met WHO GNT by achieving at least 50% EBF prevalence. Areas in green highlight locations that achieved a 1.2% annual relative increase in addition to meeting WHO GNT of at least 50% EBF prevalence. Maps reflect administrative boundaries, land cover, lakes and population; gray-colored grid cells had fewer than ten people per 1 km × 1 km grid cell and were classified as ‘barren or sparsely vegetated’, or were not included in this analysis.

tive increase in addition to meeting WHO GNT of at least 50% EBF prevalence. Maps reflect administrative boundaries, land cover, lakes and population; gray-colored grid cells had fewer than ten people per 1 km × 1 km grid cell and were classified as ‘barren or sparsely vegetated’, or were not included in this analysis. Many areas will require substantial acceleration of current rates of improvement or reversals in trends to meet WHO GNT, including much of North Africa and western SSA, and parts of all other African regions (Fig. 1i). Given the rates of change we estimated for the 2000–2017 period, a band of areas along the Sahel in western SSA need an estimated 400% or more increase of existing AROC to achieve 50% EBF prevalence by 2025. Most East African Rift and bordering countries are on track to achieve targets. Despite large gains between 2000 and 2017, and high AROC, reaching WHO GNT remains unlikely (<50% probability) by projections for some countries, such as in Mali and Côte d’Ivoire (Fig. 4). At subnational levels, just 6.3% (412 of 6,499) of second administrative subdivisions across Africa have a high probability (>95%) of reaching WHO GNT by 2025, while 43.3% (2,817 of 6,499) were almost certain to not reach the target (<5% probability). Local-level variation of this probability can be broad; within Senegal, Angola, Ethiopia and Tanzania, areas with <5% and areas with >95% probabilities of meeting WHO GNT were estimated. Despite a higher probability (>50%) of national achievement to meet the 50% prevalence target by 2025, this goal was not within reach in Ethiopia’s or Tanzania’s south-eastern areas (<5% probability), indicating vulnerable populations left behind in general progress.Fig. 4 Probability of meeting WHO GNT for EBF by 2025 at different levels of spatial resolution. a–d, Probability of meeting WHO GNT of at least 50% EBF prevalence by 2025 at the national (a), first administrative subdivision (b), second administrative subdivision (c) and 5 km × 5 km grid cell level (d). Maps reflect administrative boundaries, land cover, lakes and population; gray-colored grid cells had fewer than ten people per 1 km × 1 km grid cell and were classified as ‘barren or sparsely vegetated’, or were not included in this analysis.

, second administrative subdivision (c) and 5 km × 5 km grid cell level (d). Maps reflect administrative boundaries, land cover, lakes and population; gray-colored grid cells had fewer than ten people per 1 km × 1 km grid cell and were classified as ‘barren or sparsely vegetated’, or were not included in this analysis. These geospatial analyses add to the landscape of mapping progress towards nutrition targets and, when compared against mapped estimates of conditions associated with infant nutrition, can aid in determining where the most at-risk populations are located. Many child health conditions are inextricably linked to infants’ feeding practices. EBF is associated with reduced incidence of diarrhea and pneumonia, and reduced infant mortality rates2. Areas with low EBF prevalence and high diarrheal incidence12, such as in communities scattered throughout western and central SSA and Somalia in 2015 (Extended Data Fig. 1), could benefit from increased breastfeeding promotion, education and support. Locations with both high infant mortality13 and low EBF prevalence (Extended Data Fig. 2), such as in Somalia, and along or near the western Sahel in Guinea, Côte d’Ivoire, Nigeria, Chad and CAR require urgent attention to improve EBF practices for the greatest benefit to child survival.

omotion, education and support. Locations with both high infant mortality13 and low EBF prevalence (Extended Data Fig. 2), such as in Somalia, and along or near the western Sahel in Guinea, Côte d’Ivoire, Nigeria, Chad and CAR require urgent attention to improve EBF practices for the greatest benefit to child survival. Our results underscore substantial improvements in EBF practices across large geographic areas, as well as disparities both between and within countries. Overall, we estimated that only 18 African countries out of 49 (36.7%) are nationally on track to meet WHO GNT by 2025, agreeing with the United Nations Children’s Fund (UNICEF) and WHO conclusions that the majority of nations are not on track to achieve nutrition targets8,9. Our projections offer the additional capacity to identify which countries and areas are predicted to fall short of WHO GNT, with estimates disentangled from the tracking of other nutritional targets. Even within countries projected to meet WHO GNT, pockets of slow uptake remain; only three countries are predicted to reach 50% EBF prevalence by 2025 in all second administrative subdivisions, and each region had countries with poor-performing areas (<25% estimated prevalence). By mapping local-level EBF prevalence and trends, we reveal considerable geographic heterogeneity, provide a tool to aid decision-makers in visualizing where populations with the greatest needs may reside, and allow for the aggregation of estimates within meaningful catchment areas or administrative levels. When compared against additional data on interventions employed, this information could aid in identifying program or policy successes and failures.

in visualizing where populations with the greatest needs may reside, and allow for the aggregation of estimates within meaningful catchment areas or administrative levels. When compared against additional data on interventions employed, this information could aid in identifying program or policy successes and failures. The varied success in implementing national policies—including the adoption and enforcement of the International Code of Marketing Breast-Milk Substitutes14 (the Code), paid maternity leave and breastfeeding-related workplace programs15, and the Baby-Friendly Hospital Initiative (BFHI)16—may contribute to variation in EBF levels across Africa. In 1973, the publication of The Baby Killer, an exposé on the manipulative marketing tactics used by some breast-milk substitute (BMS) companies, raised widespread alarm17. Controversial BMS promotion strategies, such as free samples and dressing saleswomen as nurses, may have led to increased usage and dependence on unaffordable formula products lacking the nutritional and immunological benefits of breast-milk, and the exposure of infants to increased risks of pathogens and mortality17. In 1981, the Code14 was created at the World Health Assembly to encourage breastfeeding and regulation of the safety and promotion of alternatives. However, as of 2018, only 12 African countries have comprehensive legislation on the Code, and 17 of 47 have no legal measures in place to protect consumers from aggressive BMS marketing18. Even with provisions enacted into national law, active monitoring systems and implementation may be weak or inadequate, yet evidence suggests that more restrictive policies may be associated with less pervasive promotion of unnecessary BMS usage19,20. Additional legislation and more thorough enforcement of existing legislation is needed in some countries.

ional law, active monitoring systems and implementation may be weak or inadequate, yet evidence suggests that more restrictive policies may be associated with less pervasive promotion of unnecessary BMS usage19,20. Additional legislation and more thorough enforcement of existing legislation is needed in some countries. Maternal protection policies—such as onsite child care, physical areas for breastfeeding or pumping and storing breast-milk, and paid maternity leave policies—offer additional support and autonomy to mothers working outside the home who may otherwise turn to BMS19,21–23. The International Labour Organization advocates for legislation requiring 14 weeks of paid maternity leave to support breastfeeding by working mothers9. Additionally, a recent analysis across 38 low- and middle-income countries found a significant and positive association between a 1-month increase in legislated maternity leave and a 5.9 percentage-point difference in EBF15. A study in Ghana showed that working mothers with shorter maternity leave were less likely to practice EBF24. In 2018, however, only ten African countries met the basic provisions of maternity leave25.

e association between a 1-month increase in legislated maternity leave and a 5.9 percentage-point difference in EBF15. A study in Ghana showed that working mothers with shorter maternity leave were less likely to practice EBF24. In 2018, however, only ten African countries met the basic provisions of maternity leave25. BFHI is a WHO- and UNICEF-led effort to ensure that all hospital-based and free-standing maternity units are breastfeeding support centers16. ‘Baby friendly’-designated facilities do not accept free or low-cost BMS, and implement ten steps for successful breastfeeding—a package of clinical practices and management procedures that have been demonstrated to improve EBF rates26. While the majority of African countries have adopted BFHI, only two have reported more than 50% implementation of the ten steps across their facilities25. Many have reported 0% implementation or have not assessed facilities in the past 5 years, suggesting that the initiative has become dormant9,25. As with any program, BFHI delivery efficiency varies across space and time; thus, local-level monitoring to gauge progress is needed.

the ten steps across their facilities25. Many have reported 0% implementation or have not assessed facilities in the past 5 years, suggesting that the initiative has become dormant9,25. As with any program, BFHI delivery efficiency varies across space and time; thus, local-level monitoring to gauge progress is needed. Many of EBF’s primary barriers involve cultural perceptions and misinformation, which can be highly variable across communities, contributing to local variation in EBF practices and the need for community-based interventions. Women’s perceptions of insufficient breast-milk, beliefs about infant thirst and need for water, and the cultural and family norms that support the early introduction of food and liquid are a few examples of barriers that vary broadly across communities21–23,27. Mothers who perceive their breast-milk to be nutritionally inadequate are more likely to discontinue EBF22. In many settings, mothers’ early breast-milk (colostrum) is considered sour and difficult to digest, and is discarded and replaced by prelacteal feeding of water, formula or animal milk, making it difficult to establish breastfeeding21–23. The early introduction of water and porridges is common practice across the continent23,28, inhibiting EBF practice and exposing infants to disease and nutritional risks from pathogens; plain water is the greatest obstacle in the western and central regions28, and women along the Sahel have cited the high heat index as a reason for feeding their infants water23. Generational feeding practices are passed on, and mothers can be influenced by community and family members’ attitudes towards breastfeeding23,28,29.

ter is the greatest obstacle in the western and central regions28, and women along the Sahel have cited the high heat index as a reason for feeding their infants water23. Generational feeding practices are passed on, and mothers can be influenced by community and family members’ attitudes towards breastfeeding23,28,29. Although pervasive, the aforementioned issues can be addressed through lactation management, breastfeeding support, and social and behavior-change communication approaches27. Through intensive home-based support or participatory women’s groups, health workers can dispel breastfeeding myths, increase confidence and equip mothers with the necessary skills to address breastfeeding issues, including infant suckling difficulties or pain22. A study in Ghana showed that women who received infant feeding recommendations from health workers were more likely to practice EBF24. An integrated approach combining promotion, counseling and education on EBF in communities and health facilities has been found to be significantly more effective than counselling as a single intervention29. In a systematic review of 46 studies, all forms of extra breastfeeding support—including face-to-face or telephone interactions by professional or lay support staff—led to a decrease in EBF cessation (risk ratio 0.88 (0.85–0.92)) when analyzed together30. As of 2018, however, only 18 of the 49 African countries in our analyses offered community-based breastfeeding programs in all of their districts, and 21 report offering individual infant and young child feeding counselling in all of their primary health care facilities25; no information on the quality of services or number of women reached by these programs is available9. Funding for such interventions is limited, as only 17 African countries currently receive at least US$2 per birth towards breastfeeding programs25.

feeding counselling in all of their primary health care facilities25; no information on the quality of services or number of women reached by these programs is available9. Funding for such interventions is limited, as only 17 African countries currently receive at least US$2 per birth towards breastfeeding programs25. Government buy-in and combined approaches are key to increasing the likelihood of success of community-based programs. The Alive & Thrive Initiative has shown that improving EBF is possible at scale through a combination of advocacy, interpersonal communication, community mobilization and mass media21,31. In Ghana and Madagascar, EBF rates significantly improved when training, as well as social and behavior-change activities, were delivered via partnerships between government and non-governmental organizations27. Furthermore, a meta-analysis found that combining health system, home and community-based approaches was most effective at improving EBF rates29. Impact evaluations of similar community-based projects with health systems integration in Ethiopia, Kenya and Senegal found EBF counselling from both facility-based and community personnel to significantly increase the odds (odds ratio = 2.90; P < 0.001) of EBF32.

approaches was most effective at improving EBF rates29. Impact evaluations of similar community-based projects with health systems integration in Ethiopia, Kenya and Senegal found EBF counselling from both facility-based and community personnel to significantly increase the odds (odds ratio = 2.90; P < 0.001) of EBF32. It is difficult to interpret EBF trends in Africa without considering the human immunodeficiency virus (HIV) epidemic and the impact that evolving recommendations may have had on infant feeding practices in high-burden countries. In 1997, WHO was advising HIV-infected mothers to avoid breastfeeding—to prevent mother-to-child transmission—if replacement feeding could be practised safely33. However, several studies in the early 2000s reported lower HIV transmission risk among exclusively breastfed infants compared with those who were mixed-fed (that is, breastfed and given solid foods or infant formula)34–36. These studies, coupled with subsequent research indicating the efficacy of antiretroviral therapy (ART)37, led to new guidance in 2010 in favor of EBF for children of HIV-infected women on ART38—a recommendation reiterated in 201639. These changing recommendations may have contributed to low initial EBF rates in 2000 and subsequent improvements through 2017 in some countries, particularly in eastern and southern SSA, where HIV prevalence is high and access to HIV testing, counselling and ART increased during the 2000–2017 period. Continued access to HIV treatment, along with support for breastfeeding in high-burden areas, are urgent priorities to further EBF and optimize maternal and child health40.

rly in eastern and southern SSA, where HIV prevalence is high and access to HIV testing, counselling and ART increased during the 2000–2017 period. Continued access to HIV treatment, along with support for breastfeeding in high-burden areas, are urgent priorities to further EBF and optimize maternal and child health40. Widely considered one of the most effective behaviors in preventing child mortality, and a key component of WHO’s Global Action Plan for the Prevention and Control of Pneumonia and Diarrhoea1, EBF can save and improve the quality of lives3. While facilities may be collecting data on breastfeeding interventions and EBF rates for monitoring purposes in some locations, improved data collection efforts are needed across many African locales, and the estimates here are supplemental to those efforts. The EBF estimates presented here, at various spatial resolutions, can assist public health practitioners and policymakers in visualizing and identifying disparities across and within countries—informing decisions on where existing interventions and policies may need to be bolstered, or new strategies considered, to ensure that all infants have the opportunity to survive and thrive.

, can assist public health practitioners and policymakers in visualizing and identifying disparities across and within countries—informing decisions on where existing interventions and policies may need to be bolstered, or new strategies considered, to ensure that all infants have the opportunity to survive and thrive. Methods Overview Our analyses provide annual estimates of EBF prevalence among infants under 6 months of age during the period of 2000–2017 across Africa at the national, first and second administrative (for example, state and district level, respectively), and 5 km × 5 km grid cell levels. EBF prevalence is defined as the proportion of children who receive only breast-milk, oral rehydration salts or other medicines or vitamins, without receiving additional food or drink (including water) between birth and 6 months of age. Our primary goal was to provide prevalence predictions at a high spatial resolution across the African continent with the best out-of-sample predictive performance. The methodology used here is similar to that used for previous analyses of diarrhea incidence12, under 5 years mortality13, child growth failure41, educational attainment42 and HIV43 in Africa. We first mapped our estimates on a 5 km × 5 km grid to remain consistent with these previous analyses, align with the resolutions available for pre-existing covariates incorporated in these analyses, and maintain flexibility in aggregating these estimates to other levels of interest (for example, first and second administrative subdivisions). Our analyses of 49 countries include mainland Africa and the islands of Madagascar, Comoros, and São Tomé and Príncipe. We do not provide estimates for Libya, Djibouti or island nations where survey data were not available (Mauritius, Seychelles and Cape Verde). This study follows the Guidelines for Accurate and Transparent Health Estimates Reporting (http://gather-statement.org; Supplementary Table 1).

moros, and São Tomé and Príncipe. We do not provide estimates for Libya, Djibouti or island nations where survey data were not available (Mauritius, Seychelles and Cape Verde). This study follows the Guidelines for Accurate and Transparent Health Estimates Reporting (http://gather-statement.org; Supplementary Table 1). Data extraction and processing Extended Data Fig. 3a describes the detailed steps performed during the data extraction and data processing workflow. We extracted data from the Demographic and Health Surveys (DHS) program, UNICEF's Multiple Indicator Cluster Surveys (MICS), and country-specific and other multinational surveys conducted in the years 1998–2017 for African countries. Though we model estimates for the years 2000 to 2017, we assigned data from 14 surveys in the years 1998–1999 to the year 2000 to address data scaricity in earlier years and to help establish a baseline. We searched the Global Health Data Exchange (GHDx: http://ghdx.healthdata.org/) for all surveys in African countries tagged as containing EBF indicators of interest; designed and tested a codebook, or survey data extraction framework, for breastfeeding variables present in the household surveys; extracted and geo-matched (either to geospatial coordinates or administrative subdivisions) all surveys available for Africa; and refreshed our query of the GHDx for surveys performed in African countries.

ted a codebook, or survey data extraction framework, for breastfeeding variables present in the household surveys; extracted and geo-matched (either to geospatial coordinates or administrative subdivisions) all surveys available for Africa; and refreshed our query of the GHDx for surveys performed in African countries. Data inclusion and exclusion criteria As our goal was to estimate the prevalence of EBF among infants under 6 months of age, we only included data regarding the feeding of children less than 6 months old at the time of survey (0–5 months in survey data). Specifically, our inclusion criteria for survey microdata (that is, surveys with individual-level responses) were the following: (1) the survey must have been conducted in an African country between 1998 and 2017; (2) survey responses must be available at the individual level; (3) the survey must contain subnational geographic identifiers, which could include either subnational areal units (typically administrative subdivisions) or Global Positioning System (GPS) coordinates, and data referenced to subnational units must also contain survey weights for each observation; and (4) the survey must contain questions about the age of the child, whether the child is still being breastfed and whether the child has consumed other food or liquid items. Typically, consumption during the past 24 h was recorded. In eight out of 181 surveys with microdata, the question about food or liquid items did not specify a particular recall period. After performing sensitivity analysis, we decided to keep those eight surveys in our model. In cases where survey microdata were not available, we searched for survey report estimates. Our inclusion criteria for these survey reports were the following: (1) the survey must have been conducted in an African country between 1998 and 2017; (2) the survey must contain subnational identifiers, which could include subnational areal units (typically administrative subdivisions); and (3) the survey must contain the prevalence of EBF, with a sample size or the lower and upper bounds for the 95% confidence interval.

in an African country between 1998 and 2017; (2) the survey must contain subnational identifiers, which could include subnational areal units (typically administrative subdivisions); and (3) the survey must contain the prevalence of EBF, with a sample size or the lower and upper bounds for the 95% confidence interval. Very few surveys directly asked about EBF practice; as such, we derived breastfeeding status from questions asking about the consumption of breast-milk and other foods, liquids and medicines consumed in a set period before the survey, typically within the 24-h period before survey completion. We excluded surveys that only asked mothers and caregivers whether infants had been exclusively breastfed (for example, ‘did you exclusively breastfeed?’) without ascertaining further information. This exclusion criterion was established after finding, by comparing the responses in surveys containing both types of questions, that many mothers and caregivers stated that infants had exclusively breastfed but also answered that they had received food or water in the 24-h recall questions. This may have been due to the respondent misunderstanding the meaning of ‘exclusive breastfeeding’, or the question may have been misinterpreted with translation. Instead, we classified children as exclusively breastfed if survey responses indicated that they received only breast-milk and medicines (oral rehydration salts, vitamins or other medicines) without other foods or liquids during the 24-h period before the survey.

question may have been misinterpreted with translation. Instead, we classified children as exclusively breastfed if survey responses indicated that they received only breast-milk and medicines (oral rehydration salts, vitamins or other medicines) without other foods or liquids during the 24-h period before the survey. To identify potential survey biases, we reviewed national-level survey estimates for each country and compared them with national-level estimates from the DHS program, the 2017 Global Burden of Disease (GBD) study5 and the geospatial model. In cases where a survey’s estimates appeared implausible compared with other existing survey-based data sources, we inspected differences in definitions, data collection or other methodological explanations. Identified data sources As a result, we identified and used 188 household surveys that had complete records of questions relating to infant feeding and geographical information; 102 were from the DHS series, 79 were from the MICS series and seven were from other sources. Extended Data Fig. 4 shows the spatial and temporal extent of data availability by country, and Supplementary Tables 2 and 3 provide information on the names, citations and geographic detail of surveys of the underlying data sources of our models.

series, 79 were from the MICS series and seven were from other sources. Extended Data Fig. 4 shows the spatial and temporal extent of data availability by country, and Supplementary Tables 2 and 3 provide information on the names, citations and geographic detail of surveys of the underlying data sources of our models. Supplementary Table 4 provides a list of surveys that were excluded from both the geostatistical model and GBD 2017 estimates5. Supplementary Table 5 provides a list of surveys that were included in the geostatistical model but excluded from the GBD estimates (in cases where surveys were non-nationally representative but could provide spatial information for the geostatistical model). Supplementary Table 6 provides a list of surveys that were included in GBD estimates but excluded from the geostatistical model.

in the geostatistical model but excluded from the GBD estimates (in cases where surveys were non-nationally representative but could provide spatial information for the geostatistical model). Supplementary Table 6 provides a list of surveys that were included in GBD estimates but excluded from the geostatistical model. Data processing After data identification and extraction, we aggregated the individual-level responses from survey microdata to calculate EBF prevalence and the effective sample size at the finest possible spatial resolution available, incorporating individual-level sample weights and using the Kish approximation44 for the effective sample size. Each individual child record was associated with a cluster, a group of neighboring households or a ‘village’ that acts as a primary sampling unit (a census enumeration area). For surveys where a latitude and longitude pair representing the location of each survey cluster were available (‘point data’), data were aggregated to these specific coordinates. Geographic coordinates or place names for each cluster were included in 101 surveys (33,341 clusters).

ing unit (a census enumeration area). For surveys where a latitude and longitude pair representing the location of each survey cluster were available (‘point data’), data were aggregated to these specific coordinates. Geographic coordinates or place names for each cluster were included in 101 surveys (33,341 clusters). In the case of survey microdata where geographical coordinates were not available and in the case of survey reports, we assigned data to the smallest available administrative unit in the survey (‘polygon data’)45. We ‘resampled’ data matched to polygons to generate pseudo-point data based on the underlying population distribution within the polygon. The methods for resampling were consistent with those previously used in geospatial modeling of under 5 years mortality13. Specifically, for each polygon-level observation, we randomly sampled 10,000 locations among grid cells in the given polygon with probability proportional to grid cell population. A grid cell was assigned to a polygon if its centroid fell within the geographic boundary. We performed k-means clustering (with k set to 1 per 40 grid cells) on the sampled points to generate a reduced set of locations to be used in modeling based on the k-means cluster centroids. Weights were assigned to each pseudo-point proportional to the number of sampled points contained in each of the k-means clusters (that is, the number of sampled points divided by 10,000). Each pseudo-point generated by this process was assigned the EBF prevalence and sample size observed for the polygon as a whole, and the weights associated with each pseudo-point were applied during all stages of model fitting.

d in each of the k-means clusters (that is, the number of sampled points divided by 10,000). Each pseudo-point generated by this process was assigned the EBF prevalence and sample size observed for the polygon as a whole, and the weights associated with each pseudo-point were applied during all stages of model fitting. After performing the data processing described above, our final dataset consisted of 60,083 clusters (33,341 of which were GPS-located data points and 26,742 of which were polygon data) from 188 surveys (181 surveys with microdata and seven survey reports) representing 153,465 children across 49 African countries. Statistical analysis Covariates In these analyses, we included the following socioeconomic, environmental and health-related covariates to improve the predictions of EBF: urbanicity, night-time lights, travel time to the nearest settlement with >50,000 inhabitants, total population, human development index (HDI), educational attainment in women of reproductive age (15–49 years old), nutritional yield for vitamin A, and HIV prevalence. These covariates were selected because they are factors or proxies for factors that previous literature has identified to be associated (not necessarily causally) with EBF prevalence.

index (HDI), educational attainment in women of reproductive age (15–49 years old), nutritional yield for vitamin A, and HIV prevalence. These covariates were selected because they are factors or proxies for factors that previous literature has identified to be associated (not necessarily causally) with EBF prevalence. The first four covariates were included as measures or proxies for connectedness and urbanicity, as EBF is typically found to be different in urban areas compared with rural locations. HDI—a composite indicator of key aspects of development (namely, education, economy and health)—was chosen based on previous studies relating country development to EBF. Educational attainment in women of reproductive age (15–49 years old) was included because previous studies highlight education as a maternal factor influencing the decision to initiate and continue EBF. Nutritional yield for vitamin A was chosen as a proxy of maternal nutrition while breastfeeding. HIV was included given the known risks of mother-to-child transmission of HIV and consequent potential avoidance of breastfeeding in hyper-endemic settings over the study period. These covariates underwent spatial and temporal processing in preparation for their inclusion in analysis. See Supplementary Table 7 for references to the covariate data used in the models, as well as references supporting our rationale for using these covariates.

eeding in hyper-endemic settings over the study period. These covariates underwent spatial and temporal processing in preparation for their inclusion in analysis. See Supplementary Table 7 for references to the covariate data used in the models, as well as references supporting our rationale for using these covariates. Spatial processing involved resampling the input covariate raster to align the spatial resolution of the covariate to the 5 km × 5 km resolution used in modeling. For covariates that were originally at a finer resolution, we resampled the raster by taking the neighborhood average (that is, for the covariates ‘travel time to the nearest settlement of >50,000 inhabitants’ and ‘night-time lights’) or using the nearest neighbor (that is, for the covariate ‘urbanicity’) or sum (that is, for the covariate ‘total population’) of the finer covariate raster to produce one at a 5 km × 5 km resolution. Educational attainment in women of reproductive age and HIV covariates were produced at a 5 km × 5 km resolution in our previous studies, and thus did not require additional spatial processing. For covariates that were originally at lower resolutions (that is, the covariates ‘HDI’ and ‘nutritional yield for vitamin A’), we resampled the raster using bilinear interpolation, with the effect of smoothing some of the hard pixel boundaries in the raw data to make for a 5-km × 5-km-resolution raster.

al processing. For covariates that were originally at lower resolutions (that is, the covariates ‘HDI’ and ‘nutritional yield for vitamin A’), we resampled the raster using bilinear interpolation, with the effect of smoothing some of the hard pixel boundaries in the raw data to make for a 5-km × 5-km-resolution raster. Temporal processing was required in instances where the original temporal resolution of the covariate was anything other than annual. To resolve from a coarser time period to an annual time period, we filled the intervening years with the value from the nearest neighboring year (that is, for the covariate ‘urbanicity’) or used an exponential growth rate model (that is, for the covariate ‘total population’). Night-time lights, educational attainment and HIV prevalence were available at a 1-year temporal resolution and did not require interpolation. As the travel time to the nearest settlement of >50,000 inhabitants and nutritional yield for vitamin A covariates were available only for a single representative year (2015 and 2005, respectively), these covariates were set to be unchanged over time. After interpolation, the covariates of night-time lights, HDI and urbanicity were still missing information for the most recent years of the 2000–2017 period, and in these instances we filled out the end of the time series carrying forward the most recent year without modification.

s were set to be unchanged over time. After interpolation, the covariates of night-time lights, HDI and urbanicity were still missing information for the most recent years of the 2000–2017 period, and in these instances we filled out the end of the time series carrying forward the most recent year without modification. We list detailed information on the temporal resolution and source(s) for each of the eight included covariates in Supplementary Table 7. In addition, the calendar year was used as a covariate in our model. See Extended Data Fig. 5 for maps of spatial covariate raster layers for 2017.

s were set to be unchanged over time. After interpolation, the covariates of night-time lights, HDI and urbanicity were still missing information for the most recent years of the 2000–2017 period, and in these instances we filled out the end of the time series carrying forward the most recent year without modification. We list detailed information on the temporal resolution and source(s) for each of the eight included covariates in Supplementary Table 7. In addition, the calendar year was used as a covariate in our model. See Extended Data Fig. 5 for maps of spatial covariate raster layers for 2017. Spatial covariate stacking Our primary goal was to provide prevalence predictions across the African continent at a high resolution, and we used methods designed to provide the best out-of-sample predictive performance at the cost of inferential understanding. An ensemble covariate modeling method was implemented to both select covariates and capture possible nonlinear effects and complex interactions between them46. We fit separate models for five African regions based on the geographical regions defined for the GBD47 (central, eastern, northern, southern or western, as seen in Extended Data Fig. 3b). For each region, three submodels were fitted to our dataset, using all of our covariate data as explanatory predictors: generalized additive models, boosted regression trees and lasso regression. We selected these three submodels based on the ease of implementation through existing software packages, the fundamental differences in their approaches and a proven track record in predictive accuracy46. Submodels were fit in R using the mgcv, xgboost, glmnet and caret packages.

regression trees and lasso regression. We selected these three submodels based on the ease of implementation through existing software packages, the fundamental differences in their approaches and a proven track record in predictive accuracy46. Submodels were fit in R using the mgcv, xgboost, glmnet and caret packages. Each submodel was fit using fivefold cross-validation to avoid overfitting, and hyper-parameter fitting was performed to maximize the predictive power. For each submodel, we produced two sets of predictions: out of sample and in sample. Out-of-sample predictions for each model were generated by compiling the predictions from the five holdouts from each cross-validation fold, and in-sample predictions were generated by refitting the submodels using all available data. The out-of-sample submodel predictions were used as explanatory covariates when fitting the geostatistical model described below, and the in-sample predictions were used when generating predictions from the geostatistical model, to maximize data use. In both cases, the logit transformation of the predictions was used to put these predictions on the same scale as the linear predictor in the geostatistical model. Maps of in-sample predictions from each stacker are presented in Extended Data Fig. 6. A recent study has shown that this ensemble approach can improve predictive validity by up to 25% over an individual model46.

ns was used to put these predictions on the same scale as the linear predictor in the geostatistical model. Maps of in-sample predictions from each stacker are presented in Extended Data Fig. 6. A recent study has shown that this ensemble approach can improve predictive validity by up to 25% over an individual model46. Geostatistical model As a second step, we fit the geostatistical model below separately for the five African regions. For each region, we write the hierarchy that defines our Bayesian model as follows: EBFipi,Ni~binomial(pi,Ni)logitpi=β0+Xiβ+γci+ϵGPi+ϵi∑β=1γci~normal(0,σcountry2)ϵi~normal(0,σnug2)ϵGPΣspace,Σtime~GP(0,Σspace⊗Σtime) We modeled the number of children who were categorized as ‘exclusively breastfed’ (EBFi) among a sample size (Ni) at space–time location (i) as a binomial random variable. The logit-transformed prevalence of EBF (pi) was specified as a linear combination of a regional intercept (β0), the logit-transformed predictions from the three submodels (Xi), country-level random effects (γci), a correlated spatiotemporal error term (ϵGPi) and an independent and identically distributed nugget (uncorrelated error term) effect ϵi. Weighting coefficients (β) were constrained to sum to 1 (ref. 46). The spatial covariance (Σspace) was modeled using an isotropic and stationary Matérn function48. The temporal covariance (Σtime) was an annual first-order autoregressive function.

t and identically distributed nugget (uncorrelated error term) effect ϵi. Weighting coefficients (β) were constrained to sum to 1 (ref. 46). The spatial covariance (Σspace) was modeled using an isotropic and stationary Matérn function48. The temporal covariance (Σtime) was an annual first-order autoregressive function. The intercept captures the overall mean level of EBF prevalence, while the covariate effects capture the spatial and temporal variation in EBF prevalence that can be described as a function of spatial and temporal variation in the included covariates. The country random effects capture additional variation between countries. Spatially and temporally correlated random effects capture additional variation by location (within and between countries) and time. Finally, the uncorrelated error term (or nugget effect) captures any additional, non-structured variation by location and time. The Matérn covariance function is associated with two hyper-parameters, κ and τ (ν is fixed at 1), while a temporal first-order autoregressive (AR1) covariance function is associated with one hyper-parameter, ρ. The following hyper-priors were set for each of these parameters: θ1= logτ~normal(μθ1,σθ12)θ2= logκ~normal(μθ2,σθ22)log1+ρ∕1-ρ~normal(4,1.22)

associated with two hyper-parameters, κ and τ (ν is fixed at 1), while a temporal first-order autoregressive (AR1) covariance function is associated with one hyper-parameter, ρ. The following hyper-priors were set for each of these parameters: θ1= logτ~normal(μθ1,σθ12)θ2= logκ~normal(μθ2,σθ22)log1+ρ∕1-ρ~normal(4,1.22) The prior for the temporal correlation parameter, ρ, corresponds to a mean of 0.96 and a distribution that is wide enough to include approximately 0.2 to 1.0 within three standard deviations of the mean. This relatively informative prior was chosen because temporal correlation was expected to be high. μθ1, σθ1, μθ2 and σθ2 were automatically determined by integrated nested Laplace approximation (INLA). Priors for fixed effects and hyper-priors for other random effects were set as: β0~normal0,321∕σcountry2~gammarate=1,shape=0.000051∕σnug2~gammarate=1,shape=0.00005

The prior for the temporal correlation parameter, ρ, corresponds to a mean of 0.96 and a distribution that is wide enough to include approximately 0.2 to 1.0 within three standard deviations of the mean. This relatively informative prior was chosen because temporal correlation was expected to be high. μθ1, σθ1, μθ2 and σθ2 were automatically determined by integrated nested Laplace approximation (INLA). Priors for fixed effects and hyper-priors for other random effects were set as: β0~normal0,321∕σcountry2~gammarate=1,shape=0.000051∕σnug2~gammarate=1,shape=0.00005 This model was fit in R-INLA49 using the stochastic partial differential equations50 approach to approximate the continuous spatiotemporal Gaussian random fields ϵGPi. We constructed a finite-elements mesh for the stochastic partial differential equations approximation to the Gaussian process regression using a simplified polygon boundary (as seen in Supplementary Fig. 12 of our previous publication of geospatial estimates of child growth failure41). We set the inner mesh triangle maximum edge length (the mesh size for areas over land) to be 0.25 decimal degrees, and the buffer maximum edge length (the mesh size for areas over the ocean) to be 5.0 decimal degrees. Fitted model parameters are listed in Supplementary Table 8.

ial estimates of child growth failure41). We set the inner mesh triangle maximum edge length (the mesh size for areas over land) to be 0.25 decimal degrees, and the buffer maximum edge length (the mesh size for areas over the ocean) to be 5.0 decimal degrees. Fitted model parameters are listed in Supplementary Table 8. After fitting each model based on regional classification, we generated 1,000 draws of all model parameters from the approximated joint posterior distribution using the inla.posterior.sample() function in R-INLA. For each draw, s, of the model parameters, we constructed a draw of pi(s) as: pi(s)=logit-1β0(s)+Xiβ(s)+γci(s)+ϵGPi(s)+ϵi(s) Additional processing of the output from inla.posterior.sample() is required for the correlated spatiotemporal error term ϵGPi(s) and the nugget effect ϵi(s) before constructing pi(s) according to the equation above. Specifically, for ϵGPi(s), draws are generated initially only at the vertices of the finite element mesh, so we project from this mesh to each location i desired for prediction (that is, the centroid of each grid cell on a 5 km × 5 km grid, as well as years from 2000–2017). For the nugget effect, we generate ϵi(s) for each i by sampling from normal (0,σnug2(s)). At the end of this process, we have 1,000 draws of pi for each grid cell and year.

is mesh to each location i desired for prediction (that is, the centroid of each grid cell on a 5 km × 5 km grid, as well as years from 2000–2017). For the nugget effect, we generate ϵi(s) for each i by sampling from normal (0,σnug2(s)). At the end of this process, we have 1,000 draws of pi for each grid cell and year. Model validation Validation strategy We used fivefold cross-validation to assess the performance of the modeling framework described above. To do so, we first split all survey data into five groups by randomly sorting a list of unique identifiers for each survey, calculating the cumulative effective sample size represented by the surveys in this list, and then dividing the list into five parts at the point where this cumulative sample size was closest to 20, 40, 60 and 80% of the total. This resulted in five groups that were approximately equal in terms of the total effective sample size, and which contained entire surveys (that is, all of the data points derived from each survey were contained exclusively within only one fold). We then fit the model described above five times, excluding each of the five groups of data in turn.

s that were approximately equal in terms of the total effective sample size, and which contained entire surveys (that is, all of the data points derived from each survey were contained exclusively within only one fold). We then fit the model described above five times, excluding each of the five groups of data in turn. After fitting the model five times, the data withheld from each model were matched with predictions from that model, and then these data–prediction pairs were compiled across all five models, resulting in a complete dataset of out-of-sample predictions corresponding to all survey data included in the analysis. EBF prevalence estimates based on single survey clusters are generally quite noisy due to very small sample sizes, and were consequently insufficient as a ‘gold standard’ for evaluating the model predictions13. To address this issue, we aggregated both the observed data and the corresponding out-of-sample predictions within countries and within first- and second-level administrative subdivisions, by calculating a weighted mean of each using the effective sample sizes as the weights. Then, across all data–estimate pairs, we calculated two summary measures: the mean error (a measure of bias) and the root-mean-square error (RMSE; a measure of total variance). In addition, for each data–estimate pair, we constructed 95% prediction intervals from the 2.5th and 97.5th percentiles of 1,000 draws from a binomial distribution corresponding to each of the 1,000 posterior draws of EBF prevalence, with p equal to EBF prevalence in a given posterior draw and N equal to the effective sample size for the data point type. We then calculated coverage as the percentage of data–estimate pairs where the data point was contained within this 95% prediction interval.

to each of the 1,000 posterior draws of EBF prevalence, with p equal to EBF prevalence in a given posterior draw and N equal to the effective sample size for the data point type. We then calculated coverage as the percentage of data–estimate pairs where the data point was contained within this 95% prediction interval. Sensitivity analyses To assess the utility of the stacking ensemble, we ran five fivefold cross-validation holdout experiments, using different combinations of covariates and random effects. The following five models were compared:raw covariates: logitpi=β0+Xiβraw+γci+ϵi stacking predictions as covariates: logitpi=β0+Xiβstack+γci+ϵi a Gaussian process: logit(pi)=β0+γci+ϵGPi+ϵi raw covariates + a Gaussian process: logitpi=β0+Xiβraw+γci+ϵGPi+ϵi stacking covariates + a Gaussian process (final model): logitpi=β0+Xiβstack+γci+ϵGPi+ϵi

Sensitivity analyses To assess the utility of the stacking ensemble, we ran five fivefold cross-validation holdout experiments, using different combinations of covariates and random effects. The following five models were compared:raw covariates: logitpi=β0+Xiβraw+γci+ϵi stacking predictions as covariates: logitpi=β0+Xiβstack+γci+ϵi a Gaussian process: logit(pi)=β0+γci+ϵGPi+ϵi raw covariates + a Gaussian process: logitpi=β0+Xiβraw+γci+ϵGPi+ϵi stacking covariates + a Gaussian process (final model): logitpi=β0+Xiβstack+γci+ϵGPi+ϵi Supplementary Table 9 compares the results of this cross-validation exercise in terms of the performance of these five different modeling strategies, and Extended Data Fig. 7 provides a comparison of the estimates derived from these different models. At all three levels of aggregation, and both in and out of sample, mean error (bias) is relatively low, ranging from −0.49 to 0.42 percentage points. Out-of-sample RMSE is relatively similar for all five models, while in sample, model 1 (raw covariates only) has noticeably worse RMSE compared with the other models. Overall, model 5 (a two-stage model including stacked regression for the covariates) has the lowest out-of-sample RMSE value across three levels of aggregation. The coverage of the 95% prediction intervals showed that the models with a Gaussian process (models 3, 4 and 5) outperformed those without (models 1 and 2). For all models with a Gaussian process, coverage of the prediction intervals was close to 98% in sample, and between 88 and 92% out of sample. From the results of these sensitivity analyses, we chose model 5: a two-stage model including stacked regression for the covariates.

outperformed those without (models 1 and 2). For all models with a Gaussian process, coverage of the prediction intervals was close to 98% in sample, and between 88 and 92% out of sample. From the results of these sensitivity analyses, we chose model 5: a two-stage model including stacked regression for the covariates. Additionally, to assess the impact of including surveys that do not explicitly state a 24-h recall period in questions asking about food and liquid given to a child, we considered models with and without the data included from those surveys. Supplementary Table 9 also compares the cross-validation performance of the two models: one containing only surveys that specify a 24-h recall period; and one containing all available surveys. Both in-sample and out-of-sample metrics are reasonably comparable across these two models. Since the model that includes all surveys does not produce additional bias or underestimate the degree of uncertainty (there were only eight out of 188 surveys that did not specify 24 h as a recall period), we chose to keep all surveys (Extended Data Fig. 8).

e metrics are reasonably comparable across these two models. Since the model that includes all surveys does not produce additional bias or underestimate the degree of uncertainty (there were only eight out of 188 surveys that did not specify 24 h as a recall period), we chose to keep all surveys (Extended Data Fig. 8). Post-estimation To take advantage of the extensive data gathering and analysis of GBD 20175, which in some cases included data sources outside of the scope of our geospatial modeling framework, we performed post-hoc calibration of our estimates to the GBD estimates5 (please refer to Supplementary Tables 2, 3 and 6 for the data sources used). First, each grid cell in our 5 km × 5 km grid was assigned to a GBD geography based on the location of the grid cell centroid. Then, for each country and year, we defined a raking factor that was the ratio of the GBD estimate for this geography and year to the population-weighted posterior mean EBF prevalence across all grid cells within this geography and year. Finally, this raking factor was used to scale each draw of EBF prevalence for each grid cell within the GBD geography and year. The corresponding mean raking factor across all countries was 0.96 (interquartile range: 0.82–1.08), indicating close agreement with GBD estimates. National time series plots of the post-GBD calibration final estimates (including uncertainty ranges) are presented along with the aggregated input data (classified by survey series, data type and sample size) in Extended Data Fig. 9.

0.96 (interquartile range: 0.82–1.08), indicating close agreement with GBD estimates. National time series plots of the post-GBD calibration final estimates (including uncertainty ranges) are presented along with the aggregated input data (classified by survey series, data type and sample size) in Extended Data Fig. 9. After calibration to GBD 20175, grid cell level estimates were aggregated to the second administrative subdivision, first administrative subdivision and national levels using population-weighted averages at the draw level. This was carried out for each of the 1,000 posterior draws (after calibration to GBD 20175, as described above), and then point estimates and uncertainty intervals were derived from the mean, 2.5th percentile and 97.5th percentile of these draws, respectively. In cases where an administrative subdivision did not contain the centroid of any grid cell, the nearest grid cell to it was assigned as its proxy prevalence. Since the publication of GBD 20175, recently released survey microdata (Senegal 2017, Sierra Leone 2017 and South Africa 2016) and additional survey reports (Algeria 2006, Burkina Faso 2012, Burkina Faso 2016, Mali 2016, Niger 2009, São Tomé and Príncipe 2006, and Somalia 2009) were incorporated to update GBD 2017 estimates using GBD 2017 methods5. These updated GBD estimates were used for calibrating our estimates. For additional information on the names, citations and geographic details of these surveys, see Supplementary Tables 2 and 3 (records are marked with a single asterisk).

lia 2009) were incorporated to update GBD 2017 estimates using GBD 2017 methods5. These updated GBD estimates were used for calibrating our estimates. For additional information on the names, citations and geographic details of these surveys, see Supplementary Tables 2 and 3 (records are marked with a single asterisk). Although our models can predict for all locations covered by available raster covariates, we applied a mask on barren areas based on Moderate Resolution Imaging Spectroradiometer satellite data51. All maps in our figures reflect administrative boundaries, land cover, lakes and population. Gray-colored grid cells represent areas with fewer than ten people per 1 km × 1 km grid cell, and were classified as ‘barren or sparsely vegetated’, or were not included in this analysis51–55. This step was intended to be useful to policy planners and data specialists. Projections We compared our estimated rates of improvement in EBF prevalence over the past 18 years with the improvements needed between 2017 and 2025 to meet WHO GNT (50% EBF prevalence)6 by performing a simple projection calculation. First, we calculated log-additive AROC at each grid cell (i) by logit-transforming our 18 years of posterior mean prevalence, previ,yearl and calculating the AROC between each pair of adjacent years starting with 2001: AROCi,yearl=previ,yearl−previ,year−1l

EBF prevalence)6 by performing a simple projection calculation. First, we calculated log-additive AROC at each grid cell (i) by logit-transforming our 18 years of posterior mean prevalence, previ,yearl and calculating the AROC between each pair of adjacent years starting with 2001: AROCi,yearl=previ,yearl−previ,year−1l We then calculated a weighted AROC for each pixel by taking a weighted average across the years, where more recent AROC were given more weight in the average. We defined the weights to be: wyear=(year−2000)γ∑20012017(year−2000)γ where γ may be chosen to give varying amounts of weight across the years. For this set of projections, we selected γ = 1, resulting in a linear weighting scheme that has been tested and vetted for use in projecting the health-related Sustainable Development Goal)56. For any grid cell, we then calculated the weighted AROC to be: AROCi=∑20012017wyearAROCi,yearl Finally, we calculated the projections by applying the weighted AROC at each grid cell to our 2017 posterior mean prevalence: Proji,2025=logit−1(previ,2017l+AROCi,j×8) We used the same process to project country- and administrative-level AROC. This projection scheme was analogous to the methods used in the GBD 2017 measurement of progress and projected attainment of health-related Sustainable Development Goals56.

Finally, we calculated the projections by applying the weighted AROC at each grid cell to our 2017 posterior mean prevalence: Proji,2025=logit−1(previ,2017l+AROCi,j×8) We used the same process to project country- and administrative-level AROC. This projection scheme was analogous to the methods used in the GBD 2017 measurement of progress and projected attainment of health-related Sustainable Development Goals56. Limitations Data availability This work should be assessed in full acknowledgment of the data and methodological limitations. Most importantly, the accuracy of our estimates is critically dependent on the quantity and quality of the underlying data. The availability of relevant data varied both spatially and temporally across Africa (Extended Data Fig. 4), and the lack of relevant data is one of the main sources of uncertainty around our estimates (as seen in Fig. 1f). We have constructed a large database of geo-located EBF prevalence data for the purposes of this analysis; nonetheless, important gaps in data coverage—both spatial and temporal—remain. More local data are necessary to monitor health outcomes and guide quality improvement efforts, and to increase the certainty of our results. Collecting local data from all communities every year would be an insurmountable task for most countries; this study aids in filling the current knowledge gap by producing estimates for areas without data collection based on learned patterns from well-surveyed areas, using the same estimation methods for all areas for comparable results across communities.

ties every year would be an insurmountable task for most countries; this study aids in filling the current knowledge gap by producing estimates for areas without data collection based on learned patterns from well-surveyed areas, using the same estimation methods for all areas for comparable results across communities. Data accuracy In addition, there are several factors related to data quality that should be acknowledged. Data in our analyses were obtained from caregivers of infants at any time point between birth and 6 months of age. Although an infant’s EBF status was based on a single time point (the 24 h preceding the survey interview), which is known to overestimate EBF practice for the full 6-month period, as infants may be fed other foods and liquids either before or after the survey, this estimation is standard practice57,58. Following the standard approach for estimating EBF based on international guidelines57,58, the proportion of infants who are exclusively breastfed for the full 6 months is calculated by estimating the prevalence of EBF for all children under 6 months of age (though EBF is known to decline with age)57. Due to the age range (0- to 5-month-old infants) relevant to the purpose of estimating EBF prevalence, our sample sizes are relatively smaller than previous efforts mapping localized estimates for health conditions, outcomes and socioeconomic indicators12,13,41,42, further contributing to the relatively large degree of uncertainty associated with our estimates.

ants) relevant to the purpose of estimating EBF prevalence, our sample sizes are relatively smaller than previous efforts mapping localized estimates for health conditions, outcomes and socioeconomic indicators12,13,41,42, further contributing to the relatively large degree of uncertainty associated with our estimates. The location information associated with the data compiled for these analyses is subject to some error. To protect respondents’ confidentiality, most surveys that collect GPS coordinates perform some type of random displacement on those coordinates before releasing data for secondary analyses. For example, GPS coordinates for DHS data are displaced by up to 2 km for urban clusters, up to 5 km for most rural clusters, and up to 10 km in a random 1% of rural clusters59. Furthermore, data associated with polygons rather than GPS coordinates were resampled so that they could be included in the geostatistical model, but this process essentially assumes that EBF prevalence is constant over the polygon. Research on scalable methods for better integration of polygon data in geostatistical models similar to those used in this analysis is currently ongoing.

ates were resampled so that they could be included in the geostatistical model, but this process essentially assumes that EBF prevalence is constant over the polygon. Research on scalable methods for better integration of polygon data in geostatistical models similar to those used in this analysis is currently ongoing. Modeling limitations With respect to the modeling strategy, the primary limitation is the difficulty in assessing model performance at the grid cell level. We used cross-validation to assess model performance but, due to the substantial impact of sampling error on estimates derived from single survey clusters, it was necessary to aggregate both the data and predictions when assessing error. Additionally, while we attempted to propagate uncertainty from various sources through the different modeling stages, there are some sources of uncertainty that have not been propagated. In particular, it was not computationally feasible to propagate uncertainty from the submodels in stacking through the geostatistical model. Similarly, although the WorldPop population raster is also composed of estimates associated with some uncertainty, this uncertainty is difficult to quantify and not currently reported, and so we were unable to propagate this uncertainty into our estimates of EBF prevalence for administrative subdivisions that were created using population-weighted averages of grid cell estimates.

posed of estimates associated with some uncertainty, this uncertainty is difficult to quantify and not currently reported, and so we were unable to propagate this uncertainty into our estimates of EBF prevalence for administrative subdivisions that were created using population-weighted averages of grid cell estimates. Model fitting was carried out using an integrated nested Laplace approximation to the posterior distribution, as implemented in the R-INLA package49. Prediction from fitted models was subsequently carried out using the inla.posterior.sample() function, which generates samples from the approximated posterior of the fitted model. Both model fitting and prediction thus require approximations, and these approximations may introduce error. While it is difficult to assess the impact of these approximations in this particular use case, our validation analyses found that our final model has low bias and good coverage of the 95% prediction intervals, which provides some reassurance that the approximation method used, as well as other potential sources of error, are not resulting in appreciable bias or poorly described uncertainty in our reported estimates. Furthermore, our projection methods are derived from the previous spatiotemporal historical trends and based on the assumption that recent trends will continue; thus, we are not projecting underlying drivers (such as increasing urbanization or changes in population)60. Reporting Summary Further information on research design is available in the Nature Research Reporting Summary linked to this article.

Furthermore, our projection methods are derived from the previous spatiotemporal historical trends and based on the assumption that recent trends will continue; thus, we are not projecting underlying drivers (such as increasing urbanization or changes in population)60. Reporting Summary Further information on research design is available in the Nature Research Reporting Summary linked to this article. Online content Any methods, additional references, Nature Research reporting summaries, source data, statements of code and data availability and associated accession codes are available at 10.1038/s41591-019-0525-0. Supplementary information Supplementary Information Supplementary Tables 1–12 Reporting Summary Extended data Extended Data Fig. 1 Comparison of diarrhea prevalence in children under 5 years and EBF prevalence by area. Overlapping population-weighted tertiles of diarrhea prevalence (in children under 5 years)12 and EBF prevalence (in children 0–5 months) in 2000, 2005, 2010 and 2015. Cut-offs for the tertiles were 20.8 and 36.1% for the EBF prevalence axis, and 3.6 and 5.0% for the diarrhea prevalence axis. Maps reflect administrative boundaries, land cover, lakes and population; gray-colored grid cells had fewer than ten people per 1 km × 1 km grid cell and were classified as ‘barren or sparsely vegetated’, or were not included in this analysis.

the EBF prevalence axis, and 3.6 and 5.0% for the diarrhea prevalence axis. Maps reflect administrative boundaries, land cover, lakes and population; gray-colored grid cells had fewer than ten people per 1 km × 1 km grid cell and were classified as ‘barren or sparsely vegetated’, or were not included in this analysis. Extended Data Fig. 2 Comparison of risk of death (age 0–11 months) and EBF prevalence by area. Overlapping population-weighted tertiles of mortality risk (in children 0–11 months)13 and EBF prevalence (in children 0–5 months) in 2000, 2005, 2010 and 2015. Cut-offs for the tertiles were 20.8 and 36.1% for the EBF prevalence axis, and 4.3 and 6.4% for the risk of death axis. Maps reflect administrative boundaries, land cover, lakes and population; gray-colored grid cells had fewer than ten people per 1 km × 1 km grid cell and were classified as ‘barren or sparsely vegetated’, or were not included in this analysis.

1% for the EBF prevalence axis, and 4.3 and 6.4% for the risk of death axis. Maps reflect administrative boundaries, land cover, lakes and population; gray-colored grid cells had fewer than ten people per 1 km × 1 km grid cell and were classified as ‘barren or sparsely vegetated’, or were not included in this analysis. Extended Data Fig. 3 Analytic process overview and map of modeling regions. a, The process used to produce EBF prevalence estimates in Africa involved three main parts. In the data-processing steps (peach), data were identified, extracted and prepared for use in the model. In the modeling phase (orange), we used these data and covariates in a stacked generalization ensemble model and spatiotemporal Gaussian process model. In post-processing (red), we calibrated the prevalence estimates to match GBD 20175 estimates and aggregated the estimates to the first- and second-level administrative subdivisions in each country. b, Modeling regions were defined as the five GBD regions for Africa: central (central SSA), east (eastern SSA), north (North Africa and the Middle East), south (southern SSA) and west (western SSA). As this study was limited to mainland Africa and African island nations (except Mauritius, Seychelles, Cape Verde Islands, Libya and Djibouti, where relevant data were not available or did not meet our inclusion and exclusion criteria), Middle East countries were excluded (Afghanistan, Bahrain, Iran, Iraq, Jordan, Kuwait, Lebanon, Oman, Palestine, Qatar, Saudi Arabia, Syria, Turkey, United Arab Emirates and Yemen).

les, Cape Verde Islands, Libya and Djibouti, where relevant data were not available or did not meet our inclusion and exclusion criteria), Middle East countries were excluded (Afghanistan, Bahrain, Iran, Iraq, Jordan, Kuwait, Lebanon, Oman, Palestine, Qatar, Saudi Arabia, Syria, Turkey, United Arab Emirates and Yemen). Extended Data Fig. 4 Data availability for EBF among infants under 6 months by type and country, 1998–2017. a, EBF data used in this study, by region and country. Color indicates the data source: DHS; MICS; or other survey type. Shape type indicates whether a data source has point (GPS) or polygon (for example, aggregated to an administrative level) location information. Size indicates the relative effective sample size for each source. A full list of data sources, with additional details about data type (such as survey microdata and survey reports) and geographical details, is provided in Supplementary Tables 2 and 3. b, Maps of EBF data coverage displayed at 5-year intervals. Maps show the spatial resolution of the underlying data in our models, and the color indicates the EBF prevalence as estimated from the data sources. Countries in white have no available survey data in the given time range.

ded in Supplementary Tables 2 and 3. b, Maps of EBF data coverage displayed at 5-year intervals. Maps show the spatial resolution of the underlying data in our models, and the color indicates the EBF prevalence as estimated from the data sources. Countries in white have no available survey data in the given time range. Extended Data Fig. 5 Map of covariates. Covariate raster layers of possible socioeconomic, environmental and health-related covariates used as inputs for the stacking modeling process. Time-varying covariates are presented for the year 2017. For additional detail on the year of production of non-time-varying covariates, see the individual covariate citation in Supplementary Table 7. Maps reflect administrative boundaries, land cover, lakes and population; gray-colored grid cells had fewer than ten people per 1 km × 1 km grid cell and were classified as ‘barren or sparsely vegetated’, or were not included in this analysis. Extended Data Fig. 6 Maps of in-sample predictions from the ensemble covariate modeling process. a–c, Each map represents the in-sample predicted prevalence of EBF generated from the three submodels: (a) a generalized additive model; (b) a boosted regression trees model; and (c) lasso regression, for 2000. Maps reflect administrative boundaries, land cover, lakes and population; gray-colored grid cells had fewer than ten people per 1 km × 1 km grid cell and were classified as ‘barren or sparsely vegetated’, or were not included in this analysis.

l; (b) a boosted regression trees model; and (c) lasso regression, for 2000. Maps reflect administrative boundaries, land cover, lakes and population; gray-colored grid cells had fewer than ten people per 1 km × 1 km grid cell and were classified as ‘barren or sparsely vegetated’, or were not included in this analysis. Extended Data Fig. 7 Predictions comparison from the covariate sensitivity analysis. a–e, EBF prevalence among infants under 6 months in 2017 at the 5 km × 5 km grid cell level, based on models with no covariates and including: a Gaussian process (a); raw covariates with no Gaussian process (b); raw covariates with a Gaussian process (c); stacked covariates with no Gaussian process (d) and stacked covariates with a Gaussian process (e; the final model). Estimates are shown without calibration to GBD 20175, to better highlight the differences between the models (that would have been masked after calibration). Maps reflect administrative boundaries, land cover, lakes and population; gray-colored grid cells had fewer than ten people per 1 km × 1 km grid cell and were classified as ‘barren or sparsely vegetated’, or were not included in this analysis.

ferences between the models (that would have been masked after calibration). Maps reflect administrative boundaries, land cover, lakes and population; gray-colored grid cells had fewer than ten people per 1 km × 1 km grid cell and were classified as ‘barren or sparsely vegetated’, or were not included in this analysis. Extended Data Fig. 8 Predictions comparison from the recall period sensitivity analysis. a,b, EBF prevalence among infants under 6 months in 2000 at the 5 km × 5 km grid cell level, based on models containing only surveys that specify a 24-h recall period (a) and containing all available surveys (b). Estimates are shown after calibration to GBD 20175, to better highlight the differences between the final maps when the models include all surveys or only surveys that specify a 24-h recall period. Maps reflect administrative boundaries, land cover, lakes and population; gray-colored grid cells had fewer than ten people per 1 km × 1 km grid cell and were classified as ‘barren or sparsely vegetated’, or were not included in this analysis. Extended Data Fig. 9 National time series plots and aggregated input data. National time series plots of the post-GBD calibration final estimates by country during 2000–2017. Uncertainty ranges are presented in gray, and aggregated input data are classified by survey series (purple, country specific; green, DHS; yellow, MICS), data type (square, polygon; circle, point) and whether the survey is nationally or subnationally representative. A list of subnationlly representative surveys in given in Supplementary Table 10.

nted in gray, and aggregated input data are classified by survey series (purple, country specific; green, DHS; yellow, MICS), data type (square, polygon; circle, point) and whether the survey is nationally or subnationally representative. A list of subnationlly representative surveys in given in Supplementary Table 10. Peer review information: Jennifer Sargent and Joao Monteiro were the primary editors on this article and managed its editorial process and peer review in collaboration with the rest of the editorial team. Publisher’s note: Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. These authors jointly supervised this work: Laura Dwyer-Lindgren, Simon I. Hay. Change history 8/27/2019 An amendment to this paper has been published and can be accessed via a link at the top of the paper. Extended data is available for this paper at 10.1038/s41591-019-0525-0. Supplementary information is available for this paper at 10.1038/s41591-019-0525-0. Acknowledgements This work was primarily supported by grant OPP1132415 from the Bill & Melinda Gates Foundation.

An amendment to this paper has been published and can be accessed via a link at the top of the paper. Extended data is available for this paper at 10.1038/s41591-019-0525-0. Supplementary information is available for this paper at 10.1038/s41591-019-0525-0. Acknowledgements This work was primarily supported by grant OPP1132415 from the Bill & Melinda Gates Foundation. Author contributions S.I.H. and L.D.-L. conceived and planned the study. S.J.S., J.A., W.M.G., B.V.P., C.L., D.L., E.G.P., R.R. and B.S. identified and obtained data for analysis. S.J.S., J.A., W.M.G., N.J.H., C.L., A.L.-A., B.V.P. and D.L. extracted, processed and geo-positioned the data. N.V.B. carried out the statistical analyses. D.K., A.O.-Z., N.J.H., M.A.C., J.F.M., A.D., R.B., L.P.W., J.D.V., K.E.W., R.C.R. and L.D.-L. provided input on the methods. N.V.B., L.E.S., L.B.M., J.M.R., S.J.S., J.A., W.M.G., C.S., A.S., M.F.S., B.V.P., N.J.H., K.B.J., C.L., M.A.C., K.M.S., A.L.-A., D.L., D.K., A.O.-Z., L.E., J.F.M., A.D., R.B., L.P.W., K.F.W., J.D.V., K.E.W., R.C.R., E.G.P., R.R., B.S., N.D.W., M.R.N., D.L.S., N.J.K., E.G., S.S.L., A.H.M., C.J.L.M., L.D.-L., and S.I.H. provided intellectual input into aspects of this study. N.V.B., J.A., N.J.H., C.L., M.A.C., K.M.S., A.L.-A., D.L., K.B.J. and L.E. prepared the figures and tables. N.V.B., L.E.S., L.B.M. and J.M.R. wrote the first draft of the manuscript with assistance from C.S., A.S. and M.F.S. S.J.S., W.M.G., B.V.P., N.J.H., D.K., A.O.-Z., J.F.M., A.D., L.P.W., K.F.W., J.D.V., K.E.W., R.C.R., E.G.P., R.R., B.S., N.D.W., N.J.K., L.D.-L. and S.I.H. contributed to the subsequent revisions.

ables. N.V.B., L.E.S., L.B.M. and J.M.R. wrote the first draft of the manuscript with assistance from C.S., A.S. and M.F.S. S.J.S., W.M.G., B.V.P., N.J.H., D.K., A.O.-Z., J.F.M., A.D., L.P.W., K.F.W., J.D.V., K.E.W., R.C.R., E.G.P., R.R., B.S., N.D.W., N.J.K., L.D.-L. and S.I.H. contributed to the subsequent revisions. Data availability The findings of this study are supported by data that are available in public online repositories, data that are publicly available on request from the data provider, and data that are not publicly available due to restrictions by the data provider and which were used under license for the current study (including select data sources in Botswana, Eritrea, Ghana, Kenya, South Africa and Zambia, as indicated in Supplementary Tables 2 and 10). Detailed data sources can be found in Supplementary Tables 2–6 and 10. More information about each data source is available on the GHDx (http://ghdx.healthdata.org/), including information about the data provider and links to where the data can be accessed or requested (where available).

Supplementary Tables 2 and 10). Detailed data sources can be found in Supplementary Tables 2–6 and 10. More information about each data source is available on the GHDx (http://ghdx.healthdata.org/), including information about the data provider and links to where the data can be accessed or requested (where available). Administrative boundaries were retrieved from the Global Administrative Unit Layers dataset, implemented by the FAO within the CountrySTAT and Agricultural Market Information System projects52. Land cover was retrieved from the online Data Pool, courtesy of the NASA EOSDIS Land Processes Distributed Active Archive Center, USGS/Earth Resources Observation and Science Center, Sioux Falls, South Dakota51. Lakes were retrieved from the Global Lakes and Wetlands Database, courtesy of the World Wildlife Fund and the Center for Environmental Systems Research, University of Kassel53. Populations were retrieved from WorldPop55. Outputs of these EBF analyses at national, administrative and 5 km × 5 km levels throughout Africa are publicly available at the GHDx (http://ghdx.healthdata.org/record/ihme-data/africa-exclusive-breastfeeding-prevalence-geospatial-estimates-2000-2017) and can be explored through our customized visualization tools (https://vizhub.healthdata.org/lbd/ebf).

ional, administrative and 5 km × 5 km levels throughout Africa are publicly available at the GHDx (http://ghdx.healthdata.org/record/ihme-data/africa-exclusive-breastfeeding-prevalence-geospatial-estimates-2000-2017) and can be explored through our customized visualization tools (https://vizhub.healthdata.org/lbd/ebf). EBF estimates, at various spatial levels, can be explored using custom online data visualization tools (http://vizhub.healthdata.org/lbd/ebf), and are publicly available at the GHDx (http://ghdx.healthdata.org/record/ihme-data/africa-exclusive-breastfeeding-prevalence-geospatial-estimates-2000-2017). The data that support the findings of this study are available on the GHDx; however, some of these data were used under licenses for the current study and are not publically available. All data sources are indicated in Supplementary Table 2, and data with restrictions are indicated with an obelisk symbol. Code availability Our study follows the Guidelines for Accurate and Transparent Health Estimates Reporting (Supplementary Table 1). All code used for these analyses is publicly available online at https://github.com/ihmeuw/lbd/tree/ebf-africa-2019.

EBF estimates, at various spatial levels, can be explored using custom online data visualization tools (http://vizhub.healthdata.org/lbd/ebf), and are publicly available at the GHDx (http://ghdx.healthdata.org/record/ihme-data/africa-exclusive-breastfeeding-prevalence-geospatial-estimates-2000-2017). The data that support the findings of this study are available on the GHDx; however, some of these data were used under licenses for the current study and are not publically available. All data sources are indicated in Supplementary Table 2, and data with restrictions are indicated with an obelisk symbol. Code availability Our study follows the Guidelines for Accurate and Transparent Health Estimates Reporting (Supplementary Table 1). All code used for these analyses is publicly available online at https://github.com/ihmeuw/lbd/tree/ebf-africa-2019. Competing interests This study was funded by the Bill & Melinda Gates Foundation. Co-authors employed by the Bill & Melinda Gates Foundation provided feedback on initial maps and drafts of this manuscript. Otherwise, the funders had no role in study design, data collection, data analysis, data interpretation, writing of the final report, or decision to publish. The corresponding author had full access to all of the data in the study, and had final responsibility for the decision to submit for publication.

Although genomics has shaped the current scope of precision medicine, it is becoming increasingly clear that molecular phenotypes, such as DNA and RNA profiles and, in particular, protein abundance profiles, are essential to our understanding of biology and for enhancing our ability to achieve the promise of precision medicine for patients. Hence, simultaneous generation and integration of multidimensional multi-omics datasets from a large set of tumor samples, such as those used in the National Cancer Institute’s (NCI) The Cancer Genome Atlas (TCGA; https://cancergenome.nih.gov) and the Clinical Proteomic Tumor Analysis Consortium (CPTAC; https://proteomics.cancer.gov) projects1–4, is becoming a powerful approach to understanding the molecular basis of diseases and speeding the translation of new discoveries to patient care. This development has been largely enabled by the rapid technological advancement, standardization and harmonization in tumor molecular profiling in recent years. Consequently, several initiatives have been launched to leverage this development for application to clinical practice, including the International Cancer Proteogenome Consortium5 and the Applied Proteogenomics Organizational Learning and Outcomes6 programs. These efforts promise to revolutionize our understanding of cancer biology and change the way cancer is treated.

d to leverage this development for application to clinical practice, including the International Cancer Proteogenome Consortium5 and the Applied Proteogenomics Organizational Learning and Outcomes6 programs. These efforts promise to revolutionize our understanding of cancer biology and change the way cancer is treated. The value of multi-omics technologies and datasets lies in the possibility of accurately extracting rich information to help understand the molecular complexities specific to individual patients through use of sophisticated integrative computational algorithms. Such information can be used to reach a deeper understanding of a disease, which then can be applied clinically, for example, to elucidate the relationship between the genome and proteome of a patient’s tumor or to deconvolute tumor heterogeneity associated with clinical outcome. Ideally, individual and population data would ultimately serve to inform a physician and a patient and to help determine the most appropriate treatment options. Furthermore, the comprehensive information obtained on the same sample in multiple dimensions can add value in pinpointing and correcting problems that can be encountered, such as sample mislabeling by accidental swapping of patient samples or data mislabeling (accidental swapping of patient omics data), which could lead to multiple patients receiving the wrong medical treatment, resulting in severe, irreversible consequences.

pinpointing and correcting problems that can be encountered, such as sample mislabeling by accidental swapping of patient samples or data mislabeling (accidental swapping of patient omics data), which could lead to multiple patients receiving the wrong medical treatment, resulting in severe, irreversible consequences. Sample mislabeling that contributes to irreproducible results and invalid conclusions is known to be one of the obstacles in basic and translational research7. This is also prevalent in data-rich large-scale omics studies8,9, in which human errors could arise anywhere in the data production and analysis pipeline—either sample mislabeling (early in the pipeline) or data mislabeling (later in the pipeline). The Food and Drug Administration (FDA) and NCI-CPTAC, with a history of collaboration10, also have experience in building challenges, such as the precisionFDA Challenges (https://precision.fda.gov/challenges) and NCI–CPTAC DREAM Proteogenomics Challenge (https://www.synapse.org/#!Synapse:syn8228304/wiki/413428), to solve complex problems. Now they are joining forces to launch a Multi-omics Enabled Sample Mislabeling and Correction Challenge (https://precision.fda.gov/mislabeling) in September 2018. The objective of this challenge is to encourage development and evaluation of computational algorithms that can accurately detect and correct mislabeled samples using rich multi-omics datasets, enhancing the assurance that the right data is attributed to the right patient.

cision.fda.gov/mislabeling) in September 2018. The objective of this challenge is to encourage development and evaluation of computational algorithms that can accurately detect and correct mislabeled samples using rich multi-omics datasets, enhancing the assurance that the right data is attributed to the right patient. Challenge design The challenge comprises two subchallenges to be conducted sequentially. In Subchallenge 1, participants will be asked to detect mislabeled samples. Participants will be presented with a training dataset and a test dataset, comprising real-world clinical and proteomics data. Mislabeled samples will be known in the training dataset and not known in the test dataset. Using the training dataset, participants will develop computational models to distinguish samples of matched and nonmatched clinical and proteomics data. The computational models will then be used to identify mislabeled samples in the test dataset.

ill be known in the training dataset and not known in the test dataset. Using the training dataset, participants will develop computational models to distinguish samples of matched and nonmatched clinical and proteomics data. The computational models will then be used to identify mislabeled samples in the test dataset. In Subchallenge 2, participants will be asked to correct mislabeled samples in richer data. Participants will be presented with real-world RNA profiling data for all samples in both the training and test datasets. Similar to the clinical and proteomics data, newly introduced RNA profiling data will also include mislabeled samples. As with Subchallenge 1, this information will be known in the training dataset, but not in the test dataset. Participants will develop computational algorithms to model the relationships among the three data types in the training dataset and then will apply the computational model to identify and correct instances of single data type sample mislabeling among the trio of data types in the test dataset. Subchallenge results will be independently evaluated (Fig. 1).Fig. 1 Challenge design and timelines.

s among the three data types in the training dataset and then will apply the computational model to identify and correct instances of single data type sample mislabeling among the trio of data types in the test dataset. Subchallenge results will be independently evaluated (Fig. 1).Fig. 1 Challenge design and timelines. Anticipated outcome and impact An immediate outcome envisioned is a flagship challenge manuscript that gives an overview of the challenge data, questions, design, and outcomes11. Additionally, the algorithms that the participants propose will be aggregated with the aim of refining a final open-source product to be incorporated into an analysis pipeline and ultimately as part of a quality-management system to reduce errors. This could help speed the translation of multidimensional omics technologies and datasets to the clinic. Meanwhile, NCI and FDA hope to build and expand a community of scientists that will collaborate to solve important problems that prevent the translation of multi-omics data to the clinical labs. Competing interests The authors declare no competing interests.

Main Several preclinical studies have described the structure and function of the typhoid toxin in vitro and in small-animal models2–6. Systemic administration of typhoid toxin to C57BL/6 mice results in the reproduction of many characteristic symptoms of typhoid fever; other studies have suggested that typhoid toxin may contribute to the establishment of chronic infection4,5. It remains unclear how the surrogate end points of illness in mice—such as lethargy, weight loss, behavioral and motor changes3,8—are representative of acute typhoid fever in humans. The toxin is also encoded by >40 clade B non-typhoidal Salmonella (NTS) serovars that display a broad host range and a distinct clinical phenotype to S. Typhi and Paratyphi9–11, although some typhoid toxin-expressing NTS serovars appear to cause an enteric fever-like syndrome12. Importantly, no previous studies have characterized the role of typhoid toxin in a human model of disease.

(NTS) serovars that display a broad host range and a distinct clinical phenotype to S. Typhi and Paratyphi9–11, although some typhoid toxin-expressing NTS serovars appear to cause an enteric fever-like syndrome12. Importantly, no previous studies have characterized the role of typhoid toxin in a human model of disease. We aimed to characterize the role of typhoid toxin in human infection and pathogenesis using an S. Typhi human challenge model7. This model has previously been used to test novel live-attenuated (MO1ZH09) (ref. 13) and Vi-conjugate (Typbar-TCV) typhoid vaccines14. We manufactured two challenge strains of S. Typhi to good manufacturing practice (GMP) standards. We used the wild-type S. Typhi Quailes strain (genotype 3.0.1 (ref. 15)) as the parent strain to generate an isogenic typhoid toxin-deficient knockout strain (TN), as described previously16. Whole-genome sequencing confirmed the absence of the typhoid toxin pathogenicity islet in the TN strain (Supplementary Information). The wild-type and TN strains harbored no other differences in relation to known key virulence factors (Supplementary Information). In particular, there were no differences identified in Salmonella pathogenicity island 7, a region encoding genes required for expression of the Vi-capsule—a key virulence factor in the pathogenesis of S. Typhi. Differences between strains were confined to highly variable regions encoding phage proteins, which were not known to impact on bacterial survival in the environment or persistence in the human host. However, deletion of the entire typhoid toxin pathogenicity island was associated with increased bacterial burden in a mouse model of S. Typhi infection, compared with a strain expressing a catalytic mutant of typhoid toxin (cdtBH160Q pltBS35A pltAE133A; Fig. 1). Otherwise, the wild-type and TN challenge strain variants displayed comparable phenotypic properties with regards to Vi-capsule expression, cellular invasion, in vitro growth characteristics, antibiotic susceptibility and survival in environmental water and soil samples (data not shown). Cell cycle arrest in vitro was observed with the wild-type but not the TN strain (Fig. 1).Fig. 1 Trial design. a, Schematic of trial design (Supplementary Information); comparison of wild-type and TN strains. b, Cellular invasion assessed using a gentamicin protection assay. MOI = 50, n = 3 independent replicates, two-sided Mann–Whitney U-test.

was observed with the wild-type but not the TN strain (Fig. 1).Fig. 1 Trial design. a, Schematic of trial design (Supplementary Information); comparison of wild-type and TN strains. b, Cellular invasion assessed using a gentamicin protection assay. MOI = 50, n = 3 independent replicates, two-sided Mann–Whitney U-test. c, Induction of cell cycle arrest in Henle-407 cells infected with the wild-type or TN strain (MOI = 50). Intoxicated cells show a larger proportion of cells in the G2/M phase of the cell cycle. n = 3 independent replicates. d, Comparison of the bacterial loads of S. Typhi Quailes typhoid toxin-null mutant (ST-ΔTT, ∆pltB, ∆pltA, ∆cdtB) with S. Typhi Quailes typhoid toxin catalytic mutant (ST-CM, cdtBH160Q, pltBS35A, pltAE133A). n = 28, Wilcoxon signed-rank test. The box plots display the median and IQR, with the upper whiskers extending to the largest value ≤1.5 × IQR from 75th percentile and the lower whiskers extending to smallest values ≤1.5 × IQR from 25th percentile. The Fig. 1a images were sourced from Servier Medical Art and reproduced and adapted under a Creative Commons 3.0 unported license29. Source data

c, Induction of cell cycle arrest in Henle-407 cells infected with the wild-type or TN strain (MOI = 50). Intoxicated cells show a larger proportion of cells in the G2/M phase of the cell cycle. n = 3 independent replicates. d, Comparison of the bacterial loads of S. Typhi Quailes typhoid toxin-null mutant (ST-ΔTT, ∆pltB, ∆pltA, ∆cdtB) with S. Typhi Quailes typhoid toxin catalytic mutant (ST-CM, cdtBH160Q, pltBS35A, pltAE133A). n = 28, Wilcoxon signed-rank test. The box plots display the median and IQR, with the upper whiskers extending to the largest value ≤1.5 × IQR from 75th percentile and the lower whiskers extending to smallest values ≤1.5 × IQR from 25th percentile. The Fig. 1a images were sourced from Servier Medical Art and reproduced and adapted under a Creative Commons 3.0 unported license29. Source data We enrolled a total of 41 healthy adults (aged 18–60 years) into a randomized, double-blind, human challenge study between 10 April and 1 August 2017. One volunteer withdrew prior to challenge, and 40 completed the challenge protocol (Extended Data Fig. 1). The study was undertaken in a cohort of healthy adult volunteers in a setting non-endemic for typhoid fever (Oxford14; see the Life Sciences Reporting Summary). Groups were well matched at baseline (Extended Data Fig. 2 and Supplementary Data). Participants fasted for 90 min before oral challenge with 1–5 × 104 colony-forming units (CFUs) of either wild-type or TN strains administered 2 min after sodium bicarbonate pretreatment (Supplementary Information). Study visits were scheduled for 12 h after challenge, and then daily for 14 d, when daily blood cultures were collected (Fig. 1) (ref. 7). Antibiotic treatment (ciprofloxacin 500 mg twice daily) was initiated at typhoid diagnosis or at day 14 for those without illness.

arbonate pretreatment (Supplementary Information). Study visits were scheduled for 12 h after challenge, and then daily for 14 d, when daily blood cultures were collected (Fig. 1) (ref. 7). Antibiotic treatment (ciprofloxacin 500 mg twice daily) was initiated at typhoid diagnosis or at day 14 for those without illness. Using the primary composite diagnostic end point of fever ≥38 °C for ≥12 h and/or S. Typhi bacteremia, we observed no significant difference in the rate of typhoid disease between participants challenged with wild-type or TN strains (15 out of 21 (71%) versus 15 out of 19 (79%); relative risk 1.11 (95% confidence interval (CI) 0.8–1.6); P = 0.58; Fig. 2 and Supplementary Data). The attack rate (ndiagnosed/nchallenged) in the wild-type group met the target range of 60–75% and was consistent with earlier studies7,13,14. There was no significant difference in the attack rate when we applied alternative diagnostic criteria (Supplementary Data). The challenge dose administered did not impact the outcome of the challenge (Fig. 2). Furthermore, there was no significant difference in time to diagnosis between wild-type and TN groups (median (interquartile range (IQR)) days to diagnosis 7.05 (5.08–8.83) versus 5.25 (5.01–6.14); P = 0.23; Fig. 2).Fig. 2 Clinical response to challenge with wild-type and TN S. Typhi. a, Time to diagnosis after challenge. Cumulative proportion of participants meeting the composite diagnostic end point defined as S. Typhi bacteremia and/or fever ≥38 C° persisting ≥12 h. Participants not meeting the diagnostic criteria for typhoid diagnosis were censored at day 14. log-rank test. b, Challenge dose administered between wild-type and TN challenge groups. Two-sided Mann–Whitney U-test, nTN = 19, nWild-type = 21. c, Challenge dose administered according to outcome and challenge strain. Mann–Whitney U-test. d, Time to first fever >38 °C. e, Fever clearance time categorized according to study group. Kaplan–Meier survival curve showing the cumulative proportion of participants with any fever >38 °C by challenge group. Participants with no recorded fever were censored at day 14. log-rank test. f, Cumulative symptom severity scores in all participants challenged30; nTN = 19, nWild-type = 21. Two-sided Mann–Whitney U-test. g, Maximum symptom severity score (day 0–21) in participants diagnosed with typhoid fever according to study group. Percentage of participants reporting one or more events, graded as mild, moderate or severe30.

m severity scores in all participants challenged30; nTN = 19, nWild-type = 21. Two-sided Mann–Whitney U-test. g, Maximum symptom severity score (day 0–21) in participants diagnosed with typhoid fever according to study group. Percentage of participants reporting one or more events, graded as mild, moderate or severe30. The box plots display the median and IQR, with the upper whiskers extending to largest value ≤1.5 × IQR from the 75th percentile and the lower whiskers extending to the smallest values ≤ 1.5 × IQR from the 25th percentile31. The overlaid violin plots illustrate the distribution of the data points and their probability density31. Source data

The box plots display the median and IQR, with the upper whiskers extending to largest value ≤1.5 × IQR from the 75th percentile and the lower whiskers extending to the smallest values ≤ 1.5 × IQR from the 25th percentile31. The overlaid violin plots illustrate the distribution of the data points and their probability density31. Source data To determine if absence of the typhoid toxin was associated with an altered disease phenotype, we compared the clinical profiles between challenge groups (Fig. 2). Five participants met the prespecified criteria for severe typhoid fever; of these, one participant was randomized to wild-type (1 out of 15; 7%) and four (4 out of 15; 27%; P = 0.3) were randomized to TN (Supplementary Data). Two serious adverse events were reported, neither of which was assessed as being related to S. Typhi challenge (Supplementary Information). The most common symptoms reported by participants who developed typhoid were headache (30 out of 30; 100%), malaise (30 out of 30; 100%), anorexia (26 out of 30; 87%) and abdominal pain (23 out of 30; 77%; Fig. 2 and Supplementary Information). Fever clearance time was comparable between wild-type and TN groups (median (IQR) hours 53.15 (23.0–87.4) versus 44.92 (12–96.6) hours; P = 0.71; Fig. 2). Laboratory abnormalities (elevated C-reactive protein, lymphopenia, neutropenia) were all consistent with the expected presentation of typhoid fever in the field17 (Extended Data Fig. 3). Overall, the clinical phenotype was comparable between groups.

rs 53.15 (23.0–87.4) versus 44.92 (12–96.6) hours; P = 0.71; Fig. 2). Laboratory abnormalities (elevated C-reactive protein, lymphopenia, neutropenia) were all consistent with the expected presentation of typhoid fever in the field17 (Extended Data Fig. 3). Overall, the clinical phenotype was comparable between groups. We next assessed if absence of typhoid toxin was associated with altered microbiological end points (Fig. 3). At least one stool culture was positive for S. Typhi in 13 out of 21 (62%) participants challenged with the wild-type strain and 11 out of 19 (58%) challenged with the TN strain. The pattern of stool shedding was comparable between groups, peaking 24–48 h after challenge, followed by a second peak in week 2 (Fig. 3) (ref. 18). There was no difference in the probability of shedding over the entire challenge period following challenge with TN compared with wild-type (odds ratio 0.64, 95% CI 0.17–2.47, P = 0.51). In a humanized mouse model, infection with a typhoid toxin-deficient strain of S. Typhi was associated with an increased bacterial burden compared with the wild-type strain19. Consistent with this observation, the duration of bacteremia was significantly longer in participants challenged with the TN strain compared with the wild-type strain (47.6 h (28.9–97.0) versus 30.3 (3.6–49.4); P ≤ 0.001; Fig. 3), although circulating quantitative colony counts did not differ (0.2 CFU ml−1 (0–21) versus 0.55 CFU ml−1 (0–3); P = 0.44; Fig. 3). We next performed a principal component analysis of disease severity, using all clinical, microbiological and laboratory measures collected during the course of the challenge study. When all participants were included in the analysis, participants diagnosed with typhoid fever clearly cluster separately from individuals who did not develop disease (Extended Data Fig. 4); however, there was no clustering of participants by challenge group, suggesting that challenge with a typhoid toxin-deficient strain of S. Typhi was associated with an indistinguishable clinical phenotype to that caused by wild-type S. Typhi (Extended Data Fig. 5).Fig. 3 Microbiological response to challenge with wild-type and TN S. Typhi. a,b, Pattern of stool shedding after TN (a) and wild-type (b) challenge. The rows correspond to individual participants. Gray squares, negative sample; brown squares, positive stool culture; white squares = no sample collected. Tx is the day of treatment initiation. c, Probability of stool shedding S.

and TN S. Typhi. a,b, Pattern of stool shedding after TN (a) and wild-type (b) challenge. The rows correspond to individual participants. Gray squares, negative sample; brown squares, positive stool culture; white squares = no sample collected. Tx is the day of treatment initiation. c, Probability of stool shedding S. Typhi over time after challenge. Samples were classified as culture-positive or culture-negative for S. Typhi and combined in mixed effects logistic regression models, as described previously18. nWild-type = 21, nTN = 19. d,e, Pattern of bacteremia after TN (d) and wild-type (e) challenge. Red squares = positive blood culture. Participants above the dotted lines did not meet the composite criteria for typhoid diagnosis. f, Quantitative blood culture at time of typhoid diagnosis. Samples with no colonies were assigned an arbitrary value corresponding to half the lower limit of detection (0.05 CFU ml−1). nWild-type = 15, nTN = 15, two-sided Mann–Whitney U-test. g,h, Kaplan–Meier survival curve showing the cumulative proportion of participants with bacteremia after challenge (g) and time to S. Typhi bacteremia (h). Participants not meeting the diagnostic criteria were censored at day 14. Cumulative proportion of participants with ongoing bacteremia were measured from time of treatment initiation to first persistently negative blood culture, according to challenge group. log-rank test. The box plots display the median and IQR, the upper whiskers extending to the largest value ≤1.5 × IQR from the 75th percentile and the lower whiskers extending to the smallest values ≤1.5 × IQR from the 25th percentile. The overlaid violin plots illustrate the distribution of the data points and their probability density31.

x plots display the median and IQR, the upper whiskers extending to the largest value ≤1.5 × IQR from the 75th percentile and the lower whiskers extending to the smallest values ≤1.5 × IQR from the 25th percentile. The overlaid violin plots illustrate the distribution of the data points and their probability density31. Source data

x plots display the median and IQR, the upper whiskers extending to the largest value ≤1.5 × IQR from the 75th percentile and the lower whiskers extending to the smallest values ≤1.5 × IQR from the 25th percentile. The overlaid violin plots illustrate the distribution of the data points and their probability density31. Source data To determine if absence of the typhoid toxin modulated host immune responses to infection, we measured T-cell and antibody-secreting cell (ASC) responses between challenge groups. Interferon-γ (IFN-γ)-producing T-cell responses to peptide pools comprising the typhoid toxin subunits CdtB, PltA and PltB were detectable in participants challenged with the wild-type strain, but not the TN strain, and peaked at day 28 post-challenge (Fig. 4). We observed a significant increase in circulating ASCs specific to the S. Typhi surface antigens O9:LPS and Hd at the time of typhoid diagnosis in both challenge groups (Extended Data Fig. 6) (ref. 13). The magnitude of the O9:LPS-antigen- and Hd-antigen-specific ASC response at typhoid diagnosis was generally greater in participants challenged with the TN strain. In particular, S. Typhi O9:LPS-specific immunoglobulin A ASC responses at the time of typhoid diagnosis were significantly increased in the TN group (Fig. 4).Fig. 4 Host response to challenge with wild-type and TN S. Typhi. a, CdtB-, PltA- and PltB-specific IFN-γ-producing PBMCs at baseline, day 14 and 28 after challenge. nWild-type = 21, nTN = 19, Wilcoxon signed-rank test for within-group comparisons on paired samples. b, Magnitude of ASC response at typhoid diagnosis. nWild-type = 11, nTN = 10; two-sided Mann–Whitney U-test. The box plots display the median and IQR. c, Plasma cytokine profiles after challenge with wild-type and TN S. Typhi. Heatmap showing log2 fold change in MFI for each cytokine (rows) and participant (columns) at time of diagnosis relative to baseline. nWild-type = 15, nTN = 15. Rows are annotated by significance of cytokine up- or downregulation relative to baseline in each challenge group (white, significant after adjustment for multiple testing; light gray, significant before adjustment; dark gray, non-significant). Two-sided moderated t-test with Benjamin–Hochberg correction. Clustering by Euclidean distance. FDR, false discovery rate. d, Volcano plots illustrate plasma cytokine up/downregulation at typhoid diagnosis in wild-type and TN challenge groups, with adjustment for baseline.

efore adjustment; dark gray, non-significant). Two-sided moderated t-test with Benjamin–Hochberg correction. Clustering by Euclidean distance. FDR, false discovery rate. d, Volcano plots illustrate plasma cytokine up/downregulation at typhoid diagnosis in wild-type and TN challenge groups, with adjustment for baseline. The size of each point reflects average abundance (log2MFI) in the plasma. nWild-type = 15, blue; nTN = 15, red. e, Principal component plot of log2 fold change in MFI relative to baseline for each participant. Ellipses are drawn with a 95% confidence level. nWild-type = 15, nTN = 15. FGFB, basic fibroblast growth factor; GM-CSF, granulocyte-macrophage colony-stimulating factor; M-CSF, macrophage colony-stimulating factor. Source data

The size of each point reflects average abundance (log2MFI) in the plasma. nWild-type = 15, blue; nTN = 15, red. e, Principal component plot of log2 fold change in MFI relative to baseline for each participant. Ellipses are drawn with a 95% confidence level. nWild-type = 15, nTN = 15. FGFB, basic fibroblast growth factor; GM-CSF, granulocyte-macrophage colony-stimulating factor; M-CSF, macrophage colony-stimulating factor. Source data We next aimed to determine if the presence or absence of typhoid toxin was associated with a distinct plasma cytokine profile, measured using a 62-plex bead-based cytokine platform (Luminex) at baseline and during acute typhoid disease (Fig. 4). At the time of typhoid diagnosis, the plasma cytokines 10 kDa interferon-gamma-induced protein (IP-10), monokine induced by interferon-gamma (MIG) and interleukin-1 receptor antagonist protein (IL-1RA) were significantly increased relative to baseline in both groups (Fig. 4). Hierarchical clustering and principal component analysis showed no separation of challenge groups by cytokine profile during acute typhoid disease (Fig. 4). Following adjustment for multiple testing, linear modeling found no cytokines to be significantly different between groups, although interleukin 8 (IL-8) was marginally downregulated in the toxin-negative but not the wild-type group (Fig. 4).

f challenge groups by cytokine profile during acute typhoid disease (Fig. 4). Following adjustment for multiple testing, linear modeling found no cytokines to be significantly different between groups, although interleukin 8 (IL-8) was marginally downregulated in the toxin-negative but not the wild-type group (Fig. 4). These data suggest that the typhoid toxin is not essential for S. Typhi infection nor the early acute presentation of typhoid fever. This study represents the first application of a typhoid human challenge model to prospectively study the role of a specific virulence factor in the pathogenesis of typhoid fever. Previous trials of live-attenuated S. Typhi vaccines have offered insights into the importance of other Salmonella genes to human disease (including aroC/aroD, htrA, phoP, phoQ, ssaV and cya)20. Overall, in this study, the clinical presentation was indistinguishable between the TN and wild-type groups. Counterintuitively, there was a trend toward a more severe disease phenotype in the TN group, including a shorter time to diagnosis, higher number of cases meeting the criteria for severe enteric fever, elevated ASC response and prolonged duration of bacteremia. These observations suggest that the typhoid toxin may have an important role in modifying host immune responses to infection.

henotype in the TN group, including a shorter time to diagnosis, higher number of cases meeting the criteria for severe enteric fever, elevated ASC response and prolonged duration of bacteremia. These observations suggest that the typhoid toxin may have an important role in modifying host immune responses to infection. These data raise questions as to the utility of targeting typhoid toxin in the development of novel therapeutics or vaccine strategies. Notably, currently utilized typhoid vaccines, including Vi capsular polysaccharide/conjugate vaccines, are capable of inducing protection despite targeting a virulence factor that is not strictly necessary for the establishment of enteric fever21–23. Antibody and T-cell responses to typhoid toxin components have been detected in patients with typhoid fever24–27. Further studies are required to correlate host responses to typhoid toxin with protection against disease and to further characterize its function in the context of natural S. Typhi infection.

c fever21–23. Antibody and T-cell responses to typhoid toxin components have been detected in patients with typhoid fever24–27. Further studies are required to correlate host responses to typhoid toxin with protection against disease and to further characterize its function in the context of natural S. Typhi infection. We acknowledge the limitations of our experimental approach. Due to ethical considerations, this model is not suited to assess the role of typhoid toxin in severe typhoid fever, including typhoid encephalopathy, which has been associated with typhoid toxin in animal models4,8. The primary diagnostic criteria minimizes risk to study participants by early treatment initiation7, but could mask differences between groups by treating self-limiting disease. The study population may not be generalizable to typhoid endemic countries, owing to differences in prior immune-priming and/or baseline genetic differences28.

ostic criteria minimizes risk to study participants by early treatment initiation7, but could mask differences between groups by treating self-limiting disease. The study population may not be generalizable to typhoid endemic countries, owing to differences in prior immune-priming and/or baseline genetic differences28. The absence of experiments showing reversion to virulence after complementation of the typhoid toxin genes in vitro is a limitation of this study. Additional studies with a strain expressing inactive components of typhoid toxin (for example, cdtBH160Q, pltBS35A, pltAE133A) could address whether deletion of typhoid toxin genes is associated with an altered phenotype beyond loss of toxicity. Importantly, the study was underpowered to detect anything other than a large effect size (80% power to detect at 72% relative reduction). Additional in vitro studies and deeper analysis of challenge samples are ongoing to further characterize the potential immunobiological role of typhoid toxin in the pathogenesis of typhoid fever. These data indicate that typhoid toxin is not essential for the development of early acute typhoid fever within the context of a controlled human infection model. These data highlight some of the benefits and challenges of studying bacterial virulence factors using controlled human infection models, in particular for the screening of potential vaccine and therapeutic targets.

r the development of early acute typhoid fever within the context of a controlled human infection model. These data highlight some of the benefits and challenges of studying bacterial virulence factors using controlled human infection models, in particular for the screening of potential vaccine and therapeutic targets. Methods Study design and participants The OVG2016/03 (TYGER) study was a randomized, double-blind, controlled human infection study comparing the response to challenge with wild-type S. Typhi with a typhoid toxin-deficient isogenic mutant strain S. Typhi (SB6000). The study was designed as an outpatient challenge study, conducted in a cohort of healthy community adult volunteers in a setting non-endemic for typhoid fever (Oxford). Challenge strains To facilitate comparisons with earlier challenge studies7,13,14 and to minimize the risk to study participants, we used the S. Typhi Quailes strain (genotype 3.0.1 (ref. 15)) as the parent strain to generate the typhoid toxin-deficient knockout strain. Deletion of typhoid toxin subunit genes was carried as described previously16.

te comparisons with earlier challenge studies7,13,14 and to minimize the risk to study participants, we used the S. Typhi Quailes strain (genotype 3.0.1 (ref. 15)) as the parent strain to generate the typhoid toxin-deficient knockout strain. Deletion of typhoid toxin subunit genes was carried as described previously16. Briefly, deletion of typhoid toxin subunit genes was carried out using the R6K-derived, suicide vector pSB890. The pSB890 plasmid cannot replicate in S. Typhi since it requires the bacteriophage λpir protein to replicate. The plasmid vector also encodes a counterselectable marker sacB, which encodes an enzyme that is lethal to bacteria when grown in the presence of sucrose. The pSB890 plasmid vector is maintained in a specially constructed strain of Escherichia coli, which encodes the bacteriophage λpir protein. This E. coli strain also carries a deletion mutation in the asd gene(∆asd), which encodes the aspartate-semialdehyde dehydrogenase required for peptidoglycan synthesis—growth of this strain will only occur in media supplemented with lysine diaminoheptanedioate.

richia coli, which encodes the bacteriophage λpir protein. This E. coli strain also carries a deletion mutation in the asd gene(∆asd), which encodes the aspartate-semialdehyde dehydrogenase required for peptidoglycan synthesis—growth of this strain will only occur in media supplemented with lysine diaminoheptanedioate. Due to the genomic organization of the typhoid toxin pathogenicity island, the cdtB gene was deleted first, followed by simultaneous deletion of the pltA and pltB genes (encoded immediately adjacent to one another) in the ∆cdtB strain. Chromosomal DNA fragments encoding sequences upstream and downstream of the target genes were expanded by PCR and cloned into the pSB890 plasmid, maintained in the E. coli ∆asd λpir strain. The plasmid vector encoding the cloned sequences was then transferred to S. Typhi by conjugation, counterselecting the donor E. coli strain by plating the transconjugants in media lacking lysine diaminoheptanedioate (counterselecting for the donor E. coli). Transconjugants of S. Typhi possessing deletions of the toxin integrated into the chromosome were identified by plating in sucrose that counterselects for the plasmid vector. Colonies were screened by PCR to identified mutants carrying the specific deletions. Challenge strains were manufactured to a GMP standard at the Walter Reed Army Institute of Research (Silver Spring) and stored as a frozen suspension in soya tryptone medium containing 10% sucrose at −80 °C before use.

Due to the genomic organization of the typhoid toxin pathogenicity island, the cdtB gene was deleted first, followed by simultaneous deletion of the pltA and pltB genes (encoded immediately adjacent to one another) in the ∆cdtB strain. Chromosomal DNA fragments encoding sequences upstream and downstream of the target genes were expanded by PCR and cloned into the pSB890 plasmid, maintained in the E. coli ∆asd λpir strain. The plasmid vector encoding the cloned sequences was then transferred to S. Typhi by conjugation, counterselecting the donor E. coli strain by plating the transconjugants in media lacking lysine diaminoheptanedioate (counterselecting for the donor E. coli). Transconjugants of S. Typhi possessing deletions of the toxin integrated into the chromosome were identified by plating in sucrose that counterselects for the plasmid vector. Colonies were screened by PCR to identified mutants carrying the specific deletions. Challenge strains were manufactured to a GMP standard at the Walter Reed Army Institute of Research (Silver Spring) and stored as a frozen suspension in soya tryptone medium containing 10% sucrose at −80 °C before use. Strain characterization Growth curves of wild-type and toxin-deficient strains of S. Typhi were performed in lysogeny broth (LB) using wild-type and TN strains. Isolates were inoculated into 10 ml of LB and grown overnight in a shaking incubator at 200–220 r.p.m. and 37 °C. The following day, cultures were vortexed and diluted 1:10 in fresh LB (100 µl culture + 900 µl LB). The OD600 was read in a cuvette using a mini photospectrometer against an LB-only blank, multiplying by the dilution factor (×10) to give the actual OD of the overnight culture. The dilution required to reduce the culture OD to 0.05 in a 30 ml volume of LB was calculated using the following equation (equation (1)): OD1×V1=OD2×V2

was read in a cuvette using a mini photospectrometer against an LB-only blank, multiplying by the dilution factor (×10) to give the actual OD of the overnight culture. The dilution required to reduce the culture OD to 0.05 in a 30 ml volume of LB was calculated using the following equation (equation (1)): OD1×V1=OD2×V2 where OD1 is the OD of the overnight culture, V1 is the volume of the overnight culture to be added to the new mix; OD2 is theOD of the new inoculum (0.05) and V2 is the volume of new inocula (30 ml). The calculated volume of overnight cultures (V1) was added to 30 ml of fresh LB, the OD600 read to ensure the OD of the new culture was 0.05 and was subsequently returned to the shaking incubator. Samples of culture were removed from the incubator at regular intervals and the OD600 was measured in a 50:50 mix of the culture and fresh LB against an LB blank. OD600 readings were multiplied by the dilution factor (×2) to give values for the undiluted culture and plotted against time to give the growth curve.

bator. Samples of culture were removed from the incubator at regular intervals and the OD600 was measured in a 50:50 mix of the culture and fresh LB against an LB blank. OD600 readings were multiplied by the dilution factor (×2) to give values for the undiluted culture and plotted against time to give the growth curve. Typhoid toxin activity Activity of the typhoid toxin was assessed using previously published methods4,6. Briefly, Henle-407 intestinal epithelial cells were infected with wild-type Quailes or typhoid toxin-deficient SB6000 S. Typhi for 1 h. Cells were washed and culture medium containing gentamicin (50 μg ml−1) was added and then incubated for 2 h. Cell were then washed and medium containing 5 μg ml−1 gentamicin was added and infection continued for 48 h. Cells were collected from dishes by trypsinization (subsequently neutralized with serum-containing media). The cell suspensions were centrifuged for 5 min at 1500 r.p.m., the supernatant discarded and cell pellets resuspended in 0.5 ml of PBS at room temperature. Cell suspensions were slowly added to tubes containing 4 ml of cold 90% ethanol solution with continuous mixing. Cells were kept in fixative for 2 h on ice. The fixed cells were collected by centrifugation and the fixative decanted thoroughly. The pellets were washed once with 5 ml PBS and the cell pellet was resuspended in 1 ml of a solution containing 0.1% Triton X-100, DNase-free ribonuclease A (20 mg ml−1) and propidium iodide (20 μg ml−1; Molecular Probes) in PBS. The stained cells were analyzed by flow cytometry with a FACStar Plus flow cytometer (BD Biosciences). Intoxicated cells showed a larger proportion of cells in the G2/M phase of the cell cycle and thus exhibited a larger amount of DNA content.

e A (20 mg ml−1) and propidium iodide (20 μg ml−1; Molecular Probes) in PBS. The stained cells were analyzed by flow cytometry with a FACStar Plus flow cytometer (BD Biosciences). Intoxicated cells showed a larger proportion of cells in the G2/M phase of the cell cycle and thus exhibited a larger amount of DNA content. Cellular invasion assay Henle-407 intestinal epithelial cells were infected with wild-type S. Typhi Quailes or the toxin-deficient SB6000 derivative for 2 h at three multiplicity of infection (MOI) levels (50,100 and 500). Cells were washed and gentamicin (50 μg ml−1) was added to the culture medium. After 2 h, cells were washed again, lysed and colony counts of both strains were determined by plating dilutions of the cell lysates. The invasive ability was expressed as the percentage of the bacterial inoculum that survived gentamicin treatment.

ere washed and gentamicin (50 μg ml−1) was added to the culture medium. After 2 h, cells were washed again, lysed and colony counts of both strains were determined by plating dilutions of the cell lysates. The invasive ability was expressed as the percentage of the bacterial inoculum that survived gentamicin treatment. Comparison of the bacterial loads of S. Typhi Quailes ∆pltB, ∆pltA and ∆cdtB with S. Typhi Quailes cdtBH160Q, pltBS35A and pltAE133A Strains used for comparison of the bacterial loads of S. Typhi Quailes ∆pltB, ∆pltA and ∆cdtB with S. Typhi Quailes cdtBH160Q, pltBS35A and pltAE133A were derived from S. Typhi Quailes and were constructed by standard recombinant DNA techniques as described previously16. CmaH−/− bloc3−/− mice, which are susceptible to S. Typhi infection32, were intraperitoneally infected with equal numbers (105 CFUs) of S. Typhi Quailes derivative mutant strains carrying either deletions in the pltB, pltA and cdtB genes (S. Typhi Quailes ∆pltB, ∆pltA and ∆cdtB) or expressing an inactivated version of typhoid toxin by virtue of catalytic mutations in its active subunits PltA and CdtB and a mutation in the receptor-binding site of PltB (S. Typhi Quailes cdtBH160Q, pltBS35A and pltAE133A). The strains were alternatively marked by a chloramphenicol (cmR) or kanamycin (kanR) resistance genes, as indicated, inserted within the STY4607 gene, which previous studies have shown not to affect virulence32. All animal experiments were conducted in accordance with protocols approved by Yale University’s Institutional Animal Care and Use Committee. Seven-to-ten week old cmaH−/− bloc3−/− mice were injected intraperitoneally with 105 CFUs each of the two strains. The inoculum was plated to confirm the equivalent ratio of the bacterial strains. Mice were killed at day 5 post-infection and the CFUs of each strain in the spleens of infected animals were determined by plating on LB plates containing chloramphenicol (30 µg ml−1) or kanamycin (50 µg ml−1).

Us each of the two strains. The inoculum was plated to confirm the equivalent ratio of the bacterial strains. Mice were killed at day 5 post-infection and the CFUs of each strain in the spleens of infected animals were determined by plating on LB plates containing chloramphenicol (30 µg ml−1) or kanamycin (50 µg ml−1). Phenotypic characterization of the wild-type and TN strains comprised growth characteristics in liquid culture, agglutination, cellular invasion assays and cell intoxication assays4. Whole-genome sequencing using both the MiSeq (Illumina) and PacBio (Pacific Biosciences) platforms was performed by the Wellcome Sanger Institute (Hinxton). DNA for MiSeq sequencing was extracted using the Wizard Genomic DNA purification kit according to the manufacturer’s instructions33. Sequence reads were assembled using HGAP v.3 of the SMRT analysis software v.2.3.0 (Supplementary Information). Establishment of challenge dose Challenge agents were prepared in batches for a maximum of six participants at any one time. All work was performed in the containment level 3 facility at the Centre for Clinical Vaccinology and Tropical Medicine (Oxford) in a class II biological safety hood dedicated for challenge agent preparation.

of challenge dose Challenge agents were prepared in batches for a maximum of six participants at any one time. All work was performed in the containment level 3 facility at the Centre for Clinical Vaccinology and Tropical Medicine (Oxford) in a class II biological safety hood dedicated for challenge agent preparation. Two GMP master stock vials of typhoid toxin-negative S. Typhi TN strain (BPR–1218-00, lot 1977; cell concentration 1 × 106) or wild-type S. Typhi Quailes strain (BPR–1218-00, lot 1977; cell concentration 9.8 × 105) were selected at random from stocks stored in a −80 °C freezer. Vials were thawed at room temperature for approximately 10 min and mixed by vortexing. The contents of two GMP master stock vials were transferred to a master stock tube and mixed for 6–10 s by vortexing. A 1:10 dilution in sodium bicarbonate was performed by transferring 1,600 μl from the tube labeled ‘master stock’ to a fresh 50 ml falcon labeled as ‘master stock 1:10’. To create the challenge inoculum of the toxin-negative strain, 1.74 ml from the ‘master stock 1:10 dilution’ was transferred to a sterile culture flask containing 4.2 g sodium bicarbonate dissolved in 240 ml bottled mineral water (‘challenge flask’). The challenge inoculum for the wild-type strain was generated by transferring 1.85 ml from the 1:10 dilution into an equivalent challenge flask. The challenge agents were then prepared by transferring 30 ml from the challenge flask to prelabelled 50 ml falcon tubes, sealed and stored on ice.

water (‘challenge flask’). The challenge inoculum for the wild-type strain was generated by transferring 1.85 ml from the 1:10 dilution into an equivalent challenge flask. The challenge agents were then prepared by transferring 30 ml from the challenge flask to prelabelled 50 ml falcon tubes, sealed and stored on ice. The challenge dose was confirmed by pipetting 200 μl from the challenge dose onto six Tryptone Soya Agar plates (code no. PO0163A; Oxoid). The bacterial suspension was spread over the source of the agar using an L-shaped spreader and cultured in an incubator overnight at 37 °C, 5% CO2. On the following day, colonies were manually counted and checked by a second operator. The CFUs of the challenge inoculum were calculated by multiplying the mean of the CFU counts for the plates by the dilution factor of the volume plated (×150 for a total challenge inoculum of 30 ml for a plating of 0.2 ml (30/0.2 = ×150)). Sodium bicarbonate was prepared by dissolving 2.1 g sodium bicarbonate in 120 ml bottled mineral water. Participant characteristics Healthy adults aged 18–60 years, without prior residency in an enteric fever endemic country for ≥6 months, were considered eligible for enrollment. Key exclusion criteria included significant medical, surgical or psychiatric history and gallbladder disease. A full description of the inclusion and exclusion criteria is provided in the Nature Research Reporting Summary.

dency in an enteric fever endemic country for ≥6 months, were considered eligible for enrollment. Key exclusion criteria included significant medical, surgical or psychiatric history and gallbladder disease. A full description of the inclusion and exclusion criteria is provided in the Nature Research Reporting Summary. Randomization and masking Participants were randomized 1:1 to challenge with either wild-type strain S. Typhi or toxin-negative strain S. Typhi (TN) in varying block sizes. Anti-Vi IgG was measured at screening using a commercial ELISA kit (VaccZyme; The Binding Site Ltd) according to the manufacturer’s instructions14. Randomization was stratified by anti-Vi IgG measured (low (<7.4 EU ml−1) or high (≥7.4 EU ml−1)). The exception was a sentinel group of two participants who were randomized 1:1 to receive the wild-type strain or TN knockout strain using a block size of two. Randomization was performed at the prechallenge visit, one week before challenge. We generated a randomization list in STATA v.14.2 (StataCorp), which was implemented in the computerized randomization software Sortition (Nuffield Department of Primary Care, Clinical Trials Unit, University of Oxford), which matched a masked allocation group to each participant. The software generated a randomization number, corresponding to the challenge allocation group. A locked, challenge agent randomization allocation list was maintained by the study statistician and unblinded laboratory team responsible for challenge agent preparation.

tched a masked allocation group to each participant. The software generated a randomization number, corresponding to the challenge allocation group. A locked, challenge agent randomization allocation list was maintained by the study statistician and unblinded laboratory team responsible for challenge agent preparation. The study was conducted double-blind from the time of randomization until participant unblinding, such that participants, and clinical or laboratory staff undertaking follow-up procedures, were unaware of challenge agent allocation. Both wild-type and TN strains were prepared suspended in sodium bicarbonate and had an indistinguishable appearance (transparent, colorless liquid). Procedures Participants fasted for 90 min before challenge. Two minutes before challenge, participants drank a sodium bicarbonate solution (2.1 g 120 ml−1) to neutralize stomach acid. The oral challenge inoculum was administered suspended in sodium bicarbonate (0.53 g 30 ml−1) and was kept on ice before administration within 3 h of preparation. Participants were observed for 90 min post-challenge. The challenge dose administered was 1–5 × 104 CFUs calculated as described previously7,13,14. Participants attended the clinical site 12 h after challenge and then daily for 14 d, as described previously7. Daily visits comprised continued consent check, oral temperature measurement, heart rate and blood pressure measurement and sample collection, as outlined in the study protocol.

s described previously7,13,14. Participants attended the clinical site 12 h after challenge and then daily for 14 d, as described previously7. Daily visits comprised continued consent check, oral temperature measurement, heart rate and blood pressure measurement and sample collection, as outlined in the study protocol. Solicited symptoms and twice-daily temperature measurements were recorded in an electronic diary for 21 d after challenge. Symptoms were categorized as not present, mild, moderate or severe (Supplementary Information). Antibiotic treatment was initiated on fulfillment of composite diagnostic criteria or at day 14 for those without illness. First-line treatment was oral ciprofloxacin 500 mg twice daily for 14 d. Outcomes The primary objective of this study was to compare the proportion of participants meeting the composite diagnostic end point for typhoid fever (attack rate) following oral challenge with (1–5) × 104 CFUs wild-type S. Typhi Quailes strain, compared to challenge with (1–5) × 104 CFUs of a typhoid toxin-deficient isogenic mutant of S. Typhi Quailes strain SB6000 (TN). The composite diagnostic end point for typhoid fever was defined as a temperature ≥38 °C persisting for ≥12 h and/or S. Typhi bacteremia collected ≥72 h after oral challenge.

Quailes strain, compared to challenge with (1–5) × 104 CFUs of a typhoid toxin-deficient isogenic mutant of S. Typhi Quailes strain SB6000 (TN). The composite diagnostic end point for typhoid fever was defined as a temperature ≥38 °C persisting for ≥12 h and/or S. Typhi bacteremia collected ≥72 h after oral challenge. Secondary end points were: mode of diagnosis; time to typhoid diagnosis; time to first temperature ≥38 °C; fever clearance time; time to bacteremia; duration of bacteremia; and quantitative blood culture (for definitions, see the Life Sciences Reporting Summary). Descriptive end points included: severe adverse events; solicited symptom profiles; proportion of participants meeting the criteria for severe enteric fever; hematological and biochemical measures; plasma cytokine profiles; pattern of bacteremia; and pattern of stool shedding (see Life Sciences Reporting Summary). Stool samples for culture, blood samples for culture (10 ml), and hematological and biochemical testing were processed by the local hospital’s accredited pathology laboratory as described previously7. Criteria for severe enteric fever Severe enteric fever was defined as participants meeting any of the following criteria: oral temperature >40 °C; systolic blood pressure <85 mmHg; significant lethargy or confusion; gastrointestinal bleeding; gastrointestinal perforation; or any grade 4 laboratory abnormality34.

Stool samples for culture, blood samples for culture (10 ml), and hematological and biochemical testing were processed by the local hospital’s accredited pathology laboratory as described previously7. Criteria for severe enteric fever Severe enteric fever was defined as participants meeting any of the following criteria: oral temperature >40 °C; systolic blood pressure <85 mmHg; significant lethargy or confusion; gastrointestinal bleeding; gastrointestinal perforation; or any grade 4 laboratory abnormality34. Ex vivo ASC enzyme-linked immune absorbent spot (ELISpot) Ex vivo IgG-, IgA- and IgM-producing ASC responses against O- and H-antigen were measured at baseline and 24–48 h after typhoid diagnosis in those meeting the diagnostic criteria as described previously13.

Criteria for severe enteric fever Severe enteric fever was defined as participants meeting any of the following criteria: oral temperature >40 °C; systolic blood pressure <85 mmHg; significant lethargy or confusion; gastrointestinal bleeding; gastrointestinal perforation; or any grade 4 laboratory abnormality34. Ex vivo ASC enzyme-linked immune absorbent spot (ELISpot) Ex vivo IgG-, IgA- and IgM-producing ASC responses against O- and H-antigen were measured at baseline and 24–48 h after typhoid diagnosis in those meeting the diagnostic criteria as described previously13. Multiscreen filtration ELISpot plates (catalog no. MAHAS4510; Merck Millipore) were coated with S. Typhi O9:LPS, S. Typhi Hd antigen (University of Maryland) and Pan goat anti-human immunoglobulin (catalog no. H17000; Invitrogen) each at a final concentration of 10 μg ml−1 in carbonate-bicarbonate buffer and incubated overnight at 4 °C. Plates were blocked with 200 μl per well of R10 medium for 1 h before use at 37 °C, 5% CO2. Peripheral blood mononuclear cells (PBMCs) were separated using ACCUSPIN tubes (Sigma-Aldrich), counted and resuspended in R10 media. PBMCs at a concentration of 2.5 × 105 were added in duplicate to the ELISpot plate (100 μl per well) and incubated overnight at 37 °C, 5% CO2. Plates were washed four times with PBS-0.25% Tween, once with PBS and soaked with PBS for 5 min. Goat anti-human IgG, IgA and IgM secondary antibodies conjugated to alkaline phosphatase (catalog nos. 401442, 401132 and 401902, respectively; Sigma-Aldrich) were diluted to 1:5,000 in PBS/FBS and incubated for 4 h at room temperature. After incubation, the plates were washed five times with PBS-0.25% Tween and four times with dsH2O. Alkaline phosphatase substrate (catalog no. 170-6432; Bio-Rad) was added at 50 μl per well, allowed to develop over approximately 10 min and stopped with dsH2O as spots began to develop.

ated for 4 h at room temperature. After incubation, the plates were washed five times with PBS-0.25% Tween and four times with dsH2O. Alkaline phosphatase substrate (catalog no. 170-6432; Bio-Rad) was added at 50 μl per well, allowed to develop over approximately 10 min and stopped with dsH2O as spots began to develop. ELISpot plates were read using an automated ELISpot reader (ELR03/ELR030408215; Autoimmun Diagnostika) and the AID ELISpot software v.5.0. Study- and antigen-specific count settings for spot intensity, size and gradient were applied to the plate counts and manually verified to remove artifacts. Raw counts (spots per 2.5 × 105 PBMCs) were averaged across duplicate wells and multiplied by four to give the number of spot-forming units (SFUs) per 106 PBMCs.

. Study- and antigen-specific count settings for spot intensity, size and gradient were applied to the plate counts and manually verified to remove artifacts. Raw counts (spots per 2.5 × 105 PBMCs) were averaged across duplicate wells and multiplied by four to give the number of spot-forming units (SFUs) per 106 PBMCs. Fluorospots Measurements were taken from frozen PBMCs collected at baseline, and on day 14 and 28 post-challenge. Precoated plates (catalog no. FSP-010308-10; Mabtech) were blocked before adding 50 µl per well toxin peptide pools consisting of 15-mer sequences with 11-amino acid overlaps and covering the sequence of proteins CdtB, PltA and PltB (thinkpeptides). The peptides were dissolved in 100% DMSO (Sigma-Aldrich) and arranged in three pools. Concentration was adjusted at 0.6 mg ml−1 and used in the fluorospot assay at a final concentration of 3 µg ml−1 of each peptide. DMSO and concanavalin A (Sigma-Aldrich) were used as negative and positive controls, respectively. After defrosting and resting for 1 h, 50 µl per well of PBMCs were added to the peptide wells at a concentration of 4 × 106 cells ml−1 in triplicate and incubated overnight at 37 °C, 5% CO2, 95% humidity. Detection of spots was carried out according to the manufacturer’s instructions (Mabtech) and analyzed with the iSpot EliSpot reader (Autoimmun Diagnostika).

l per well of PBMCs were added to the peptide wells at a concentration of 4 × 106 cells ml−1 in triplicate and incubated overnight at 37 °C, 5% CO2, 95% humidity. Detection of spots was carried out according to the manufacturer’s instructions (Mabtech) and analyzed with the iSpot EliSpot reader (Autoimmun Diagnostika). Plasma cytokine analysis Plasma was isolated from heparinized blood by centrifugation. Protease inhibitor was added in a 1:40 dilution before storage at −80 °C. Longitudinal cytokine quantification was carried out for all 40 challenged participants by the Human Immune Monitoring Center at Stanford University using a 62-plex Luminex system (brain-derived neurotrophic factor, beta-nerve growth factor, CD40 ligand, epidermal growth factor, ENA-78 (CXCL5), eotaxin, fibroblast growth factor 2, granulocyte-colony-stimulating factor, granulocyte-macrophage colony-stimulating factor, GRO-α (CXCL1), hepatocyte growth factor, IFN-α, IFN-β, IFN-γ, IL-1-α, IL-1-β, IL-10, IL-12p40, IL-12p70, IL-13, IL-15, IL-17A, IL-17F, IL-18, IL-1RA, IL-2, IL-21, IL-22, IL-23, IL-27, IL-31, IL-4, IL-5, IL-6, IL-7, IL-8, IL-9, IP-10, leptin, leukemia inhibitory factor, macrophage colony-stimulating factor 1, monocyte chemoattractant protein 1 (MCP-1), MCP-3, MIG, macrophage inflammatory protein-1-α (MIP-1-α), MIP-1-β, plasminogen activator inhibitor 1, platelet-derived growth factor subunit B, RANTES, resistin, stem cell factor, stromal cell-derived factor 1, soluble Fas ligand, soluble intercellular adhesion molecule 1, soluble vascular cell adhesion protein 1, transforming growth factor-α (TGF-α), TGF-β, tumor necrosis factor-α (TNF-α), TNF-β, tumor necrosis factor ligand superfamily member 10, vascular endothelial growth factor A (VEGF-A) and VEGF-D). Samples were run in duplicate and the mean fluorescence intensity (MFI) of duplicates was used for analysis. To minimize plate-to-plate variation, samples across time points for each individual were run on the same plates and each plate contained an equal mix of individuals allocated to wild-type or toxin-negative challenge. Control beads (CHEX 1–4) and control sera were used per plate.

I) of duplicates was used for analysis. To minimize plate-to-plate variation, samples across time points for each individual were run on the same plates and each plate contained an equal mix of individuals allocated to wild-type or toxin-negative challenge. Control beads (CHEX 1–4) and control sera were used per plate. The MFI of all samples were examined by principal component analysis to confirm consistency between duplicates and identify outliers. One participant was excluded on this basis. Duplicates were then averaged and the MFI quantile normalized. Significance testing was performed using linear modeling in limma v3.34.9 (ref. 35), incorporating plate, dose and sex as covariates. P values were corrected for multiple testing using the Benjamin–Hochberg correction. Hierarchal clustering was carried out based on Euclidean distance.

d and the MFI quantile normalized. Significance testing was performed using linear modeling in limma v3.34.9 (ref. 35), incorporating plate, dose and sex as covariates. P values were corrected for multiple testing using the Benjamin–Hochberg correction. Hierarchal clustering was carried out based on Euclidean distance. Sample size The sample size was dictated primarily by the number of participants that could be feasibly enrolled within the time frame and budget of the study; therefore, it represents a convenience sample. Assuming typhoid toxin is central to the clinical presentation of acute typhoid fever, it was anticipated that the attack rate following challenge with the TN strain would be reduced compared with the wild-type strain, although the effect size was unknown. Assuming an attack rate of 65% following wild-type challenge (as observed in previous studies) and 50% attack rate following TN challenge, and accounting for a 10% dropout, 20 participants in each group had 95% CIs for attack rate of 41–85% in the wild-type group and 27–73% in the TN group. Twenty participants per arm provided 95% power to detect an absolute reduction in attack rate of 55% (65% with the wild-type strain versus 10% with the TN strain, corresponding to an 85% relative risk reduction) and 80% power to detect an absolute reduction in attack rate of 47% (65% with the S. Typhi wild-type strain versus 18% with the S. Typhi toxin-negative strain, corresponding to a 72% relative risk reduction) based on Fisher’s exact test with 5% alpha.

with the TN strain, corresponding to an 85% relative risk reduction) and 80% power to detect an absolute reduction in attack rate of 47% (65% with the S. Typhi wild-type strain versus 18% with the S. Typhi toxin-negative strain, corresponding to a 72% relative risk reduction) based on Fisher’s exact test with 5% alpha. Statistical considerations Attack rates and 95% CIs were calculated for each challenge group for the per-protocol population (that is, participants who completed the 14-d challenge period) as the primary end point. All participants were included in the analyses if they were successfully challenged on day 0 and had at least one post-challenge assessment. The difference in attack rate (and other categorical variables) between naïve and rechallenge groups was tested using Fisher’s exact test. Time-to-event data were summarized using the Kaplan–Meier method, with participants censored at day 14. Group comparisons were performed using a log-rank test. Continuous variables were compared using the Mann–Whitney U-test for unpaired samples and the Wilcoxon signed-rank test for paired samples. All statistical tests were two-sided.

a were summarized using the Kaplan–Meier method, with participants censored at day 14. Group comparisons were performed using a log-rank test. Continuous variables were compared using the Mann–Whitney U-test for unpaired samples and the Wilcoxon signed-rank test for paired samples. All statistical tests were two-sided. Paired samples across time points were compared using the Wilcoxon signed-rank test. Comparisons between groups were performed using the Mann–Whitney U-test. ELISpot/FluoroSpot data were log10-transformed to approximate a normal distribution; wells with no spots were assigned an arbitrary value of 0.5, corresponding to half the lower limit of detection. Raw counts were averaged across replicate wells. The number of background spots detected in blank wells were subtracted from the test samples to give the final cell count per sample. Clinical data were recorded on a web-based database (OpenClinica Enterprise v3.13). Symptom and ELISpot were extracted using Microsoft Excel. Data analysis was performed using R v.3.4.4. Variables were normalized by z-score before inclusion in the principal component analysis, which was performed using the FactoMineR package v1.41 (ref. 36).

rded on a web-based database (OpenClinica Enterprise v3.13). Symptom and ELISpot were extracted using Microsoft Excel. Data analysis was performed using R v.3.4.4. Variables were normalized by z-score before inclusion in the principal component analysis, which was performed using the FactoMineR package v1.41 (ref. 36). Approvals The OVG2016/03 study was sponsored by the University of Oxford (Clinical Trials & Research Governance). Ethical approvals for the primary protocol, and any study amendments, were obtained from the South Central-Oxford A Research Ethics Committee (16/SC/0358). In the UK, legislation governing the deliberate release of genetically modified organisms is currently provided by the Environmental Protection Act 1990, sections 111 and 112 (ref. 37), and the Genetically Modified Organisms (Deliberate Release) Regulations 2002 (ref. 38). Approvals for deliberate release of the genetically modified strain of S. Typhi were obtained from the United Kingdom Department for Environment, Food & Rural Affairs (16/R48/01) (ref. 39). The study was registered with clinicaltrials.gov (NCT03067961) and was performed according to the provisions of the Declaration of Helsinki (2013) and Good Clinical Practice guidelines. This work is licensed under the Creative Commons Attribution 4.0 International License. Reporting Summary Further information on research design is available in the Nature Research Reporting Summary linked to this article.

Approvals The OVG2016/03 study was sponsored by the University of Oxford (Clinical Trials & Research Governance). Ethical approvals for the primary protocol, and any study amendments, were obtained from the South Central-Oxford A Research Ethics Committee (16/SC/0358). In the UK, legislation governing the deliberate release of genetically modified organisms is currently provided by the Environmental Protection Act 1990, sections 111 and 112 (ref. 37), and the Genetically Modified Organisms (Deliberate Release) Regulations 2002 (ref. 38). Approvals for deliberate release of the genetically modified strain of S. Typhi were obtained from the United Kingdom Department for Environment, Food & Rural Affairs (16/R48/01) (ref. 39). The study was registered with clinicaltrials.gov (NCT03067961) and was performed according to the provisions of the Declaration of Helsinki (2013) and Good Clinical Practice guidelines. This work is licensed under the Creative Commons Attribution 4.0 International License. Reporting Summary Further information on research design is available in the Nature Research Reporting Summary linked to this article. Online content Any methods, additional references, Nature Research reporting summaries, source data, statements of code and data availability and associated accession codes are available at 10.1038/s41591-019-0505-4. Supplementary information Supplementary Information Supplementsary Method Tables 1–3, Supplementary Figs. 1 and 2; Details of strain genotyping Reporting summary Supplementary Tables Supplementary Data Tables 1–6 Study Protocol Study protocol Extended data

Online content Any methods, additional references, Nature Research reporting summaries, source data, statements of code and data availability and associated accession codes are available at 10.1038/s41591-019-0505-4. Supplementary information Supplementary Information Supplementsary Method Tables 1–3, Supplementary Figs. 1 and 2; Details of strain genotyping Reporting summary Supplementary Tables Supplementary Data Tables 1–6 Study Protocol Study protocol Extended data Extended Data Fig. 1 Trial profile. One participant randomized to wild-type S. Typhi withdrew before challenge and was excluded from all further analyses, leaving 40 participants in the per-protocol analysis. Extended Data Fig. 2 Baseline anti-Vi IgG. S. Typhi Quailes strain (n = 21). Two-sided Mann–Whitney U-test. The box plots represent the median and IQR. The overlaid violin plots illustrate the distribution of the data points and their probability density. Samples below the lower limit of detection of the ELISA (7.4 EU ml−1) were assigned a value equating to half the lower limit of detection (3.7 EU ml−1)14. The box plots display the median and IQR, with the upper whiskers extending to the largest value ≤1.5 × IQR from the 75th percentile and the lower whiskers extending to the smallest values ≤1.5 × IQR from the 25th percentile. The overlaid violin plots illustrate the distribution of the data points and their probability density31. TN S. Typhi (n = 19). Source data

Extended Data Fig. 2 Baseline anti-Vi IgG. S. Typhi Quailes strain (n = 21). Two-sided Mann–Whitney U-test. The box plots represent the median and IQR. The overlaid violin plots illustrate the distribution of the data points and their probability density. Samples below the lower limit of detection of the ELISA (7.4 EU ml−1) were assigned a value equating to half the lower limit of detection (3.7 EU ml−1)14. The box plots display the median and IQR, with the upper whiskers extending to the largest value ≤1.5 × IQR from the 75th percentile and the lower whiskers extending to the smallest values ≤1.5 × IQR from the 25th percentile. The overlaid violin plots illustrate the distribution of the data points and their probability density31. TN S. Typhi (n = 19). Source data Extended Data Fig. 3 Hematological and biochemical laboratory measures after challenge. Participants challenged with TN (red) and wild-type (blue) S. Typhi and diagnosed with typhoid fever, nTN = 15, nWild-type = 15. Data are presented relative to the time of typhoid diagnosis (time point 0 = day of diagnosis). The dot plots corresponds to individual values colored according to challenge group (TN, red; wild-type, blue). The solid, colored lines connect the median values at each time point. The dashed lines represent the upper and lower reference limits for the individual parameters measured. The box plots display the median and IQR, with the upper whiskers extending to the largest value ≤1.5 × IQR from the 75th percentile and the lower whiskers extending to the smallest values ≤1.5 × IQR from the 25th percentile31, and are colored according to challenge group.

limits for the individual parameters measured. The box plots display the median and IQR, with the upper whiskers extending to the largest value ≤1.5 × IQR from the 75th percentile and the lower whiskers extending to the smallest values ≤1.5 × IQR from the 25th percentile31, and are colored according to challenge group. Source data Extended Data Fig. 4 Principal component analysis (PCA) biplot of disease severity by outcome of challenge. Participants challenged with TN (n = 19) or wild-type (n = 21) S. Typhi based on the typhoid severity score. The data points are colored by outcome of challenge (yellow, typhoid diagnosis; blue, no typhoid diagnosis). The ellipses represent the 95% confidence levels for diagnosis status. The biplot arrows represent the contributions of individual variables to a given principal component, scaled according to their relative contribution36. Source data Extended Data Fig. 5 PCA biplot of disease severity by challenge agent. PCA biplot of participants challenged with typhoid fever following challenge with TN (n = 19) or wild-type (n = 21) S. Typhi based on all clinical, microbiological and laboratory measures. The data points are colored by challenge agent allocation. The ellipses represent the 95% confidence levels for challenge agent. The biplot arrows represent the contributions of individual variables to a given principal component, scaled according to their relative contribution36. Source data

Extended Data Fig. 5 PCA biplot of disease severity by challenge agent. PCA biplot of participants challenged with typhoid fever following challenge with TN (n = 19) or wild-type (n = 21) S. Typhi based on all clinical, microbiological and laboratory measures. The data points are colored by challenge agent allocation. The ellipses represent the 95% confidence levels for challenge agent. The biplot arrows represent the contributions of individual variables to a given principal component, scaled according to their relative contribution36. Source data Extended Data Fig. 6 09:LPS- and Hd-specific IgG, IgA and IgM ex vivo ASC responses. a, 09:LPS. b, Hd. Participants diagnosed with typhoid, illustrated as log10 SFUs per 106 PBMCs. D0, baseline; TD, time of diagnosis (samples processed 24–48 h after initiation of treatment). Two-sided matched-pairs Wilcoxon signed-rank test. nWild-type = 11, nTN = 10. The box plots display the median and IQR, with the upper whiskers extending to the largest value ≤1.5 × IQR from the 75th percentile and the lower whiskers extending to the smallest values ≤1.5 × IQR from the 25th percentile. Source data Source data Source Data Fig. 1 Statistical Source Data Source Data Fig. 2 Statistical Source Data Source Data Fig. 3 Statistical Source Data Source Data Fig. 4 Statistical Source Data Source Data Extended Data Fig. 2 Statistical Source Data Source Data Extended Data Fig. 3 Statistical Source Data Source Data Extended Data Fig. 4 Statistical Source Data Source Data Extended Data Fig. 5 Statistical Source Data Source Data Extended Data Fig. 6 Statistical Source Data

Source Data Fig. 4 Statistical Source Data Source Data Extended Data Fig. 2 Statistical Source Data Source Data Extended Data Fig. 3 Statistical Source Data Source Data Extended Data Fig. 4 Statistical Source Data Source Data Extended Data Fig. 5 Statistical Source Data Source Data Extended Data Fig. 6 Statistical Source Data Peer review information: Alison Farrell is the primary editor on this article and managed its editorial process and peer review in collaboration with the rest of the editorial team. Publisher’s note: Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. Extended data is available for this paper at 10.1038/s41591-019-0505-4. Supplementary information is available for this paper at 10.1038/s41591-019-0505-4.

Peer review information: Alison Farrell is the primary editor on this article and managed its editorial process and peer review in collaboration with the rest of the editorial team. Publisher’s note: Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. Extended data is available for this paper at 10.1038/s41591-019-0505-4. Supplementary information is available for this paper at 10.1038/s41591-019-0505-4. Acknowledgements The authors wish to acknowledge the contribution of all study participants. The authors acknowledge the support of the Wellcome Trust (Wellcome Trust Strategic Translational Award grant no. 092661) in the development of the typhoid human challenge model that was used for this study, and the support of the NIHR Oxford Biomedical Research Centre. Cytokine studies were supported by the Bill & Melinda Gates Foundation: Global Health Vaccine Accelerator Platform grant to the Center for Human Systems Immunology at Stanford (no. OPP1113682). M.M.G. is supported in part by the NIHR Imperial Biomedical Research Centre. In addition, the authors wish to thank the following persons: R. Milward and the Walter Reed Army Institute of Research Pilot BioProduction Facility, who manufactured the GMP challenge lots; the Data Safety Monitoring Committee (D. Lalloo, D. Hill, P. Monk); M. Morgan and the microbiology laboratory at the Oxford University Hospital NHS foundation Trust; M. McClure; M. Raymond; T. Darton; C. Waddington; M. M. Levine; and the University of Maryland for provision of the original S. Typhi Quailes challenge strain. The study was funded by the Bill & Melinda Gates Foundation (no. OPP1126235). The study funders had no role in study design, data collection or analysis.

for melanoma subtype, TMB was not a significant predictor (P = 0.24). Strikingly, responders with mucosal or acral melanoma had a lower TMB than progressors with cutaneous or occult melanoma (MWW, P = 0.03; Fig. 1f), suggesting that disease subtype confounds the association between TMB and response to anti-PD1 therapy. Genomic and transcriptomic features associated with response Higher tumor purity and heterogeneity were associated with progression (MWW, P = 0.04 and P = 0.02, respectively; Fig. 2a,c), whereas ploidy was lower in progressors (MWW, P = 0.04; Fig. 2b). The proportion of the tumor genome with copy number alterations (CNAs) trended toward being higher in patients with PD (MWW, P = 0.09; Extended Data Fig. 2c).Fig. 2 Genomic and transcriptomic features associated with response. All P values are unadjusted, unless otherwise indicated. a, Tumor heterogeneity, defined as the proportion of subclonal mutations in each tumor (Methods), in responders (CR or PR) versus progressors (PD). Progressors (n = 65 patients) had greater heterogeneity than responders (n = 55 patients; two-sided MWW, P = 0.02). b, Tumor ploidy, defined as the overall genomic copy number (a normal diploid cell has a copy number of 2; Methods), in responders versus progressors. Responders (n = 55 patients) had higher tumor ploidy than progressors (n = 65 patients; two-sided MWW, P = 0.04). c, Tumor purity, defined as the proportion of DNA from tumor versus other cells in the sample (Methods), in responders versus progressors. Progressors (n = 65 patients) had higher tumor purity than responders (n = 55 patients; two-sided MWW, P = 0.04). d, Response versus progression in TAP2-amplified tumors versus other tumors. TAP2 amplification (n = 6 patients) was associated with response (two-sided Fisher’s exact test, P = 0.008). e, Response versus progression in tumors with amplified MHC-I HLA genes (HLA-A, HLA-B or HLA-C) versus other tumors. MHC-I HLA amplification (n = 6 patients) was associated with response (two-sided Fisher’s exact test, P = 0.008). f, Venn diagram showing the overlap of TAP2-amplified tumors and tumors with amplification of MHC-I HLA genes. Four tumors had amplifications on chromosome 6, including the MHC-I genes HLA-A, HLA-B, HLA-C and TAP2 and two tumors each had amplifications in one but not the other region, for a total of eight tumors with amplifications in either.

Trust; M. McClure; M. Raymond; T. Darton; C. Waddington; M. M. Levine; and the University of Maryland for provision of the original S. Typhi Quailes challenge strain. The study was funded by the Bill & Melinda Gates Foundation (no. OPP1126235). The study funders had no role in study design, data collection or analysis. Author contributions J.G. and A.J.P. conceived the project. M.M.G. and A.J.P. designed the clinical study. M.M.G., C.Jin., J.M., D.C., E.J., S.C., C.J.B., C.Dold, C.Darlow, L.B., J.H., H.T.-B. and L.S.R. collected the data. M.M.G, J.M, C.Black and C.Jones co-ordinated study approvals and recruitment. The plasma cytokine assays were performed by Y.H.-R. and G.O. A.B. performed the cytokine analysis. M.L.-T. and X. J. constructed the typhoid toxin mutant strain and carried out its in vitro characterization. G. S. conducted the mouse infection studies. E.H. and G.D. analyzed the sequencing data. U.G. provided the statistical oversight. B.A. and. A.J.P provided clinical oversight. M.M.G. wrote the first draft of the manuscript and all authors reviewed and edited the manuscript and approved the final version. Data availability The datasets generated and/or analyzed during the current study are attached. Any additional data are available from the corresponding author. No participant identifiable information will be disclosed. The raw sequence reads for the wild-type and TN strains used in the challenge are available under accession nos. ERS3381923 and ERS3381927.

Data availability The datasets generated and/or analyzed during the current study are attached. Any additional data are available from the corresponding author. No participant identifiable information will be disclosed. The raw sequence reads for the wild-type and TN strains used in the challenge are available under accession nos. ERS3381923 and ERS3381927. Competing interests A.J.P. chairs the UK Department of Health’s (DH) Joint Committee on Vaccination and Immunisation (JCVI) and the European Medicines Agency Scientific Advisory Group on Vaccines, and is a member of the World Health Organization’s (WHO) Strategic Advisory Group of Experts. A.J.P. previously received grant funding from Okairos, which ended in 2016. The views expressed in this manuscript are those of the authors and do not necessarily reflect the views of the JCVI, DH or WHO. All other authors declare no competing interests.

Main While ICB has resulted in durable clinical response in multiple tumor types1–7, only a subset of patients respond, and predictors of response are not fully characterized. Both tumor-intrinsic and tumor-extrinsic biomarkers of response and resistance to ICB in melanoma have been proposed, including tumor mutational burden (TMB) and neoantigen load8–11, immunohistological detection of PD-L1 and CD812 and genetic alterations affecting antigen presentation13,14, interferon (IFN)-γ signaling pathways15, alternative survival and proliferation pathways13,16,17 and aneuploidy18,19. Gene expression signatures expressed in tumors20 and the tumor immune microenvironment21 have also been implicated. However, these observations have often been made in preclinical models or in small clinical cohorts without validation in larger, independent cohorts of patients with melanoma. Furthermore, whether these observations are exclusive to a specific ICB regimen (that is, anti-PD1, anti-CTLA4 or a combination of these) is incompletely characterized. Broadly, the expanding suite of pathways that has been invoked to mediate selective ICB response in melanoma indicates that integrated systems biology models to predict response and survival are necessary, but these have yet to be well developed.

PD1, anti-CTLA4 or a combination of these) is incompletely characterized. Broadly, the expanding suite of pathways that has been invoked to mediate selective ICB response in melanoma indicates that integrated systems biology models to predict response and survival are necessary, but these have yet to be well developed. Clinically, the optimal role of anti-CTLA4 in conjunction3 or sequentially22 with anti-PD1 ICB is unclear. Understanding the differential biology underlying the response to anti-PD1 ICB in tumors with and without previous anti-CTLA4 therapy may inform the rational design of combination therapies and optimize therapy selection for individual patients. Thus, we performed an integrative study employing genomic, transcriptomic and clinical data from a comprehensively clinically annotated and sequenced cohort of 144 patients with advanced melanoma undergoing anti-PD1 ICB with and without previous anti-CTLA4 ICB to discover biomarkers of response and resistance, and develop clinically applicable parsimonious predictive models.

, transcriptomic and clinical data from a comprehensively clinically annotated and sequenced cohort of 144 patients with advanced melanoma undergoing anti-PD1 ICB with and without previous anti-CTLA4 ICB to discover biomarkers of response and resistance, and develop clinically applicable parsimonious predictive models. Results Genomic and clinical cohort characteristics and melanoma subtypes We identified 206 patients diagnosed with advanced melanoma and treated with anti-PD1 ICB, and performed whole-exome sequencing (WES) on matched pretreatment tumor samples and normal tissue23, and whole-transcriptome sequencing (RNA-seq) on available pretreatment tumor tissue. After quality control (Methods), WES data from 144 patients and RNA-seq data from 121 patients were available for final evaluation (Extended Data Fig. 1). Best objective response (BOR) to anti-PD1 ICB using RECIST (v.1.1) criteria (Methods) included 45% with progressive disease (PD), 14% with stable disease (SD), 3% with mixed response (MR), 26% with partial response (PR) and 12% with complete response (CR; Fig. 1a), for an overall response rate of 38%. Overall, 73% were cutaneous melanomas, 13% were of occult origin, 7% were mucosal and 7% were acral in origin. A total of 44% (n = 64) of patients had previous treatment with ipilimumab, whereas 56% (n = 80) were naive to ipilimumab. The median follow-up for survival was 29.9 months. Other clinical characteristics are detailed in Table 1.Fig. 1 Cohort genomic and clinical characteristics and association of TMB with response. a, CoMut plot showing association between clinical and genomic characteristics. Each column represents a tumor. Tumors are ordered by best RECIST criteria response (CR, PR, PD, SD or MR), and within each response subgroup by decreasing nonsynonymous (Nonsyn) mutational load (top row). Nonsynonymous mutational burden is further subdivided into clonal (purple) and subclonal (light purple) mutational load. Mutational signatures (sig) refer to the inferred relative contribution of UV-induced mutations, alkylating DNA damage process and other mutational signatures (aging+). The primary type of melanoma (skin, occult, acral or mucosal) is indicated. Tumor purity is the inferred proportion of the tumor sample that is from cancer cells compared to other cell types (Methods). The dominant mutational signature (that is, the mutational signature associated with the highest proportion of mutations) is indicated.

of melanoma (skin, occult, acral or mucosal) is indicated. Tumor purity is the inferred proportion of the tumor sample that is from cancer cells compared to other cell types (Methods). The dominant mutational signature (that is, the mutational signature associated with the highest proportion of mutations) is indicated. Mutations in BRAF, NRAS and NF1 are shown for each tumor. b, Mutational load (mut load) in progressors (n = 65 patients), responders (n = 55 patients) and patients with SD and MR (n = 24 patients). Nonsynonymous mutational load is higher in responders (CR and PR) than in progressors (two-sided MWW, P = 0.026), but is not significantly different between responders and patients having SD or MR as the best RECIST response (two-sided MWW, P = 0.14). c, Mutational load by melanoma type. Different melanoma types have different mutational loads (Kruskal–Wallis, P = 2.4 × 10−5): mutational load is higher in cutaneous and occult melanomas (n = 124 patients) than in acral and mucosal melanomas (n = 20 patients; median 297.5 versus 58; two-sided MWW, P = 1.1 × 10−6). d, Response to anti-PD1 ICB by melanoma type. Cutaneous and occult melanomas (n = 124 patients) have higher response rates (~40% CR and PR) versus acral and mucosal melanomas (n = 20 patients, 20%; two-sided Fisher’s exact test, P = 0.06). e, Mutational load in responders versus progressors stratified by melanoma type. There was no significant difference between responder and progressor mutational loads when melanomas were stratified by type (two-sided MWW; progressors versus responders (PD/R): skin (n = 42/43), P = 0.27; occult (n = 10/8), P = 0.35; acral (n = 7/2), P = 0.19; mucosal (n = 6/2), P = 0.40). Mutational load was also not a significant predictor of response in combined logistic regression after adjusting for melanoma type (P = 0.24). f, TMB in responders versus nonresponders, stratified by skin or occult melanomas versus mucosal or acral melanomas. Within each subgroup, responders trended toward having higher TMB than nonresponders (cutaneous/occult (n = 52 progressors and 51 responders): MWW, P = 0.14; mucosal/acral (n = 13 progressors and 4 responders): MWW, P = 0.08). Notably, responders with mucosal or acral melanoma (n = 4) had a lower mutational load than progressors with cutaneous or occult melanoma (n = 52; MWW, P = 0.03). Boxplots: box limits indicate the IQR (25th to 75th percentiles), with a center line indicating the median.

13 progressors and 4 responders): MWW, P = 0.08). Notably, responders with mucosal or acral melanoma (n = 4) had a lower mutational load than progressors with cutaneous or occult melanoma (n = 52; MWW, P = 0.03). Boxplots: box limits indicate the IQR (25th to 75th percentiles), with a center line indicating the median. Whiskers show the value ranges up to 1.5 × IQR above the 75th or below the 25th percentiles, with outliers beyond those ranges shown as individual points. *P < 0.05, **P < 0.01, ***P < 0.001. NS, not significant. Table 1 Cohort clinical characteristics n (%) Total cohort 144 (100) Drug received Nivolumab 59 (41.0) Pembrolizumab 85 (59.0) Sex Female 60 (41.7) Male 84 (58.3) Stage Unresectable stage III 10 (6.9) M1a 8 (5.6) M1b 18 (12.5) M1c 108 (75.0) Active brain metastases Yes 16 (11.1) No 128 (88.9) Elevated LDH Yes 71 (49.3) No 70 (48.6) Unknown 3 (2.1) ECOG performance status 0 99 (68.8) 1 37 (25.7) 2 2 (1.4) 3 1 (0.7) Unknown 5 (3.5) Primary melanoma Cutaneous 105 (72.9) Occult 19 (13.2) Acral 10 (6.9) Mucosal 10 (6.9) Received anti-PD1 ICB First line 71 (49.3) Second line or later 73 (50.7) Previous ipilimumab Yes 60 (41.7) No 84 (58.3) The number of patients with the given characteristic is shown, with the number in parentheses indicating the percentage of patients represented. ECOG, Eastern Cooperative Oncology Group.

sal 10 (6.9) Received anti-PD1 ICB First line 71 (49.3) Second line or later 73 (50.7) Previous ipilimumab Yes 60 (41.7) No 84 (58.3) The number of patients with the given characteristic is shown, with the number in parentheses indicating the percentage of patients represented. ECOG, Eastern Cooperative Oncology Group. Overall the median nonsynonymous TMB was 6.5 mutations per Mb (250.5 mutations per exome), with an interquartile range (IQR) of 2.0–14.4 mutations per Mb (77.75–578.5 mutations per exome). Overall, 39% of tumors had BRAF mutations, 30% had NRAS mutations and 17% had NF1 mutations (Fig. 1a). The median tumor purity (the proportion of sample DNA from tumor cells) was 0.67 (IQR 0.46–0.83) and the median tumor heterogeneity (the proportion of subclonal mutations) was 0.17 (IQR 0.12–0.25). The median purity-corrected tumor ploidy (Methods) was 2.15 (IQR 2.01–3.12), with 38% of tumors inferred to have genome doubling, consistent with previous reports24. The predominant mutational signature in most tumors was related to ultraviolet (UV) exposure25 (69% related to UV, 3% related to alkylating chemotherapy25 and 28% related to another predominant mutational signature, mostly associated with aging25; Fig. 1a). Individual tumor characteristics are detailed in Supplementary Table 1.

minant mutational signature in most tumors was related to ultraviolet (UV) exposure25 (69% related to UV, 3% related to alkylating chemotherapy25 and 28% related to another predominant mutational signature, mostly associated with aging25; Fig. 1a). Individual tumor characteristics are detailed in Supplementary Table 1. To discover differential features associated with response, we compared clinical responders (n = 55) to progressors (n = 65), excluding patients with SD (n = 20) and MR (n = 4) as the BOR. Overall survival (OS) and progression-free survival (PFS) were significantly different between these groups (log-rank P < 0.00001 for both comparisons; Extended Data Fig. 2a,b).

ith response, we compared clinical responders (n = 55) to progressors (n = 65), excluding patients with SD (n = 20) and MR (n = 4) as the BOR. Overall survival (OS) and progression-free survival (PFS) were significantly different between these groups (log-rank P < 0.00001 for both comparisons; Extended Data Fig. 2a,b). TMB was higher in responders than in progressors (Mann–Whitney–Wilcoxon (MWW), P = 0.026; Fig. 1b), but there was substantial overlap between responders and progressors. We hypothesized that the relationship between response and TMB might further be confounded by melanoma subtype. TMB was significantly different between different melanoma subtypes (Kruskal–Wallis, P = 2.4 × 10−5; Fig. 1c), with cutaneous and occult melanomas having similar and higher TMB than acral and mucosal melanomas26 (median of 297.5 versus 58, nonsynonymous mutations; MWW, P = 1.1 × 10−6), with a higher response rate (~40% versus ~20%; Fisher’s exact test, P = 0.06; Fig. 1d). When stratified by melanoma subtype, responders did not have significantly higher TMB than nonresponders (Fig. 1e), and, in multivariate logistic regression adjusting for melanoma subtype, TMB was not a significant predictor (P = 0.24). Strikingly, responders with mucosal or acral melanoma had a lower TMB than progressors with cutaneous or occult melanoma (MWW, P = 0.03; Fig. 1f), suggesting that disease subtype confounds the association between TMB and response to anti-PD1 therapy.

the overlap of TAP2-amplified tumors and tumors with amplification of MHC-I HLA genes. Four tumors had amplifications on chromosome 6, including the MHC-I genes HLA-A, HLA-B, HLA-C and TAP2 and two tumors each had amplifications in one but not the other region, for a total of eight tumors with amplifications in either. g, Difference in the median expression and two-sided MWW P value of association between 938 immune-related genes56 and features in responders versus progressors. Expression levels of MHC-II HLA genes (red), MHC-I HLA genes and antigen-presentation machinery (APM)-related genes (orange) are shown. h, Hierarchical clustering of the correlation matrix between genomic, clinical and transcriptomic features associated with response. Color indicates the Pearson correlation between features, from perfect negative correlation (Pearson, r = −1, blue) to perfect positive correlation (Pearson, r = 1, red). An immune-related cluster of MHC-I- and MHC-II-related gene expression is observed, with subclusters of MHC-I and MHC-II genes. Mutational and neoantigen load are highly correlated and form a cluster independently from the immune cluster (for example, Pearson correlation, r = 0.15, P = 0.11 between ssGSEA of MHC-II HLA genes and nonsynonymous mutational load). Purity is negatively correlated with the immune cluster and is independent of ploidy and heterogeneity. The sample size for each correlation depended on the number of available data points: correlations involving exclusively genomic or clinical data had n = 144 tumor samples, whereas correlations involving transcriptomic features had n = 121 tumor samples with data available. Tx, treatment. Boxplots: box limits indicate the IQR (25th to 75th percentile), with a center line indicating the median. Whiskers show the value ranges up to 1.5 × IQR above the 75th or below the 25th percentile, with outliers beyond those ranges shown as individual points. *P < 0.05, **P < 0.01, ***P < 0.001; NS, not significant.

. Boxplots: box limits indicate the IQR (25th to 75th percentile), with a center line indicating the median. Whiskers show the value ranges up to 1.5 × IQR above the 75th or below the 25th percentile, with outliers beyond those ranges shown as individual points. *P < 0.05, **P < 0.01, ***P < 0.001; NS, not significant. Given these observations, we performed an unbiased analysis for single-gene predictors of response to anti-PD1 ICB across all mutated genes detected in this cohort. After multiple-hypothesis test correction, no genes were significantly associated with response or resistance to therapy (Extended Data Fig. 3a and Supplementary Table 2), highlighting the large sample sizes needed for adequate power to detect these associations27. We observed only rare mutations in major histocompatibility complex class I (MHC-I) antigen-presentation genes (TAP1, TAP2, B2M, HLA-A, HLA-B and HLA-C)14,28. Mutations in SERPINB3 or SERPINB429 were not associated with response (Fisher’s exact test, P = 0.51 and P = 1.0, respectively). Loss of heterozygosity (LOH) in B2M14 was found in 9 of 55 (16%) responders and 16 of 65 (25%) progressors (odds ratio (OR) = 0.6) but was not significantly associated with resistance (Fisher’s exact test, P = 0.37). LOH in HLA-A, HLA-B or HLA-C28 was not associated with response or resistance to therapy (Fisher’s exact test, P = 0.52, P = 0.57 and P = 0.84, respectively). Confirming previous findings, LOH of JAK113,30,31 was associated with resistance (OR = 0.33; Fisher’s exact test, P = 0.02). Biallelic CDKN2A alteration27 was found in 15 of 55 (27%) responders and 25 of 65 (38%) progressors (OR = 0.6), but was not significant (Fisher’s exact test, P = 0.24).

d P = 0.84, respectively). Confirming previous findings, LOH of JAK113,30,31 was associated with resistance (OR = 0.33; Fisher’s exact test, P = 0.02). Biallelic CDKN2A alteration27 was found in 15 of 55 (27%) responders and 25 of 65 (38%) progressors (OR = 0.6), but was not significant (Fisher’s exact test, P = 0.24). We also performed an unbiased analysis of the association between focal gene amplifications and response to therapy. While no gene amplification was significant after multiple-hypothesis test correction (Extended Data Fig. 3b and Supplementary Table 3), amplification of TAP2, an integral part of the MHC-I antigen-loading pathway, was found exclusively in responders to therapy (n = 6; Fisher’s exact test, P = 0.008; Fig. 2d). TAP2 is located at 6p21 in a region encoding both MHC-I and MHC-II human leukocyte antigen (HLA) loci, and four out of six TAP2 amplifications were associated with larger amplifications across the region, while two out of six amplifications were more focal (Extended Data Fig. 4). Notably, tumors with amplifications in this region encompassing the MHC-I-related HLA-A, HLA-B and HLA-C genes (a region of approximately 1.5 Mb; n = 6) were also associated with response to therapy (Fisher’s exact test, P = 0.008; Fig. 2e), with four out of six also having amplifications in TAP2. Altogether, eight patients had amplification of either TAP2 or HLA-A, HLA-B or HLA-C amplification (Fig. 2f and Extended Data Fig. 4), and were exclusively responders (Fisher’s exact test, P = 0.001).

response to therapy (Fisher’s exact test, P = 0.008; Fig. 2e), with four out of six also having amplifications in TAP2. Altogether, eight patients had amplification of either TAP2 or HLA-A, HLA-B or HLA-C amplification (Fig. 2f and Extended Data Fig. 4), and were exclusively responders (Fisher’s exact test, P = 0.001). We then examined the expression of antigen-presentation molecules and their association with response. Interestingly, expression of all 13 MHC-II-associated HLA genes was higher in responders (collective two-sided binomial test, P = 0.0002; Fig. 2g), with four genes (HLA-DMA, HLA-DMB, HLA-DOB and HLA-DOB) individually passing a statistical significance threshold (MWW, P < 0.05; Supplementary Table 4). MHC-I antigen-presentation genes all trended toward having higher expression in responders (collective two-sided binomial test, P = 0.02; Fig. 2g and Supplementary Table 4), but none passed the statistical significance threshold. To examine pathways differentially enriched in responders versus progressors, we next performed unbiased gene set enrichment analysis (GSEA32; Methods) using the Hallmark gene sets33. A total of 24 pathways were enriched (false discovery rate (FDR), q < 0.1) in responders, and 5 of the top 6 enriched pathways were immune related, including IFN-γ response, genes involved in allograft rejection, complement, the inflammatory response and interleukin (IL)-6–JAK–STAT3 signaling (Supplementary Table 5). No pathways were significantly enriched in progressors.

rate (FDR), q < 0.1) in responders, and 5 of the top 6 enriched pathways were immune related, including IFN-γ response, genes involved in allograft rejection, complement, the inflammatory response and interleukin (IL)-6–JAK–STAT3 signaling (Supplementary Table 5). No pathways were significantly enriched in progressors. We further evaluated various transcriptomic signatures18,19,21,34–40 that had been proposed and demonstrated in various settings to be associated with response to immunotherapy (Methods), but we found no significant differences (P < 0.05) in these signatures between responders and progressors within our cohort (Supplementary Table 6).

ated various transcriptomic signatures18,19,21,34–40 that had been proposed and demonstrated in various settings to be associated with response to immunotherapy (Methods), but we found no significant differences (P < 0.05) in these signatures between responders and progressors within our cohort (Supplementary Table 6). Immune infiltrate has been associated with response to immunotherapy across multiple cancer types and immune therapies12,34. We inferred the absolute level of immune infiltrate within each sample using an immune deconvolution algorithm (CIBERSORT41 using the LM22 signature matrix). We found no significant difference in the total immune infiltrate or abundance of individual immune cell subsets in responders versus progressors (Supplementary Table 7). We also generated profiles of expression of immune cell subset signatures derived from single-cell analyses42 and found that the expression of multiple signatures was significantly higher (unadjusted MWW, P < 0.05) in responders versus progressors, including for signatures of overall immune infiltrate, T cells, B cells, macrophages, CD8+ cytotoxic exhausted T cells and CD4+ exhausted T cells (Supplementary Table 7). Although the strength of association differed by deconvolution method, both approaches generally agreed on the direction of association, providing evidence for a moderate association of immune infiltrate with the response to anti-PD1 ICB.

+ cytotoxic exhausted T cells and CD4+ exhausted T cells (Supplementary Table 7). Although the strength of association differed by deconvolution method, both approaches generally agreed on the direction of association, providing evidence for a moderate association of immune infiltrate with the response to anti-PD1 ICB. Correlations between molecular features To understand the relationship between predictors of response, we performed hierarchical clustering of the correlation coefficients between associated predictors (Fig. 2h). Clustered features are correlated and may reflect the same underlying biology, whereas separate clusters may reflect independent feature categories. A large immune-related cluster with MHC-II- and MHC-I-related subclusters was observed, with a separate independent cluster of mutation- and neoantigen-load-related features, suggesting independent feature categories. Tumor purity was negatively correlated with the immune cluster, suggesting that low tumor purity may be a proxy for higher immune infiltrates within the tumor sample. Other features, including tumor heterogeneity and ploidy, were independent from these two clusters. Extending this analysis to previously hypothesized signatures and Hallmark gene set signatures (Extended Data Fig. 5), immune activity signatures, including signatures of cytolytic21 and cytotoxic19 activity, IFN-γ and T effector cells18,35, immune chemokines38 and single-cell-derived immune cell signatures42, clustered together. A tumor-intrinsic resistance program signature42, signatures of T cell dysfunction and exclusion37 and comparative immune-checkpoint gene expression34 were distinct from the immune cluster. Overall, these findings suggested that multiple previously hypothesized predictors of ICB response reflect the same underlying biological state and additional independent classes of predictors exist that may provide additional predictive power.

immune-checkpoint gene expression34 were distinct from the immune cluster. Overall, these findings suggested that multiple previously hypothesized predictors of ICB response reflect the same underlying biological state and additional independent classes of predictors exist that may provide additional predictive power. Previous exposure to anti-CTLA4 ICB Our cohort contained patients with previous exposure to ipilimumab in anti-CTLA4 ICB (n = 60) and patients who were naive to ipilimumab (n = 84; Fig. 3a). Despite the groups having similar response rates (Fig. 3b), we hypothesized that these two groups might have differential predictors of response and resistance to anti-PD1 ICB. We performed a focused analysis of patients who were treated with ipilimumab and biopsied after treatment (n = 44 with WES and n = 34 with RNA-seq) versus patients who were naive to ipilimumab (n = 84 with WES and n = 71 with RNA-seq). A composite, rank-based score of MHC-II HLA expression (single-sample GSEA (ssGSEA)43; Methods) was higher in responders than in progressors in the overall cohort and the subgroup treated with ipilimumab, but was not significantly different in the subgroup that was naive to ipilimumab (Fig. 3c–e; MWW, P = 0.03, P = 0.03 and P = 0.31, respectively). We found very similar results in the largest available independent validation cohort44 with information on previous ipilimumab treatment (Fig. 3f–h), although the difference was not significant in this smaller cohort (n = 32 patients, 15 of whom were treated with ipilimumab and 17 of whom were naive for ipilimumab).Fig. 3 Differential predictors of response and progression in ipilimumab-treated tumors versus ipilimumab-naive tumors. a, Timeline showing when sequenced biopsies were obtained from tumors that were treated with ipilimumab or naive to ipilimumab in the course of therapy. Subsequent analyses focused on comparing tumor biopsies obtained after ipilimumab treatment (n = 45 WES, n = 34 RNA-seq) to ipilimumab-naive tumor biopsies (n = 84 WES, n = 74 RNA-seq). b, Best RECIST response by ipilimumab pretreatment status. There was no difference between the distribution of responses in naive (n = 84) and pretreated (n = 60) patients (two-sided chi-squared test, P = 0.44; degrees of freedom (d.f.) = 3). c, ssGSEA of MHC-II HLA genes (Methods) in responders (n = 47 patients) versus progressors (n = 56 patients) in the overall cohort. MHC-II scores were higher in responders than in progressors (two-sided MWW, P = 0.03).

d pretreated (n = 60) patients (two-sided chi-squared test, P = 0.44; degrees of freedom (d.f.) = 3). c, ssGSEA of MHC-II HLA genes (Methods) in responders (n = 47 patients) versus progressors (n = 56 patients) in the overall cohort. MHC-II scores were higher in responders than in progressors (two-sided MWW, P = 0.03). d, ssGSEA of MHC-II HLA genes in responders versus progressors in the post-ipilimumab-treatment subgroup. MHC-II scores were higher in responders (n = 11 patients) than in progressors (n = 16 patients; two-sided MWW, P = 0.03). e, ssGSEA of MHC-II HLA genes in responders (n = 31 patients) versus progressors (n = 34 patients) in the ipilimumab-naive subgroup. There was no significant difference in MHC-II scores between responders and progressors (two-sided MWW, P = 0.31). f, MHC-II HLA gene set scores (ssGSEA) in responders (n = 10 patients) versus progressors (n = 22 patients) in a validation cohort (Methods; two-sided MWW, P = 0.34). g, MHC-II HLA gene set scores (ssGSEA) in responders (n = 4 patients) versus progressors (n = 11 patients) in the ipilimumab-treated subgroup of a validation cohort (two-sided MWW, P = 0.10). h, MHC-II HLA gene set scores (ssGSEA) in responders (n = 6 patients) versus progressors (n = 11 patients) in the ipilimumab-naive subgroup of a validation cohort (two-sided MWW, P = 0.80). i, Selected Cancer Hallmark gene sets (GSEA) enriched in responders versus progressors in the overall (n = 47 responders and 56 progressors), post-ipilimumab-treatment (n = 11 responders and 16 progressors), and ipilimumab-naive (n = 31 responders and 34 progressors) subgroups of our discovery cohort. IFN-γ and IFN-α response pathways were enriched in responders in the overall (FDR, q < 0.001 and q = 0.02, respectively) and ipilimumab-treated (q < 0.001, both) subgroups but not in the ipilimumab-naive subgroups (q = 0.13 and q = 0.997, respectively) in the discovery cohort (empiric, P = 0.183 and P = 0.18, respectively for the difference in q values between the subgroups; Methods). j, Selected Cancer Hallmark gene sets (GSEA) enriched in responders versus progressors in the overall, post-ipilimumab-treatment and ipilimumab-naive subgroups in an independent validation cohort.

overy cohort (empiric, P = 0.183 and P = 0.18, respectively for the difference in q values between the subgroups; Methods). j, Selected Cancer Hallmark gene sets (GSEA) enriched in responders versus progressors in the overall, post-ipilimumab-treatment and ipilimumab-naive subgroups in an independent validation cohort. IFN-γ and IFN-α response pathways were enriched in responders in the overall (FDR, q < 0.0001 and q = 0.0034, respectively) and ipilimumab-treated (q < 0.0001, both) subgroups but not in the ipilimumab-naive subgroup (q = 0.87 and q = 0.03 (enriched in progressors), respectively) in the discovery cohort. Pathways favoring enrichment in progressors (as opposed to responders) are visualized here with an FDR q value of 1. All Hallmark pathways and their GSEA enrichment scores are shown in Supplementary Table 5. Boxplots: box limits indicate the IQR (25th to 75th percentile), with a center line indicating the median. Whiskers show the value ranges up to 1.5 × IQR above the 75th or below the 25th percentile, with outliers beyond those ranges shown as individual points. *P < 0.05, **P < 0.01, ***P < 0.001. NS, not significant.

. Boxplots: box limits indicate the IQR (25th to 75th percentile), with a center line indicating the median. Whiskers show the value ranges up to 1.5 × IQR above the 75th or below the 25th percentile, with outliers beyond those ranges shown as individual points. *P < 0.05, **P < 0.01, ***P < 0.001. NS, not significant. We next examined the association of TMB, purity, ploidy and heterogeneity with response stratified by previous ipilimumab therapy (Extended Data Fig. 6a). Unlike previous studies44,45, we found no specific association of a higher TMB with response in the ipilimumab-naive versus ipilimumab-treated subgroup (MWW, P = 0.15, both). However, higher heterogeneity and lower ploidy were associated with progressors only in the ipilimumab-naive subgroup (MWW, P = 0.06 and P = 0.004, respectively).

tudies44,45, we found no specific association of a higher TMB with response in the ipilimumab-naive versus ipilimumab-treated subgroup (MWW, P = 0.15, both). However, higher heterogeneity and lower ploidy were associated with progressors only in the ipilimumab-naive subgroup (MWW, P = 0.06 and P = 0.004, respectively). We analyzed the differential expression of specific immune-related genes in responders versus progressors in ipilimumab-treated and ipilimumab-naive subgroups and found that higher expression of various immune-related pathways distinguished responders from progressors in ipilimumab-treated but not ipilimumab-naive subgroups (all P values are unadjusted). Examples included the leukocyte chemoattractants CXCL9 and CXCL10 and their receptor CXCR3 (MWW, P = 0.05, P = 0.08 and P = 0.02, respectively, in the ipilimumab-treated subgroup), CD3D (MWW, P = 0.02), B cell markers CD19 (MWW, P = 0.04) and CD20 (MS4A1; MWW, P = 0.002) and macrophage marker CD163 (MWW, P = 0.03). Interestingly, CD4, FOXP3 and CTLA4 also followed this pattern of higher expression in responders in the ipilimumab-treated subgroup (MWW, P = 0.06, P = 0.06 and P = 0.008, respectively), but CD8A and CD8B had less evidence of association with response in either ipilimumab-treated (MWW, P = 0.17 and P = 0.27, respectively) or ipilimumab-naive (MWW, P = 0.93 and P = 0.49, respectively) subgroups. Expression of TAP2 was higher (MWW, P = 0.02) in responders than in progressors in the ipilimumab-treated subgroup but not in the ipilimumab-naive subgroup (MWW, P = 0.98). In contrast, TGFB2 expression was higher in progressors in the ipilimumab-naive subgroup (MWW, P = 0.002) but not in the ipilimumab-treated subgroup (MWW, P = 0.43). The complete set of gene expression comparisons in the overall, ipilimumab-treated and ipilimumab-naive cohorts is in Supplementary Table 4.

W, P = 0.98). In contrast, TGFB2 expression was higher in progressors in the ipilimumab-naive subgroup (MWW, P = 0.002) but not in the ipilimumab-treated subgroup (MWW, P = 0.43). The complete set of gene expression comparisons in the overall, ipilimumab-treated and ipilimumab-naive cohorts is in Supplementary Table 4. We then repeated GSEA to examine the pathways differentially enriched in responders versus progressors between ipilimumab-treated and ipilimumab-naive subgroups. The most differentially enriched pathways were related to immune response: the IFN-γ and IFN-α responses were significantly enriched in responders in the ipilimumab-treated subgroup (FDR, q < 0.0001, both), but not in the ipilimumab-naive subgroup (FDR, q = 0.13 and q = 0.996, respectively; Fig. 3i). Using permutation testing (Methods), we found a nonsignificant empiric P value of 0.183 and 0.18, respectively, for this difference in enriched pathways in these subgroups in our discovery cohort. However, we repeated the analysis in an independent validation cohort46 and found similar results (Fig. 3j). Complete GSEA results are provided in Supplementary Table 5.

a nonsignificant empiric P value of 0.183 and 0.18, respectively, for this difference in enriched pathways in these subgroups in our discovery cohort. However, we repeated the analysis in an independent validation cohort46 and found similar results (Fig. 3j). Complete GSEA results are provided in Supplementary Table 5. To further dissect the impact of MHC-II expression on patient response, we stratified the cohort into patients with high and low MHC-II expression (ssGSEA, median split). In the overall cohort, low MHC-II expression was associated with primary resistance (Fig. 4a; Fisher’s exact test, P = 0.01; OR = 2.9, 95% confidence interval (CI) 1.3–6.5), but this association was largely driven by the ipilimumab-treated subgroup (Fig. 4b; Fisher’s exact test, P = 0.02; OR = 9.9, 95% CI 1.5–63.7), with a nonsignificant association in the ipilimumab-naive subgroup (Fig. 4c; Fisher’s exact test, P = 0.32; OR = 1.9, 95% CI 0.7–5.1). A formal interaction test of previous ipilimumab treatment status with MHC-II expression for predicting response was consistent with a subgroup-specific effect of low MHC-II expression on the ipilimumab-experienced subgroup (OR = 0.20, 95% CI 0.02–1.64), although this was nonsignificant (P = 0.13) in this small cohort.Fig. 4 Progression versus response by immune infiltrate and MHC-II HLA expression stratified by ipilimumab treatment. a, Proportion of responders versus progressors in subgroups with high and low MHC-II HLA score (ssGSEA; divided by the median). Overall, a high MHC-II HLA score was associated with higher response (MHC-II high: n = 29 responders and 20 progressors; MHC-II low: n = 18 responders and 36 progressors; two-sided Fisher’s exact test, P = 0.01; OR = 2.9, 95% CI 1.3–6.5). b, As in a but in the ipilimumab-treated subgroup. Tumors with low MHC-II HLA score (ssGSEA) were associated with PD (MHC-II high: n = 9 responders and 5 progressors; MHC-II low: n = 2 responders and 11 progressors; two-sided Fisher’s exact test, P = 0.02; OR = 9.9, 95% CI 1.5–63.7). c, As in a but in the ipilimumab-naive subgroup. There was no significant difference between tumors with high and low MHC-II HLA scores (ssGSEA; MHC-II high: n = 18 responders and 14 progressors; MHC-II low: n = 13 responders and 19 progressors; two-sided Fisher’s exact test, P = 0.32; OR = 1.9, 95% CI 0.7–5.1). d, Proportion of responders versus progressors in subgroups with high and low immune infiltrate scores (divided by the median).

-II HLA scores (ssGSEA; MHC-II high: n = 18 responders and 14 progressors; MHC-II low: n = 13 responders and 19 progressors; two-sided Fisher’s exact test, P = 0.32; OR = 1.9, 95% CI 0.7–5.1). d, Proportion of responders versus progressors in subgroups with high and low immune infiltrate scores (divided by the median). Overall, there was no statistically significant difference in the proportion of responders versus progressors in tumors with high versus low immune infiltrate (infiltrate high: n = 27 responders and 24 progressors; infiltrate low: n = 20 responders and 32 progressors; two-sided Fisher’s exact test, P = 0.17; OR = 1.8, 95% CI 0.8–3.9). e, As in d but in the ipilimumab-treated subgroup. Tumors with low immune infiltrate scores were strongly associated with PD (infiltrate high: n = 10 responders and 7 progressors; infiltrate low: n = 1 responder and 9 progressors; two-sided Fisher’s exact test, P = 0.02; OR = 12.9, 95% CI 1.3–125.8). f, As in d but in the ipilimumab-naive subgroup. There was no statistically significant difference between tumors with high and low immune infiltrate scores (infiltrate high: n = 14 responders and 15 progressors; infiltrate low: n = 17 responders and 18 progressors; two-sided Fisher’s exact test, P = 1.0; OR = 0.99, 95% CI 0.4–2.6). g, Schematic of the hypothesized effect of ipilimumab treatment on immune response as a predictor of subsequent response to anti-PD1 ICB. Tumors that were treated with ipilimumab but failed to have an immune response or infiltrate in the tumor microenvironment were strongly predicted to have intrinsic resistance to anti-PD1 ICB. However, having an immune response did not guarantee a subsequent response to anti-PD1 ICB. In ipilimumab-naive tumors, neither the presence nor the absence of immune infiltrate was a good predictor of anti-PD1 ICB response.

he tumor microenvironment were strongly predicted to have intrinsic resistance to anti-PD1 ICB. However, having an immune response did not guarantee a subsequent response to anti-PD1 ICB. In ipilimumab-naive tumors, neither the presence nor the absence of immune infiltrate was a good predictor of anti-PD1 ICB response. A similar analysis of estimated total immune infiltrate levels41 (with tumors split by the median into high and low groups) showed that low immune infiltrate was significantly associated with intrinsic resistance in the ipilimumab-treated subgroup (Fig. 4e; Fisher’s exact test, P = 0.02; OR = 12.9, 95% CI 1.3–125.8), but not in the ipilimumab-naive subgroup (Fig. 4f; Fisher’s exact test, P = 1.00; OR = 0.99, 95% CI 0.4–2.6) or overall cohort (Fig. 4d; Fisher’s exact test, P = 0.17; OR = 1.8, 95% CI 0.8–3.9), indicating a subgroup-specific association in the ipilimumab-experienced subgroup of low immune infiltrate with resistance to therapy (OR = 0.08, 95% CI 0.007–0.97; P = 0.047). Taken together, these findings suggest that evidence of immune response in the tumor microenvironment at the time of progression following anti-CTLA4 ICB is a necessary but not sufficient marker for response to anti-PD1 ICB therapy; patients without immune response to anti-CTLA4 ICB are very likely to also be intrinsically resistant to anti-PD1 ICB, highlighting a high-risk and poor-prognosis subgroup of patients (Fig. 4g).

e time of progression following anti-CTLA4 ICB is a necessary but not sufficient marker for response to anti-PD1 ICB therapy; patients without immune response to anti-CTLA4 ICB are very likely to also be intrinsically resistant to anti-PD1 ICB, highlighting a high-risk and poor-prognosis subgroup of patients (Fig. 4g). Integrative predictive modeling of primary resistance Patients with primary resistance to anti-PD1 ICB have poor survival (Extended Data Fig. 7), and the ability to predict these patients would enable individualized management regimens (for example combination ICB) to improve outcomes. Thus, we set out to develop parsimonious predictive models integrating clinical, genomic and transcriptomic features to predict PD (primary resistance) versus non-PD (CR, PR, SD and MR) and developed separate predictive models in ipilimumab-treated and ipilimumab-naive subgroups.

e combination ICB) to improve outcomes. Thus, we set out to develop parsimonious predictive models integrating clinical, genomic and transcriptomic features to predict PD (primary resistance) versus non-PD (CR, PR, SD and MR) and developed separate predictive models in ipilimumab-treated and ipilimumab-naive subgroups. In the ipilimumab-treated group (n = 34 with WES and RNA-seq), there were 16 patients who had PD and 18 who had non-PD. Using a forward-selection approach to choose the features of a parsimonious predictive model (Methods), low MHC-II HLA expression was most strongly predictive of PD (Fig. 5a) and was correlated with MHC-I HLA, IFN-α and IFN-γ response pathway scores. In the final multivariate model, high MHC-II expression, low lactate dehydrogenase (LDH; below the median of 247 U l−1) and the presence of lymph node metastases were independent predictors of non-PD (P = 0.03, P = 0.02 and P = 0.04, respectively, Supplementary Table 8 and Extended Data Fig. 8a,c), and the model had an area under the curve (AUC) of 0.90 in our discovery cohort (fivefold cross-validation mean AUC = 0.83; empiric P < 0.001; Fig. 5b, Extended Data Fig. 8f and Methods). Notably, TMB did not significantly improve model fit (log-likelihood ratio, P = 0.10), was not an independently predictive feature (P = 0.18) and did not meet Bayesian information criteria (BIC; Extended Data Fig. 8c) when added to the model.Fig. 5 Integrative predictive modeling of intrinsic resistance to anti-PD1 ICB. a, CoMut plot showing the relationship between response and predictive features in the ipilimumab-treated subgroup. Each column represents a patient, and the top row indicates whether the patient had PD (intrinsic resistance) or non-PD (CR, PR, SD or MR) as the best response. Patients are sorted by MHC-II HLA score, which was the most predictive feature and was correlated with MHC-I and IFN response pathway scores. MHC-II HLA, LDH at treatment initiation and the presence of lymph node metastases (LN met) were features used for our logistic regression model, chosen using forward selection (Methods). b, A receiver–operator characteristic (ROC) curve for our predictive model for ipilimumab-treated tumors (n = 34 patients) using MHC-II HLA, LDH and lymph node metastases as features. The AUC was 0.9, and the log-likelihood ratio was P = 0.0003. The fivefold cross-validation mean AUC was 0.83. c, CoMut plot showing the relationship between response and predictive features in the ipilimumab-naive subgroup.

b-treated tumors (n = 34 patients) using MHC-II HLA, LDH and lymph node metastases as features. The AUC was 0.9, and the log-likelihood ratio was P = 0.0003. The fivefold cross-validation mean AUC was 0.83. c, CoMut plot showing the relationship between response and predictive features in the ipilimumab-naive subgroup. Each column represents a patient, and the top row indicates whether the patient had PD (intrinsic resistance) or non-PD as the best response. Patients are sorted by tumor heterogeneity (top), tumor ploidy (middle) and tumor purity (bottom), which were the three features chosen in our predictive model (Methods). Purity, ploidy and heterogeneity were independent predictors in the multivariate model (Supplementary Table 9). d, The ROC curve for our predictive model for ipilimumab-naive tumors (n = 84 tumors) using heterogeneity, purity and ploidy as features. The AUC was 0.77, and the log-likelihood ratio was P = 0.0003. The average tenfold cross-validation mean AUC was 0.73. e, Survival (PFS and OS as indicated) stratified by high versus low predictive model score (split by the median) in ipilimumab-treated tumors in our discovery cohort. Tumors with high scores had worse PFS and OS (two-sided KM log-rank test, P = 0.0001 and P = 0.001, respectively). f, Survival (PFS and OS as indicated) stratified by high versus low predictive model score (split by the median) in ipilimumab-naive tumors in our discovery cohort. Tumors with high scores had worse PFS and OS (two-sided KM log-rank test, P = 0.003 and P = 6.3 × 10−5, respectively).

0.0001 and P = 0.001, respectively). f, Survival (PFS and OS as indicated) stratified by high versus low predictive model score (split by the median) in ipilimumab-naive tumors in our discovery cohort. Tumors with high scores had worse PFS and OS (two-sided KM log-rank test, P = 0.003 and P = 6.3 × 10−5, respectively). In the ipilimumab-naive group, there were 34 patients with PD and 41 with non-PD. In the multivariate predictive model, higher heterogeneity, lower ploidy and higher purity were independently predictive of PD (P = 0.025, P = 0.014 and P = 0.046, respectively; Fig. 5c, Supplementary Table 9 and Extended Data Fig. 8b,d), with an AUC of 0.77 (tenfold cross-validation mean AUC of 0.73; empiric P = 0.036; Fig. 5d, Extended Data Fig. 8e and Methods). TMB did not significantly improve the model fit (log-likelihood ratio test, P = 0.63), was not an independently predictive feature (P = 0.63) and did not meet BIC criteria (Extended Data Fig. 8d) when added to this multivariate model. Further, each model’s performance was specific to its subgroup (P = 0.004 and P = 0.018 for the interaction between ipilimumab-experienced and ipilimumab-naive model scores, respectively, and previous ipilimumab therapy). Each model had poor performance when applied to the opposing subgroup (AUC = 0.49 and AUC = 0.54, respectively; Extended Data Fig. 9a,b), suggesting that previous ipilimumab treatment status may stratify the appropriate predictors and predictive models to be applied in these subgroups.

-negative matrix factorization72. Cosine similarity was used to compare the discovered signatures to the 30 existing discovered and validated signatures in COSMIC25,73, with a threshold of 0.85, and we also manually visualized and inspected similarities in mutational motifs between our signatures and COSMIC signatures. Transcriptomic analysis Whole-transcriptome sequencing data from FFPE tissues were aligned using STAR74 and quantified with RSEM75 to yield gene-level expression in transcripts per million (TPM). For RNA-seq quality control, sequencing- and alignment-specific metrics were considered for each sample. The following alignment metrics (output by the STAR alignment method) were considered: percentage of uniquely mapped reads, average mapped read length, number of splices, mismatch rate per base, percentage of multimapped reads, percentage of reads mapped to too many locations, percentage of unmapped reads due to too many mismatches, percentage of unmapped reads due to reads being too short and percentage of unmapped reads due to other reasons. Additionally, we considered the raw number of reads, the average read length, the read duplication rate and DV200 for each sample. Samples were clustered across quality-control metrics using principal-component analysis, and outlier samples were manually evaluated and discarded. Three samples were removed owing to poor quality: patient 143 was excluded owing to an abnormally low absolute number of reads (number of reads <1 million); patient 107 was excluded owing to an abnormally high percentage of reads mapped to too many locations (>10% of reads), likely indicating high numbers of short or degraded reads; and patient 61 was excluded owing to multiple aberrant quality-control metrics resulting in overall poor quality when considering all metrics in aggregate as well as an aberrant expression profile compared to all other samples. Only transcriptomes from tumors whose WES also passed quality control were included; the final patient cohort for RNA-seq analysis included n = 121 transcriptomes.

and previous ipilimumab therapy). Each model had poor performance when applied to the opposing subgroup (AUC = 0.49 and AUC = 0.54, respectively; Extended Data Fig. 9a,b), suggesting that previous ipilimumab treatment status may stratify the appropriate predictors and predictive models to be applied in these subgroups. We attempted to validate our models in independent cohorts but were limited by a lack of publicly available cohorts with all the molecular features and relevant clinical data on previous ipilimumab therapy and biopsy timing used in our integrated model. In a limited validation, we tested predictive models incorporating individual features where data were available in an independent validation cohort44 (Methods), and found concordant predictions of primary resistance with low MHC-II HLA expression in ipilimumab-treated tumors and higher heterogeneity in ipilimumab-naive tumors (Extended Data Fig. 10), although neither of these predictors was significant in this small cohort (empiric P = 0.21 and P = 0.066, respectively; Methods). In an exploratory analysis, we predicted PD in response to anti-PD1 ICB in tumors where the tumor biopsy was taken before ipilimumab treatment (n = 15; Fig. 3a) using our ipilimumab-naive predictive model. Of the eight tumors with PD in response to anti-PD1 ICB, five had the highest model scores (Extended Data Fig. 9c), with an overall AUC of 0.71. Interestingly, of the three poorly discriminated tumors with low model scores but PD responses, one (patient 82) was from a brain metastasis and one (patient 80) was an acral melanoma.

ght tumors with PD in response to anti-PD1 ICB, five had the highest model scores (Extended Data Fig. 9c), with an overall AUC of 0.71. Interestingly, of the three poorly discriminated tumors with low model scores but PD responses, one (patient 82) was from a brain metastasis and one (patient 80) was an acral melanoma. Finally, splitting the cohort into subsets with high and low model scores (split by the median), we found large differences in OS and PFS in both the ipilimumab-treated subgroup (Fig. 5e; median PFS: 38.1 months versus 2.8 months, log-rank P = 0.0001; median OS: unreached versus 9.0 months, log-rank P = 0.001) and ipilimumab-naive subgroup (Fig. 5f; median PFS: 24.7 months versus 3.1 months, log-rank P = 0.005; median OS: unreached versus 15.0 months, log-rank P < 0.0001). Discussion In this study, we analyzed a uniformly clinically annotated cohort of patients with advanced melanoma treated with anti-PD1 ICB monotherapy for whom WES and RNA-seq data were available. While we observed an association between response and TMB, this observation was confounded by disease subtype, strongly suggesting that TMB cannot be applied generically across melanoma subtypes as a predictive biomarker for anti-PD1 ICB.

eated with anti-PD1 ICB monotherapy for whom WES and RNA-seq data were available. While we observed an association between response and TMB, this observation was confounded by disease subtype, strongly suggesting that TMB cannot be applied generically across melanoma subtypes as a predictive biomarker for anti-PD1 ICB. Beyond TMB, we found that MHC-II expression, tumor heterogeneity, purity and ploidy were associated with ICB response. In two previous studies47,48, MHC-II expression on tumor cells by immunohistochemistry was found to be predictive for response to anti-PD1 ICB and was hypothesized to represent a subset of tumors that could stimulate CD4+ helper T cell or cytotoxic activity. Consistent with this hypothesis, we found that MHC-II transcriptomic expression was correlated with expression of CD4 and the cytolytic molecules PRF1 and GZMA in our cohort. However, whether MHC-II expression represents expression on tumor cells or antigen-presenting cells within the tumor microenvironment cannot be determined from our bulk transcriptome data, and whether the association of PD1 ICB response with MHC-II expression is limited to tumor-cell-specific MHC-II expression is unclear. Notably, CD8+ T cell markers were not higher in responders versus progressors in our cohort, and although MHC-II, MHC-I, IFN-γ and IFN-α response pathway expression was correlated, MHC-II expression was the best predictor of response in our cohort. Further, we found evidence for the involvement of other immune compartments (for example, B cell markers enriched in ipilimumab-experienced responders) in ICB response, consistent with data from a recent trial demonstrating higher B cell infiltrate in responders to neoadjuvant immunotherapy46, although the specific cell types, functional states and tumor immune interactions are not yet well characterized.

, B cell markers enriched in ipilimumab-experienced responders) in ICB response, consistent with data from a recent trial demonstrating higher B cell infiltrate in responders to neoadjuvant immunotherapy46, although the specific cell types, functional states and tumor immune interactions are not yet well characterized. Tumor heterogeneity (the proportion of subclonal mutations) has previously been associated with poor prognosis across multiple tumor types and therapies49–51. High heterogeneity suggests a highly mutagenic disease and a high degree of subclonality, with a higher likelihood of preexisting or rapidly evolving resistant clones. Interestingly, in our cohort, four patients had a very high TMB with an alkylating chemotherapy mutational signature (and known previous alkylating chemotherapy); low tumor heterogeneity distinguished the two responders from the two nonresponders (with SD and PD as the best response), who had high tumor heterogeneity with a majority of subclonal mutations. This association has also been observed in dacarbazine-experienced patients with melanoma treated with anti-CTLA4 ICB49, suggesting that tumor heterogeneity may be significantly correlated with ICB resistance.

with SD and PD as the best response), who had high tumor heterogeneity with a majority of subclonal mutations. This association has also been observed in dacarbazine-experienced patients with melanoma treated with anti-CTLA4 ICB49, suggesting that tumor heterogeneity may be significantly correlated with ICB resistance. In our cohort, higher ploidy and lower purity were associated with ICB response, but the biological basis of these relationships is unclear. Genome doubling events are common in cancer and may accelerate genome evolution by increasing the tolerance of genome instability52, and higher aneuploidy has been associated with a worse response to ICB18,19. However, whether genome doubling (and higher ploidy) is also associated with increased immunogenicity is unclear. Purity is negatively correlated with expression of markers of immune response and may be a proxy for the level of immune response in the tumor microenvironment in this setting rather than an artifact of tumor sample processing. However, tumor purity may also reflect differences in tumor biology leading to intrinsic resistance.

s negatively correlated with expression of markers of immune response and may be a proxy for the level of immune response in the tumor microenvironment in this setting rather than an artifact of tumor sample processing. However, tumor purity may also reflect differences in tumor biology leading to intrinsic resistance. Notably, we found that previous exposure to anti-CTLA4 ICB affected the predictors of response to anti-PD1 ICB, although patients with and without exposure had similar response rates to anti-PD1 ICB. Immune-related markers are strongly enriched in responders compared to progressors with previous ipilimumab exposure, but this relationship is less clear in ipilimumab-naive patients. Specifically, post-ipilimumab tumors with poor immune response at progression were resistant to further anti-PD1 ICB, whereas ipilimumab-naive immune-poor and immune-infiltrated tumors were similarly likely to respond to anti-PD1 ICB. Whether anti-CTLA4 ICB induces or, alternatively, reveals an immune-resistant state in a subset of melanomas is an important question that deserves further evaluation. Further, cross-resistance to sequential ICB may also predict resistance to simultaneous combination ICB; this hypothesis should be evaluated in an appropriate cohort.

TLA4 ICB induces or, alternatively, reveals an immune-resistant state in a subset of melanomas is an important question that deserves further evaluation. Further, cross-resistance to sequential ICB may also predict resistance to simultaneous combination ICB; this hypothesis should be evaluated in an appropriate cohort. Building on these findings, we constructed predictive models integrating clinical, genomic and transcriptomic characteristics to identify patients with melanoma with intrinsic resistance to anti-PD1 ICB. Integrating multiple clinical and molecular features resulted in superior discrimination compared to models with any single feature or modality. In patients treated with ipilimumab, low MHC-II expression and high LDH predicted intrinsic resistance, whereas lymph node metastasis predicted improved response to therapy. MHC-II47,48 and LDH53 have previously been implicated in predicting anti-PD1 responsiveness. Lymph node metastases might provide a reservoir of tumor-specific immune cells, facilitating their activation by physiologic lymph node function; recent experimental data in a murine model suggests that lymph node metastases are necessary for PD1 response54, and recent clinical data showing greater tumor-resident T cell clone response to neoadjuvant compared to adjuvant immunotherapy further supports this hypothesis55. Similarly, integrating tumor heterogeneity, ploidy and purity for ipilimumab-naive disease resulted in a higher AUC than was obtained with any single-feature model alone. Beyond predicting response, these parsimonious models strongly stratified patients by PFS and OS, suggesting potential clinical applicability in identifying patients at high and low risk.

eity, ploidy and purity for ipilimumab-naive disease resulted in a higher AUC than was obtained with any single-feature model alone. Beyond predicting response, these parsimonious models strongly stratified patients by PFS and OS, suggesting potential clinical applicability in identifying patients at high and low risk. These findings will require validation in independent and larger cohorts; at the time of our study, limited data were publicly available where molecularly sequenced tumors with previous treatment data and all relevant clinical parameters were available for validation. Further, heterogeneity in sequencing approaches and data normalization between cohorts hindered our ability to develop standardized features to create and validate models. However, our results highlight the value of integrating rich clinical data with molecular tumor characterization and the need to generate such multimodal data.

terogeneity in sequencing approaches and data normalization between cohorts hindered our ability to develop standardized features to create and validate models. However, our results highlight the value of integrating rich clinical data with molecular tumor characterization and the need to generate such multimodal data. Methods Patient cohort and clinical end points Patients were identified in databases of participating sites. For enrollment, patients were required to have advanced melanoma and to have received PD1 blockade as a palliative treatment. Tissue obtained before PD1 blockade was required for enrollment and was collected during routine medical care. Clinicopathological and demographic data were collected from patient records locally and are shown in Table 1. Age, stage and ECOG performance status were documented before the first application of anti-PD1 ICB. LDH was measured within 28 d of the first application of nivolumab or pembrolizumab. OS was defined as the time between the first application of anti-PD1 ICB and the date of death (any cause). For subjects without documentation of death, OS was censored on the last date the patient was known to be alive. BOR to anti-PD1 ICB was assessed according to RECIST criteria v.1.1 by the participating sites. Patients achieving CR or PR as BOR were grouped as responders, whereas patients showing PD as the best response were referred to as progressors. Patients were classified as MR when achieving unequivocal responses in individual existing lesions but also progression in others or new lesions. PFS was defined as the time between the first application of anti-PD1 ICB and the date of documented disease progression. For patients without documentation of progression, PFS was censored on the last date the patient was known to be without progression.

existing lesions but also progression in others or new lesions. PFS was defined as the time between the first application of anti-PD1 ICB and the date of documented disease progression. For patients without documentation of progression, PFS was censored on the last date the patient was known to be without progression. This retrospective study and associated informed consent procedures were approved by the central Ethics Committee (EC) of the University Hospital Essen (12-5152-BO and 11-4715). Approval by the local EC was obtained by investigators if required by local regulations. Samples Samples were collected retrospectively and obtained by excision or biopsy of melanoma tissue, collected locally at the participating sites and provided formalin-fixed and paraffin-embedded (FFPE). Samples were collected between January 2013 and June 2016. The median time from biopsy to initiation of anti-PD1 blockade was 2.1 months with 90% of samples being collected 6 months before the first application of nivolumab or pembrolizumab. All biopsies were from metastatic sites, with the exception of eight biopsies; seven were from a primary lesion and one was from a recurrence at a primary site, representing less than 6% of the overall cohort.

onths with 90% of samples being collected 6 months before the first application of nivolumab or pembrolizumab. All biopsies were from metastatic sites, with the exception of eight biopsies; seven were from a primary lesion and one was from a recurrence at a primary site, representing less than 6% of the overall cohort. Whole-exome and whole-transcriptome sequencing DNA extraction, whole exome library preparation and sequencing were performed for samples as previously described10,23. Slides were cut from FFPE blocks and macrodissected for tumor-enriched tissue. Paraffin was removed from FFPE sections and cores using CitriSolv (Fisher Scientific), followed by ethanol washes and tissue lysis overnight at 56 °C. Samples were then incubated at 90 °C to remove DNA cross links. Extraction of DNA, and, when possible, RNA was performed using the QIAGEN AllPrep DNA/RNA mini kit (51306). Germline DNA was obtained from peripheral blood mononuclear cells and adjacent normal tissue.

by ethanol washes and tissue lysis overnight at 56 °C. Samples were then incubated at 90 °C to remove DNA cross links. Extraction of DNA, and, when possible, RNA was performed using the QIAGEN AllPrep DNA/RNA mini kit (51306). Germline DNA was obtained from peripheral blood mononuclear cells and adjacent normal tissue. Whole-exome capture libraries were constructed from 100 ng of DNA from tumor and normal tissue after sample shearing, end repair and phosphorylation and ligation to barcoded sequencing adaptors. Ligated DNA was size selected for lengths of 200–350 bp and subjected to exonic hybrid capture using Illumina library preps. The sample was multiplexed and sequenced using Illumina HiSeq technology. The Illumina exome sequencing approach uses Illumina’s in-solution DNA-probe-based hybrid selection method that applies principles similar to those of Broad Institute–Agilent Technologies’ in-solution RNA-probe-based hybrid selection method57,58 to generate Illumina exome sequencing libraries. Total RNA was assessed for quality using the Caliper LabChip GX2. The percentage of fragments with a size greater than 200 nucleotides (DV200) was calculated using software. An aliquot of 200 ng of RNA was used as the input for first-strand cDNA synthesis using Illumina’s TruSeq RNA Access Library Prep kit. Synthesis of the second strand of cDNA was followed by indexed adaptor ligation. Subsequent PCR amplification enriched for adaptor-ligated fragments. The amplified libraries were quantified using an automated PicoGreen assay.

the input for first-strand cDNA synthesis using Illumina’s TruSeq RNA Access Library Prep kit. Synthesis of the second strand of cDNA was followed by indexed adaptor ligation. Subsequent PCR amplification enriched for adaptor-ligated fragments. The amplified libraries were quantified using an automated PicoGreen assay. A total of 200 ng of each cDNA library, not including controls, was combined into four-plex pools. Capture probes that target the exome were added and hybridized for the recommended time. Following hybridization, streptavidin magnetic beads were used to capture the library-bound probes from the previous step. Two wash steps effectively removed any nonspecifically bound products. These same hybridization, capture and wash steps were repeated to assure high specificity. A second round of amplification enriched the captured libraries. After enrichment the libraries were quantified with qPCR using the KAPA Library Quantification kit for Illumina sequencing platforms and were then pooled at an equimolar ratio. The entire process was in a 96-well format and all pipetting was performed using Agilent Bravo or Hamilton Starlet. Pooled libraries were normalized to 2 nM and denatured using 0.2 N NaOH before sequencing. Flowcell cluster amplification and sequencing were performed according to the manufacturer’s protocols using either the HiSeq 2000 v.3 or HiSeq 2500. Each run generated 76-bp paired-end reads with a dual eight-base index barcode. Data were analyzed using the Broad Picard Pipeline, which includes demultiplexing and data aggregation.

ter amplification and sequencing were performed according to the manufacturer’s protocols using either the HiSeq 2000 v.3 or HiSeq 2500. Each run generated 76-bp paired-end reads with a dual eight-base index barcode. Data were analyzed using the Broad Picard Pipeline, which includes demultiplexing and data aggregation. Quality control and variant calling Initial exome sequence data processing and analysis were performed using pipelines at the Broad Institute. After alignment from the Broad Picard Pipeline, BAM files were uploaded into the Firehose infrastructure (https://software.broadinstitute.org/cancer/cga/Firehose), which managed intermediate analysis files executed by analysis pipelines. Sequencing data were incorporated into quality-control modules in Firehose to compare the tumor and normal genotypes and ensure concordance between samples. Quality-control cutoffs were as follows: mean target coverage >50× (tumor) and >20× (matched normal), cross-contamination of samples estimation (ContEst59) <5% and tumor purity ≥10% (Extended Data Fig. 1). Power calculation quality control To limit our analysis to samples where we had adequate power to call somatic variants, we performed a downstream per-sample power calculation. For each sample, we performed a Monte Carlo simulation of 1,000 true clonal mutations using the following procedures:Sample the number of reads from the sample-specific coverage distribution Draw the number of tumor reads from a binomial distribution using the estimated tumor purity

metrics resulting in overall poor quality when considering all metrics in aggregate as well as an aberrant expression profile compared to all other samples. Only transcriptomes from tumors whose WES also passed quality control were included; the final patient cohort for RNA-seq analysis included n = 121 transcriptomes. We excluded certain classes of noncoding genes that constituted a large (>10%) proportion of TPM in the majority of samples. Specifically we excluded genes characterized as snoRNA, using biomaRt76 to download Ensembl biotype annotations (using the dataset ‘hsapiens_gene_ensembl’ and version ‘GRCh38.p10’) and excluding genes whose biotype was ‘snoRNA’ (n = 380 genes). We then regenerated a new TPM metric for each sample to normalize the total transcriptome sum to 1 million. For analysis, only genes with TPM > 0 in 25% or more of the samples were included. This excluded 6,158 genes, with 20,848 genes passing this threshold. GSEA and ssGSEA GSEA32 was performed using the Cancer Hallmarks gene sets33 from MSigDB at https://cloud.genepattern.org/. We used default settings with 10,000 gene set permutations to generate P and q values, and we compared progressors and responders in the overall cohort, the ipilimumab-treated subgroup and the ipilimumab-naive subgroup separately.

Power calculation quality control To limit our analysis to samples where we had adequate power to call somatic variants, we performed a downstream per-sample power calculation. For each sample, we performed a Monte Carlo simulation of 1,000 true clonal mutations using the following procedures:Sample the number of reads from the sample-specific coverage distribution Draw the number of tumor reads from a binomial distribution using the estimated tumor purity Draw the number of mutation reads from a binomial distribution given the assumption of a heterozygous mutation and no copy number variation Characterize the mutation as detected or not on the basis of a log odds threshold of 6.3 (consistent with the MuTect60 implementation). The estimated power to detect clonal mutations is the proportion of simulated mutations detected (for example, 800 detected out of 1,000 simulated clonal mutations is 80% power), which is a function of both sample-specific sequencing depth of coverage and tumor purity. Three tumors were excluded using this threshold (Extended Data Fig. 1). Variant calling The MuTect algorithm60 was applied to identify somatic single-nucleotide variants in targeted exons, with computational filtering of artifacts introduced by DNA oxidation during sequencing61 or FFPE-based DNA extraction using a filter-based method. Strelka62 was applied to identify small insertions or deletions. Identified alterations were annotated using Oncotator63.

atic single-nucleotide variants in targeted exons, with computational filtering of artifacts introduced by DNA oxidation during sequencing61 or FFPE-based DNA extraction using a filter-based method. Strelka62 was applied to identify small insertions or deletions. Identified alterations were annotated using Oncotator63. TMB and neoantigen load For purposes of analysis, the TMB was calculated as the log of the number of nonsynonymous mutations detected from WES. Mutations per Mb was calculated by dividing the total number of nonsynonymous mutations by the number of bases with sufficient coverage in the tumor and normal samples (≥14× and ≥8×, respectively) to call mutations (https://software.broadinstitute.org/cancer/cga/mutect). Neoantigen prediction was performed as previously described10. Briefly, HLA type was inferred using POLYSOLVER64, which uses a Bayesian classifier to determine the genotype of each patient. Neoantigens were predicted for each patient by defining all novel nine and ten amino-acid sequences resulting from mutations and determining whether the predicted binding affinity to the patient’s germline HLA alleles was <500 nM using NetMHCpan (v.2.4)65.

Bayesian classifier to determine the genotype of each patient. Neoantigens were predicted for each patient by defining all novel nine and ten amino-acid sequences resulting from mutations and determining whether the predicted binding affinity to the patient’s germline HLA alleles was <500 nM using NetMHCpan (v.2.4)65. Copy number variants The total number of copy number alterations for individual tumors was inferred using adaptations of a binary segmentation algorithm66 (CapSeg) comparing fractional exon coverage for tumor segments to a panel of normal samples, generating exomic segments and segment copy number. Copy number data were inspected visually and manually for focal amplifications and deletions and genes were annotated with Oncotator63. For allelic copy number alterations, heterozygous single-nucleotide polymorphisms were identified and integrated into the binary segmentation algorithm (Allelic CapSeg) and allelic segments were further adjusted for tumor purity and ploidy using estimates derived from ABSOLUTE67. We then called allelic amplifications and deletions, following previously described methodology and criteria68 integrating segment focality and the purity- and ploidy-corrected allelic copy number.

ic CapSeg) and allelic segments were further adjusted for tumor purity and ploidy using estimates derived from ABSOLUTE67. We then called allelic amplifications and deletions, following previously described methodology and criteria68 integrating segment focality and the purity- and ploidy-corrected allelic copy number. Purity and ploidy Purity and ploidy were estimated using the ABSOLUTE algorithm67, which integrates variant allele frequency distributions and copy number variants to estimate absolute tumor purity and ploidy and infer the cancer cell fraction, the proportion of cancer cells in the sample that contain each mutation. Allelic segments following purity and ploidy correction were used to estimate allelic copy number. Heterogeneity and aneuploidy Heterogeneity was estimated as the proportion of mutations in each sample that were inferred to be subclonal. Clonal mutations were defined as having a cancer cell fraction ≥0.8, while other mutations were defined as subclonal; we chose this definition as a simple conservative approach with high specificity. To estimate aneuploidy, we used the proportion of the genome inferred to have an allelic amplification or deletion, using the allelic segmentation described above.

Heterogeneity and aneuploidy Heterogeneity was estimated as the proportion of mutations in each sample that were inferred to be subclonal. Clonal mutations were defined as having a cancer cell fraction ≥0.8, while other mutations were defined as subclonal; we chose this definition as a simple conservative approach with high specificity. To estimate aneuploidy, we used the proportion of the genome inferred to have an allelic amplification or deletion, using the allelic segmentation described above. Mutational signatures De novo mutational signatures were generated in this cohort using an adaptation of non-negative matrix factorization69 via the Brunet update method70, as previously described in detail51, with the R package SomaticSignatures71 and non-negative matrix factorization72. Cosine similarity was used to compare the discovered signatures to the 30 existing discovered and validated signatures in COSMIC25,73, with a threshold of 0.85, and we also manually visualized and inspected similarities in mutational motifs between our signatures and COSMIC signatures.

med using the Cancer Hallmarks gene sets33 from MSigDB at https://cloud.genepattern.org/. We used default settings with 10,000 gene set permutations to generate P and q values, and we compared progressors and responders in the overall cohort, the ipilimumab-treated subgroup and the ipilimumab-naive subgroup separately. To generate nonparametric gene set scores in individual samples, we generated ssGSEA projections43 for gene sets using rank normalization, including the MHC-II HLA genes (HLA-DMA, HLA-DMB, HLA-DOA, HLA-DOB, HLA-DPA1, HLA-DPB1, HLA-DQA1, HLA-DQA2, HLA-DQB1, HLA-DQB2, HLA-DRA, HLA-DRB1 and HLA-DRB5) and MHC-I HLA genes (HLA-A, HLA-B, HLA-C, HLA-E, HLA-F, TAP1, TAP2 and B2M). ssGSEA scores were also generated for Cancer Hallmarks gene sets. Absolute immune infiltrate and immune subsets Estimation of the total immune infiltrate in each sample and immune cell subsets was performed using CIBERSORT with the LM22 gene set41 on the CIBERSORT website (http://cibersort.stanford.edu). The absolute mode was enabled and quantile normalization was disabled using the RNA-seq TPM matrix for the cohort. A separate immune infiltrate and immune cell subsets score analysis was performed using single-cell-derived signatures, and the methodology used for generating normalized signature scores was as described by Jerby-Arnon and colleagues42.

Absolute immune infiltrate and immune subsets Estimation of the total immune infiltrate in each sample and immune cell subsets was performed using CIBERSORT with the LM22 gene set41 on the CIBERSORT website (http://cibersort.stanford.edu). The absolute mode was enabled and quantile normalization was disabled using the RNA-seq TPM matrix for the cohort. A separate immune infiltrate and immune cell subsets score analysis was performed using single-cell-derived signatures, and the methodology used for generating normalized signature scores was as described by Jerby-Arnon and colleagues42. As there is no gold standard for inferring immune infiltrate from bulk RNA-seq data, we chose to use CIBERSORT-inferred values throughout the analysis to allow comparisons with other studies, as CIBERSORT has been widely used (for example, in a pan-TCGA immune landscape analysis77).

A separate immune infiltrate and immune cell subsets score analysis was performed using single-cell-derived signatures, and the methodology used for generating normalized signature scores was as described by Jerby-Arnon and colleagues42. As there is no gold standard for inferring immune infiltrate from bulk RNA-seq data, we chose to use CIBERSORT-inferred values throughout the analysis to allow comparisons with other studies, as CIBERSORT has been widely used (for example, in a pan-TCGA immune landscape analysis77). Gene expression signatures Published gene expression signatures related to immune-checkpoint response were collected from the literature and validated in our cohort18,19,21,34–40. Sample-wise scores for these gene signatures were calculated from RNA-seq data using TPM values and following the methodology described in corresponding studies. Genes with unavailable expression data were excluded from calculations of gene signature scores. In two gene signatures39,40, genes were incorporated independently (that is, weighted) into the published model, but neither the direction nor the coefficient was available, so these signatures were excluded from evaluation. Differences in immune signature scores between responders (CR and PR) and progressors (PD) across all samples, in the ipilimumab-naive subset and in the ipilimumab-treated subset were tested using the Mann–Whitney U-test. The predictive utility of these immune signatures was evaluated with AUC values derived from ROC curves of gene signature scores in the complete cohort, ipilimumab-naive subset and the ipilimumab-treated subset. Results and detailed descriptions of evaluated gene signatures are provided in Supplementary Table 6.

-test. The predictive utility of these immune signatures was evaluated with AUC values derived from ROC curves of gene signature scores in the complete cohort, ipilimumab-naive subset and the ipilimumab-treated subset. Results and detailed descriptions of evaluated gene signatures are provided in Supplementary Table 6. Analysis Two primary response comparisons were made: (1) responders (defined as having CR or PR as the best RECIST response) versus progressors (defined as having PD as the best RECIST response) and (2) progressors (defined as having non-PD as the best RECIST response) versus non-progressors. Statistical tests were performed utilizing the Python scipy.stats package. To compare numeric features between response categories, including transcriptome expression, a nonparametric MWW rank-sum test (mannwhitneyu() function) was used to minimize the effects of outliers. For comparison of the proportion between response categories, a chi-squared test (chi2_contingency() function) was utilized. For association of binary variables (for example, association of gene alteration with responders versus progressors), a Fisher’s exact test (fisher_exact() function) was utilized to generate a P value. A conservative adjusted OR was generated by repeating the Fisher’s exact test, adding one to both the number of gene-mutant responders and progressors. All tests were two sided unless otherwise indicated.

responders versus progressors), a Fisher’s exact test (fisher_exact() function) was utilized to generate a P value. A conservative adjusted OR was generated by repeating the Fisher’s exact test, adding one to both the number of gene-mutant responders and progressors. All tests were two sided unless otherwise indicated. Survival analyses were performed utilizing the Python lifelines package78. For Kaplan–Meier curve survival analysis, a log-rank test (logrank_test() function) was used to compare survival curves. Hierarchical clustering was performed using the clustermap() function from the Python seaborn package79, with default settings including a Euclidean distance metric and the ‘single’ method of calculating cluster distance (minimization of the nearest point between clusters). Validation For validation, we reviewed the literature and found three studies18,20,44 of advanced melanoma treated with anti-PD1 ICB with response, WES and RNA-seq data. However, one did not have information on previous ipilimumab treatment20, and another18 had only two patients who were naive to ipilimumab and nine who were treated with ipilimumab with post-ipilimumab tumor biopsies and available WES and NanoString data; thus, we used the remaining cohort44 as our primary validation cohort.

r, one did not have information on previous ipilimumab treatment20, and another18 had only two patients who were naive to ipilimumab and nine who were treated with ipilimumab with post-ipilimumab tumor biopsies and available WES and NanoString data; thus, we used the remaining cohort44 as our primary validation cohort. To allow appropriate validation, only cutaneous, occult, acral and mucosal samples were included from validation cohorts; specifically, uveal and ocular melanomas were excluded (Riaz cohort, n = 5 excluded). Only patients with evaluable response criteria were included (Riaz cohort, n = 2 excluded). WES, transcriptomic and heterogeneity data were obtained from https://github.com/riazn/bms038_analysis. Fragments per kilobase of transcript per million mapped reads values were converted to TPM to be consistent with our cohort normalization.

with evaluable response criteria were included (Riaz cohort, n = 2 excluded). WES, transcriptomic and heterogeneity data were obtained from https://github.com/riazn/bms038_analysis. Fragments per kilobase of transcript per million mapped reads values were converted to TPM to be consistent with our cohort normalization. Predictive model generation and cross-validation We used logistic regression for our model to predict PD as the best RECIST response versus non-PD rather than responder versus progressor to better reflect the real-world setting where all outcomes (PD, SD, MR, PR and CR) are possible. We evaluated genomic, transcriptomic and clinical features. Categorical features were converted to binary features for each categorical value. To be conservative, no gene-level mutations or expression values were individually considered. Global genomic tumor characteristics such as TMB, purity, ploidy, heterogeneity and aneuploidy were considered. Features were generated from the transcriptome, including ssGSEA values for gene sets representing Cancer Hallmarks pathways, and MHC-II and MHC-I antigen-presentation genes, as well as gene expression signatures following the methodology in the respective publications, as described above and in Supplementary Table 6. Clinical characteristics including LDH and ECOG at the start of anti-PD1 ICB, the number of metastatic organs, sex, M stage, the number of different metastatic sites, metastatic sites and melanoma subtype were evaluated (Supplementary Table 1). Features were chosen in a forward-selection-based process, where features that were significantly predictive (P < 0.05) when added to the base model were ranked on the basis of the ability of the combined model to discriminate outcomes (using ROC curve AUC as the metric), and the best feature was chosen to be added to the base model. Potential features were evaluated on the basis of a manual review considering biological interpretability and clinical applicability. This process was iterated with the new base model and stopped when no features under consideration were significantly predictive.

e best feature was chosen to be added to the base model. Potential features were evaluated on the basis of a manual review considering biological interpretability and clinical applicability. This process was iterated with the new base model and stopped when no features under consideration were significantly predictive. The set of tumors with both WES and RNA-seq data was smaller than the set of tumors with only WES data; when the features chosen in model development for ipilimumab-naive tumors resulted in only WES features being chosen, model development was repeated in the superset of tumors requiring only clinical and WES data, and this model in the larger set is reported in the main text. To estimate the ‘out-of-bag’ AUC, we used k-fold cross-validation (splitting the dataset into k-subsets, training on k − 1 subsets and calculating AUC on the holdout subset) and calculated the mean cross-validation AUC. Given the partially manual review of features, feature selection was not included in cross-validation. For the ipilimumab-treated subset (n = 34), we chose k = 5 folds, and for the larger ipilimumab-naive subset (n = 85), we chose k = 10 folds, to maintain a cross-validation holdout set of >5 tumors. Cross-validation scores were calculated using the cross_val_score function from the Python sklearn package.

s-validation. For the ipilimumab-treated subset (n = 34), we chose k = 5 folds, and for the larger ipilimumab-naive subset (n = 85), we chose k = 10 folds, to maintain a cross-validation holdout set of >5 tumors. Cross-validation scores were calculated using the cross_val_score function from the Python sklearn package. To further evaluate the statistical support for our models, we calculated the Akaike information criteria and BIC of each subsequent model after adding an additional feature in forward selection in the ipilimumab-experienced and ipilimumab-naive subgroups (Extended Data Fig. 8c,d), and we also evaluated the addition of TMB as an additional feature to the selected models.

odels, we calculated the Akaike information criteria and BIC of each subsequent model after adding an additional feature in forward selection in the ipilimumab-experienced and ipilimumab-naive subgroups (Extended Data Fig. 8c,d), and we also evaluated the addition of TMB as an additional feature to the selected models. Permutation testing To test the statistical significance of differences in FDR q values of enriched pathways between ipilimumab-naive versus ipilimumab-experienced subgroups, we performed a permutation analysis. Briefly, we shuffled the ipilimumab-experienced and ipilimumab-naive labels for each tumor, keeping each subgroup size the same and keeping the same number of responders and progressors in each subgroup, and we reran GSEA on each new simulated subgroup. We repeated this 1,000 times to generate a distribution of enriched pathways in each subgroup under the null hypothesis of no relationship between subgroup and pathway enrichment. We then compared our observed outcome within this null distribution to generate an empiric P value. For example, for each pathway enriched in ipilimumab-experienced patients, the proportion of simulations with a difference in log q value between ipilimumab-experienced and ipilimumab-naive subgroups (equivalent to the ratio of q values) greater than or equal to our observed difference, and with an FDR q value in the ipilimumab-experienced subgroup equal to or more extreme than our observed ipilimumab-experienced q value would represent the empiric P value.

ween ipilimumab-experienced and ipilimumab-naive subgroups (equivalent to the ratio of q values) greater than or equal to our observed difference, and with an FDR q value in the ipilimumab-experienced subgroup equal to or more extreme than our observed ipilimumab-experienced q value would represent the empiric P value. We performed a similar permutation test to generate an empiric P value for the predictive models for ipilimumab-experienced and ipilimumab-naive subgroups. Briefly, we permuted the outcome labels (progressors and non-progressors) within each subgroup, and generated an AUC and cross-validation AUC for the predictive model with the specified features (that is, MHC-II, LDH and lymph node metastasis for ipilimumab-experienced tumors; purity, ploidy and heterogeneity for ipilimumab-naive tumors) to generate a null distribution of AUCs and cross-validation AUCs under the null hypothesis that these predictors are not associated with outcomes. By permuting the phenotype rather than the predictors, we preserved the inter-predictor structure. Then, the proportion of simulations with AUC and cross-validation AUC greater than our observed AUC represented an empiric P value. Reporting Summary Further information on research design is available in the Nature Research Reporting Summary linked to this article.

We performed a similar permutation test to generate an empiric P value for the predictive models for ipilimumab-experienced and ipilimumab-naive subgroups. Briefly, we permuted the outcome labels (progressors and non-progressors) within each subgroup, and generated an AUC and cross-validation AUC for the predictive model with the specified features (that is, MHC-II, LDH and lymph node metastasis for ipilimumab-experienced tumors; purity, ploidy and heterogeneity for ipilimumab-naive tumors) to generate a null distribution of AUCs and cross-validation AUCs under the null hypothesis that these predictors are not associated with outcomes. By permuting the phenotype rather than the predictors, we preserved the inter-predictor structure. Then, the proportion of simulations with AUC and cross-validation AUC greater than our observed AUC represented an empiric P value. Reporting Summary Further information on research design is available in the Nature Research Reporting Summary linked to this article. Online content Any methods, additional references, Nature Research reporting summaries, source data, extended data, supplementary information, acknowledgements, peer review information; details of author contributions and competing interests; and statements of data and code availability are available at 10.1038/s41591-019-0654-5. Supplementary information Reporting Summary Supplementary Data 1 Gene-level mutation calls; copy number alterations; immune cell signatures; mutational signatures; and ssGSEA signatures using Hallmarks gene sets. Supplementary Data 2 RNA-seq TPM matrix.

Online content Any methods, additional references, Nature Research reporting summaries, source data, extended data, supplementary information, acknowledgements, peer review information; details of author contributions and competing interests; and statements of data and code availability are available at 10.1038/s41591-019-0654-5. Supplementary information Reporting Summary Supplementary Data 1 Gene-level mutation calls; copy number alterations; immune cell signatures; mutational signatures; and ssGSEA signatures using Hallmarks gene sets. Supplementary Data 2 RNA-seq TPM matrix. Supplementary Tables Supplementary Tables 1–9. Extended data Extended Data Fig. 1 Consort Diagram showing inclusion, exclusion, and quality control criteria for patients/tumors included in analysis. Extended Data Fig. 2 Survival and genomic copy number characteristics of responders (n = 55 patients) (defined as CR or PR as best response) vs. progressors (n = 65 patients) (defined as PD as best response). (a) PFS Kaplan Meier survival curves; two-sided log-rank p = 2.1e-28 (b) OS Kaplan Meier survival curves; two-sided log-rank p = 5.1e-18 (c) Proportion of tumor genome with copy number alterations (two-sided MWW p = 0.09). Boxplots: Box limits indicate the interquartile range (25th-75th percentile), with a center line indicating the median. Whiskers show the value ranges up to 1.5xIQR above the 75th or below the 25th percentile, with outliers beyond those ranges shown as individual points. PD = progressive disease. CR = complete response. PR = partial response.

icate the interquartile range (25th-75th percentile), with a center line indicating the median. Whiskers show the value ranges up to 1.5xIQR above the 75th or below the 25th percentile, with outliers beyond those ranges shown as individual points. PD = progressive disease. CR = complete response. PR = partial response. Extended Data Fig. 3 Association of Genomic Alterations in Genes with Response vs. Progression. For each gene, the association of a genomic alteration with response was tested using a two-sided Fisher’s Exact test, and the odds ratio (OR) and p-value calculated. The adjusted OR is a conservative estimate generated by adding 1 to both mutant gene responders and progressors, thus moving the OR estimate closer to 1. The x- and y-axis are on log scales, but labeled in the original units. (a) Association of gene mutations with response (n = 55 patients) vs. progression (n = 65 patients). (b) Association of gene amplifications with response (n = 55 patients) vs. progression (n = 65 patients). Extended Data Fig. 4 Amplification of Chromosome 6 Regions in TAP2, HLA-A/B/C amplified tumors. Patients with amplifications in TAP2 or HLA-A/B/C are shown, with inferred amplifications in the relevant section of Chromosome 6 shown. 6 patients had amplifications in TAP2, 6 patients had amplifications in HLA-A/B/C, and 4 patients had amplifications of a chromosomal region including all four genes.

mplified tumors. Patients with amplifications in TAP2 or HLA-A/B/C are shown, with inferred amplifications in the relevant section of Chromosome 6 shown. 6 patients had amplifications in TAP2, 6 patients had amplifications in HLA-A/B/C, and 4 patients had amplifications of a chromosomal region including all four genes. Extended Data Fig. 5 Hierarchical clustering of the correlation matrix between transcriptional signatures previously associated with immunotherapy response and hallmark gene-sets, genomic, clinical, and transcriptomic features associated with response in our cohort. The color indicates the Pearson correlation between features, from perfect negative correlation (Pearson r = -1, blue) to perfect positive correlation (Pearson r = 1, red). Most previously hypothesized signatures cluster within an immune-activity related group with immune-related gene-sets. Sample size for each correlation depends on the number of available data points: correlations involving exclusively genomic or clinical data had n = 144 tumor samples, while correlations involving transcriptomic features had n = 121 tumor samples with data available.

ctivity related group with immune-related gene-sets. Sample size for each correlation depends on the number of available data points: correlations involving exclusively genomic or clinical data had n = 144 tumor samples, while correlations involving transcriptomic features had n = 121 tumor samples with data available. Extended Data Fig. 6 Expression of Genomic Features in Responders vs. Progressors in ipilimumab-treated and -naive subsets. Two-sided Mann-Whitney-Wilcoxon tests were used to compare responders to progressors (ipilimumab-treated: n = 16 responders/20 progressors; ipilimumab-naïve: n = 34 responders/37 progressors). All p-values are unadjusted. Log Nonsyn Mutload: Ipi-treated p = 0.15; Ipi-naïve p = 0.15. Log Clonal Mutload: Ipi-treated p = 0.17; Ipi-naïve p = 0.08. Heterogeneity: Ipi-treated p = 0.51; Ipi-naïve p = 0.057. Ploidy: Ipi-treated p = 0.45; Ipi-naïve p = 0.002. Boxplots: Box limits indicate the interquartile range (25th-75th percentile), with a center line indicating the median. Whiskers show the value ranges up to 1.5xIQR above the 75th or below the 25th percentile, with outliers beyond those ranges shown as individual points. Extended Data Fig. 7 Survival of patients stratified by progressive disease (PD) or non-PD (SD, MR, PR, or CR) best response (RECIST 1.1) to PD-1 ICB. (a) PFS (two-sided logrank p = 4.0e-38 (b) OS (two-sided logrank p = 1.2e-19).

Extended Data Fig. 6 Expression of Genomic Features in Responders vs. Progressors in ipilimumab-treated and -naive subsets. Two-sided Mann-Whitney-Wilcoxon tests were used to compare responders to progressors (ipilimumab-treated: n = 16 responders/20 progressors; ipilimumab-naïve: n = 34 responders/37 progressors). All p-values are unadjusted. Log Nonsyn Mutload: Ipi-treated p = 0.15; Ipi-naïve p = 0.15. Log Clonal Mutload: Ipi-treated p = 0.17; Ipi-naïve p = 0.08. Heterogeneity: Ipi-treated p = 0.51; Ipi-naïve p = 0.057. Ploidy: Ipi-treated p = 0.45; Ipi-naïve p = 0.002. Boxplots: Box limits indicate the interquartile range (25th-75th percentile), with a center line indicating the median. Whiskers show the value ranges up to 1.5xIQR above the 75th or below the 25th percentile, with outliers beyond those ranges shown as individual points. Extended Data Fig. 7 Survival of patients stratified by progressive disease (PD) or non-PD (SD, MR, PR, or CR) best response (RECIST 1.1) to PD-1 ICB. (a) PFS (two-sided logrank p = 4.0e-38 (b) OS (two-sided logrank p = 1.2e-19). Extended Data Fig. 8 Model selection criteria with increasing number of features by forward selection; cross-validation AUC distribution of final models. (a) Model AUC for the ipi-treated subgroup. Each point is labeled with the additional feature, and the final model includes MHC-II, Lymph node metastasis, and LDH below the median as features. (b) Model AUC for the ipi-naive subgroup. Each point is labeled with the additional feature, and the final model includes ploidy, heterogeneity, and purity as features. (c) Model AIC/BIC for the ipi-experienced subgroup. (d) Model AIC/BIC for the ipi-naïve subgroup. (e) Using 10-fold crossvalidation in the ipilimumab-naïve subset, mean crossvalidation AUC is 0.73, with a SD of 0.25. Each fold split the cohort into a training (n = 74 patients) and test set (n = 10 patients) ((f) In the ipi-experienced subset, mean crossvalidation AUC is 0.85, with a SD of 0.18. Each fold split the cohort into a training (n = 27 patients) and test set (n = 7 patients).

crossvalidation AUC is 0.73, with a SD of 0.25. Each fold split the cohort into a training (n = 74 patients) and test set (n = 10 patients) ((f) In the ipi-experienced subset, mean crossvalidation AUC is 0.85, with a SD of 0.18. Each fold split the cohort into a training (n = 27 patients) and test set (n = 7 patients). Extended Data Fig. 9 Subgroup Model ROC curves applied to other subgroups. (a) Ipilimumab-treated subgroup model applied to ipilimumab-naive (n = 74 patients) subgroup; (b) Ipilimumab-naive model applied to ipilimumab-treated (n = 45 patients) subgroup; (c) Ipi-naïve model predicting PD as best response to PD-1 ICB applied to pre-ipilimumab-treated (n = 15 patients) tumors. (top) Comut plot overlaying best response and the ipi-naïve model score. Each column is one of 15 ipilimumab-treated tumors biopsied prior to ipilimumab therapy, ordered by the predictive model score. The 5 highest scoring tumors were all PD. (bottom) Receiver-operator curve for our ipilimumab-naive predictive model applied to pre-ipilimumab tumors (n = 15), with an AUC of 0.71.

odel score. Each column is one of 15 ipilimumab-treated tumors biopsied prior to ipilimumab therapy, ordered by the predictive model score. The 5 highest scoring tumors were all PD. (bottom) Receiver-operator curve for our ipilimumab-naive predictive model applied to pre-ipilimumab tumors (n = 15), with an AUC of 0.71. Extended Data Fig. 10 Limited validation testing single feature models in validation cohort. (a) Relationship between response and MHC-II HLA score in the ipilimumab-treated subgroup of a validation cohort. Each column is a patient, and patients are sorted by MHC-II HLA score. (b) Validation of Predictive Model for ipilimumab-treated tumors. In a validation cohort (n = 23), a model using the only common available feature (MHC-II HLA score) had an AUC of 0.65 (log likelihood ratio p = 0.09, empiric p = 0.21, Methods). (c) Relationship between response and MHC-II HLA score in the ipilimumab-treated subgroup of a validation cohort. Each column is a patient, and patients are sorted by MHC-II HLA score. (d) Validation of Predictive Model for ipilimumab-naive Tumors. In a validation cohort (n = 20), a model using the only common available feature (heterogeneity) had an AUC of 0.73 (log likelihood ratio p = 0.17, empiric p = 0.066, Methods). Peer review information Saheli Sadanand was the primary editor on this article and managed its editorial process and peer review in collaboration with the rest of the editorial team. Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.

Extended Data Fig. 10 Limited validation testing single feature models in validation cohort. (a) Relationship between response and MHC-II HLA score in the ipilimumab-treated subgroup of a validation cohort. Each column is a patient, and patients are sorted by MHC-II HLA score. (b) Validation of Predictive Model for ipilimumab-treated tumors. In a validation cohort (n = 23), a model using the only common available feature (MHC-II HLA score) had an AUC of 0.65 (log likelihood ratio p = 0.09, empiric p = 0.21, Methods). (c) Relationship between response and MHC-II HLA score in the ipilimumab-treated subgroup of a validation cohort. Each column is a patient, and patients are sorted by MHC-II HLA score. (d) Validation of Predictive Model for ipilimumab-naive Tumors. In a validation cohort (n = 20), a model using the only common available feature (heterogeneity) had an AUC of 0.73 (log likelihood ratio p = 0.17, empiric p = 0.066, Methods). Peer review information Saheli Sadanand was the primary editor on this article and managed its editorial process and peer review in collaboration with the rest of the editorial team. Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. These authors contributed equally: David Liu, Bastian Schilling. Extended data is available for this paper at 10.1038/s41591-019-0654-5. Supplementary information is available for this paper at 10.1038/s41591-019-0654-5.

Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. These authors contributed equally: David Liu, Bastian Schilling. Extended data is available for this paper at 10.1038/s41591-019-0654-5. Supplementary information is available for this paper at 10.1038/s41591-019-0654-5. Acknowledgements The authors thank the active investigators A. Gesierich (University Hospital Würzburg), J.C. Hassel (NCT Heidelberg), C. Pföhler (Saarland University Medical Center), E. Dabrowski (Ludwigshafen Medical Center), L.A. Schneider (University Medical Center Ulm), C. Weishaupt (University Hospital of Münster), K.G. Griewank (University Hospital Essen), E. Hadaschik (University Hospital Essen), G. Kyriakakis (Laikon General Hospital), F. Meier (NCT Dresden), M.H. Geukes Foppen (The Netherlands Cancer Institute), R. Dummer (University Hospital Zürich), E. Bräunlein (Technical University Munich) and M. Boxberg (Technical University Munich) for providing additional patient samples and clinical data. The authors also thank A. Giobbie-Hurder (Dana-Farber Cancer Institute) and J. Weirather (Dana-Farber Cancer Institute) for useful discussions regarding statistical testing. This work was supported by the Adelson Medical Research Foundation (L.G.), the Conquer Cancer Foundation (David Liu), the Society for Immunotherapy of Cancers (David Liu), the Damon Runyon Cancer Research Foundation (David Liu), the BroadNext10 (E.M.V.), the National Institutes of Health (K08 CA234458 (David Liu), R01 CA227388 (E.M.V.), U01 CA233100 (E.M.V.), T32 GM008313 (M.X.H.)), the Deutsche Forschungsgemeinschaft (German Research Foundation), SCHA 422/17-1, PA 2376/1-1 and HO 6389/2-1 (KFO 337; D.S, A.P.) and National Science Foundation Graduate Research Fellowship Program DGE1144152 (M.X.H.).

lth (K08 CA234458 (David Liu), R01 CA227388 (E.M.V.), U01 CA233100 (E.M.V.), T32 GM008313 (M.X.H.)), the Deutsche Forschungsgemeinschaft (German Research Foundation), SCHA 422/17-1, PA 2376/1-1 and HO 6389/2-1 (KFO 337; D.S, A.P.) and National Science Foundation Graduate Research Fellowship Program DGE1144152 (M.X.H.). Author contributions David Liu, B.S., L.G., A.R., K.F., E.M.V.A. and D.S. conceived and designed the overall study. B.S. collected and reviewed all clinical data. A.S. performed sample processing and shipping. B.S., E.L., L.Z., R.G., I.S., C.L., S.G., H.G., S.M.G., J.U., C.U.B, R.R., D.v.B., A.K., B.W., S.H. and F.K. provided samples and clinical annotation. A.T. oversaw sample processing and sequencing. David Liu, C.A.M., N.V., J.C., D.M., F.D. and M.X.H. designed and performed the mutational and copy number analyses. David Liu, Derek Liu, C.A.M., M.X.H., D.M., L.J.-A. and B.I. designed and performed RNA-seq analysis, including quality control, normalization, signature analysis, CIBERSORT and GSEA. David Liu, H.E. and M.X.H. designed and constructed the predictive modeling. David Liu, B.S., A.P., E.M.V.A. and D.S. interpreted the data. David Liu, B.S., E.M.V.A. and D.S. wrote the manuscript, and all authors reviewed and approved the final manuscript.

control, normalization, signature analysis, CIBERSORT and GSEA. David Liu, H.E. and M.X.H. designed and constructed the predictive modeling. David Liu, B.S., A.P., E.M.V.A. and D.S. interpreted the data. David Liu, B.S., E.M.V.A. and D.S. wrote the manuscript, and all authors reviewed and approved the final manuscript. Data availability All reasonable requests for raw and analyzed data and materials will be promptly reviewed by the senior authors to determine whether the request is subject to any intellectual property or confidentiality obligations. Patient-related data not included in the paper may be subject to patient confidentiality. Any data and materials that can be shared will be released via a material transfer agreement. All analyzed sequencing data are in supplementary tables or data available online. Raw sequencing data are available in dbGaP (accession number phs000452.v3.p1). Code availability Python code is available in packages as described in the manuscript. Code to regenerate figures using supplementary data and tables is available at GitHub at https://github.com/vanallenlab/schadendorf-pd1. Additional reasonable requests for code will be promptly reviewed by the senior authors to verify whether the request is subject to any intellectual property or confidentiality obligations, and shared to the extent permissible by these obligations.

available at GitHub at https://github.com/vanallenlab/schadendorf-pd1. Additional reasonable requests for code will be promptly reviewed by the senior authors to verify whether the request is subject to any intellectual property or confidentiality obligations, and shared to the extent permissible by these obligations. Competing interests David Liu reports funding by a postdoctoral fellowship from the Society for Immunotherapy of Cancers, which is funded in part by an educational grant from Bristol-Meyers Squibb (BMS). BMS has had no input into the conception, conduct or reporting of the submitted work. B.S. is on the advisory board or has received honoraria from Incyte, Novartis, Roche, BMS and Merck Sharp & Dohme (MSD), research funding from Pierre-Fabre Pharmaceuticals, BMS and MSD and travel support from Novartis, Roche, BMS, Pierre-Fabre Pharmaceuticals, MSD and Amgen, outside the scope of the submitted work. D.S. reports grants, personal fees and nonfinancial support from BMS, personal fees and nonfinancial support from Roche, grants, personal fees and nonfinancial support from Novartis, nonfinancial support from Regeneron, personal fees from Sanofi, personal fees and nonfinancial support from MSD, personal fees and nonfinancial support from Amgen, personal fees and nonfinancial support from 4SC, personal fees and nonfinancial support from Merck-EMD, personal fees from Array, personal fees and nonfinancial support from Pierre-Fabre, personal fees and nonfinancial support from Philogen, personal fees and nonfinancial support from Incyte and personal fees from Pfizer, outside the scope of the submitted work. E.M.V.A. reports advisory relationships and consulting with Tango Therapeutics, Genome Medical, Invitae, Illumina and Ervaxx; research support from Novartis and BMS; equity in Tango Therapeutics, Genome Medical, Syapse, Ervaxx and Microsoft; and travel reimbursement from Roche and Genentech, outside the submitted work. S.G. reports personal fees from MSD, personal fees from BMS, personal fees from Novartis and personal fees from Roche, outside the scope of the submitted work. L.Z. reports personal fees and others from BMS, personal fees and others from Novartis, personal fees and others from Pierre-Fabre, personal fees and others from MSD, personal fees from Roche, other fees from Amgen and personal fees from Sanofi, outside the scope of the submitted work. F.K.

scope of the submitted work. L.Z. reports personal fees and others from BMS, personal fees and others from Novartis, personal fees and others from Pierre-Fabre, personal fees and others from MSD, personal fees from Roche, other fees from Amgen and personal fees from Sanofi, outside the scope of the submitted work. F.K. reports personal fees from Amgen, personal fees from BMS, grants and personal fees from Novartis, personal fees from Roche, personal fees from MSD and personal fees from Merck, outside the submitted work. J.U. reports personal fees and nonfinancial support from Amgen, personal fees and nonfinancial support from BMS, personal fees and nonfinancial support from MSD, personal fees and nonfinancial support from Novartis, personal fees and nonfinancial support from Pierre-Fabre, and personal fees and nonfinancial support from Roche, outside the scope of the submitted work. H.G. reports grants and personal fees from BMS, grants and personal fees from Roche, grants and personal fees from MSD, grants and personal fees from Novartis, personal fees from Amgen and personal fees from Pierre-Fabre, outside the scope of the submitted work. R.G. reports personal fees and nonfinancial support from BMS, personal fees and nonfinancial support from Roche Pharma, personal fees and nonfinancial support from Merck Serono, grants, personal fees and nonfinancial support from Amgen, personal fees and nonfinancial support from Pierre-Fabre, personal fees and nonfinancial support from Sanofi Regeneron, personal fees from MSD, grants, personal fees and nonfinancial support from Novartis, personal fees and nonfinancial support from Almirall Hermal, grants and personal fees from Pfizer, personal fees from LEO, personal fees from AstraZeneca, personal fees from Sun Pharma, personal fees from 4SC and grants from Johnson & Johnson, outside the scope of the submitted work. I.S. reports grants and personal fees from Novartis, grants from Pfizer, personal fees from Roche, personal fees BMS and personal fees from MSD, outside the scope of the submitted work. C.L. reports personal fees from Roche, personal fees from Novartis, personal fees from Pierre-Fabre, personal fees from BMS, personal fees from MSD, personal fees from Amgen and personal fees from Idera and Sun Pharma, outside the scope of the submitted work. S.G. reports personal fees from BMS, personal fees from MSD and personal fees from Merck KGaA, outside the scope of the submitted work. E.L.

s from Pierre-Fabre, personal fees from BMS, personal fees from MSD, personal fees from Amgen and personal fees from Idera and Sun Pharma, outside the scope of the submitted work. S.G. reports personal fees from BMS, personal fees from MSD and personal fees from Merck KGaA, outside the scope of the submitted work. E.L. reports personal fees and others from Amgen, personal fees and others from BMS, personal fees and others from MSD, personal fees and others from Novartis, personal fees and others from Roche, personal fees from medac, personal fees from Janssen, other fees from Actelion and other fees from Pierre-Fabre, outside the scope of the submitted work. A.K. reports grants from BMS and Kiadis, personal fees from BMS, Sanofi, Novartis, Roche, Vaccibody and nonfinancial support from Sanofi and BMS, outside the scope of the submitted work. C.B. reports personal fees from BMS, MSD, Roche, Novartis, GSK, Pfizer, Lilly, Pierre-Fabre and GenMab and grants from BMS, Novartis and NanoString, outside the scope of the submitted work. S.H. reports personal fees from Novartis, personal fees from BMS, personal fees from Amgen, personal fees from Pierre-Fabre and personal fees from Roche, outside the scope of the submitted work. B.W. reports grants and personal fees from Philogen, personal fees from Curevac, grants and personal fees from MSD, grants and personal fees from BMS and personal fees from GSK, outside the scope of the submitted work. R.R. reports personal fees from Novartis, personal fees and nonfinancial support from Amgen and nonfinancial support from BMS, outside the scope of the submitted work. D.v.B., A.S. and A.P. have nothing to disclose.

Main Anomalously warm and cold weather events are an important public health concern in today’s world, and one of the key drivers for seeking adaptation measures against anthropogenic climate change3–5. Current assessments of the health effects of weather and climate, and by extension of global climate change, largely focus on parasitic and infectious diseases, and cardiorespiratory and other chronic diseases3–8. Less research has been conducted on injuries9–12, especially in a consistent way across injury types and demographic subgroups of the population. There are two reasons for investigating a potential role for temperature anomalies on injury mortality. First, death rates from injuries vary seasonally and the seasonality varies by age group13,14, which motivates investigating whether temperature contributes to their pathogenesis. Second, there are plausible behavioral and physiological pathways for a relationship between temperature and injury—for example, changes in alcohol drinking15, driving patterns and performance12,16–24, and levels of anger25–27—that motivate testing whether injury deaths are affected by temperature anomalies. Our aim was to evaluate how deaths from various injuries in the USA might be affected by anomalously warm temperatures that occur today and are expected to become increasingly common as a result of global climate change1.

s of anger25–27—that motivate testing whether injury deaths are affected by temperature anomalies. Our aim was to evaluate how deaths from various injuries in the USA might be affected by anomalously warm temperatures that occur today and are expected to become increasingly common as a result of global climate change1. We used vital registration data on all injury deaths in the contiguous USA (that is, excluding Alaska and Hawaii) from 1980 to 2017, with information on sex, age at death, underlying cause of death, and county and state of residence. From 1980 to 2017, 4,145,963 boys and men and 1,825,817 girls and women died from an injury in the contiguous USA, accounting for 9.3% and 4.2% of all male and female deaths, respectively; 95.7% of male injury deaths and 94% of female injury deaths were in those aged 15 years and older, and over half (52.3%) of male injury deaths were in those aged 15–44 years (Fig. 1). By contrast with males, there was less of an age gradient in females after age 15 years.Fig. 1 Number of injury deaths, by type of unintentional (transport, falls, drownings and other) and intentional (assault and suicide) injury, by sex and age group in the contiguous USA for 1980–2017. The top row shows the breakdown by type of injury and age group for males. The bottom row shows the breakdown by type of injury and age group for females.

type of unintentional (transport, falls, drownings and other) and intentional (assault and suicide) injury, by sex and age group in the contiguous USA for 1980–2017. The top row shows the breakdown by type of injury and age group for males. The bottom row shows the breakdown by type of injury and age group for females. Injuries from transport, falls, drownings, assault and suicide accounted for 78.6% of injury deaths in males and 71.8% in females. The remainder were from a heterogeneous group of ‘other injuries’ (Fig. 1), within which the composition of injuries that led to death varied by sex and age group. Transport was the leading cause of death by injury in women younger than 75 years and in men younger than 35 years. Between ages 35 and 74 years, more men died of suicide than any other injury. Above 75 years of age, falls were the largest cause of injury-related death in both men and women.

aried by sex and age group. Transport was the leading cause of death by injury in women younger than 75 years and in men younger than 35 years. Between ages 35 and 74 years, more men died of suicide than any other injury. Above 75 years of age, falls were the largest cause of injury-related death in both men and women. There was a decline in age-standardized death rates of three out of five major injuries (transport, drownings and assault) from 1980 to 2017, although assault death rates have increased more recently (since 2014) (Fig. 2). By contrast, age-standardized death rates from falls increased over time whereas those from suicide initially decreased, followed by an increase to surpass 1980 levels. The largest overall decline over time was for transport deaths in both sexes and for deaths from drownings in men, which declined by more than 50% from 1980 to 2017. Age-standardized death rates for transport injuries and drownings peaked in the summer months, but deaths from other major injuries did not have clear seasonal patterns.Fig. 2 National age-standardized death rates from 1980 to 2017, by type of injury, sex and month. The top row shows the breakdown by type of injury for males. The bottom row shows the breakdown by type of injury for females.

in the summer months, but deaths from other major injuries did not have clear seasonal patterns.Fig. 2 National age-standardized death rates from 1980 to 2017, by type of injury, sex and month. The top row shows the breakdown by type of injury for males. The bottom row shows the breakdown by type of injury for females. We defined a measure of anomalous temperature for each county and month, which represents the deviation from the county’s average temperature in that month over the entire analysis period (see Extended Data Fig. 1). County-level anomalies were aggregated to the state level with the use of population weights. This generated a number for each state and month that measured deviation from long-term average of the state in that month. Average size of anomaly over the study period (1980–2017), a measure of how variable temperatures are around their state–month long-term average, ranged from 0.4 °C for Florida in September to 3.4 °C for North Dakota in February (see Extended Data Fig. 2). Taken across all states and months, the average size of anomaly had a median value of 1.2 °C. Temperature anomalies were largest in January and December and smallest in August and September. In addition, they were larger in northern and central states than in southern and coastal ones.

ruary (see Extended Data Fig. 2). Taken across all states and months, the average size of anomaly had a median value of 1.2 °C. Temperature anomalies were largest in January and December and smallest in August and September. In addition, they were larger in northern and central states than in southern and coastal ones. We analyzed the association of monthly injury death rates with anomalous temperature using a Bayesian spatio-temporal model, described in detail in Methods. We used the resultant risk estimates, and the age- and sex-specific death rates from each injury in 2017, to estimate additional deaths if each month in each state were +1.5 °C above its long-term average, as envisioned under the Paris Climate Agreement2. We present additional results, based on +2 °C, which is the upper boundary of the Paris Climate Agreement, as Extended Data Figs. 3 and 4. Based on this analysis, there would be an estimated 1,601 (95% credible interval 1,430–1,776) excess injury deaths, equivalent to 0.75% of all injury deaths in 2017, in a year in which each month in each state was +1.5 °C warmer than its long-term average (Fig. 3). The number of excess injury deaths would increase to 2,135 (1,906–2,368), equivalent to 1.0% of all injury deaths in 2017, in each year in which each month in each state was +2 °C warmer than its long-term average (see Extended Data Fig. 3).Fig. 3 Additional annual injury deaths for the 2017 US population in the year in which each month was +1.5 °C warmer compared with 1980–2017 average temperatures. The top row shows the breakdown by type of injury, sex and age group. The bottom row shows the breakdown by type of injury, sex and month. Black dots represent net changes in deaths for each set of bars. See Extended Data Fig. 3 for results of the scenario of 2 °C warmer.

r compared with 1980–2017 average temperatures. The top row shows the breakdown by type of injury, sex and age group. The bottom row shows the breakdown by type of injury, sex and month. Black dots represent net changes in deaths for each set of bars. See Extended Data Fig. 3 for results of the scenario of 2 °C warmer. Deaths from drownings, transport, assault and suicide would increase, partly offset by a decline in deaths from falls in middle and older ages and in winter months (Fig. 3). Most excess deaths would be from transport injuries (739; 650–814 in the +1.5 °C warmer scenario) followed closely by suicide (540; 445–631). Of the excess deaths, 84% would occur in males and 16% in females. Of all male excess deaths, 92% would occur in those aged 15–64 years, who have higher rates of deaths from transport and suicide. In those aged 85 years and older, there would be an estimated decline in injury deaths, because deaths from falls are expected to decline in a warmer year.

occur in males and 16% in females. Of all male excess deaths, 92% would occur in those aged 15–64 years, who have higher rates of deaths from transport and suicide. In those aged 85 years and older, there would be an estimated decline in injury deaths, because deaths from falls are expected to decline in a warmer year. Proportionally, deaths from drownings are estimated to increase more than those of other injury types—by as much as 13.7% (12.5, 15.2) for a +1.5 °C anomaly in men aged 15–24 years (Fig. 4). The smallest proportional increase was that of assault and suicide (less than 3% in all age and sex groups). There was a larger percentage increase in transport deaths for males than for females, especially in young and middle ages (for example, 2.0% (1.6, 2.6) for 25- to 34-year-old men versus 0.5% (−0.3, 1.4) for women of the same age) (Fig. 4). We present additional results, based on +2 °C, in Extended Data Fig. 4.Fig. 4 Percentage change in death rates in year in which each month was +1.5 °C compared with 1980–2017 average temperatures by type of injury, sex and age group or month. a, Percentage change in death rates by injury, sex and age group. b, Percentage change in death rates by injury, sex and month. Colored dots show the posterior means and error bars represent 95% credible intervals, both obtained at the posterior draw level. See Extended Data Fig. 4 for the scenario of 2 °C warmer.

nth. a, Percentage change in death rates by injury, sex and age group. b, Percentage change in death rates by injury, sex and month. Colored dots show the posterior means and error bars represent 95% credible intervals, both obtained at the posterior draw level. See Extended Data Fig. 4 for the scenario of 2 °C warmer. That anomalously warm temperature influences deaths from drownings, although not previously quantified, is highly plausible because swimming is likely to be more common when the temperature is higher. The higher relative and absolute impacts on men compared with women may reflect differences in their behaviors. For example, over half of swimming deaths for males occur in natural water, compared with about a quarter for females28. The former may rise more in warmer weather. Similarly, deaths from falls declined in older ages, but increased in young ages, because falls in elderly people are more likely to be due to slipping on ice than falls in younger people29–31.

ths for males occur in natural water, compared with about a quarter for females28. The former may rise more in warmer weather. Similarly, deaths from falls declined in older ages, but increased in young ages, because falls in elderly people are more likely to be due to slipping on ice than falls in younger people29–31. The pathways from anomalous temperature to transport injury are more varied. First, driving performance deteriorates at higher temperatures20–23. Furthermore, alcohol consumption increases in warm temperatures15, which also provides an explanation for why teenagers, who are more likely than other age groups to crash while intoxicated32, could experience a larger proportional rise in deaths from transport, when temperatures are anomalously warm, than older adults. Last, warmer temperatures generally increase road traffic in North America12,16–19,24; coupled with more people outdoors in warmer weather33, this increase could lead to more fatal collisions.

ould experience a larger proportional rise in deaths from transport, when temperatures are anomalously warm, than older adults. Last, warmer temperatures generally increase road traffic in North America12,16–19,24; coupled with more people outdoors in warmer weather33, this increase could lead to more fatal collisions. Pathways linking anomalously high temperatures and deaths from assault and suicide are less established. One hypothesis is that more time spent outdoors in anomalously warmer temperatures leads to an increased number of face-to-face interactions, and hence arguments, confrontations and ultimately assaults34,35. These effects could be compounded by the greater anger levels linked to higher temperatures25–27. However, further research on the association of temperature and assault, and the factors mediating it, is needed36. Regarding suicide, it has been hypothesized that a higher temperature is associated with higher levels of distress in younger people37. Nevertheless, the mechanisms for the links between temperature and mental health require further investigation, including whether the relationship varies by age and sex, as indicated by our results. Future research should also investigate the extent to which the increased risk of injury death as a result of anomalous temperature depends on community characteristics, such as poverty and deprivation, social connectivity and cohesion, quality of roads and housing, public transportation options, emergency response and social services.

hould also investigate the extent to which the increased risk of injury death as a result of anomalous temperature depends on community characteristics, such as poverty and deprivation, social connectivity and cohesion, quality of roads and housing, public transportation options, emergency response and social services. The major strength of our study is that we have comprehensively modeled the association of temperature anomaly with injury by type of injury, month, age group and sex. Our measure of temperature anomaly internalizes the long-term historical experience of each state, and is closer to what climate change may bring about than solely examining daily episodes, or average temperature to which people have adapted. To utilize this metric, we integrated two large disparate national datasets on mortality (vital statistics) and meteorology (ERA5), and developed a bespoke Bayesian spatio-temporal model. A limitation of our study is that, like all observation studies, we cannot rule out confounding of results due to other factors. As described above, our statistical model by design adjusts for factors related to month, state and state–month that either are invariant over time or change linearly. Rather, the confounding factors would be those with anomalies similar to those of the monthly temperature in each state, such as air pollution. However, to our knowledge, there is currently no evidence of an association between air pollution and injury mortality. We analyzed the associations between anomalous temperature and injury mortality at the state level, because the small number of events and computational demands made county-level analyses unfeasible. Analyses at finer spatial resolution, such as county or district level38, would be ideal because the impacts of anomalously warm and cold temperature on deaths from injuries may depend on socioeconomic (for example, poverty, social connectivity and cohesion, availability of guns), environmental (for example, availability of swimming pools, distance to bodies of water), infrastructure (for example, quality and safety of roads, public transportation options), and health and social services (for example, counseling and mental health services, emergency response). We used categories of injuries that are relevant to public health purposes and for designing and implementing interventions. It may be possible to further split each category.

roads, public transportation options), and health and social services (for example, counseling and mental health services, emergency response). We used categories of injuries that are relevant to public health purposes and for designing and implementing interventions. It may be possible to further split each category. For example, 92% of all transport injuries in males and 96% in females are from road traffic injuries, with the remainder being classified as other transport injuries (see Extended Data Fig. 5). Similarly, suicides can be classified based on the means of suicide. To the extent that these subcategories are relevant for interventions, they should be separately analyzed in future studies. Finally, as with any Bayesian model, choices of prior distributions and hyper-parameters are necessary. There are alternatives to the priors we used. For example, our weakly informative gamma priors could have been replaced by penalized complexity priors39 or uniform priors on the standard deviation scale40. We tested a limited number of alternatives and found that our results were robust to such specifications.

cessary. There are alternatives to the priors we used. For example, our weakly informative gamma priors could have been replaced by penalized complexity priors39 or uniform priors on the standard deviation scale40. We tested a limited number of alternatives and found that our results were robust to such specifications. Our work highlights how deaths from injuries are currently susceptible to temperature anomalies and could also be modified by rising temperatures resulting from climate change, unless countered by social infrastructure and health system interventions that mitigate these impacts. Although absolute impacts on mortality are modest, some groups, especially men who are young to middle aged, experience larger impacts than other age and sex groups. Therefore, a combination of public health interventions that broadly target injuries in these groups—for example, targeted messaging for younger males on the risks of transport injury and drowning—and those that trigger in relation to forecast high-temperature periods—for example, additional targeted blood alcohol level checks—should be a public health priority.

terventions that broadly target injuries in these groups—for example, targeted messaging for younger males on the risks of transport injury and drowning—and those that trigger in relation to forecast high-temperature periods—for example, additional targeted blood alcohol level checks—should be a public health priority. Methods Data sources We used data on deaths by sex, age, underlying cause of death and state of residence in the contiguous USA from 1980 to 2017 through the National Center for Health Statistics (NCHS) (https://www.cdc.gov/nchs/nvss/dvs_data_release.htm) and on the population from the NCHS bridged-race dataset for 1990–2017 (https://www.cdc.gov/nchs/nvss/bridged_race.htm) and from the US Census Bureau before 1990 (https://www.census.gov/data/tables/time-series/demo/popest/1980s-county.html). We did not include Alaska and Hawaii (which together made up 0.5% of the US population in 2017) because their climates and environment are distinct from other states due to their substantial physical distance. We calculated monthly population counts through linear interpolation, assigning each yearly count to July.

l). We did not include Alaska and Hawaii (which together made up 0.5% of the US population in 2017) because their climates and environment are distinct from other states due to their substantial physical distance. We calculated monthly population counts through linear interpolation, assigning each yearly count to July. The underlying cause of death was coded according to the International Classification of Diseases (ICD) system (the 9th revision from 1980 to 1998 and the 10th revision thereafter). The six million injury deaths fell into six categories: transport, falls, drownings, assault, suicide and an aggregate set of other injuries (see Supplementary Table 1). We report the results of all of these categories except other injuries (1,402,941 deaths or 23% of total injury deaths during 1980–2017), because the composition of this aggregate group varies by sex, age group, state and time. We obtained data on temperature from ERA5, which uses data from global in situ and satellite measurements to generate a worldwide meteorological dataset, with full space and time coverage over our analysis period41. We used gridded estimates measured four times daily at a resolution of 30 km to generate monthly temperatures by county.

on temperature from ERA5, which uses data from global in situ and satellite measurements to generate a worldwide meteorological dataset, with full space and time coverage over our analysis period41. We used gridded estimates measured four times daily at a resolution of 30 km to generate monthly temperatures by county. Anomalous temperature metric With few exceptions9,42, current climate change risk assessments extrapolate from associations of daily mortality with daily temperature7,8,43–45. Climate change will, however, fundamentally modify weather, including seasonal weather patterns, compared with long-term averages, and hence can disrupt existing forms of adaptation. To mimic the conditions that may arise with global climate change, we developed methodology to examine how deviations from the long-term average temperature may impact injury death rates.

weather, including seasonal weather patterns, compared with long-term averages, and hence can disrupt existing forms of adaptation. To mimic the conditions that may arise with global climate change, we developed methodology to examine how deviations from the long-term average temperature may impact injury death rates. We first defined a measure of anomalous temperature for each county and month, which represents the deviation from the average temperature of the county in that month over the entire analysis period. To calculate the magnitude of temperature anomaly, we first calculated average temperatures for each month in each county over the entire 38 years of analysis. We subtracted these long-term average temperatures from respective monthly temperature values to generate a temperature anomaly time series for each month and year in each county (see Extended Data Fig. 1). The temperature anomaly metric measures the extent to which the temperature experienced in a specific month, year and county is warmer or cooler than the long-term average to which the population has acclimatized. These values can be different for different months in the same county, and different counties in the same month. Furthermore, a county with a higher, but more stable, temperature in a specific month has smaller anomalies than one with a lower, but more inter-annually variable, temperature. County-level anomalies were aggregated to the state level with the use of population weights for analyzing their associations with mortality.

Furthermore, a county with a higher, but more stable, temperature in a specific month has smaller anomalies than one with a lower, but more inter-annually variable, temperature. County-level anomalies were aggregated to the state level with the use of population weights for analyzing their associations with mortality. Statistical methods We analyzed the association of monthly injury death rates with anomalous temperature using a Bayesian spatio-temporal model, which leveraged variations over space and time to infer associations. We modeled the number of deaths in each month in each year as following a Poisson distribution: deathsstate-time~Poisson(deathratestate-time×populationstate-time) with the log-transformed death rates modeled as a sum of components that depend on location (state) of death, month of year, overall time (in months) and temperature anomaly: logdeathratestate-time=(α0+β0×time)+(αstate+βstate×time)+(αmonth+βmonth×time)+ζstate−month+(ψstate-month×time)+νtime+(γmonth×anomalystate-time)+εstate-time The model contained terms that represent the national level and trend in mortality, with α0 as the common intercept and β0 the common slope with overall time. Death rates also vary by month, which may be partly related to temperature and partly due to other monthly factors; monthly variations tend to be smooth across adjacent months13. Therefore, we allowed each month of the year to systematically have a different mortality level and trend, with αmonth the month-specific intercept and βmonth the month-specific slope with overall time. We used a first-order random walk prior for the monthly random intercepts and slopes, widely used to characterize smoothly varying trends46. The random walk had a cyclic structure, so that December was adjacent to January.

, with αmonth the month-specific intercept and βmonth the month-specific slope with overall time. We used a first-order random walk prior for the monthly random intercepts and slopes, widely used to characterize smoothly varying trends46. The random walk had a cyclic structure, so that December was adjacent to January. We also included state random intercepts and slopes for death rates, with αstate as the state-specific intercept and βstate the state-specific slope with overall time. These terms measure deviations of each state from national values, and allow variation in level and trend in mortality by state. We modeled the state-level random intercepts and slopes using the Besag, York and Mollie spatial model47, which includes both spatially structured random effects with an intrinsic conditional autoregressive prior and spatially unstructured, independent and identically distributed gaussian random effects. The extent to which information is shared between neighboring states depends on the uncertainty of death rates in a state and the empirical similarity of death rates in neighboring states. We also included state–month interactions for intercepts and slopes (ζstate–month and ψstate–month), to allow variation in mortality levels and trends in a particular state for different months and vice versa. These state–month interactions were modeled as independent and identically distributed, and therefore were of type I space–time interactions48. Non-linear change over overall time (in months) was captured by a first-order random walk, vtime46. To ensure identifiability, each set of random walk terms or state random effects was constrained to sum to zero.

modeled as independent and identically distributed, and therefore were of type I space–time interactions48. Non-linear change over overall time (in months) was captured by a first-order random walk, vtime46. To ensure identifiability, each set of random walk terms or state random effects was constrained to sum to zero. Finally, we included a term that relates log-transformed death rate to the above-defined state–month temperature anomaly, γmonth × anomalystate–time. The coefficients of γmonth represent the logarithm of the monthly death rate ratio per 1 °C increase in anomaly. There was a separate coefficient for each month, which means that an anomaly of the same magnitude could have different associations with injury mortality in different months. As with the month-specific intercepts and trends, we used a cyclic first-order random walk to smooth the coefficient of the temperature anomaly across months. An over-dispersion term (εstate–time) captured the variation unaccounted for by other terms in the model, modeled as N(0,σϵ2). We used weakly informative priors so that parameter estimation was driven by the data. As in previous analyses49,50, hyper-priors were defined on the logarithm of the precisions of the random effects, in other words on log(1/𝜎2). These were modeled as logGamma(𝜃, 𝜹) distributions with shape 𝜃 = 1 and rate 𝜹 = 0.001. The same hyper-priors were used for all precision parameters of the random effects in the model. For the common slope, we used N(0, 1,000) and for the common intercept a flat prior.

f the random effects, in other words on log(1/𝜎2). These were modeled as logGamma(𝜃, 𝜹) distributions with shape 𝜃 = 1 and rate 𝜹 = 0.001. The same hyper-priors were used for all precision parameters of the random effects in the model. For the common slope, we used N(0, 1,000) and for the common intercept a flat prior. In addition to representing the spatial (across states) and temporal (across months and years) patterns of mortality, the intercept terms (αmonth, αstate, ζstate–month) in our statistical model implicitly adjust for unobserved factors that influence mortality at the state, month and state–month level; the slope terms (βmonth, βstate, ψstate–month) do so for changes in these factors over time49. This means that the only confounding factors would be those that have the same state–month anomaly as temperature. We fitted the models using integrated nested Laplace approximation (INLA), and the R-INLA software, which is computationally more efficient than traditional Markov Chain Monte Carlo for Bayesian inference51. The uncertainty in our results was obtained from 5,000 draws from the posterior marginal of each month’s excess relative risk. The reported 95% credible intervals are the 2.5th to 97.5th percentiles of the sampled values.

computationally more efficient than traditional Markov Chain Monte Carlo for Bayesian inference51. The uncertainty in our results was obtained from 5,000 draws from the posterior marginal of each month’s excess relative risk. The reported 95% credible intervals are the 2.5th to 97.5th percentiles of the sampled values. Analyses were done separately by injury type, because different injuries can have differing associations with anomalously warm and cold temperatures. Analyses were also done separately by sex and age group (0–4 years, 10-year age groups from 5 years to 84 years, and 85+ years) because injury death rates vary by age group and sex (see Fig. 1 and Supplementary Table 2), as might their associations with temperature. We used the resultant risk estimates, and the age–sex-specific death rates from each injury in 2017, to calculate additional deaths if each month in each state was +1.5 °C above its long-term average, not only realistic in our lifetime under the current projections of global climate change, but as agreed under the Paris Climate Agreement2,52. This +1.5 °C rise is also within the range of the size of the anomaly experienced by some states (see Extended Data Fig. 2). For these calculations, we multiplied the actual death counts for each month, sex, state and age group in 2017 by the corresponding excess relative risk, which was calculated as the exponential of the coefficient of the temperature anomaly term from the above analysis. We did similar calculations for +2 °C, which is the upper boundary of the Paris Climate Agreement, and present these as Extended Data Figs. 3 and 4.

roup in 2017 by the corresponding excess relative risk, which was calculated as the exponential of the coefficient of the temperature anomaly term from the above analysis. We did similar calculations for +2 °C, which is the upper boundary of the Paris Climate Agreement, and present these as Extended Data Figs. 3 and 4. Sensitivity analyses We conducted sensitivity analyses to assess how much our results might depend on the temperature metric used to generate anomalous temperature. First, rather than building our monthly temperature anomalies based on daily mean temperatures, we used daily maxima and minima. These measures were strongly correlated to those generated from daily means (see Supplementary Table 3), and therefore we did not run models using these alternatives.

e anomalous temperature. First, rather than building our monthly temperature anomalies based on daily mean temperatures, we used daily maxima and minima. These measures were strongly correlated to those generated from daily means (see Supplementary Table 3), and therefore we did not run models using these alternatives. Second, together with temperature anomaly based on daily mean temperatures, we also included a second measure of anomaly in the model. We tested three different measures for this sensitivity analysis: (1) temperature anomaly calculated based on the 90th percentile (°C) of daily mean temperatures within a month, compared with the average of the 90th percentiles for each state and month; (2) number of days in a month above the long-term 90th percentile of average temperature for each state and month (adjusted for length of month); and (3) number of episodes of 3+ day episodes above the long-term 90th percentile of average temperature for each state and month (adjusted for length of month). These additional measures were related to more extreme anomalous situations, which may be relevant if the impacts on injuries are related to more extreme temperatures and their frequency in each month.

es above the long-term 90th percentile of average temperature for each state and month (adjusted for length of month). These additional measures were related to more extreme anomalous situations, which may be relevant if the impacts on injuries are related to more extreme temperatures and their frequency in each month. The correlations among these variables and anomaly based on the mean were between 0.60 and 0.89 (see Supplementary Table 4). The estimated rate ratios of the temperature anomaly based on daily means (that is, the anomaly measure used in the main analysis) were robust to the addition of alternative measures of anomaly, whereas the coefficients of the additional measures were generally not statistically significant and with large credible intervals. Therefore, we did not include the alternative additional measures of extreme anomalous temperature in the main analysis.

is) were robust to the addition of alternative measures of anomaly, whereas the coefficients of the additional measures were generally not statistically significant and with large credible intervals. Therefore, we did not include the alternative additional measures of extreme anomalous temperature in the main analysis. Comparison with previous studies Although there are no previous studies of how deviations of monthly temperature from the long-term average are associated with injury mortality, our results are broadly in agreement with both those that have analyzed associations with absolute temperature and those for specific injury types. A study of suicide in US counties over 37 years (1968–2004) estimated that a 1 °C higher monthly temperature would lead to a 0.7% rise in suicides9, compared with our findings of 0.7–1.5% in males and 0.5–2.9% in females at different ages for a +1.5 °C anomaly. A cross-sectional analysis in 100 US counties found that a 1 °C higher temperature would lead to a 1.3% increase in death rates from road traffic injuries24, compared with our finding of 0.6–3.1% in males and 0.5–2.0% in females for a +1.5 °C anomaly. In a study of six French heatwaves during 1971–2003, mortality from unintentional injuries rose by up to 4% during a heatwave period compared with a non-heatwave baseline10. A study of daily mortality from all injuries from Estonia found a 1.24% increase in mortality when the daily maximum temperature went from the 75th to the 99th percentile of long-term distribution11.

mortality from unintentional injuries rose by up to 4% during a heatwave period compared with a non-heatwave baseline10. A study of daily mortality from all injuries from Estonia found a 1.24% increase in mortality when the daily maximum temperature went from the 75th to the 99th percentile of long-term distribution11. Reporting Summary Further information on research design is available in the Nature Research Reporting Summary linked to this article. Online content Any methods, additional references, Nature Research reporting summaries, source data, extended data, supplementary information, acknowledgements, peer review information; details of author contributions and competing interests; and statements of data and code availability are available at 10.1038/s41591-019-0721-y. Supplementary information Supplementary Information Supplementary Tables 1–4. Reporting Summary Extended data Extended Data Fig. 1 Graphic representation of temperature anomaly measure used in the analysis. The graph shows how monthly temperatures in July two example states (Florida in red and Minnesota in blue) (left panel) for 1980–2017 are used to calculate temperature anomalies. As seen, a warmer state like Florida (top right) can have a smaller inter-annual variation in a particular month (here, July) compared with a cooler state like Minnesota (bottom right).

July two example states (Florida in red and Minnesota in blue) (left panel) for 1980–2017 are used to calculate temperature anomalies. As seen, a warmer state like Florida (top right) can have a smaller inter-annual variation in a particular month (here, July) compared with a cooler state like Minnesota (bottom right). Extended Data Fig. 2 Average size of temperature anomaly (°C) from 1980 to 2017, by state and month. The value for each state and month is the mean of the absolute size of anomaly, be it cold or warm, and hence gives an indication of the scale of anomalies around the local average temperatures. Extended Data Fig. 3 Additional annual injury deaths for the 2017 US population in year in which each month was +2 °C warmer compared with 1980–2017 average temperatures. The top row shows breakdown by type of injury, sex and age group. The bottom row shows the break down by type of injury, sex and month. Black dots represent net changes in deaths for each set of bars. Extended Data Fig. 4 Percent change in death rates in year in which each month was +2 °C compared with 1980–2017 average temperatures by type of injury, sex and (A) age group or (B) month. Colored dots show the posterior means and error bars represent 95% credible intervals, both obtained at the posterior draw level. Extended Data Fig. 5 Number of deaths by type of transport injury, month, sex and age group in the contiguous United States for 1980–2017.

Extended Data Fig. 4 Percent change in death rates in year in which each month was +2 °C compared with 1980–2017 average temperatures by type of injury, sex and (A) age group or (B) month. Colored dots show the posterior means and error bars represent 95% credible intervals, both obtained at the posterior draw level. Extended Data Fig. 5 Number of deaths by type of transport injury, month, sex and age group in the contiguous United States for 1980–2017. Peer review information Jennifer Sargent was the primary editor on this article, and managed its editorial process and peer review in collaboration with the rest of the editorial team. Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. Extended data is available for this paper at 10.1038/s41591-019-0721-y. Supplementary information is available for this paper at 10.1038/s41591-019-0721-y.

Peer review information Jennifer Sargent was the primary editor on this article, and managed its editorial process and peer review in collaboration with the rest of the editorial team. Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. Extended data is available for this paper at 10.1038/s41591-019-0721-y. Supplementary information is available for this paper at 10.1038/s41591-019-0721-y. Acknowledgements R.M.P. was supported by a Wellcome Trust ISSF Studentship. The development of statistical methods was supported by grants from the Wellcome Trust (grant no. 209376/Z/17/Z). Work on the US mortality data was supported by a grant from the US Environmental Protection Agency (EPA), as part of the Center for Clean Air Climate Solution (assistance agreement no. R835873). This article has not been formally reviewed by the EPA. The views expressed in this document are solely those of the authors and do not necessarily reflect those of the EPA. The EPA does not endorse any products or commercial services mentioned in this publication. We thank M. Blangiardo, S. Flaxman and C. Paciorek for discussions on the statistical model, and K. Bhalla, H. Frumkin, A. Haines and T. Kjellstrom for suggestions of relevant studies.

o not necessarily reflect those of the EPA. The EPA does not endorse any products or commercial services mentioned in this publication. We thank M. Blangiardo, S. Flaxman and C. Paciorek for discussions on the statistical model, and K. Bhalla, H. Frumkin, A. Haines and T. Kjellstrom for suggestions of relevant studies. Author contributions All authors contributed to the study concept and interpretation of results. R.M.P., G.D., R.T. and M.E. collated and organized temperature and mortality files. R.M.P., J.E.B., V.K., H.T.-W. and M.E. developed the statistical model. R.M.P., J.E.B. and V.K. implemented the statistical model. R.M.P. performed the analysis, with input from J.E.B., H.T.-W., V.K., R.T., G.D. and M.E. R.M.P. and M.E. wrote the first draft of the paper. J.E.B., H.T.-W., V.K., R.T. and G.D. contributed to revising and finalizing the paper. Data availability ERA5 temperature data are downloadable from https://www.ecmwf.int/en/forecasts/datasets/reanalysis-datasets/era5. Vital statistics files with geographical information can be requested through submission of a proposal to the NCHS (https://www.cdc.gov/nchs/nvss/nvss-restricted-data.htm). Code availability The computer code for the Bayesian model used in this work is available at http://globalenvhealth.org/code-data-download. Competing interests M.E. reports a charitable grant from AstraZeneca Young Health Programme, and personal fees from Prudential, Scor and Third Bridge, all outside the submitted work; all other authors declare no competing interests.

It is now well established that a delicate balance between inflammatory versus regulatory responses underlies disease progression in many autoimmune disorders. DCs have emerged as central players in initiating and regulating adaptive immune responses1–5. Emerging evidence suggests that DCs are also vital in suppressing immune responses through the generation of anergic or regulatory T cells6. However, the mechanisms by which DCs can be programmed to induce regulatory T cells are poorly understood. DCs express several Toll-like receptors (TLRs) and the C-type lectins, which are critical in sensing and initiating immune response against pathogens7–9. Engagement of such receptors programs gene expression that orchestrate DC maturation and activation2,7,8. The types of cytokines secreted by the DCs can regulate the differentiation of CD4+ T cells into TH-1, TH-2, TH-17 or T regulatory (Treg) responses. For example stimuli that induce IL-12(p70) promote IFN-γ producing TH-1 cells, stimuli that induce IL-10 favor TH-2 or Treg responses, whereas stimuli that induce TGF-β, IL-6 and IL-23 promote TH-17 differentiation.

an regulate the differentiation of CD4+ T cells into TH-1, TH-2, TH-17 or T regulatory (Treg) responses. For example stimuli that induce IL-12(p70) promote IFN-γ producing TH-1 cells, stimuli that induce IL-10 favor TH-2 or Treg responses, whereas stimuli that induce TGF-β, IL-6 and IL-23 promote TH-17 differentiation. Zymosan, a yeast cell wall derivative, is a complex microbial stimulus that is recognized by many innate immune receptors, including TLR2 and dectin-1, a C-type lectin receptor for β-gulcans10–15. How signaling via both TLR2 and dectin-1 is integrated and influences adaptive immunity is poorly understood. Collaboration between TLR2 and dectin-1 results in the induction of pro-inflammatory cytokines in macrophages and DCs14, as well as robust IL-10 production in DCs16–18. Consistent with this, our previous work demonstrates that zymosan conditions splenic DCs to secrete IL-10 and induce tolerogenic T cell responses18. Furthermore, zymosan is also known to induce splenic macrophages to secrete TGF-β18, a cytokine critical for the generation of regulatory T cells, as well as TH-17 cells11,19–21. In contrast, other studies have shown that dectin-1 mediated signaling in DCs induces TH-17 cells and IFN-γ producing TH-1 cells 22,23.

es18. Furthermore, zymosan is also known to induce splenic macrophages to secrete TGF-β18, a cytokine critical for the generation of regulatory T cells, as well as TH-17 cells11,19–21. In contrast, other studies have shown that dectin-1 mediated signaling in DCs induces TH-17 cells and IFN-γ producing TH-1 cells 22,23. Here we show that TLR2 and dectin-1 mediate divergent programs of DC activation, resulting in distinct adaptive responses to zymosan. Thus, zymosan induces Raldh2 expression in DCs via a mechanism dependent largely on TLR2-mediated activation of ERK MAPK. Raldh2 converts vitamin A derived retinal to RA, which acts in an autocrine manner to induce the expression of Socs3, and suppress activation of p38 MAPK and pro-inflammatory cytokines. Consistent with this, TLR2 signaling was critical for zymosan mediated induction of Treg cells, and suppression of TH-1 and TH-17 responses mediated autoimmunity in vivo; in the absence of TLR2 signaling, dectin-1 mediated signaling alone induced potent TH-1 and TH-17 responses and exacerbated autoimmunity.

cytokines. Consistent with this, TLR2 signaling was critical for zymosan mediated induction of Treg cells, and suppression of TH-1 and TH-17 responses mediated autoimmunity in vivo; in the absence of TLR2 signaling, dectin-1 mediated signaling alone induced potent TH-1 and TH-17 responses and exacerbated autoimmunity. Results Mechanism of induction of vitamin-A metabolizing enzymes Zymosan is known to induce both pro-inflammatory10–14, and anti-inflammatory cytokines in DCs and macrophages16–18. Our previous study demonstrated that zymosan stimulates regulatory DCs and macrophages, which produce IL-10 and TGF-β, respectively18. One important variable in the different studies is the source and type of DCs. Therefore, it was important to determine whether zymosan stimulated similar cytokine profiles in both in vitro generated bone marrow derived DCs (BM-DCs), as well as from splenic DCs. In agreement with published reports23,24, zymosan induced the pro-inflammatory cytokines IL-23, IL-6, IL-12 and TNF-α , as well as IL-10 from BM-DCs (Supplementary Fig. 1). In contrast, zymosan induced robust IL-10, and little IL-23, IL-6, IL-12 and TNF-α in splenic DCs in vitro (Supplementary Fig. 2) and in vivo (Supplementary Fig. 3a, b). Thus, zymosan induces distinct cytokine profiles depending on the type DCs.

-6, IL-12 and TNF-α , as well as IL-10 from BM-DCs (Supplementary Fig. 1). In contrast, zymosan induced robust IL-10, and little IL-23, IL-6, IL-12 and TNF-α in splenic DCs in vitro (Supplementary Fig. 2) and in vivo (Supplementary Fig. 3a, b). Thus, zymosan induces distinct cytokine profiles depending on the type DCs. To obtain insights into how zymosan activates splenic DCs, we performed a microarray analysis. Zymosan induced robust expression of the anti-inflammatory gene Il-10 and much lower expression of pro-inflammatory genes such Il-6, Tnf-α and Il-12, relative to LPS (data not shown). Surprisingly, zymosan also induced expression of genes involved in the biosynthesis of retinoic acid (RA) such as Adh1 and Aldh1a2 (Raldh2) (data not shown). This was unexpected as splenic DCs, unlike intestinal DCs25–30, are not thought to express these enzymes. Consistent with the microarray data, RT-PCR analysis demonstrated that splenic DCs constitutively express the RA metabolizing enzymes Adh class I (Adh1), Adh class III (Adh5), and low levels of Raldh1 (data not shown); upon zymosan treatment their expressions levels were further increased (data not shown). Importantly, zymosan stimulated an 80-fold increase in the expression of Raldh2 mRNA (Fig. 1a). In contrast, BM-DCs stimulated with zymosan showed a modest increase in Raldh2 mRNA expression compared to the splenic DCs stimulated with zymosan (Supplementary Fig. 4).

sions levels were further increased (data not shown). Importantly, zymosan stimulated an 80-fold increase in the expression of Raldh2 mRNA (Fig. 1a). In contrast, BM-DCs stimulated with zymosan showed a modest increase in Raldh2 mRNA expression compared to the splenic DCs stimulated with zymosan (Supplementary Fig. 4). We then examined whether other TLR ligands could induce Raldh2 in DCs. Neither LPS nor CpG induced Raldh2 (Fig. 1b). However, other ligands specific for TLR2/TLR6, such as Pam-2-cys and FSL induced substantial expression of Raldh2 (Fig. 1b). Consistent with this, injection of zymosan into mice induced robust expression of Raldh2 mRNA in DCs (Fig. 1c). The expression of Raldh2 protein in splenic DCs in vivo was confirmed by western blot and immunohistochemical analysis (Fig. 1d, e). These data demonstrate that Raldh2 is robustly induced in splenic DCs upon stimulation with zymosan, or other TLR2/6 ligands.

ce induced robust expression of Raldh2 mRNA in DCs (Fig. 1c). The expression of Raldh2 protein in splenic DCs in vivo was confirmed by western blot and immunohistochemical analysis (Fig. 1d, e). These data demonstrate that Raldh2 is robustly induced in splenic DCs upon stimulation with zymosan, or other TLR2/6 ligands. Since zymosan signals via both TLR2 and dectin-110–12,14,16,18, we determined whether induction of Raldh2 is dependent on both TLR2 and dectin-1. DCs from Tlr2−/− mice expressed lower levels of Raldh2 mRNA compared to wild type DCs, upon zymosan stimulation (Fig. 1f). Consistent with this, DCs from Tlr2−/− mice showed a significant reduction in Raldh2 expression upon zymosan injection in vivo (Fig. 1f). Curdlan, a β-glucan ligand, that is thought to be a specific ligand for dectin-116, induced Raldh2, albeit at a much lower level (12-fold, Fig. 1b) than zymosan. However, DCs from dectin-1-deficient mice displayed only a very modest reduction in Raldh2 mRNA level compared to the wild type DCs upon zymosan stimulation (Fig. 1g), suggesting that the effects of curdlan may have been mediated in part by contaminants that triggered TLR2; thus dectin-1 likely does not play a major role in inducing Raldh2.

in-1-deficient mice displayed only a very modest reduction in Raldh2 mRNA level compared to the wild type DCs upon zymosan stimulation (Fig. 1g), suggesting that the effects of curdlan may have been mediated in part by contaminants that triggered TLR2; thus dectin-1 likely does not play a major role in inducing Raldh2. We next investigated the downstream signaling pathways through which TLR2 induced Raldh2 expression. TLR2 stimulates rapid induction of ERK, which mediates IL-10 production by DCs16,18,23. Induction of Raldh2 mRNA expression was largely abrogated by inhibitors against ERK (Fig. 1h). Additionally DCs from Erk1−/− mice had substantially reduced Raldh2 expression upon zymosan stimulation (data not shown). Furthermore, induction of Raldh2 was also inhibited substantially by a syk inhibitor (data not shown). Thus, Raldh2 induction by zymosan is syk-dependent, but largely dectin-1 independent, suggesting that an alternative syk-dependent receptor is likely involved. In summary therefore, zymosan induces Raldh2 expression in DCs via TLR2-mediated activation of ERK, likely acting in concert with syk-dependent signaling via another receptor.

duction by zymosan is syk-dependent, but largely dectin-1 independent, suggesting that an alternative syk-dependent receptor is likely involved. In summary therefore, zymosan induces Raldh2 expression in DCs via TLR2-mediated activation of ERK, likely acting in concert with syk-dependent signaling via another receptor. Next, we determined whether production of Raldh2 and IL-10 in DCs were interdependent. Notably, induction of Il-10 mRNA preceded that of Raldh2, raising the possibility that IL-10 may promote Raldh2 induction (Supplementary Fig. 5a). However, induction of Raldh2 was unaffected in DCs from IL-10 deficient mice (Supplementary Fig. 5b). Conversely, blocking RA synthesis or RA signaling using RAR antagonists had no effect on IL-10 production upon zymosan treatment (Supplementary Fig. 5c). Therefore, induction of IL-10 and Raldh2 do not seem to be interdependent. Finally, we explored whether the mechanisms of Raldh2 induction observed with zymosan would also operate in response to a live yeast infection. Thus, we investigated whether the live pathogenic fungus Candida albicans triggered Raldh2 in splenic DCs. Stimulation of splenic DCs with live C. albicans induced Raldh2 mRNA and protein in WT DCs (Fig. 1i), but this induction was significantly reduced in Tlr2−/− DCs. Collectively, these results suggest an important role for TLR2 mediated ERK activation, in the induction of Raldh2 in splenic DCs upon stimulation with zymosan, or live yeast.

enic DCs with live C. albicans induced Raldh2 mRNA and protein in WT DCs (Fig. 1i), but this induction was significantly reduced in Tlr2−/− DCs. Collectively, these results suggest an important role for TLR2 mediated ERK activation, in the induction of Raldh2 in splenic DCs upon stimulation with zymosan, or live yeast. RA and IL-10 synergize to induce Treg cells Next, we determined the role of RA and IL-10 in the induction of Treg cells. In the presence of TGF-β, RA can promote conversion of naïve T cells to Foxp3 expressing Treg cells25–27,29,30. Since zymosan induces enzymes involved in RA synthesis, we determined whether these DCs can metabolize vitamin A (retinol) to RA, and induce Treg cells. Splenic DCs were stimulated with zymosan, or LPS or curdlan for 10h, and then washed and pulsed with an I-Ab-restricted ovalbumin peptide OVA323–339 (OVA), and cultured with or without TGF-β, with naive OT-II CD4+ T cells, which express a transgenic T cell receptor specific for the OVA18. After 4 days, OT-II T cells were restimulated with antibodies against CD3 and CD28. Zymosan-treated DCs in the absence of TGF-β induced mostly IL-10-producing Tr1 cells, whereas in the presence of TGF-β, they induced both Treg cells, and Tr1 cells (Fig. 2a). Zymosan stimulated DCs did not induce robust TH-1 or TH-17 cells (Fig. 2a). In contrast, LPS stimulated DCs induced robust TH1 responses, and curdlan stimulated DCs induced both TH1 cells and TH-17 cells (Fig. 2a). Interestingly, in contrast to splenic DCs, zymsoan treated BM-DCs induced TH-17 cells (Supplementary Fig. 6).

san stimulated DCs did not induce robust TH-1 or TH-17 cells (Fig. 2a). In contrast, LPS stimulated DCs induced robust TH1 responses, and curdlan stimulated DCs induced both TH1 cells and TH-17 cells (Fig. 2a). Interestingly, in contrast to splenic DCs, zymsoan treated BM-DCs induced TH-17 cells (Supplementary Fig. 6). The proportion of Treg cells stimulated by zymosan treated DCs was increased in the presence of retinol (Fig. 2b). To determine, whether this effect was mediated through RA synthesis, we stimulated DCs with zymosan, in the presence or absence of the Raldh inhibitor disulphiram, which inhibits RA synthesis by blocking its conversion from retinal25–30. After 10hrs, DCs were washed and cultured with naïve T cells in the presence of retinol and OVA. Inhibition of de novo RA synthesis in DCs suppressed zymosan-induced Treg cells (Fig. 2b). Interestingly, disulphiram also inhibited the frequency of Treg cells which were induced in the absence of zymosan (data not shown). This may reflect expression of low levels of Raldh1 in splenic DCs, in response to signals from activated CD4+ T cells. This basal level of Raldh1 may play a role in converting pre-stored retinol to retinal in splenic DCs. Nevertheless, taken together, these results suggest zymosan induces splenic DCs to express RA synthesizing genes and stimulate Treg cells.

of Raldh1 in splenic DCs, in response to signals from activated CD4+ T cells. This basal level of Raldh1 may play a role in converting pre-stored retinol to retinal in splenic DCs. Nevertheless, taken together, these results suggest zymosan induces splenic DCs to express RA synthesizing genes and stimulate Treg cells. Interestingly, zymosan-treated DCs from Tlr2−/− mice induced a significantly lower frequency of Treg cells compared to wild type DCs (Fig. 2c). Furthermore, zymosan stimulated DCs treated with an ERK inhibitor were compromised in their ability to induce Treg cells (Fig. 2d). However, addition of exogenous RA to the culture restored their ability to induce Treg cells (Fig. 2c). These results are consistent with the effects of TLR2-mediated ERK signaling in inducing Raldh2 in DCs (Fig. 1f – h). Thus, zymosan activates DCs via TLR2 to stimulate retinol metabolizing enzymes which induce Treg cells.

er, addition of exogenous RA to the culture restored their ability to induce Treg cells (Fig. 2c). These results are consistent with the effects of TLR2-mediated ERK signaling in inducing Raldh2 in DCs (Fig. 1f – h). Thus, zymosan activates DCs via TLR2 to stimulate retinol metabolizing enzymes which induce Treg cells. In addition to the effects of RA on the induction of Treg cells, IL-10 was also observed to play a role, in experiments using a neutralizing antibody against IL-10 and its receptor (Supplementary Fig. 7a), and DCs from Il-10−/− mice (Supplementary Fig. 7b). However, addition of retinol to these cultures significantly increased the proportion of Treg cells (Supplementary Fig. 7b). Consistent with these results, blocking IL-10, or RA mediated signaling in DCs, converts them from regulatory DCs to stimulatory DCs that induce enhanced TH1 and TH17 responses (Supplementary Fig. 8a, b). These results demonstrate that IL-10 deficient DCs can metabolize vitamin A as efficiently as wild type DCs, and that synthesis of IL-10 and RA are interdependent. Importantly, inhibition of RA or IL-10 alone results in a pronounced diminution of Treg cells, suggesting that IL-10 and RA act synergistically to induce Treg cells.

ults demonstrate that IL-10 deficient DCs can metabolize vitamin A as efficiently as wild type DCs, and that synthesis of IL-10 and RA are interdependent. Importantly, inhibition of RA or IL-10 alone results in a pronounced diminution of Treg cells, suggesting that IL-10 and RA act synergistically to induce Treg cells. Autocrine effects of RA and IL-10 on DCs induces Socs3 IL-10 exerts autocrine effects on DCs to suppress zymosan induced pro-inflammatory cytokines18. Given the synergistic effects of RA and IL-10 in stimulating Treg cells, we hypothesized that RA might exert a similar effect on DCs. Thus, we determined the expression of the RA nuclear receptors RARαβ and γ. We observed that all three receptors were expressed on DCs (Fig. 3a). We stimulated DCs with zymosan in the presence or absence of retinol, and determined the induction pro-inflammatory cytokines. While zymosan-treated DCs produced little IL-6, IL-12, TNF-α, addition of retinol further reduced the levels of these cytokines (Fig. 3b). To determine whether this effect was dependent on RARs, we stimulated DCs with zymosan in the presence of retinol plus the RA receptor antagonist (LE135/540). Addition of LE135/540 significantly increased the production of pro-inflammatory cytokines (Fig. 3b), compared to the retinol treated DCs. Furthermore, neutralization of IL-10 with an antibody (Fig. 3b), or DCs from Il-10−/− mice (Fig. 3c), produced enhanced levels of pro-inflammatory cytokines in response to zymosan; however the addition of retinol or RA receptor antagonists produced the same effects observed with wild type DCs (Fig. 3b & c). These results demonstrate that RA produced by zymosan-treated DCs can act in an autocrine manner to suppress the production of pro-inflammatory cytokines.

ory cytokines in response to zymosan; however the addition of retinol or RA receptor antagonists produced the same effects observed with wild type DCs (Fig. 3b & c). These results demonstrate that RA produced by zymosan-treated DCs can act in an autocrine manner to suppress the production of pro-inflammatory cytokines. In exploring the mechanisms by which RA suppressed the production of pro-inflammatory cytokines in zymosan-treated DCs, we observed an increase in the expression of Socs3 in zymosan-treated DCs, relative to the untreated DCs in our microarray analysis (data not shown). This was confirmed by RT-PCR (Fig. 3d). Addition of retinol to the culture further increased Socs3 expression to approximately 20 fold; in contrast, addition of the RAR antagonist significantly reduced the Socs3 mRNA expression (Fig. 3d). Thus Socs3 is inducible upon zymosan stimulation and is partly dependent on RAR-mediated signaling. Consistent with this, knock-down of Raldh2 in DCs using siRNA significantly reduced the induction of Socs3 in response to zymosan (Supplementary Fig. 9). Furthermore, IL-10 also enhanced the induction of Socs3 expression by zymosan (Fig. 3e). DCs from Il-10−/− mice stimulated with zymosan showed a significant reduction in Socs3 expression, relative to wild type DCs (Fig. 3d, e), but this defect could be corrected by the addition of exogenous IL-10 or retinol to the culture (Fig. 3e). Consistent with these in vitro observations, zymosan induced a significant increase in Socs3 expression in DCs within 3 hours in vivo (Fig. 3f). Furthermore, treatment of mice with disulphiram or LE135/540 reduced the level of Socs3 expression in vivo upon zymosan injection (Fig. 3g). Induction of Socs3 was dependent on TLR2, as zymosan induced much lower levels of Socs3 in Tlr2−/− mice, relative to wild type mice (Fig. 3h).

Cs within 3 hours in vivo (Fig. 3f). Furthermore, treatment of mice with disulphiram or LE135/540 reduced the level of Socs3 expression in vivo upon zymosan injection (Fig. 3g). Induction of Socs3 was dependent on TLR2, as zymosan induced much lower levels of Socs3 in Tlr2−/− mice, relative to wild type mice (Fig. 3h). Next, we determined the effect of RA on the activation of ERK and p38 MAPK. Blocking RAR-mediated signaling with LE135/540 led to a sustained activation of p38, but had no effect on ERK activation (Supplementary Fig. 10a, c). Furthermore, IL-10 deficiency also resulted in a marked effect on the sustained expression of p38 MAPK in DCs (Supplementary Fig. 10d). This suggests that similar to IL-10, RA signaling attenuates activation of p38 MAPK. To further address the role of Socs3 in zymosan mediated suppression of pro-inflammatory cytokines we knocked-down Socs3 in DCs using siRNA (Supplementary Fig. 11). DCs transfected with siRNA against Socs3 produced significantly higher levels of pro-inflammatory cytokines upon zymosan treatment compared to DCs transfected with control siRNA (Supplementary Fig. 11). Collectively, these results show that RA meditated autocrine signaling is critical for the induction of Socs3, and for regulating the activity of p38 MAPK and pro-inflammatory cytokines, in zymosan-stimulated DCs.

y cytokines upon zymosan treatment compared to DCs transfected with control siRNA (Supplementary Fig. 11). Collectively, these results show that RA meditated autocrine signaling is critical for the induction of Socs3, and for regulating the activity of p38 MAPK and pro-inflammatory cytokines, in zymosan-stimulated DCs. TLR2 suppression of IL-23 and TH-17 mediated autoimmunity To determine whether zymosan induces Treg cells in vivo, we adoptively transferred naïve OT-II cells into wild-type, Tlr2−/− or Il-10−/− mice, and then immunized the mice with OVA alone, or OVA mixed with either zymosan, or curdlan or LPS. After 4 days, we analyzed the proliferation, and cytokine production of splenic OT-II T cells restimulated in vitro. Injection of zymosan + OVA resulted in a weak clonal expansion of antigen-specific T cells, relative to injection of OVA + LPS, or OVA + curdlan (data not shown). Consistent with this, stimulation with OVA+ zymosan resulted in a robust induction of antigen specific Treg cells and IL-10-producing Tr1 cells, compared to stimulation with OVA+PBS or LPS or curdlan (Fig. 4a). Intracellular cytokine staining revealed robust induction of TH1 responses in mice injected with LPS, and induction of both TH1 and TH-17 responses in mice injected with curdlan. In contrast, mice injected with zymosan showed relatively weak TH-1 and TH-17 responses (Fig. 4a). In Tlr2−/− and Il-10−/− mice, induction of Treg cells was markedly reduced, relative to wild type mice (Fig. 4b). Furthermore, in the knockout mice, zymosan induced enhanced TH1 and TH17 responses, and reduced IL-10 producing cells (Fig. 4c). These data suggest that TLR2 mediated signaling is critical for the induction of both Treg cells and IL-10+ Tr1 cells.

f Treg cells was markedly reduced, relative to wild type mice (Fig. 4b). Furthermore, in the knockout mice, zymosan induced enhanced TH1 and TH17 responses, and reduced IL-10 producing cells (Fig. 4c). These data suggest that TLR2 mediated signaling is critical for the induction of both Treg cells and IL-10+ Tr1 cells. We next determined the therapeutic potential of zymosan stimulated Treg cells on autoimmune disease, using a mouse model of EAE. First, we immunized mice with myelin oligodendrocytes glycoprotein peptide, MOG35–55 (MOG) emulsified in CFA, or MOG + zymosan, subcutaneously, on days 0 and 7. As previously described24, MOG + CFA showed onset of neurological impairment starting at around day 14 (Fig. 5a, left panel). In contrast MOG + zymosan induced a relatively attenuated and transient disease course (Fig. 5a, left panel). In Tlr2−/− mice, injection of MOG + zymosan resulted in as severe a disease as observed in the positive control, demonstrating a regulatory role for TLR2 (Fig. 5a, right panel). We next determined whether zymosan was capable of actively suppressing disease. Thus in a separate experiment, we immunized mice with MOG + CFA, and simultaneously injected zymosan, or curdlan. Control mice received PBS or curdlan, and displayed disease symptoms, starting around day 14 (Fig. 5b, left panel). In contrast, mice treated with zymosan developed significantly lower clinical scores compared to the PBS or curdlan treated mice (Fig. 5b, left panel). Consistent with a regulatory role for TLR2, zymosan induced enhanced disease progression in Tlr2−/− mice relative to wild type mice (Fig. 5b, right panel).

left panel). In contrast, mice treated with zymosan developed significantly lower clinical scores compared to the PBS or curdlan treated mice (Fig. 5b, left panel). Consistent with a regulatory role for TLR2, zymosan induced enhanced disease progression in Tlr2−/− mice relative to wild type mice (Fig. 5b, right panel). We then determined the phenotype of CNS infiltrated CD4+ T cells at day 18. We observed that zymosan treatment resulted in enhanced induction of Treg cells and Tr1 cells, relative to mice treated with PBS or curdlan (Fig. 5c). In contrast, control or curdlan treated mice showed a significant increase in the number of TH1 and TH17 cells compared to the zymosan treated mice (Fig. 5c). In Tlr2−/− mice, there was a diminished frequency of Treg cells and Tr1 cells, and greatly enhanced TH1 and TH-17 responses (Fig. 5c). Thus, TLR2 mediated signaling is essential for the induction of Treg cells and Tr1 cells, and suppresses TH-1 and TH-17 responses.

cells compared to the zymosan treated mice (Fig. 5c). In Tlr2−/− mice, there was a diminished frequency of Treg cells and Tr1 cells, and greatly enhanced TH1 and TH-17 responses (Fig. 5c). Thus, TLR2 mediated signaling is essential for the induction of Treg cells and Tr1 cells, and suppresses TH-1 and TH-17 responses. IL-23 plays a pivotal role in the expansion of TH-17 cells, and the pathogenesis of EAE. We thus evaluated the IL-23p19 mRNA expression in DCs isolated ex vivo from wild type versus Tlr2−/− mice (Fig. 5d). We also measured the serum cytokine levels of IL-23 in wild type and Tlr2−/− mice (Fig. 5e) upon zymosan injection. Zymosan induced much higher expression of IL-23 in Tlr2−/− mice, relative to wild type mice (Fig. 5d & e). Similarly, DCs isolated from Tlr2−/− mice, produced significantly higher levels of IL-23 compared to wild type DCs (Fig. 5f). Collectively, these results suggest that TLR2 mediated signaling suppresses IL-23 production in DCs (Fig. 6).

igher expression of IL-23 in Tlr2−/− mice, relative to wild type mice (Fig. 5d & e). Similarly, DCs isolated from Tlr2−/− mice, produced significantly higher levels of IL-23 compared to wild type DCs (Fig. 5f). Collectively, these results suggest that TLR2 mediated signaling suppresses IL-23 production in DCs (Fig. 6). Finally, we investigated the relative roles of IL-10 and RA in zymsoan mediated suppression of EAE using IL-10−/− mice and Raldh inhibitor. Injection of disulphiram resulted in a significant (p<0.01), but transient, enhancement in the acceleration of the disease (Supplementary Fig. 12a). Furthermore, IL-10−/− mice were also more susceptible to EAE compared to the wild type mice (Supplementary Fig. 12b). Even in IL-10−/− mice, injection of disulphiram resulted in a marked enhancement in the disease severity (Supplementary Fig. 12b). Similarly, wild type mice injected with RALDH inhibitor-encapsulated in microparticles, that specifically targets to antigen-presentig cells in vivo 31,32, displayed enhanced EAE upon zymosan treatment compared to the mice injected with empty particles (Supplementary Fig. 12 c). These results collectively suggest both RA and IL-10 contribute to the suppressive effect of zymosan (Fig. 6).

n microparticles, that specifically targets to antigen-presentig cells in vivo 31,32, displayed enhanced EAE upon zymosan treatment compared to the mice injected with empty particles (Supplementary Fig. 12 c). These results collectively suggest both RA and IL-10 contribute to the suppressive effect of zymosan (Fig. 6). Discussion Several aspects of these findings deserve comment. The first concerns the mechanisms of generation and action of RA. Although recent studies have highlighted an important role for RA in the induction of Treg cells in the intestine26,27,33,34, its role in systemic immune responses is poorly understood. The present data demonstrate that Raldh2 can be induced in splenic DCs, via TLR2-dependent ERK signaling, which programs DCs to induce Treg cells. Dectin-1 signaling seems largely irrelevant, although the dependency on syk suggests that other syk-dependent receptors14 may act in concert with TLR2 to induce Raldh2. With respect to the target of RA, it acts directly on T cells25–30, but there is emerging evidence that RA can act also on DCs35–38. Retinoids signal via two groups of nuclear receptors, the RAR receptors (RARα,β. and γ) and the retinoid X receptors (RXRα, βand γ )39. Our data demonstrate that DCs express RARα, RARβ, RARγ and RXRα, and that signaling via RARs, is critical for enhanced Raldh2 expression. This is consistent with reports on other cell types that RA regulates its own synthesis through RAR receptor mediated signaling and RALDH expression40–44. Furthermore, Raldh2 suppresses pro-inflammatory cytokines in DCs via induction of Socs3, a well known regulatory of pro-inflammatory responses45–47.

ldh2 expression. This is consistent with reports on other cell types that RA regulates its own synthesis through RAR receptor mediated signaling and RALDH expression40–44. Furthermore, Raldh2 suppresses pro-inflammatory cytokines in DCs via induction of Socs3, a well known regulatory of pro-inflammatory responses45–47. The second point concerns the conflicting reports on the effects of zymosan on innate and adaptive responses. Zymosan induces pro-inflammatory responses from macrophages and DCs10–14,16,17, but also induces robust IL-10 from DCs16–18. Our work demonstrates that specific TLR2 ligands bias towards the Th2 pathway48,49, and that zymosan in particular induces tolerogenic responses18, which contrast with recent findings that zymosan induces TH-17 responses23,24,50. One key variable in these studies is the type of DCs used. Our data suggest that splenic DCs produce robust IL-10, but little or no pro-inflammatory cytokines, while BM-DCs produce robust pro-inflammatory cytokines. Furthermore, zymosan induces splenic DCs to stimulate a predominantly Treg response, while it induces BM-DCs to stimulate a TH-17 response. Consistent with this, BM-DCs expressed high levels of dectin-1 compared to splenic DCs, whereas TLR2 expression levels were comparable between BM-DCs and splenic-DCs (data not shown). Another variable is dose of zymosan. Stockinger et al 24 showed that injection of 500µg of zymosan induced TH17 responses and EAE, but only transiently. Here, we observed that 100µg of zymosan, resulted in an active suppression of disease progression, and a striking reduction in the frequency of TH-1/TH-17 cells, and an enhancement in the frequency of Treg cells and Tr1 cells (Fig. 5). When we injected 500µg of zymosan, we could only see a mild and transient disease, of similar severity and kinetics observed with the lower dose (data not shown). The reasons for the difference between our study and those of Stockinger et al24 are not clear, but may lie with differences in the mouse colonies.

(Fig. 5). When we injected 500µg of zymosan, we could only see a mild and transient disease, of similar severity and kinetics observed with the lower dose (data not shown). The reasons for the difference between our study and those of Stockinger et al24 are not clear, but may lie with differences in the mouse colonies. Finally, it is now clear that the immune system senses microbes not with a single innate immune receptor, but with a combination of several receptors51 . Combinatorial triggering of specific TLRs on DCs induces synergistic responses51, and cooperation between TLRs and non TLRs is known9, but its impact on the adaptive immunity is poorly understood. The present data demonstrate that two receptors that sense the same microbe can mediate divergent innate and adaptive immune responses, with distinct effects on disease progression. What evolutionary benefit might accrue to the microbe or to the host, from these mixed signals? From the microbe’s perspective, the induction of Treg cells could represent an immune evasion strategy; from the host’s perspective a “balanced” response could ensure immune defense, without collateral damage caused by excessive inflammation. It is tempting to speculate that pathogens that cause chronic infections like tuberculosis, HIV, and HCV might exploit this balance. Therefore immune interventions against chronic infections should focus on strategies that not only enhance TH-1/TH-17 responses, but which simultaneously inhibit Treg cells. Furthermore, vaccine adjuvants that engage multiple innate receptors to simultaneously promote TH-1/TH-17 and Treg responses might induce effective, but safe immunity in humans.

against chronic infections should focus on strategies that not only enhance TH-1/TH-17 responses, but which simultaneously inhibit Treg cells. Furthermore, vaccine adjuvants that engage multiple innate receptors to simultaneously promote TH-1/TH-17 and Treg responses might induce effective, but safe immunity in humans. Methods Mice C57BL/6, OT-II and B6.PL mice were from Jackson Laboratories. Il-10−/− mice (B6.129P2-Il10tm1Cgn/J) were from Jackson Laboratories. Mice were maintained in specific pathogen–free conditions in the Emory Vaccine Center vivarium. All animal protocols were reviewed and approved by the Institute Animal Care and Use Committee of Emory University. Purification of splenic DCs In brief, spleens from mice cut into small fragments, and then digested with collagenase type 4 (1 mg ml−1) in complete DMEM plus 2% FBS for 30 minutes at 37°C. Cells were washed twice and the CD11c+ DCs were enriched using the CD11c+ microbeads from Miltenyi Biotech. The resulting purity of CD11c+ DCs was approximately 95%. TLR stimulation of DCs CD11c+ splenic DCs (106 cells ml−1) were cultured for 24 h with Escherichia coli LPS (5 µg/ml), Pam-3-cys (100 ng ml−1), Pam-2-cys (100 ng ml−1), CpG dinucleotides (1 µg ml−1), zymosan (25 µg ml−1,) or curdlan (25 µg ml−1). Supernatants were collected and ELISA measured cytokines. In some experiments, anti-IL-10 and anti-IL-10R (1 µg ml−1) or disulphiram (100 nM) or LE 135/LE 540 (1 µM) were added for the duration of the stimulation.

1), Pam-2-cys (100 ng ml−1), CpG dinucleotides (1 µg ml−1), zymosan (25 µg ml−1,) or curdlan (25 µg ml−1). Supernatants were collected and ELISA measured cytokines. In some experiments, anti-IL-10 and anti-IL-10R (1 µg ml−1) or disulphiram (100 nM) or LE 135/LE 540 (1 µM) were added for the duration of the stimulation. In vitro cultures of murine DCs and T cells For in vitro stimulation, purified splenic CD11c+ DCs (106 cells ml−1) were stimulated with LPS (5 µg ml−1) or zymsoan (25 µg ml−1) or curdlan (25 µg ml−1) for 10 hrs and washed with media three times. In some experiments DCs were cultured with disulphiram (100 nM) or Erk inhibitor (100 nM) or LE135/LE540 (1 µM) for the duration of stimulation. Activated DCs (2 × 104) were washed, and then cultured together with naive CD4+CD62L+ OT-II CD4+ T cells (105) and OVA (2 µg ml−1) in 200 ml RPMI complete medium in 96-well round-bottomed plates. Supernatants were analyzed after 90 h and cells were collected and analyzed directly or were restimulated after 90 h. In some experiments, 500 nM retinol (Sigma) and/or 1ng ml−1 TGF-β (R&D Systems) were added to cultures. Antibody against mouse IL-10 (JES5-16E3; Becton Dickinson), antibody against IL-10R (1B1.3a; Becton Dickinson), antibody against human TGF-β (MAB1835; R&D Systems) or rat IgG isotype control antibody (A95-1; Becton Dickinson) was added to cultures at a final concentration of 10 µg ml−1. LE135/LE540 was added to some cultures at a concentration of 1 µM. For secondary restimulation, cells were collected after 90 h of primary culture, then were restimulated for 6 h with plate-bound antibody against CD3 (10 µg ml−1; 145.2C11 from Becton Dickinson) and antibody against CD28 (2 µg ml−1; 37.51 from Becton Dickinson) in the presence of brefeldin A (Becton Dickinson) for intracellular cytokine detection or were restimulated with OVA (2 µg ml−1) for 48 h for analysis of proliferation and cytokine production in cell supernatants.

g ml−1; 145.2C11 from Becton Dickinson) and antibody against CD28 (2 µg ml−1; 37.51 from Becton Dickinson) in the presence of brefeldin A (Becton Dickinson) for intracellular cytokine detection or were restimulated with OVA (2 µg ml−1) for 48 h for analysis of proliferation and cytokine production in cell supernatants. Statistical analysis Statistical analyses were conducted using GraphPad Prism. Mean clinical scores were analyzed using the Mann-Whitney non-parametric t test. The statistical significance of differences in the means + s. d. of cytokines released by cells of various groups were calculated with the Student's t-test (one-tailed). Supplementary Material 1 Note: Supplementary information is available on the Nature Medicine website. ACKNOWLEDGEMENTS We thank S. Aguilar Mertens and L. Bonner for assistance with cell sorting; D. Levesque (Emory Vaccine Center, Atlanta, U.S.A) for assistance with animal husbandry; Dr. S. Akira (Research Institute for Microbial Diseases, Osaka University, Osaka, Japan), for supply of various knockout mice; and Dr. G. Brown (University of Cape Town, South Africa) for Dectin-1−/− mice; and Dr. G. Landreth (Case Western Reserve University, Cleveland, OH, USA) for Erk1−/− mice. Supported by funding from the U.S. National Institutes of Health (R01 DK057665, R01 AI048638, U19 AI057266, U54 AI057157, N01 AI50019, N01 AI50025), and from the Bill & Melinda Gates Foundation to B.P.

ica) for Dectin-1−/− mice; and Dr. G. Landreth (Case Western Reserve University, Cleveland, OH, USA) for Erk1−/− mice. Supported by funding from the U.S. National Institutes of Health (R01 DK057665, R01 AI048638, U19 AI057266, U54 AI057157, N01 AI50019, N01 AI50025), and from the Bill & Melinda Gates Foundation to B.P. Figure 1 Mechanism of induction of vitamin-A metabolizing enzymes in splenic DCs. (a) Expression of Raldh (Aldh1a) mRNA in splenic DCs cultured in vitro for 24 h with (black bars), or without (grey bars) zymosan. Expression of Raldh mRNA relative to expression of mRNA encoding glyceraldehyde phosphate dehydrogenase (GAPDH) was analyzed by quantitative RT-PCR in this and all figures below. (b) Induction of Raldh2 (Aldh1a2) in splenic DCs cultured in vitro with TLR ligands or curdlan for 24 h. (c) Induction of Raldh2 in splenic DCs in vivo, in C57BL/6 mice injected with zymosan. (d) Western blot analysis of expression of Raldh2 protein in splenic DCs in vivo, 5 h after C57BL/6 mice were injected with zymosan. (e) Immunofluorescence microscopy of frozen tissue section of spleens of C57BL/6 mice injected with PBS or Zymosan, fixed and stained with antibodies specific for mouse CD11c (red), Raldh (green) and B220 (blue). (f) Induction of Raldh2 is dependent on TLR2. Left panel: Splenic DCs from WT or Tlr2−/− mice were cultured in media alone or with zymosan and, 24 h later expression of Raldh2 mRNA was analyzed. Right panel: C57BL/6 mice or Tlr2−/− were injected with zymosan and at various time points, splenic DCs were isolated, and expression of Raldh2 mRNA was analyzed. (g) Raldh2 induction is not dependent on dectin-1. Splenic DCs from wild type or dectin-1−/− mice were cultured in media alone or with zymosan, and 24 h later expression of Raldh2 mRNA was analzed. Difference between wild type and dectin-1−/− mice is not significant. (h) Induction of Raldh2 is dependent on ERK MAPK signaling. Splenic DCs were cultured as described above, with or without U0126. (i) Induction of Raldh2 in splenic DCs by Candida albicans is dependent on TLR2. Splenic-DCs from WT or Tlr2−/− mice were cultured in media alone or with Candida albicans. In all figures results are means + s. d. of 2 – 3 mice per group in one representative experiment out of two or three. *,P < 0.05; **,P < 0.005; ***P < 0.0001 in all figures.

splenic DCs by Candida albicans is dependent on TLR2. Splenic-DCs from WT or Tlr2−/− mice were cultured in media alone or with Candida albicans. In all figures results are means + s. d. of 2 – 3 mice per group in one representative experiment out of two or three. *,P < 0.05; **,P < 0.005; ***P < 0.0001 in all figures. Figure 2 RA and IL-10 synergize to induce Foxp3+ Treg cells. (a) Zymosan stimulates splenic DCs to induce Treg cells. Splenic DCs were stimulated with zymosan or curdlan or LPS for 12 h, and washed and cultured with naïve (CD4+CD62L+) OT-II T cells with OVA323–339 peptide (OVA) in the presence or absence of TGF-β. After 4 d, OT-II cells were restimulated for 6 h with plated bound antibodies to CD3 and CD28. Foxp3 expression and, intracellular production of IL-17, IFN-γ and IL-10 by CD4+ T cell were assessed by intracellular staining and flow cytometry. Data are from one representative experiment of three. (b) Induction of Treg cells by zymosan stimulated splenic DCs is dependent on RA. Splenic DCs were stimulated with zymosan in the presence of disulphiram or vehicle for 10 h, and washed and cultured with naïve OT-II T cells with OVA and TGF-β in the presence or absence of retinol. After 4 d, Foxp3 expression by CD4+ T cell was assessed by intracellular staining and flow cytometry. Data are representative of one experiment of three. (c) Induction of Treg cells by zymosan stimulated splenic DCs, is dependent on TLR2. Splenic DCs from wild type or Tlr2−/− mice were stimulated with zymosan for 10h, and washed and cultured with naïve OT-II T cells with OVA and TGF-β in the presence or absence of retinol. After 4 d, Foxp3 expression by CD4+ T cell was assessed by intracellular staining and flow cytometry. Data are representative of one experiment of three. (d) Effect of blocking ERK activation in zymosan stimulated DCs on Treg induction. Data are representative of one experiment of three.

presence or absence of retinol. After 4 d, Foxp3 expression by CD4+ T cell was assessed by intracellular staining and flow cytometry. Data are representative of one experiment of three. (d) Effect of blocking ERK activation in zymosan stimulated DCs on Treg induction. Data are representative of one experiment of three. Figure 3 RA and IL-10 exert autocrine effects on DCs to induce Socs3 which regulates activation of p38 MAPK and pro-inflammatory cytokines. (a) Expression of RA nuclear receptors (RARs and RXRs) in splenic DCs by western blot. (b, c) Cytokines secreted in supernatants obtained after culture of splenic DCs from wild-type mice (b), and Il-10−/− mice (c), with zymosan for 24 h, in the presence or absence of antibodies against IL-10 receptor (aIL-10R) IL-10R, retinol or retinol plus LE135/ LE540, or IL-10. Representative of four experiments. (d, e) RA dependent induction of Socs3 mRNA expression in splenic DCs from wild type (d) or IL-10−/− mice (e) stimulated with zymosan. (f, g) RA-dependent induction of Socs3 in splenic DCs in vivo. C57BL/6 mice were injected with zymosan or zymosan plus disulphiram (g), or zymosan plus LE135/ LE540 (g), and spleens were harvested at different time points. RNA was isolated from purified splenic DCs, and expressions of Socs1 and Socs3 mRNA were analyzed by quantitative RT-PCR. (h) Induction of Socs3 in splenic DCs in vivo is dependent on TLR2. Wild type or Tlr2−/− mice were injected with zymosan, splenic DCs isolated and mRNA expression of Socs3 evaluated by RT-PCR. Representative of three experiments. *P < 0.01; **P < 0.001; ***P < 0.0001 in all figures.

ere analyzed by quantitative RT-PCR. (h) Induction of Socs3 in splenic DCs in vivo is dependent on TLR2. Wild type or Tlr2−/− mice were injected with zymosan, splenic DCs isolated and mRNA expression of Socs3 evaluated by RT-PCR. Representative of three experiments. *P < 0.01; **P < 0.001; ***P < 0.0001 in all figures. Figure 4 Induction of antigen specific IL-10+ Tr1 and Treg cells in vivo. (a) B6.PL (Thy1.1+) mice reconstituted with OT-II TCR transgenic T cells were injected i.v. with class II–restricted OVA323–339 peptide (OVA) alone, or OVA plus LPS, OVA plus zymosan, or OVA plus curdlan. Four days after challenge, the splenocytes were isolated and expression of Foxp3, IL-17, IFN-γ and IL-10 by CD4+ T Thy1.2+ cell was assessed by intracellular staining and flow cytometry. Data are from one experiment representative of two. (b) C57BL/6 or Tlr2−/− or Il-10−/− mouse were reconstituted with OT-II TCR transgenic T cells, and on the following day, injected with OVA or OVA plus zymosan. Five days later, splenocytes were isolated and induction of OVA specific Foxp3+ T cells was assessed by intracellular staining and flow cytometry. Means + s. d. of 3 or 4 mice per group. (c) Splenocytes from immunized mice described in (b) were restimulated with OVA in culture for 48 h and cytokines in the supernatants were analyzed by ELISA. Means + s. d. of 3 or 4 mice per group. *P < 0.01; **P < 0.001; ***P < 0.0001 in all figures.

staining and flow cytometry. Means + s. d. of 3 or 4 mice per group. (c) Splenocytes from immunized mice described in (b) were restimulated with OVA in culture for 48 h and cytokines in the supernatants were analyzed by ELISA. Means + s. d. of 3 or 4 mice per group. *P < 0.01; **P < 0.001; ***P < 0.0001 in all figures. Figure 5 Zymosan suppresses IL-23 and TH-17 mediated EAE. (a) In wild type mice (left panel), immunization with MOG + zymosan resulted in substantially reduced EAE, relative to immunization with MOG + CFA. In Tlr2−/− mice, immunization with MOG + zymosan resulted in enhanced disease, relative to wild type mice (P < 0.0001). Representative experiment of two. (b) Wild type or Tlr2−/− mice were immunized s.c with MOG + CFA, MOG + CFA plus zymosan (i.v) or MOG + CFA plus curdlan (i. v), and monitored for disease. Injection of zymosan suppressed the severity of disease, relative to immunization with MOG + CFA alone, in wild type mice (left panel), but not in Tlr2−/− mice (right panel). Representative experiment of two. (c) Mononuclear cells were isolated from CNS tissue on day 18 after immunization and induction of IFN-γ, IL-17, IL-10 and Foxp3 was assessed by intracellular staining and flow cytometry, as described in Supplementary Methods, online. Representative experiment of two. (d) Wild type or Tlr2−/− mice were injected with zymosan, and expression of Il23p19 mRNA in splenic DCs analyzed by quantitative RT-PCR. Representative of three experiments. (e) IL-23 induction in the serum of the mice described in (d) was assayed by ELISA. (f) TLR2 regulates IL-23 production in splenic DCs in response to zymosan. IL-23 secretion by splenic DCs from wild type or Tlr2−/− mice cultured in vitro with zymosan. *P < 0.01; **P < 0.001; ***P < 0.0001.

tative of three experiments. (e) IL-23 induction in the serum of the mice described in (d) was assayed by ELISA. (f) TLR2 regulates IL-23 production in splenic DCs in response to zymosan. IL-23 secretion by splenic DCs from wild type or Tlr2−/− mice cultured in vitro with zymosan. *P < 0.01; **P < 0.001; ***P < 0.0001. Figure 6 Mechanism of induction of Raldh2 in DCs. Innate sensing of zymosan via TLR2 efficiently (thick arrows) induces ERK activation and Raldh2. Dectin-1 does not play a major role in Raldh2 induction (thin arrow), although signaling via syk is critical, suggesting the involvement of an additional syk-dependent receptor (X). Thus, the combinatorial activation of TLR2 dependent ERK and syk, likely orchestrates induction of Raldh2. This results in the conversion of retinal to RA, which then exerts an autocrine effect on DCs via RAR/RXR to induce SOSC3, which suppresses activation of p38 MAPK and pro-inflammatory cytokines. In contrast, dectin-1 promotes induction of pro-inflammatory cytokines.

INTRODUCTION Organelle-mediated stress, particularly endoplasmic reticulum (ER) stress, has recently emerged as an important pathophysiological paradigm underlying chronic metabolic diseases1–8. In conjunction with its central role in protein synthesis, folding and transportation, the ER serves as a critical site for integrating cellular responses to stress9. The presence of misfolded proteins and other stresses lead to the activation of an adaptive program by the ER, known as the unfolded protein response (UPR), to reestablish equilibrium9. Initiation of the canonical UPR engages three distinct signaling branches mediated by pancreatic ER kinase (PERK), inositol-requiring transmembrane kinase/endonuclease 1 (IRE1) and activating transcription factor 6 (ATF6). The UPR is also linked to the activation of stress kinases such as the c-Jun N-terminal kinase (JNK). The combined action of these pathways leads to inhibition of protein translation, stimulation of protein degradation, and the production of chaperone proteins resulting in either recovery of ER function or cell death10.

he UPR is also linked to the activation of stress kinases such as the c-Jun N-terminal kinase (JNK). The combined action of these pathways leads to inhibition of protein translation, stimulation of protein degradation, and the production of chaperone proteins resulting in either recovery of ER function or cell death10. In obesity, activation of the UPR in metabolic tissue contributes to insulin resistance, at least in part through IRE-1-dependent, JNK-1-mediated inhibition of insulin action5. Promotion of ER stress by genetic X-box binding protein 1 (XBP-1) haploinsufficiency, which functions in the UPR-induced transcriptional program, also leads to systemic insulin resistance, while alleviation of ER stress by chemical or molecular chaperones protects mice against insulin resistance and type 2 diabetes5–7. Activation of ER stress response pathways is also a characteristic of lipid-laden macrophages in atherosclerotic lesions in mice and humans and is proposed to play a role in plaque vulnerability and acute cardiac death3,4,8. However, the role of ER stress in macrophages and cardiovascular disease remains obscure and it is unknown whether the modulation of ER stress pathways could alter the function and survival of macrophages and the course of atherosclerosis. Moreover, it is unclear how accumulation of excess lipids in macrophages can engage the ER stress response pathways. Despite continuing debate, it is likely that the biological effects of toxic lipids such as those prevalent in dyslipidemia are signaled through specific pathways rather than lipotoxicity representing a non-specific demise of cellular function and viability. Several pieces of evidence suggest a connection between lipid metabolism and the UPR. For example, XBP-1 plays a role in ER phosphatidylcholine synthesis and endomembrane expansion, and has been linked to transcriptional regulation of several lipogenic genes in the liver11,12. ER stress can induce lipogenesis and promote hepatic steatosis12–15. On the other hand, inhibition of phospholipid synthesis or increasing phospholipase activity exacerbates ER stress responses and sphingolipid levels can influence ER function16,17. Additionally, ER stress was identified as a mechanism driving free cholesterol-induced cell death, in a model of cholesterol loading3. Hence, it is possible that the ER may serve as an important target organelle that senses stresses related to lipid status and exposure10.

ngolipid levels can influence ER function16,17. Additionally, ER stress was identified as a mechanism driving free cholesterol-induced cell death, in a model of cholesterol loading3. Hence, it is possible that the ER may serve as an important target organelle that senses stresses related to lipid status and exposure10. However, the signaling networks linking ER function, lipid metabolism and the physiological outcomes are not known. Cellular lipid metabolism and reception of lipid signals are regulated by cytosolic lipid chaperones, particularly fatty acid binding protein-4 (aP2), which exhibit profound effects on chronic metabolic diseases and whose function is relevant to human disease10,18,19–23. The dramatic impact of aP2 on atherosclerosis is related exclusively to its action in the macrophages although the underlying mechanisms are not fully resolved22. The fact that macrophage aP2-deficiency can mediate protection from atherosclerosis in the setting of severe dyslipidemia raised the possibility that lipid chaperones may be a link between toxic lipids and organelle stress in macrophages. Here, we explored the mechanisms related to lipotoxic macrophage ER stress and the impact of ER dysfunction on atherosclerosis utilizing a chemical chaperone and lipid chaperone-deficient mouse model. We demonstrate that mitigation of ER stress is protective against atherosclerosis and that aP2 is an obligatory intermediate for macrophage ER stress responses to lipids.

ic macrophage ER stress and the impact of ER dysfunction on atherosclerosis utilizing a chemical chaperone and lipid chaperone-deficient mouse model. We demonstrate that mitigation of ER stress is protective against atherosclerosis and that aP2 is an obligatory intermediate for macrophage ER stress responses to lipids. RESULTS I: Blocking macrophage ER stress and atherosclerosis The chemical chaperone, 4-phenyl butyric acid (PBA) can alleviate ER stress, and hence, provides an experimental opportunity to approach the role of ER stress in atherosclerosis7,24. We first tested whether PBA can alter ER stress induced in macrophages upon exposure to saturated fatty acids25,26. Treatment with palmitate (PA) induced ER stress in macrophages as determined by phosphorylation of PERK (P-PERK) and eukaryotic translation initiation factor 2α (P-eIF2–α). However, co-treatment with PBA resulted in essentially complete protection against PA-induced ER stress (Fig.1a). PBA treatment also suppressed PA-induced splicing of XBP-1 (sXBP-1) and C/EBP homologous protein (CHOP) expression, two elements of the UPR-induced transcriptional program (Fig.1b,c). Since saturated fatty acids or modified lipoproteins can induce apoptotic pathways we next asked whether modifying ER stress in this setting could prevent lipotoxic death in macrophages27. Treatment with PBA resulted in marked protection against PA-induced apoptosis in macrophages, as determined by TUNEL assays (Fig.1d and Supplementary Fig.1a). These results demonstrate that PBA can protect cultured macrophages against lipid-induced ER stress and apoptosis in vitro.

ent lipotoxic death in macrophages27. Treatment with PBA resulted in marked protection against PA-induced apoptosis in macrophages, as determined by TUNEL assays (Fig.1d and Supplementary Fig.1a). These results demonstrate that PBA can protect cultured macrophages against lipid-induced ER stress and apoptosis in vitro. Next, to elucidate whether PBA could mitigate ER stress in atherosclerotic lesions in vivo, we analyzed ER stress indicators and apoptosis in serial sections from aortic sinuses of atherosclerotic mice briefly treated with PBA. Six-week-old, male, ApoE−/− mice were fed with a Western diet for 8 weeks, and given daily doses (10 mg kg−1 or 100 mg kg−1) of PBA or vehicle during the final 2 weeks. Examination of the sections of the aortic sinus for proximal lesions demonstrated that ApoE−/− mice receiving PBA showed a dose-dependent reduction (9%, NS and 32%, p<0.05, respectively) in atherosclerosis (Fig.1e). At this early stage of atherosclerosis, en face analysis was similar between the groups (Supplementary Fig.1b). Of note, suppression of macrophage ER stress and reduction in vascular lesions by PBA treatment occurred in the absence of any impact on lipids, lipoprotein profiles, glucose and insulin levels in the circulation and body weight (30.6±0.4g vs. 29.5±0.6g; p=0.19) (Supplementary Fig.1c–f and data not shown). Hence, this relatively short treatment period provided a suitable experimental design to examine the status of ER stress indicators and apoptosis in atherosclerotic lesions without significant changes in systemic metabolic parameters, insulin sensitivity or dramatic alterations in total burden of lesions that occurs with longer PBA treatment (data not shown)7,24. Immunohistochemical analysis revealed that all mice developed early lesions that predominantly contained macrophages and macrophage-derived foam cells (shown by staining with monocyte/macrophage-specific antibody, MOMA-2) (Fig.1f). There was no significant reduction in total macrophage area upon PBA treatment despite reduction in lesion size (macrophage area in control: 123097±13711 and in PBA treatment: 170545±30249µm2). The atherosclerotic lesions of ApoE−/− (control) mice stained positive for the ER stress markers, P-eIF2–α and P-PERK in macrophage dense areas (Fig.1f). In contrast, P-eIF2–α and P-PERK staining was significantly diminished in atherosclerotic lesions of mice treated with PBA (Fig.1f).

PBA treatment: 170545±30249µm2). The atherosclerotic lesions of ApoE−/− (control) mice stained positive for the ER stress markers, P-eIF2–α and P-PERK in macrophage dense areas (Fig.1f). In contrast, P-eIF2–α and P-PERK staining was significantly diminished in atherosclerotic lesions of mice treated with PBA (Fig.1f). Immunofluorescent staining revealed a 44% reduction in ATF3 expression (p<0.05) and a 53% reduction in P-eIF2–α expression (p<0.05) in the macrophage-dense areas of the lesions following PBA treatment (Fig.1g and Supplementary Fig.1g). We also examined the extent of apoptosis in these atherosclerotic lesions using TUNEL staining. Atherosclerotic lesions from control mice contained abundant apoptotic cells whereas lesions from PBA treated mice exhibited significant reduction in apoptotic cells in a dose dependent manner (29% and 42%, respectively; p<0.05 in the 100 mg kg−1 dose) (Fig.1h). These results demonstrate that PBA treatment leads to marked reduction in macrophage ER stress and apoptosis in atherosclerotic lesions in vivo indicating that improvement of ER chaperoning function can protect against the deleterious effects of toxic lipids in promoting atherosclerotic lesions.

mg kg−1 dose) (Fig.1h). These results demonstrate that PBA treatment leads to marked reduction in macrophage ER stress and apoptosis in atherosclerotic lesions in vivo indicating that improvement of ER chaperoning function can protect against the deleterious effects of toxic lipids in promoting atherosclerotic lesions. II: Regulation of lipid-induced ER stress by a lipid chaperone It has been shown that aP2-deficiency in macrophages protects against atherosclerosis despite a highly unfavorable lipid profile22,28,29. Since, toxic lipids fail to trigger ER stress in the presence of PBA, we investigated whether lipid chaperoning activity is linked to ER stress and PBA activity in macrophages. Notably, aP2 immunostaining was significantly suppressed in vascular lesions of ApoE−/− mice treated with PBA when compared to vehicle treated animals (71%, p<0.05) (Fig.2a,b). Furthermore, PA induced rapid and marked upregulation of aP2 protein, but not aP2 mRNA levels (Fig.2c). Induction of aP2 by PA was prevented upon co-treatment with PBA (Fig.2c). Taken together, these results indicate that aP2 expression is directly related to lipid-induced ER stress and strongly inhibited by PBA in macrophages in vitro and in vivo. This observation led us to ask whether aP2 regulates ER responses to lipid stress in macrophages. Treatment of wild type (WT) macrophages with long-chain saturated fatty acids such as PA or stearate (STE), but not their monounsaturated counterparts, led to ER stress as judged by the robust phosphorylation of PERK and eIF2–α, activation of JNK, and induction Ddit3 and sXBP-1 expression (Fig.2d,e and Supplementary Fig.2b,d and g). However, PA failed to induce ER stress in aP2-deficient (aP2−/−) macrophages (Fig.2d,e and Supplementary Fig.2g). In these cell lines, there were no alterations in PA uptake (Supplementary Fig.2e). The aP2−/− cells also maintained the ability to respond to tunicamycin, an inhibitor of protein glycosylation that leads to ER stress, indicating that they do not suffer from a general defect in mounting ER stress responses (Fig.2d,e).

In these cell lines, there were no alterations in PA uptake (Supplementary Fig.2e). The aP2−/− cells also maintained the ability to respond to tunicamycin, an inhibitor of protein glycosylation that leads to ER stress, indicating that they do not suffer from a general defect in mounting ER stress responses (Fig.2d,e). We then compared ER stress responses in aP2−/− macrophages reconstituted with a lipid-binding mutant (LM) of aP2 (R126L, Y128F; aP2−/−LM) to those reconstituted with WT aP2 (aP2−/−REC)30. Under the conditions tested, WT- and LM-aP2 proteins were expressed at comparable levels to each other and did not lead to any alterations in lipid uptake (Supplementary Fig.2e and data not shown). Reconstitution of WT-aP2 rendered aP2−/− macrophages responsive to PA as indicated by the induction of P-PERK, P-eIF2–α and JNK activity, while aP2−/−LM macrophages remained markedly resistant to PA-induced ER stress (Fig.2d). PA also induced the expression of Ddit3 and sXBP-1 in WT but not in aP2−/− macrophages (Supplementary Fig.2g and Fig.2e). In aP2−/− macrophages, reconstitution of aP2 restored the induction of UPR target genes by PA (Fig.2d,e and Supplementary Fig.2g). Consistent with their ER stress-resistant phenotype, aP2−/− macrophages were also significantly protected against PA-induced apoptosis as indicated by suppression of caspase-3 activity and PARP cleavage (Fig.2f and Supplementary Fig.2h). Reconstitution of WT-aP2 into the aP2−/− macrophages rendered these cells susceptible to lipid-induced apoptosis whereas cells expressing LM-aP2 remained refractory to apoptosis, demonstrating the requirement for the lipid binding activity of aP2 in regulating ER stress responses (Fig.2f and Supplementary Fig.2h).

2h). Reconstitution of WT-aP2 into the aP2−/− macrophages rendered these cells susceptible to lipid-induced apoptosis whereas cells expressing LM-aP2 remained refractory to apoptosis, demonstrating the requirement for the lipid binding activity of aP2 in regulating ER stress responses (Fig.2f and Supplementary Fig.2h). Next, we exposed macrophages to free cholesterol (FC) loading in order to examine ER stress responses in another setting of lipotoxicity associated with the pathogenesis of atherosclerosis4. In WT, but not aP2−/− macrophages, FC induced ER stress as indicated by the induction of P-PERK and P-eIF2–α, sXBP-1, and JNK activity (Fig.2g). These observations were also independent of compromised cholesterol uptake; in fact, aP2−/− macrophages exhibit increased cholesterol influx31. Responsiveness to FC-induced ER stress was restored in aP2−/− macrophages, upon reconstitution of aP2, demonstrating that aP2 mediates FC-induced ER stress in macrophages (Fig.2g). Furthermore, aP2−/− macrophages were resistant to FC-induced induced apoptosis, determined by activation of caspase 3 and cleavage of PARP, that normally occurs in FC-treated WT macrophages (Supplementary Fig.2i)3. Taken together our results show that aP2 is necessary for toxic lipids to trigger ER stress and apoptosis in macrophages.

−/− macrophages were resistant to FC-induced induced apoptosis, determined by activation of caspase 3 and cleavage of PARP, that normally occurs in FC-treated WT macrophages (Supplementary Fig.2i)3. Taken together our results show that aP2 is necessary for toxic lipids to trigger ER stress and apoptosis in macrophages. III: Regulation of macrophage ER stress by aP2 in vivo We next investigated whether aP2-deficiency in mice can modulate ER stress responses in vascular lesions in vivo. The aP2−/− mouse model provides an ideal setting to examine the link between macrophage ER stress and atherosclerosis since aP2-deficiency does not alter the hyperlipidemia or any other metabolic parameters in the ApoE−/− background, and furthermore, aP2’s impact on atherosclerosis is predominantly, if not completely, mediated by its action in the macrophages22. The early stage atherosclerotic plaques from ApoE−/− mice showed induction of ER stress as indicated by elevated P-PERK, P-eIF2–α, and Ddit3 mRNA in the infiltrating macrophages within the lesions (Fig.3a). Notably, the vascular lesions in the aP2−/− ApoE−/− mice were essentially devoid of staining for markers of ER stress (Fig.3a). Quantitative analysis of ER stress by immunofluorescent staining demonstrated a significant reduction in P-eIF2–α and ATF3 (55% and 67%, respectively; p<0.05) expression in macrophage-rich areas of the lesions (Fig.3b,c). Furthermore, TUNEL assays demonstrated a significant reduction in the number of apoptotic macrophages in lesions of aP2−/−ApoE−/− mice compared to ApoE−/− animals (7.8% and 18.4%, respectively; p<0.05) (Fig.3d,e and Supplementary Fig.3a). This confirmed the critical role of aP2 in mediating macrophage ER stress response to toxic lipids in vivo, similar to its actions in cultured macrophages in vitro.

ptotic macrophages in lesions of aP2−/−ApoE−/− mice compared to ApoE−/− animals (7.8% and 18.4%, respectively; p<0.05) (Fig.3d,e and Supplementary Fig.3a). This confirmed the critical role of aP2 in mediating macrophage ER stress response to toxic lipids in vivo, similar to its actions in cultured macrophages in vitro. The impact of genetic aP2-deficiency can be mimicked by a specific chemical inhibitor for aP2 (aP2-i) in vitro and in vivo29. Treatment of WT macrophages with aP2-i also led to marked protection against PA-induced ER stress as assessed by diminished P-PERK, P-eIF2–α, Ddit3 and sXBP-1 expression, without any effects on fatty acid uptake (Fig.3f,g and Supplementary Fig.3b). Treatment with aP2-i also reduced ATF3 expression in macrophage-dense areas of the plaques in ApoE−/− mice in vivo, without altering hyperlipidemia (Fig.3h)29. These results clearly demonstrate that genetic or chemical ablation of aP2 protects macrophages from ER stress in vivo in the context of hypercholesterolemic atherosclerosis.

eated aP2−/− macrophages, but not in scrambled (control) siRNA treated cells, as evidenced by increased P-PERK and P-eIF2–α, active caspase 3 and cleaved PARP levels (Fig.5g). These results demonstrate that aP2-mediated regulation of SCD-1 activity is causally linked to lipid-induced ER stress responses in macrophages. Unraveling the molecular mechanisms by which aP2 regulates SCD-1 is important for understanding how lipid-stress signals impinge on the lipid synthetic pathways. Both SCD-1 and Fasn are direct transcriptional targets of the nuclear receptor LXR (LXR-responsive elements, LXRE, are located on Fasn, between positions −669 and −665, and SCD-1, between positions −1263 and −1248) (Fig.6a)35,36. Thus, we analyzed whether LXR activity is altered in aP2-deficient macrophages by utilizing an LXRE-driven reporter. The aP2−/− macrophages displayed a significant elevation in stimulated LXR activity when compared to control, aP2−/−REC cells (Fig.6b). These results indicate that aP2 negatively regulates LXR activity in WT macrophages. Consistently, we observed increased expression of LXR target genes Abca1, Abcg1 and CD51 (AIM) in aP2−/− macrophages (Supplementary Fig.4a–c)37–39. Further examination revealed marked elevation of Nr1h3 (LXR–α) mRNA and protein levels in the absence of aP2, while Nr1h2 (LXR–β) expression remained unchanged between the genotypes (Fig.6c and Supplementary Fig.4d,e). Hence, the LXR–α expression appeared to be the main driver of the alterations seen in LXR target gene expression. In order to definitively link LXR–α activity to SCD-1 regulation, we next suppressed Nr1h3 expression in aP2−/− macrophages using a siRNA-mediated knock down approach (Supplementary Fig.5a). Reduction of Nr1h3 expression had only a partial effect on the expression genes regulated by both LXR–α and –β, such as Abca1 and Abcg1, but generated a profound effect on the expression of an LXR–α exclusive target gene, CD51 (AIM) (Supplementary Fig.5b–d)37–39. The reduction in LXR–α also led to significant inhibition of both Fasn and SCD-1 mRNA levels, demonstrating that LXR–α is mainly responsible for the upregulation of these genes in the absence of aP2 (Fig.6d and Supplementary Fig.5e). Suppression of LXR–α also restored sensitivity to PA-induced ER stress and apoptosis as determined by induction of P-PERK and P-eIF2–α and cleaved PARP levels in aP2−/− macrophages (Fig.6e).

aP2-i also reduced ATF3 expression in macrophage-dense areas of the plaques in ApoE−/− mice in vivo, without altering hyperlipidemia (Fig.3h)29. These results clearly demonstrate that genetic or chemical ablation of aP2 protects macrophages from ER stress in vivo in the context of hypercholesterolemic atherosclerosis. IV: Mechanisms linking aP2 to macrophage ER stress In order to identify the metabolic pathways that control the observed tolerance to toxic lipids, we next studied the impact of aP2 on macrophage lipid composition and metabolism. We analyzed the profiles of individual fatty acids (FA) in macrophages in a systematic manner using high resolution, quantitative lipidomic analysis. aP2−/− macrophages contained elevated levels of monounsaturated fatty acids (MUFA), indicating that a greater proportion of the lipids in these cells were produced de novo (Supplementary Table 1). Furthermore, we saw evidence of increased delta 9 desaturase/steaoryl CoA desaturase (Δ9D/SCD) activity in aP2−/− macrophages, reflected in the elevated C16:1n7/C16:0; C14:1n5/C14:0, and C18:1n9/C18:0 ratios of FA present in various lipid classes (Fig.4a and Supplementary Table 1). Enhancement of de novo lipogenesis in aP2−/− macrophages resulted in marked increase in C16:1n7 levels and its direct elongation product, C18:1n7 in addition to a modest elevation in C14:1n5 and C18:1n9 (Fig.4b and Supplementary Table 1). The reconstitution of aP2 dramatically shifted the FA profile from one of active de novo synthesis, with enhanced desaturase activity, to high elongase activity with little desaturase action, demonstrating that macrophage de novo lipogenesis is strongly regulated by aP2 (Fig.4a,b and Fig.5a).

nd Supplementary Table 1). The reconstitution of aP2 dramatically shifted the FA profile from one of active de novo synthesis, with enhanced desaturase activity, to high elongase activity with little desaturase action, demonstrating that macrophage de novo lipogenesis is strongly regulated by aP2 (Fig.4a,b and Fig.5a). We next asked which lipid classes are enriched by these newly synthesized fatty acids and determined the distribution of all major classes of lipids. When compared to WT macrophages, aP2−/− macrophages had elevated phospholipids (PL;%138), triglycerides (TG;140%), diacylglycerol (DG;143%) and free fatty acid (FFA;224%) concentrations and lower levels of cholesterol esters (CE;79%) (Fig.4c,d). Reconstitution of aP2 suppressed PL levels by 31% and dramatically increased CE concentrations by 197%, demonstrating a crucial role for aP2 in the regulation of macrophage phospholipid and cholesterol production (Fig.4c,d). The marked increase in the phospholipid-to-cholesterol ratio seen in the aP2−/− macrophages suggests that one potential way aP2 can modulate stress responses to toxic lipids may be through alteration in membrane lipid composition and metabolic properties.

f macrophage phospholipid and cholesterol production (Fig.4c,d). The marked increase in the phospholipid-to-cholesterol ratio seen in the aP2−/− macrophages suggests that one potential way aP2 can modulate stress responses to toxic lipids may be through alteration in membrane lipid composition and metabolic properties. The results of lipidomic analysis implicated regulation of de novo lipogenesis and desaturation, a rate-limiting step catalyzed by SCD, as a potential mechanism underlying the aP2-driven compositional changes in macrophages (Fig.5a). The desaturase activity of SCD converts saturated FA to MUFA which are then incorporated into phospholipids, triglycerides, and cholesterol esters32,33. Indeed, aP2−/− macrophages are enriched with MUFA, most significantly in C16:1n7-palmitoleate and its direct elongation product, C18:1n7, across all major lipid classes (Fig.4b and Supplementary Table 1). Consistently, SCD-1 expression in aP2−/− macrophages was ~50 fold-higher than that of WT controls (Fig.5b). Next, we examined aP2-regulated SCD-1 expression in macrophages in vivo and found that its expression was markedly upregulated in the peritoneal macrophages isolated from mice treated with aP2-i (Fig.5c). Fatty acid synthase (Fasn) expression was also significantly elevated in aP2−/− macrophages and induced upon aP2-i treatment in vivo (data not shown). Collectively, these data indicate that aP2 action in macrophages is linked to the regulation of key enzymes involved in the de novo synthesis and desaturation of fatty acids.

acid synthase (Fasn) expression was also significantly elevated in aP2−/− macrophages and induced upon aP2-i treatment in vivo (data not shown). Collectively, these data indicate that aP2 action in macrophages is linked to the regulation of key enzymes involved in the de novo synthesis and desaturation of fatty acids. To examine the impact of aP2-regulated SCD-1 activity in the resistance to ER stress in aP2−/− macrophages, we took two distinct but related approaches. First, we asked whether C16:1n7-palmitoleate, a product of de novo lipogenesis elevated in aP2−/− macrophages, could modify ER responses to lipids34. Strikingly, we found that WT macrophages pre-treated with C16:1n7-palmitoleate became resistant to PA-induced ER stress and apoptosis, but not to tunicamycin-induced ER stress, determined by the P-PERK, P-eIF2–α, cleaved PARP and Ddit3 and sXBP-1 levels. This pattern induced by C16:1n7-palmitoleate was highly reminiscent of genetic or chemical aP2-deficiency (Fig.5d,e). Oleate, another FA product of desaturation, which is not regulated to the same extent as palmitoleate in aP2−/− macrophages, could also protect against PA or STE- induced ER stress (Supplementary Fig.2b–d).

pattern induced by C16:1n7-palmitoleate was highly reminiscent of genetic or chemical aP2-deficiency (Fig.5d,e). Oleate, another FA product of desaturation, which is not regulated to the same extent as palmitoleate in aP2−/− macrophages, could also protect against PA or STE- induced ER stress (Supplementary Fig.2b–d). Second, we utilized a siRNA-mediated approach to significantly deplete SCD-1 protein and activity in aP2−/− macrophages (Fig.5f). Sensitivity to PA-induced ER stress and apoptosis was re-established in the SCD-1 siRNA treated aP2−/− macrophages, but not in scrambled (control) siRNA treated cells, as evidenced by increased P-PERK and P-eIF2–α, active caspase 3 and cleaved PARP levels (Fig.5g). These results demonstrate that aP2-mediated regulation of SCD-1 activity is causally linked to lipid-induced ER stress responses in macrophages.

ing that LXR–α is mainly responsible for the upregulation of these genes in the absence of aP2 (Fig.6d and Supplementary Fig.5e). Suppression of LXR–α also restored sensitivity to PA-induced ER stress and apoptosis as determined by induction of P-PERK and P-eIF2–α and cleaved PARP levels in aP2−/− macrophages (Fig.6e). To validate these links in a genetic setting, we examined lipid-induced ER stress response in primary peritoneal macrophages derived from mice with aP2 and Nr1h3 combined genetic deficiency. The expression of Fasn and SCD-1 was also markedly down-regulated in aP2−/−Nr1h3−/− macrophages compared to aP2−/− cells (Supplementary Fig.6b,c)40. The expression of CD51 was profoundly suppressed in double mutant cells, while Abca1 and Abcg1 were only partially affected (Supplementary Fig.6d–f)37–39. Consequently, the protection against PA-induced ER stress was also lost in the aP2−/−Nr1h3−/− cells as determined by P-PERK, P-eIF2–α and cleaved PARP induction (Fig.6f). Consistently, treatment of WT macrophages with a specific LXR agonist, T0901317, promoted resistance to lipid-induced ER stress, similar to aP2-deficiency (Supplementary Fig.5f). These results illustrate that LXR–α is responsible for the upregulation of SCD-1 in the absence of aP2 and provides one crucial mechanism for how lipid stress signals may impinge on macrophage lipid metabolism and ER stress (Fig.6g).

sistance to lipid-induced ER stress, similar to aP2-deficiency (Supplementary Fig.5f). These results illustrate that LXR–α is responsible for the upregulation of SCD-1 in the absence of aP2 and provides one crucial mechanism for how lipid stress signals may impinge on macrophage lipid metabolism and ER stress (Fig.6g). DISCUSSION Macrophages are particularly vulnerable to lipid-induced toxicity and contribute to the pathogenesis of several metabolic derangements where exposure to lipids is increased, such as the foam cells in hypercholesterolemic atherosclerosis and in adipose tissue-associated macrophages in obesity2,10,41. Previous studies have shown expression of UPR markers in macrophages infiltrating the atherosclerotic lesions of both mice and humans4,8. These findings are complemented by in vitro studies demonstrating that accumulation of free cholesterol leads to apoptosis via activation of ER stress in macrophages3. Elevation of ER stress-associated macrophage apoptosis has been proposed to contribute to advanced atherosclerotic lesions in macrophages defective in insulin signaling42. While these findings have sparked interest, demonstration of a link between macrophage ER stress and atherogenesis in vivo and the mechanisms integrating lipid signals to ER function in macrophages has been challenging. Hence, modulation of ER stress, especially upstream of the apoptotic execution pathways, becomes critical in understanding the extent of its contribution to the pathogenesis of atherosclerosis. The data presented in this paper provide insight into these critical questions. First, we demonstrate that the lipid chaperone aP2 is an obligatory mediator coupling toxic lipids to ER stress in macrophages in vitro and in vivo. Second, we show that alleviating ER stress, either by the use of a chemical chaperone or through the inhibition of a lipid chaperone, provides significant protection against macrophage ER stress, cell death, and atherosclerosis.

obligatory mediator coupling toxic lipids to ER stress in macrophages in vitro and in vivo. Second, we show that alleviating ER stress, either by the use of a chemical chaperone or through the inhibition of a lipid chaperone, provides significant protection against macrophage ER stress, cell death, and atherosclerosis. The surprising and striking upregulation of aP2 by saturated fatty acids and downregulation by PBA in macrophages led to the uncovering of a previously unknown function of the lipid chaperones in mitigating ER stress in macrophages in vitro and in atherosclerotic lesions in vivo. The observations place aP2 as a central modulator of lipid—induced ER stress responses—a process for which there had been previously little mechanistic insight. It is also established that the role of macrophage aP2 action on atherosclerosis is not related to other metabolic alterations, such as changes in insulin sensitivity or dyslipidemia, and is therefore intrinsic to macrophages22,43. Similarly, the treatment with PBA dose used in this study, and as reported earlier, does not yield significant metabolic alterations7. Therefore, our present findings unravel a mechanism by which aP2 could mediate its anti-atherosclerotic effects through modulating the UPR, helping to clarify a much sought after but very challenging aspect of lipid chaperone biology in macrophages and in atherosclerosis4,8. Finally, our findings offer tools to modulate ER stress responses associated with dyslipidemia in vitro and in vivo that may facilitate therapeutic applications including that of aP2.

helping to clarify a much sought after but very challenging aspect of lipid chaperone biology in macrophages and in atherosclerosis4,8. Finally, our findings offer tools to modulate ER stress responses associated with dyslipidemia in vitro and in vivo that may facilitate therapeutic applications including that of aP2. Dramatic resistance to ER stress can be achieved by blocking aP2 action, which is dependent on SCD-1 activity. aP2 prevents enrichment of macrophages in desaturation products such as C16:1n7-palmitoleate, a molecule which provides relief from lipid-induced ER stress, in addition to its other reported beneficial endocrine effects34. It will be interesting to determine whether C16:1n7 supplementation or a diet enriched with palmitoleate could confer resistance to macrophage ER stress and reduction in atherosclerosis in future studies. In this study, we showed that enhanced LXR–α activity in aP2−/− macrophages drives Fasn and SCD-1 transcription and resistance to lipotoxic ER stress and identified the transcriptional mechanism underlying this function of aP2. These observations also raise the possibility of a specific link between the nuclear hormone receptor LXR–α and ER stress responses. Overall, this study demonstrates that de novo fatty acid synthesis and desaturation can be highly beneficial, if not essential, for defending ER function when macrophages are exposed to toxic lipids. Accordingly, inhibition of SCD-1 in the whole body may not have great therapeutic prospects due to adverse metabolic effects seen in the macrophages and pancreatic β cells, although elevated liver Δ9D activity has been linked to obesity and diabetes10,44. Indeed, systemic inhibition of SCD-1 leads to severe atherosclerosis despite protection from obesity and hepatosteatosis45. Similarly, activation of LXR in the whole body leads to undesirable metabolic side effects, particularly in the liver. On the other hand selective upregulation of LXR and SCD-1 activity in macrophages and adipocytes, and downregulation in liver may be an optimal strategy and yield the most beneficial overall metabolic outcome, including protection against diabetes and atherosclerosis29.

etabolic side effects, particularly in the liver. On the other hand selective upregulation of LXR and SCD-1 activity in macrophages and adipocytes, and downregulation in liver may be an optimal strategy and yield the most beneficial overall metabolic outcome, including protection against diabetes and atherosclerosis29. We unraveled a mechanism by which aP2 could mediate its anti-atherosclerotic effects, at least in part, through modulating the UPR. It is important to note that our findings do not exclude other LXR regulated macrophage responses as additional contributors to aP2’s impact on macrophage function and atherosclerosis. Since the most significant upregulation seen in the aP2−/− macrophages is that of SCD-1 and CD51 genes and knocking down SCD-1 alone is sufficient to re-establish sensitivity to lipid-induced ER dysfunction in the aP2−/− macrophages, LXR-mediated lipogenesis pathways are likely to dominate the aP2-deficient phenotype related to lipotoxic ER stress. Yet, the impact of aP2 or LXR on inflammatory functions of macrophages may also play a role in preventing atherosclerosis. aP2 expression is limited to macrophages and dendritic cells in normal conditions as well as in atherosclerotic lesions (Fig.2B and data not shown)10,22,46,47. While aP2 does not alter antigen presentation, studies have shown that it can affect T cell priming and cytokine production19,46,47. Moreover, ER stress responses and inflammation are integrated at several levels and modulate each other; inflammation can compromise ER function and ER stress can promote inflammation10,. The links between inflammatory pathways and ER stress are of great interest in chronic metabolic disease and understanding the intricate links between lipid metabolism and the immune system and identification of molecular targets like aP2 at this interface are critical for developing effective therapeutics against the metabolic disease cluster.

tory pathways and ER stress are of great interest in chronic metabolic disease and understanding the intricate links between lipid metabolism and the immune system and identification of molecular targets like aP2 at this interface are critical for developing effective therapeutics against the metabolic disease cluster. In conclusion, we uncovered a previously unknown function for aP2 in lipid-induced ER stress signaling in macrophages, while addressing the essential role of ER stress in vascular disease progression (Fig.6g). The ability to defend against the lipids disrupting ER function illustrates a novel metabolic adaptation capacity of macrophages that is governed by the lipid chaperones. aP2 in particular, and perhaps cytosolic lipid chaperones in general, could function as a molecular sensor for fatty acids and as a central coordinator of metabolic-ER stress. Since aP2-deficiency can alleviate the ER stress that occurs during atherosclerosis, similar to the actions of a chemical chaperone, our findings offer insights into the detrimental role of macrophage ER stress in atherosclerosis and the benefits of addressing this target to treat cardiovascular disease.

lic-ER stress. Since aP2-deficiency can alleviate the ER stress that occurs during atherosclerosis, similar to the actions of a chemical chaperone, our findings offer insights into the detrimental role of macrophage ER stress in atherosclerosis and the benefits of addressing this target to treat cardiovascular disease. MATERIALS AND METHODS Mice, immunhistochemistry and quantification of arterial lesions Colonies for aP2−/−ApoE−/−, ApoE−/− and aP2−/−Nr1h3−/− mice on the C57BL/6 background were established in our facilities. Nr1h3−/− mice used in this study were previously published40. Harvard Medical Area Standing Committee on Animals approved the animal handling procedures. Mice were fed with Western diet for 16 weeks as previously described29. After sacrifice, aortas were flushed through the left ventricle and dissected as described earlier29. To detect macrophages in arterial lesions, 5µm serial cryosections of the proximal aorta were fixed in acetone and incubated with antibodies following manufacturer’s recommendations and previously established protocols48. The sections were incubated with biotinylated secondary antibodies and then with alkaline phosphatase-labeled ABC (Vector Lab). Lesion areas were quantified using Imaging System KS 300 (2.0; Kontron Electronik GmbH). For immunofluorescence stainings, Alexa Fluor 488- and Alexa Fluor 647-conjugated secondary antibodies were applied after an overnight incubation of the lesions with the primary antibody. DAPI was used for counterstaining. The mean fluorescent intensity was measured for each corresponding ER stress marker in the macrophage-dense areas using Auxiovision 4.6 software (n≥3). For TUNEL analysis, serial the sections were pretreated with 3% citric acid, fixed in 4% paraformaldehyde and stained using an in situ cell death detection AP kit (Roche). After visualization of alkaline phosphatase with Fast Red TR/Naphthol AS-NX substrate (Sigma), TUNEL-positive cells were counted (n=18 per aorta).

. For TUNEL analysis, serial the sections were pretreated with 3% citric acid, fixed in 4% paraformaldehyde and stained using an in situ cell death detection AP kit (Roche). After visualization of alkaline phosphatase with Fast Red TR/Naphthol AS-NX substrate (Sigma), TUNEL-positive cells were counted (n=18 per aorta). Analysis of serum lipids Mice were fasted (4h) and the serum cholesterol and triglycerides were measured by conventional enzymatic methods using the reagents from Raichem and the SoftMax Pro5 software (Molecular Devices). Quantitative real-time polymerase chain reaction (qRT-PCR) and immunoblot analysis RNA was isolated from macrophages using RNeasy kit (Qiagen), cDNA was synthesized using iScript (Biorad) and qRT-PCR was performed using an ABI Thermocycler. Macrophage proteins were harvested and total protein content was assessed as previously described31. Equal amounts of total protein per sample were subjected to SDS-PAGE and Western blots were performed as described previously49.

as synthesized using iScript (Biorad) and qRT-PCR was performed using an ABI Thermocycler. Macrophage proteins were harvested and total protein content was assessed as previously described31. Equal amounts of total protein per sample were subjected to SDS-PAGE and Western blots were performed as described previously49. Cell culture, knock-down, TUNEL, caspase and reporter assays Unless otherwise indicated, bone-marrow derived mouse macrophage lines were used. Reconstitution experiments were carried out in bone-marrow derived macrophages that were immortalized29. When indicated, peritoneal macrophages derived by 4% thioglycolate elicitation were used. Macrophages were maintained in RPMI supplemented with 10% fetal bovine serum. For caspase 3/7 activity assays, cells were plated on 96 well plates, and upon reaching confluency, treated with the appropriate reagents to activate caspase activity over a time course. Caspase 3/7 activity was measured using the caspase 3/7-Glo kit (Promega) according to the manufacturer’s instructions. For loss-of-function experiments, siRNAs were transfected into 80% confluent macrophages using sImporter (Upstate) according to the manufacturer’s instructions. For reporter assays, macrophages plated on 12-well plates, grown to 80% confluency, and transfected with the appropriate plasmids using Superfect (Qiagen). 24 hours post-transfection, cells were treated with LXR ligands T0901317 (10µM) and 25 hydroxy cholesterol (10µM), or DMSO followed by luciferase assays using the Dual-Luciferase assays system (Promega) according to the manufacturer’s instructions.

and transfected with the appropriate plasmids using Superfect (Qiagen). 24 hours post-transfection, cells were treated with LXR ligands T0901317 (10µM) and 25 hydroxy cholesterol (10µM), or DMSO followed by luciferase assays using the Dual-Luciferase assays system (Promega) according to the manufacturer’s instructions. Lipidomics Fatty acids (FA) were measured in 1 experiment (n=5) by Lipomics, Inc (Sacramento, CA) in each neutral lipid class CE, DAG, PL, FFA and TG. The lipids from plasma and tissues were extracted in the presence of authentic internal standards by the method of Folch et al. using chloroform:methanol (2:1 v/v). Individual lipid classes within each extract are separated by liquid chromatography (Agilent Technologies model 1100 Series)34. Each lipid class was trans-esterified in 1% sulfuric acid in methanol in a sealed vial under a nitrogen atmosphere at 100 °C for 45 min. The resulting FA methyl esters were extracted from the mixture with hexane containing 0.05% butylated hydroxytoluene and prepared for gas chromatography by sealing the hexane extracts under nitrogen. Fatty acid methyl esters were separated and quantified by capillary gas chromatography (Agilent Technologies 6890) equipped with a 30 m DB-88MS capillary column (Agilent Technologies) and a flame-ionization detector.

lated hydroxytoluene and prepared for gas chromatography by sealing the hexane extracts under nitrogen. Fatty acid methyl esters were separated and quantified by capillary gas chromatography (Agilent Technologies 6890) equipped with a 30 m DB-88MS capillary column (Agilent Technologies) and a flame-ionization detector. Statistics One-way ANOVA was used to determine significance in lesion size differences. TUNEL assays: Statistical differences were determined by one-way ANOVA multiple comparisons versus the control group (Dunn’s method) in Fig.1 and. the statistical differences in mean of TUNEL(+)/MOMA-2(+) lesion area between the groups were determined by the Mann-Whitney rank sum test in Fig.3. Lipidomic analyses: The distributions of each fatty acid within each lipid class were examined for extreme outliers or poor measurement. If a fatty acid was missing more than 30% of its observations, it was removed from further analysis. Initial statistical analysis included a one-way ANOVA to identify fatty acids within each lipid class that differed between the genotypes. Results of the one-way ANOVA were visualized using lipid class composition analysis. Direct comparisons between groups were made using a Wilcoxon rank test and visualized using bar plots.

l statistical analysis included a one-way ANOVA to identify fatty acids within each lipid class that differed between the genotypes. Results of the one-way ANOVA were visualized using lipid class composition analysis. Direct comparisons between groups were made using a Wilcoxon rank test and visualized using bar plots. Supplementary Material 1 ACKNOWLEDGEMENTS This project has been supported by grants from National Institutes of Health to GSH DK DK52539 (to GSH), HL65405 (to MFL and GSH) and DK59637 (Lipid, Lipoprotein and Atherosclerosis Core of the Vanderbilt Mouse Metabolic Phenotype Center). EE is supported by the Ruth Kirschstein National Research Award. We are grateful to the members of the Hotamisligil lab, Jie Chen and Rebecca Bachman for their scientific input and contributions, to Aybike Onur for technical assistance, to Rebecca Foote and Kristen Gilbert for administrative support, to David Mangelsdorf (UT Southwestern) for TK-LXRE-X3luc reporter and Nr1h3−/− mice, and Alan Edgar (Fournier, France) for the ACAT inhibitor. CONTRIBUTIONS

Supplementary Material 1 ACKNOWLEDGEMENTS This project has been supported by grants from National Institutes of Health to GSH DK DK52539 (to GSH), HL65405 (to MFL and GSH) and DK59637 (Lipid, Lipoprotein and Atherosclerosis Core of the Vanderbilt Mouse Metabolic Phenotype Center). EE is supported by the Ruth Kirschstein National Research Award. We are grateful to the members of the Hotamisligil lab, Jie Chen and Rebecca Bachman for their scientific input and contributions, to Aybike Onur for technical assistance, to Rebecca Foote and Kristen Gilbert for administrative support, to David Mangelsdorf (UT Southwestern) for TK-LXRE-X3luc reporter and Nr1h3−/− mice, and Alan Edgar (Fournier, France) for the ACAT inhibitor. CONTRIBUTIONS E.E. developed the hypothesis, designed and performed the bulk of the experiments, analyzed all data and wrote the manuscript; V.R.B. contributed to the in vivo studies; J.R.M. contributed to the in vitro studies and conducted data analysis, K.N.C. and M.S. contributed to the in vitro studies; L.M., M.M.W, S.M.W. contributed to the analysis of lipidomic data, S.F. and M.F.L. contributed to the analysis of in vivo data and assisted with writing; and G.S.H. developed the hypothesis, designed and analyzed all data, wrote the manuscript and supervised the project and the peer review process.

he in vitro studies; L.M., M.M.W, S.M.W. contributed to the analysis of lipidomic data, S.F. and M.F.L. contributed to the analysis of in vivo data and assisted with writing; and G.S.H. developed the hypothesis, designed and analyzed all data, wrote the manuscript and supervised the project and the peer review process. Figure 1 PBA treatment protects against macrophage ER stress and reduces vascular disease progression (a–d) We induced ER stress in WT macrophages by 500 µM PA or 300 nM thapsigargin (Thaps) in the presence or absence of 3 mM PBA. P-PERK and P-eIF2–α were examined by Western blotting (a), mRNA levels of Ddit3 (b) and sXBP-1 were examined by qRT-PCR (c), and apoptosis were examined by TUNEL assays (d) (data represent mean±SEM; * p<0.05). (e–g) Atherosclerotic lesion area in the aortic sinus from mice treated with either control or PBA for 2 weeks (n≥13) is reported (data represents mean±SD; * indicates p<0.05) (e) and the serial sections were stained with antibodies against P-PERK (f), P-eIF2–α (f), ATF3 (g) and MOMA-2 (f,g) (Red indicates positive staining with antibody. Scale bars represent 200 µm). (g) Relative fluorescent intensity was calculated for antibody staining corresponding to ATF3 and P- eIF2–α expression (n≥3) in the macrophage-dense areas (data represent mean±SD; * indicates p<0.05). (h) Percent of apoptotic cells (TUNEL positive) in the aortic sinus area are shown for PBS (control) or 100 mg kg−1 PBA (PBA-100) treated mice (Arrows point to apoptotic cells. Scale bars represent 200 µm. *indicates p<0.05).

α expression (n≥3) in the macrophage-dense areas (data represent mean±SD; * indicates p<0.05). (h) Percent of apoptotic cells (TUNEL positive) in the aortic sinus area are shown for PBS (control) or 100 mg kg−1 PBA (PBA-100) treated mice (Arrows point to apoptotic cells. Scale bars represent 200 µm. *indicates p<0.05). Figure 2 Requirement for aP2 in lipid-induced ER stress and apoptosis (a–b) Serial sections from aortic sinus of PBS or PBA treated ApoE−/− mice were stained for antibodies against MOMA-2 or aP2 (Arrows point to red staining for these antibodies in macrophage-dense areas. Scale bar represents 200 µm) (b) Relative fluorescence intensity for aP2 expression (green) in the macrophage-dense areas (red) of the lesions was calculated (Scale bar represents 100 µm. Data represent mean±SD; * indicates p<0.05 and n≥3). (c) aP2 expression was determined by Western blotting in PA (500 µM) treated macrophages in the presence or absence of 3 µM PBA or 100 µg mL−1 cycloheximide (CHX). (d) P-PERK, P-eIF2–α was examined by Western blotting and JNK activity (P-cJun) by an in vitro kinase assay from PA (500 µM) treated macorphages. (e) sXBP-1 was examined by qRT-PCR in (500 µM) PA- or (2 µg/mL) tunicamycin (Tunic) treated (12 h) macrophages. (f) Cleaved PARP and tubulin expression were examined by Western blotting in PA-treated (24 h) macrophages (relative band intensities were quantified and data represent mean±SEM; *p<0.05). (g) P-PERK, P-eIF2–α were examined by Western blotting, JNK activity (P-c-Jun) by a kinase assay, and sXBP-1 by RT-PCR from macrophages treated with 10 µM ACAT inhibitor (Ai) and 100 µg/mL Ac-LDL (24 hours) or 300 nM thapsgiargin (Thaps) (The ratio of relative intensities corresponding to spliced (s-XBP1) and unspliced (u-XBP1) were calculated).

xamined by Western blotting, JNK activity (P-c-Jun) by a kinase assay, and sXBP-1 by RT-PCR from macrophages treated with 10 µM ACAT inhibitor (Ai) and 100 µg/mL Ac-LDL (24 hours) or 300 nM thapsgiargin (Thaps) (The ratio of relative intensities corresponding to spliced (s-XBP1) and unspliced (u-XBP1) were calculated). Figure 3 aP2 deficiency protects from hypercholesteremia induced macrophage ER stress and apoptosis in atherosclerotic lesions (a–b) Immunohistochemical staining with antibodies against MOMA-2, P-PERK, P-eIF2–α (a,b), CHOP (a), and ATF3 (b) were performed in atherosclerotic lesions from the proximal aorta of ApoE−/− and aP2−/−ApoE−/− mice fed a Western diet for 16 weeks (Arrows point to ATF3 and P-eIF2–α (green), expressed in the MOMA-2 positive (red) areas of the lesions. Scale bars in (a) represent 50 µm and in (b) represent 100 µm). (c) Relative fluorescent intensity was calculated for stainings corresponding to ATF3 and P-eIF2–α in the macrophage-dense areas (data represent mean±SD; * indicates p<0.05, n≥3). (d–e) Apoptotic macrophages in the lesions from ApoE−/− and aP2−/−ApoE−/− mice were determined by TUNEL assay (Arrows point to apoptotic cells. Scale bars represent 100 µm. * indicates p<0.05). (f–g) Macrophage lines were stressed with PA (500 µM) in the presence of vehicle (−) or varying does of the aP2-i (0.1–50 µM). P-PERK and P-eIF2–α was examined by Western blotting (f) and sXBP-1 and Ddit3 mRNA were analyzed by qRT-PCR from macrophages treated with 25 µM of aP2-i (g). (h) Double immunofluorescent staining was performed using antibodies against MOMA-2 and ATF-3 in the atherosclerotic lesions from ApoE−/− mice treated with vehicle or aP2-i (15 mg kg−1 for 14 weeks) (Arrows indicate staining for ATF3 (green) in MOMA-2 positive areas (red). Scale bars represent 100µm).

µM of aP2-i (g). (h) Double immunofluorescent staining was performed using antibodies against MOMA-2 and ATF-3 in the atherosclerotic lesions from ApoE−/− mice treated with vehicle or aP2-i (15 mg kg−1 for 14 weeks) (Arrows indicate staining for ATF3 (green) in MOMA-2 positive areas (red). Scale bars represent 100µm). Figure 4 Regulation of macrophage lipid composition by aP2 (a) Lipid class composition analysis for TG, PL and FA was performed. The F statistics from one-way ANOVA are displayed as red diamonds over the distribution of F statistics from permuted data. The black line is the 95th percentile of the F statistics over 1000 permutations. The higher the value of the F-statistics from ANOVA, the more different the groups are. The heat map displays the observed data, centered to the mean of the WT genotype and scaled by the standard deviation of all observations. (b) The mean concentration of C16:1n7 and C18:1n7 was determined for each lipid class in the various macrophage lines. (c–d) Total lipid composition: Percent total lipids (c) and bar plots of the mean concentration of lipids (d) in the macrophage lines.

genotype and scaled by the standard deviation of all observations. (b) The mean concentration of C16:1n7 and C18:1n7 was determined for each lipid class in the various macrophage lines. (c–d) Total lipid composition: Percent total lipids (c) and bar plots of the mean concentration of lipids (d) in the macrophage lines. Figure 5 A central role for SCD and C16:1n7 in aP2 mediated lipotoxic signaling (A) A summary of the lipid changes that occur as a result of aP2 deficiency in macrophages (LCE and ELOVL; fatty acid elongase for long chain fatty acid). (b)SCD-1 mRNA levels were examined by qRT-PCR in primary peritoneal macrophages at the base line or (c) after treatment of animals aP2-i for 6 weeks (n=6) (data represent mean±SEM; * indicates p<0.05). (d–e) ER stress was induced in macrophages by PA (300 µM) or tunicamycin (2 mM) treatment for 3 hours. Cells were pretreated with PAO (300 µM) PAO, where indicated. P-PERK, P-eIF2–α and cleaved PARP were examined by Western blotting (d) and Ddit3 and sXBP-1 mRNA were examined by qRT-PCR (E). (F) From macrophage lines treated with SCD-1 siRNA (50–100 nM) or scrambled (−) siRNA, SCD activity (upper panel) was examined by an enzymatic assay and SCD protein expression (lower panel) was examined by Western blotting (G) P-PERK, P-eIF2–α and cleaved PARP were examined by Western blotting from lysates of aP2−/− macrophages treated with negative (−) siRNA or SCD-1 specific siRNA (100nM) and treated with or without PA (500 µM) (data represent mean±; SEM; * indicates p<0.05).

in expression (lower panel) was examined by Western blotting (G) P-PERK, P-eIF2–α and cleaved PARP were examined by Western blotting from lysates of aP2−/− macrophages treated with negative (−) siRNA or SCD-1 specific siRNA (100nM) and treated with or without PA (500 µM) (data represent mean±; SEM; * indicates p<0.05). Figure 6 Linking toxic lipids to ER stress and atherosclerosis through aP2-LXR–α crosstalk (a) Alignment of LXR responsive element (LXRE) on Fasn and SCD-1 promoters. (b) LXR–driven transcriptional activity was determined from various macrophage lines upon stimulation with a synthetic T0901317 (10 µM) or endogenous 25-hydroxycholesterol (10 µM) LXR ligand (luciferase activity is reported after normalizing to transfection efficiency). (c) Relative LXR–α and LXR–β protein levels in various macrophage lines were examined by Western blotting (d) Relative SCD-1 and Fasn mRNA levels from aP2−/− macrophage treated with a scrambled (−) or Nr1h3-specific siRNA were examined by qRT-PCR. (e) Lysates from various macrophages treated with scrambled (−) siRNA or a specific siRNA against SCD-1 or Nr1h3 and stressed with or without PA (500 µM) were examined for P-PERK and P-eIF2–α by Western blotting. (f) Lysates from peritoneal macrophages from aP2−/−Nr1h3−/− or WT mice stressed with or without PA (500 µM) were examined for P-PERK, P-eIF2–α and cleaved PARP by Western blotting. (g) A cellular lipotoxicity model: Toxic levels of lipids are sensed by the ER through an aP2-dependent pathway and induce the UPR and lead to macrophage apoptosis. The absence of aP2 serves to reactivate macrophage de novo lipogenesis pathways and promotes desaturation, particularly through LXRα-mediated activation of SCD-1, leading to increased production of bioactive lipids and resistance to ER stress. Our findings indicate that alleviation of macrophage ER stress, either through aP2 inhibition or enhancing ER function, is protective against atherosclerosis.

Introduction In lissencephaly patients, mutation of two genes, LIS1 or DCX, account for the majority of classical lissencephaly (ILS)1–3. Consistent with an important role for this protein in neuronal migration, mice with decreased Lis1 exhibited disorganization of cortical layers, hippocampus and olfactory bulb in dose dependent fashion, and are a good model for the human disorder4. LIS1 was first identified as a subunit of the brain platelet-activating factor acetylhydrolase (PAFAH1B1)5, but extensive studies in a variety of model organisms from bread molds to mammals led to the conclusion that LIS1 is essential for the proper regulation and localization of cytoplasmic dynein6–8. Many studies have further delineated the role of LIS1 on neuronal morphogenesis and the maintenance of cell integrity. However, no studies have addressed potential therapeutic approaches for lissencephaly, a devastating human disorder.

S1 is essential for the proper regulation and localization of cytoplasmic dynein6–8. Many studies have further delineated the role of LIS1 on neuronal morphogenesis and the maintenance of cell integrity. However, no studies have addressed potential therapeutic approaches for lissencephaly, a devastating human disorder. We previously demonstrated that LIS1 is required for anterograde transport of cytoplasmic dynein in a kinesin dependent fashion8. Interestingly, we found that a substantial fraction of LIS1 is degraded at the periphery (cortex) of the cell. We probed for molecules that were involved in LIS1 degradation using inhibitors, and found that calpain inhibitors efficiently prevented the degradation of LIS1, suggesting that LIS1 is degraded by calpain dependent proteolysis. Here, we report that inhibition of calpain rescued various phenotypes that were observed in cells and in the whole animal using our Lis1-deficient mice that are a good model of this disorder. Our work provides a proof-of-principle for the treatment of lissencephaly die to haploinsufficiency of LIS1, and also suggests a unique therapeutic strategy for diseases associated with haploinsufficiency.

re observed in cells and in the whole animal using our Lis1-deficient mice that are a good model of this disorder. Our work provides a proof-of-principle for the treatment of lissencephaly die to haploinsufficiency of LIS1, and also suggests a unique therapeutic strategy for diseases associated with haploinsufficiency. Results LIS1 is degraded by calpain and inhibition of calpain rescued defective distribution of cytoplasmic dynein and membranous components in the cell First, we examined the response of LIS1 protein levels to the inhibition of calpain by ALLN using mouse embryonic fibroblast (MEF) cells4. LIS1 protein gradually increased in the Lis1+/− MEF cells, and reached a plateau two hours after the start of ALLN treatment (data not shown). After two hours of administration of ALLN, LIS1 protein was increased from 0.5 (where 1.0 is equivalent to wild type levels of LIS1) to 1.1 (Supplementary Fig. 1, Fig. 1a). We also examined the effect of E64d, another calpain inhibitor and obtained similar results (Fig. 1a: from 0.5 to 0.9). The amount of cytoplasmic dynein intermediate chain 1(DIC1) was also increased, suggesting that the protein stability of cytoplasmic dynein was reduced in the Lis1 mutated cells (Fig. 1a), which may be attributed to the direct prevention of degradation of cytoplasmic dynein or the indirect stabilization through normalization of its distribution. We also examined the effect of ALLN or E64d treatment on dorsal root ganglia (DRG) neurons, and obtained similar results in the Lis1+/− DRG neurons by ALLN (LIS1: from 0.4 to 0.7, DIC1: 0.8 to 1.6), and by E64d (LIS1: from 0.4 to 0.9, DIC1: 0.8 to 1.5) (Fig. 1b). In contrast, there was no significant effect of calpain inhibitors on LIS1 or DIC1 in Lis1+/+ MEF cells or DRG neurons (Supplementary Fig. 2a, b). We next determined whether preventing the degradation of LIS1 rescued the aberrant distribution of LIS1 and cytoplasmic dynein within the Lis1+/− MEF cells. Treatment of Lis1+/− MEF cells by ALLN or E64d clearly improved the reduction of centrosomal concentration of LIS1 after 2 hours of the treatment (Supplementary Fig. 2c). In addition, the abnormal accumulation of cytoplasmic dynein around the centrosome was rescued by ALLN or E64d treatment (Supplementary Fig. 2d). These improvements were also observed in the Lis1+/− DRG neurons (Supplementary Fig. 2e, f), whereas there was no significant effect in Lis1+/+ DRG neurons (Supplementary Fig. 2e, f).

on, the abnormal accumulation of cytoplasmic dynein around the centrosome was rescued by ALLN or E64d treatment (Supplementary Fig. 2d). These improvements were also observed in the Lis1+/− DRG neurons (Supplementary Fig. 2e, f), whereas there was no significant effect in Lis1+/+ DRG neurons (Supplementary Fig. 2e, f). We next addressed whether ALLN or E64d was able to rescue the aberrant distribution of cell components transported by cytoplasmic dynein in Lis1+/− MEF cells. Mitochondria displayed dispersed distribution in Lis1+/+ MEF cells. By contrast, they clustered in the perinuclear region of Lis1+/− MEF cells (Supplementary Fig. 2g). This aberrant clustering was rescued by ALLN or E64d treatment (Supplementary Fig. 2g). Immunofluorescence demonstrated that β-COP-positive vesicles displayed a predominantly juxtanuclear staining pattern in Lis1+/+ MEF cells (Supplementary Fig. 2h). In Lis1+/− MEF cells, this juxtanuclear clustering was disrupted, and β-COP displayed punctuate clustering9 (Supplementary Fig. 2h). This aberrant distribution of β-COP positive vesicles in Lis1+/− MEF cells was also rescued by ALLN or E64d treatment (Supplementary Fig. 2h). These effects of calpain inhibitors were not observed in Lis1+/+ MEF cells (Supplementary Fig. 2i–l). These observations suggest that inhibition of calpains improves the functional defects of cytoplasmic dynein in Lis1+/− MEF cells.

n Lis1+/− MEF cells was also rescued by ALLN or E64d treatment (Supplementary Fig. 2h). These effects of calpain inhibitors were not observed in Lis1+/+ MEF cells (Supplementary Fig. 2i–l). These observations suggest that inhibition of calpains improves the functional defects of cytoplasmic dynein in Lis1+/− MEF cells. Inhibition of LIS1 degradation rescued defective migration in Lis1+/− neurons To define the effect of calpain inhibitors on mammalian neuronal migration, we used mouse cerebellar granule neurons in an in vitro migration assay combined with ALLN or E64d treatment9–12. As heterozygous loss of LIS1 leads to lissencephaly in humans, graded reduction of Lis1 results in increased severity of migration defects in mice4. We first examined whether inhibition of calpain might affect neuronal migration in wild type cells, and found that calpain inhibition slightly facilitated neuronal migration (Fig. 2a, b, c). We next confirmed that Lis1+/− cerebellar granule neurons displayed a leftward shift of the bin distribution of migration distance, and the mean distance decreased by approximately half from the wild type level (Fig. 2a, b, c). We then tested whether ALLN or E64d treatment was sufficient to rescue the migration defect in Lis1+/− cerebellar granule neurons. Lis1+/− cerebellar granule neurons in the presence of ALLN or E64d displayed a shift in migration distance bins back toward the right, with a similar distribution to that of wild type neurons (Fig. 2a, b). Quantitation of mean migration distance of cerebellar granule neurons demonstrated rescue of defective migration in Lis1+/− cerebellar granule neurons by ALLN or E64d treatment (Fig. 2c). Specifically, inhibition of LIS1 degradation by ALLN or E64d restored migration to 83.1% (ALLN) or 84.4% (E64d) of wild type levels in Lis1+/− cerebellar granule neurons. Calpain inhibitors also slightly but significantly facilitated neuronal migration in wild type cells. It was previously shown that overexpression of LIS1 facilitates neuronal migration11, so it is possible that calpain inhibition might stabilize LIS1 locally, and/or may function by other mechanisms, including modulation of focal adhesion kinase (FAK) and/or Cdk5/p35.

y facilitated neuronal migration in wild type cells. It was previously shown that overexpression of LIS1 facilitates neuronal migration11, so it is possible that calpain inhibition might stabilize LIS1 locally, and/or may function by other mechanisms, including modulation of focal adhesion kinase (FAK) and/or Cdk5/p35. FAK that is a tyrosine kinase localized to focal adhesions has been shown to be critical for cell migration13,14. FAK levels are regulated by calpain-dependent cleavage15–19. FAK is also a physiological substrate of Cdk5 during neocortical development20–22. We therefore examined whether inhibition of calpain might modify distribution and/or expression of focal adhesion complex by migration assay using granular neurons, and did not observe obvious differences of distribution and expression of FAK and vinculin by inhibition of calpains (Supplementary Fig. 3a–e). While we cannot completely exclude the possibility that inhibition of calpain might modify signal transduction from focal adhesion, tour findings do not support this possibility.

d did not observe obvious differences of distribution and expression of FAK and vinculin by inhibition of calpains (Supplementary Fig. 3a–e). While we cannot completely exclude the possibility that inhibition of calpain might modify signal transduction from focal adhesion, tour findings do not support this possibility. Knockdown of calpain by siRNA restored LIS1 protein resulting in rescue of aberrant distribution of cytoplasmic dynein and membranous components in the cell ALLN and E64d are broad cysteine protease inhibitors, and can inhibit other cysteine proteases other than calpain, including cathepsins and papain23,24. To address whether calpain is a major enzyme for the degradation of LIS1 protein, we investigated the effect siRNA against calpain. The ubiquitous calpains, μ-calpain (calpain I) and m-calpain (calpain II), are heterodimers consisting of large catalytic subunits encoded by the Capn1 and Capn2 genes, respectively, and the small regulatory subunit encoded by Capns123,24. Inactivation of μ-calpain and m-calpain simultaneously by siRNA was technically challenging. Therefore, we knocked down Capns1 by siRNA, resulting in depletion of the both of μ-calpain and m-calpain25 (Supplementary Fig. 4a, b). After transfection of siRNA against Capns1, μ-calpain and m-calpain were gradually decreased over 48 hrs, and were almost undetectable after 96 hrs (Supplementary Fig. 4a, b). This decrease in calpain was associated with an increase of LIS1 and DIC1 levels in Lis1+/− MEFs (Fig. 3), consistent with the effects of calpain inhibitors shown above. The subcellular distribution abnormalities of LIS1 and DIC1 found in Lis1+/− cells was rescued by depletion of μ-calpain and m-calpain (Supplementary Fig. 4c, d). The aberrant distributions of β-COP and mitochondria were also rescued by depletion of μ-calpain and m-calpain (Supplementary Fig. 4e, f). These observations suggest that LIS1 is specifically degraded by calpain, and selective inhibition of calpain is sufficient to increase LIS1 levels for improvement of the cellular phenotypes.

nt distributions of β-COP and mitochondria were also rescued by depletion of μ-calpain and m-calpain (Supplementary Fig. 4e, f). These observations suggest that LIS1 is specifically degraded by calpain, and selective inhibition of calpain is sufficient to increase LIS1 levels for improvement of the cellular phenotypes. Intra-peritonial administration of ALLN partially rescued apoptotic cell death and defective neuronal migration These observations prompted us to examine whether the administration of ALLN to pregnant Lis1+/− dams rescued defective neuronal migration in vivo4. We first examined the effect of LIS1 protection from degradation by intraperitoneal injection of ALLN. We injected ALLN (38.3 µg/g) into E12.5 pregnant Lis1+/− dams, and examined LIS1 in embryonic brains by Western blotting at various times after injections. We found that LIS1 increased from one hour after injection and reached a plateau 6–12 hrs later (Supplementary Fig. 5a). The effect of ALLN decreased thereafter, and returned to the original level after 24 hrs, suggesting that ALLN is relatively short acting. Thus, we performed intraperitoneal injection of pregnant dams between E9.5–E17.5 every day, and observed cell survival and neuronal migration in the brains of in utero treated offspring. We previously reported a mild reduction of the density of cells in the neocortex of the Lis1+/− mice due to apoptotic cell death in the ventricular zone9. Lis1+/− mice displayed a reduction of brain weight compared to the wild type control pups, an effect that was partially rescued by administration of ALLN (Fig. 4a). In Lis1+/− mice, apoptotic cell death was increased at E15.59, whereas it was clearly suppressed by administration of ALLN (Fig. 4b). LIS1 is essential for neuroepithelial stem cell proliferation4,26. To trace proliferation of stem cells, we performed BrdU pulse labeling of E13.5 embyros, and found that BrdU incorporation was not significantly different among the four groups (Supplementary Fig. 5b). Although calpain inhibitors might facilitate proliferation of neuroepithelial stem cell in Lis1+/− mice, the effects on heterozygotes may be too small to measure. Thus, we believe that suppression of apoptotic cell death more likely contributes to the rescue of brain size by ALLN.

e four groups (Supplementary Fig. 5b). Although calpain inhibitors might facilitate proliferation of neuroepithelial stem cell in Lis1+/− mice, the effects on heterozygotes may be too small to measure. Thus, we believe that suppression of apoptotic cell death more likely contributes to the rescue of brain size by ALLN. We next examined the effect of ALLN on the cortical and hippocampal layering of neurons. Lis1+/− mice exhibited laminar splitting and discontinuities of pyramidal cells in the CA3 and CA2 region4 (Fig. 4c). After administration of ALLN in utero, Lis1+/− mice also displayed splitting and diffuse packing of pyramidal cells, but these defects were markedly improved (Fig. 4c, and Supplementary Fig. 5c). To examine cortical lamination, we analyzed Brn-1 immunoreactivity, to label interneurons of layer 2 and 327. In Lis1+/− mice, Brn-1 positive cells exhibited a broader distribution compared to Lis1+/+ mice. Administration of ALLN resulted in a more tightly packed lamination in Lis1+/− mice (Fig. 4d). To confirm the morphological improvement by daily ALLN treatment in utero, we performed quantitative BrdU birthdating analysis. In Lis1+/− mice, the distribution of labeled cells was shifted downward toward the ventricular zone in the cortex, and BrdU-labeling was more diffusely localized4 (Fig. 4e). The migration defects associated with the disruption of Lis1 were partially rescued by ALLN treatment (Fig. 4e). In migration assay using granular neurons in vitro, calpain inhibitors slightly but significantly facilitated neuronal migration in wild type, whereas in utero treament with calpain did not have a significant effect on the wild type embryo. Examination of neuronal migration in vivo in the embryo was performed 13–15 days after the start of injection of a calpain inhibitor, whereas the in vitro migration assay of granular neurons was performed 16 hr after treatment of calpain inhibitors. Over several days of migration in vivo, slight differences in speeds of neuronal migration may not result in detectable phenotypic difference, since once migrating neurons reach the proper position in the brain, they receive stop signals for migration28. Examination of neuronal migration in vivo in the embryo may not be sensitive enough for the detection of this statistically significant but small difference (73.9 µm/69.4 µm: 6.5%). In contrast, in Lis1 heterozygous granular neurons, the difference between the absence or presence of ALLN is significant (57.7 µm/37.4 µm: 54.3%).

on of neuronal migration in vivo in the embryo may not be sensitive enough for the detection of this statistically significant but small difference (73.9 µm/69.4 µm: 6.5%). In contrast, in Lis1 heterozygous granular neurons, the difference between the absence or presence of ALLN is significant (57.7 µm/37.4 µm: 54.3%). To examine the effect of calpain knockdown in neuronal migration in vivo, we introduced shRNA against calpain small subunit1 into the E14 mouse neocortex by in utero electroporation. ShRNA against calpain small subunit depleted the both of calpain 1 and calpain 2 as with siRNA against calpain small subunit (Supplementary Fig. 6). In P4 Lis1+/+ mice that had been transfected with control GFP, many cells labeled in the ventricular zone migrated out towards the pial surface (Fig. 4f). By contrast, neurons in P4 Lis1+/− displayed defective placement, consistent with decreased motility (Fig. 4f). Inactivation of μ-calpain and m-calpain partially rescued defective neuronal migration, resulting in larger fraction of neurons that reached more superficial layers. These data are consistent with our histological examinations, and support that calpain inhibition in Lis1+/− embryos is effective in facilitating neuronal migration.

n of μ-calpain and m-calpain partially rescued defective neuronal migration, resulting in larger fraction of neurons that reached more superficial layers. These data are consistent with our histological examinations, and support that calpain inhibition in Lis1+/− embryos is effective in facilitating neuronal migration. Intra-peritonial administration of ALLN partially rescued impaired motor behavior Lis1+/− mice displayed abnormal behavior and impaired in the spatial learning, including hindpaw clutching responses, a rotarod test and the Morris water maze task29. Therefore, we examined whether administration of ALLN in utero during embryonic development of Lis1+/− mice (ALLN-plus group) is effective in improving motor behavior compared to untreated Lis1+/− mice (ALLN-minus group). Standard measurements were not significantly different between the ALLN-minus and ALLN-plus groups, including body weight and body temperature (Supplementary Table 1). Lis1+/+ mice, ALLN-minus group and the ALLN-plus group displayed similar grip strength. Importantly, despite similar grip strength between the ALLN-minus group and the ALLN-plus group, the ALLN-plus group displayed longer time to hang on the wire before falling (Supplementary Fig. 7, Fig. 5a), suggesting that motor coordination in the ALLN-plus group was improved. Next, we examined rotarod performance. The time that Lis1+/+ mice and ALLN-plus group maintained their balance on the top of the rotating rod increased significantly over the six trials. However, the latency to fall for the ALLN-minus group was significantly less than that recorded for wild type mice as we previously reported29 (Fig. 5b). Importantly, impaired performance on the rotarod test in the ALLN-minus group was significantly improved in the ALLN-plus group (Fig. 5b). Finally, we analyzed gait dynamics to address quantitative neurological traits of Lis1+/− mice. Stride length variability of forelimbs was increased in the ALLN-minus group (Fig. 5c)30. In the ALLN-plus group, this variability was improved (Fig. 5c). Normal paw angles are ~5 degree in fore paws and ~10 degree in hind paws. More open angles of the hind paws are associated with ataxia, spinal cord injury and demeyelinating diseases31. The ALLN-minus group displayed more open angles of the fore paws (Fig. 5d), an abnormality that returned to normal angles in the ALLN-plus group. Interestingly, these aberrant gait parameters were more conspicuous in the fore paws, and less remarkable in the hind paws.

spinal cord injury and demeyelinating diseases31. The ALLN-minus group displayed more open angles of the fore paws (Fig. 5d), an abnormality that returned to normal angles in the ALLN-plus group. Interestingly, these aberrant gait parameters were more conspicuous in the fore paws, and less remarkable in the hind paws. Restoration of normal parameters in gait analysis also supported the functional improvement of Lis1+/− mice after injection of ALLN.

spinal cord injury and demeyelinating diseases31. The ALLN-minus group displayed more open angles of the fore paws (Fig. 5d), an abnormality that returned to normal angles in the ALLN-plus group. Interestingly, these aberrant gait parameters were more conspicuous in the fore paws, and less remarkable in the hind paws. Restoration of normal parameters in gait analysis also supported the functional improvement of Lis1+/− mice after injection of ALLN. Discussion In this report, we have presented proof-of-principle for a novel and potentially effective therapeutic strategy for human lissencephaly, using our Lis1-deficient mice that are a good model of this disorder. Therapeutic strategies for lissencephaly are a daunting consideration for several reasons. First, given the nature of lissencephaly, one would have to treat all neurons throughout development. Second, LIS1 mutations in humans are de novo, so that detection of the disorder at an early enough time point to allow effective therapy is difficult. In spite of these difficulties, there are some advantages to considering the treatment of lissencephaly that results from LIS1 haploinsufficiency. First, LIS1 protein is present and can potentially be manipulated, since individuals display heterozygous, not complete loss of LIS1. Second, there are dosage dependent effects of LIS1, so any augmentation of LIS1 protein levels will likely have a beneficial effect. Third, a great deal is known about the pathogenesis and mechanism of action of LIS1 and its pathway, so the effects of any therapeutic modality can be assessed directly with quantitative measures in vivo and in vitro.

dent effects of LIS1, so any augmentation of LIS1 protein levels will likely have a beneficial effect. Third, a great deal is known about the pathogenesis and mechanism of action of LIS1 and its pathway, so the effects of any therapeutic modality can be assessed directly with quantitative measures in vivo and in vitro. We based this therapeutic strategy on our recent observations that LIS1 is degraded after anterograde transport to the nerve terminals in a calpain dependent fashion. Inhibition of calpain resulted in the augmentation of LIS1 protein, which led to the rescue of aberrant distribution of cytoplasmic dynein from Lis1-deficient mice. We further demonstrated that inhibition of calpain rescued neuronal migration from granule neurons from the Lis1 mutants. Most importantly, we demonstrated that daily ALLN administration in utero was partially effective in improving the defective migration phenotypes in vivo in the Lis1-deficient mice, which was associated with improvement in motor function.

of calpain rescued neuronal migration from granule neurons from the Lis1 mutants. Most importantly, we demonstrated that daily ALLN administration in utero was partially effective in improving the defective migration phenotypes in vivo in the Lis1-deficient mice, which was associated with improvement in motor function. Recently, it was shown that increased LIS1 expression affects human and mouse brain development32. In our case, inhibition of calpain activity results in normalization close to the wild type levels rather than accumulation of LIS1 in excess. We believe that the restoration of more normal LIS1 levels was one reason that calpain inhibition resulted in phenotypic improvement of Lis1+/− mice. We cannot rule out the possibility that other effects of calpain may also play some roles in the observed phenotypic rescue, including suppression of spectrin/neurofilaments/MT breakdown, cleavage of p35, a Cdk5 activator important for neuronal migration33–35, prevention of degradation of other proteins included in the Lis1/Ndel1/Dynein complex and/or acetylated tubulin or FAK complex18,20, which will be the subjects of further investigation.

ncluding suppression of spectrin/neurofilaments/MT breakdown, cleavage of p35, a Cdk5 activator important for neuronal migration33–35, prevention of degradation of other proteins included in the Lis1/Ndel1/Dynein complex and/or acetylated tubulin or FAK complex18,20, which will be the subjects of further investigation. Several problems remain and must be overcome if this promising avenue of therapy can be eventually tested in human lissencephaly, including further refinement of the use of calpain inhibitors for the effective inhibition of LIS1 degradation as well as the safe and effective delivery of such drugs for clinically effective treatment of human lissencephaly. In spite of the challenges, our work provides a potential avenue to consider therapeutic strategies for severe, early brain developmental defects such as lissencephaly due to LIS1 mutations, as well as any other disorder that results from haploinsufficiency.

gs for clinically effective treatment of human lissencephaly. In spite of the challenges, our work provides a potential avenue to consider therapeutic strategies for severe, early brain developmental defects such as lissencephaly due to LIS1 mutations, as well as any other disorder that results from haploinsufficiency. Methods DRG preparation, culture, fluorescence recovery measurement after photobleaching (FRAP) and reaggregate neuronal migration assay DRGs from postnatal mice were dissociated using a previously described method37. Cerebellar granule cells were isolated as described previously9,10,12 and cultured at 106 cells/ml for 12 hrs, resulting in uniform-sized reaggregates (100–150 µm in diameter), which were then transferred to poly-L-lysine– (Sigma-Aldrich) and laminin- (Sigma-Aldrich) treated slides and incubated for 8 hrs. Then, 10 µM ALLN (Calbiochem), 20 µM E64d (Calbiochem) or control DMSO was added and the cultures were further incubated for 16 h. Images were obtained using a 20x objective lens and images were analyzed using a confocal microscope (TCS-SP5, Leica).

) and laminin- (Sigma-Aldrich) treated slides and incubated for 8 hrs. Then, 10 µM ALLN (Calbiochem), 20 µM E64d (Calbiochem) or control DMSO was added and the cultures were further incubated for 16 h. Images were obtained using a 20x objective lens and images were analyzed using a confocal microscope (TCS-SP5, Leica). BrdU birthdating study For bromodeoxyuridine (BrdU) experiments, pregnant dams (E15.5) were injected with BrdU (50 µg/g, i.p.). Subsequently, the distribution of BrdU-positive cells was determined at P5. For pulse labeling to trace proliferation of neuroepithelial stem cell, pregnant dams (E13.5) were injected with BrdU (150 µg/g, i.p.). Subsequently, the distribution of BrdU-positive cells was determined one hour after the injection. The incorporation of BrdU in cells was detected with a mouse anti-BrdU monoclonal primary antibody (Roche) followed by an alkaline phosphatase-conjugated secondary antibody (Boehringer Mannheim). We analyzed three independent mice for each genotype.

on of BrdU-positive cells was determined one hour after the injection. The incorporation of BrdU in cells was detected with a mouse anti-BrdU monoclonal primary antibody (Roche) followed by an alkaline phosphatase-conjugated secondary antibody (Boehringer Mannheim). We analyzed three independent mice for each genotype. Histological examination and immunohistochemistry After perfusion with 4% PFA fixative, tissues from wild type and various mutant mice were subsequently embedded in paraffin and sectioned at 5 µm thickness. After deparaffination, endogenous peroxidase activity was blocked by incubating the sections in 1.5% peroxide in methanol for 20 min. The sections were then boiled in 0.01 M citrate buffer, pH 6.0, for 20 min and cooled slowly. Before staining, the sections were blocked with rodent block (LabVision) for 60 min. The sections were washed in PBS and incubated with an anti-Brn-1 antibody (Santa Cruz).

the sections in 1.5% peroxide in methanol for 20 min. The sections were then boiled in 0.01 M citrate buffer, pH 6.0, for 20 min and cooled slowly. Before staining, the sections were blocked with rodent block (LabVision) for 60 min. The sections were washed in PBS and incubated with an anti-Brn-1 antibody (Santa Cruz). Cell culture and immunocytochemistry Establishment of mouse embryonic fibroblast (MEF) cells was performed as previously described9,12. MEF cells were grown in D-MEM supplemented with 10% FBS. To inhibit calpain, MEF cells were incubated with 10 µM ALLN (Calbiochem), 20 µM E64d (Calbiochem) or control DMSO for 2 hrs. Cells were fixed in 4% PFA in PBS followed by permealization with 0.2% Triton X-100 in PBS. Coverslips were blocked for one hour with Block Ace (Yukijirushi) in PBS supplemented with 5% BSA, and were incubated for one hour in primary antibody, washed, and incubated for 1 hr using Alexa 546-conjugated secondary antibodies (Molecular Probes). Primary antibodies were an anti-β COP antibody (Sigma) and an anti-DIC1 antibody (Chemicon). Mitochondria were stained by MitoTracker Green FM (Molecular Probes). The anti-calpain antibody (1D10A7) was provided from Seiichi Kawashima, which recognized conventional calpain36.

ted secondary antibodies (Molecular Probes). Primary antibodies were an anti-β COP antibody (Sigma) and an anti-DIC1 antibody (Chemicon). Mitochondria were stained by MitoTracker Green FM (Molecular Probes). The anti-calpain antibody (1D10A7) was provided from Seiichi Kawashima, which recognized conventional calpain36. SiRNA, shRNA and transfection Used siRNA and shRNA to target the mouse Capns1 (calpain small subunit 1) was purchased from Sigma (MISSION® siRNA: SASI_Mm01_00127701, and MISSION™ TRC shRNA: TRCN0000087168). siRNA was transfected with Lipofectamine RNAi MAX reagents (Invitrogen, Carlsbad, CA). shRNA was in utero electroporation-mediated gene transfer method38,39. shRNA plasmid and pCAGGS-GFP control plasmid were dissolved in HBS (21 mM HEPES, pH 7.0, 137 mM NaCl, 5 mM KCl, 0.7 mM Na2HPO4, 1 mg/l glucose) at a final concentration of 10mg/ml together with Fast Green (final concentration 0.01%). For cotransfection, a molar ratio of 1 (pCAGGS-GFP) to 3–6 (shRNA) was used.

hod38,39. shRNA plasmid and pCAGGS-GFP control plasmid were dissolved in HBS (21 mM HEPES, pH 7.0, 137 mM NaCl, 5 mM KCl, 0.7 mM Na2HPO4, 1 mg/l glucose) at a final concentration of 10mg/ml together with Fast Green (final concentration 0.01%). For cotransfection, a molar ratio of 1 (pCAGGS-GFP) to 3–6 (shRNA) was used. Behavior analysis Forty one (25 males and 16 females) wild type, fourteen (5 males and 9 females) Lis1+/− mice and fifteen (12 males and 3 females) Lis1+/− mice that were treated in utero with ALLN were used for behavioral experiments. Lis1+/− mice had a single Lis1 mutant allele. In this study mice were from a mixed genetic background (129SvEv×FVB). All animal experiments were carried out under protocols approved by Kyoto University. In this screen, several physical features and several motor responses of the mice were recorded including body weight and core temperature. Rotarod test was conducted as previously described40.

mixed genetic background (129SvEv×FVB). All animal experiments were carried out under protocols approved by Kyoto University. In this screen, several physical features and several motor responses of the mice were recorded including body weight and core temperature. Rotarod test was conducted as previously described40. Eighteen (11 males and 7 females) wild type, thirteen (4 males and 9 females) Lis1+/− mice and fifteen (12 males and 3 females) Lis1+/− mice that were administrated with ALLN were used for regular behavioral experiments. Gait analysis was performed using ventral plane videography as described41. Briefly, mice were placed on the treadmill belt that moves at a speed of 24.7 cm/s. Digital video images of the underside of mice were collected at 150 frames per second. The paw area indicates the temporal placement of the paw relative to the treadmill belt. The color images were converted to their binary matrix equivalents, and the areas (in pixels) of the approaching or retreating paws relative to the belt and camera were calculated throughout each stride. Plotting the area of each digital paw print (paw contact area) imaged sequentially in time provides a dynamic gait signal, representing the temporal record of paw placement relative to the treadmill belt.

xels) of the approaching or retreating paws relative to the belt and camera were calculated throughout each stride. Plotting the area of each digital paw print (paw contact area) imaged sequentially in time provides a dynamic gait signal, representing the temporal record of paw placement relative to the treadmill belt. Supplementary Material 1 2 3 Acknowledgements We thank Seiichi Kawashima for providing us an anti-calpain antibody (1D10A7)36. We thank Yoshihiko Funae, Hiroshi Iwao, Toshio Yamauch, Masami Muramatsu and Yoshitaka Nagai for generous support and encouragement. We also thank Yukimi Kira, Yoriko Yabunaka and Ryusei Zako for technical support, Hiromichi Nishimura and Keiko Fujimoto for mouse breeding, Teruko Bando for in utero injection and Kazuo Nakanishi for behavior study. This work was supported by Grant-in-Aid for Scientific Research from the Ministry of Education, Science, Sports and Culture of Japan from the Ministry of Education, Science, Sports and Culture of Japan to Makoto Sato and Shinji Hirotsune. This work was also supported by The Sagawa Foundation for Promotion of Cancer Research, The Cell Science Research Foundation, The Japan Spina Bifida & Hydrocephalus Research Foundation, Takeda Science Foundation, The Hoh-ansha Foundation and Knowledge Cluster Initiative (Stage-2) Research Foundation to Shinji Hirotsune, and NIH grants NS41030 and HD47380 to Anthony Wynshaw-Boris. This work was also supported by Grant-in-Aid for Scientific Research on Priority Areas -Integrative Brain Research (Shien)- from MEXT and Grant-in-Aid from Neuroinformatics Japan Center (NIJC), Japan to Tsuyoshi Miyakawa.

earch Foundation to Shinji Hirotsune, and NIH grants NS41030 and HD47380 to Anthony Wynshaw-Boris. This work was also supported by Grant-in-Aid for Scientific Research on Priority Areas -Integrative Brain Research (Shien)- from MEXT and Grant-in-Aid from Neuroinformatics Japan Center (NIJC), Japan to Tsuyoshi Miyakawa. Figure 1 Western blotting analysis and distribution of LIS1, dynein intermediate chain (DIC1), and cellular components after administration of calpain inhibitors in MEF cells We examined LIS1 or DIC1 protein level after administration of 10 µM ALLN or 20 µM E64d by Western blotting in mouse embryonic fibroblast (MEF) cells (a) or dorsal root ganglia (DRG) neurons (b). Western blotting was performed 2 hrs after the start of treatment. Protein levels were normalized by comparison with the β-actin control and are indicated at the bottom of each panel. Statistical examination was performed by unpaired Student’s t-test, which is shown at the bottom, with *P<0.05. Error bars in graphs were expressed as mean±SEM. We performed three independent sets of experiments. One representative data set is shown. Note: LIS1 and DIC1 were augmented by ALLN or E64d treatment.

panel. Statistical examination was performed by unpaired Student’s t-test, which is shown at the bottom, with *P<0.05. Error bars in graphs were expressed as mean±SEM. We performed three independent sets of experiments. One representative data set is shown. Note: LIS1 and DIC1 were augmented by ALLN or E64d treatment. Figure 2 Rescue of neuronal migration by administration of calpain inhibitors Migration assay using cerebellar granule neurons. Images of granule neuron clusters are shown (a). The migration distance of each neuron 16 hrs after 10 µM ALLN or 20 µM E64d treatment was binned (b). Wild type neurons displayed normal migration distances, whereas Lis1+/− neurons displayed a shift in the distribution of bins toward the left. Lis1+/− neurons in the presence of 10 µM ALLN or 20 µM E64d clearly showed improvement of migration defects. Mean migration distances are summarized at the bottom (c). n is the number of neurons measured for each examination. Statistical analysis was performed by the unpaired Student’s t-test, with *P<0.05 and ***P<0.001. Error bars in graphs were expressed as mean±SEM. We performed three independent sets of experiments, and obtained reproducible results. Note; calpain inhibitors moderately facilitated neuronal migration in wild type cells, and rescued defective neuronal migrations in Lis1+/− neurons.

-test, with *P<0.05 and ***P<0.001. Error bars in graphs were expressed as mean±SEM. We performed three independent sets of experiments, and obtained reproducible results. Note; calpain inhibitors moderately facilitated neuronal migration in wild type cells, and rescued defective neuronal migrations in Lis1+/− neurons. Figure 3 Knockdown of calpain by siRNA MEF cells were transfected with siRNA against calpain small subunit 1 (Capns1). Western blotting was performed 120 hrs after transfection of siRNA. Note: depletion of calpain small subunit 1 resulted in reduction of μ-calpain and m-calpain accompanied by increase of LIS1 and DIC1. Statistical analysis was performed by the unpaired Student’s t-test, which is shown at the bottom, with *P<0.05 or **P<0.01. We performed three independent sets of experiments. One representative data set is shown.

mall subunit 1 resulted in reduction of μ-calpain and m-calpain accompanied by increase of LIS1 and DIC1. Statistical analysis was performed by the unpaired Student’s t-test, which is shown at the bottom, with *P<0.05 or **P<0.01. We performed three independent sets of experiments. One representative data set is shown. Figure 4 Rescue of defective corticogenesis in Lis1+/− mice by intra-peritoneal injection of ALLN (a) Measurement of brain weight at d5. Genotyping and injection of ALLN are indicated at the bottom. n is the number of brains examined. All statistical analyses were performed by the unpaired Student’s t-test. Error bars: ±SEM. Statistical significance was defined as *P<0.05 or **P<0.01. (b) Apoptotic cell death was examined by TUNEL staining at E15.5. Histogram plots of the relative frequency of TUNEL positive cell to the total number of cells are shown at the bottom. n is the number of brains examined. Error bars: ±SEM. (c) Neu-N staining of mid-sagittal sections of the hippocampus is shown. Severe cell dispersion and splitting of CA3 region were observed in the Lis1+/− mouse. (d) Cortical phenotypes were examined by a layer specific maker, Brn-1 (layer 2 and 3). The distribution of Brn-1 positive cells is indicated at the right side of each panel. Quantitation of the thickness of Brn-1 positive cells is summarized at the bottom. n is the number of brains examined. Error bars: ±SEM. (e) BrdU birthdating analysis Quantitative analysis was performed by measuring the distribution of BrdU labeled cells in each bin that equally divided the cortex from ML to SP. (f) In utero injection of shRNA against Capns1. The distribution of migrated neurons is shown at lower panels. Cortex was divided into ten compartments, followed by counting of the neurons located at each compartment, and summarized.

ution of BrdU labeled cells in each bin that equally divided the cortex from ML to SP. (f) In utero injection of shRNA against Capns1. The distribution of migrated neurons is shown at lower panels. Cortex was divided into ten compartments, followed by counting of the neurons located at each compartment, and summarized. Figure 5 Rescue of impaired behavior in Lis1+/− mice by intra-peritoneal injection of ALLN Neurological screen: wire hang test (a). Note: there were no obvious differences in body weight, rectal temperature (Supplementary Table 1) and grip strength in each group. Lis1+/− mice displayed clear shorter time to falling in the wire hang test. P-values are shown at the upper parts of bars. Statistical analysis was conducted using Stat View (SAS institute). Data were analyzed by two-way ANOVA. Error bars in graphs were expressed as mean±SEM. All P-values indicated are two tailed. Statistical significance was defined as *P<0.05 or **P<0.01. (b) Examination of motor function by the rotarod test. Time spent balanced on top of the rotating rod was measured across six test trials for Lis1+/+ mice (open circle), Lis1+/− mice without ALLN treatment (open triangle) and Lis1+/− mice with ALLN treatment (closed circle). Significant differences between Lis1+/+ mice and Lis1+/− mice (***P<0.001) were observed. Lis1+/− mice with ALLN treatment displayed improvement of rotarod performance. Data were analyzed by two-way repeated measures. (c) Examination of stride length variability and (d) fore paw angle in gait analysis. Lis1+/− mice with ALLN treatment displayed improvement of gait parameters. Data were analyzed by two-way ANOVA.

Tuberculosis (TB) continues to be a major public health threat around the world. The estimate that more lives may have been lost in 2009 due to TB than in any year in history is alarming 4. An increasing number of cases reporting infection with multi-(MDR) and extensively drug-resistant (XDR) strains of M. tuberculosis has diminished our capability to respond effectively against this threat. A recent study reporting high mortality rates of patients co-infected with HIV and XDR-TB illustrates the need for new drugs to treat TB 5. A major reason for emergence of drug resistance is thought to be poor compliance to treatment regimens as the current therapy requires a combination of drugs to be taken daily for 6 months or more 4. While >99% of M. tuberculosis bacilli are killed within two weeks, it takes the remainder of the therapy to effectively kill the surviving population 6. These bacilli, broadly defined as ‘persisters’, are able to transiently tolerate drugs. The phenomenon of persistence is poorly understood. In vitro models designed to mimic the physiology of persisters are based on exposure to nitric oxide and depletion of oxygen and nutrients as these conditions are thought to prevail in a persisting infection in vivo 7,8,9.

ers’, are able to transiently tolerate drugs. The phenomenon of persistence is poorly understood. In vitro models designed to mimic the physiology of persisters are based on exposure to nitric oxide and depletion of oxygen and nutrients as these conditions are thought to prevail in a persisting infection in vivo 7,8,9. A higher percentage of E. coli bacilli are able to survive exposure to drugs at stationary phase and persist compared to exponential phase of growth 10. Little is known about changes in the cell wall during chronic phase of infection and whether it regulates persistence of M. tuberculosis in the host. Understanding the regulation of cell wall physiology and its consequences may enable us to effectively target and kill persisters by interfering with this process. The cell wall of M. tuberculosis accounts for up to 40% of cell’s dry mass compared to 5% and 10% in gram-positive and negative bacteria 11. It is approximated that the degree of peptidoglycan cross–linking is ~50% in E. coli 12 whereas it is ~70–80% in Mycobacterium spp. 13. Mycobacterial peptidoglycan layer is cross–linked with both 4→3 and 3→3 linkages 1. Recently it was shown that 80% of the transpeptide linkages in the peptidoglycan of M. tuberculosis at stationary phase are of the non–classical 3→3 type 2. In this study, we report identification of a gene MT2594(Rv2518c) that encodes an L,D–transpeptidase for synthesis of non–classical 3→3 cross-linkages and show that inactivation of the gene results in altered colony morphology, attenuation of persistence, and increased susceptibility to amoxicillin/clavulanate both in vitro and in the mouse model of TB.

tion of a gene MT2594(Rv2518c) that encodes an L,D–transpeptidase for synthesis of non–classical 3→3 cross-linkages and show that inactivation of the gene results in altered colony morphology, attenuation of persistence, and increased susceptibility to amoxicillin/clavulanate both in vitro and in the mouse model of TB. A mutant M. tuberculosis resulting from inactivation of gene MT2594 (Rv2518c), hereafter referred to ldtMt2, was isolated by screening a collection of 5,100 unique transposon insertion mutants for growth attenuation 14. Colonies of this mutant (ldtMt2::Tn) were smaller, smooth and had punctuated aerial growth rather than the typical large, rough and laterally diffuse morphology observed in the parent strain (Fig. 1a). In liquid broth, culture of the mutant strain reached lower optical and cell densities compared with the parental strain (Fig. 1b). The wild–type phenotypes were restored upon complementation of the mutant with a single copy of the gene. The ratio of colony forming units (CFU) of the mutant to WT at the beginning of the growth assay was 10:1 whereas after 192 h it was 1:100. This data shows that there was ~1,000 fold larger increase in the wild-type population over 192 h compared to the mutant strain. The doubling times of 18.1, 14.5 and 14.8 h were derived from CFU data for the mutant, wild-type and the complemented strains, respectively.

h assay was 10:1 whereas after 192 h it was 1:100. This data shows that there was ~1,000 fold larger increase in the wild-type population over 192 h compared to the mutant strain. The doubling times of 18.1, 14.5 and 14.8 h were derived from CFU data for the mutant, wild-type and the complemented strains, respectively. MT2594 (LdtMt2) is annotated as a hypothetical protein with an unknown function. The N–terminus contains a putative single transmembrane domain spanning residues (positions 20–42) anchoring the remainder of the protein that is predicted to protrude outside the cell membrane into the cell wall (www.ch.embnet.org). The C–terminal region of MT2594 is similar to the catalytic domain of the prototypic peptidoglycan L,D–transpeptidase from E. faecium (29% identity) including the active site cysteine residue within the invariant SHGC motif 15 (Fig. 2a). To determine if ldtMt2 encodes a functional L,D–transpeptidase, we produced a soluble fragment of the protein in E. coli that was purified and assayed for cross-linking activity (Fig. 2b). The substrate was a disaccharide–tetrapeptide monomer isolated from the peptidoglycan of C. jeikeium, which has the same structure as the predominant monomer of M. tuberculosis 16. Electrospray mass spectrometry analysis of the reaction products revealed the formation of a peptidoglycan dimer (m/z = 1,786.87 [M+H]1+, Fig. 2c) from two disaccharide–tetrapeptide monomers (m/z = 938.44 [M+H]1+). Tandem mass spectrometry of the dimer confirmed the presence of a 3→3 cross-link connecting two meso–diaminopimelic acids at the third position of the stem peptides (Fig. 2c). LdtMt2 did not catalyze formation of dimers with disaccharide–peptide containing a stem pentapeptide. Thus, LdtMt2 is specific for stem tetrapeptides as LdtMt1 2 and the prototypic L,D–transpeptidase from E. faecium 15. These results show that MT2594 is an L,D–transpeptidase that catalyzes formation of 3→3 peptidoglycan cross–links.

catalyze formation of dimers with disaccharide–peptide containing a stem pentapeptide. Thus, LdtMt2 is specific for stem tetrapeptides as LdtMt1 2 and the prototypic L,D–transpeptidase from E. faecium 15. These results show that MT2594 is an L,D–transpeptidase that catalyzes formation of 3→3 peptidoglycan cross–links. Further investigation revealed four putative paralogs of ldtMt2 in the genome of M. tuberculosis (MT0125, MT0202, MT0501 and MT1477). To gain an insight into the level of expression of the five paralogs, we performed quantitative RT–PCR analyses on eight RNA samples prepared from exponential and stationary phases of growth (Supplementary Fig. 2). The ldtMt2 mRNA was at least five fold more abundant than the combined expression of the four paralogs. Next, we assessed their functional relevance to L,D–transpeptidation. Mutants lacking MT0202, MT0501 and MT1477 have morphologies and growth phenotypes similar to that of the parent wild-type strain (Supplementary Fig. 3). We purified MT0202, MT0501 and MT1477 but did not detect any L,D–transpeptidase activity using the peptidoglycan precursor as the substrate. MT0125 of M. tuberculosis strain CDC1551 is identical to Rv0116c of strain H37Rv, a gene designated ldtMt1 that was recently shown to also encode a peptidoglycan L,D–transpeptidase 2. Between the two L,D–transpeptidases, MT2594 was expressed at a level at least 10 fold higher than that of ldtMt1 at all phases of growth (Fig. 2d). The morphology, growth and virulence deficient phenotype of our mutant indicates that low-level expression of ldtMt1 did not compensate for the loss of the L,D–transpeptidase activity of MT2594 although analysis of the peptidoglycan structure showed that 3→3 cross–linkages were synthesized by LdtMt1 in stationary phase culture of the mutant (Supplementary Fig. 4).

phenotype of our mutant indicates that low-level expression of ldtMt1 did not compensate for the loss of the L,D–transpeptidase activity of MT2594 although analysis of the peptidoglycan structure showed that 3→3 cross–linkages were synthesized by LdtMt1 in stationary phase culture of the mutant (Supplementary Fig. 4). Next, we assessed if inactivation of ldtMt2 affected in vivo growth and virulence of M. tuberculosis. Approximately 3.5 log10 of CFUs of each strain was implanted in the lungs of three groups of immunocompetent BALB/c mice. The mutant established infection and exhibited a normal growth pattern during the first two weeks but discontinued proliferation beyond this stage of infection (Fig. 3a). The wild–type and the complemented strains proliferated rapidly until four weeks of infection, at which stage a heavy bacterial burden led to death of the animals. The median survival time for the wild–type and complement infected group were 38 and 30 days, respectively (Fig. 3b). The mice infected with the mutant strain did not die and signs of morbidity were not observed despite the presence of ca. 104 CFU in the lungs during the persistent stage of infection.

h of the animals. The median survival time for the wild–type and complement infected group were 38 and 30 days, respectively (Fig. 3b). The mice infected with the mutant strain did not die and signs of morbidity were not observed despite the presence of ca. 104 CFU in the lungs during the persistent stage of infection. The non-classical 3→3 linkages comprise the majority of the transpeptide linkages in non–replicating M. tuberculosis 2. We hypothesized that the loss of LdtMt2 may compromise the mutant’s ability to adapt during the chronic phase of infection, a critical stage in the pathogenesis of TB. If the failure in adaptation was the result of a defect in peptidoglycan cross-linking by LdtMt2, another consequence could be an increased susceptibility to β–lactams. We tested this hypothesis by assessing susceptibility of the LdtMt2 mutant to amoxicillin. M. tuberculosis produces a β–lactamase that is inactivated by clavulanic acid 17. The MIC of the commercial clavulanic acid–amoxicillin combination (Augmentin®) was 1.2 and 0.14 µg ml−1 for wild-type and the mutant strain respectively. Loss of MT2594 did not alter susceptibility to isoniazid (MIC = 0.03 µg ml−1) and D–cycloserine (5 µg ml−1). Thus, loss of LdtMt2 was associated with increased susceptibility to amoxicillin in the presence of clavulanic acid that inhibited the β–lactamase. In a recent study authors assessed susceptibility of drug sensitive laboratory strain H37Rv and 13 XDR strains of M. tuberculosis to amoxicillin 3. While H37Rv was found to be resistant to amoxicillin–clavulanate, all 13 XDR strains were highly susceptible with an MIC ranging between 0.32 and 10 ug ml−1. It has been reported that amoxicillin–clavulanate lacks early bactericidal activity, a measure of effectiveness of the drugs during the first two days of treatment 18. However, amoxicillin–clavulanate has also been used to treat MDR–TB patients 19. An explanation for these observations is that amoxicillin–clavulanate lacks potency during the early phase but shows activity during the extended phase of treatment.

of effectiveness of the drugs during the first two days of treatment 18. However, amoxicillin–clavulanate has also been used to treat MDR–TB patients 19. An explanation for these observations is that amoxicillin–clavulanate lacks potency during the early phase but shows activity during the extended phase of treatment. Next we assessed susceptibility of the mutant strain to amoxicillin in the mouse. Mice infected with either the wild–type M. tuberculosis or the mutant strain were treated daily with either phosphate buffered saline (PBS) as a placebo or 25mg kg−1 of isoniazid or 200 mg kg−1 amoxicillin in combination with 50 mg kg−1 clavulanate (Fig. 3c). Isoniazid was similarly effective against both wild-type and the mutant strain (Fig. 3e). Bacterial burden in mice infected with the wild–type strain treated with amoxicillin–clavulanate was similar to the group that received no treatment placebo (Fig. 3d, f). Amoxicillin–clavulanate was ineffective during the first two weeks of treatment in mice infected with the mutant as the bacterial burden remained unchanged. However, a decrease of more than 2 log10 in CFU was observed between two and four weeks of treatment illustrating that the mutant is selectively susceptible to amoxicillin–clavulanate during the chronic phase of infection (Fig. 3f).

treatment in mice infected with the mutant as the bacterial burden remained unchanged. However, a decrease of more than 2 log10 in CFU was observed between two and four weeks of treatment illustrating that the mutant is selectively susceptible to amoxicillin–clavulanate during the chronic phase of infection (Fig. 3f). All bacteria possess an elaborate peptidoglycan layer in their cell wall. In E. coli the main inter–peptide cross-linking occurs between the penultimate D–alanine (D–ala) at the fourth position of the donor and mDAP at the third position of the acceptor. Formation of these 4→3 bond is catalyzed by the D,D-transpeptidases that are the essential target of β–lactam antibiotics 20. The drugs are structural analogs of the D–Ala–D–Ala extremity of the peptidoglycan precursors and act as suicide substrate in an acylation reaction 21. E. coli has served as the model organism for studying peptidoglycan metabolism but the existing paradigm built on this organism is incomplete. In this classical model the peptidoglycan layer is regarded as a static network involving polymerization of glycan chains and cross–linking of adjacent chains by formation of 4→3 peptide bonds (Fig. 4a) 22. Based on the data presented in this report and our recent findings 2 we propose a model describing the peptidoglycan layer as a dynamic structure whose inter–peptide linkages are altered as an adaptive response to a change in the environment and growth phase (Fig. 4b). The peptide chains of the peptidoglycan layer are linked with both 4→3 and 3→3 bonds.

report and our recent findings 2 we propose a model describing the peptidoglycan layer as a dynamic structure whose inter–peptide linkages are altered as an adaptive response to a change in the environment and growth phase (Fig. 4b). The peptide chains of the peptidoglycan layer are linked with both 4→3 and 3→3 bonds. In addition to transpeptidases, endo– and carboxypeptidases are also present in bacteria and function to modify the peptidoglycan network 23. These enzymes have yet to be definitively identified in M. tuberculosis. A recent report showed gene pgdA to encode a N–deacetylase that is involved in modification of the peptidoglycan layer in L. monocytogenes 24. A putative homolog of pgdA exists in M. tuberculosis. This gene, MT1128 (Rv1096), has yet to be characterized and its in vivo function identified. Although MT2594 is a 3→3 transpeptidase, its loss and accompanying changes in the cross-linking may have pleiotropic effects on the metabolism of the peptidoglycan layer.

4. A putative homolog of pgdA exists in M. tuberculosis. This gene, MT1128 (Rv1096), has yet to be characterized and its in vivo function identified. Although MT2594 is a 3→3 transpeptidase, its loss and accompanying changes in the cross-linking may have pleiotropic effects on the metabolism of the peptidoglycan layer. In this report we have shown a novel molecular basis of 3→3 linkages and their physiological role for viability and virulence of M. tuberculosis. We have also shown that L,D–transpeptidation is required to resist killing by amoxicillin–clavulanate and that inhibition of LdtMt2 alone may be sufficient to target the 3→3 linkages despite presence of redundancy. It may be inferred from our findings that both 3→3 and 4→3 trans-peptide linkages need to be destroyed to effectively kill M. tuberculosis. We have presented an unexploited enzyme in the pathway that has been targeted by one the most successful antibiotics in human clinical use, namely the β–lactams. Therefore, a regimen containing a combination of inhibitors of L,D–transpeptidase and β–lactamase, and a β–lactam may be able to kill M. tuberculosis by comprehensively destroying the peptidoglycan layer. As peptidoglycan layer is a vital structure of the bacterial cell wall, insight and applications resulting from studies in M. tuberculosis is likely broadly applicable to other bacteria.

speptidase and β–lactamase, and a β–lactam may be able to kill M. tuberculosis by comprehensively destroying the peptidoglycan layer. As peptidoglycan layer is a vital structure of the bacterial cell wall, insight and applications resulting from studies in M. tuberculosis is likely broadly applicable to other bacteria. METHODS Bacterial strains and Culture Conditions We used M. tuberculosis CDC1551, a clinical isolate, as the host strain to generate a transposon insertion mutant in MT2594 (ldtMt2::Tn) as described 14. This mutant carries a Himar1 transposon insertion at +872 base from the putative translation start site of the gene. Next, we generated a complemented strain by transforming the mutant with pGS202_2594. This is a single copy integrating plasmid based on pMH94 backbone 25, which we modified into a GATEWAY compatible destination vector (Invitrogen). We cloned a wild-type copy of MT2594 along with its promoter into this destination vector pGS202 to generate pGS202_2594. We verified genotypes of the strains by Southern blotting. We used Middlebrook 7H9 liquid medium supplemented with 0.2% glycerol, 0.05% Tween–80, 10% vol/vol oleic acid–albumin–dextrose–catalase (OADC) and 50 ug ml−1 cycloheximide for in vitro growth, and Middlebrook 7H11 solid medium (Becton-Dickinson) for enumerating colony forming units (CFU) in in vitro and in vivo growth studies.

Middlebrook 7H9 liquid medium supplemented with 0.2% glycerol, 0.05% Tween–80, 10% vol/vol oleic acid–albumin–dextrose–catalase (OADC) and 50 ug ml−1 cycloheximide for in vitro growth, and Middlebrook 7H11 solid medium (Becton-Dickinson) for enumerating colony forming units (CFU) in in vitro and in vivo growth studies. Production and Purification of Recombinant LdtMt2 We amplified a portion of ltdMt2 with primers 5’-TTTTCATGATCGCCGATCTGCTGGTGC-3’ and 5’-TTGGATCCCGCCTTGGCGTTACCGGC-3’, digested with BspHI–BamHI (underlined) and cloned into pET2818 15. The resulting plasmid, encodes a fusion protein consisting of a methionine specified by the ATG initiation codon of pET2818, residues 55 to 408 of LdtMt2, and a C-terminal polyhistidine tag with the sequence GSH6. We grew E. coli BL21(DE3) harboring pREP4GroESL 26 and pET2818ΩldtMt2 at 37 °C in 3 L of brain heart infusion broth containing 150 µg ml−1 ampicillin, induced expression using Isopropyl–D–thiogalactopyranoside. We purified LdtMt2 from a clarified lysate by affinity chromatography on Ni2+-nitrilotriacetate-agarose resin (Qiagen GmbH, Germany) followed by anion exchange chromatography (MonoQ HR5/5, Amersham Pharmacia) with a NaCl gradient in 50 mM Tris–HCl pH 8.5. We performed an additional size exclusion chromatography on a Superdex HR10/30 column equilibrated with 50 mM Tris–HCl (pH 7.5) containing 300 mM NaCl. Finally, we concentrated the protein by ultrafiltration (Amicon Ultra–4 centrifugal filter devices, Millipore) and stored at −20 °C in the same buffer supplemented with 20% glycerol.

n additional size exclusion chromatography on a Superdex HR10/30 column equilibrated with 50 mM Tris–HCl (pH 7.5) containing 300 mM NaCl. Finally, we concentrated the protein by ultrafiltration (Amicon Ultra–4 centrifugal filter devices, Millipore) and stored at −20 °C in the same buffer supplemented with 20% glycerol. L,D-transpeptidase assays We purified disaccharide-tetrapeptide containing amidated meso-diaminopimelic acid (GlcNAc–MurNAc–L–Ala1–D–iGln2–mesoDapNH23–D–Ala4) from C. jeikeium strain CIP103337 and determined the concentration after acid hydrolysis 27,28. Next, we tested in vitro formation of muropeptide dimers in 10 µL of 50 mM Tris–HCl (pH 7.5) containing 300 mM NaCl, 5 µM LdtMt2, and 280 µM disaccharide–tetrapeptide. We incubated he reaction mixture for two hours at 37 °C and analyzed the resulting muropeptides by nanoelectrospray tandem mass spectrometry using N2 as the collision gas 28.

ro formation of muropeptide dimers in 10 µL of 50 mM Tris–HCl (pH 7.5) containing 300 mM NaCl, 5 µM LdtMt2, and 280 µM disaccharide–tetrapeptide. We incubated he reaction mixture for two hours at 37 °C and analyzed the resulting muropeptides by nanoelectrospray tandem mass spectrometry using N2 as the collision gas 28. Growth and Virulence Analysis in Mice We used 4–5 week old, female BALB/c mice (Charles River Laboratories) to study in vivo virulence of the strains and their susceptibility to drugs. We infected mice with a log phase culture of wild–type M. tuberculosis, or ldtMt2::Tn or the complemented strain in an aerosol chamber. For assessing in vivo growth of each strain, we sacrificed four mice per group at days 1, 7, 14, 28, 56 and 98 following infection, obtained lungs and spleen, homogenized and cultured appropriate dilutions on Middlebrook 7H11 medium to determine CFU. We allocated 12 mice for each infection group to assess virulence of each strain, for which we determined median-survival-time that mice from each group survived following infection. Protocols for experiments involving mice were approved by Johns Hopkins University Animal Care and Use Committee.

edium to determine CFU. We allocated 12 mice for each infection group to assess virulence of each strain, for which we determined median-survival-time that mice from each group survived following infection. Protocols for experiments involving mice were approved by Johns Hopkins University Animal Care and Use Committee. Drug Susceptibility Testing in Mice We determined minimum inhibitory concentrations for amoxicillin–clavulanate, imipenem (Merck), isoniazid (Sigma) and cycloserine (Sigma) using the broth dilution method 29. We used Augmentin (GlaxoSmithKline), a preparation containing amoxicillin and calvulanate as M. tuberculosis contains β–lactamases 17. For this we inoculated 105 M. tuberculosis bacilli in 2.5 ml of 7H9 broth and added drugs at different concentrations. We incubated these cultures at 37 °C and evaluated for growth by visual inspection at 7 and 14 days. For those samples with diminished growth compared to no-drug control, we determined CFU. We used four-five week old, female BALB/c mice for in vivo assessment of susceptibility of M. tuberculosis lacking LdtMt2 to amoxicillin. We infected two groups of mice, 36 per group, with aerosolized cultures of either wild-type M. tuberculosis or ldtMt2::Tn. We allocated 12 mice from each group into 3 sub–groups and initiated daily treatment at 2 weeks following infection. We provided each sub–group either 25 mg kg−1 isoniazid, or 200 mg kg−1 amoxicillin and 50 mg kg−1 clavulanate, or no drug at all by oral gavage. For analysis, we sacrificed four mice from each treatment sub-group at 1, 2 and 4 weeks following initiation of therapy, obtained lungs, homogenized in 1 ml of PBS and determined CFU in each organ by plating appropriate dilutions of the homogenates on Middlebrook 7H11 selective plates.

or no drug at all by oral gavage. For analysis, we sacrificed four mice from each treatment sub-group at 1, 2 and 4 weeks following initiation of therapy, obtained lungs, homogenized in 1 ml of PBS and determined CFU in each organ by plating appropriate dilutions of the homogenates on Middlebrook 7H11 selective plates. Supplementary Material 1 2 ACKNOWLEDGMENT The support of NIH award AI30036 is gratefully acknowledged. This work was also supported by the Foundation pour la Recherche Médicale (Equipe FRM 2006 (DEQ200661107918). M. Lavollay is the recipient of INSERM PhD fellowship (Poste d’Accueil pour Pharmacien, Médecin, et Vétérinaire). AUTHOR CONTRIBUTIONS: R. Gupta, W. R. Bishai, and G. Lamichhane designed the project. J-L. Mainardi, and M. Arthur, designed the biochemical characterization of MT2594. M. Lavollay and J-L. Mainardi performed biochemistry and analyzed data. R. Gupta and G. Lamichhane conducted genetics, microbiology and mouse experiments. G. Lamichhane wrote the manuscript with contributions from co-authors. Figure 1 Morphology and growth in vitro. (a) Morphologies of wild-type M. tuberculosis (WT), LdtMt2 mutant (MUT) and the complemented strain (COMP) on solid media after 21 days of growth at 37 °C. (scale = 1 cm).(b) Growth of wild-type M. tuberculosis (WT), LdtMt2 mutant (MUT) and the complemented strain (COMP) in Middlebrook 7H9 liquid medium at 37 °C. The decrease in optical density and CFU at the final time point for WT and COMP is due to clumped cultures.

OMP) on solid media after 21 days of growth at 37 °C. (scale = 1 cm).(b) Growth of wild-type M. tuberculosis (WT), LdtMt2 mutant (MUT) and the complemented strain (COMP) in Middlebrook 7H9 liquid medium at 37 °C. The decrease in optical density and CFU at the final time point for WT and COMP is due to clumped cultures. Figure 2 Characterization of LdtMt2 from M. tuberculosis. (a) Domain composition of L,D-transpeptidases from E. faecium (Ldtfm) and M. tuberculosis (LdtMt2). Residues 250–377 of LdtMt2 share homology with the catalytic domain of Ldtfm (Domain II, 338–466). (b) Purification of a soluble fragment of LdtMt2 produced in E. coli. (c) Structure and inferred fragmentation pattern of the peptidoglycan dimer formed in vitro by LdtMt2. The ion at m/z 974.51 was generated by losses of the two GlcNAc-MurNAc residues following fragmentations of the ether links connecting the lactoyl group to the C–3 position of MurNAc. Loss of each sugar decreased the m/z by 203. Cleavage of additional peptide bonds from the ion at m/z 974.51 gave ions at 703.36, 532.28, 433.22 and 272.15 as indicated in the inset. (d) Transcription profile of LdtMt2 (MT2594) and LdtMt1 (MT0125). RNA isolated from wild-type Mtb cultures at growth phases T1 (A600nm=0.2), T2 (A600nm=0.5), T3 (A600nm=0.8), T4 (A600nm=0.9), T5 (A600nm=01.9), T6 (A600nm=3.0), T7 (2 days post A600nm =3.0) and T8 (3 days post clumping).

272.15 as indicated in the inset. (d) Transcription profile of LdtMt2 (MT2594) and LdtMt1 (MT0125). RNA isolated from wild-type Mtb cultures at growth phases T1 (A600nm=0.2), T2 (A600nm=0.5), T3 (A600nm=0.8), T4 (A600nm=0.9), T5 (A600nm=01.9), T6 (A600nm=3.0), T7 (2 days post A600nm =3.0) and T8 (3 days post clumping). Figure 3 Assessment of growth, virulence and susceptibility to amoxicillin in vivo. (a) Bacterial burden in the lungs of mice infected with wild-type M. tuberculosis (WT), LdtMt2 mutant (MUT) or the complemented (COMP) strain are shown. (b) Virulence of each strain was assessed by determining time-to-death following infection of mice, 12 per group, with the three strains. (c) Mice were infected with either WT or the MUT strain. Following two weeks of infection, mice were treated daily with either no drug placebo (d), or 25 mg kg−1 isoniazid (e), or 200 mg kg−1 amoxicillin and 50 mg kg−1 clavulanate (f). Bacterial burden was determined by enumerating CFU from the lungs of mice. Figure 4 Proposed model for physiology of the peptidoglycan layer in M. tuberculosis. (a) classical model of peptidoglycan crosslinking containing 4→3 inter peptide bonds. (b) model based on recent data: the peptidoglycan is cross-linked with classical 4→3 and non-classical 3→3 inter peptide bonds. Both 4→3 (orange) and 3→3 (green) transpeptidases are involved in maintenance and remodeling of the peptidoglycan layer in the proposed model.

Introduction Despite the host immune response and treatment with highly active antiretroviral therapy (HAART), HIV causes a persistent infection. Viral persistence is due in part to latent HIV reservoirs in resting CD4+ T cells 1 that do not express viral proteins but can be induced to active infection by a variety of stimuli. However, recent studies of viral genetics have revealed that additional reservoirs likely exist 2. Hematopoietic progenitor cells (HPCs) have been considered as a possible reservoir, but it has been difficult to establish that these cells are infected by HIV 3-6 because HPCs are difficult to maintain in culture and indirect measurements of infection may be confounded by contamination with other cell types. Here, we utilized flow cytometry and recently developed culture conditions 7 that have allowed us to conclude that a proportion of HPCs become infected following exposure to HIV both in vivo and in vitro. RESULTS HIV infects HPCs To assess the susceptibility of HPCs to HIV, intracellular Gag expression was examined using purified bone marrow (BM) CD34+ cells treated with HIV 89.6ΔEenv89.6 (Fig. 1a). After 3 d in culture, 6% of CD34+ HPCs expressed intracellular HIV Gag (Fig. 2a, middle right panel). Antiretroviral treatment blocked Gag expression (Fig. 2a, lower right panel) and experiments with five other HIVs yielded similar results (Supplementary Fig. 1a). As previously reported for HIV–infected T cells 8,9, infected CD34+ cells downmodulated MHC–I (Fig. 2b).

intracellular HIV Gag (Fig. 2a, middle right panel). Antiretroviral treatment blocked Gag expression (Fig. 2a, lower right panel) and experiments with five other HIVs yielded similar results (Supplementary Fig. 1a). As previously reported for HIV–infected T cells 8,9, infected CD34+ cells downmodulated MHC–I (Fig. 2b). HPCs are a heterogeneous collection of cells that include multipotent HPCs and stem cells (HSCs). Multipotent HPCs have a Lin−CD34+CD133+CD38− surface phenotype, where “Lin” represents markers of specific hematopoietic lineages. Following treatment with wild type HIV 89.6 (Fig. 1b), both Lin+ and Lin− cells expressed intracellular Gag (Fig. 2c). HIV is cytotoxic to infected HPCs A time course analysis revealed that Gag+ cells were lost rapidly in culture (Fig. 2d and Supplementary Fig. 1b). Moreover, infected cells displayed increased annexin V reactivity (Fig. 2e) and a high fraction of Gag+ cells had light scatter properties of dead cells (Supplementary Fig. 1c). Cell death required active viral gene expression as transduction of the cells with a reporter virus (Fig. 1c) pseudotyped with an HIV envelope did not result in cell loss unless the HIV LTR actively expressed HIV genes (Supplementary Fig. 1d).

g+ cells had light scatter properties of dead cells (Supplementary Fig. 1c). Cell death required active viral gene expression as transduction of the cells with a reporter virus (Fig. 1c) pseudotyped with an HIV envelope did not result in cell loss unless the HIV LTR actively expressed HIV genes (Supplementary Fig. 1d). Multipotent HPCs are susceptible to HIV infection To assess the developmental capacity of infected HPCs, we used a minimal HIV genome (HIV–7SF–GFP, Fig. 1d) pseudotyped with 89.6 Env, which “tagged” infected cells without causing cell death. Using this system, we found that a proportion of CD34+ cells were infected (GFP+) [1 to 6% in replicate experiments (e.g. Fig. 3a; initial sort purity shown in Supplementary Fig. 2a)] and a more primitive subset of these cells (CD34+CD38−CD133+) had a similar infection rate (Fig. 3b). Infection of CD133+ HPCs purified from BM yielded similar results (Supplementary Fig. 2f and g). These infection rates were comparable to the fraction of CD34+ cells expressing both HIV co–receptors (Supplementary Fig. 3a and b). CD133+ HPCs from UCB infected with HIV–7SF–GFPenv89.6 generated GFP+ colonies of erythroid (CFU–E), myeloid (CFU–M and CFU–GM) and multi–lineage (CFU–GEMM) origin, demonstrating that HIV can infect multipotent HPCs (Fig. 3c). Quantitation revealed similar numbers of total colonies from uninfected and infected cells (Fig. 3d). Similar results were obtained using a full–length HIV reporter (89.6–SIΔE–SF–GFP, Fig. 1e) that did not express HIV genes because of an LTR mutation (Fig. 3e,f).

strating that HIV can infect multipotent HPCs (Fig. 3c). Quantitation revealed similar numbers of total colonies from uninfected and infected cells (Fig. 3d). Similar results were obtained using a full–length HIV reporter (89.6–SIΔE–SF–GFP, Fig. 1e) that did not express HIV genes because of an LTR mutation (Fig. 3e,f). Induction of latent HIV from infected HPCs To assess latent infection, we asked whether induction of differentiation induced viral gene expression. Indeed, BM–derived HPCs (99.5% CD34+, Supplementary Fig. 2b) infected with replication defective [HIV HXB–ePLAPenvVSV–G (Fig. 1f)] and treated with PMA expressed the reporter in 12-fold more cells (Fig. 4a) and produced more viral particles (Supplementary Fig. 4a) than the control. BM immunodepleted for CD34+ cells were not viable under these conditions (Supplementary Fig. 4b). We found similar numbers of integrated genomes plus or minus PMA (Fig. 4b), indicating that PMA–induced gene expression was not due to effects on integration. Consistent with these results, the integrase inhibitor raltegravir blocked initial infection but not PMA–induced gene expression (Supplementary Fig. 5). Similar results were obtained using HXB–ePLAPenv89.6, albeit with lower infection rates (Fig. 4c).

t PMA–induced gene expression was not due to effects on integration. Consistent with these results, the integrase inhibitor raltegravir blocked initial infection but not PMA–induced gene expression (Supplementary Fig. 5). Similar results were obtained using HXB–ePLAPenv89.6, albeit with lower infection rates (Fig. 4c). Using wild type HIV–89.6 (Fig. 1b), we infected purified BM–derived HPCs (98% CD34+, Supplementary Fig. 2d) and cultured them plus or minus GM–CSF and TNF–α to induce myeloid differentiation 10. GM–CSF–TNF–α–treatment of infected HPCs resulted in rapid release of HIV into the culture supernatant (Fig. 4d). In contrast, BMMCs immunodepleted for CD34+ cells did not release HIV (Fig 4d) and rapidly died (Supplementary Fig 4 b, c, d and f). Flow cytometric analysis of the cells confirmed that GM–CSF–TNF–α stimulated intracellular HIV Gag expression (Fig. 4e) and that cells cultured in GM–CSF–TNF–αacquired myeloid markers (CD83+) (Fig. 4f).

MCs immunodepleted for CD34+ cells did not release HIV (Fig 4d) and rapidly died (Supplementary Fig 4 b, c, d and f). Flow cytometric analysis of the cells confirmed that GM–CSF–TNF–α stimulated intracellular HIV Gag expression (Fig. 4e) and that cells cultured in GM–CSF–TNF–αacquired myeloid markers (CD83+) (Fig. 4f). To assess the stability of latent HIV in HPCs, we infected CD34+ BM–derived HPCs (99% pure, Supplementary Fig. 2e) with wild type 89.6. After 7 d, when the culture was uniformly Gag–negative, GM–CSF–TNF–α was added to half the culture. GM-CSF-TNF-α resulted in a resurgence of HIV gene expression compared with the untreated culture (Fig 4g, h). 89.6 is a dual tropic HIV that can utilize both CCR5 and CXCR4 to enter cells. Similar results were obtained with a wild type virus that only uses CXCR4 although, as expected, there was less viral spread in the differentiated myeloid cells (Supplementary Fig. 6b). Spread of infection in the culture was inhibited by antiretrovirals and supernatant from infected cells could be used to infect T cell lines (Supplementary Fig. 6).

with a wild type virus that only uses CXCR4 although, as expected, there was less viral spread in the differentiated myeloid cells (Supplementary Fig. 6b). Spread of infection in the culture was inhibited by antiretrovirals and supernatant from infected cells could be used to infect T cell lines (Supplementary Fig. 6). Direct detection of latency To detect latent infection in situ without inducing changes in the infected cells, we developed a novel latency reporter virus [89.6–ΔE–SF–GFP (Fig. 1c)] that expresses GFP independently of the HIV LTR. Infection of T cells with 89.6–ΔE–SF–GFPenv89.6 yields some cells expressing Gag and others expressing only GFP (Fig. 5a). To confirm that GFP+Gag− cells were latently infected, we demonstrated that CD4 downmodulation, which occurs only when HIV Nef, Vpu or Env is expressed, 11 occurred in Gag+ but not GFP+Gag− cells (Fig. 5a). In contrast, when cells were infected with a virus that expressed GFP from the HIV LTR (89.6–ΔE–IRES–GFP, Fig. 1g), GFP–expressing cells downmodulated CD4 (Fig. 5a). Similar results were obtained using peripheral blood mononuclear cells (PBMCs) infected with 89.6ΔE–SF–GFPenv89.6 (Fig. 5b). Moreover, PMA and ionomycin treatment of Jurkat cells infected with the reporter virus increased Gag+ cell frequency and reduced GFP+Gag− cell frequency (Fig. 5c).

ls downmodulated CD4 (Fig. 5a). Similar results were obtained using peripheral blood mononuclear cells (PBMCs) infected with 89.6ΔE–SF–GFPenv89.6 (Fig. 5b). Moreover, PMA and ionomycin treatment of Jurkat cells infected with the reporter virus increased Gag+ cell frequency and reduced GFP+Gag− cell frequency (Fig. 5c). We observed separate populations of Gag+ and GFP+ cells in UCB–derived CD34+ HPCs infected with the latency reporter virus, indicating that active and latent infection occurred in this cell type (Fig. 5d). In culture, the Gag+ cells were rapidly lost whereas the GFP+Gag−cells persisted at least 20 d (Fig. 5e, Supplementary Fig 1d). Analysis of these cells revealed that many had a cell surface phenotype consistent with primitive HPCs (CD34+Lin− or CD34+CD38−)(Fig. 5f).

atent infection occurred in this cell type (Fig. 5d). In culture, the Gag+ cells were rapidly lost whereas the GFP+Gag−cells persisted at least 20 d (Fig. 5e, Supplementary Fig 1d). Analysis of these cells revealed that many had a cell surface phenotype consistent with primitive HPCs (CD34+Lin− or CD34+CD38−)(Fig. 5f). CD34+ cells from donors with HIV are Gag+ We obtained samples from HIV–infected people (Supplementary Table 1) and found that we could detect Gag+CD34+ cells in three of seven freshly isolated samples (Supplementary Table 1, Fig. 6a, bottom right panel). When the cells were cultured in GM–CSF–TNF–α, Gag expression could be detected in samples from all seven donors (Fig. 6b,c). In contrast, donor BMMCs specifically depleted of CD34+ cells did not express Gag after culturing (Fig. 6c, d). The addition of the anti–HIV drug raltegravir [which inhibits new in vitro infection in T cells (Supplementary Fig. 7)] partially suppressed the induction of Gag+ cells (Fig 6c. d), confirming that a component of the infection we observed was from viral spread. Similar results were also obtained from a donor (number seven) who had undetectable viral loads for 2 years (Supplementary Table 1, Fig. 6e).

n in T cells (Supplementary Fig. 7)] partially suppressed the induction of Gag+ cells (Fig 6c. d), confirming that a component of the infection we observed was from viral spread. Similar results were also obtained from a donor (number seven) who had undetectable viral loads for 2 years (Supplementary Table 1, Fig. 6e). CD34+ cells from HIV+ donors with undetectable viral loads contain HIV genomic DNA Using a real–time PCR assay for integrated HIV DNA, we detected viral genomes in freshly isolated CD34+ cells from four of nine donors with undetectable viral loads on HAART for more than 6 months (44%) (Fig 6f, g). In these donors, 40 (donor 7), 3.1 (donor 12), 39 (donor 14), and 2.5 (donor 15) HIV genomes per 10,000 CD34+ cells were detected. We detected HIV genomes in BMMCs immunodepleted of CD34+ cells only for donor 12, for whom 1.2 HIV genomes per 10,000 CD34− cells were detected. The limit of detection for this assay varied by donor, but was approximately 1 genome per 10,000 cells due to the limited number of CD34+ cells obtained from each donor. Thus, it is likely that the proportion of donors in which we detected HIV genomes underestimates the percentage of HIV+, HAART–treated individuals harboring integrated HIV genomes in CD34+ cells. DISCUSSION Long–lived cellular reservoirs of latent HIV genomes are a critical obstacle to viral eradication. Here, we demonstrate that HIV can infect hematopoietic progenitor cells in vivo and in vitro to cause an active, cytotoxic infection as well as a latent infection that can be induced to active infection by cytokine treatment.

–lived cellular reservoirs of latent HIV genomes are a critical obstacle to viral eradication. Here, we demonstrate that HIV can infect hematopoietic progenitor cells in vivo and in vitro to cause an active, cytotoxic infection as well as a latent infection that can be induced to active infection by cytokine treatment. Our finding that HIV infects HPCs with an immature phenotype, has clear ramifications for HIV disease because some of these cells may be long–lived and could carry latent HIV for extended periods of time. While further studies are needed to demonstrate that CD34+ stem cells per se are infected, the detection of HIV genomes in HPCs isolated from people effectively treated with HAART for more than 6 months confirms that HIV targets some long–lived HPCs. One might expect these results to predict the presence of identifiable proviral records in differentiated lineages that are known not to be susceptible. However, we show that actively infected HPCs are rapidly killed. Therefore, we expect latently infected HPCs will be killed by viral activation shortly after differentiation is induced. Further studies are now needed to demonstrate that residual circulating virus in individuals on HAART is derived in part from HPCs, as previously demonstrated for resting memory T cells 2. Additionally, studies examining the factors influencing HIV infection and latency in CD34+ cells, as well as limiting dilution experiments to determine the fraction of proviral genomes in these cells that can be reactivated, would further our understanding of this viral reservoir.

nstrated for resting memory T cells 2. Additionally, studies examining the factors influencing HIV infection and latency in CD34+ cells, as well as limiting dilution experiments to determine the fraction of proviral genomes in these cells that can be reactivated, would further our understanding of this viral reservoir. Supplementary Material 1 Acknowledgements Funded by NIH grant RO1 AI051192, MO1–RR000042, the Burroughs Wellcome Foundation, University of Michigan Molecular Mechanisms in Microbial Pathogenesis Training Grant (CCC), Rackham Predoctoral Fellowship (CCC), a National Science Foundation predoctoral fellowship (LAM), and a Bernard Maas Fellowship (LAM). We thank C. van de Ven for umbilical cord blood and C. McIntyre–Ramm for assistance with human studies. The following reagents were obtained through the AIDS Research and Reference Reagent Program, Division of AIDS, NIAID, NIH: p89.6 from R. G. Collman (University of Pennsylvania), pNL4–3 from M. Martin (Viral Pathogenesis and Vaccine Section at NIAID), p94UG–114.1 and pYU-2 from B. Hahn (University of Alabama at Birmingham), and pMJ4 from T. Ndung’u (Nelson Mandela School of Medicine), B. Renjifo (Harvard School of Public Health), and M. Essex (Harvard School of Public Health). pCMV–G was from N. Hopkins, (Massachusetts Institute of Technology), pCMV–HIV–1 and pHIV–7/SF–GFP were from Shin–Jiing–Kuan Yee (City of Hope National Medical Center), and pEGFP–N2–LAMP1 was from Norma Andrews (Yale University). C.C.C., L.A.M. and A.N. conducted in vitro experiments and data analysis and assisted with writing the manuscript

Supplementary Material 1 Acknowledgements Funded by NIH grant RO1 AI051192, MO1–RR000042, the Burroughs Wellcome Foundation, University of Michigan Molecular Mechanisms in Microbial Pathogenesis Training Grant (CCC), Rackham Predoctoral Fellowship (CCC), a National Science Foundation predoctoral fellowship (LAM), and a Bernard Maas Fellowship (LAM). We thank C. van de Ven for umbilical cord blood and C. McIntyre–Ramm for assistance with human studies. The following reagents were obtained through the AIDS Research and Reference Reagent Program, Division of AIDS, NIAID, NIH: p89.6 from R. G. Collman (University of Pennsylvania), pNL4–3 from M. Martin (Viral Pathogenesis and Vaccine Section at NIAID), p94UG–114.1 and pYU-2 from B. Hahn (University of Alabama at Birmingham), and pMJ4 from T. Ndung’u (Nelson Mandela School of Medicine), B. Renjifo (Harvard School of Public Health), and M. Essex (Harvard School of Public Health). pCMV–G was from N. Hopkins, (Massachusetts Institute of Technology), pCMV–HIV–1 and pHIV–7/SF–GFP were from Shin–Jiing–Kuan Yee (City of Hope National Medical Center), and pEGFP–N2–LAMP1 was from Norma Andrews (Yale University). C.C.C., L.A.M. and A.N. conducted in vitro experiments and data analysis and assisted with writing the manuscript J.R. assisted with human subjects M.R.S. and D.B. obtained bone marrow aspirates and assisted with human subjects K.L.C. supervised the project and wrote the manuscript

Supplementary Material 1 Acknowledgements Funded by NIH grant RO1 AI051192, MO1–RR000042, the Burroughs Wellcome Foundation, University of Michigan Molecular Mechanisms in Microbial Pathogenesis Training Grant (CCC), Rackham Predoctoral Fellowship (CCC), a National Science Foundation predoctoral fellowship (LAM), and a Bernard Maas Fellowship (LAM). We thank C. van de Ven for umbilical cord blood and C. McIntyre–Ramm for assistance with human studies. The following reagents were obtained through the AIDS Research and Reference Reagent Program, Division of AIDS, NIAID, NIH: p89.6 from R. G. Collman (University of Pennsylvania), pNL4–3 from M. Martin (Viral Pathogenesis and Vaccine Section at NIAID), p94UG–114.1 and pYU-2 from B. Hahn (University of Alabama at Birmingham), and pMJ4 from T. Ndung’u (Nelson Mandela School of Medicine), B. Renjifo (Harvard School of Public Health), and M. Essex (Harvard School of Public Health). pCMV–G was from N. Hopkins, (Massachusetts Institute of Technology), pCMV–HIV–1 and pHIV–7/SF–GFP were from Shin–Jiing–Kuan Yee (City of Hope National Medical Center), and pEGFP–N2–LAMP1 was from Norma Andrews (Yale University). C.C.C., L.A.M. and A.N. conducted in vitro experiments and data analysis and assisted with writing the manuscript J.R. assisted with human subjects M.R.S. and D.B. obtained bone marrow aspirates and assisted with human subjects K.L.C. supervised the project and wrote the manuscript Appendix METHODS Antibodies CD4: OKT4 conjugated to FITC (Fluorotag, Sigma–Aldrich); CD34: FITC–conjugated (BD Biosciences) or APC–conjugated (Caltag); CD38: PE–Cy7–conjugated (eBioscience); CD133: PE–conjugated (Miltenyi); CCR5: PE–conjugated (eBioscience); CXCR4: PE–conjugated (eBioscience); Gag: clone KC57 conjugated to FITC or PE (Coulter); lineage markers: FITC-conjugated (BD Biosciences) or lineage cocktail (BD Biosciences) conjugated to biotin (EZ Link Sulfo–NHS–Biotin, Pierce Biotechnology) or PE–Cy5.5 (Lightning Link PE–Cy5.5, Innova Biosciences); PLAP (serotec); goat antibody to mouse immunoglobulin: PE–conjugated (Invitrogen).

FITC or PE (Coulter); lineage markers: FITC-conjugated (BD Biosciences) or lineage cocktail (BD Biosciences) conjugated to biotin (EZ Link Sulfo–NHS–Biotin, Pierce Biotechnology) or PE–Cy5.5 (Lightning Link PE–Cy5.5, Innova Biosciences); PLAP (serotec); goat antibody to mouse immunoglobulin: PE–conjugated (Invitrogen). Plasmid constructs p89.6ΔE was generated by dropping out a BsaBI–StuI fragment from p89.6. The 89.6 env expression vector pCDNA–89env was created by cloning the HindIII and EcoRV fragment into pcDNA3.1(+). To construct p89.6–ΔE–SF–GFP, a BamH–KpnI fragment containing the SFFV promoter from pHIV–7/SF–GFP was sub–cloned into pCDNA3.1(+), replacing the original CMV promoter. A BamHI–XbaI fragment from pEGFP–N2–LAMP1 was inserted downstream of the promoter. Next, a nef89.6 PCR product was ligated into MfeI cut pSFDNA–EGFP, generating pNef–SFFV–EGFP. Then a PCR product generated using the primers: 5′–CACCATTATCGTTTCAGACCCT–3 ′ and 5 ′–TCTCGAGTTTAAACTTATAGCAAAGCCCTTTCCA–3′ with p89.6 as a template was ligated into p89.6 (p89.6ΔNE). Finally, a PmeI fragment from pNef–SFFV–EGFP was cloned into PmeI–cut p89.6ΔNE. Standard PCR mutagenesis approaches were used to delete the 3′ U3 region of the viral LTR of p89.6–ΔE–GFP, analogous to those previously described 15. The resulting U3 deletion eliminates all consensus AP–1, NF–AT, NF–κB SP–1 and TATA motifs, but maintains the polyA site found in the R region.

To construct p89.6–ΔE–SF–GFP, a BamH–KpnI fragment containing the SFFV promoter from pHIV–7/SF–GFP was sub–cloned into pCDNA3.1(+), replacing the original CMV promoter. A BamHI–XbaI fragment from pEGFP–N2–LAMP1 was inserted downstream of the promoter. Next, a nef89.6 PCR product was ligated into MfeI cut pSFDNA–EGFP, generating pNef–SFFV–EGFP. Then a PCR product generated using the primers: 5′–CACCATTATCGTTTCAGACCCT–3 ′ and 5 ′–TCTCGAGTTTAAACTTATAGCAAAGCCCTTTCCA–3′ with p89.6 as a template was ligated into p89.6 (p89.6ΔNE). Finally, a PmeI fragment from pNef–SFFV–EGFP was cloned into PmeI–cut p89.6ΔNE. Standard PCR mutagenesis approaches were used to delete the 3′ U3 region of the viral LTR of p89.6–ΔE–GFP, analogous to those previously described 15. The resulting U3 deletion eliminates all consensus AP–1, NF–AT, NF–κB SP–1 and TATA motifs, but maintains the polyA site found in the R region. Cell culture Fresh whole bone marrow aspirates were obtained commercially (AllCells ltd.). Pre-existing umbilical cord blood lacking patient identifiers was obtained following scheduled cesarean section procedures. Bone marrow mononuclear cells (BMMCs) and cord blood mononuclear cells (UCB) were prepared by density separation using Ficoll–Paque (GE healthcare) according to the manufacturer’s instructions.

). Pre-existing umbilical cord blood lacking patient identifiers was obtained following scheduled cesarean section procedures. Bone marrow mononuclear cells (BMMCs) and cord blood mononuclear cells (UCB) were prepared by density separation using Ficoll–Paque (GE healthcare) according to the manufacturer’s instructions. CD34+ cells were prepared from adherence depleted mononuclear cells using commercially available kits [EasySep CD34 positive selection kit, StemCell Technologies and CD34 MACS positive selection, Miltenyi, biotin selection kit with CD34–biotin antibody (Invitrogen)] according to the manufacturer’s instructions. CD133+ cells were isolated similarly, using CD133 positive selection MACS (Miltenyi). Following isolation, CD34+ cells were maintained in StemSpan (StemCell Technologies) or Stemline (Sigma–Aldrich) medium supplemented with CC110 (StemCell Technologies) cytokine cocktail (100 ng ml−1 SCF, 100 ng ml−1 Flt3–L and 100 ng ml−1 TPO). To expand immature progenitor cells for colony formation assays, media was supplemented with STIF cytokine cocktail (CC110 plus 100 ng ml−1 IGFBP–27). For chemical stimulation, washed cells were incubated with 10 ng ml−1 PMA (Sigma–Aldrich) in DMEM supplemented with 10% FBS as described elsewhere 16,17. For cytokine stimulation of HPCs, washed cells were cultured with 100 ng ml−1 rhGM–CSF (R&D systems) and 2.5 ng ml−1 TNF–α (Biolegend). Methylcellulose colony forming assays were conducted according to the manufacturer’s recommendation (Methocult H4034, StemCell Technologies).

Following isolation, CD34+ cells were maintained in StemSpan (StemCell Technologies) or Stemline (Sigma–Aldrich) medium supplemented with CC110 (StemCell Technologies) cytokine cocktail (100 ng ml−1 SCF, 100 ng ml−1 Flt3–L and 100 ng ml−1 TPO). To expand immature progenitor cells for colony formation assays, media was supplemented with STIF cytokine cocktail (CC110 plus 100 ng ml−1 IGFBP–27). For chemical stimulation, washed cells were incubated with 10 ng ml−1 PMA (Sigma–Aldrich) in DMEM supplemented with 10% FBS as described elsewhere 16,17. For cytokine stimulation of HPCs, washed cells were cultured with 100 ng ml−1 rhGM–CSF (R&D systems) and 2.5 ng ml−1 TNF–α (Biolegend). Methylcellulose colony forming assays were conducted according to the manufacturer’s recommendation (Methocult H4034, StemCell Technologies). HIV preparation Infectious supernatants were prepared by transfection of 293T cells using polyethylenimine 17. For pseudotyped and internal promoter viruses, supernatants were concentrated using high–molecular weight polyethylene glycol (PEG) precipitation 18. Pellets were resuspended in StemSpan medium and MOIs were calculated using MOI = −Ln(1−p) where p is the proportion of CEM T cells infected. HIV infections were conducted using a standard spin infection technique for primary cells.

ere concentrated using high–molecular weight polyethylene glycol (PEG) precipitation 18. Pellets were resuspended in StemSpan medium and MOIs were calculated using MOI = −Ln(1−p) where p is the proportion of CEM T cells infected. HIV infections were conducted using a standard spin infection technique for primary cells. Flow cytometry Cells were stained in FACS buffer (2% FBS, 1% human serum, 2 mM HEPES, 0.025% NaN3 and PBS) for 20 min on ice, washed and fixed in 2% paraformaldehyde and PBS. For intracellular Gag staining, cells were then incubated for 5 min in 0.1% Triton–X100 in PBS at room temperature. Washed cells were incubated with antibody to Gag in FACS buffer for 30 min on ice and analyzed on a FACScan or FACSCanto flow cytometer. Cell sorting was performed using a FACSVantage SE in Normal–R mode with a sorted drop envelope of 1.0. qPCR integration assay Genomic DNA was isolated with Qiagen DNeasy Blood and Tissue Kit. Assays were similar to those previously described [25]. One primer was modified to utilize a more conserved sequence (2nd–LTR–F–univ, GTGTIGAAAATCTCTAGCAGTGGC). Copy number per sample was calculated using a standard curve based on ACH–2 cell DNA 19. Isolation of CD34+ cells from HIV–infected donors HIV+ individuals were recruited from The University of Michigan HIV–AIDS Clinic and informed consent was obtained according to a Univeristy of Michigan IRB–approved protocol. 10 ml of marrow was obtained in preservative–free heparin. CD34+ cells were isolated as described above.

of CD34+ cells from HIV–infected donors HIV+ individuals were recruited from The University of Michigan HIV–AIDS Clinic and informed consent was obtained according to a Univeristy of Michigan IRB–approved protocol. 10 ml of marrow was obtained in preservative–free heparin. CD34+ cells were isolated as described above. Figure 1 HIV genomes. (a), (b), (c), (e) and (g) are derived from the molecular clone p89.6. (d) and (f) have been described elsewhere and are derived from HXB and NL4–3 11-13. Expressed viral genes are shown in white, deletions and additions to the genome are shown in black, and non–functional genes are shaded in gray. Figure 2 HIV actively infects HPCs, leading to cell death. (a). Intracellular Gag in BM–derived CD34+ HPCs infected with HIV 89.6ΔEenv89.6 (Fig. 1a) for 3 d. Gray histograms are isotype controls. (b) CD34, MHC-I and intracellular Gag expression in UCB–derived CD34+ HPCs infected with 89.6ΔEenv89.6 for 48 h. (c) Gag, CD34 and Lin staining in BM–derived CD34+ HPCs infected with 89.6. (d) Time course of intracellular Gag expression in UCB–derived CD34+ HPCs infected with 89.6. (e) Annexin V reactivity in UCB–derived CD34+ HPCs infected with 89.6ΔEenv89.6 for 48 h. The right panel shows CD34+Gag− (gray gate and histograms) and CD34+Gag+ cells (black gate and histogram).

CD34+ HPCs infected with 89.6. (d) Time course of intracellular Gag expression in UCB–derived CD34+ HPCs infected with 89.6. (e) Annexin V reactivity in UCB–derived CD34+ HPCs infected with 89.6ΔEenv89.6 for 48 h. The right panel shows CD34+Gag− (gray gate and histograms) and CD34+Gag+ cells (black gate and histogram). Figure 3 HIV infects multipotent HPCs. (a) GFP expression in CD34+ UCB cells infected with HIV–7SF–GFPenv89.6 for 3 d. (b) The percentage of CD133+,CD34+,CD38− cells expressing GFP 3 d post–infection with HIV–7SF–GFPenv89.6. Gray histograms and events represent isotype control staining. (c) and (e), Colony formation by GFP+CD133+ UCB–derived HPCs infected with HIV–7SF–GFPenv89.6(c) or 89.6–SIΔE–GFPenv89.6 (e). (d) and (f), The relative number of colonies formed by equal numbers of sorted GFP+ and GFP−, CD133+ UCBs infected with HIV–7SF–GFPenv89.6 (d) or 89.6–SIΔE–GFPenv89 (f). (erythroid (CFU–E), granulocyte–macrophage (CFU–GM), multi–lineage (CFU–GEMM)).

HPCs infected with HIV–7SF–GFPenv89.6(c) or 89.6–SIΔE–GFPenv89.6 (e). (d) and (f), The relative number of colonies formed by equal numbers of sorted GFP+ and GFP−, CD133+ UCBs infected with HIV–7SF–GFPenv89.6 (d) or 89.6–SIΔE–GFPenv89 (f). (erythroid (CFU–E), granulocyte–macrophage (CFU–GM), multi–lineage (CFU–GEMM)). Figure 4 Induction of HIV from latency. (a) The percentage of BM CD34+ HPCs (See Supplementary Fig. 2b for initial cell purity) expressing an HIV marker gene (PLAP) following infection with HXB–ePLAPenvVSVG plus or minus PMA. (b) qPCR 14 of integrated DNA from BM CD34+ cells infected and cultured as in (a). Pol–minus samples lack polymerase in the first round. Data are displayed as mean relative amount of integrated HIV DNA ± standard deviation (sd), n = 3. (c) The percentage of BM–derived CD34+ HPCs (Supplementary Fig. 2c) expressing intracellular Gag following infection with HXB–ePLAPenv89.6 plus or minus 10 ng ml−1 PMA. (d) Reverse transcriptase activity of CD34+ BM HPCs (Supplementary Fig. 2c) infected with HIV 89.6, plus or minus GM–CSF–TNF–α. The mean ± sd, n = 3 is shown. Control is BMMC immunodepleted for CD34. (e) Intracellular Gag expression 14 d post–infection for BM–derived HPCs infected and cultured as in (d). (f) CD34 and CD83 expression (right panels) after 2 weeks in CC110 or GM–CSF–TNF–α. Isotype–matched controls are shown (Iso–FITC and Iso–PE). (g) Intracellular Gag expression for BM–derived HPCs (Supplementary Fig. 2e) infected with HIV 89.6 and cultured in CC110 cocktail. On day 7, the cells were divided either into CC110 cocktail or GM–CSF–TNF–α. Asterisks indicate Gag reactivity < mock treated cells. (h) Graphical representation of the experiment depicted in part (g).

ular Gag expression for BM–derived HPCs (Supplementary Fig. 2e) infected with HIV 89.6 and cultured in CC110 cocktail. On day 7, the cells were divided either into CC110 cocktail or GM–CSF–TNF–α. Asterisks indicate Gag reactivity < mock treated cells. (h) Graphical representation of the experiment depicted in part (g). Figure 5 Active and latent infection in T cells and HPCs. (a) Gag, GFP and CD4 expression 7 d after infection in CEM–SS cells infected with 89.6–ΔE–SF–GFPenv89.6 or 89.6–ΔE–IRES–GFPenv89.6. Histogram shading corresponds to cell gate. (b) Flow cytometric analysis of PHA–activated PBMC infected with 89.6–ΔE–SF–GFPenv89.6 for 48 h. The histogram is shaded to match the gated cells. In the left panel, the isotype control is shown in gray. (c) Flow cytometric analysis of Jurkat cells infected with 89.6–ΔE–SF–GFPenv89.6 for 7 d, then split into PMA and ionomycin or DMSO control for 48 h. (d) Flow cytometric analysis of UCB–derived CD34+ HPCs infected with 89.6–ΔE–SF–GFPenv89.6 for 3 d. (e) Time course analysis of Gag+ and Gag− GFP+ UCB–derived CD34+ HPCs infected as above and cultured in CC110. (f) Flow cytometric analysis of UCB–derived CD34+ HPCs infected as above and cultured 3 d in CC110 medium. Gag+ cells and Gag−GFP+ cells were gated on the left plot and overlaid (black dots) on plots of CD34 vs. Lin (middle panels) or CD38 plots (right panels). The grey background shows the total population.

in CC110. (f) Flow cytometric analysis of UCB–derived CD34+ HPCs infected as above and cultured 3 d in CC110 medium. Gag+ cells and Gag−GFP+ cells were gated on the left plot and overlaid (black dots) on plots of CD34 vs. Lin (middle panels) or CD38 plots (right panels). The grey background shows the total population. Figure 6 Active and inducible infection in HPCs from HIV+ people. (a) HIV–1 Gag expression in freshly isolated adherence–depleted Lin−CD34+CD133+ BMMCs. The middle panel shows background staining using an isotype control for the Gag antibody only. (b) CD34 and intracellular Gag expression in CD34+ cells stained immediately or after culturing (14 d). Control shows background staining with an isotype control antibody. (c) Gag expression before and after culturing in GM–CSF–TNF–α plus or minus raltegravir for donors 1-6. (d) Summary graph of Gag induction plus or minus raltegravir. Fold induction = (% Gag+ in cultured cells) ÷ (initial % Gag+)]. Mean ± sd is shown. (e) Intracellular Gag expression in Donor 7 CD34+ or CD34–immunodepleted BMMCs cultured as described in c. [GM (GM–CSF–TNF–α; GMR (GM–CSF–TNF–α plus raltegravir)]. (f) Real–time PCR of HIV genomes ng−1 DNA isolated from fresh CD34+ or immunodepleted BMMCs. Mean ± sd is shown, n = 3. (g) Real–time PCR of HIV genomes from donor CD34+ or immunodepleted cells. The limit of detection was approximately one HIV genome per 10,000 cells. Means ± sd, n = two independent experiments with three replicates each.

Breast cancer recurs at distant sites in a significant number of women who receive adjuvant chemotherapy after surgical removal of the primary breast tumor1. De novo resistance mechanisms present within tumor cells prior to treatment are key factors leading to failure of chemotherapeutic drugs to prevent metastatic recurrence. Therefore, discovery of the genomic alterations and genes contributing to de novo chemo-resistance to specific drugs is an important goal2. Although a number of multidrug resistance genes have been discovered, their over-expression is often induced during drug treatment3,4 and not useful for initial guidance of drug selection. Gene signatures generated from drug responses of tumor cell lines are reported to predict drug response in patients5–7; however, others found cell-line derived signatures are not predictive of response in clinical cases8. Repeatedly observed genomic gains or losses have identified genomic regions that may harbor genes contributing to malignant behavior and poor outcome9–12. Which genomic region(s) harbors genes that may contribute to de novo resistance to therapy is currently unknown.

s are not predictive of response in clinical cases8. Repeatedly observed genomic gains or losses have identified genomic regions that may harbor genes contributing to malignant behavior and poor outcome9–12. Which genomic region(s) harbors genes that may contribute to de novo resistance to therapy is currently unknown. We analyzed gene expression profiles of 115 breast carcinomas from women diagnosed between 2000 and 2003 and treated according to current guidelines including adjuvant chemotherapy if indicated. We performed predictive analysis of microarrays (PAM; 13) and identified 114 probes, encoding 75 known genes, differentially expressed between cases with early distant metastatic recurrence and cases without distant recurrence (Supplementary Table 1). Fifteen percent of these probes, corresponding to 12 different genes, mapped to chromosome 8q22, the only chromosomal region with statistically significant enrichment (P < 2.1e-09) of probes associated with metastatic recurrence (Fig. 1a). We applied Cox proportional hazard regression 13,14 which also demonstrated differential over-expression of these 8q22 genes in tumors with distant recurrence. These genes included CCNE2 and MTDH which are reported associated with metastatic recurrence and poor prognosis of breast cancer 15,16. The coordinate over-expression of neighboring genes often reflects chromosomal amplification. Indeed, 8q22 amplification was observed by SNP array analysis in 50 breast cancers (Supplementary Fig. 1) and expression of the 8q22 genes correlated with DNA copy number (Supplementary Table 1).

gnosis of breast cancer 15,16. The coordinate over-expression of neighboring genes often reflects chromosomal amplification. Indeed, 8q22 amplification was observed by SNP array analysis in 50 breast cancers (Supplementary Fig. 1) and expression of the 8q22 genes correlated with DNA copy number (Supplementary Table 1). We confirmed 8q22 amplification by DNA interphase fluorescence in situ hybridization (FISH) (Fig. 1b,c) and found it in 21% of 85 breast cancers. Degree of copy gain was correlated with average expression of the 12 recurrence-associated 8q22 genes (8q gene expression index, 8qEI) (Fig. 1d and Supplementary Table 1). Kaplan-Meier analysis showed 8q22 amplification was associated with reduced metastasis-free survival in the entire cohort evaluated by FISH (Fig. 1e), in the ER− cases (Supplementary Fig. 2a), and in the women who had received anthracycline-based adjuvant chemotherapy (Fig. 1f). In multivariate analysis, amplification of 8q22 was a strong independent prognostic factor for breast cancer recurrence (Supplementary Table 2).

the entire cohort evaluated by FISH (Fig. 1e), in the ER− cases (Supplementary Fig. 2a), and in the women who had received anthracycline-based adjuvant chemotherapy (Fig. 1f). In multivariate analysis, amplification of 8q22 was a strong independent prognostic factor for breast cancer recurrence (Supplementary Table 2). We sought validation in a meta analysis of six independent cohorts annotated with treatment and outcome 12,14,17–20. Kaplan-Meier analysis demonstrated a significant difference in disease-free survival between 8qEI low-expression and high-expression groups in either chemo-treated (Fig. 1g) or untreated cases (Supplemental Fig.2b). These results indicate that 8q22 amplification promotes over-expression of 8q22 genes in tumor tissue, which are associated with poor prognosis in untreated cases and inferior disease-free survival despite adjuvant chemotherapy.

pression groups in either chemo-treated (Fig. 1g) or untreated cases (Supplemental Fig.2b). These results indicate that 8q22 amplification promotes over-expression of 8q22 genes in tumor tissue, which are associated with poor prognosis in untreated cases and inferior disease-free survival despite adjuvant chemotherapy. To determine if 8q22 genes influence sensitivity to chemotherapy, we treated the breast cancer cell line BT549 harboring 8q gain with siRNA against the 12 candidate genes (Supplementary Fig. 3) and screened for alteration of sensitivity to chemotherapeutic drugs (Fig. 2a). Depletion of two genes significantly increased the sensitivity to anthracyclines (Fig. 2a). One of these, YWHAZ, codes for a known anti-apoptotic protein 14-3-3ζ 21,22. The second gene codes a novel lysosomal protein LAPTM4B (Lysosomal Associated Protein Transmembrane 4B), about which little is known in breast cancer. We examined 16 breast cancer cell lines and found a strong positive correlation between higher endogenous LAPTM4B mRNA level and higher IC50 (relative resistance) to anthracyclines (P < 0.00034, Supplementary Fig. 4a), a weaker or no correlation with IC50 to cisplatin and paclitaxel (P = 0.008 and 0.4, respectively; data not shown). The expression of YWHAZ in cell lines also correlated with the IC50 to doxorubicin (Supplementary Fig 4). Specific knockdown of LAPTM4B and YWHAZ in three cell lines (Fig. 2b) increased sensitivity to the anthracyclines doxorubicin (Fig. 2c upper panels) and daunorubicin (data not shown), but had weaker or no effect on sensitivity to cisplatin and paclitaxel (Fig. 2c lower panels). Knockdown of either LAPTM4B or YWHAZ significantly increased doxorubicin-induced apoptosis (Fig. 2d). Induction of apoptosis was less apparent in response to cisplatin treatment (Fig. 2d).

) and daunorubicin (data not shown), but had weaker or no effect on sensitivity to cisplatin and paclitaxel (Fig. 2c lower panels). Knockdown of either LAPTM4B or YWHAZ significantly increased doxorubicin-induced apoptosis (Fig. 2d). Induction of apoptosis was less apparent in response to cisplatin treatment (Fig. 2d). To investigate mechanism, intracellular localization of anthracyclines was tracked by following the autofluorescence of doxorubicin. LAPTM4B expression in cell lines correlated with both IC50 of doxorubicin (Fig. 3a) and with appearance of doxorubicin in the nucleus within 24 hours (Fig. 3b). Knockdown of LAPTM4B by siRNA in MDA-MB-231 cells resulted in a significant increase in nuclear localization of doxorubicin, detectable within 12 h to 24 h of treatment, maximal at 24 h to 36 h (12 h after withdrawal of drug), and sustained at 48 h (Fig. 3c). Knockdown of LAPTM4B in BT549 cells resulted in a similar increase in nuclear localization at 24 h (data not shown). Decreased distribution of doxorubicin into the nucleus is associated with lower phospho-H2AX (Fig. 3c,d), a marker of DNA damage response, and consistent with reduced doxorubicin-induced apoptosis and increased IC50 (Fig. 2). YWHAZ levels correlated with IC50 of doxorubicin (Fig. 3a) and knockdown of YWHAZ increased slightly the phospho-H2AX levels in drug treated cells (Fig. 3d). These results suggest one mechanism of poorer outcome for women harboring breast cancers with 8q22 amplification is increased expression of LAPTM4B and interference with nuclear accumulation of anthracyclines. YWHAZ may effect drug sensitivity through inhibition of apoptosis, consistent with reports of others21,22.

d). These results suggest one mechanism of poorer outcome for women harboring breast cancers with 8q22 amplification is increased expression of LAPTM4B and interference with nuclear accumulation of anthracyclines. YWHAZ may effect drug sensitivity through inhibition of apoptosis, consistent with reports of others21,22. We introduced HA-tagged full length LAPTM4B and YWHAZ into partially transformed human mammary epithelial cells (HMECs) 23. Expression of either exogenous LAPTM4B or YWHAZ increased the IC50 of doxorubicin (270% increase, P < 0.002 and 394% increase, P = 0.0001, respectively) but had no significant effect on sensitivity to paclitaxel or cisplatin (Fig. 3e, f). The LAPTM4B-induced decrease in drug sensitivity parallels and is consistent with delayed appearance of anthracycline in the nucleus of LAPTM4B-overexpressing HMEC (Fig. 3g). Kaplan Meier analysis of women treated with adjuvant chemotherapy showed that the expression of YWHAZ and LAPTM4B above median level was associated with shorter disease-free survival (Supplementary Fig. 2c). The association with poor outcome after adjuvant chemotherapy is consistent with either a prognostic effect or a role of these two genes in chemotherapy resistance.

chemotherapy showed that the expression of YWHAZ and LAPTM4B above median level was associated with shorter disease-free survival (Supplementary Fig. 2c). The association with poor outcome after adjuvant chemotherapy is consistent with either a prognostic effect or a role of these two genes in chemotherapy resistance. Finally, LAPTM4B and YWHAZ were tested for their association with response to anthracyclines in a neoadjuvant (pre-operative) treatment trial of epirubicin monotherapy. The average of LAPTM4B and YWHAZ expression levels from expression array data of pretreatment tumor biopsies was evaluated for association with pathologic complete response (pCR) to epirubicin. The two-gene expression levels displayed a coherent pattern with higher levels of expression associated with absence of pCR (presence of residual disease) after epirubicin treatment (Fig. 4a, b). We evaluated the capability of the two genes to predict pCR by measuring the area under receiver operator characteristic (ROC) curves24,25(AUC); which demonstrated higher expression of the two genes is associated with absence of pCR after anthracycline chemotherapy in the cohort of 118 breast tumors (AUC 0.315, p < 0.00058, Fig. 4d). The association is more significant in 87 ER− HER2− tumors (AUC 0.241, p < 0.000062, Fig. 4e). When the two genes were analyzed separately, both genes were significantly predictive of poor response in the whole cohort and in the ER− HER2− subset, but only LAPTM4B level was predictive in the ER− HER2+ subset (Supplementary Table 3). In contrast, the expression of these two genes was not associated with treatment response to cisplatin monotherapy in a separate neo-adjuvant clinical trial in triple negative (ER− PR− HER2−) cancer cases26 (AUC 0.675, P > 0.3, Fig. 4c,f; Supplementary Table 3). Although MTDH, one of the twelve 8q22 genes, was reported to induce chemo-resistance to a broad spectrum of drugs in experimental models16, its expression was not predictive for pCR in either the epirubicin or cisplatin trials (data not shown). We analyzed a third single agent neoadjuvant therapy trial of predominantly ER+ tumors treated with docetaxel27 and found that higher levels of expression of LAPTM4B and YWHAZ were not associated with an inferior clinical response to therapy (data not shown). The results support the notion that LAPTM4B and YWHAZ over-expression is associated preferentially with poor response to anthracyclines.

ER+ tumors treated with docetaxel27 and found that higher levels of expression of LAPTM4B and YWHAZ were not associated with an inferior clinical response to therapy (data not shown). The results support the notion that LAPTM4B and YWHAZ over-expression is associated preferentially with poor response to anthracyclines. LAPTM4B is similar to its family member LAPTM4A that promotes selective resistance to anthracyclines and not cisplatin in Saccharomyces cerevisiae 28. Our results show that LAPTM4B acts on anthracycline trafficking by reducing drug entry into the nucleus and decreasing drug-induced DNA damage. Higher YWHAZ expression protects cells from drug-induced apoptosis. Because they reside in proximity, amplification produces coordinated up-regulation of their various functions, together resulting in preferential resistance to anthracyclines. The results from three clinical trials support this contention, and suggest clinical options for the treatment of primary breast cancers might depend upon the status of 8q22 amplification and over-expression of these two genes in tumors. Anthracyclines appear to be reasonable treatment in tumors without 8q22 amplification, and alternatives might be selected for those whose cancers harbor amplification.

options for the treatment of primary breast cancers might depend upon the status of 8q22 amplification and over-expression of these two genes in tumors. Anthracyclines appear to be reasonable treatment in tumors without 8q22 amplification, and alternatives might be selected for those whose cancers harbor amplification. Microarray data Microarray data sets are deposited in the NCBI GEO database under the following accession numbers: Gene expression data from the DF/HCC cases, GSE19615; SNP array data from 50 of the DF/HCC cases, GSE19594; gene expression data from neo-adjuvant epirubicin “Trial of Principle”, GSE16446; gene expression data from neo-adjuvant cisplatin trial, GSE18864. Methods Cohort Primary breast tumors of 115 subjects were obtained from the NCI-Harvard Breast SPORE blood and tissue repository under Dana-Farber/Harvard Cancer Center Institutional Review Board approved protocols, with informed consent from subjects. Affymetrix U133 plus 2 gene expression array analysis was performed as described 29,30 A subset of 85 tumors were represented in tissue microarrays and used for FISH analysis (see Supplementary Methods). Fifty of the cases were analyzed by Affymetrix 10K SNP array as described31,32. A portion of the SNP and gene expression data were reported previously 29,30,31,32. Clinical and pathologic characteristics for each sample in the cohort are provided in Supplementary Table 4.

for FISH analysis (see Supplementary Methods). Fifty of the cases were analyzed by Affymetrix 10K SNP array as described31,32. A portion of the SNP and gene expression data were reported previously 29,30,31,32. Clinical and pathologic characteristics for each sample in the cohort are provided in Supplementary Table 4. Neoadjuvant Clinical Trials The neoadjuvant “Trial of Principle” for breast cancer is conducted in European hospitals and coordinated at the Institut Jules Bordet. This trial is registered on the clinical trials site of the US National Cancer Institute website http://clinicaltrials.gov/ct2/show/NCT00162812?term=NCT00162812&rank=1. Single agent epirubicin was given as neo-adjuvant (pre-operative) chemotherapy to 118 reportedly ER− cases. After central review, 4 of the cases were found to be ER+. Of the remaining 114 ER−, 87 cases were classified as HER2− based on low ERBB2 module score33. Pretreatment core biopsy of the primary breast tumor was performed for diagnosis and RNA isolation. At completion of chemotherapy, pathologic response was determined by microscopic examination of the excised tumor and nodes. Pathological complete response (pCR) was defined by the absence of residual invasive breast carcinoma in the breast and axillary nodes. This study has been approved by the medical ethics committee of Institute Jules Bordet and all women given written informed consent prior to study entry. Gene expression data of U133plus 2 were generated from RNA of pretreatment core biopsies.

ence of residual invasive breast carcinoma in the breast and axillary nodes. This study has been approved by the medical ethics committee of Institute Jules Bordet and all women given written informed consent prior to study entry. Gene expression data of U133plus 2 were generated from RNA of pretreatment core biopsies. The trial of single agent cisplatin given as neo-adjuvant chemotherapy to women with triple negative breast cancer26. Gene expression array data from pre-treatment biopsies was available for 24 cases26. Pathological response was determined by microscopy examination after chemotherapy as described above. ROC (Receiver Operating Characteristic) curve analysis was performed to evaluate the mean level of combined LAPTM4B and YWHAZ expression, or the levels of each individual gene, for their capacity to predict pathological complete response (pCR). The association of ranked gene expression levels with pCR was evaluated by determining the area under the curve (AUC) estimated through the concordance index34; the corresponding p-value is from one-sided Wilcoxon's rank test35. The ROC curves are plotted for prediction of pCR, so a curve below the midline in low-right area of the graph indicates the ranked gene expression is associated with absence of pCR.

area under the curve (AUC) estimated through the concordance index34; the corresponding p-value is from one-sided Wilcoxon's rank test35. The ROC curves are plotted for prediction of pCR, so a curve below the midline in low-right area of the graph indicates the ranked gene expression is associated with absence of pCR. Statistical analysis of gene expression arrays Affymetrix U133plus2.0 array data from tumor samples of 115 cases were classified into those with distant recurrence within 36 months of diagnosis or those without distant recurrence and at least 36 months of follow up. The 115 arrays were log-transformed, normalized and invariable genes were removed by filtering using dChip software (p-value < 0.05). Differentially expressed genes were determined by Prediction Analysis of Microarrays (PAM) 13 using the pamr package implemented in R language (http://cran.r-project.org/web/packages/pamr/index.html). PAMR implements the shrunken nearest centroid method 13. Genes were selected at a false discovery threshold that minimized a 10-fold cross-validation and test errors near the shrinkage parameter Δ = 2. The PAM score indicates the degree of statistical association for expression of each gene and metastatic recurrence. Cox proportional hazard regression analysis 13,14 was used to discover differentially expressed genes associated with time-to-recurrence.

ss-validation and test errors near the shrinkage parameter Δ = 2. The PAM score indicates the degree of statistical association for expression of each gene and metastatic recurrence. Cox proportional hazard regression analysis 13,14 was used to discover differentially expressed genes associated with time-to-recurrence. Analysis in independent cohorts Pooling six independent gene expression array datasets 12,14,17–20 for analysis was performed as described 36. 8q22 gene expression-based Index (8qEI) is calculated as the mean expression value of the twelve 8q22 genes identified in PAM analysis, and was scored in each of the six independent cohorts. We defined a simple median-based classifier: 8qEI level higher than the median as 8qEI high and lower than the median as 8qEI low. The classification was performed in 6 cohorts separately, and then combined into one data sheet with 1348 samples. Of these, 1130 had annotation for treatment and followup, 361 cases received adjuvant chemotherapy with or without hormonal therapy and 725 cases received no adjuvant hormonal or chemotherapy. Follow-up time was constrained to a maximum of 10 years with annotation for disease free survival. Kaplan-Meier analyses were carried out using the survival package within the R statistical package. P-values are derived using Mantel-Cox logrank test. Additional methods are provided in supplementary material online.

Analysis in independent cohorts Pooling six independent gene expression array datasets 12,14,17–20 for analysis was performed as described 36. 8q22 gene expression-based Index (8qEI) is calculated as the mean expression value of the twelve 8q22 genes identified in PAM analysis, and was scored in each of the six independent cohorts. We defined a simple median-based classifier: 8qEI level higher than the median as 8qEI high and lower than the median as 8qEI low. The classification was performed in 6 cohorts separately, and then combined into one data sheet with 1348 samples. Of these, 1130 had annotation for treatment and followup, 361 cases received adjuvant chemotherapy with or without hormonal therapy and 725 cases received no adjuvant hormonal or chemotherapy. Follow-up time was constrained to a maximum of 10 years with annotation for disease free survival. Kaplan-Meier analyses were carried out using the survival package within the R statistical package. P-values are derived using Mantel-Cox logrank test. Additional methods are provided in supplementary material online. Supplementary Material 1 2 3 ACKNOWLEDGMENTS We thank Charles Lee, Abha Aggarwal (DF/HCC cytogenetics core facility); Edward Fox, Pamela Hollasch, Maura Berkeley (DF/HCC microarray core facility); Rebecca Gelman, Weixiu Luo, Xin Lu (Biostatisticians); and members of the Richardson-Wang lab for their advice and assistance. We also thank Jean Zhao for kindly providing vectors and cell lines, Aron C. Eklund, Nicolai Juul, and Rou-Li Zhou for helpful discussions and advice, and Daniel Silver and Ji-Young Kim for their critical review of this manuscript. Human mammary epithelial cells (HMECs) carrying a dominant negative allele of p53 were generously provided by Dr. Jean Zhao, Dana-Farber Cancer Institute, Boston, MA. This work was supported by the Breast Cancer Research Foundation (BCRF) in New York. Support also came from the NCI SPORE in Breast Cancer at Harvard (CA89393) and a DOD Concept Award (BC053041). The Trial of Principle (TOP) trial was supported by the Fondation Luxembourgeoise contre le Cancer, by the Fonds National de la Recherche Scientifique (CS, BH, CD), by the Brussels Region, as well as by the European Commission through the “Advancing Clinico-Genomic Trials” project (FP6-2005-IST-026996).

ard (BC053041). The Trial of Principle (TOP) trial was supported by the Fondation Luxembourgeoise contre le Cancer, by the Fonds National de la Recherche Scientifique (CS, BH, CD), by the Brussels Region, as well as by the European Commission through the “Advancing Clinico-Genomic Trials” project (FP6-2005-IST-026996). AUTHOR CONTRIBUTIONS A.L.R. and Z.C.W. designed the experiment and supervised the project. Y.L. (Yang Li) performed the in vitro laboratory experiments including generating cDNA vector constructs, gene transfer and knockdown, RT-PCR and Western blot analysis, and drug sensitivity and localization studies. L.Z. performed the PAM data analysis and statistics. Q.L., under supervision of Z.S., performed bioinformatic analysis on validation data sets and the cisplatin trial data set. R.T. performed apoptosis assays. Y.L. (Yan Li) contributed to the preparation of cDNA vector constructs. C.D., C.S., and B.H.-K. performed the epirubicin trial and provided clinical data and analysis. A.L.R. provided the DFHCC and cisplatin trial clinical samples and performed gene expression array analysis. Z.C.W. performed the SNP array analysis and scored the FISH assays. A.L.R., J.D.I. and Z.C.W. wrote the manuscript with comments from Y.L.(Yang Li), C.D., C.S., and Z.S.

and provided clinical data and analysis. A.L.R. provided the DFHCC and cisplatin trial clinical samples and performed gene expression array analysis. Z.C.W. performed the SNP array analysis and scored the FISH assays. A.L.R., J.D.I. and Z.C.W. wrote the manuscript with comments from Y.L.(Yang Li), C.D., C.S., and Z.S. Figure 1 8q22 amplification, gene expression and cancer recurrence (a) Values on the ordinate are the PAM score. Genes with significant PAM scores are displayed according to their annotated location on chromosome 8 (b) metaphase FISH in normal human lymphoid cells using 8q22 probe RP11-347C18 (orange) (c) interphase FISH with 8q22 probe RP11-347C18 (orange), chromosome 8 centromere probe (green), and DAPI nuclear stain (blue) in a human tumor sample (d) Correlation between 8q22 copy number detected by FISH and the 8q22 twelve gene expression-based index (8qEI) (regression R = 0.65). (e) Kaplan-Meier analysis for distant recurrence-free survival (DRFS) of 85 Boston cases with or without 8q22 amplification identified by FISH (Cox Hazard Ratio (HR) = 7.77). (f) DRFS for 52 Boston cases treated with doxorubicin and cyclophosphamide, according to 8q22 amplification identified by FISH (HR = 7.66). (g) Disease free survival (DFS) for 361 patients from six independent cohorts who received adjuvant chemotherapy, according to 8qEI high and low based on 8qEI levels above or below median level in each of the six cohorts (HR = 1.9).

n and cyclophosphamide, according to 8q22 amplification identified by FISH (HR = 7.66). (g) Disease free survival (DFS) for 361 patients from six independent cohorts who received adjuvant chemotherapy, according to 8qEI high and low based on 8qEI levels above or below median level in each of the six cohorts (HR = 1.9). Figure 2 Knockdown of 8q22 genes by siRNA in tumor cell lines to determine effect on sensitivity to anthracycline chemotherapy (a) siRNA knockdown of the twelve 8q22 genes in BT549 breast tumor cells. Bars indicate the ratio of IC50 for the indicated drugs in gene-specific siRNA treated cells, relative to control siRNA treated cells. Percent reduction in IC50 is indicated as follows: ** 63% reduction, *** 82%, **** >85%. (b) Relative mRNA levels for LAPTM4B and YWHAZ in the three breast cancer cell lines, after treatment with control or specific siRNAs as indicated. Levels are log 2 scale and relative to HMEC reference sample. (c) Drug concentration-dependent cell survival curves for doxorubicin (upper panels), cisplatin (middle panels), and paclitaxel (lower panels) in cells transfected with control siRNA or gene-specific siRNAs, as indicated. The cell lines are shown at top. Approximate percent reduction in IC50 is indicated as follows: *** 85%, ** 75%, and * 50%. (d) Western blot for pro-caspase 3 (pr-cas3), active caspase 3 (ac-cas) and cleaved lower molecular weight PARP (cl-PARP), in MB231 and BT549 cells transfected with siRNA for LAPTM4B (L), YWHAZ (Y), or scramble control (C) as indicated across the top. Cells were treated with carrier (no marker) or with doxorubicin or cisplatin (+) for 48 hours, as indicated above each lane.

pase 3 (ac-cas) and cleaved lower molecular weight PARP (cl-PARP), in MB231 and BT549 cells transfected with siRNA for LAPTM4B (L), YWHAZ (Y), or scramble control (C) as indicated across the top. Cells were treated with carrier (no marker) or with doxorubicin or cisplatin (+) for 48 hours, as indicated above each lane. Figure 3 LAPTM4B expression, intracellular doxorubicin distribution, and drug-induced DNA damage in breast cancer cell lines (a) mRNA levels of LAPTM4B and YWHAZ relative to a reference sample (log2 scale, left axis) and doxorubicin IC50 concentration (µM, right axis) in three breast cancer cell lines, as indicated below the bars. (b, c, and g) Merged fluorescence analysis for doxorubicin (drug autofluorescence, red) and nuclear staining (DAPI, blue). Doxorubicin localized in the nucleus appears purple. Scale bars indicate 10 µm. (b) Intracellular doxorubicin distribution after 24 h of drug exposure in three breast tumor cell lines. (c) Intracellular doxorubicin distribution in MDA-MB-231 cells transfected with control siRNA and LAPTM4B-specific siRNA as indicated to the right. Time points during drug exposure and after removal of drug from culture medium are shown above the panels. (d) Western blot with anti-phospho-H2AX antibody and β-actin control antibody in lysates of MDA-MB-231 cells transfected with control and gene-specific siRNA oligonucleotides, as indicated along the left side, at the indicated time points of drug exposure and after removal of drug from culture medium. (e) Plot of inhibition of cell growth by doxorubicin, paclitaxel and cisplatin in HMEC cells transfected with vectors containing GFP control, LAPTM4B-HA, or YWHAZ-HA. The percent viable cells, compared to transfected cells without drug treatment, are indicated on the Y-axis (mean of triplicates ± S.D.). ** indicates >250% increase in IC50. (f) Western blot of HMEC cells transfected with vectors containing LacZ control, LAPTM4B-HA and YWHAZ-HA, as indicated above, using HA tag antibody and β-actin antibody. (g) Intracellular doxorubicin distribution in HMEC cells transfected with LacZ control vector or with LAPTM4B-HA vector after 16 hours of doxorubicin exposure.

rn blot of HMEC cells transfected with vectors containing LacZ control, LAPTM4B-HA and YWHAZ-HA, as indicated above, using HA tag antibody and β-actin antibody. (g) Intracellular doxorubicin distribution in HMEC cells transfected with LacZ control vector or with LAPTM4B-HA vector after 16 hours of doxorubicin exposure. Figure 4 LAPTM4B and YWHAZ expression and pathologic complete response (pCR) to neoadjuvant chemotherapy (a, b, c) Cases ranked according to mean sum expression of LAPTM4B and YWHAZ and (d, e, f) Receiver Operating Characteristic plots for the performance of the LAPTM4B and YWHAZ genes to predict pCR for (a, d) epirubicin in the whole breast cancer cohort (pCR, n = 17; no pCR, n = 101), (b, e) epirubicin in the ER− and HER2− sub-cohort (pCR, n = 9; no pCR, n = 78), and (c, f) cisplatin in ER− PR− HER2− breast cancer patients (pCR, n = 4; no pCR, n = 20). In the ROC plots (e, f, g), the solid diagonal lines indicate the performance of a random predictor and the dashed line indicates the performance of the mean LAPTM4B and YWHAZ levels to predict pCR.

Introduction The gastro-intestinal tract is constantly exposed to a vast number of commensal bacteria and their inflammatory products. Essential to intestinal homeostasis are pattern recognition receptors (PRR) such as TLR1. Engagement of TLR with their cognate ligands in the intestinal mucosa provokes the production of pro-inflammatory, pro-angiogenic and growth factors that support IEC differentiation and proliferation2. In a genetically susceptible host, an on-going intestinal inflammation provokes an uncontrolled growth of IEC leading to neoplasia3,4,5. Likewise, it was proposed that signaling through TLR regulates IEC tumor development, in mice heterozygous for a mutant form of the tumor suppressor gene, adenomatous polyposis coli (Apc)6. However, the molecular mechanisms and its relationship to intestinal inflammation have not been identified. The Apcmin/+ mouse is an animal model of human familial adenomatous polyposis7. These mice develop multiple intestinal neoplasia (Min), after they lose the heterozygote wild type Apc allele and consequently die when they reach 6 months of age8.

s and its relationship to intestinal inflammation have not been identified. The Apcmin/+ mouse is an animal model of human familial adenomatous polyposis7. These mice develop multiple intestinal neoplasia (Min), after they lose the heterozygote wild type Apc allele and consequently die when they reach 6 months of age8. The survival and growth of certain tumors are dependent on the continued activation of certain oncogenes. This phenomenon that was termed “oncogene addiction”, explains tumor suppression due to the inactivation of a single gene product9. The oncogene c-myc is critical for Apc-mediated tumorigenesis10,11. The genetic deletion of c-myc results in the inhibition of tumor growth 11 and as low as a two-fold reduction in c-myc expression in IEC is sufficient to inhibit tumorigenesis in Apcmin/+ mice12-14. Here we identified that a MyD88-dependent activation of ERK in IEC is essential to drive intestinal tumor growth in Apcmin/+ mice. Consequently, the inhibition of pERK abrogates the Min phenotype in these animals.

wo-fold reduction in c-myc expression in IEC is sufficient to inhibit tumorigenesis in Apcmin/+ mice12-14. Here we identified that a MyD88-dependent activation of ERK in IEC is essential to drive intestinal tumor growth in Apcmin/+ mice. Consequently, the inhibition of pERK abrogates the Min phenotype in these animals. Results MyD88 signaling is essential for polyp growth in Apcmin/+ mice TLRs signal mainly through either MyD88 or TRIF. To explore the potential impact of TLR signaling on IEC tumors we crossed Apcmin/+ mice to Myd88-/- or TrifLps2/Lps2 (Lps2) mice 15. The average survival was 23 weeks for Apcmin/+ mice and 28 weeks for Apcmin/+/Lps2 mice. In contrast, all of the Apcmin/+/Myd88-/- mice survived the 45-week study (Supp. Fig. 1A). We then determined the role of each adapter protein on tumor (polyp) formation at 20 weeks of age. Apcmin/+/Myd88-/- mice had fewer polyps throughout the small and large intestines compared to Apcmin/+ or Apcmin/+/Lps2 mice (Supp. Fig. 1B and Supp. Fig. 1C), but they displayed circular raised lesions (microadenomas) in both the distal small intestine (DSI) and the colon (Supp. Fig. 1C-1F).

mation at 20 weeks of age. Apcmin/+/Myd88-/- mice had fewer polyps throughout the small and large intestines compared to Apcmin/+ or Apcmin/+/Lps2 mice (Supp. Fig. 1B and Supp. Fig. 1C), but they displayed circular raised lesions (microadenomas) in both the distal small intestine (DSI) and the colon (Supp. Fig. 1C-1F). MyD88 signaling enhances IEC proliferation and suppresses IEC apoptosis in Apcmin/+ mice As the polyps in the Apcmin/+/Myd88-/- mice failed to grow (Supp. Fig. 1C), we investigated whether the deletion of Myd88 affected IEC proliferation. The proliferation and the migration rate of IEC along the crypt-villus axis, as analyzed by BrdU incorporation, were decreased as compared to those in Apcmin/+ mice (Fig. 1A). We also observed a significantly higher number of apoptotic IEC in Apcmin/+/Myd88-/- (Fig. 1B) as well as increased levels of cleaved poly(ADP-ribose) polymerase (PARP), a substrate of caspase-316 (Fig. 1C). Taken together, these data indicate that TLR signaling via MyD88 enhances IEC proliferation and inhibits IEC apoptosis, and suggest that these two effects synergize in enhancing IEC tumor growth in the Apcmin/+ mice.

s increased levels of cleaved poly(ADP-ribose) polymerase (PARP), a substrate of caspase-316 (Fig. 1C). Taken together, these data indicate that TLR signaling via MyD88 enhances IEC proliferation and inhibits IEC apoptosis, and suggest that these two effects synergize in enhancing IEC tumor growth in the Apcmin/+ mice. Myd88 signaling in IEC, but not in hematopoietic cells, controls IEC tumor growth in Apcmin/+ mice In the intestinal mucosa, both IEC and bone marrow (BM)-derived cells have functional TLR that utilize MyD88 for signaling 17,18,19. To further identify the role of BM-derived cells in IEC tumorigensis, we generated BM chimeras: both Apcmin/+ and Apcmin/+/Myd88-/- recipients were reconstituted with BM harvested from either WT or Myd88-/- donors 20. Reconstitution of Apcmin/+ recipients with either Myd88-/- or WT BM did not significantly alter polyp count and growth in either the DSI or the colon. Similarly, the number of polyps did not significantly change in Apcmin/+/Myd88-/- recipients reconstituted with WT or Myd88-/- BM (Fig. 2A). These results indicate that polyp growth in Apcmin/+ mice does not depend on TLR-MyD88 signaling in BM-derived cells and highly suggests its dependence on TLR-MyD88 activation of IEC.

larly, the number of polyps did not significantly change in Apcmin/+/Myd88-/- recipients reconstituted with WT or Myd88-/- BM (Fig. 2A). These results indicate that polyp growth in Apcmin/+ mice does not depend on TLR-MyD88 signaling in BM-derived cells and highly suggests its dependence on TLR-MyD88 activation of IEC. To explore whether host-derived or microbial-derived TLR ligands play a role in IEC tumorigenesis, we crossed Apcmin/+ mice with Il1r1-/- or with Caspase1-/- mice, which are limited in processing IL-1 and IL-18 21. As presented in Fig. 2B, there was no significant difference in the numbers of polyps in Apcmin/+/Il1r1-/- or Apcmin/+/Caspase1-/- as compared to Apcmin/+ mice. In addition, administration of the IL-1R antagonist, Anakinra, did not affect the extent of IEC tumorigenesis in Apcmin/+ mice (Fig. 2C). Collectively, these data strongly suggest that MyD88-dependent TLR activation by microbial ligands is responsible for IEC tumor growth in Apcmin/+ mice.

as compared to Apcmin/+ mice. In addition, administration of the IL-1R antagonist, Anakinra, did not affect the extent of IEC tumorigenesis in Apcmin/+ mice (Fig. 2C). Collectively, these data strongly suggest that MyD88-dependent TLR activation by microbial ligands is responsible for IEC tumor growth in Apcmin/+ mice. A MyD88-dependent TLR signaling upregulates c-myc in IEC The decrease in IEC proliferation and the increase in IEC apoptosis in Apcmin/+/Myd88-/- mice suggested the involvement of a MyD88-dependent oncogene or mitogen in IEC tumorigenesis. Since c-myc is essential for tumorigenesis in Apcmin/+ mice 11,12,14, we tested whether MyD88 regulates the expression of c-myc. MyD88-deficiency resulted in a significant decrease in the c-myc protein level in IEC. While c-myc was expressed throughout the crypt in both the DSI and the colon of Apcmin/+ mice, its expression in Apcmin/+/Myd88-/- mice was restricted to the base of the crypt (Fig. 3A). Immunoblotting analysis of c-myc in isolated IEC (DSI) confirmed the reduced expression of not only c-myc, but also pERK in Apcmin/+/Myd88-/- mice (Fig. 3B and Supp. Fig. 2A). The decreased c-myc level in Apcmin/+/Myd88-/- IEC was observed in both normal and tumor regions (Supp. Fig. 2B). However, the c-myc mRNA levels in IEC did not differ significantly between Apcmin/+ and Apcmin/+/Myd88-/- mice (Fig. 3C). Inactivation of Apc activates β–catenin, which induces transcription of c-myc. The deletion of MyD88 did not affect the β-catenin level in vivo (Fig. 3B) or Wnt3-induced activation of β-catenin in vitro (Fig. 3D). Collectively, these data indicate that MyD88 signaling affects tumorigenesis independently of the Wnt-APC-β-catenin pathway.

c activates β–catenin, which induces transcription of c-myc. The deletion of MyD88 did not affect the β-catenin level in vivo (Fig. 3B) or Wnt3-induced activation of β-catenin in vitro (Fig. 3D). Collectively, these data indicate that MyD88 signaling affects tumorigenesis independently of the Wnt-APC-β-catenin pathway. A posttranslational modification of c-myc by TLR-MyD88-ERK pathway stabilizes c-myc expression The data suggested that MyD88-mediated signaling in IEC provokes tumor growth. Since IEC express functional TLRs 18,19 and Supp. Fig. 3A, we tested whether activation of a TLR-MyD88 pathway directly induces c-myc. Indeed, activation of TLR2 enhanced the protein level of c-myc in an IEC line RKO (Apc wild type) (Fig. 4A) in a MyD88-dependent manner (Supp. Fig. 3B). TLR5 (a MyD88-dependent TLR) activation in RKO produced a similar result (Supp. Fig. 3C). Consistent with the results obtained in vivo (Fig. 3C), the level of c-myc mRNA was not affected by TLR2 triggering, while the levels of IL-8 and IκBα were increased 19 (Fig. 4A).

. 4A) in a MyD88-dependent manner (Supp. Fig. 3B). TLR5 (a MyD88-dependent TLR) activation in RKO produced a similar result (Supp. Fig. 3C). Consistent with the results obtained in vivo (Fig. 3C), the level of c-myc mRNA was not affected by TLR2 triggering, while the levels of IL-8 and IκBα were increased 19 (Fig. 4A). The increase in c-myc protein level without a concomitant increase in mRNA level upon TLR stimulation, suggested that c-myc protein is subjected to post-translational modifications 22. Indeed, inhibition of proteasomal function by MG-132 enhanced the c-myc protein levels in RKO cells without affecting the mRNA level (Fig. 4B), indicating a steady state degradation of c-myc. We therefore tested whether TLR stimulation in IEC stabilizes c-myc protein. While the c-myc-ubiquitin conjugates were easily detected even in the absence of a proteasome inhibitor in RKO cells, they rapidly disappeared upon TLR2 stimulation with a concomitant increase in unconjugated c-myc protein (Fig. 4C). These data indicate that the TLR-MyD88-mediated signaling pathway stabilizes c-myc protein in IEC by inhibiting its proteasomal degradation.

ted even in the absence of a proteasome inhibitor in RKO cells, they rapidly disappeared upon TLR2 stimulation with a concomitant increase in unconjugated c-myc protein (Fig. 4C). These data indicate that the TLR-MyD88-mediated signaling pathway stabilizes c-myc protein in IEC by inhibiting its proteasomal degradation. MEK/ERK pathway phosphorylates c-myc on Serine 62, which stabilizes c-myc by preventing ubiquitin/proteasomal degradation 23,24,25. We examined whether Myd88-mediated activation of ERK is responsible for the stabilization of c-myc. Indeed TLR2 activation induced the phosphorylation of ERK as well as of c-myc on Serine 62 (Fig. 4A). In addition, blocking ERK activation with pharmacological inhibitors rapidly reduced c-myc level (Fig. 4D). Caco-2, another IEC line, expresses a truncated APC protein similar to that observed in Apcmin/+ mice 26. We therefore tested whether MyD88-dependent ERK activation can stabilize c-myc in these Apc mutant cells. Indeed, TLR2 stimulation increased the c-myc protein level with concomitant ERK activation and a decrease in the polyubiquitinated c-myc (Supp. Fig. 4A and B). Similarly, stimulation of either TLR2 or EGFR in a non-transformed IEC line derived from the small intestine (RIE-1) also activated ERK and c-myc (Supp. Fig. 4C). These data indicate that c-myc level in Apcmin/+ mice is maintained by two independent mechanisms, 1) a transcriptional activation of c-myc by β-catenin signaling initiated by Apc inactivation and 2) a post-translational stabilization of c-myc by MyD88-dependent ERK activation.

activated ERK and c-myc (Supp. Fig. 4C). These data indicate that c-myc level in Apcmin/+ mice is maintained by two independent mechanisms, 1) a transcriptional activation of c-myc by β-catenin signaling initiated by Apc inactivation and 2) a post-translational stabilization of c-myc by MyD88-dependent ERK activation. ERK signaling drives the Min phenotype We tested whether a MyD88-independent activation of ERK increases c-myc protein level in Apcmin/+/Myd88-/- mice and restores the Min phenotype. As EGF activates ERK and enhances c-myc levels in non-transformed IEC (Supp. Fig. 4C), we treated Apcmin/+/Myd88-/- mice with either EGF alone or with EGF plus a MEK1/2 inhibitor (PD0325901, PD). The latter is a specific and an effective pharmacological inhibitor of ERK 27 and is in phase II clinical trials. The administration of EGF significantly increased the number of polyps in the DSI (for comparison see Supp. Fig. 1B), and this induction was abrogated by PD treatment (Fig. 5A). Serum hemoglobin and body weights drop significantly in Apcmin/+ mice over time, due to the increase in numbers and the sizes of the exophytic polypoid intestinal tumors minimizing food absorption, with subsequent intestinal obstruction and intestinal bleeding 28. The inhibition of tumor growth in Apcmin/+/Myd88-/-, PD-treated animals coincided with increased serum hemoglobin levels (Fig. 5B) and increased body weight (Fig. 5C), indicating that these were healthier animals. As expected, EGF administration enhanced levels of c-myc and pERK in IEC, which was reversed by PD treatment (Fig. 5D). Taken together, the inhibition of IEC tumors in PD treated mice further validated the regulatory role of ERK on tumorigenesis in Apcmin/+/Myd88-/- mice (Fig. 3B) and could suggest that TLR-MyD88 pathway contributes significantly to ERK activation in Apcmin/+ mice, under the steady state conditions.

treatment (Fig. 5D). Taken together, the inhibition of IEC tumors in PD treated mice further validated the regulatory role of ERK on tumorigenesis in Apcmin/+/Myd88-/- mice (Fig. 3B) and could suggest that TLR-MyD88 pathway contributes significantly to ERK activation in Apcmin/+ mice, under the steady state conditions. These results indicated a pivotal role for ERK activation in the Min phenotype. We therefore tested its tumorigenic role in 10-week old Apcmin/+ mice. PD treatment for 14 weeks of these animals resulted in complete inhibition of polyp growth (Fig. 6A) with the concomitant increase in serum hemoglobin levels (Fig. 6B) and body weight (Fig. 6C). PD treatment inhibited the levels of both c-myc and pERK in IEC of these mice (Fig. 6D and 6E). Furthermore, PD treatment resulted in 100% survival whereas treatment of control animals with vehicle resulted in 100% mortality during the 17 weeks treatment period of Apcmin/+ mice (Fig. 6F). To detect the long-term effects of PD treatment, we delivered it or vehicle to already 17-week PD-treated Apcmin/+ mice, for additional 15 weeks. Continuous PD treatment inhibited tumorigenesis while its discontinuation provoked high tumor count (Fig. 6G). Collectively, these results indicate that the regulation of ERK pathway in Apcmin/+ mice controls intestinal tumorigenesis and the subsequent manifestation of the Min phenotype, most likely via post-translational modifications of c-myc protein.

umorigenesis while its discontinuation provoked high tumor count (Fig. 6G). Collectively, these results indicate that the regulation of ERK pathway in Apcmin/+ mice controls intestinal tumorigenesis and the subsequent manifestation of the Min phenotype, most likely via post-translational modifications of c-myc protein. Discussion Overt inflammation can promote neoplasia 29,30,31,32. TLR activation of innate immune cells (e.g., macrophages) in the intestinal mucosa provokes the production of various pro-inflammatory mediators 3,5. This mechanism was proposed to enhance tumorigenesis in the Apcmin/+ mice 6. However, our study indicates that MyD88 in non-hematopoietic cells, such as IEC, is required for intestinal tumor growth in the Apcmin/+ mouse. Furthermore, we identified that TLR ligands presumably from intestinal flora (Fig. 4), and not from the host (Fig. 2), mediate IEC tumor growth under the steady-state conditions. In this setting, MyD88-mediated signaling, induces ERK activation that stabilizes and hence, increases the protein level of the oncogene c-myc in IEC 23. This sequence of events enhances IEC proliferation and reduces IEC apoptosis and therefore promotes IEC tumor growth in Apcmin/+ mice.

er the steady-state conditions. In this setting, MyD88-mediated signaling, induces ERK activation that stabilizes and hence, increases the protein level of the oncogene c-myc in IEC 23. This sequence of events enhances IEC proliferation and reduces IEC apoptosis and therefore promotes IEC tumor growth in Apcmin/+ mice. The oncogene c-myc is a Wnt target gene 33,34. While β-catenin/TCF signaling induces c-myc transcriptionally, its expression levels are heavily regulated by ubiquitin-mediated proteasomal degradation 35,36 which can be antagonized by pERK phosphorylation of c-myc 23,24,25. Our findings indicate that Myd88-dependent, ERK activation is essential to stabilize c-myc levels (Fig. 4), that the activation of ERK by a MyD88-independent ligand, EGF 37, increases c-myc levels and restores the Min phenotype in the Apcmin/+/Myd88-/- mice (Fig. 5), and that treatment with a specific ERK inhibitor suppresses tumor development in both Apcmin/+ and EGF-treated Apcmin/+/Myd88-/- mice (Fig. 5 and Fig. 6). Collectively, these data indicate that 1) the loss of heterozygosity of Apc is insufficient to drive the Min phenotype in the Apcmin/+ mouse, 2) that the synergy between c-myc transcription and post-translational modifications are required for tumor growth in this model, 3) activation of ERK is essential for IEC tumorigenesis in the Apcmin/+ mouse and 4) that ERK functions as a major regulator of c-myc expression in the intestinal epithelium (Fig. 6H).

ouse, 2) that the synergy between c-myc transcription and post-translational modifications are required for tumor growth in this model, 3) activation of ERK is essential for IEC tumorigenesis in the Apcmin/+ mouse and 4) that ERK functions as a major regulator of c-myc expression in the intestinal epithelium (Fig. 6H). One mechanism that explains tumor suppression due to the inactivation of a single gene product is termed oncogene addiction. This phenomenon occurs when tumors require sustained activation of a single oncogene for their growth and survival, despite other oncogenic events 9. Our data reveal that the IEC tumor growth in the Apcmin/+ mice is due to pERK “addiction”. ERK addiction was shown recently to drive the survival of certain intestinal epithelial cell lines in vitro 38, although via a different pathway. Activation of ERK in this setting is most likely induced by a TLR-MyD88-dependent pathway (e.g., microfora, Fig. 3) and by a TLR-Myd88-independent pathway (e.g, growth factors) (Fig. 5). Consequently, the inhibition of ERK prevents tumorigenesis in Apcmin/+ mice, most likely via the generation of an unstable c-myc protein (Fig. 5-6) leading to low c-myc levels in IEC (Fig. 3). Although the regulation of the ERK-c-myc pathway is sufficient for the inhibition of the Min phenotype under the steady state conditions, and its reversal upon EGF administration, in Apcmin/+/Myd88-/- mice, we can't rule out other anti-apoptotic effects provoked by pERK 38 in IEC of these animals.

vels in IEC (Fig. 3). Although the regulation of the ERK-c-myc pathway is sufficient for the inhibition of the Min phenotype under the steady state conditions, and its reversal upon EGF administration, in Apcmin/+/Myd88-/- mice, we can't rule out other anti-apoptotic effects provoked by pERK 38 in IEC of these animals. The dichotomy in tumor numbers between Apcmin/+ and Apcmin/+/Myd88-/- mice (Supp. Fig. 1), as well as the biochemical evidence presented above in vitro (Fig. 4) and in vivo (Fig. 5-6), highly suggest the inductive role of microflora-derived MyD88 signaling on IEC tumorigenesis in Apcmin/+ mice. These observations reveal a new facet of oncogene-environment interactions, which might explain why a germline mutation in Apc results primarily in tumors originating from the intestinal epithelium (Fig. 6H) and not in other organs. Since pERK is a major player in the induction of the Min phenotype (Fig. 5-6), we propose that interventions aimed at inhibiting ERK activation in IEC (Fig. 6) may help suppress the induction of IEC neoplasia in humans with variant Apc genes.

originating from the intestinal epithelium (Fig. 6H) and not in other organs. Since pERK is a major player in the induction of the Min phenotype (Fig. 5-6), we propose that interventions aimed at inhibiting ERK activation in IEC (Fig. 6) may help suppress the induction of IEC neoplasia in humans with variant Apc genes. Materials and Methods Materials The following antibodies were obtained from Cell Signaling Technology (Danvers, MA): anti-phospho ERK1/2, anti-ERK1/2, anti-c-myc, anti-PARP, anti-β-catenin, anti-PCNA and anti-MyD88. Anti-ubiquitin antibody was purchased from Santa Cruz Biotechnology (Santa Cruz, CA) and anti-phospho-c-myc (Ser62) antibody for IB and anti-c-myc antibody for immunohistochemistry from Abcam (Cambridge, MA). InSolution™ MG-132 was purchased from Calbiochem (San Diego, CA), the MEK1/2 inhibitor (U0126) from Promega (Madison, WI) and the MEK1/2 inhibitor, PD0325901, from Stemgent (San Diego, CA). Anakinra was purchased from Amgen (Kineret®, CA), recombinant mEGF from PeproTech, Inc. (Rocky Hill, NJ), the TLR2 ligand, Pam3Cys (P3C) from InvivoGen (San Diego, CA) and the Wnt3a from R&D system (Minneapolis, MN).

(U0126) from Promega (Madison, WI) and the MEK1/2 inhibitor, PD0325901, from Stemgent (San Diego, CA). Anakinra was purchased from Amgen (Kineret®, CA), recombinant mEGF from PeproTech, Inc. (Rocky Hill, NJ), the TLR2 ligand, Pam3Cys (P3C) from InvivoGen (San Diego, CA) and the Wnt3a from R&D system (Minneapolis, MN). Mice C57Bl/6J, Apcmin/+ and Il1r1-/- mice were purchased from The Jackson Laboratory (Bar Harbor, ME). Myd88-/- mice were kindly provided by Dr. S. Akira (Osaka University, Japan), and were backcrossed 10 generations onto C57Bl/6, Lps2 by Dr. B. Beutler (TSRI, San Diego, CA) and Caspase1-/- mice by Dr. R. Flavell (Yale University, CT). All these mice strain were crossed to Apcmin/+ mice. All animal protocols received prior approval by the Institutional Review Board. In vivo treatment with Anakinra Eight to 10 week-old mice Apcmin/+ mice were injected i.p with 50 mg/kg of Anakinra, 5 times/week for ten weeks and analyzed when they reached 20 weeks of age. In vivo treatment with EGF Eight to 10 week-old mice were injected i.p with EGF (2 μg/mouse), 3 times/week for 10 weeks and analyzed when they reached 20 weeks of age.

In vivo treatment with Anakinra Eight to 10 week-old mice Apcmin/+ mice were injected i.p with 50 mg/kg of Anakinra, 5 times/week for ten weeks and analyzed when they reached 20 weeks of age. In vivo treatment with EGF Eight to 10 week-old mice were injected i.p with EGF (2 μg/mouse), 3 times/week for 10 weeks and analyzed when they reached 20 weeks of age. In vivo treatment with an ERK inhibitor PD0325901 was dissolved initially in DMSO (50 mg/ml) as a stock solution. The stock solution was then diluted fresh in water containing 0.05% (Hydroxypropyl)methycellulose and 0.02% Tween 80. The formulation containing PD0325901 in 250 μl at the 25 mg/kg dose was administered by gavage three times a week to EGF-treated Apcmin/+/Myd88-/- mice or five times a week to Apcmin/+ mice, for the duration of each study. Controls mice were treated with vehicle (gavage). Bone marrow (BM) chimeras were generated by reconstituting irradiated (9 Gy of γ-radiation) 6-10 week-old Apcmin/+ and Apcmin/+/Myd88-/- mice with BM cells (1.5 × 107, i.v.) from sex-matched WT or Myd88-/- donor mice. Chimerism was verified by qPCR of peripheral blood cells. Polyp counts were performed when mice reached 20 weeks of age. BrdU staining was performed using a BrdU in situ staining kit (BD Biosciences, San Diego, CA). Mice were injected i.p. with 2 mg of BrdU solution. Intestinal tissue samples were fixed with formalin and embedded in paraffin. Immunostaining for labeled BrdU was performed according to the manufacturer's instruction. The enumeration of BrdU positioning was performed as described 39.

(BD Biosciences, San Diego, CA). Mice were injected i.p. with 2 mg of BrdU solution. Intestinal tissue samples were fixed with formalin and embedded in paraffin. Immunostaining for labeled BrdU was performed according to the manufacturer's instruction. The enumeration of BrdU positioning was performed as described 39. TUNEL assay was performed on paraffinized intestinal tissues according to the manufacturer's instruction (BD Biosciences). Nuclei were stained with Hoechst 33258 (Invitrogen, Carlsbad, CA). Isolation of intestinal epithelial cells, RT-PCR, Immunoblotting and immunoprecipitation were performed as previously described 19. Cell Culture The human IEC cell lines RKO and Caco-2 were cultured in DMEM supplemented with 4.0 mM glutamine, 10% fetal calf serum, 50 U/ml penicillin and 50 μg/ml streptomycin. siRNA-mediated knockdown Myd88 siRNA or c-myc siRNA from Santa Cruz Biotechnology (Santa Cruz, CA). Briefly, siRNA (40 μM) in 50 μl of Opti-MEM (Invitrogen) was mixed with 5 μl of Dharmafect 4 (Dharmacon, Chicago, IL) in 50 μl of Opti-MEM. After 30 min incubation at RT, the transfection mixture was combined with 1 × 106 cells in culture medium. Non-targeting siRNA #2 (luciferase targeting siRNA) from Dharmacon was used as a control.

40 μM) in 50 μl of Opti-MEM (Invitrogen) was mixed with 5 μl of Dharmafect 4 (Dharmacon, Chicago, IL) in 50 μl of Opti-MEM. After 30 min incubation at RT, the transfection mixture was combined with 1 × 106 cells in culture medium. Non-targeting siRNA #2 (luciferase targeting siRNA) from Dharmacon was used as a control. Histology and Immunohistochemistry (IHC) DSI and colon were fixed in 10% formalin, paraffin embedded, and sectioned at 3 to 6 μm for H&E staining or immunostaining. The tissue sections were incubated with rabbit anti-c-myc ab (1:50), rabbit anti-pERK ab, or with control ab, overnight at 4°C. After washing with PBS, sections were incubated in HRP-conjugated secondary antibody for an hour and the staining was visualized with AEC peroxidase substrate kit (Vector Laboratories, Inc., Burlingame, CA), with hematoxylin nuclear counterstaining. Blood hemoglobin was measured on a MS9 Blood Analyzer (Melet Schloesing Laboratories, El Cajon, CA) according to the manufacturer's instructions. Statistical analysis was performed by Student's t test for paire samples or two-way ANOVA for multiple comparisons and by log-rank analysis for survival curves. Data are presented as means ± s.d. Supplementary Material 1 2 The authors thank Patty Charos for animal breeding and Steve Shenouda for tissue processing. This work was supported by NIH grants AI068685, CA133702, DK35108 and DK080506. The authors declare that they have no competing financial interest.

Statistical analysis was performed by Student's t test for paire samples or two-way ANOVA for multiple comparisons and by log-rank analysis for survival curves. Data are presented as means ± s.d. Supplementary Material 1 2 The authors thank Patty Charos for animal breeding and Steve Shenouda for tissue processing. This work was supported by NIH grants AI068685, CA133702, DK35108 and DK080506. The authors declare that they have no competing financial interest. Author contributions. E.R. designed the study, S.H.L. and J.L. performed the signaling experiments, C.S., L.H., S.H. and G.S.S. performed the in vivo studies, M.C. generated the bone marrow chimeras, J.B., J.L. and J.G. performed immunohistochemistry and flow cytometry, S.H.L., M.P.C, N.V., J.L. and E.R. analyzed the data, and S.H.L., J.L. and E.R. wrote the manuscript. Figure 1 Genetic disruption of Myd88 in Apcmin/+ mice suppresses proliferation and enhances apoptosis of IEC A. IHC and BrdU incorporation in IEC (DSI) after i.p. BrdU injection (scale bars - 20 μm, magnification ×200). BrdU-positive cells, per time point, were enumerated for each indicated position in a crypt (10 crypt-villi units/time point), position 0 being the base of the crypt 39. B. Apoptotic IEC (DSI) were determined by TUNEL assay (scale bars - 40 μm, magnification ×100). C. Cleaved product of poly(ADP-ribose) polymerase (PARP) in IEC (DSI) harvested from the indicated mice (n=2/group).

indicated position in a crypt (10 crypt-villi units/time point), position 0 being the base of the crypt 39. B. Apoptotic IEC (DSI) were determined by TUNEL assay (scale bars - 40 μm, magnification ×100). C. Cleaved product of poly(ADP-ribose) polymerase (PARP) in IEC (DSI) harvested from the indicated mice (n=2/group). Figure 2 Myd88 signaling in hematopoietic cells is not required for tumorigenesis in Apcmin/+ mice A. Polyp count in BM chimeras in the DSI and colon (P=n.s, n=7-9 mice/group). B. Polyp count in the small intestine in Apcmin/+/Il1r1-/- and Apcmin/+/caspase-1-/- mice at 20 weeks of age (n=7/group). C. Polyp count in Anakinra-treated Apcmin/+ mice (DSI) (P=n.s., n=7/group). Figure 3 MyD88 regulates c-myc expression levels A. IHC analysis of c-myc protein in IEC from the DSI and colon from 20-week old mice (scale bars, 10 μm, magnification ×200). B. IB analysis of the indicated proteins in IEC (DSI) of 20-weeks mice (n=2). C. Transcript levels of c-myc in IEC (DSI) (P=n.s, n=3/group). D. RKO cells transfected with either control or Myd88 siRNA, were stimulated with Wnt3a (100ng/ml) and subjected to IB analysis.

old mice (scale bars, 10 μm, magnification ×200). B. IB analysis of the indicated proteins in IEC (DSI) of 20-weeks mice (n=2). C. Transcript levels of c-myc in IEC (DSI) (P=n.s, n=3/group). D. RKO cells transfected with either control or Myd88 siRNA, were stimulated with Wnt3a (100ng/ml) and subjected to IB analysis. Figure 4 TLR signaling via MyD88 stabilizes c-myc protein in IEC through activation of ERK A. Upper panel: RKO cells were stimulated with P3C (2 μg/ml), lysed and analyzed by IB. Lower panel: Transcript levels after TLR2 stimulation (qPCR). B. Protein levels (IB) (Upper panel) and transcript level (qPCR) (Lower panel) in MG-132 treated (10 μM) RKO cells. C-myc was immunoprecipitated followed by IB with anti-ubiquitin (Ub) ab. C. RKO cells were treated with P3C (2 μg/ml) and ubiquitinated c-myc level was measured by IP followed by IB. D. Phospho-ERK and c-myc levels (IB) in U0126- or PD0325901-treated RKO cells. Figure 5 Activation of ERK restores the Min phenotype in Apcmin/+/Myd88-/- mice A. PD reduces the number of polyps in EGF-treated Apcmin/+/Myd88-/- mice (DSI) (n=8/group). B. Blood hemoglobin levels and (C) body weight of these mice. D. Top panel: H&E of DSI in control, EGF-treated, and EGF + PD-treated mice. The arrows indicate intestinal polyps. Middle panel: C-myc expression in IEC. Bottom panel: Phospho-ERK levels in IEC of these mice.

reated Apcmin/+/Myd88-/- mice (DSI) (n=8/group). B. Blood hemoglobin levels and (C) body weight of these mice. D. Top panel: H&E of DSI in control, EGF-treated, and EGF + PD-treated mice. The arrows indicate intestinal polyps. Middle panel: C-myc expression in IEC. Bottom panel: Phospho-ERK levels in IEC of these mice. Figure 6 Activation of ERK is essential for the Min phenotype in Apcmin/+ mice A. Polyp count, B hemoglobin level, and C body weight in PD-treated Apcmin/+ mice (n=6 for vehicle, n=9 for PD group). D. Upper panel: H&E of DSI in control and PD-treated Apcmin/+ mice. Arrows indicate intestinal polyps. Lower panel: C-myc expression in IEC. E. IB analysis of c-myc and pERK levels in IEC (DSI) of these mice. F. Survival in PD-treated or vehicle-treated Apcmin/+ mice for 17 weeks (n=8). G. The PD-treated Apcmin/+ mice mentioned in F were split to PD- and vehicle-treated groups (n=4/group). Polyp count (DSI) was performed 15 weeks later. H. The microflora induces tumorigenesis in Apcmin/+ mice by triggering TLR-ERK pathway in IEC. This stabilizes c-myc and inhibits its proteasomal degradation. Increased c-myc levels induce the Min phenotype. Additional signals such as growth factors, utilize the MEK-ERK pathway and similarly to TLR ligands, can enhance c-myc levels. Of note, sterile food and water still contain TLR ligands (e.g., LPS) that are capable of stimulating IEC. This mechanism may account for the Min phenotype observed in Apcmin/+ mice housed under germ-free conditions 40.

D-1 inhibition. Silencing BATF in T cells from chronic viremic patients rescued HIV-specific T cell function. Thus inhibitory receptors can cause T cell exhaustion by upregulating genes – such as BATF – that inhibit T cell function. Such genes may provide new therapeutic opportunities to improve T cell immunity to HIV. We hypothesized that receptors like PD-1 function to inhibit T cells not only by reducing TCR signaling, but also by inducing the expression of genes that impair T cell function. To test this hypothesis, we queried gene expression profiles from HIV-specific CD8+ T cells for upregulation of PD-1 induced genes. The majority of individuals infected with HIV show chronic elevation of viral load in the absence of anti-retroviral therapy (progressors), associated with defects in HIV-specific T cell cytokine secretion, proliferation and survival6,7. In contrast, spontaneous control of viral replication has been documented for a small minority of individuals (controllers)8. Analysis of CD8+ T cell responses to HIV in progressors and controllers therefore allows a comparison of populations of human antigen-specific T cells at the extremes of functional competence.

. In contrast, spontaneous control of viral replication has been documented for a small minority of individuals (controllers)8. Analysis of CD8+ T cell responses to HIV in progressors and controllers therefore allows a comparison of populations of human antigen-specific T cells at the extremes of functional competence. We sorted CD8+ T cells specific for epitopes from the Gag protein (hereafter termed HIV-specific CD8+ T cells) from 18 progressors and 24 controllers (Fig. 1a, Supplementary Fig. 1, and Supplementary Table 1). The gene expression profiles of HIV-specific CD8+ T cells from progressors showed marked differences to those from controllers (n = 518 genes, moderated t-statistic < −2.0, Fig. 1b and Supplementary Table 2). Genes upregulated in HIV-specific CD8+ T cells from progressors were enriched for those involved with the interferon response and MHC expression (Supplementary Table 3), consistent with a higher viral load in progressors. HIV-specific CD8+ T cells from controllers were enriched for genes involved in mRNA transcription and protein translation, consistent with prior observations of defects seen in the mouse model of chronic LCMV infection9 (Supplementary Table 4). We therefore compared the expression profiles of HIV-specific CD8+ T cells to exhausted LCMV-specific CD8+ T cells from the mouse model9. Using an analytic technique called gene set enrichment analysis (GSEA)10–12 (Supplementary Methods) we analyzed the expression profiles of murine virus-specific CD8+ T cells during infection with each of two strains of LCMV: clone 13 (Cl13), which gives rise to chronic infection with T cell exhaustion; and Armstrong (Arm), an acute infection that does not cause T cell exhaustion9,13. We found that the HIV progressor signature was significantly enriched in the profiles of exhausted LCMV-specific CD8+ T cells from Cl13 infection (P = 4.8 × 10−5, Supplementary Fig. 2), suggesting global similarity between the transcriptional profiles of exhausted CD8+ T cells in humans and in the mouse model.

Introductory Paragraph CD8+ T cells in chronic viral infections like HIV develop functional defects such as loss of IL-2 secretion and decreased proliferative potential that are collectively termed exhaustion1. Exhausted T cells express increased levels of multiple inhibitory receptors, such as Programmed Death 1 (PD-1)2,3 that contribute to impaired virus-specific T cell function. Reversing PD-1 inhibition is therefore an attractive therapeutic target, but the cellular mechanisms by which PD-1 ligation results in T cell inhibition are not fully understood. PD-1 is thought to limit T cell activation by attenuating T cell receptor (TCR) signaling4,5. It is not known whether PD-1 also acts by upregulating genes in exhausted T cells that impair their function. Here, we analyzed gene-expression profiles from HIV-specific CD8+ T cells in patients with HIV and show that PD-1 coordinately upregulates a program of genes in exhausted CD8+ T cells from humans and mice. This program includes upregulation of basic leucine transcription factor, ATF-like (BATF), a transcription factor in the AP-1 family. Enforced expression of BATF was sufficient to impair T cell proliferation and cytokine secretion, while BATF knockdown reduced PD-1 inhibition. Silencing BATF in T cells from chronic viremic patients rescued HIV-specific T cell function. Thus inhibitory receptors can cause T cell exhaustion by upregulating genes – such as BATF – that inhibit T cell function. Such genes may provide new therapeutic opportunities to improve T cell immunity to HIV.

ion9,13. We found that the HIV progressor signature was significantly enriched in the profiles of exhausted LCMV-specific CD8+ T cells from Cl13 infection (P = 4.8 × 10−5, Supplementary Fig. 2), suggesting global similarity between the transcriptional profiles of exhausted CD8+ T cells in humans and in the mouse model. We next asked if this exhausted CD8+ signature was influenced by PD-1 signaling. To do this, we first identified the genes upregulated following PD-1 ligation. We incubated PD-1 expressing Jurkat cells with beads coated with a cross-linking antibody to PD-1 together with antibodies to CD3 and CD28 (PD-1/CD3/CD28 beads); or with beads coated with equivalent amounts of control antibody together with CD3 and CD28 (CD3/CD28 beads). Incubation with PD-1/CD3/CD28 beads significantly decreased production of IL-2 compared to cells incubated with CD3/CD28 beads (P = 0.007, Fig. 1c) as previously observed14,15. Microarray analysis identified over one thousand genes that were significantly upregulated in cells functionally inhibited by PD-1 (n = 1179, t > 2.0, Fig. 1d and Supplementary Table 5). A similar number of genes was reduced in expression following PD-1 ligation (n = 1361, t < −2.0, Fig. 1d and Supplementary Table 5). We validated 13 representative genes that were upregulation in PD-1-ligated Jurkat cells. Incubation of human CD4+ T cells with PDL1-Ig/CD3/CD28 beads led to the coordinate upregulation of these representative PD-1 signature genes in a PDL1-Ig dose-dependent manner (Supplementary Fig. 3). Thus ligation of PD-1 in CD3/CD28 stimulated cells induces a specific transcriptional program in both Jurkat cells and primary human T cells.

D4+ T cells with PDL1-Ig/CD3/CD28 beads led to the coordinate upregulation of these representative PD-1 signature genes in a PDL1-Ig dose-dependent manner (Supplementary Fig. 3). Thus ligation of PD-1 in CD3/CD28 stimulated cells induces a specific transcriptional program in both Jurkat cells and primary human T cells. We sought to determine whether the transcriptional program induced by PD-1 signaling defined in vitro could be detected in gene expression profiles from exhausted CD8+ T cells ex vivo. We therefore tested whether PD-1 induced genes were coordinately upregulated in HIV-specific CD8+ T cells from HIV Progressors. Using enrichment analysis, we found that a set of PD-1 signature genes was significantly upregulated in the HIV progressors compared with controllers (P = 5 × 10−6, Fig. 1e). Similarly, we found that PD-1 signature genes were significantly upregulated in exhausted LCMV-specific CD8+ T cells from Cl13 infection compared with Armstrong infection (P = 2 × 10−4, Supplementary Fig. 4). Thus PD-1 ligation results in upregulation of a consistent pattern of genes in exhausted CD8+ T cells in humans and mice.

we found that PD-1 signature genes were significantly upregulated in exhausted LCMV-specific CD8+ T cells from Cl13 infection compared with Armstrong infection (P = 2 × 10−4, Supplementary Fig. 4). Thus PD-1 ligation results in upregulation of a consistent pattern of genes in exhausted CD8+ T cells in humans and mice. The upregulation of PD-1 signature genes in exhausted CD8+ T cells contrasted with that seen in profiles of human virus-specific CD8+ T cells associated with functional T cell responses. Using single-sample GSEA (Supplementary Methods), we found that the PD-1 signature was significantly more enriched in HIV-specific CD8+ T cells than in antigen-specific CD8+ T cells specific for CMV (P < 0.01), EBV (P < 0.001), or influenza virus (P < 0.001) from healthy HIV-uninfected donors (Fig. 1f). Notably, the PD-1 signature was significantly more enriched in HIV-specific T cells than in EBV-specific T cells, despite the fact EBV-specific T cells express PD-116,17. This suggests that the upregulation of PD-1 signature genes may not occur in all cells that express PD-1, but may reflect increased strength and duration of PD-1 signaling experienced by CD8+ T cells in the setting of chronic infection.

n in EBV-specific T cells, despite the fact EBV-specific T cells express PD-116,17. This suggests that the upregulation of PD-1 signature genes may not occur in all cells that express PD-1, but may reflect increased strength and duration of PD-1 signaling experienced by CD8+ T cells in the setting of chronic infection. We reasoned that genes upregulated by PD-1 in exhausted CD8+ T cells might include those involved in the inhibition of T cell function. To refine the list of candidate genes, we identified genes that were both upregulated by PD-1 ligation in Jurkat cells and increased in HIV progressors (Fig. 2a and Supplementary Table 6). We focused on transcription factors because of their broad effect on the cell state. Of the 75 genes common to both gene sets, only three were transcription factors: BATF, STAT1 and IRF9 (Fig. 2a and Supplementary Table 6). We selected BATF for further analysis because it has been shown to function as a negative regulator of AP-1 activity18,19. Moreover, we have previously observed it to be upregulated during CD8+ memory differentiation in humans and mice20, suggesting that it may play a conserved role in regulating T cell function. BATF expression showed a 2 – 3 fold increase in both primary human CD4+ and CD8+ T cells after incubation with PDL1/CD3/CD28 beads compared with CD3/CD28 beads, indicating that BATF expression is increased by PD-1 ligation in vitro (P = 0.001 and P = 0.02, respectively, Fig. 2b).

ved role in regulating T cell function. BATF expression showed a 2 – 3 fold increase in both primary human CD4+ and CD8+ T cells after incubation with PDL1/CD3/CD28 beads compared with CD3/CD28 beads, indicating that BATF expression is increased by PD-1 ligation in vitro (P = 0.001 and P = 0.02, respectively, Fig. 2b). High BATF levels were seen in antigen-specific T cells with the greatest degree of dysfunction. BATF expression measured by microarray was significantly higher in exhausted HIV-specific CD8+ T cells from progressors than in HIV-specific T cells from controllers (P = 0.003, Fig. 2c). However BATF expression was not significantly different in naïve CD8+ T cells from controllers or progressors (P = NS, Fig. 2c). A significant correlation existed between microarray and RT-PCR measurements of BATF expression (Rs 0.53, P = 0.02, n = 18, Supplementary Fig. 5).

cells from controllers (P = 0.003, Fig. 2c). However BATF expression was not significantly different in naïve CD8+ T cells from controllers or progressors (P = NS, Fig. 2c). A significant correlation existed between microarray and RT-PCR measurements of BATF expression (Rs 0.53, P = 0.02, n = 18, Supplementary Fig. 5). In order to define the kinetics of BATF expression following infection with a persistent virus, we compared BATF expression in murine virus-specific CD8+ T cell during acute and chronic infection (Fig. 2d). As early as day 8 post-infection, DbGP33-specific CD8+ T cells in Cl13 infection expressed significantly higher levels of BATF than in Arm infection (P = 0.02). BATF expression was maintained at higher levels in virus-specific cells in Cl13 infection at day 15 and by day 30 was ~7 fold higher than in DbGP33-specific CD8+ T cells generated during LCMV Arm infection (P = 0.02). The increased expression of BATF during acute and chronic infection was coincident with the upregulation of PD-1 as GP33-specific T cells showed increased levels of both PD-1 and BATF by day 8 (Fig. 2d). Increased BATF expression is therefore an early and persistent feature of exhausted CD8+ T cells in the setting of chronic viral infection in vivo and is temporally correlated with upregulation of PD-1.

h the upregulation of PD-1 as GP33-specific T cells showed increased levels of both PD-1 and BATF by day 8 (Fig. 2d). Increased BATF expression is therefore an early and persistent feature of exhausted CD8+ T cells in the setting of chronic viral infection in vivo and is temporally correlated with upregulation of PD-1. We next tested whether BATF could inhibit T cell function. Overexpression of BATF in primary human T cells (Supplementary Fig. 6a) markedly reduced proliferation in response to CD3/CD28 (P = 0.002, Fig. 3a,b). Apoptosis was also slightly increased in BATF overexpressing cells following stimulation (29% vs. 20%, P = 0.013, Fig. 3a,b), consistent with the previous defined role of PD-1 signaling in reducing cell survival21. However, the majority of BATF overexpressing cells were viable, suggesting that reduced proliferation was not simply from cell death. Overexpression of BATF also significantly reduced IL-2 secretion following CD3/CD28 stimulation, (P = 4.5 × 10−5, Fig. 3c), but was not overtly toxic because IFN-γ secretion was not significantly reduced compared with vector controls. Thus increased expression of BATF reduces proliferation and IL-2 secretion in primary human T cells.

of BATF also significantly reduced IL-2 secretion following CD3/CD28 stimulation, (P = 4.5 × 10−5, Fig. 3c), but was not overtly toxic because IFN-γ secretion was not significantly reduced compared with vector controls. Thus increased expression of BATF reduces proliferation and IL-2 secretion in primary human T cells. Enforced expression of BATF in vitro did not increase the expression of PD-1 itself, or of two other inhibitory receptors (CD244 or CD160) in CD8+ or CD4+ T cells (Supplementary Fig. 6b–e). This suggests that BATF does not mediate inhibition by modulating the expression of these inhibitory receptors. Future studies will be required to determine if BATF regulates other components of the PD-1 induced expression signature. We next asked whether depletion of BATF enhanced T cell function, using shRNA-mediated gene-silencing (Fig. 3d,e). Compared with control hairpins, depletion of BATF in Jurkat cells with two different shRNA sequences (Fig. 3d) significantly increased IL-2 expression in cells cultured with PD-1/CD3/CD28 (P < 0.01, Fig. 3e), reversing inhibition to levels seen in CD3/CD28 stimulated cells. Testing additional hairpin sequences showed that there was a strong correlation between the extent of knockdown and degree of increase in IL-2 secretion, confirming the on-target specificity of BATF silencing (Rs −0.82, P = 0.056; Fig. 3f).

01, Fig. 3e), reversing inhibition to levels seen in CD3/CD28 stimulated cells. Testing additional hairpin sequences showed that there was a strong correlation between the extent of knockdown and degree of increase in IL-2 secretion, confirming the on-target specificity of BATF silencing (Rs −0.82, P = 0.056; Fig. 3f). BATF silencing also increased IL-2 expression in cells stimulated with CD3/CD28 without exogenous PD-1 cross-linking (P < 0.01, Fig. 3e), suggesting that pathways in addition to PD-1 could inhibit cell activation via BATF. Silencing BATF may therefore have the effect of increasing IL-2 expression not only by relieving PD-1-mediated inhibition, but also by impairing other negative feedback pathways. Consistent with this, the expression of BATF across 42 samples of HIV-specific CD8+ T cells was significantly correlated with expression levels of several receptors with known or putative inhibitory function (Supplementary Fig. 7)22,23.

g PD-1-mediated inhibition, but also by impairing other negative feedback pathways. Consistent with this, the expression of BATF across 42 samples of HIV-specific CD8+ T cells was significantly correlated with expression levels of several receptors with known or putative inhibitory function (Supplementary Fig. 7)22,23. Finally, we tested whether silencing BATF would improve the function of HIV-specific T cells. HIV-specific T cell function after BATF knockdown (Fig. 4a, b) was assessed by measuring cytokine secretion or proliferation in response to Gag peptides. BATF knockdown caused a significant increase in CD8+ Gag-specific IFN-γ secretion (Fig. 4c) compared to a control siRNA pool, increasing IFN-γ secretion an average of 60% (P < 0.0001). Similar results were seen in HIV-specific CD4+ T cells where silencing BATF caused a two-fold increase in Gag-specific IL-2 secretion (P = 0.008, Fig. 4d) and a trend towards increase IFN-γ secretion (P = 0.078, Fig. 4e). HIV-specific CD8+ T cell proliferation was also increased by BATF knockdown, with a 5-fold increase in proliferating cells incubated with optimal Gag peptides (P = 0.004, Fig. 4f). Reducing BATF expression therefore increases the function of exhausted HIV-specific T cells.

increase IFN-γ secretion (P = 0.078, Fig. 4e). HIV-specific CD8+ T cell proliferation was also increased by BATF knockdown, with a 5-fold increase in proliferating cells incubated with optimal Gag peptides (P = 0.004, Fig. 4f). Reducing BATF expression therefore increases the function of exhausted HIV-specific T cells. We show that exhausted T cells specific for HIV in humans and LCMV in mice share a common expression signature that reflects the transcriptional consequences of PD-1 receptor ligation. Our data therefore suggest a model in which inhibitory receptors such as PD-1 may mediate T cell exhaustion not only by limiting TCR signaling, but also by inducing the expression of genes such as BATF that inhibit T cell function. BATF is a highly conserved member of the AP-1/ATF family, a group of transcription factors that regulate many aspects of cellular function in the immune system24. Recent studies show that BATF is required for Th17 and follicular T helper cell differentiation25,26. BATF may therefore be one of a number of transcription factors, such as Blimp-127,28, that have distinct, context-dependent roles both in regulating the function of T cells responding to chronic viral infection and in CD4+ lineage decisions. Our studies do not identify the mechanism by which PD-1 ligation induces BATF upregulation, or whether this happens by direct or indirect pathways. However, our findings give impetus to further studies of how BATF regulates T cell state.

on of T cells responding to chronic viral infection and in CD4+ lineage decisions. Our studies do not identify the mechanism by which PD-1 ligation induces BATF upregulation, or whether this happens by direct or indirect pathways. However, our findings give impetus to further studies of how BATF regulates T cell state. Blockade of PD-1:PDL-1 interactions partially reverses T cell dysfunction16,17 and improves control of viral replication2,29, indicating that the function of exhausted T cells can be rescued even in settings of viral persistence. BATF and the pathways that control its activity may provide new opportunities to reverse CD8+ T cell exhaustion. Integrated genomic analysis of the T cell response in humans may therefore provide a general approach to identifying novel regulators of T cell function that are potential therapeutic targets for improving T cell immunity in chronic infection.

ts activity may provide new opportunities to reverse CD8+ T cell exhaustion. Integrated genomic analysis of the T cell response in humans may therefore provide a general approach to identifying novel regulators of T cell function that are potential therapeutic targets for improving T cell immunity in chronic infection. Methods Subjects Subjects were recruited from outpatient clinics at local Boston hospitals, or referred from providers throughout the US, following institutional review board approval (Partners IRB) and written informed consent. HIV controllers included elite controllers (n = 20) with HIV RNA below the level of detection for the respective available ultrasensitive assay (< 75 copies ml−1 by DNA or < 50 copies ml−1 by ultrasensitive PCR); and viremic controllers (n = 4) with HIV RNA levels < 2000 copies ml−1. Chronic progressors (n = 18) were defined as having HIV RNA levels above 2,000 copies (Supplementary Table 1). All patients were off therapy and had detectable HIV-specific CD8+ T cells in the peripheral blood, allowing a median of 21,500 HIV Gag tetramer+ T cells (range 3,000 – 85,000 cells) to be isolated for microarray analysis from each patient.

re defined as having HIV RNA levels above 2,000 copies (Supplementary Table 1). All patients were off therapy and had detectable HIV-specific CD8+ T cells in the peripheral blood, allowing a median of 21,500 HIV Gag tetramer+ T cells (range 3,000 – 85,000 cells) to be isolated for microarray analysis from each patient. Flow cytometry and sorting PBMC were isolated via density centrifugation and were stained with a cocktail of antibodies to exclude irrelevant lineages and dead cells, anti-CD8 and MHC Class I HIV-Gag-specific tetramers to identify the antigen-specific populations, and antibodies against CD62L and CD45RA for memory phenotype characterization of the tetramer+ fraction. CD8+tetramer+ cells were sorted using a FACSAria Cell Sorter (BD Biosciences). All experiments examining proliferation via CFSE dilution and survival via Annexin V (BD Biosciences) staining were collected on a FC500 flow cytometer (Beckman Coulter). Analysis of flow cytometry data was carried out using FlowJo software (version 8.8.6, Tree Star)

orted using a FACSAria Cell Sorter (BD Biosciences). All experiments examining proliferation via CFSE dilution and survival via Annexin V (BD Biosciences) staining were collected on a FC500 flow cytometer (Beckman Coulter). Analysis of flow cytometry data was carried out using FlowJo software (version 8.8.6, Tree Star) Microarray data acquisition and analysis Tetramer-sorted human CD8+ T cells or Jurkat cells following 18h of bead stimulation were pelleted and resuspended in TRIzol. RNA extraction was performed using the RNAdvance Tissue Isolation kit (Agencourt). Concentrations of total RNA were determined using a Nanodrop spectrophotometer or Ribogreen RNA quantitation kit (Molecular Probes/Invitrogen). RNA purity was determined by Bioanalyzer 2100 traces (Agilent Technologies). Total RNA was amplified using the WT-Ovation Pico RNA Amplification system (NuGEN) according to the manufacturer’s instructions. Following fragmentation and biotinylation, cDNA was hybridized to Affymetrix HT HG-U133A or HG-U133A2.0 microarrays. Microarray data for CMV, EBV, and influenza specific CD8+ T cells and for LCMV-specific T cells were obtained from previous studies9,20. Detailed description of microarray data analysis can be found in Supplementary Methods.

on and biotinylation, cDNA was hybridized to Affymetrix HT HG-U133A or HG-U133A2.0 microarrays. Microarray data for CMV, EBV, and influenza specific CD8+ T cells and for LCMV-specific T cells were obtained from previous studies9,20. Detailed description of microarray data analysis can be found in Supplementary Methods. Quantitative PCR Expression of BATF following in vitro stimulation of primary human T cells, shRNA and overexpression experiments was determined by real-time quantitative PCR using Taqman gene expression assays for BATF (assay #Hs00232390_m1) and β-ACTIN (Hs00357333_g1) which served as a loading and normalization control. For LCMV mouse experiments, RNA was converted to cDNA using a high capacity cDNA kit and RT-PCR primers were Taqman assays (Batf, Gapdh, Hprt-1; Applied Biosystems). Expression levels were compared using the relative quantification method, comparing Batf expression to either Gapdh or Hprt-1 housekeeping genes. Quantitative multiplex RT-PCR via ligation-mediated amplification was carried out as previously described30. PD-1 signature genes were selected for the multiplex validation panel based using criteria previously established30 and sequences for primer sets are available upon request.

r Gapdh or Hprt-1 housekeeping genes. Quantitative multiplex RT-PCR via ligation-mediated amplification was carried out as previously described30. PD-1 signature genes were selected for the multiplex validation panel based using criteria previously established30 and sequences for primer sets are available upon request. Mouse model of LCMV infection Groups of mice were infected with either LCMV Armstrong or LCMV clone-13. At 8, 15, or 30 days post-infection, spleens were harvested and splenocytes pooled prior to CD19 depletion and sorting. LCMV tetramer specific (DbGP33) CD8+ T cells were sorted using a BD FACSAria directly into siliconized 1.5 ml tubes containing Trizol LS, followed by RNA extraction. Separate aliquots of cells were stained for PD-1 expression as described previously9.

arvested and splenocytes pooled prior to CD19 depletion and sorting. LCMV tetramer specific (DbGP33) CD8+ T cells were sorted using a BD FACSAria directly into siliconized 1.5 ml tubes containing Trizol LS, followed by RNA extraction. Separate aliquots of cells were stained for PD-1 expression as described previously9. BATF siRNA knockdown in HIV samples PBMCs from untreated, chronically HIV-infected individuals were isolated with density gradient centrifugation. Inhibition of BATF expression was achieved through siRNA transfection by electroporation on a Gene Pulser XCell (BioRad). Fifteen million cells were resuspended in 300 µl of Opti-MEM in a 2-mm cuvette and pulsed with 1 nmol of siRNA (ON-TARGET Non-Targeting pool or BATF ON-TARGETplus SMARTpool, Dharmacon). Pulse conditions were designed to maximize electroporation efficiency in T cells (a unique square wave with a pulse of 360 V and a duration of 5 ms), and transfection efficiency assayed with siGlo fluorescent oligonucleotides (Dharmacon). For assessment of CD4+ T cell cytokine responses, CD8+ T cell depleted PBMCs (RosetteSep CD8+ depletion reagents; StemCell) were stimulated with an HIV Gag peptide pool (1 µg ml−1 peptide−1), or left unstimulated. For HIV-specific CD8+ T cell cytokine responses, non-depleted PBMCs were stimulated with 0.2 µg ml−1 of HIV optimal epitopes. After a 96-hour incubation, IFN-γ and IL-2 levels were measured. Proliferation of CD8+ T cells was measured 6 days after transfection and stimulation using CFSE assay as published before18.

ted. For HIV-specific CD8+ T cell cytokine responses, non-depleted PBMCs were stimulated with 0.2 µg ml−1 of HIV optimal epitopes. After a 96-hour incubation, IFN-γ and IL-2 levels were measured. Proliferation of CD8+ T cells was measured 6 days after transfection and stimulation using CFSE assay as published before18. Additional methods Detailed methodology is described in the Supplementary Methods. Accession Codes Gene Expression Omnibus [data submitted, awaiting accession number] Supplementary Material 1 Acknowledgments The authors would like to thank the subjects for taking part in the study; E. Cutrell, B. Baker, K. Moss, A. Rathod and C. Brume for coordinating sample management; and A. Sharpe, T. Golub, G. Lauer, M. Altfeld, and H. Joffe for valuable discussions. This work was supported by National Institutes of Health grants AI082630, AI56299, HHSN26620050030C, HL092565, the International HIV Controllers Study (http://www.hivcontrollers.org), and the Foundation for the National Institutes of Health through the Grand Challenges in Global Health Initiative. Author attributions

Supplementary Material 1 Acknowledgments The authors would like to thank the subjects for taking part in the study; E. Cutrell, B. Baker, K. Moss, A. Rathod and C. Brume for coordinating sample management; and A. Sharpe, T. Golub, G. Lauer, M. Altfeld, and H. Joffe for valuable discussions. This work was supported by National Institutes of Health grants AI082630, AI56299, HHSN26620050030C, HL092565, the International HIV Controllers Study (http://www.hivcontrollers.org), and the Foundation for the National Institutes of Health through the Grand Challenges in Global Health Initiative. Author attributions MQ designed and performed experiments, analyzed data and helped write the paper. FP designed the clinical components of the study. BN and JLJ designed and performed computational experiments. FP, DSK, JZ and DEK designed and performed siRNA experiments in samples from HIV patients. CF, QE, BJ, KB, SI, KR, IT, AP-T, DD, LF all performed experiments. GF designed experiments and developed PD-L1-Ig. JA, AC, HS, EJW designed and performed animal experiments, and analyzed data. WNH, BE and BDW conceived of the study and designed the experiments. WNH analyzed data and wrote the paper.

from HIV patients. CF, QE, BJ, KB, SI, KR, IT, AP-T, DD, LF all performed experiments. GF designed experiments and developed PD-L1-Ig. JA, AC, HS, EJW designed and performed animal experiments, and analyzed data. WNH, BE and BDW conceived of the study and designed the experiments. WNH analyzed data and wrote the paper. Figure 1 Transcriptional profiles of HIV-specific CD8+ T cells show coordinate upregulation of genes induced by PD-1 signaling (a) HIV viral load in controllers (grey circles) and progressors (black circles). Horizontal lines indicate median viral load in each cohort. (b) Genes differentially expressed in Gag-specific CD8+ T cells from controllers (grey bars) or progressors (black bars) ranked by moderated t–statistic. Each column represents an individual sample and each row an individual gene, colored to indicate normalized expression. The top 200 genes in either direction are shown. (c) IL-2 secretion from PD-1 expressing Jurkat cells cultured with inhibitory PD-1/CD3/CD28 beads (black bar) or control CD3/CD28 beads (grey bar) measured by ELISA (**P = 0.007). (d) Differentially expressed genes in PD-1 Jurkat cells cultured as in c. The top 100 differentially expressed genes from either condition are shown. (e) Enrichment analysis of PD-1 signature in HIV-specific CD8+ T cell profiles. The top 200 genes in the PD-1-specific signature were tested for enrichment in the rank-ordered list of genes differentially expressed in progressor vs. controller HIV-specific CD8+ T cells. X-axis indicates the t–statistic measured for each of the ~20,000 genes assayed in HIV-specific T cells, ranked in order of their differential expression in progressor vs. controller classes. Y-axis indicates the cumulative distribution of all genes (dotted lines) or of a set of 200 PD-1 signature genes (black line). Gene sets that are related to the class distinction on the X-axis would be expected to deviate from the dotted line (i.e. shifted towards the left if enriched in profiles of CD8+ T cells from progressors). (f) Enrichment of PD-1 signature in tetramer-sorted CD8+ T cells specific for different human viral pathogens. PD-1 signature genes were tested for enrichment by single-sample enrichment analysis in gene expression profiles from sorted tetramer+ CD8+ T cells specific for pathogens shown, or in naive CD8+ T cells. Each point represents the relative enrichment of PD-1 signature genes in an individual sample.

uman viral pathogens. PD-1 signature genes were tested for enrichment by single-sample enrichment analysis in gene expression profiles from sorted tetramer+ CD8+ T cells specific for pathogens shown, or in naive CD8+ T cells. Each point represents the relative enrichment of PD-1 signature genes in an individual sample. Y-axis indicates normalized enrichment score (*indicates P < 0.05; **P < 0.01; ***P < 0.001 by Wilcoxon ranked-sum test).

uman viral pathogens. PD-1 signature genes were tested for enrichment by single-sample enrichment analysis in gene expression profiles from sorted tetramer+ CD8+ T cells specific for pathogens shown, or in naive CD8+ T cells. Each point represents the relative enrichment of PD-1 signature genes in an individual sample. Y-axis indicates normalized enrichment score (*indicates P < 0.05; **P < 0.01; ***P < 0.001 by Wilcoxon ranked-sum test). Figure 2 Expression of BATF is upregulated by PD-1 and increased in exhausted T cells (a) Venn diagram representation of three transcription factors upregulated in Gag-specific T cells from HIV progressors and Jurkat cells after PD-1 ligation (t > 2.0). (b) BATF expression measured by real-time quantitative PCR in primary human CD4+ (left bars) and CD8+ (right bars) T cells cultured with CD3/CD28 beads (grey bars) or PDL1/CD3/CD28 beads (black bars) for 4 days. Data represent independent experiments with 4 – 10 normal donors and displayed as expression relative to CD3/28 condition (**P = 0.001; *P = 0.02; paired t test). (c) Relative BATF expression in arbitrary expression units from Affymetrix analysis of sorted naïve (CD62L+CD45RA+, white bars) or HIV gag-specific CD8+ T cell populations from controllers (grey bar) and progressors (black bar). (d) Batf expression measured by real-time quantitative PCR in LCMV-specific CD8+ T cells from mice infected with LCMV Armstrong (grey symbols), or LCMV clone 13 (black symbols) relative to naive (*P < 0.05; **P < 0.01). (e) PD-1 expression on LCMV-specific CD8+ T cells measured by flow cytometry following infection with the viruses indicated.

ured by real-time quantitative PCR in LCMV-specific CD8+ T cells from mice infected with LCMV Armstrong (grey symbols), or LCMV clone 13 (black symbols) relative to naive (*P < 0.05; **P < 0.01). (e) PD-1 expression on LCMV-specific CD8+ T cells measured by flow cytometry following infection with the viruses indicated. Figure 3 BATF inhibits T cell function (a) CFSE-labeled primary human CD4+ or CD8+ T cells from healthy volunteers transduced with a lentivirus expressing BATF (lower plot) or with control vector (upper plot) and cultured for 4 days with CD3/CD28 beads. (b) Summary data of proliferation (percent CFSEdimAnnexin−, upper plot), and cell death (percent Annexin V+, lower plot) in primary human CD4+ or CD8+ T cells (n = 14) transduced as in (a) and cultured for 4 days with CD3/CD28 beads. (c) IL-2 (left bars, P = 4.5 × 10−5) and IFN-γ (right bars, P=NS) secretion by primary human CD4+ T cells (n = 10) transduced as in (a) and cultured with CD3/CD28 beads. Data is shown normalized to the empty vector condition. (d) BATF expression in PD-1 expressing Jurkat cells lentivirally transduced with shGFP (control) or two separate shBATF sequences measured by western blot (upper panel) or quantitative PCR (lower panel). (e) IL-2 expression by PD-1 Jurkat cells transduced with shRFP (control, white bar) or shBATF (black, grey bars) cultured with no beads or either PD-1/CD3/CD28 or CD3/CD28 beads as indicated for 18 hours. Data shows IL-2 expression normalized (+/− SEM) measured by quantitative PCR (***P < 0.001; **P = 0.01). Data are normalized to β-ACTIN and presented as fold change with respect to unstimulated conditions. (f) Correlation between BATF silencing and IL-2 secretion in PD-1 expressing Jurkat cells transfected with five sequence independent shBATF constructs (black symbols) or a control hairpin (grey symbol) and cultured with PD-1/CD3/CD28 beads. BATF expression was measured by quantitative PCR and presented as fold-change relative to control hairpin.

F silencing and IL-2 secretion in PD-1 expressing Jurkat cells transfected with five sequence independent shBATF constructs (black symbols) or a control hairpin (grey symbol) and cultured with PD-1/CD3/CD28 beads. BATF expression was measured by quantitative PCR and presented as fold-change relative to control hairpin. Figure 4 BATF silencing improves HIV-specific T cell function (a) Efficacy of siRNA uptake in CD3+ T cells cultured with mixture of siRNA pool and fluorescent oligonucleotides (to monitor transduction) either with (black histogram) or without (grey histogram) electroporation. (b) Silencing of BATF by siRNA sequences targeting BATF in CD3+ T cells from a representative chronic progressor measured by quantitative PCR. Expression (mean, SEM) normalized to a housekeeping gene is presented as fold change relative to control siRNA (c–e) BATF silencing enhances HIV-specific cytokine secretion in CD8+ (c) and CD4+ (d,e) T cells from chronic progressors. PBMC depleted of CD4+ (a) or CD8+ (b,c) T cells were electroporated with siRNA pools targeting the genes indicated and cultured with or without HIV Gag peptides for four days, and IFN-γ (c,e) or IL-2 (d) measured using a highly-sensitive cytokine bead assay. In each figure the left panel shows a representative patient, and the right panel summary data (CD8+ responses, 26 HIV epitope responses in 4 subjects; CD4+ responses, HIV Gag peptide pool in 7 subjects). Cytokine levels shown are adjusted for background secretion, and statistical significance evaluated with the paired t test. (f) Proliferation of CFSE CD8+ T cells was measured by fraction of CFSEdim, CD25+ cells six days after transfection and peptide stimulation of PBMCs. Data represent nine HIV epitope-specific responses in 4 subjects.

Cell-based therapies, such as hematopoietic stem cell (HSC), islet cell, or hepatocyte transplants are in routine clinical practice1,2, while new treatment strategies implementing adult, embryonic, or induced pluripotent stem cells are in various stages of development3,4. In the field of cancer immunotherapy, early clinical trials infusing ex vivo-expanded tumor-specific T-lymphocytes have yielded promising results for the treatment of cancer and chronic infections5-7. Notably, following cell transfer, therapeutic cells often rely on the co-delivery of adjuvant drugs. These agents are designed to maximize donor cell efficacy and in vivo persistence, offset suppressive molecules at cell homing sites, or promote the differentiation of transferred cells into a therapeutically optimal phenotype. Examples include γc receptor cytokines5,8 or TGF-β signaling inhibitors9 in adoptive T-cell therapy, or the use of small-molecule drugs to boost immune reconstitution following HSC transplants10. However, these agents often require high and sustained systemic levels for efficacy. This leads to dose-limiting toxicities for these drugs due to their generally pleiotropic activity, which has restricted their clinical use11,12. One approach to focus adjuvant drug action on the transferred cells is to genetically engineer donor cells to secrete their own supporting factors13. However, regulatory and cost barriers of large-scale clinical grade vector production and safety testing, costly and lengthy cell culture, and technical challenges of efficient gene transfer hinder the implementation of clinical gene therapy protocols. More importantly, several emerging adjuvant therapies are based on small-molecule drugs that cannot be genetically encoded9,10. Here we describe an alternate strategy for adjuvant drug delivery in cell therapies, based on chemical conjugation of submicron-sized drug-loaded synthetic particles directly onto the plasma membrane of donor cells, enabling continuous pseudo-autocrine stimulation of transferred cells in vivo.

genetically encoded9,10. Here we describe an alternate strategy for adjuvant drug delivery in cell therapies, based on chemical conjugation of submicron-sized drug-loaded synthetic particles directly onto the plasma membrane of donor cells, enabling continuous pseudo-autocrine stimulation of transferred cells in vivo. Results Stable nanoparticle (NP) attachment to cell surfaces To stably couple synthetic drug carrier NPs to the surface of therapeutic cells, we exploited the fact that many cells exhibit high levels of reduced thiol groups on their surfaces14. Confirming prior reports, we detected substantial levels of free thiols on the surfaces of T-cells, B-cells, and HSCs, but low amounts on red blood cells (Fig. 1a). To link synthetic drug carriers to cells using these surface thiols, we utilized liposomes and liposome-like synthetic NPs 100-300 nm in diameter with a drug-loaded core and phospholipid surface layer, where the lipid bilayer surface of the particles included thiol-reactive maleimide headgroups (Supplementary Fig. 1). We achieved particle conjugation by a simple two-step process (Fig. 1b): donor cells were first incubated with NPs to permit maleimide-thiol coupling, followed by in situ PEGylation with thiol-terminated poly(ethylene glycol) (PEG) to quench residual reactive groups of the particles (Supplementary Fig. 2). With this approach, we could covalently link a substantial number of NPs with diameters in the 100-300 nm range to cell types used commonly in cell therapy, including CD8+ T lymphocytes or lineage-Sca-1+c-kit+ HSCs (Fig. 1c, left panels). Particles ranging from simple liposomes (with an aqueous drug-loaded core), to more complex multilamellar lipid NPs or lipid-coated polymer NPs15 (Fig. 1c, and Supplementary Figs. 1 and 3) were stably attached to live cells. Importantly, particle coupling was benign; coupling of up to 140 (±30) ∼200 nm-diameter multilamellar lipid NPs to the surface of cells was nontoxic (Supplementary Fig. 4), and blocked only 17.2% (± 8.7%) of the total available cell surface thiol groups (Supplementary Fig. 5). These findings are consistent with a simple calculation of the surface area occupied by the NPs: attachment of 150 particles each 200 nm in diameter would occlude only 3% of the surface of a typical 7 μm-diameter T-cell.

locked only 17.2% (± 8.7%) of the total available cell surface thiol groups (Supplementary Fig. 5). These findings are consistent with a simple calculation of the surface area occupied by the NPs: attachment of 150 particles each 200 nm in diameter would occlude only 3% of the surface of a typical 7 μm-diameter T-cell. Although liposomes and lipid-coated polymer particles spontaneously adsorbed to cell surfaces, we found that physically-adsorbed particles were removed during mild cell washing steps, while maleimide-linked particles remained stably bound to cells (Fig. 1d). Attachment of NPs to T-cells did not trigger spontaneous activation of the cells (Supplementary Fig. 6), and strikingly, particles bound to lymphocytes or HSCs remained localized at the cell surface as revealed by optical sectioning with confocal microscopy (Fig. 1c, and Supplementary Movies 1 and 2), and by flow cytometry internalization assays (Fig. 1e), even following extended in vitro stimulation (Fig. 1c, right panels). In contrast, we observed that phagocytic cells such as immature dendritic cells efficiently internalized maleimide-functionalized NPs after a short incubation (Fig. 1e). Although all three types of NPs tested here conjugated to lymphocytes with comparable efficiency, we chose to focus on ∼300 nm-diameter multilamellar lipid NPs (Supplementary Fig. 1b) for our subsequent in vitro functional and in vivo therapeutic studies, based on their high drug encapsulation efficiencies, week-long drug release profiles, and the lack of inflammatory responses elicited from innate immune cells exposed to the “empty” particles (Supplementary Figs. 7 and 8).

s (Supplementary Fig. 1b) for our subsequent in vitro functional and in vivo therapeutic studies, based on their high drug encapsulation efficiencies, week-long drug release profiles, and the lack of inflammatory responses elicited from innate immune cells exposed to the “empty” particles (Supplementary Figs. 7 and 8). NP conjugation does not compromise key cellular functions We next determined the maximal number of particles (without encapsulated drug cargo) that could be linked to cells without compromising key cellular functions, focusing on therapeutic cytotoxic T-cells that must be capable of forming an immunological synapse and killing target cells, proliferating, and secreting cytokines as part of their normal function. TCR-transgenic OT-1 CD8+ T-cells conjugated with up to 100 (±20) NPs per cell retained an unmodified proliferative response after co-culture with ovalbumin-pulsed dendritic cells; higher surface densities of particles began to inhibit T-cell proliferation (Fig. 2a, and Supplementary Fig. 9a,b). During cell division, surface-attached NPs segregated equally to daughter cells, reflected by a stepwise decrease in the mean fluorescence from cell-conjugated NPs with increasing number of divisions (Figs. 1c and 2a). Attachment of at least ∼100 particles/cell also did not impact T-cell recognition/killing of ovalbumin peptide-pulsed target cells or cytokine release profiles (Fig. 2b, and Supplementary Figure 9c). We next assessed the impact of cell surface-tethered NPs on T-cell transmigration across endothelial monolayers – a key capability of any therapeutic cell to efficiently infiltrate its target tissue. We first utilized an in vitro transwell co-culture system and quantified the migration of NP-conjugated T-lymphocytes across a membrane-supported confluent endothelial monolayer in response to a chemoattractant placed in the lower chamber. T-cells carrying 100 NPs/cell exhibited unaltered transmigration efficiencies compared to unmodified cells (Fig. 2c). After crossing the endothelial barrier, T-cells retained 83% (±3%) of their original NP cargo physically attached (Fig. 2d). (In comparative experiments, liposomes and lipid-coated PLGA particles could also be carried through endothelial layers by T-cells, though PLGA particles were not retained as well by transmigrating cells and showed a tendency to inhibit T-cell transmigration at high particle/cell loadings, Supplementary Fig. 10)

. 2d). (In comparative experiments, liposomes and lipid-coated PLGA particles could also be carried through endothelial layers by T-cells, though PLGA particles were not retained as well by transmigrating cells and showed a tendency to inhibit T-cell transmigration at high particle/cell loadings, Supplementary Fig. 10) To determine whether in vivo tissue homing of T-cells was affected by NP conjugation, we evaluated the tumor-homing properties of particle-conjugated lymphocytes. Subcutaneous EL4 tumors expressing membrane-bound Gaussia luciferase (extG-luc) and ovalbumin (EG7-OVA) or exG-luc alone were established on opposite flanks of C57Bl/6 mice. Tumor-bearing mice then received adoptive transfers of Firefly luciferase (F-luc)-transgenic OT-1 T-cells with or without surface-conjugated red-fluorescent NPs, or an i.v. injection of an equivalent dose of fluorescent particles alone. Particle-carrying OT-1 T-cells specifically trafficked to EL4-OVA tumors (Fig. 3a), and no difference in the tumor homing potential of particle-conjugated compared to unmodified OT-1 T-cells was observed (Fig. 3b, upper panel). Quantitative fluorescent particle imaging of EG7-OVA tumors demonstrated that NPs accumulated a mean 176-fold more efficiently at the tumor site when surface-attached to OT-1 T-cells compared to systemically infused free NPs, which were rapidly scavenged by the liver and the spleen (Fig. 3b,d). Flow cytometry analysis independently verified that T-cell infiltration of EG7-OVA tumors was quantitatively identical for particle-decorated and control OT-1 cells, and that the majority of particle-conjugated cells recovered from tumors still retained their NP cargo (Fig. 3a). In separate experiments using fluorescently-labeled OT-1 T-cells, we confirmed prominent infiltration of NP-decorated T-cells into EG7-OVA tumors in histological tumor sections examined by confocal microscopy, and NPs appeared surface-localized as observed in vitro (Fig. 3c, Supplementary Fig. 11, Supplementary Movie 3,4). Of note, the ability of lymphocytes to efficiently transfer surface-tethered NPs across endothelial barriers in vivo was not restricted to the abnormal endothelial lining16 found in tumor vasculature. When we linked particles to resting CCR7+CD62L+ B-cells (Supplementary Fig.

ementary Fig. 11, Supplementary Movie 3,4). Of note, the ability of lymphocytes to efficiently transfer surface-tethered NPs across endothelial barriers in vivo was not restricted to the abnormal endothelial lining16 found in tumor vasculature. When we linked particles to resting CCR7+CD62L+ B-cells (Supplementary Fig. 12) or central memory CD8+ T-cells (data not shown), particles were transported across the intercellular boundaries of high endothelial venules into lymph nodes – a poorly accessible compartment for systemically infused free NPs.

ementary Fig. 11, Supplementary Movie 3,4). Of note, the ability of lymphocytes to efficiently transfer surface-tethered NPs across endothelial barriers in vivo was not restricted to the abnormal endothelial lining16 found in tumor vasculature. When we linked particles to resting CCR7+CD62L+ B-cells (Supplementary Fig. 12) or central memory CD8+ T-cells (data not shown), particles were transported across the intercellular boundaries of high endothelial venules into lymph nodes – a poorly accessible compartment for systemically infused free NPs. Cell-Bound NPs enhance cytokine support of anti-tumor T-cells We next tested whether cell-bound adjuvant drug-loaded NPs could directly impart amplified therapeutic functions to their cellular carriers, using self/tumor-reactive CD8+ T-cell receptor transgenic Pmel-1 melanoma-specific T-cells to treat established, disseminated B16F10 melanoma lung and bone marrow mestastases17. We encapsulated a mixture of the cytokines IL-15 (converted to a superagonist (IL-15Sa) by pre-complexing with soluble IL-15Rα18) and IL-21 into multilamellar lipid NPs. These two interleukins cooperatively promote in vivo T-cell expansion and effector function when administered daily at high doses8. Particles ∼300 nm in diam. efficiently entrapped IL-15Sa and IL-21 and released bioactive cytokine over a seven-day period (Supplementary Fig. 13). These cytokine-loaded particles were conjugated to Click bettle red (CBR)-luciferase expressing CD8+ Pmel-1 effector T-cells. Particle-conjugated or control T-cells were infused into lymphodepleted mice bearing established Gaussia luciferase-expressing B16F10 melanoma lung and bone marrow tumors (Fig. 4a). Serial imaging of non-adjuvanted Pmel-1 T-cells showed a gradual CBR-luc signal decline following T-cell injection, consistent with poor in vivo T-cell expansion and persistence (Fig. 4a–c). Whereas a single systemic infusion of 5 μg free cytokine (4.03 μg IL-15Sa + 0.93 μg IL-21) given on the day of adoptive transfer did not significantly boost Pmel-1 proliferation (1.4-fold-higher CBR-luc signal on day 6, P = 0.32), the same cytokine dose loaded in cell-bound NPs elicited markedly amplified proliferation by Pmel-1 cells (81-fold higher peak photon count relative to unmodified Pmel-1 T-cells on day 6, P < 0.0001, Fig. 4a,c). Subsequent to a contraction period, cytokine NP-carrying T-cells displayed enhanced long-term persistence (14.8-fold and 4.7-fold higher photon counts than Pmel-1 T-cells alone at 16 and 30 days after T-cell infusion, respectively, P < 0.0001) and homed as CD44+CD62L+ central memory T-cells to lymph nodes and spleen (Fig. 4a,b, and Supplementary Fig. 14).

ytokine NP-carrying T-cells displayed enhanced long-term persistence (14.8-fold and 4.7-fold higher photon counts than Pmel-1 T-cells alone at 16 and 30 days after T-cell infusion, respectively, P < 0.0001) and homed as CD44+CD62L+ central memory T-cells to lymph nodes and spleen (Fig. 4a,b, and Supplementary Fig. 14). Notably, experiments comparing the in vivo proliferative response of T-cells bearing cytokine-loaded NPs vs. bystander tumor-homing T-cells showed that NP-released cytokines activated T-cells primarily in cis with limited paracrine stimulation of bystander cells (Supplementary Fig. 15). The adjuvant effect of T-cell-conjugated cytokine NPs was largely tumor antigen-independent (Supplementary Fig. 16b,d), consistent with earlier studies demonstrating antigen-independent proliferation of T-cells in response to IL-15,19 but there was no evidence of progressive T-cell clonality or leukemia formation in any treated animal imaged at late time points (data not shown). Importantly, cytokine-loaded particles co-injected but not attached to T-cells elicited a 4.9-fold higher peak Pmel-1 T-cell proliferation compared to the same cytokine dose administered in a non-encapsulated soluble form (day 6, P = 0.0052), but this stimulatory effect was still 11-fold (P < 0.0001) less than that obtained by linking the same number of cytokine-loaded NPs directly to the surface of the adoptively transferred T-cells (Supplementary Fig. 16c,d). Pmel-1 T-cells conjugated with “empty” NPs exhibited the same expansion/decline in vivo as unmodified Pmel-1 cells (Supplementary Figure 16a,d). All mice receiving cytokine NP-decorated Pmel-1 T-cells achieved complete tumor clearance (Fig. 4a,d), whereas treatment with Pmel-1 T-cells with or without systemic cytokine infusion at the same doses yielded only modest survival advantages (Fig. 4a,d). The in vivo tumor eradication potential of cytokine NP-conjugated Pmel-1 T lymphocytes was also investigated in animals bearing large, established subcutaneous B16F10 flank tumors. Animals treated with unmodified Pmel-1 T-lymphocytes uniformly succumbed to tumors within 30 d, whereas the infusion of cytokine NP-decorated Pmel-1 T-cells prevented tumor growth, with all animals alive 30 d after T-cell treatment (Supplementary Fig. 17).

ls bearing large, established subcutaneous B16F10 flank tumors. Animals treated with unmodified Pmel-1 T-lymphocytes uniformly succumbed to tumors within 30 d, whereas the infusion of cytokine NP-decorated Pmel-1 T-cells prevented tumor growth, with all animals alive 30 d after T-cell treatment (Supplementary Fig. 17). Enhanced HSC reconstitution via cell-bound adjuvant NPs Prompted by the substantial therapeutic benefits achieved with conjugation of cytokine-loaded particles to tumor-specific T-cells, we further examined the utility of this new adjuvant delivery approach in the context of hematopoietic stem cell transplantations. We chose the glycogen synthase kinase-3 β (GSK-3β) inhibitor TWS11920 as therapeutic cargo, based on reports that repeated high-dose bolus therapy of transplant recipients with glycogen synthase kinase-3 (GSK-3) inhibitors enhances the repopulation kinetics of donor HSCs10. Multilamellar lipid nanoparticles efficiently encapsulated this small-molecule drug, and slowly released it over a seven-day time window (Supplementary Fig. 13). We evaluated the in vivo repopulation capabilities of hematopoietic grafts supported by cell-bound TWS119-loaded NPs based on the whole body photon emission from Firefly luciferase-transgenic donor HSCs, and in separate experiments, by tracing the frequencies of GFP+ donor HSCs by flow cytometry. Following transplantation of lineage-Sca-1+c-kit+ HSCs from luciferase-transgenic donors into lethally-irradiated syngeneic recipients, a steady increase in whole body bioluminescent emission was observed originating from discrete foci over anatomic sites corresponding to the femurs, humeri, sternum and the spleen (Fig. 5a). Whereas a systemic TWS119 bolus injection (1.6 ng) at the time of transplantation did not significantly alter measured engraftment kinetics (Fig. 5a,b), the same TWS119 dose encapsulated in NPs surface-tethered to donor HSCs markedly enhanced reconstitution by HSC grafts (median 5.7-fold higher bioluminescence than systemic TWS119 after one week, P < 0.0001, Fig. 5a–c). Notably, animals in all treatment groups initially engrafted HSCs in both femurs and the sternum, indicating that NP conjugation did not compromise the intrinsic homing properties of donor HSCs.

nstitution by HSC grafts (median 5.7-fold higher bioluminescence than systemic TWS119 after one week, P < 0.0001, Fig. 5a–c). Notably, animals in all treatment groups initially engrafted HSCs in both femurs and the sternum, indicating that NP conjugation did not compromise the intrinsic homing properties of donor HSCs. While increasing the rate of initial reconstitution, conjugating TWS119 NPs onto HSCs did not affect their multilineage differentiation potential, reflected by a similar frequency of donor-derived GFP+ reconstituted cell types compared to control HSC grafts three months after transplantation (Fig. 5d). Thus, this simple approach for donor cell modification just prior to cell transfer can also augment hematopoietic stem cell transplants, a procedure in routine clinical practice. Discussion Cell therapies are in common clinical practice for certain indications (e.g., HSC and islet cell transplants) and are also being aggressively developed in other areas of medicine, such as adoptive T-cell therapy of cancer5-7. However, many cell therapy protocols rely on adjuvant drugs that act directly on the transferred therapeutic cells to maintain their function, phenotype, and/or lifespan. A ubiquitous challenge is the pleiotropic activity of many biological and small-molecule drugs, leading to toxicity or unwanted side effects following systemic exposure. This problem is illustrated by the use of interleukin-2 in the support of adoptive T-cell therapy of melanoma, where IL-2 provides important adjuvant signals to donor T-cells, but also elicits severe dose-limiting toxicity12.

mall-molecule drugs, leading to toxicity or unwanted side effects following systemic exposure. This problem is illustrated by the use of interleukin-2 in the support of adoptive T-cell therapy of melanoma, where IL-2 provides important adjuvant signals to donor T-cells, but also elicits severe dose-limiting toxicity12. Here, we devised a facile and generalizable strategy to robustly augment the therapeutic potential of cytoreagents, while limiting the potential for side effects from adjuvant drugs. We showed that adjuvant agent-releasing particles can be stably conjugated to cells without toxicity or interference with intrinsic cell functions, follow the characteristic in vivo migration patterns of their cellular vehicles and, ultimately, endow their carrier cells with substantially enhanced function using low drug doses that have no effect when given by traditional systemic routes. Prolonged retention of the particles on the surfaces of donor cells as shown here enables sustained drug release without concerns of premature degradation of the particle carrier or cargo due to internalization into degradative intracellular compartments. Notably, prior work has shown that particles ∼200 nm in diameter coated with anti-CD3 are readily internalized by T-cell lines21, suggesting that internalization of particles in the size range studied here is not impossible for lymphocytes per se, but rather that internalization may be tightly regulated at the cell surface– elucidating the mechanism(s) for prolonged particle retention on T-cell and HSC surfaces is an area for future study. Numerous reports have illustrated the potential of systemically-infused nanoparticle materials slowly releasing drug cargos to enhance the efficacy of therapeutic drugs, and this has led to the development of clinical products such as anthracycline-loaded liposomes for cancer therapy22. However, in the context of support for cell therapy, our data demonstrate that conjugation of drug-loaded particles directly to the donor cells increases their therapeutic impact significantly (here, ∼10-fold increases in peak T-cell expansion in an adoptive T-cell therapy model for particles attached to cells vs. the same particles systemically infused). This strategy does not require cell preconditioning and complements traditional genetic engineering or chemical biology approaches23 to augment or reprogram cell function.

increases in peak T-cell expansion in an adoptive T-cell therapy model for particles attached to cells vs. the same particles systemically infused). This strategy does not require cell preconditioning and complements traditional genetic engineering or chemical biology approaches23 to augment or reprogram cell function. Based on the wealth of available nanoparticle formulations tailored to deliver small molecule drugs, proteins, siRNA, or magnetic imaging agents24-27, the range of therapeutic or diagnostic cargos that can be attached to therapeutic cells likely extends far beyond the small molecules and recombinant proteins illustrated here.

n the wealth of available nanoparticle formulations tailored to deliver small molecule drugs, proteins, siRNA, or magnetic imaging agents24-27, the range of therapeutic or diagnostic cargos that can be attached to therapeutic cells likely extends far beyond the small molecules and recombinant proteins illustrated here. Our study further demonstrates the concept of cells as chaperones that actively direct drug-loaded nanoparticles into poorly accessible anatomical compartments. In the field of cancer therapy, targeting strategies functionalizing drugs or biomaterials with specific tumor-targeting ligands, such as antibodies, aptamers, small molecules or folic acid have been demonstrated to improve therapeutic efficacy28-30. However, these approaches generally rely on the initially passive accumulation of targeted therapeutics at tumor sites via the enhanced permeation and retention effect27, and it has been shown in some systems that targeting ligands do not change the overall tissue biodistribution of i.v.-delivered nanoparticle drug carriers, but rather enable those particles that do reach tumors to be more efficiently internalized by target cells28,31. In contrast, cellular nanoparticle vectors actively transmigrate the endothelial barrier and accumulate cell-attached cargo in tissues at >100-fold greater levels than systemically infused free particles. This profoundly altered biodistribution opens new venues, beyond existing cell therapies, for applications of cell products as actively targeting drug delivery “pharmacytes” or vaccine delivery tools.

ier and accumulate cell-attached cargo in tissues at >100-fold greater levels than systemically infused free particles. This profoundly altered biodistribution opens new venues, beyond existing cell therapies, for applications of cell products as actively targeting drug delivery “pharmacytes” or vaccine delivery tools. Methods Cell lines The murine melanoma cell line B16F10, the pancreatic islet endothelial cell line MS1, the thymoma cell line EL4 and EG7-OVA, an EL4 cell line stably transfected with the plasmid pAc-neo-OVA which carries a complete copy of chicken ovalbumin (OVA) mRNA, were all purchased from the American Type Culture Collection (ATCC). We purchased the Phoenix™ Eco retroviral packaging cell line from Orbigen. For bioluminescent in vivo tumor imaging we retrovirally transduced the B16F10, EL4 and EG7-OVA cell lines with a membrane-anchored form of the Gaussia luciferase (extG-Luc), provided to us by M. Sadelain (Memorial Sloan-Kettering Cancer Center), as described in the Supplementary Methods.

kaging cell line from Orbigen. For bioluminescent in vivo tumor imaging we retrovirally transduced the B16F10, EL4 and EG7-OVA cell lines with a membrane-anchored form of the Gaussia luciferase (extG-Luc), provided to us by M. Sadelain (Memorial Sloan-Kettering Cancer Center), as described in the Supplementary Methods. Mice and in vivo tumor models Animals were housed in the MIT Animal Facility. We performed all mouse studies in the context of an animal protocol approved by the MIT Division of Comparative Medicine following federal, state, and local guidelines. C57Bl/6 mice, C57Bl/6-Pmel-1-Thy1.1 mice, OT-1 OVA-TCR transgenic mice, and C57Bl/6-GFP-transgenic mice were all obtained from Jackson Laboratories. C57Bl/6 (H-2Kb, Thy-1.1) firefly luciferase (F-luc)-transgenic mice 32 were provided to us by M. van den Brink (Memorial Sloan-Kettering Cancer Center). For adoptive T-cell experiments with OVA-specific transgenic T cells, we subcutaneously injected C57Bl/6 mice with 4 × 106 EG7-OVA tumor cells into the right flank and 2 × 106 control EL4 cells into the left flank to generate equally sized s.c. tumors seven days later. We retrovirally transduced both tumor cell lines with extG-luc for bioluminescent imaging. To establish melanoma lung tumor metastases, we injected 1 × 106 B16F10-extG-luc tumor cells i.v. via the tail vein into C57Bl/6 mice one week before T cell treatment. On the day of adoptive Pmel-1 T cell transfer, we sublethally irradiated recipient mice with 500 cGy of total body irradiation from a 137Cs source. All mice were treated with a single infusion of 15 × 106 effector CD8+ T cells.

G-luc tumor cells i.v. via the tail vein into C57Bl/6 mice one week before T cell treatment. On the day of adoptive Pmel-1 T cell transfer, we sublethally irradiated recipient mice with 500 cGy of total body irradiation from a 137Cs source. All mice were treated with a single infusion of 15 × 106 effector CD8+ T cells. Preparation of primed T-cells for adoptive transfer and retroviral transduction Spleens were harvested, macerated over a filter, and resuspended in ACK lysing Buffer (Biosource, Rockville, MD). In all, we placed 3 × 106 splenocytes per milliliter in complete RPMI 1640 with 1 ng mL-1 IL-7 and 2 μg mL-1 Concavalin A (Calbiochem, La Jolla, CA), and incubated at 37°C. Two days later, we removed dead cells by Ficoll gradient (GE Healthcare) and isolated CD8+ cells using a mouse CD8 Negative Isolation Kit (Stemcell Technologies). We then preloaded 1 mL per well of concentrated retrovirus (see Supplementary Methods) on six-well non-tissue culture treated dishes coated with RetroNectin (TakiraBio) and incubated them at 37°C incubation for 1 h. An equal volume of isolated T cells (3 × 106 cells mL-1 substituted with 50 IU hIL-2 mL-1) was added and centrifuged at 2000 × g for 30 min. 6 h after spinoculation, 1 mL of fresh, prewarmed RPMI, containing 50 IU hIL-2 (Chiron) was added. We used T cells for adoptive transfer experiments one day after gene transfer.

. An equal volume of isolated T cells (3 × 106 cells mL-1 substituted with 50 IU hIL-2 mL-1) was added and centrifuged at 2000 × g for 30 min. 6 h after spinoculation, 1 mL of fresh, prewarmed RPMI, containing 50 IU hIL-2 (Chiron) was added. We used T cells for adoptive transfer experiments one day after gene transfer. NP conjugation with cells and in situ PEGylation Detailed information on nanoparticle and liposome synthesis as well as cytokine/small molecule particle loading is included in the Supplementary Methods. We resuspended 60 × 106 cells mL-1 in serum-free X-Vivo 10 medium (Cambrex) following two PBS washes. We then added an equal volume of NPs in nuclease-free water, with 1200/600/300/or 150 NPs/T cell (resulting in 139±29/128±23/100±21 or 75±32 surface-tethered particles/T-cell, respectively, after cell washes and PEGylation), and incubated at 37°C for 30 min with gentle agitation every 10 min. After a PBS wash to separate cells from unbound particles, we quenched residual maleimide groups present on cell-bound particles by incubation of 3 × 106 cells mL-1 with 1 mg mL-1 thiol-terminated 2Kda poly(ethylene glycol) (PEG, Laysan Bio) at 37°C for 30 min in complete RPMI medium, followed by 2 PBS washes to remove unbound PEG.

ash to separate cells from unbound particles, we quenched residual maleimide groups present on cell-bound particles by incubation of 3 × 106 cells mL-1 with 1 mg mL-1 thiol-terminated 2Kda poly(ethylene glycol) (PEG, Laysan Bio) at 37°C for 30 min in complete RPMI medium, followed by 2 PBS washes to remove unbound PEG. We determined 1 mg mL-1 thiol-PEG as the optimal concentration required to quench all remaining maleimide groups displayed on NPs after cell conjugation based on no significant FACS signal following a 30 min incubation with 70 mg ml-1 bodipy-tagged cysteine (generated from reduction of disulfide bond in bodipy L-cystine (Invitrogen; Carlsbad, CA) with 15 molar excess of TCEP (Thermo Scientific; Rockford, IL) for 45 min at RT). The nanoparticle binding efficiency of maleimide-functionalized (50 mole% maleimide MPB-PE in the lipid fraction) multilamellar lipid NPs to effector T lymphocytes was 33.4% (± 6.9%), when incubating 500 particles/T-cell, as determined by high magnification confocal microscopy imaging of 30 single T cell z-stacks. We distinguished between surface-conjugated and internalized NPs from by flow cytometry internalization assay, described in the Supplementary Methods. Functional in vitro T-cell and HSC assays, HSC transplantation, in vivo bioluminescence and fluorescence imaging, NP biodistribution assay, flow cytometry and confocal microscopy Detailed information on in vitro T cell and HSC assay, transplantation, serial bioluminescent imaging and confocal microscopy assays are included in the Supplementary Methods.

assays, HSC transplantation, in vivo bioluminescence and fluorescence imaging, NP biodistribution assay, flow cytometry and confocal microscopy Detailed information on in vitro T cell and HSC assay, transplantation, serial bioluminescent imaging and confocal microscopy assays are included in the Supplementary Methods. Supplementary Material 1 2 This work was supported in part by the National Institute of Health (CA140476), the National Science Foundation (MRSEC award DMR-02-13282), Cancer Center Support (core) grant P30-CA14051 from the National Cancer Institute, and a gift to the Koch Institute by Curtis and Cathy Marble. DJI is an investigator of the Howard Hughes Medical Institute. Author Contributions: M.T.S. designed and conducted all experiments and wrote the manuscript. J.J.M. assisted in T-cell transmigration assays, optimization of multilamellar lipid NP synthesis, and in vivo NP biodistribution assays. S.H.U. assisted optimization of multilamellar lipid NP synthesis. A.B. assisted in initial in vitro T cell assays, collected electron microscopy images, and contributed experimental suggestions. D.J.I. supervised all experiments and wrote the manuscript. Competing financial interest: The authors declare no competing financial interests.

Author Contributions: M.T.S. designed and conducted all experiments and wrote the manuscript. J.J.M. assisted in T-cell transmigration assays, optimization of multilamellar lipid NP synthesis, and in vivo NP biodistribution assays. S.H.U. assisted optimization of multilamellar lipid NP synthesis. A.B. assisted in initial in vitro T cell assays, collected electron microscopy images, and contributed experimental suggestions. D.J.I. supervised all experiments and wrote the manuscript. Competing financial interest: The authors declare no competing financial interests. Figure 1 Stable conjugation of nanoparticles (NPs) to the surfaces of T-cells and HSCs via cell-surface thiols. (a) Flow cytometry analysis of cell surface thiols on mouse splenocytes detected by fluorophore-conjugated malemide co-staining with lineage surface markers for erythrocytes (Ter-119), T-cells (CD3), B-cells (B220) and hematopoietic stem cells (c-kit). (b) Schematic of maleimide-based conjugation to cell surface thiols. (c) Confocal microscopy images of CD8+ effector T-cells and lineage-Sca-1+c-kit+ HSCs immediately following conjugation with fluorescent DiD-labeled multilamellar lipid NPs (left panel) and after four day in vitro expansion (right panel). Scale bars, 2 μm. (d) Flow cytometry analysis of CD8+ T-cells after incubation with DiD-labeled multilamellar lipid NPs synthesized with or without maleimide-headgroup lipids. (e) Quantification of nanoparticle internalization. Immature dendritic cells (DCs), effector CD8+ T-cells, or HSCs were conjugated with carboxyfluorescein-tagged maleimide-bearing liposomes. Extracellular trypan blue quenching was used to differentiate surface-bound and internalized liposomes immediately following conjugation or after four days in culture.

ernalization. Immature dendritic cells (DCs), effector CD8+ T-cells, or HSCs were conjugated with carboxyfluorescein-tagged maleimide-bearing liposomes. Extracellular trypan blue quenching was used to differentiate surface-bound and internalized liposomes immediately following conjugation or after four days in culture. Figure 2 Nanoparticle conjugation does not impact key T-cell functions. OT-1 ova-specific CD8+ effector T-cells were conjugated with 100 DiD-labeled multilamellar lipid NPs per cell or left unmanipulated as controls. (a) CFSE dilution of unmodified or NP-conjugated T-cells stimulated in vitro with mature ova peptide-pulsed dendritic cells. DiD Mean Fluorescence Intensities (MFI) for distinct CFSE lymphocytes populations are indicated on the right. (b) Standard 4 h 51Cr release assay comparing cytotoxicity of unmanipulated (open symbols) and particle-conjugated (filled symbols) OT-1 cells targeting ova peptide-pulsed (circles) or control (triangles) EL4 tumor cells. (c,d) Transmigration of OT-1 T-cells (with or without surface-bound particles) seeded onto MS1 endothelial cell monolayers in the upper well of a transwell chamber, following addition of the chemoattractant MCP-1 to the lower chamber. The fraction of transmigrating T-cells (c) and the profile of cell-bound NP fluorescence before (UW) and after (LW) transmigration (d) were quantitated by flow cytometry. (DiD MFI±s.e.m. from triplicate samples shown in blue).

of a transwell chamber, following addition of the chemoattractant MCP-1 to the lower chamber. The fraction of transmigrating T-cells (c) and the profile of cell-bound NP fluorescence before (UW) and after (LW) transmigration (d) were quantitated by flow cytometry. (DiD MFI±s.e.m. from triplicate samples shown in blue). Figure 3 Nanoparticle-decorated T cells efficiently carry surface-tethered NPs into antigen-expressing tumors. (a,b) Comparative whole-animal in vivo bioluminescence (tumors, T-cells) and fluorescence imaging (NPs) of mice bearing established s.c. Gaussia luc-expressing EG7-OVA and EL4 tumors on opposite flanks, two days after i.v. infusion of firefly luc-transgenic Thy1.1+ effector OT-1 T-cells (with or without attached DiD-labeled NPs), or an equivalent number of free NPs. Thy1.1+ OT-1 T-cells recovered from the EG7-OVA tumors were analyzed for surface-bound DiD NPs by flow cytometry (a), and the mean bioluminescent T-cell and fluorescent NP signals from groups of 6 mice are shown in (b). NS, no significance. (c) In an independent experiment, CellTracker green-labeled OT-1 T-cells conjugated with rhodamine-labeled NPs were transferred into mice bearing established s.c. EG7-OVA tumors, and tumors were excised and sectioned for confocal histological analysis two days later. Scale bar, 10 μm. A higher magnification image of NP-carrying tumor infiltrating T-cells is shown in the right panel. Scale bar, 1.5 μm. Yellow arrowheads highlight evidence for surface localization of NPs. Shown is 1 of 2 independent experiments. (d) Groups of 3 C57Bl/6 mice bearing s.c. EG7-OVA tumors were i.v. injected with 15 × 106 OT-1 effector T-cells bearing surface-conjugated with DiD-labeled NPs (100 NPs/cell, filled bars), an equivalent number of DiD-labeled particles alone (open bars). After 48 h indicated tissues were removed, weighed, and macerated with scissors. We quantified specific DiD tissue fluorescence for each organ using the IVIS Spectrum imaging system and calculated the mean percentage of injected dose per gram of tissue (%ID g-1) as final readout (d). Data shown are pooled from three independent experiments.

issues were removed, weighed, and macerated with scissors. We quantified specific DiD tissue fluorescence for each organ using the IVIS Spectrum imaging system and calculated the mean percentage of injected dose per gram of tissue (%ID g-1) as final readout (d). Data shown are pooled from three independent experiments. Figure 4 Pmel-1 T-cells conjugated with IL-15Sa/IL-21-releasing NPs robustly proliferate in vivo and eradicate established B16 melanomas. Lung and bone marrow tumors were established by tail vein injection of 1×106 Gaussia luciferase-expressing B16F10 cells in C57Bl/6 mice. Tumor-bearing animals were treated after 1 week by sublethal irradiation followed by i.v. infusion of 10×106 Click beetle red luciferase-expressing Vβ13+CD8+ Pmel-1 T-cells. One group of mice received Pmel-1 T-cells conjugated with 100 NPs/cell carrying a total dose of 5 μg IL-15Sa/IL-21 (4.03 μg IL-15Sa + 0.93 μg IL-21), control groups received unmodified Pmel-1 T-cells and a single systemic injection of the same doses of IL-15Sa/IL-21 or Pmel-1 T-cells alone. (a) Dual longitudinal in vivo bioluminescence imaging of Gaussia luc-expressing B16F10 tumors and CBR-luc-expressing Pmel-1 T-cells. (b) Frequencies of Vβ13+CD8+ Pmel-1 T-cells recovered from pooled lymph nodes of representative animals 16 days after T-cell transfer. (c) CBR-luc T-cell signal intensities from sequential bioluminescence imaging every two days after T-cell transfer. Every line represents one animal with each dot showing the whole animal photon count. (d) Survival of animals following T-cell therapy illustrated by Kaplan-Meier curves. Shown are six mice/treatment group pooled from three independent experiments.

equential bioluminescence imaging every two days after T-cell transfer. Every line represents one animal with each dot showing the whole animal photon count. (d) Survival of animals following T-cell therapy illustrated by Kaplan-Meier curves. Shown are six mice/treatment group pooled from three independent experiments. Figure 5 HSCs carrying GSK-3β inhibitor-loaded nanoparticles reconstitute recipient animals with rapid kinetics following bone marrow transplants without affecting multilineage differentiation potential. (a,b) Engraftment kinetics of luciferase-transgenic HSC grafts in lethally-irradiated nontransgenic syngeneic recipients. Mice were treated with a single bolus injection of the GSK-3β inhibitor TWS119 (1.6 ng) on the day of transplantation, an equivalent TWS119 dose encapsulated in HSC-attached NPs, or no exogenous adjuvant compounds. Transplanted mice were imaged for whole-body bioluminescence every seven days for three weeks. Shown are representative IVIS images (a) and whole animal photon counts (b) for nine mice total/treatment condition. (c) Percentage of donor-derived cells two weeks after transplantation of GFP+ HSCs into lethally-irradiated recipients with or without TWS119 adjuvant drug. *P < 0.001. (d) Average frequency of donor-derived GFP+ B-cells, T-cells, and myeloid cells in recipient mice three months after transplantation. five mice/group were analyzed.

ibitor currently in clinical development15. PF02341066 abrogated both p-Akt and p-S6RP signals as well as tumor growth (Fig. 3c,d). These results suggest that c-Met elevation is one mechanism underlying the growth of recurrent tumors that have escaped oncogenic PIK3CA addiction but remain dependent on the PI3K pathway. Because two of three GDC-0941-insensitive recurrent tumors featured c-Myc amplification (Fig. 3b) and overexpression (Supplementary Fig. 14), and given the known role of c-Myc functioning downstream of the PI3K pathway16, we hypothesized that c-Myc elevation might contribute to the recurrence of tumors that were resistant to PI3K inhibition. Further analyses of c-Myc for DNA copy number as well as both mRNA and protein levels in a large cohort of recurrent tumors (Supplementary Fig. 7–10) demonstrated that c-Myc elevation is a frequent event selected in recurrent tumors following sustained PIK3CAH1047R inactivation. To test whether c-Myc elevation contributes to tumor recurrence in a PI3K pathway-independent manner, we examined the effects of c-Myc knockdown by short hairpin RNAs (sh-Myc1 and sh-Myc2) on the growth of recurrent tumors transplanted in the mammary fat pads of immunodeficient mice. Knockdown of c-Myc dramatically reduced tumor incidence and extended the time to tumor onset (Fig. 4a,b). Conversely, enforced expression of c-Myc or c-MycT58A, a more stable version of c-Myc17, rendered otherwise PIK3CAH1047R-dependent tumors able to grow in the absence of doxycycline (Fig. 4c). Moreover, these c-Myc- or c-MycT58A-expressing tumors were resistant to GDC-0941 treatment (Fig. 4f and Supplementary Fig. 15). Together, these data suggest that c-Myc elevation is a mechanism that renders tumors free of addiction to PIK3CAH1047R and provides resistance to PI3K inhibition.

tumors on doxycycline, a markedly increased number of TUNEL-positive cells were observed in tumors after doxycycline removal (Fig. 1e). These data indicate that reduced cellular proliferation and increased apoptosis are responsible for the initial phase of tumor regression following downregulation of oncogenic PIK3CA. To determine whether the continued inactivation of oncogenic p110αH1047R resulted in sustained regression of mammary carcinomas initiated by the expression of PIK3CAH1047R, we followed a large cohort of tumors after doxycycline withdrawal for up to 6 months. We found that one-third of tumors rapidly and completely regressed to a non-palpable state within 1–2 months following doxycycline withdrawal with no re-growth (Fig. 2a and Supplementary Fig. 4a), indicating that these tumors remained dependent on p110αH1047R for their maintenance. While a small fraction of tumors regressed partially and did not resume growth following doxycycline removal, about two-thirds of tumors partially regressed but then resumed growth in the absence of doxycycline (Fig. 2a and Supplementary Fig. 4b). We confirmed that all recurrent tumors showed sustained downregulation of the PIK3CAH1047R transgene and its protein product (Fig. 2b). Thus, PIK3CAH1047R-initiated mammary tumors frequently failed to regress completely upon PIK3CAH1047R inactivation and recurred in a PIK3CAH1047R-independent manner.

nce of doxycycline (Fig. 4c). Moreover, these c-Myc- or c-MycT58A-expressing tumors were resistant to GDC-0941 treatment (Fig. 4f and Supplementary Fig. 15). Together, these data suggest that c-Myc elevation is a mechanism that renders tumors free of addiction to PIK3CAH1047R and provides resistance to PI3K inhibition. In our model, PIK3CAH1047R-induced tumors have three potential outcomes in response to PI3K inhibition (Fig. 4g). For those tumors that escape oncogene addiction and recur, c-Myc elevation represents a potential resistance mechanism with respect to current PI3K-targeted therapies in clinical trials. To explore whether PIK3CA mutations and c-MYC elevation coexist in human breast cancer, we analyzed several breast cancer datasets containing both PIK3CA mutation status and c-MYC copy number or expression data18–21. Among these cohorts, substantial fractions of PIK3CA mutation positive tumors have increased c-MYC copy number as well as mRNA and c-MYC protein levels2223 (Supplementary Fig. 16 and Supplementary Table 3). Taken together, our findings suggest that aberrant elevation of c-MYC represents a potential mechanism by which tumors develop resistance to PI3K inhibition, and thus combination therapies targeting both PI3K and c-MYC may be necessary to circumvent resistance to PI3K-targeted therapy.

g. 16 and Supplementary Table 3). Taken together, our findings suggest that aberrant elevation of c-MYC represents a potential mechanism by which tumors develop resistance to PI3K inhibition, and thus combination therapies targeting both PI3K and c-MYC may be necessary to circumvent resistance to PI3K-targeted therapy. METHODS Transgenic mice We cloned human PIK3CAH1047R into the BamHI site of pTRE2 (Clontech) and inserted an IRES-firefly luciferase sequence downstream of PIK3CAH1047R to generate the TetO-PIK3CAH1047R-IRES-luciferase plasmid. We linearized the plasmid and gel-purified the released fragment for injection into fertilized oocytes from superovulated FVB mice at the transgenic core facility at the Brigham & Women’s Hospital, Boston. We crossed TetO-PIK3CAH1047R mice with MMTV-rtTA (MTB) mice (generously provided by L. Chodosh) to produce mice with inducible PIK3CAH1047R transgene expression in mammary glands (iPIK3CAH1047R). We administered iPIK3CAH1047R mice with doxycycline in their drinking water (2mg/ml). We performed all mouse experiments in accordance with protocols approved by the Institutional Animal Care and Use Committees of Dana-Farber Cancer Institute and Harvard Medical School. Bioluminescence imaging We anesthetized mice with ketamine and xylazine, and administeredmice with D-luciferin (Promega) intraperitoneally to monitored luciferase gene expression in vivo. We analyzed images using KODAK Molecular Imaging Software (version 4.5.0b6 SE).

ementary Fig. 4b). We confirmed that all recurrent tumors showed sustained downregulation of the PIK3CAH1047R transgene and its protein product (Fig. 2b). Thus, PIK3CAH1047R-initiated mammary tumors frequently failed to regress completely upon PIK3CAH1047R inactivation and recurred in a PIK3CAH1047R-independent manner. We next examined whether the PI3K pathway remained active in recurrent tumors, thus compensating for the loss of PIK3CAH1047R expression. Western blot analyses of six paired primary and recurrent tumors revealed that, while both p-AKT and p-S6RP signals were robust in all six primary tumors maintained on doxycycline, in three recurrent tumors these signals were maintained at comparably high levels, but were reduced substantially in the other three recurrent tumors (Fig. 2b). These six recurrent tumors were then transplanted into the mammary fat pads of athymic mice, and the tumor-bearing recipients treated with GDC-0941, a pan-Class I PI3K inhibitor currently in clinical trials11,12. Three recurrent tumors (RCT-D782, RCT-E565 and RCT-E302), all of which retained high levels of both p-AKT and p-S6RP, were sensitive to GDC-0941 treatment (Fig. 2c upper panels). In contrast, the three recurrent tumors (RCT-E473, RCT-D419 and RCT-C658), which showed reduced p-AKT and p-S6RP signals, were resistant to GDC-0941 (Fig. 2c lower panels). These data suggest that some recurrent tumors escaped addiction to the oncogenic PIK3CA but remained dependent on the PI3K pathway, while others acquired the ability to grow independently of both the PIK3CA oncogene and the PI3K pathway.

METHODS Transgenic mice We cloned human PIK3CAH1047R into the BamHI site of pTRE2 (Clontech) and inserted an IRES-firefly luciferase sequence downstream of PIK3CAH1047R to generate the TetO-PIK3CAH1047R-IRES-luciferase plasmid. We linearized the plasmid and gel-purified the released fragment for injection into fertilized oocytes from superovulated FVB mice at the transgenic core facility at the Brigham & Women’s Hospital, Boston. We crossed TetO-PIK3CAH1047R mice with MMTV-rtTA (MTB) mice (generously provided by L. Chodosh) to produce mice with inducible PIK3CAH1047R transgene expression in mammary glands (iPIK3CAH1047R). We administered iPIK3CAH1047R mice with doxycycline in their drinking water (2mg/ml). We performed all mouse experiments in accordance with protocols approved by the Institutional Animal Care and Use Committees of Dana-Farber Cancer Institute and Harvard Medical School. Bioluminescence imaging We anesthetized mice with ketamine and xylazine, and administeredmice with D-luciferin (Promega) intraperitoneally to monitored luciferase gene expression in vivo. We analyzed images using KODAK Molecular Imaging Software (version 4.5.0b6 SE). Western blotting We prepared lysates for mammary glands, mammary tumors or tumor cells in ice-cold RIPA buffer (Sigma-Aldrich) containing protease inhibitor cocktail (Roche). We cleared lysates by centrifugation before subjecting them to separation on SDS-PAGE gels and performed western blot assays as described previously5 with antibodies against phospho-AKT (Ser473 or Thr 308), AKT, phospho-S6 ribosomal protein (Ser235/Ser236), S6 ribosomal protein, and c-Met (Cell Signaling Technology), c-Myc (Santa Cruz Biotechnology) and vinculin (Sigma-Aldrich). We used immunofluorescently labelled anti-mouse IgG (Rockland Immunochemicals) and anti-rabbit IgG (Molecular Probes) to visualize western blots on an Odyssey scanner (Li-Cor, Lincoln, NE).

duced p-AKT and p-S6RP signals, were resistant to GDC-0941 (Fig. 2c lower panels). These data suggest that some recurrent tumors escaped addiction to the oncogenic PIK3CA but remained dependent on the PI3K pathway, while others acquired the ability to grow independently of both the PIK3CA oncogene and the PI3K pathway. To search for genomic aberrations associated with this recurrence, we carried out mouse SNP array analyses of six recurrent tumors. A GDC-0941 sensitive tumor, RCT-E565, had a narrow amplification region encompassing c-Met (Fig. 3a) and also harbored a single copy loss of the tumor suppressor gene Cdkn2a (Supplementary Fig. 5a). Notably, two of three tumors that were resistant to GDC-0941 had a common amplification on chromosome 15 spanning 1.48 Mb (Chromosome 15:61,271,320–62,750,432), which contains the coding sequence for a single gene, c-Myc (Fig. 3b and Supplementary Fig. 6). In addition to c-Myc amplification, RCT-C658 also carried an amplification encompassing the Mdm2 oncogene (Supplementary Fig. 5b). Further analyses of c-Met, c-Myc and Mdm2 in a large collection of recurrent tumors showed that these oncogenes were upregulated in various fractions of recurrent tumors (Supplementary Fig. 7–12 and Supplementary Table 2). These data demonstrate that several of the most common gain- or loss-of-function genetic events in human cancers were recapitulated in this mouse tumor model.

on of recurrent tumors showed that these oncogenes were upregulated in various fractions of recurrent tumors (Supplementary Fig. 7–12 and Supplementary Table 2). These data demonstrate that several of the most common gain- or loss-of-function genetic events in human cancers were recapitulated in this mouse tumor model. Since c-Met is a receptor tyrosine kinase known to activate the PI3K pathway via ERBB3 and GAB113,14, we tested whether c-Met amplification contributes to increased PI3K activity and tumor growth in the absence of PIK3CAH1047R expression. We confirmed that the RCT-E565 tumor, but not its parental primary tumor PMT-E565, had elevated c-Met mRNA and protein levels (Supplementary Fig. 13). We then treated mice bearing RCT-E565 tumor transplants with PF02341066, a c-Met inhibitor currently in clinical development15. PF02341066 abrogated both p-Akt and p-S6RP signals as well as tumor growth (Fig. 3c,d). These results suggest that c-Met elevation is one mechanism underlying the growth of recurrent tumors that have escaped oncogenic PIK3CA addiction but remain dependent on the PI3K pathway.

More than 25% of breast cancers harbor somatic mutations in the PIK3CA-encoded p110α catalytic subunit of phosphatidylinositol 3-kinase (PI3K)1–4. These mutations usually occur in the helical region (E545K and E542K) or the kinase domain (H1047R) of p110α, with H1047R being the most common mutation (>50% of cases). Several experimental models have demonstrated that these tumor-associated PIK3CA mutations result in constitutive p110α activation and oncogenic transformation5–9, making the PIK3CA oncogene a major target for cancer therapy.

e kinase domain (H1047R) of p110α, with H1047R being the most common mutation (>50% of cases). Several experimental models have demonstrated that these tumor-associated PIK3CA mutations result in constitutive p110α activation and oncogenic transformation5–9, making the PIK3CA oncogene a major target for cancer therapy. To study the effects of mutational activation of PI3K on breast tumorigenesis in vivo and to identify potential mechanisms of resistance to PI3K inhibition, we generated a transgenic mouse line expressing human PIK3CAH1047R in which transgene expression is under the control of a tetracycline-inducible promoter (TetO). PIK3CAH1047R expression is coupled with a luciferase reporter allowing transgene expression to be followed in vivo (Fig. 1a). To drive mammary-specific expression of PIK3CAH1047R, we crossed two tetO-PIK3CAH1047R founders (HR-2239 and HR-2251) to a previously described MMTV-rtTA (MTB) line10. The resulting bitransgenic MTB/tetO-PIK3CAH1047R mice were designated iPIK3CAH1047R. Quantitative RT-PCR analyses of mammary tissues isolated from bitransgenic females revealed that doxycycline treatment led to a substantial increase in PIK3CAH1047R expression as well as luciferase reporter activity, whereas endogenous mouse Pik3ca expression remained unaffected (Supplementary Fig. 1a,b). As mice derived from both iPIK3CAH1047R founder lines showed comparable mammary gland-specific and doxycycline-dependent transgene expression, the MTB/HR-2239 line was used for all subsequent experiments.

ferase reporter activity, whereas endogenous mouse Pik3ca expression remained unaffected (Supplementary Fig. 1a,b). As mice derived from both iPIK3CAH1047R founder lines showed comparable mammary gland-specific and doxycycline-dependent transgene expression, the MTB/HR-2239 line was used for all subsequent experiments. To determine whether expression of PIK3CAH1047R can initiate transformation of mammary epithelium, we analyzed mammary glands isolated from iPIK3CAH1047R females treated with doxycycline for 4 weeks. Histological examination showed increased mammary ductal side-branching and enlarged focal nodular structures filled with hyperproliferative cells characteristic of early neoplastic lesions (Supplementary Fig. 2a,b). Immunohistochemical (IHC) analyses demonstrated strong p-AKT signals in proliferating epithelial cells in the mammary glands from doxycycline-treated mice (Supplementary Fig. 2c), indicating activation of PI3K signaling in response to the induction of PIK3CAH1047R. Consistent with the phenotype noted above, chronic doxycycline induction of the PIK3CAH1047R transgene in bitransgenic mice resulted in mammary tumors with 95% penetrance and a mean latency of 7 months (Fig. 1b). These primary tumors displayed heterogeneous pathological phenotypes, including adenocarcinomas and adenosquamous carcinomas (Fig. 1c and Supplementary Table 1). In contrast, no tumors were observed in any of the control groups over the same time period (Fig. 1b). Thus, sustained induction of oncogenic PIK3CA expression leads to mammary tumor formation.

rogeneous pathological phenotypes, including adenocarcinomas and adenosquamous carcinomas (Fig. 1c and Supplementary Table 1). In contrast, no tumors were observed in any of the control groups over the same time period (Fig. 1b). Thus, sustained induction of oncogenic PIK3CA expression leads to mammary tumor formation. To examine whether established tumors require continued PIK3CAH1047R expression to maintain their malignant state, we withdrew doxycycline from a cohort of tumor-bearing mice. All tumors exhibited regression during the first week following doxycycline removal. The suppression of PIK3CAH1047R expression following doxycycline withdrawal was confirmed by RT-PCR in primary tumors (Supplementary Fig. 3). IHC analyses revealed dramatically reduced levels of both p-Akt and p-S6RP in doxcycycline-off tumors as compared to those maintained on doxycycline (Fig. 1d). Moreover, while a robust Ki67 signal was detected in tumors maintained on doxycycline, the number of proliferating cells significantly decreased in tumors following doxycycline withdrawal (Fig. 1e). Conversely, while only a few apoptotic cells were detected in tumors on doxycycline, a markedly increased number of TUNEL-positive cells were observed in tumors after doxycycline removal (Fig. 1e). These data indicate that reduced cellular proliferation and increased apoptosis are responsible for the initial phase of tumor regression following downregulation of oncogenic PIK3CA.

r236), S6 ribosomal protein, and c-Met (Cell Signaling Technology), c-Myc (Santa Cruz Biotechnology) and vinculin (Sigma-Aldrich). We used immunofluorescently labelled anti-mouse IgG (Rockland Immunochemicals) and anti-rabbit IgG (Molecular Probes) to visualize western blots on an Odyssey scanner (Li-Cor, Lincoln, NE). Histology and immunohistochemistry We fixed tumors in formalin overnight before paraffin embedding. Paraffin blocks were sectioned, and stained with hematoxylin and eosin at the DF/HCC Rodent Histopathology Core. We performed immunohistochemistry using the antibodies: Ki67 (Vector), phospho-AktSer473 (Invitrogen) and phospho-S6 Ribosomal Protein (Cell Signaling). We performed TUNEL assay using the ApoTag Plus Peroxidase in situ TUNEL Apoptosis Kit (Millipore) according to the manufacturer’s instructions. Mouse SNP analyses We isolated genomic DNAs from mammary tissues or tumors using the Allprep DNA/RNA Kit (Qiagen). SNP array analyses with Mouse Diversity Genotyping Arrays (Affymetrix) were performed at the Microarray Core at Dana-Farber Cancer Institute. The SNP data (GEO accession number, GSE27691) were analyzed using a SNP microarray copy number application24 in the software suite, dChip (http://biosun1.harvard.edu/complab/dchip/), to compare positions of copy difference between a normal tissue sample from the inbred strain of mouse used in this study, and each of the tumor samples from the same inbred strain.

) were analyzed using a SNP microarray copy number application24 in the software suite, dChip (http://biosun1.harvard.edu/complab/dchip/), to compare positions of copy difference between a normal tissue sample from the inbred strain of mouse used in this study, and each of the tumor samples from the same inbred strain. Tumor cell culture and viral transduction We isolated tumors and dissociated them into single cells as described25 with the exception that the cells were cultured in DMEM/F12 supplemented with 0.5% FBS and 10ng/ml EGF and doxycycline (2µg/ml). We produced retrovirus or lentivirus and infected cells according to the methods previously described26,27. Infected cells were selected in culture medium plus puromycin (0.5 µg/ml) for 2 days. Cells were passaged no more than twice before being used for injection or further analysis. The retroviral vectors used in this study were MSCV-PIG (Puro IRES GFP) (used as a control vector, Addgene plasmid 18751), MSCV-MYC-T58A (Addgene plasmid 20076) and MSCV-MYC (derived from MSCV-MYC-T58A by site-directed mutagenesis (Stratagene)). The lentiviral shRNA constructs, sh-Luc, shMyc-1 (ID TRCN 42513) and shMyc-2 (ID TRCN 42517) were obtained from the RNAi consortium (Broad Institute, Cambridge, MA).

vector, Addgene plasmid 18751), MSCV-MYC-T58A (Addgene plasmid 20076) and MSCV-MYC (derived from MSCV-MYC-T58A by site-directed mutagenesis (Stratagene)). The lentiviral shRNA constructs, sh-Luc, shMyc-1 (ID TRCN 42513) and shMyc-2 (ID TRCN 42517) were obtained from the RNAi consortium (Broad Institute, Cambridge, MA). Tumor transplantation and in vivo treatment studies For tumor grafting, we injected 2–5 × 105 tumor cells into the inguinal mammary glands of recipient mice (NcrNu or NOD-SCID females, 10–12 week old, Taconic). GDC-0941 was purchased from commercial sources (Sai Advantium Pharma) and was reconstituted in 0.5% methylcellulose (Sigma) and 0.2% Tween 80 (Sigma) and administered by oral gavage (120 mg/kg/day). PF02341066 (Selleck Chemicals) was administered via oral gavage at doses of 25 or 50mg/kg/day in water. Tumor volumes were measured twice a week with calipers and calculated by the following formula: Tumor volume = (length × width2)/228. Supplementary Material 1 Note: Supplementary information is available on the Nature Medicine website. AUTHOR CONTRIBUTIONS

Tumor transplantation and in vivo treatment studies For tumor grafting, we injected 2–5 × 105 tumor cells into the inguinal mammary glands of recipient mice (NcrNu or NOD-SCID females, 10–12 week old, Taconic). GDC-0941 was purchased from commercial sources (Sai Advantium Pharma) and was reconstituted in 0.5% methylcellulose (Sigma) and 0.2% Tween 80 (Sigma) and administered by oral gavage (120 mg/kg/day). PF02341066 (Selleck Chemicals) was administered via oral gavage at doses of 25 or 50mg/kg/day in water. Tumor volumes were measured twice a week with calipers and calculated by the following formula: Tumor volume = (length × width2)/228. Supplementary Material 1 Note: Supplementary information is available on the Nature Medicine website. AUTHOR CONTRIBUTIONS P.L., H.C. and J.J.Z. designed the experiments, interpreted the data and wrote the paper. P.L. and H.C. performed most of the experiments. S.S, A.I and D.J.S. assisted with biochemical analyses and mouse work. J.Y., C.C, E.A.F., J.M. and R.S. performed genome-wide DNA copy number profiling. N.S.G. provided GDC-0941 inhibitor. M.R. and R.B. analyzed co-occurrence of PIK3CA mutation with c-MYC amplification and overexpression in human breast tumors. F.Z. and G.B.M. provided the RPPA data on the co-occurrence of PIK3CA mutation with increased c-MYC protein levels in human breast tumors. COMPETING INTERESTS STATEMENT The authors declare no competing financial interests.

P.L., H.C. and J.J.Z. designed the experiments, interpreted the data and wrote the paper. P.L. and H.C. performed most of the experiments. S.S, A.I and D.J.S. assisted with biochemical analyses and mouse work. J.Y., C.C, E.A.F., J.M. and R.S. performed genome-wide DNA copy number profiling. N.S.G. provided GDC-0941 inhibitor. M.R. and R.B. analyzed co-occurrence of PIK3CA mutation with c-MYC amplification and overexpression in human breast tumors. F.Z. and G.B.M. provided the RPPA data on the co-occurrence of PIK3CA mutation with increased c-MYC protein levels in human breast tumors. COMPETING INTERESTS STATEMENT The authors declare no competing financial interests. ACKNOWLEDGMENTS We thank T. Roberts, L. Cantley and W. Sellers for scientific discussions and suggestions. We thank L. Clayton and D. Silver for critical review of this manuscript. We thank R. Bronson for pathological analyses of tumor samples. We thank C. Li and E. Allgood for technical assistance. This work was supported by US National Institute of Health (NIH) grants CA134502 (JJZ), CA148164-01 (JJZ, NG) and K08CA122833 (RB). The Stand Up To Cancer (JJZ, GBM), a Dana Farber Harvard Cancer Center breast cancer SPORE grants P50 CA089393-08S1 (JJZ), the Department of Defense (BC051565 to JJZ.), the V Foundation (JJZ, RB) and the Claudia Barr Program (JJZ). In compliance with Harvard Medical School guidelines, we disclose that JJZ and RB are consultants for Novartis Pharmaceuticals, Inc.

rber Harvard Cancer Center breast cancer SPORE grants P50 CA089393-08S1 (JJZ), the Department of Defense (BC051565 to JJZ.), the V Foundation (JJZ, RB) and the Claudia Barr Program (JJZ). In compliance with Harvard Medical School guidelines, we disclose that JJZ and RB are consultants for Novartis Pharmaceuticals, Inc. Figure 1 Mammary gland-specific expression of PIK3CAH1047R induces mammary tumors. (a) Generation of a transgenic mouse model expressing HA (haemagglutinin)-tagged human PIK3CAH1047R under the control of a tetracycline-inducible promoter (TetO). The expression of PIK3CAH1047R is coupled through an IRES with downstream expression of luciferase. These mice were crossed with MMTV-rtTA (MTB) mice to generate bi-transgenic iPIK3CAH1047R animals to drive the expression of HAPIK3CAH1047R in mammary glands. Lower panels demonstrate bioluminescence imaging of iPIK3CAH1047R mice maintained in the presence or absence of doxycycline. (b) Tumor-free survival curve for iPIK3CAH1047R mice maintained on doxycycline (n = 81, median tumor free survival 208 days), and three groups of control mice: MTB (n = 12) and tetO-PIK3CAH1047R (n = 10) mice maintained with doxycycline, and iPIK3CAH1047R (n = 14) mice maintained in the absence of doxycycyline. All three types of control mice are represented by the blue line. (c) Representative haematoxylin and eosin (H&E)-stained sections of primary mammary tumors from iPIK3CAH1047R mice subjected to chronic doxycycline treatment. Scale bars, 25 µm. (d) Immunohistochemistry (IHC) for p-AKT(Ser473) and pS6RP(S235/236) performed on tumors isolated from iPIK3CAH1047R mice maintained on doxycycline (Dox on panels) or 6 days following doxycycline withdrawal (Dox off panels). Representative images are shown. Scale bar, 50 µm. (e) IHC for Ki67 or TUNEL performed on tumors isolated from iPIK3CAH1047R mice maintained on doxycycline (Dox on panels) or 3 days following doxycycline withdrawal (Dox off panels). Representative images are shown. Scale bar, 50 µm. n = 6 for each group. * P < 0.005 (Student’s t-test).

ages are shown. Scale bar, 50 µm. (e) IHC for Ki67 or TUNEL performed on tumors isolated from iPIK3CAH1047R mice maintained on doxycycline (Dox on panels) or 3 days following doxycycline withdrawal (Dox off panels). Representative images are shown. Scale bar, 50 µm. n = 6 for each group. * P < 0.005 (Student’s t-test). Figure 2 Tumor responses to doxycycline withdrawal. (a) Primary tumors (135 primary tumors were derived from 107 tumor bearing bi-transgenic mice; 81 mice carried one tumor, 21 mice bore two tumors and 4 mice had three tumors) had three types of response to doxycycline withdrawal. 45 of 135 (33%) regressed completely without re-growth (black bar), 4 of 135 (3%) regressed partially with no re-growth (green bar) and 86 of 135 (64%) regressed partially but then re-grew (red bar). (b) Western blot analyses of HA-p110αH1047R, p-Akt and p-S6RP in six recurrent tumors (RCT) in the absence of doxycycline and their matched primary tumors (PMT) maintained on doxycycline. Mammary gland tissues from uninduced iPIK3CAH1047R mice were used as controls. (c) Responses of recurrent tumor transplants to GDC-0941 or vehicle treatment. Data are shown as mean ± S.E.M (n = 6). * P < 0.001 (Student’s t test).

e absence of doxycycline and their matched primary tumors (PMT) maintained on doxycycline. Mammary gland tissues from uninduced iPIK3CAH1047R mice were used as controls. (c) Responses of recurrent tumor transplants to GDC-0941 or vehicle treatment. Data are shown as mean ± S.E.M (n = 6). * P < 0.001 (Student’s t test). Figure 3 Genetic alterations associated with PIK3CAH1047R-independent tumor recurrence. Mouse SNP6.0 array analyses of six recurrent tumors identified an amplification region encompassing c-Met in RCT-E565 (a), and a common focal amplification at the c-Myc locus in RCT-D419 and RCT-C658 tumors (b). (c) Western blot analyses of p-Akt (Ser473) and pS6RP(S235/236) in two RCT-E565 xenograft tumors treated with vehicle or PF02341066. Samples were isolated 4 hours after the last dose from mice treated with PF02341066 for 3 days. (d) Responses of RCT-E565 xenograft tumors in NcrNu mice to PF02341066 or vehicle. Data are shown as mean ± S.E.M (each group, n = 6). *P < 0.005, ** P < 0.001 (Student’s t-test).

t tumors treated with vehicle or PF02341066. Samples were isolated 4 hours after the last dose from mice treated with PF02341066 for 3 days. (d) Responses of RCT-E565 xenograft tumors in NcrNu mice to PF02341066 or vehicle. Data are shown as mean ± S.E.M (each group, n = 6). *P < 0.005, ** P < 0.001 (Student’s t-test). Figure 4 Elevation of c-Myc drives mammary tumors to become independent of PIK3CAH1047R and resistant to PI3K inhibition. (a) shRNA knockdown of c-Myc in primary tumor cells isolated from RCT-D419. Western blot analysis of c-Myc in RCT-D419 parental cells or cells infected with the indicated lentiviral shRNAs. Vinculin was used as a loading control. (b) RCT-D419 cells expressing sh-Luc, sh-Myc1 or sh-Myc2 were transplanted into NOD-SCID mice and tumor formation monitored. Downregulation of c-Myc suppressed tumor formation. * P < 0.001 (log-rank test) (c) Western blot analysis of ectopically expressed c-Myc or c-MycT58A in D777 tumor cells isolated from a PIK3CAH1047R-dependent primary tumor that had been maintained on doxycycline. (d) Bioluminescence imaging showing tumor establishment in NOD-SCID mice transplanted with D777 cells expressing vector, c-MycT58A or c-Myc. These mice were maintained on doxycycline to sustain PIK3CAH1047R expression. (e) Tumors established by D777 cells expressing a control vector in the presence of doxycycline regressed upon doxycycline withdrawal. Tumors established by D777 cells expressing c-Myc or c-MycT58A continued to grow in the absence of doxycycline. Data are shown as mean ± S.E.M (n = 6). (f) Mice bearing D777-MycT58A tumors were treated with either GDC-0941 or vehicle and tumor growth followed. Data are shown as mean ± S.E.M (n = 6). (g) A schematic diagram summarizing three outcomes of PIK3CAH1047R-initiated tumors following inactivation of PIK3CAH1047R expression; tumors either regress (gray box) or recur (red box) via a PI3K pathway-dependent or -independent mechanism. As illustrated, these tumor outcomes affect tumor responses to drug treatment.

MicroRNAs repress gene expression by inhibiting mRNA translation or by promoting mRNA degradation and are considered master regulators of different processes, ranging from proliferation1 to apoptosis2. Altered miRNA expression in various human tumor types has been observed, and critical roles of miRNAs in cancer pathogenesis and response to therapy have been demonstrated 3,4. Non-small cell lung cancers (NSCLC) account for roughly 85% of all lung cancer cases5. Although NSCLC is a remarkably heterogeneous disease that includes distinct morphological and molecular subtypes, activation of epidermal growth factor receptor (EGFR) and MET (the receptor tyrosine kinase (RTK) for hepatocyte growth factors) is common and associated with RAS/ERK and PI3K/AKT axes stimulation, leading to NSCLC cell proliferation, survival and invasion6. Tyrosine kinase inhibitors (TKI) gefitinib and erlotinib effectively target EGFR in NSCLC patients, but these important therapeutic agents are ultimately limited by the emergence of drug resistance mutations and other putative molecular mechanisms.7

mulation, leading to NSCLC cell proliferation, survival and invasion6. Tyrosine kinase inhibitors (TKI) gefitinib and erlotinib effectively target EGFR in NSCLC patients, but these important therapeutic agents are ultimately limited by the emergence of drug resistance mutations and other putative molecular mechanisms.7 MET protein expression and phosphorylation have been associated with primary and acquired resistance to EGFR TKI therapy in NSCLC patients 8,9, strongly implicating MET as an effective therapeutic target to overcome resistance to this important class of drugs in lung cancer10. Here we show that EGF and MET receptors, by modulating specific microRNAs, control gefitinib-induced apoptosis and NSCLC tumorigenesis. Our results are the first to identify EGF and MET receptor-regulated microRNAs representing oncogenic signaling networks in NSCLC.

ce to this important class of drugs in lung cancer10. Here we show that EGF and MET receptors, by modulating specific microRNAs, control gefitinib-induced apoptosis and NSCLC tumorigenesis. Our results are the first to identify EGF and MET receptor-regulated microRNAs representing oncogenic signaling networks in NSCLC. Results MicroRNAs modulated by both EGFR and MET To identify EGFR- and MET-regulated-miRNAs, we stably silenced EGFR and MET in Calu-1 cells using shRNA lentiviral particles (Fig. 1a) and examined global microRNA expression profiles. In EGFR- and MET-knockdown (EGFR-KD and MET-KD) Calu-1 cells, 35 and 44 significantly dysregulated microRNAs were identified, respectively (Figs. 1b and Supplementary Fig. 1a). MicroRNAs with >1.5- (EGFR) and with >1.7- (MET) fold-change are shown. By comparing the two lists, only 8 microRNAs were found to be regulated by both EGFR and MET (Fig. 1c): miR-21 (fold changeEGFR-KD= -1.56; fold changeMET-KD= -1.7), -221/222 (f.c.EGFR-KD= -1.79/-1.66; f.c.MET-KD= -2.07/-1.75), -30b/c (f.c.EGFR-KD= -1.81/-2.4; f.c.MET-KD= -3.5/-4.0), -29a/c (f.c.EGFR-KD= -1.52/-1.53; f.c.MET-KD= -1.72/-1.79), and -100 (f.c.EGFR-KD= -1.55; f.c.MET-KD= -1.72). We initially focused on miR-30b/c and 221/222, downregulated after both MET and EGFR silencing and showing the highest expression level fold-change. We also investigated two microRNAs most differentially induced after MET silencing, miR-103 (f.c.= 2.45) and miR-203 (f.c.= 2.5), based on recent evidence for MET overexpression in de novo and acquired resistance to TKIs8,9. Expression levels of these six miRNAs in EGFR-KD and MET-KD Calu-1 cells were validated using qRT-PCR (Supplementary Fig. 1b) and northern blot (Fig. 1d) analysis.

MET silencing, miR-103 (f.c.= 2.45) and miR-203 (f.c.= 2.5), based on recent evidence for MET overexpression in de novo and acquired resistance to TKIs8,9. Expression levels of these six miRNAs in EGFR-KD and MET-KD Calu-1 cells were validated using qRT-PCR (Supplementary Fig. 1b) and northern blot (Fig. 1d) analysis. MiR103, 203, 30b/c and 221/222 target PKCε, SRC, BIM and APAF1 As MET and EGFR RTKs play a significant role in lung cancer tumorigenesis and progression11, we hypothesized that miR-103 and miR-203 (increased after MET knockdown) are tumor suppressors and miR-221/222 and miR-30b/c (decreased after MET and EGFR silencing) are oncogenic. Our search for mRNA targets of miR-103, -203, -221/222, and -30b/c using Targetscan and Pictar computational tools revealed that 3′UTRs of human Apaf-1, BIM, Pkc-ε and Src genes contain evolutionarily conserved binding sites specific for these miRNAs (Supplementary Fig. 2a). We focused on these genes based on their role in TKI sensitivity (BIM and APAF-1)12,13 and resistance (SRC)14 or negative allosteric modulation of EGFR signalling (PKC-ε)15. To investigate whether these miRNAs directly interact with the four putative target genes, we cotransfected pGL3-3′UTR vectors with synthetic miR-103, -203, 221/222 and -30b/c. Decreased luciferase activity indicated direct miR-PKC-ε, SRC, APAF-1 and BIM 3′ UTR interactions (Fig. 1e), and target gene repression was rescued by mutations or deletions in the complementary seed sites (Figs. 1e and Supplementary Fig. 2a). Western blot analysis showed an inverse correlation (p<0.05) between miR-221/222, -103, -203, 30b/c expression and target protein amount in a NSCLC cell panel (Supplementary Fig. 2b,c) confirmed by Pearson coefficient (Figs. 1f and Supplementary Fig. 2d). The immunoblot analysis fully agreed with data obtained using reporter gene assays (Supplementary Results, Figs. 1g–j and Supplementary Fig. 3a,b). Detection of PKC-ε, SRC, APAF-1, BIM proteins in vivo in 110 lung cancer specimens using miRNA in situ hybridization (ISH) followed by immunohistochemistry (IHC) further strengthened the negative correlation between these proteins and miR-103, -203, 221/222 and -30b/c seen in vitro (Supplementary Table 1). An inverse correlation between miR-203/SRC, miR-30c/BIM, miR-103/PKC-ε, and miR-222/APAF-1 expression was observed in the majority of the lung cancer tissues (Supplementary Fig. 4a,b).

rengthened the negative correlation between these proteins and miR-103, -203, 221/222 and -30b/c seen in vitro (Supplementary Table 1). An inverse correlation between miR-203/SRC, miR-30c/BIM, miR-103/PKC-ε, and miR-222/APAF-1 expression was observed in the majority of the lung cancer tissues (Supplementary Fig. 4a,b). Moreover, MET overexpression was seen in 52% (57/110) of the same 110 lung tumor samples (demonstrated using miRNA ISH/MET IHC; Supplementary Fig. 5a), low miR-103 and miR-203 expression and high miR-222 and miR-30c expression was observed in MET-overexpressing tumors (Fig. 2a; Supplementary Fig. 5a). Interestingly, the majority of MET overexpressing tumors had accompanying metastases (Supplementary Fig. 5b), indicating that MET-regulated microRNAs play a role in metastatic spread of lung cancer cells. We extended our analysis to 40 independent lung tumors with annotated clinical history (Supplementary Table 2), divided into two groups of “low” and “high” MET and EGFR expression (based on qRT-PCR; Supplementary Fig. 5c). ANOVA confirmed that miR-30b-c and miR-221-222 were differentially expressed between the two groups, whereas Pearson coefficient identified an inverse correlation MET-miR-103, -203 (Fig. 2b,c). The qRT-PCR results were corroborated using IHC analysis for MET and EGFR (Supplementary Fig. 5d). In addition, MET overexpression was observed in tumors that presented distant metastases versus non-metastatic cases and there was no correlation between metastases and EGFR expression in these 40 lung cancers (Fig. 2d and Supplementary Fig. 5e).

corroborated using IHC analysis for MET and EGFR (Supplementary Fig. 5d). In addition, MET overexpression was observed in tumors that presented distant metastases versus non-metastatic cases and there was no correlation between metastases and EGFR expression in these 40 lung cancers (Fig. 2d and Supplementary Fig. 5e). MiR-30b-c, 221-222, 103, 203 regulate gefitinib sensitivity Having demonstrated that EGFR regulates miR-221-222 and miR-30b-c, we investigated a role for these miRs in gefitinib-induced apoptosis in NSCLC with wild-type EGFR (Calu-1 and A549 cells) versus EGFR exon 19 deletions (PC9 and HCC827). Calu-1 and A549 cells were completely resistant to all concentrations of gefitinib tested (up to 20 μM); in contrast, PC9 and HCC827 EGFR-mutant cells were significantly growth-inhibited even by low (0.1 μM) dose gefitinib (Fig. 3a), in agreement with a previous study16. Interestingly, after gefitinib treatment, marked miR-30b-c and -221-222 downregulation and increased BIM and APAF-1 protein levels were observed, but only in PC9 and HCC827 sensitive cells (Fig. 3b,c). Moreover, as previously anticipated16, the level of phosphorylated ERKs was markedly decreased in Hcc827 and PC9 but not in Calu-1 cells versus untreated cells (Fig. 3c).

-30b-c and -221-222 downregulation and increased BIM and APAF-1 protein levels were observed, but only in PC9 and HCC827 sensitive cells (Fig. 3b,c). Moreover, as previously anticipated16, the level of phosphorylated ERKs was markedly decreased in Hcc827 and PC9 but not in Calu-1 cells versus untreated cells (Fig. 3c). To directly assess the relevance of miR-30b-c and -221-222 in gefitinib-induced apoptosis, we analyzed expression levels of these miRNAs in PC9 GR and HCC827 GR, NSCLC cells with acquired gefitinib resistance, obtained after long-term exposure to increasing drug concentrations8,17. In contrast to the gefitinib-responsive parental cells, decreased miR-30b-c and -221/222 expression and modulation of their relative targets were not observed (Fig. 3d; Supplementary Fig. 6a). Also of note, miR-30c and miR-221/222 overexpression in gefitinib-sensitive HCC827 and PC9 rendered these cells less drug responsive (Figs. 4a and Supplementary Fig. 7a), and knockdown of miR-30b-c and -221-222 increased gefitinib sensitivity of Calu-1, HCC827GR and PC9GR (Figs. 4a and Supplementary Fig. 7b), indicating that these miRs are important modulators of TKI resistance.

inib-sensitive HCC827 and PC9 rendered these cells less drug responsive (Figs. 4a and Supplementary Fig. 7a), and knockdown of miR-30b-c and -221-222 increased gefitinib sensitivity of Calu-1, HCC827GR and PC9GR (Figs. 4a and Supplementary Fig. 7b), indicating that these miRs are important modulators of TKI resistance. To investigate the contribution of miR-30b-c and 221-222-mediated APAF-1 and BIM downregulation to cellular TKI response, we overexpressed APAF-1 and BIM in A549 resistant cells. Gefitinib-induced PARP cleavage was observed in cells overexpressing BIM and APAF-1 but not in cells transfected with an empty vector plasmid (Fig. 4b). Conversely, response to gefitinib was reduced by BIM and APAF-1 silencing in gefitinib-sensitive HCC827 and PC9 cells (Fig. 4c). We cloned wild type and mutated 3′UTRs of BIM and APAF-1 (used for luciferase assays, Fig. 1e) downstream of BIM and APAF-1 coding sequences and performed caspase 3/7 and viability assays. No increase in cell death was observed after gefitinib treatment of A549 cells cotransfected with wild type 3′ UTRs for BIM and APAF-1 and miR-30b-c and -221-222; conversely, miR binding site mutation or deletion restored the apoptotic response to gefitinib, suggesting that the effect of APAF-1 and BIM on gefitinib sensitivity was directly related to miR-30b-c and -221-222-mediated knockdown of these proteins (Figs. 4d and Supplementary Fig. 7c).

AF-1 and miR-30b-c and -221-222; conversely, miR binding site mutation or deletion restored the apoptotic response to gefitinib, suggesting that the effect of APAF-1 and BIM on gefitinib sensitivity was directly related to miR-30b-c and -221-222-mediated knockdown of these proteins (Figs. 4d and Supplementary Fig. 7c). Because miR-30b-c and -221-222 are also regulated by MET, we hypothesized that MET inhibition, by modulating these miRs in common with the EGFR pathway, could overcome gefitinib-resistance in NSCLC. In support of this hypothesis, strong downregulation of miR-30b-c and -221-222 was observed in Calu-1- and A549-MET overexpressing cells (data not shown) treated with SU11274 MET inhibitor or MET knockdown (Supplementary Fig. 8a,b). Moreover, increased caspase 3/7 activity and decreased cell viability were observed in SU11274-treated Calu-1 or Calu-1-MET-KD cells exposed to different gefitinib concentrations. (Supplementary Fig. 8c,d). Taken together, these results suggest that MET inhibition restores gefitinib sensitivity in TKI-resistant Calu-1 cells through the downregulation of miR-30b-c and miR-221-222. Other miRNAs commonly modulated by EGFR-MET, including miR-21, miR-29a/c and miR-100 (Fig. 1c) were downregulated in HCC827- and PC9-gefitinib treated cells and found involved in the response to the drug (Supplementary Results, Supplementary Figs. 9, 10 and 11).

Because miR-30b-c and -221-222 are also regulated by MET, we hypothesized that MET inhibition, by modulating these miRs in common with the EGFR pathway, could overcome gefitinib-resistance in NSCLC. In support of this hypothesis, strong downregulation of miR-30b-c and -221-222 was observed in Calu-1- and A549-MET overexpressing cells (data not shown) treated with SU11274 MET inhibitor or MET knockdown (Supplementary Fig. 8a,b). Moreover, increased caspase 3/7 activity and decreased cell viability were observed in SU11274-treated Calu-1 or Calu-1-MET-KD cells exposed to different gefitinib concentrations. (Supplementary Fig. 8c,d). Taken together, these results suggest that MET inhibition restores gefitinib sensitivity in TKI-resistant Calu-1 cells through the downregulation of miR-30b-c and miR-221-222. Other miRNAs commonly modulated by EGFR-MET, including miR-21, miR-29a/c and miR-100 (Fig. 1c) were downregulated in HCC827- and PC9-gefitinib treated cells and found involved in the response to the drug (Supplementary Results, Supplementary Figs. 9, 10 and 11). Next, we focused on miR-103 and -203, strongly upregulated after MET knockdown in Calu-1 cells (Fig. 1d). As expected, treatment of Calu-1 cells with SU11274 increased (P<0.05) miR-103 and -203 expression (Supplementary Fig. 12a). SRC and PKC-ε exert pro-survival effects and contribute to gefitinib resistance by activating AKT and ERK signaling pathways14,15. Accordingly, miR-103 and -203-overexpression in A549 cells was associated with reduced phosphorylation of AKT and its substrate GSK3β and ERKs (Supplementary Fig. 13a). Engelman et al. previously reported that MET induces gefitinib resistance through persistent PI3K/AKT and ERK signaling activation8, and our results indicate that MET overexpression controls gefitinib resistance through activation of the AKT/ERKs pathway, mediated at least in part by miR-103 and -203. In agreement with Engelman et al., enforced expression of miR-103, -203 or Pkc-ε and Src silencing increased Calu-1 cell sensitivity to gefitinib (assessed by caspase 3/7 and viability assays, Supplementary Fig. 13b,c). Importantly, we found that miR-103 and -203 decreased and SRC and PKC-ε expression consequently increased in HCC827 GR cells with acquired MET amplification and gefitinib resistance8, corroborating our hypothesis that MET controls the response to TKIs, at least in part through miR-103 and -203 and their respective targets (Supplementary Fig. 13d,e). Finally, to analyze the sensitivity to gefitinib in vivo, we stably transfected A549 cells with GFP-lentivirus constructs containing either full-length miR-103, miR-203 or anti-miR-221 and -30c. MiR-103 and miR-203 overexpression or miR-221 and -30c knockdown resulted in dramatic tumor growth inhibition and increased sensitivity to gefitinib-induced apoptosis in nude mice after two weeks of treatment (Figs. 4e,f and Supplementary Fig. 14a). MiR-221-222 down-regulation and miR-103, -203 upregulation in the xenograft tumors were confirmed by qRT-PCR (Supplementary Fig. 14b).

resulted in dramatic tumor growth inhibition and increased sensitivity to gefitinib-induced apoptosis in nude mice after two weeks of treatment (Figs. 4e,f and Supplementary Fig. 14a). MiR-221-222 down-regulation and miR-103, -203 upregulation in the xenograft tumors were confirmed by qRT-PCR (Supplementary Fig. 14b). MiR-103 and -203 reduce NSCLC cell migration and proliferation The anti-proliferative activity of miR-103 and -203 in different cancers has been previously reported18,19,20. To further investigate the functional role of miR-103 and miR-203 in NSCLC tumorigenesis, we assessed the impact of miR-103 and -203 gain of function or loss of PKC-ε and SRC on cell migration and cell cycle kinetics. Migration was reduced (by 2-fold compared to controls) in cells with increased miR-103 and -203 or decreased PKC-ε and SRC expression (Fig. 5a). These results were further confirmed by wound-healing assay (Fig. 5b). In addition, A549 and Calu-1 cells transfected with miR-103, -203 or Pkc-ε and Src siRNAs showed increased G1 cell fraction and a corresponding decreased number of cells in S and G2-M phases, with miR-203 and siSRC having a slightly stronger effect compared to miR-103 and siPKC-ε (Fig. 5c).

und-healing assay (Fig. 5b). In addition, A549 and Calu-1 cells transfected with miR-103, -203 or Pkc-ε and Src siRNAs showed increased G1 cell fraction and a corresponding decreased number of cells in S and G2-M phases, with miR-203 and siSRC having a slightly stronger effect compared to miR-103 and siPKC-ε (Fig. 5c). MiR-103, 203 promote mesenchymal-to-epithelial transition An association between EMT and development of chemoresistance, including resistance to EGFR targeted therapy, leading to recurrence and metastasis has recently been reported21. Although identifying the molecular events underlying EMT is an intense area of investigation, what triggers the onset of EMT in tumor cells remains a mystery. We observed a striking change in cellular shape in MET-KD Calu-1 cells from spindle-, fibroblastoid morphology to an epithelial polarized phenotype upon MET-KO (Fig. 6a), prompting us to further investigate whether this morphological change could be due to a mesenchymal-to-epithelial transition. We assessed the expression of key EMT transition-associated markers and observed decreased expression of mesenchymal markers and increased E-cadherin expression (Fig. 6b,c.d), strongly indicating reversion of Calu-1 cells to an epithelial phenotype after MET knockdown. Importantly, in MET-KD cells, Snail protein expression was lower, localized to the cytoplasm (Fig. 6b) and presumably nonfunctional22. It is worth noting that a morphological change in EGFR-KD Calu-1 cells was not observed, and miR-200c23,24,-103 and -203 were not upregulated (Supplementary Fig. 1a). To test if miR-103 and -203 were involved in mesenchymal-epithelial transition, we overexpressed these miRs in Calu-1 cells and observed downregulation of several mesenchymal markers and increased E-cadherin expression, indicating a role for these miRs in the mesenchymal-epithelial transition (Fig. 6e–g; Supplementary Fig. 15a,b). Moreover, PKC-ε and SRC silencing increased E-cadherin and decreased SNAIL, ZEB1, ZEB2 (zinc finger E-box binding 2), vimentin and fibronectin mRNA levels (Supplementary Fig. 15c). Since it has been recently reported that miR-103 targets Dicer25, we investigated the effects of Dicer knockdown on tumorigenesis and gefitinib-induced apoptosis of NSCLC.

g increased E-cadherin and decreased SNAIL, ZEB1, ZEB2 (zinc finger E-box binding 2), vimentin and fibronectin mRNA levels (Supplementary Fig. 15c). Since it has been recently reported that miR-103 targets Dicer25, we investigated the effects of Dicer knockdown on tumorigenesis and gefitinib-induced apoptosis of NSCLC. Remarkably, almost complete Dicer knockdown reduced not only gefitinib resistance but also migration and the expression of mesenchymal markers of NSCLC cells, suggesting that miR-103 could be involved in the mesenchymal-epithelial transition process also through Dicer downregulation (Supplementary Results and Supplementary Fig. 16). In conclusion, by regulating expression of specific miRs, MET orchestrates the convergence of several EMT-associated pathways, including Dicer, SRC, PKC-ε and AKT pathways, supporting the possibility that MET targeting could provide a realistic strategy to control EMT and NSCLC progression. Discussion Although initially promising, it is now clear that EGFR inhibitors, including gefitinib and erlotinib, will not cure the majority of NSCLC patients, even in those cancers expressing mutant EGFR. Therefore, a more complete understanding of how these drugs work and the downstream molecular mediators of EGFR will provide critical information to help design strategies to augment their efficacy.