1. Computational and Systems Biology
  2. Genetics and Genomics

Risk factors affecting polygenic score performance across diverse cohorts

  1. Daniel Hui
  2. Scott Dudek
  3. Krzysztof Kiryluk
  4. Theresa L Walunas
  5. Iftikhar J Kullo
  6. Wei-Qi Wei
  7. Hemant Tiwari
  8. Josh F Peterson
  9. Wendy K Chung
  10. Brittney H Davis
  11. Atlas Khan
  12. Leah C Kottyan
  13. Nita A Limdi
  14. Qiping Feng
  15. Megan J Puckelwartz
  16. Chunhua Weng
  17. Johanna L Smith
  18. Elizabeth W Karlson
  19. Regeneron Genetics Center
  20. Penn Medicine BioBank
  21. Gail P Jarvik
  22. Marylyn D Ritchie  Is a corresponding author
  1. Department of Genetics, Perelman School of Medicine, University of Pennsylvania, United States
  2. Division of Nephrology, Department of Medicine, Columbia University, United States
  3. Department of Preventive Medicine, Northwestern University Feinberg School of Medicine, United States
  4. Department of Cardiovascular Medicine, Mayo Clinic, United States
  5. Department of Biomedical Informatics, Vanderbilt University Medical Center, United States
  6. Department of Pediatrics, University of Alabama at Birmingham, United States
  7. Departments of Pediatrics and Medicine, Columbia University Irving Medical Center, Columbia University, United States
  8. Department of Neurology, School of Medicine, University of Alabama at Birmingham, United States
  9. The Center for Autoimmune Genomics and Etiology, Division of Human Genetics, Cincinnati Children's Hospital Medical Center, United States
  10. Division of Clinical Pharmacology, Department of Medicine, Vanderbilt University Medical Center, United States
  11. Center for Genetic Medicine, Northwestern University Feinberg School of Medicine, United States
  12. Department of Biomedical Informatics, Vagelos College of Physicians & Surgeons, Columbia University, United States
  13. Division of Rheumatology, Inflammation, and Immunity, Department of Medicine, Brigham and Women's Hospital and Harvard Medical School, United States
  14. Departments of Medicine (Medical Genetics) and Genome Sciences, University of Washington Medical Center, United States
https://doi.org/10.7554/eLife.88149.3
Share this article
Cite this article
  1. Daniel Hui
  2. Scott Dudek
  3. Krzysztof Kiryluk
  4. Theresa L Walunas
  5. Iftikhar J Kullo
  6. Wei-Qi Wei
  7. Hemant Tiwari
  8. Josh F Peterson
  9. Wendy K Chung
  10. Brittney H Davis
  11. Atlas Khan
  12. Leah C Kottyan
  13. Nita A Limdi
  14. Qiping Feng
  15. Megan J Puckelwartz
  16. Chunhua Weng
  17. Johanna L Smith
  18. Elizabeth W Karlson
  19. Regeneron Genetics Center
  20. Penn Medicine BioBank
  21. Gail P Jarvik
  22. Marylyn D Ritchie
(2025)
Risk factors affecting polygenic score performance across diverse cohorts
eLife 12:RP88149.
https://doi.org/10.7554/eLife.88149.3
  2. Download BibTeX
  3. Download .RIS

eLife assessment

This study presents a convincing analysis of the effects of covariates, such as age, sex, socioeconomic status, or biomarker levels, on the predictive accuracy of polygenic scores for body mass index; the work is further supported by important approaches for improving prediction accuracy by accounting for such covariates across a variety of association studies. The authors did a commendable job addressing reviewer suggestions and comments. The work will be of interest to colleagues using and developing methods for phenotypic prediction based on polygenic scores.

https://doi.org/10.7554/eLife.88149.3.sa0

Abstract

Apart from ancestry, personal or environmental covariates may contribute to differences in polygenic score (PGS) performance. We analyzed the effects of covariate stratification and interaction on body mass index (BMI) PGS (PGSBMI) across four cohorts of European (N = 491,111) and African (N = 21,612) ancestry. Stratifying on binary covariates and quintiles for continuous covariates, 18/62 covariates had significant and replicable R2 differences among strata. Covariates with the largest differences included age, sex, blood lipids, physical activity, and alcohol consumption, with R2 being nearly double between best- and worst-performing quintiles for certain covariates. Twenty-eight covariates had significant PGSBMI–covariate interaction effects, modifying PGSBMI effects by nearly 20% per standard deviation change. We observed overlap between covariates that had significant R2 differences among strata and interaction effects – across all covariates, their main effects on BMI were correlated with their maximum R2 differences and interaction effects (0.56 and 0.58, respectively), suggesting high-PGSBMI individuals have highest R2 and increase in PGS effect. Using quantile regression, we show the effect of PGSBMI increases as BMI itself increases, and that these differences in effects are directly related to differences in R2 when stratifying by different covariates. Given significant and replicable evidence for context-specific PGSBMI performance and effects, we investigated ways to increase model performance taking into account nonlinear effects. Machine learning models (neural networks) increased relative model R2 (mean 23%) across datasets. Finally, creating PGSBMI directly from GxAge genome-wide association studies effects increased relative R2 by 7.8%. These results demonstrate that certain covariates, especially those most associated with BMI, significantly affect both PGSBMI performance and effects across diverse cohorts and ancestries, and we provide avenues to improve model performance that consider these effects.

Introduction

Polygenic scores (PGS) provide individualized genetic predictors of a phenotype by aggregating genetic effects across hundreds or thousands of loci, typically estimated from genome-wide association studies (GWAS). In recent years, it has become increasingly apparent that the transferability of PGS performance across different cohorts is poor (Martin et al., 2019). Most analyses to date have focused on ancestry differences as the main driver of this lack of portability (Wang et al., 2020; Galinsky et al., 2019; Shi et al., 2021). However, a growing body of evidence has demonstrated that PGS performance and effect estimates are influenced by differences in certain contexts, that is, environmental (classically termed ‘gene–environment’ effects or interactions) or personal-level covariates – different phenotypes seem to be differently affected by these covariates, with adiposity traits such as body mass index (BMI) having substantial evidence for these effects (Rask-Andersen et al., 2017; Robinson et al., 2017; Sulc et al., 2020; Justice et al., 2017; Helgeland et al., 2019; Vogelezang et al., 2020; Couto Alves et al., 2019; Choh et al., 2014; Mostafavi et al., 2020; Elks et al., 2012). In one previous study, they showed that GWAS stratified by sample characteristics had better PGS performance in cohorts that matched the sample characteristics of the stratified GWAS, and that differences in heritability between the stratified cohorts partially explained this observation (Mostafavi et al., 2020).

There are several gaps in current knowledge about these covariate-specific effects. Many analyses have assessed only a handful of these covariates due to the myriad of choices possible in typical large-scale biobanks. Little investigation has been done to systematically understand why certain covariates affect PGS performance, with such knowledge being useful to reduce the potential search for variables that impart context-specific effects. Furthermore, most studies investigating PGS–covariate interactions have been in European ancestry individuals; notably, comparing differences in PGS performance and prediction while controlling for differences in ancestry versus differences in context has not been assessed in previous studies. Moreover, covariate-specific effects are notorious for replicating poorly in human genetics studies, and previous studies of PGS–covariate interactions have been predominantly performed in the UK Biobank (UKBB) (Bycroft et al., 2018), where the majority of individuals are aged 40–69 (i.e., excluding young adults), are overall healthier than those from other, for example, hospital-based cohorts, and are predominantly European ancestry. Additionally, PGS performance is often assessed using linear models and in isolation of clinical covariates, which in practice would often be available. Machine learning models can have increased performance over linear models and are capable of modeling complex relationships and interactions between variables, which may serve to increase predictive performance, especially given evidence for PGS–covariate-specific effects. Finally, given evidence for context-specific effects, it should be possible to directly incorporate SNP–covariate interaction effects from a GWAS directly to improve prediction performance, instead of relying on post hoc interactions from a typical PGS calculated from main GWAS effects.

Using genetic data with linked-phenotypic information from electronic health records, we estimated the effects of covariate stratification and interaction on performance and effect estimates of PGS for BMI (PGSBMI) – a flowchart summarizing our analyses is presented in Figure 1. These analyses were done across four datasets (Supplementary file 1a): UKBB, Penn Medicine BioBank (PMBB) (Verma et al., 2022), Electronic Medical Records and Genomics (eMERGE) network dataset (Stanaway et al., 2019), and Genetic Epidemiology Research on Adult Health and Aging (GERA) (Banda et al., 2015). These datasets include participants from two ancestry groups (N = 491,111 European ancestry [EUR], N = 21,612 African ancestry [AFR]), and 62 covariates (25 present in multiple datasets) representing laboratory, survey, and biometric data types typically associated with cardiometabolic health and adiposity. After constructing PGSBMI using out-of-sample multi-ancestry BMI GWAS, we assessed the effects of covariate stratification on PGSBMI R2, the significance of PGSBMI–covariate interaction terms and their increases to model R2 over models only using main effects, as well as correlation of main effect, interaction effect, and R2 differences. We then assessed ways to increase model performance through using machine learning models, and creating PGSBMI using GxAge GWAS effects. This study addresses a plethora of open issues considering performance and effects of PGS on individuals from diverse backgrounds.

Figure 1
Download asset Open asset
A flowchart of the project.

Results

Effect of covariate stratification on PGSBMI performance

We assessed 62 covariates for PGSBMI R2 differences (25 present, or suitable proxies, in multiple datasets Supplementary file 1b) after stratifying on binary covariates and quintiles for continuous covariates. With UKBB EUR as discovery (N = 376,729), 18 covariates had significant differences (Bonferroni p<0.05/62) in R2 among groups (Figure 2a), including age, sex, alcohol consumption, different physical activity measurements, Townsend deprivation index, different dietary measurements, lipids, blood pressure, and HbA1c, with 40 covariates having suggestive (p<0.05) evidence of R2 differences. From an original PGSBMI R2 of 0.076, R2 increased to 0.094–0.088 for those in the bottom physical activity, alcohol intake, and high-density lipoprotein (HDL) cholesterol quintiles, and decreased to 0.067–0.049 for those in the top quintile, respectively, comparable to differences observed between ancestries (Martin et al., 2019). We note that the differences in R2 due to alcohol intake and HDL were larger than those of any physical activity phenotype, despite physical activity having one of the oldest and most replicable evidence of interaction with genetic effects of BMI (Kilpeläinen et al., 2011; Rampersaud et al., 2008). Despite considerable published evidence suggesting covariate-specific genetic effects between BMI and smoking behaviors (Robinson et al., 2017; Justice et al., 2017), we were only able to find suggestive evidence for R2 differences when stratifying individuals across several smoking phenotypes (minimum p=0.016, for smoking pack years). R2 differences due to educational attainment were also only suggestive (p=0.015), with published evidence on this association being conflicting (Amin et al., 2018; Li et al., 2021; Frank et al., 2019).

Figure 2
Download asset Open asset
Polygenic score (PGS) R2 stratified by quintiles for quantitative variables and by binary variables.

(a) Continuous covariates with significant (p<8.1 × 10–4) R2 differences across quintiles in UK Biobank (UKBB) European ancestry (EUR). Pork and processed meat consumption per week were excluded from this plot in favor of pork and processed meat intake. (b) Covariates with significant differences that were available in multiple cohorts. When traits had the same or directly comparable units between cohorts we show the actual trait values (and show percentiles for physical activity, alcohol intake frequency, and socioeconomic status, which had slightly differing phenotype definitions across cohorts) plotted on x-axis. Townsend index and income were used as variables for socioeconomic status in UKBB and Genetic Epidemiology Research on Adult Health and Aging (GERA), respectively. Note that the sign for Townsend index was reversed, since increasing Townsend index is lower socioeconomic status, while increasing income is higher socioeconomic status. PA, physical activity (PA); IPAQ, International Physical Activity Questionnaire.

We replicated these analyses in three additional large-scale cohorts of European and African ancestry individuals (Figure 2b, Supplementary file 1c), as well as in African ancestry UKBB individuals. Among covariates with significant performance differences in the discovery analysis, we were able to replicate significant (p<0.05) R2 differences for age, HDL cholesterol, alcohol intake frequency, physical activity, and HbA1c, despite much smaller sample sizes. We again observed mostly insignificant differences across cohorts and ancestries when stratifying due to different smoking phenotypes and educational attainment. For each covariate and ancestry combination, we combined data across cohorts and conducted a linear regression weighted by sample size, regressing R2 values on covariate values across groupings. Slopes of the regressions across cohorts had different signs between ancestries for the same covariate (triglyceride levels, HbA1c, diastolic blood pressure, and sex), although larger sample sizes may be needed to confirm these differences are statistically significant.

Several observations related to age-specific effects on PGSBMI we considered noteworthy. First, in the weighted linear regression of all R2 values across ancestries, expected R2 for African ancestry individuals can become greater than that of European ancestry individuals among individuals within bottom and top age quintiles observed in these data. For instance, the predicted R2 of 0.048 for 80-year-old European ancestry individuals would be lower than that of African ancestry individuals aged 24.7 and lower, indicating that differences in covariates can affect PGSBMI performance more than differences due to ancestry. Second, we obtained these results despite the average age of GWAS individuals being 57.8, which should increase PGSBMI R2 for individuals closest to this age (Mostafavi et al., 2020). This result suggests that PGS performance due to decreased heritability with age cannot be fully reconciled using GWAS from individuals of similar age being used to create PGSBMI (as heritability is an upper bound on PGS performance). Finally, we observed that PGSBMI R2 increases as age decreases, consistent with published evidence suggesting that the heritability of BMI decreases with age (Ge et al., 2017; Min et al., 2013).

PGS–covariate interaction effects

Next, we estimated the differences in PGS effects due to interactions with covariates by modeling interaction terms between PGSBMI and the covariate for each covariate in our list (described in ‘Materials and methods’). We implemented a correction for shared heritability between covariates of interest and outcome (which can inflate test statistics Aschard et al., 2015) to better measure the environmental component of each covariate, and show that this correction successfully reduces significance of interaction estimates (Figure 3—figure supplement 1). Again, using UKBB EUR as the discovery cohort, we observed 28 covariates with significant (Bonferroni p<0.05/62) PGS–covariate interactions (Table 1), with 38 having suggestive (p<0.05) evidence (Supplementary file 1d). We observed the largest effect of PGS–covariate interaction with alcohol drinking frequency (20.0% decrease in PGS effect per 1 standard deviation [SD] increase, p=2.62 × 10–55), with large effects for different physical activity measures (9.4–12.5% decrease/SD, minimum p=3.11 × 10–66), HDL cholesterol (15.3% decrease/SD, p=1.71 × 10–96), and total cholesterol (12.7% decrease/SD, p=1.64 × 10–71). We observed significant interactions with diastolic blood pressure (10.8% increase/SD, p=6.06 × 10–60), but interactions with systolic blood pressure were much smaller (1.17% increase/SD, p=4.41 × 10–3). Significant interactions with HbA1c (4.63% increase/SD, p=5.37 × 10–14) and type 2 diabetes (27.2% PGS effect increase in cases, p=1.83 × 10–7) were also observed. Other significant PGS–covariate interactions included lung function, age, sex, and LDL cholesterol – various dietary measurements also had significant interactions, albeit with smaller effects than other significant covariates. We were able to find significant interaction effects for smoking pack years (4.78% increase/SD, p=3.68 × 10–7), but other smoking phenotypes had insignificant interaction effects after correcting for multiple tests (minimum p=2.7 × 10–3); interactions with educational attainment were also insignificant (p=4.54 × 10–2).

Table 1
Model descriptive statistics on 28 of 62 covariates, which have significant (p<0.05/62) polygenic score (PGS)–covariate interaction terms, in UK Biobank (UKBB) European ancestry (EUR).

The third column is the percentage change in PGS effect per unit change (standard deviations for continuous variables, binary variables encoded as 0 or 1) in covariate. The fifth column is the increase in model R2 with a PGS–covariate interaction term versus a main effects only model.

Variable typeCovariate% change in βPGS per covariate unit changeInteraction pR2 increase with interaction termN
ContinuousHDL cholesterol–15.291.71 × 10–960.0012328,719
Total cholesterol–12.701.64 × 10–710.00082359,221
IPAQ–12.503.11 × 10–660.001304,951
Moderate-vigorous PA–11.418.92 × 10–650.001304,951
Diastolic BP10.846.06 × 10–600.0007352,804
Townsend Index6.782.86 × 10–580.00089376,283
Age–9.023.60 × 10–570.00061376,729
FVC–9.664.69 × 10–560.0008343,467
Drink frequency/week–19.962.62 × 10–550.0024122,281
LDL cholesterol–9.862.63 × 10–510.00058358,556
N days vigorous PA/week–9.372.42 × 10–350.0007299,963
FEV1–7.387.15 × 10–350.0005343,544
Mean alcohol consumption–7.387.65 × 10–220.00113126,756
HbA1c4.635.37 × 10–140.0002358,798
Mean drinks/week–7.661.01 × 10–130.0008112,204
Water intake4.602.97 × 10–130.00014347,472
Processed meat intake3.702.38 × 10–70.0002376,205
Starch mean5.513.15 × 10–70.00018128,346
Smoking pack years4.783.68 × 10–70.0002114,135
Protein mean4.826.52 × 10–70.00018128,181
Saturated fat mean4.921.23 × 10–60.00017127,899
Fat mean4.401.64 × 10–50.00013128,092
Saturated fat grams/week2.461.79 × 10–54.00 × 10-5364,629
Retinol mean3.773.54 × 10–49.00 × 10-5126,029
BinaryIPAQ–12.685.30 × 10–620.0009304,951
Vigorous PA/week–20.559.07 × 10–540.0009304,951
Sex–11.021.41 × 10–240.00025376,729
Diabetes27.191.83 × 10–70.0004375,903

  1. BP = blood pressure, PA = physical activity, FVC = forced vital capacity, FEV1 = forced expiratory volume in 1 s, HDL = high-density lipoprotein, LDL = high-density lipoprotein, IPAQ = International Physical Activity Questionnaire.

We replicated these analyses across ancestries and the other non-UKBB EUR cohorts (Figure 3, Supplementary file 1d). For age and sex, which were available for all cohorts, interactions were significant (p<0.05) and directionally consistent across cohorts and ancestries (except for GERA AFR, which had small sample size [N = 1,789]). We were able to test interactions with alcohol intake frequency and physical activity in GERA, and replicated significant and directionally consistent associations. We observed poor replication for LDL cholesterol, HbA1c, and smoking pack years, with insignificant and directionally inconsistent interaction effects across cohorts. Educational attainment was available in GERA, and interactions were once again insignificant. We observed significant and directionally consistent interaction effects for TG in eMERGE EUR and PMBB EUR, while the effect was inconsistent in UKBB EUR despite much larger sample size.

Figure 3 with 1 supplement see all
Download asset Open asset
Relative percentage changes in polygenic score (PGS) effect per unit change in covariate, for covariates that significantly changed PGS effect (i.e., significant interaction beta at Bonferroni p<8.1 × 10–4 – denoted by asterisks) and were present in multiple cohorts and ancestries.

Same covariate groupings and transformations were performed as in Figure 1. Similarly, actual values were used when variables had comparable units across cohorts, and standard deviations (SD) used otherwise.

However, despite significance of interaction terms, increases in model R2 when including PGS–covariate interaction terms were small. For instance, the maximum increase among all covariates in UKBB EUR was only 0.0024 from a base R2 of 0.1049 (2.1% relative increase), for alcoholic drinks per week. Across all cohorts and ancestries, the maximum increase in R2 was only 0.0058 from a base R2 of 0.09454 (6.1% relative increase), when adding a PGS–age interaction term for eMERGE EUR (p=5.40 × 10–46) – this was also the largest relative increase among models with significant interaction terms. This result suggests that, while interaction effects can significantly modify PGSBMI effect, their overall impact on model performance is relatively small, despite large differences in R2 when stratifying by covariates.

Correlations between R2 differences, interaction effects, and main effects

We next investigated the relationship between interaction effects, maximum R2 differences across quintiles, and main effects of covariates on BMI. We first estimated the main effects of each covariate on BMI (‘Materials and methods’, Supplementary file 1e), and then calculated the correlation weighted by sample size between main effects, maximum PGSBMI R2 across quintiles, and PGS–covariate interaction effects (Figure 4) across all cohorts and ancestries – GERA data were excluded from these analyses due to slightly different phenotype definition (Supplementary file 1f), as were binary variables. Interaction effects and maximum R2 differences had a 0.80 correlation (p=2.1 × 10–27), indicating that variables with larger interaction effects also had larger effects on PGSBMI performance across quintiles, and that covariates that increase PGSBMI effect also have the largest effect on PGSBMI performance, that is, individuals most at risk for obesity will have both disproportionately larger PGSBMI effect and R2. Main effects and maximum R2 differences had a 0.56 correlation (p=1.3 × 10–11), while main effects and interaction effects had a 0.58 correlation (p=7.6 × 10–12), again suggesting that PGSBMI are more predictive in individuals with higher values of BMI-associated covariates, although less predictive than estimating the interaction effects themselves directly. However, this result demonstrates that covariates that influence both PGSBMI effect and performance can be predicted just using main effects of covariates, which are often known for certain phenotypes and easier to calculate, as genetic data and PGS construction would not be required.

Figure 4
Download asset Open asset
Relationships (Pearson correlations weighted by sample size) between maximum R2 differences across strata, main effects of covariate on log(BMI), and polygenic score (PGS)–covariate interaction effects on log(BMI).

Main effect units are in standard deviations, interaction effect units are in PGS standard deviations multiplied by covariate standard deviations. Only continuous variables are plotted and modeled. Genetic Epidemiology Research on Adult Health and Aging (GERA) was excluded due to slightly different phenotype definitions. BMI, body mass index.

Increase in PGS effect for increasing percentiles of BMI itself, and its relation to R2 differences when stratifying by covariates

Given large and replicable correlations between main effects, interaction effects, and maximum R2 differences for individual covariates, it seemed these differences may be due to the differences in BMI itself, rather than any individual or combination of covariates. To assess this, we used quantile regression to evaluate the effect of PGSBMI on BMI at different deciles of BMI itself. We observed that the effect of PGSBMI consistently increases from lower deciles to higher deciles across all cohorts and ancestries (Figure 5) – for instance, in European ancestry UKBB individuals, the effect of PGSBMI (in units of log(BMI)) when predicting the bottom decile of log(BMI) was 0.716 (95% CI: 0.701–0.732), and increased to 1.31 (95% CI: 1.29–1.33) in the top decile. Across all cohorts and ancestries, the effect of PGSBMI between lowest and highest effect decile ranged from 1.43 to 2.06 times larger, with all cohorts and ancestries having nonoverlapping 95% confidence intervals between their effects (except for African ancestry eMERGE individuals, which had much smaller sample size).

Figure 5 with 2 supplements see all
Download asset Open asset
Quantile regression effects of PGSBMI (in units of log(BMI)) on log(BMI) at each decile of BMI in each cohort and ancestry.

Tau is an input parameter for quantile regression corresponding to the percentile of the BMI distribution being modeled, with lower tau values representing the lower deciles (e.g., tau = 0.1 for the 10th percentile) and higher tau values representing the upper deciles (e.g., tau = 0.9 for the 90th percentile). The effect of PGSBMI increases as BMI itself increases, suggesting that no individual covariate–PGS interaction is responsible for the nonlinear effect of PGSBMI. PGS, polygenic score; BMI, body mass index.

While this analysis showed that the effect of PGSBMI increases as BMI itself increases, which may help explain significant interaction effects between PGSBMI and different covariates, it does not directly explain differences in R2 when stratifying by different covariates – we describe several points that help explain this result and suggest they may actually be closely related. Essentially, as the magnitude of the slope of a regression line increases while the mean squared residual does not increase, model R2 will increase – we demonstrate this using simulated data (Figure 5—figure supplement 1). As the magnitude of the regression line’s slope decreases, the regression line becomes a comparatively worse predictor compared to just using the mean, which decreases R2 despite the mean error being the same across models. To demonstrate this in real data, we compared simple univariable models of log(BMI)~PGSBMI (in units of log(BMI)) between the bottom and top age quintiles in the European ancestry UKBB (Figure 5—figure supplement 2). As shown in previous sections, R2 and PGSBMI beta are higher in younger individuals (R2 = 0.088 versus R2 = 0.066, and beta = 1.12 and 0.87, respectively), which seem to be a direct consequence of one another, as the mean squared error in younger individuals is actually higher (0.027 versus 0.022, respectively). This description suggests that the use of R2 as the sole performance metric for evaluation of PGS may not always be appropriate, despite its overwhelming usage. Furthermore, this explanation helps explain the seemingly paradoxical results of significant interaction terms yet small increases in overall model R2 and comparably much larger differences in R2 in the stratified analyses.

Effects of machine learning approaches on predictive performance

Given evidence of PGS–covariate dependence, we aimed to assess increases in R2 when using machine learning models (neural networks), which can inherently model interactions and other nonlinearities, over linear models even with interaction terms. We first included age and sex as the only covariates (along with genotype PCs and PGSBMI), as age and sex were present in all datasets and had significant and replicable evidence for PGS-dependence across our analyses. Three models were assessed – L1-regularized (i.e., LASSO) linear regression without any interaction terms, LASSO including a PGS–age and PGS–sex interaction term, and neural networks (without interaction terms). When comparing neural networks to LASSO with interaction terms, the relative tenfold cross-validated R2 increased up to 67% (mean 23%) across cohorts and ancestries (Figure 6, Supplementary file 1g). The inclusion of interaction terms increased cross-validated R2 up to 12% (mean 3.9%) when comparing LASSO including interaction terms to LASSO with main effects only.

Figure 6
Download asset Open asset
Model R2 from different machine learning models across cohorts and ancestries using age and gender as covariates (along with PGSBMI and PCs 1–5).

Across all cohorts and ancestries, LASSO with PGS–age and PGS–gender interaction terms had better average tenfold cross-validation R2 than LASSO without interaction terms, while neural networks outperformed LASSO models. PGS, polygenic score; BMI, body mass index.

We then modeled all available covariates and their interactions with PGS for each cohort and did similar comparisons. Cross-validated R2 increased by up to 17.6% (mean 9.5%) when using neural networks versus LASSO with interaction terms, and up to 7.0% (mean 2.0%) when comparing LASSO with interaction terms to LASSO with main effects only. Increases in model performance using neural networks were smaller in UKBB, perhaps due to the age range being smaller than other cohorts (all participants aged 39–73). This result suggests that additional variance explained through nonlinear effects with age and sex are explained by other variables present in the remainder of the datasets. Our findings show machine learning methods can improve model R2 that include PGSBMI as variables beyond including interaction terms in linear models, even when variable selection is performed using LASSO, demonstrating that model performance can be increased beyond modeling nonlinearities through linear interaction terms and a feature selection procedure.

Calculating PGS directly from GxAge GWAS effects

Previous studies Mostafavi et al., 2020 have created PGS using GWAS stratified by different personal-level covariates, but for practical purposes this leads to a large loss of power as the full size of the GWAS is not utilized for each strata and continuous variables are forced into bins. We developed a novel strategy where PGS are instead created from a full-size GWAS that includes SNP–covariate interaction terms (‘Materials and methods’). We focused on age interactions, given their large and replicable effects based on our results – similar to a previous study (Mostafavi et al., 2020), we conducted these analyses in the European UKBB. We used a 60% random split of study individuals to conduct three sets of PGS using GWAS of the following designs: main effects only, main effects also with an SNP–age interaction term, and main effects but stratified into quartiles by age. Twenty percent of the remaining data were used for training and the final 20% were held out as a test set. The best-performing PGS created from SNP–age interaction terms (PGSGxAge) increased test R2 to 0.0771 (95% bootstrap CI: 0.0770–0.0772) from 0.0715 (95% bootstrap CI: 0.0714–0.0716), a 7.8% relative increase compared to the best-performing main effect PGS (Figure 7, Supplementary file 1h) – age-stratified PGS had much lower performance than both other strategies (unsurprising given reasons previously mentioned). Including a PGSGxAge–age interaction term only marginally increased R2 (0.0001 increase), with similar increases for the other two sets of PGS, further demonstrating that post hoc modeling of interactions cannot reconcile performance gained through directly incorporating interaction effects from the original GWAS. The strategy of creating PGS directly from full-sized SNP–covariate interactions is potentially quite useful as it increases PGS performance without the need for additional data – there are almost certainly a variety of points of improvement (described more in ‘Discussion’), but we consider their investigation outside the scope of this study.

Figure 7
Download asset Open asset
Polygenic score (PGS) R2 based on three sets of genome-wide association studies (GWAS) setups.

‘Main effects’ were from a typical main effect GWAS, ‘GxAge’ effects were from a GWAS with an SNP–age interaction term, and ‘Age stratified’ GWAS had main effects only but were conducted in four age quartiles. PGS R2 was evaluated using two models: one with main effects only and one with an additional PGS * Age interaction term.

Discussion

We uncovered replicable effects of covariates across four large-scale cohorts of diverse ancestry on both performance and effects of PGSBMI. When stratifying by quintiles of different covariates, certain covariates had significant and replicable evidence for differences in PGSBMI R2, with R2 being nearly double between top- and bottom-performing quintiles for covariates with the largest differences. When testing PGS–covariate interaction effects, we also found covariates with significant interaction effects, where, for the largest effect covariates, each standard deviation change affected PGSBMI effect by nearly 20%. Across analyses, we found age and sex had the most replicable interaction effects, with levels of serum cholesterol, physical activity, and alcohol consumption having the largest effects across cohorts. Interaction effects and R2 differences were strongly correlated, with main effects also being correlated with interaction effects and R2 differences, suggesting that covariates with the largest interaction effects also contribute to the largest R2 differences, with simple main effects also being predictive of expected differences in R2 and interaction effects. Relatedly, we observed the effect of PGSBMI increases as BMI itself increases, and reason that differences in R2 when stratifying by covariates are largely a consequence of difference in PGSBMI effects. Next, we employed machine learning methods for prediction of BMI with models that include PGSBMI and demonstrate that these methods outperform regularized linear regression models that include interaction effects. Finally, we employed a novel strategy that directly incorporates SNP interaction effects into PGS construction and demonstrate that this strategy improves PGS performance when modeling SNP–age interactions compared to PGS created only from main effects.

These observations are relevant to current research and clinical use of PGS, as individuals above a percentile cutoff are designated high risk (Ge et al., 2019), implying that individuals most at risk for obesity have both disproportionately higher predicted BMI and increased BMI prediction performance compared to the general population. More broadly, these results may likely extend to single variant effects instead of those aggregated into a PGS, which may inform the cause of previous GxE discoveries – for instance, variants near FTO that interact with physical activity discovered through GWAS of BMI are robust and well-documented. However, individuals engaging in physical activity will generally have lower BMI than those that are sedentary, and these results suggest it may not be the difference in physical activity that is driving the interaction, rather the difference in BMI itself. This concept may also apply to other traits – for instance, sex-specific analyses are commonly performed, and variants with differing effects between male and female GWAS may largely be explained by phenotypic differences, rather than any combination of biological or lifestyle differences.

Future work may include replicating these analyses across additional traits, and trying to understand why these differences occur, as well as further exploring machine learning and deep learning methods on other phenotypes to determine if this trend of inclusion of PGS, along with covariate interaction effects, outperforms linear models for risk prediction. Additionally, inclusion of a PGS for the covariate to better measure its environmental effect is potentially worth exploring further and should improve in the future as PGS performance continues to increase. A slight limitation of this method in our study is that for the UKBB analyses the GWAS used for PGS construction were also from UKBB, thus not out-of-sample, although many of the covariates only have GWAS available through UKBB individuals. Furthermore, a variety of improvements are likely possible when creating PGS directly from SNP–covariate interaction terms. First, we used the same SNPs that were selected by pruning and thresholding based on their main effect p-values, but selection of SNPs based on their interaction p-values should also be possible and would likely improve performance. Additionally, performance of pruning and thresholding-based strategies has largely been overtaken by methods that first adjust all SNP effects for LD and do not require exclusion of SNPs, and a method that could do a similar adjustment for interaction effects would likely outperform most current methods for traits with significant context-specific effects. Next, incorporating additional SNP–covariate interactions (e.g., SNP–sex) would also likely further improve prediction performance, although any SNP selection/adjustment procedures may be further complicated by additional interaction terms. Finally, if SNP effects do truly differ according to differences in phenotype, then SNP effects would differ depending on the alleles one has, implying epistatic interactions are occurring at these SNPs.

While difference in phenotype itself may be able to explain the difference in genetic effects, it still may be that specific environmental or lifestyle characteristics are driving the differences. We propose several ideas about why BMI-associated covariates have larger interaction effects and impact on R2 for PGSBMI. First, age may be a proxy for accumulated gene–environment interactions as younger individuals have less exposure to environmental influences on weight compared to older individuals; therefore, one may expect that in younger individuals their phenotype could be better explained by genetics compared to environment. Second, PGS may more readily explain high phenotype values especially for positively skewed phenotypes, as large effect variants (e.g., associated with very high weight or height; Robinson et al., 2006) may be more responsible for extremely high phenotypic values. For example, the distribution of BMI is often positively skewed, and effects in trait-increasing alleles may have a larger potential to explain trait variation compared to trait-reducing variants. This explanation would likely be better suited to positively skewed traits and is not fully satisfactory as first log-transforming or rank-normal transforming the phenotype, as was done in this study, may invalidate this explanation.

PGS is a promising technique to stratify individuals for their risk of common, complex disease. To achieve more accurate predictions as well as promote equity, further research is required regarding PGS methods and assessments. This research provides firm evidence supporting the context-specific nature of PGS and the impact of nonlinear covariate effects for improving polygenic prediction of BMI, promoting equitable use of PGS across ancestries and cohorts.

Materials and methods

Study datasets

Request a detailed protocol

Individual inclusion criteria and sample sizes per cohort are described below – additional information is available in Supplementary file 1a.

UKBB

Request a detailed protocol

Individual-level quality control (QC) and filtering have been described elsewhere (Zhang et al., 2022) for European ancestry individuals. Briefly, individuals were split by ancestry according to both genetically inferred ancestry and self-reported ethnicity (Bycroft et al., 2018). Individuals with low genotyping quality and sex mismatch were removed, only unrelated individuals (Plink pi-hat < 0.250) were retained, and variants were filtered at INFO > 0.30 and minor allele frequency > 0.01. For African ancestry, individuals were first selected based on self-reported ethnicity ‘Black or Black British’, ‘Caribbean’, ‘African’, or ‘Any other Black background’. Individuals who were low quality, that is, ‘Outliers for heterozygosity or missing rate’, and who were Caucasian from ‘Genetic ethnic grouping’ were removed. Of these individuals, those who were ±6 standard deviations from the mean of the first five genetic principal components provided by UKBB were excluded. Finally, only unrelated individuals were retained up to the second degree using plink2 (Chang et al., 2015) ‘-king-cutoff 0.125’. After QC and consideration of phenotype, a total of 7,046 individuals in the UKBB AFR data who also had BMI available were used for downstream analyses. In total, 383,775 individuals were used for analysis (NEUR = 376,729, NAFR = 7,046).

eMERGE

Request a detailed protocol

Ancestry and relatedness inference have been described elsewhere (Stanaway et al., 2019). Individuals were split into European and African ancestry cohorts, and related individuals were removed (Plink pi-hat > 0.250) such that all were unrelated. In total, 35,064 individuals (NEUR = 31,961, NAFR = 3,103) were used for analysis.

GERA

Request a detailed protocol

Ancestry inference has been described elsewhere (Banda et al., 2015), and study individuals were divided into European and African ancestry cohorts. Related individuals were removed using plink2 ‘-king-cutoff 0.125’. In total, 57,838 individuals (NEUR = 56,049, NAFR = 1,789) were used for analysis.

PMBB

Request a detailed protocol

Ancestry inference and relatedness inference have been described elsewhere (Penn Medicine BioBank, 2022). Individuals were split into European and African ancestry cohorts, and related individuals were removed at pi-hat > 0.250. In total, 36,046 individuals (NEUR = 26,372, NAFR = 9,674) were used for analysis.

Choice of covariates

Request a detailed protocol

A total of 62 covariates were included in the analyses, 25 of which were present (or similar proxies) in multiple datasets. These covariates were selected based on relevance to cardiometabolic health and obesity, and previous evidence of context-specific effects with BMI (Rask-Andersen et al., 2017; Robinson et al., 2017; Justice et al., 2017; Tyrrell et al., 2017; Young et al., 2016; Winkler et al., 2015). For UKBB, phenotype values were used from the collection that was closest to recruitment, and for PMBB the median values were used – for GERA and eMERGE, only one value was available. Additional details on covariate constructions, transformations, filtering, and cohort availability are provided in Supplementary file 1b.

PGS construction

Request a detailed protocol

PGS for BMI (PGSBMI) were constructed using PRS-CSx (Ruan et al., 2022), using GWAS summary statistics for individuals of European (Locke et al., 2015), African (Ng et al., 2017), and East Asian (Sakaue et al., 2021) ancestry that were out-of-sample of study participants. A set of 1.29 million HapMap3 SNPs provided by PRS-CSx was used for PGS calculation, which are generally well-imputed and variable frequency across global populations. Default settings for PRS-CSx (downloaded November 22, 2021) were used, which have been demonstrated to perform well for highly polygenic traits such as BMI (list of parameters is provided in Supplementary file 1i). The final PGSBMI per ancestry and cohort was calculated by regressing log(BMI) on the PGSBMI per ancestry without covariates – the combined, predicted value was used as a single PGSBMI in downstream analyses.

For GERA, BMI was not transformed as it was already binned into a categorical variable with five levels (18≤, 19–25, 26–29, 30–39, >40). Additionally, for GERA the uncombined ancestry-specific PGSBMI was used in the final models as it had higher R2 than using the combined PGSBMI (data not shown).

PGSBMI performance after covariate stratification

Request a detailed protocol

Analyses were performed separately for each cohort and ancestry. For each covariate, individuals were binned by binary covariates or quintiles for continuous covariates. Incremental PGSBMI R2 was calculated by taking the difference in R2 between:

log(BMI)∼PGSBMI+Age+Sex+PCs1−5
log(BMI)∼Age+Sex+PCs1−5

We performed 5,000 bootstrap replications to obtain a bootstrapped distribution of R2. p-Values for differences in R2 were calculated between groups by calculating the proportion of overlap between two normal distributions of the R2 value using the standard deviations of the bootstrap distributions. Again for GERA, BMI was not transformed.

PGSBMI interaction modeling

Request a detailed protocol

Evidence for interaction with each covariate with the PGSBMI was evaluated using linear regression. It has been reported that the inclusion of covariates that are genetically correlated with the outcome can inflate test statistic estimates (Aschard et al., 2015; Kerin and Marchini, 2020; Vanderweele et al., 2013). To assuage these concerns, we introduced a novel correction, where we first calculated a PGS for the covariate (PGSCovariate) and included it in the model, as well as an interaction term between PGSBMI and PGSCovariate. The PGSCovariate terms were calculated using the European ancestry Neale Lab summary statistics (URLs) and PRS-CS (Ge et al., 2019). To standardize effect sizes across analyses, PGSBMI and Covariate were first converted to mean zero and standard deviation of 1 (binary covariates were not standardized). We demonstrate inclusion of PGSCovariate terms successfully reduced significance of the PGSBMI * Covariate term (Figure 3—figure supplement 1). The final model used to evaluate PGSBMI and Covariate interactions was

log(BMI)∼PGSBMI∗Covariate+PGSBMI+Covariate+PGSCovariate+PGSBMI∗PGSCovariate+Age+Sex+PCs1−5

We report the effect estimates of the PGSBMI * Covariate term, and differences in model R2 with and without the PGSBMI * Covariate term. Again for GERA, BMI was not transformed.

Correlation between R2 differences, interaction effects, and main effects

Request a detailed protocol

We estimated the main effects of each covariate on BMI with the following model:

log(BMI)∼Covariate+Age+Sex+PCs1−5

Note that we ran new models with main effects only, instead of using the main effect from the interaction models (as the main effects in the interaction models depend on the interaction terms, and main effects used to create interaction terms are sensitive to centering of variables despite the scale invariance of linear regression itself; Afshartous and Preston, 2011). We then estimated the correlation between main effects, interaction effects, and maximum R2 differences across all cohorts and ancestries weighting by sample size, analyzing quantitative and binary variables separately.

Quantile regression to measure PGS effect across percentiles of BMI

Request a detailed protocol

The effect of PGSBMI on BMI at different deciles of BMI was assessed using quantile regression. Tau – the parameter that sets which percentile to be predicted – was set to 0.10,0.20, …,0.90. Models included age, sex, and the top 5 genetic PCs as covariates. Analyses were stratified by ancestry and cohort, and BMI was first log transformed. GERA was excluded from these analyses as a portion of the models failed to run (as BMI values from GERA were already binned, some deciles all had the same BMI value – additionally, difference in effects between bins would be harder to evaluate as BMI within each decile would be more homogeneous).

Machine learning models

Request a detailed protocol

UKBB EUR and GERA EUR models were restricted to 30,000 random individuals for computational reasons – BMI distributions did not differ from the full-sized datasets (Kolmogorov–Smirnov p-values of 0.29 and 0.57, respectively). PGSBMI and top 5 genetic principal components were included as features in all models. Two sets of models were evaluated for each cohort and ancestry: including age and sex as features, and including all available covariates in each cohort as features. Interaction terms between PGS and each covariate were included for models using interaction terms. ‘Ever Smoker’ status was used in favor of ‘Never’ versus ‘Current smoking’ status (if present) as individuals with ‘Never’ versus ‘Current’ status are a subset of those with ‘Ever Smoker’ status. UKBB AFR with all covariates was excluded due to small sample size (N = 53).

Neural networks were used as the model of choice, given their inherent ability to model interactions and other nonlinear dependencies. Prior to modeling, all features were scaled to be between 0 and 1. We used average tenfold cross-validation R2 to evaluate model performance. Separate models were trained using untransformed and log(BMI). L1-regularized linear regression was used with 18 values of lambda (1.0, 5.0 × 10–1, 2.0 × 10–1, 1.0 × 10–1, 5.0 × 10–2, 2.0 × 10–2, …, 1.0 × 10–5, 5.0 × 10–6, 2.0 × 10–6). Models were trained without inclusion of interaction terms (which neural networks can implicitly model) using 1,000 iterations of random search with the following hyperparameter ranges: size of hidden layers [10, 200], learning rate [0.01, 0.0001], type of learning rate [constant, inverse scaling], power t [0.4, 0.6], momentum [0.80, 1.0], batch size [32, 256], and number of hidden layers [1, 2].

GxAge PGSBMI creation and assessment

Request a detailed protocol

Analyses were conducted in the European UKBB (N = 376,629), as was done in a study on a similar topic (Mostafavi et al., 2020). Three sets of analyses were performed using GWAS conducted in a 60% random split of individuals using the following models (BMI was rank-normal transformed before GWAS):

  1. BMI∼SNP+Age+Sex+PCs1−5

  2. BMI∼SNP+Age∗SNP+Age+Sex+PCs1−5

  3. Using the model in (1) but stratified into quartiles by age. BMI was rank-normal transformed within each quartile.

Using each set of GWAS, PGS was first calculated in a 20% randomly selected training set of the dataset using pruning and thresholding using 10 p-value thresholds (0.50, 0.10, …, 5.0 × 10–5, 1.0 × 10–5) and remaining settings as default in Plink 1.9. For (2), GxAge PGSBMI was calculated using SNPs clumped by their main effect p-values from (1), and additionally incorporating the GxAge interaction effects per SNP. In other words, instead of typical PGS construction as

PGSi=β1k1+β2k2+...+βnkn

For an individual i’s PGS calculation, with main SNP effect β, and n SNPs, PGS incorporating GxAge effects (PGSGxAge) was calculated as

PGSGXAge,i=β1k1+βGXAge,1k1Agei+β2k2+βGXAge,2k2Agei...+βnkn+βGXAge,nknAgei

where βGxAge is the GxAge effect for each SNP n and Agei is the age for individual i.

For each of the three analyses, the parameters and models resulting in the best-performing PGS (highest incremental R2, using same main effect covariates as in the three GWAS) from the training set were evaluated in the remaining 20% of the study individuals. For (3), models were first trained within each quartile separately. To maintain sense of scale across quartiles (after rank-normal transformation), R2 between all predicted values and true values was calculated together. For R2 confidence intervals, the training set was bootstrapped and evaluated on the test set 5,000 times.

URLs

Neale Lab UKBB summary statistics: http://www.nealelab.is/uk-biobank.

Select analysis code and data are available at https://github.com/RitchieLab/BMI_PGS_eLife (copy archived at Ritchie Lab, 2024).

Appendix 1

RGC management and leadership team

Goncalo Abecasis, PhD, Aris Baras, M.D., Michael Cantor, M.D., Giovanni Coppola, M.D., Andrew Deubler, Aris Economides, Ph.D., Luca A. Lotta, M.D., Ph.D., John D. Overton, Ph.D., Jeffrey G. Reid, Ph.D., Katherine Siminovitch, M.D., Alan Shuldiner, M.D.

Sequencing and lab operations

Christina Beechert, Caitlin Forsythe, M.S., Erin D. Fuller, Zhenhua Gu, M.S., Michael Lattari, Alexander Lopez, M.S., John D. Overton, Ph.D., Maria Sotiropoulos Padilla, M.S., Manasi Pradhan, M.S., Kia Manoochehri, B.S., Thomas D. Schleicher, M.S., Louis Widom, Sarah E. Wolf, M.S., Ricardo H. Ulloa, B.S.

Clinical informatics

Amelia Averitt, Ph.D., Nilanjana Banerjee, Ph.D., Michael Cantor, M.D., Dadong Li, Ph.D., Sameer Malhotra, M.D., Deepika Sharma, MHI, Jeffrey Staples, Ph.D.

Genome informatics

Xiaodong Bai, Ph.D., Suganthi Balasubramanian, Ph.D., Suying Bao, Ph.D., Boris Boutkov, Ph.D., Siying Chen, Ph.D., Gisu Eom, B.S., Lukas Habegger, Ph.D., Alicia Hawes, B.S., Shareef Khalid, Olga Krasheninina, M.S., Rouel Lanche, B.S., Adam J. Mansfield, B.A., Evan K. Maxwell, Ph.D., George Mitra, B.A., Mona Nafde, M.S., Sean O’Keeffe, Ph.D., Max Orelus, B.B.A., Razvan Panea, Ph.D., Tommy Polanco, B.A., Ayesha Rasool, M.S., Jeffrey G. Reid, Ph.D., William Salerno, Ph.D., Jeffrey C. Staples, Ph.D., Kathie Sun, Ph.D.

Analytical genomics and data science

Goncalo Abecasis, D.Phil., Joshua Backman, Ph.D., Amy Damask, Ph.D., Lee Dobbyn, Ph.D., Manuel Allen Revez Ferreira, Ph.D., Arkopravo Ghosh, M.S., Christopher Gillies, Ph.D., Lauren Gurski, B.S., Eric Jorgenson, Ph.D., Hyun Min Kang, Ph.D., Michael Kessler, Ph.D., Jack Kosmicki, Ph.D., Alexander Li, Ph.D., Nan Lin, Ph.D., Daren Liu, M.S., Adam Locke, Ph.D., Jonathan Marchini, Ph.D., Anthony Marcketta, M.S., Joelle Mbatchou, Ph.D., Arden Moscati, Ph.D., Charles Paulding, Ph.D., Carlo Sidore, Ph.D., Eli Stahl, Ph.D., Kyoko Watanabe, Ph.D., Bin Ye, Ph.D., Blair Zhang, Ph.D., Andrey Ziyatdinov, Ph.D.

Therapeutic area genetics

Ariane Ayer, B.S., Aysegul Guvenek, Ph.D., George Hindy, Ph.D., Giovanni Coppola, M.D., Jan Freudenberg, M.D., Jonas Bovijn M.D., Katherine Siminovitch, M.D., Kavita Praveen, Ph.D., Luca A. Lotta, M.D., Manav Kapoor, Ph.D., Mary Haas, Ph.D., Moeen Riaz, Ph.D., Niek Verweij, Ph.D., Olukayode Sosina, Ph.D., Parsa Akbari, Ph.D., Priyanka Nakka, Ph.D., Sahar Gelfman, Ph.D., Sujit Gokhale, B.E., Tanima De, Ph.D., Veera Rajagopal, Ph.D., Alan Shuldiner, M.D., Bin Ye, Ph.D., Gannie Tzoneva, Ph.D., Juan Rodriguez-Flores, Ph.D.

Research program management and strategic initiatives

Esteban Chen, M.S., Marcus B. Jones, Ph.D., Michelle G. LeBlanc, Ph.D., Jason Mighty, Ph.D., Lyndon J. Mitnaul, Ph.D., Nirupama Nishtala, Ph.D., Nadia Rana, Ph.D., Jaimee Hernandez

Penn Medicine BioBank Banner Author List and Contribution Statements

PMBB Leadership Team

Daniel J. Rader, M.D., Marylyn D. Ritchie, Ph.D.

Contribution: All authors contributed to securing funding, study design and oversight. All authors reviewed the final version of the manuscript.

Patient Recruitment and Regulatory Oversight

JoEllen Weaver, Nawar Naseer, Ph.D., M.P.H., Giorgio Sirugo, M.D., P.h.D., Afiya Poindexter, Yi-An Ko, Ph.D., Kyle P. Nerz

Contributions: JW manages patient recruitment and regulatory oversight of study. NN manages participant engagement, assists with regulatory oversight, and researcher access. GS assists with researcher access. AP, YK, KPN perform recruitment and enrollment of study participants.

Lab Operations

JoEllen Weaver, Meghan Livingstone, Fred Vadivieso, Stephanie DerOhannessian, Teo Tran, Julia Stephanowski, Salma Santos, Ned Haubein, P.h.D., Joseph Dunn

Contribution: JW, ML, FV, SD conduct oversight of lab operations. ML, FV, AK, SD, TT,JS, SS perform sample processing. NH, JD are responsible for sample tracking and the laboratory information management system.

Clinical Informatics

Anurag Verma, Ph.D., Colleen Morse Kripke, M.S. DPT, MSA, Marjorie Risman, M.S., Renae Judy, B.S., Colin Wollack, M.S.

Contribution: All authors contributed to the development and validation of clinical phenotypes used to identify study subjects and (when applicable) controls.

Genome Informatics

Anurag Verma Ph.D., Shefali S. Verma, Ph.D., Scott Damrauer, M.D., Yuki Bradford, M.S., Scott Dudek, M.S., Theodore Drivas, M.D., Ph.D.

Contribution: AV, SSV, and SD are responsible for the analysis, design, and infrastructure needed to quality control genotype and exome data. YB performs the analysis. TD and AV provides variant and gene annotations and their functional interpretation of variants.

For PMBB, please use: biobank@pennmedicine.upenn.edu

For Regeneron, please use: RGCcollaborations@regeneron.com

Data availability

UK Biobank data was accessed under project #32133. eMERGE data is available at dbGaP in phs001584.v2.p2. GERA data is available at dbGaP in phs000674.v3.p3. PMBB individual level genotype and phenotype data can be accessed through a research collaboration with a Penn investigator, as long as the requestor is from a not-for-profit organization. The PMBB genetic data used in this study were generated in collaboration with Regeneron Genetics Center; as such, we are unable to share the data with for-profit organizations without a three-way research collaboration agreement. Data sharing will require a PMBB project proposal and IRB approval. Collaboration requests can be sent to biobank@upenn.edu. Select analysis code and data are available at https://github.com/RitchieLab/BMI_PGS_eLife (copy archived at Ritchie Lab, 2024).

The following previously published data sets were used
    1. Banda Y
    2. Kvale MN
    3. Hoffmann TJ
    4. Hesselson SE
    5. Ranatunga D
    6. Tang H
    (2015) dbGaP
    ID phs000674.v3.p3. Resource for Genetic Epidemiology Research on Aging (GERA).
    https://www.ncbi.nlm.nih.gov/projects/gap/cgi-bin/study.cgi?study_id=phs000674.v3.p3
    1. Stanaway IB
    2. Hall TO
    3. Rosenthal EA
    4. Palmer M
    5. Naranbhai V
    6. Knevel R
    (2020) dbGaP
    ID phs001584.v2.p2. eMERGE Network Phase III: HRC SNV and 1000 Genomes SV Imputed Array Data of 105,000 Participants.
    https://www.ncbi.nlm.nih.gov/projects/gap/cgi-bin/study.cgi?study_id=phs001584.v2.p2

References

    1. Justice AE
    2. Winkler TW
    3. Feitosa MF
    4. Graff M
    5. Fisher VA
    6. Young K
    7. Barata L
    8. Deng X
    9. Czajkowski J
    10. Hadley D
    11. Ngwa JS
    12. Ahluwalia TS
    13. Chu AY
    14. Heard-Costa NL
    15. Lim E
    16. Perez J
    17. Eicher JD
    18. Kutalik Z
    19. Xue L
    20. Mahajan A
    21. Renström F
    22. Wu J
    23. Qi Q
    24. Ahmad S
    25. Alfred T
    26. Amin N
    27. Bielak LF
    28. Bonnefond A
    29. Bragg J
    30. Cadby G
    31. Chittani M
    32. Coggeshall S
    33. Corre T
    34. Direk N
    35. Eriksson J
    36. Fischer K
    37. Gorski M
    38. Neergaard Harder M
    39. Horikoshi M
    40. Huang T
    41. Huffman JE
    42. Jackson AU
    43. Justesen JM
    44. Kanoni S
    45. Kinnunen L
    46. Kleber ME
    47. Komulainen P
    48. Kumari M
    49. Lim U
    50. Luan J
    51. Lyytikäinen LP
    52. Mangino M
    53. Manichaikul A
    54. Marten J
    55. Middelberg RPS
    56. Müller-Nurasyid M
    57. Navarro P
    58. Pérusse L
    59. Pervjakova N
    60. Sarti C
    61. Smith AV
    62. Smith JA
    63. Stančáková A
    64. Strawbridge RJ
    65. Stringham HM
    66. Sung YJ
    67. Tanaka T
    68. Teumer A
    69. Trompet S
    70. van der Laan SW
    71. van der Most PJ
    72. Van Vliet-Ostaptchouk JV
    73. Vedantam SL
    74. Verweij N
    75. Vink JM
    76. Vitart V
    77. Wu Y
    78. Yengo L
    79. Zhang W
    80. Hua Zhao J
    81. Zimmermann ME
    82. Zubair N
    83. Abecasis GR
    84. Adair LS
    85. Afaq S
    86. Afzal U
    87. Bakker SJL
    88. Bartz TM
    89. Beilby J
    90. Bergman RN
    91. Bergmann S
    92. Biffar R
    93. Blangero J
    94. Boerwinkle E
    95. Bonnycastle LL
    96. Bottinger E
    97. Braga D
    98. Buckley BM
    99. Buyske S
    100. Campbell H
    101. Chambers JC
    102. Collins FS
    103. Curran JE
    104. de Borst GJ
    105. de Craen AJM
    106. de Geus EJC
    107. Dedoussis G
    108. Delgado GE
    109. den Ruijter HM
    110. Eiriksdottir G
    111. Eriksson AL
    112. Esko T
    113. Faul JD
    114. Ford I
    115. Forrester T
    116. Gertow K
    117. Gigante B
    118. Glorioso N
    119. Gong J
    120. Grallert H
    121. Grammer TB
    122. Grarup N
    123. Haitjema S
    124. Hallmans G
    125. Hamsten A
    126. Hansen T
    127. Harris TB
    128. Hartman CA
    129. Hassinen M
    130. Hastie ND
    131. Heath AC
    132. Hernandez D
    133. Hindorff L
    134. Hocking LJ
    135. Hollensted M
    136. Holmen OL
    137. Homuth G
    138. Jan Hottenga J
    139. Huang J
    140. Hung J
    141. Hutri-Kähönen N
    142. Ingelsson E
    143. James AL
    144. Jansson JO
    145. Jarvelin MR
    146. Jhun MA
    147. Jørgensen ME
    148. Juonala M
    149. Kähönen M
    150. Karlsson M
    151. Koistinen HA
    152. Kolcic I
    153. Kolovou G
    154. Kooperberg C
    155. Krämer BK
    156. Kuusisto J
    157. Kvaløy K
    158. Lakka TA
    159. Langenberg C
    160. Launer LJ
    161. Leander K
    162. Lee NR
    163. Lind L
    164. Lindgren CM
    165. Linneberg A
    166. Lobbens S
    167. Loh M
    168. Lorentzon M
    169. Luben R
    170. Lubke G
    171. Ludolph-Donislawski A
    172. Lupoli S
    173. Madden PAF
    174. Männikkö R
    175. Marques-Vidal P
    176. Martin NG
    177. McKenzie CA
    178. McKnight B
    179. Mellström D
    180. Menni C
    181. Montgomery GW
    182. Musk AB
    183. Narisu N
    184. Nauck M
    185. Nolte IM
    186. Oldehinkel AJ
    187. Olden M
    188. Ong KK
    189. Padmanabhan S
    190. Peyser PA
    191. Pisinger C
    192. Porteous DJ
    193. Raitakari OT
    194. Rankinen T
    195. Rao DC
    196. Rasmussen-Torvik LJ
    197. Rawal R
    198. Rice T
    199. Ridker PM
    200. Rose LM
    201. Bien SA
    202. Rudan I
    203. Sanna S
    204. Sarzynski MA
    205. Sattar N
    206. Savonen K
    207. Schlessinger D
    208. Scholtens S
    209. Schurmann C
    210. Scott RA
    211. Sennblad B
    212. Siemelink MA
    213. Silbernagel G
    214. Slagboom PE
    215. Snieder H
    216. Staessen JA
    217. Stott DJ
    218. Swertz MA
    219. Swift AJ
    220. Taylor KD
    221. Tayo BO
    222. Thorand B
    223. Thuillier D
    224. Tuomilehto J
    225. Uitterlinden AG
    226. Vandenput L
    227. Vohl MC
    228. Völzke H
    229. Vonk JM
    230. Waeber G
    231. Waldenberger M
    232. Westendorp RGJ
    233. Wild S
    234. Willemsen G
    235. Wolffenbuttel BHR
    236. Wong A
    237. Wright AF
    238. Zhao W
    239. Zillikens MC
    240. Baldassarre D
    241. Balkau B
    242. Bandinelli S
    243. Böger CA
    244. Boomsma DI
    245. Bouchard C
    246. Bruinenberg M
    247. Chasman DI
    248. Chen YD
    249. Chines PS
    250. Cooper RS
    251. Cucca F
    252. Cusi D
    253. de Faire U
    254. Ferrucci L
    255. Franks PW
    256. Froguel P
    257. Gordon-Larsen P
    258. Grabe HJ
    259. Gudnason V
    260. Haiman CA
    261. Hayward C
    262. Hveem K
    263. Johnson AD
    264. Wouter Jukema J
    265. Kardia SLR
    266. Kivimaki M
    267. Kooner JS
    268. Kuh D
    269. Laakso M
    270. Lehtimäki T
    271. Marchand LL
    272. März W
    273. McCarthy MI
    274. Metspalu A
    275. Morris AP
    276. Ohlsson C
    277. Palmer LJ
    278. Pasterkamp G
    279. Pedersen O
    280. Peters A
    281. Peters U
    282. Polasek O
    283. Psaty BM
    284. Qi L
    285. Rauramaa R
    286. Smith BH
    287. Sørensen TIA
    288. Strauch K
    289. Tiemeier H
    290. Tremoli E
    291. van der Harst P
    292. Vestergaard H
    293. Vollenweider P
    294. Wareham NJ
    295. Weir DR
    296. Whitfield JB
    297. Wilson JF
    298. Tyrrell J
    299. Frayling TM
    300. Barroso I
    301. Boehnke M
    302. Deloukas P
    303. Fox CS
    304. Hirschhorn JN
    305. Hunter DJ
    306. Spector TD
    307. Strachan DP
    308. van Duijn CM
    309. Heid IM
    310. Mohlke KL
    311. Marchini J
    312. Loos RJF
    313. Kilpeläinen TO
    314. Liu CT
    315. Borecki IB
    316. North KE
    317. Cupples LA
    (2017) Genome-wide meta-analysis of 241,258 adults accounting for smoking behaviour identifies novel loci for obesity traits
    Nature Communications 8:14977.
    https://doi.org/10.1038/ncomms14977
    1. Locke AE
    2. Kahali B
    3. Berndt SI
    4. Justice AE
    5. Pers TH
    6. Day FR
    7. Powell C
    8. Vedantam S
    9. Buchkovich ML
    10. Yang J
    11. Croteau-Chonka DC
    12. Esko T
    13. Fall T
    14. Ferreira T
    15. Gustafsson S
    16. Kutalik Z
    17. Luan J
    18. Mägi R
    19. Randall JC
    20. Winkler TW
    21. Wood AR
    22. Workalemahu T
    23. Faul JD
    24. Smith JA
    25. Zhao JH
    26. Zhao W
    27. Chen J
    28. Fehrmann R
    29. Hedman ÅK
    30. Karjalainen J
    31. Schmidt EM
    32. Absher D
    33. Amin N
    34. Anderson D
    35. Beekman M
    36. Bolton JL
    37. Bragg-Gresham JL
    38. Buyske S
    39. Demirkan A
    40. Deng G
    41. Ehret GB
    42. Feenstra B
    43. Feitosa MF
    44. Fischer K
    45. Goel A
    46. Gong J
    47. Jackson AU
    48. Kanoni S
    49. Kleber ME
    50. Kristiansson K
    51. Lim U
    52. Lotay V
    53. Mangino M
    54. Leach IM
    55. Medina-Gomez C
    56. Medland SE
    57. Nalls MA
    58. Palmer CD
    59. Pasko D
    60. Pechlivanis S
    61. Peters MJ
    62. Prokopenko I
    63. Shungin D
    64. Stančáková A
    65. Strawbridge RJ
    66. Sung YJ
    67. Tanaka T
    68. Teumer A
    69. Trompet S
    70. van der Laan SW
    71. van Setten J
    72. Van Vliet-Ostaptchouk JV
    73. Wang Z
    74. Yengo L
    75. Zhang W
    76. Isaacs A
    77. Albrecht E
    78. Ärnlöv J
    79. Arscott GM
    80. Attwood AP
    81. Bandinelli S
    82. Barrett A
    83. Bas IN
    84. Bellis C
    85. Bennett AJ
    86. Berne C
    87. Blagieva R
    88. Blüher M
    89. Böhringer S
    90. Bonnycastle LL
    91. Böttcher Y
    92. Boyd HA
    93. Bruinenberg M
    94. Caspersen IH
    95. Chen Y-DI
    96. Clarke R
    97. Daw EW
    98. de Craen AJM
    99. Delgado G
    100. Dimitriou M
    101. Doney ASF
    102. Eklund N
    103. Estrada K
    104. Eury E
    105. Folkersen L
    106. Fraser RM
    107. Garcia ME
    108. Geller F
    109. Giedraitis V
    110. Gigante B
    111. Go AS
    112. Golay A
    113. Goodall AH
    114. Gordon SD
    115. Gorski M
    116. Grabe H-J
    117. Grallert H
    118. Grammer TB
    119. Gräßler J
    120. Grönberg H
    121. Groves CJ
    122. Gusto G
    123. Haessler J
    124. Hall P
    125. Haller T
    126. Hallmans G
    127. Hartman CA
    128. Hassinen M
    129. Hayward C
    130. Heard-Costa NL
    131. Helmer Q
    132. Hengstenberg C
    133. Holmen O
    134. Hottenga J-J
    135. James AL
    136. Jeff JM
    137. Johansson Å
    138. Jolley J
    139. Juliusdottir T
    140. Kinnunen L
    141. Koenig W
    142. Koskenvuo M
    143. Kratzer W
    144. Laitinen J
    145. Lamina C
    146. Leander K
    147. Lee NR
    148. Lichtner P
    149. Lind L
    150. Lindström J
    151. Lo KS
    152. Lobbens S
    153. Lorbeer R
    154. Lu Y
    155. Mach F
    156. Magnusson PKE
    157. Mahajan A
    158. McArdle WL
    159. McLachlan S
    160. Menni C
    161. Merger S
    162. Mihailov E
    163. Milani L
    164. Moayyeri A
    165. Monda KL
    166. Morken MA
    167. Mulas A
    168. Müller G
    169. Müller-Nurasyid M
    170. Musk AW
    171. Nagaraja R
    172. Nöthen MM
    173. Nolte IM
    174. Pilz S
    175. Rayner NW
    176. Renstrom F
    177. Rettig R
    178. Ried JS
    179. Ripke S
    180. Robertson NR
    181. Rose LM
    182. Sanna S
    183. Scharnagl H
    184. Scholtens S
    185. Schumacher FR
    186. Scott WR
    187. Seufferlein T
    188. Shi J
    189. Smith AV
    190. Smolonska J
    191. Stanton AV
    192. Steinthorsdottir V
    193. Stirrups K
    194. Stringham HM
    195. Sundström J
    196. Swertz MA
    197. Swift AJ
    198. Syvänen A-C
    199. Tan S-T
    200. Tayo BO
    201. Thorand B
    202. Thorleifsson G
    203. Tyrer JP
    204. Uh H-W
    205. Vandenput L
    206. Verhulst FC
    207. Vermeulen SH
    208. Verweij N
    209. Vonk JM
    210. Waite LL
    211. Warren HR
    212. Waterworth D
    213. Weedon MN
    214. Wilkens LR
    215. Willenborg C
    216. Wilsgaard T
    217. Wojczynski MK
    218. Wong A
    219. Wright AF
    220. Zhang Q
    221. LifeLines Cohort Study
    222. Brennan EP
    223. Choi M
    224. Dastani Z
    225. Drong AW
    226. Eriksson P
    227. Franco-Cereceda A
    228. Gådin JR
    229. Gharavi AG
    230. Goddard ME
    231. Handsaker RE
    232. Huang J
    233. Karpe F
    234. Kathiresan S
    235. Keildson S
    236. Kiryluk K
    237. Kubo M
    238. Lee J-Y
    239. Liang L
    240. Lifton RP
    241. Ma B
    242. McCarroll SA
    243. McKnight AJ
    244. Min JL
    245. Moffatt MF
    246. Montgomery GW
    247. Murabito JM
    248. Nicholson G
    249. Nyholt DR
    250. Okada Y
    251. Perry JRB
    252. Dorajoo R
    253. Reinmaa E
    254. Salem RM
    255. Sandholm N
    256. Scott RA
    257. Stolk L
    258. Takahashi A
    259. Tanaka T
    260. van ’t Hooft FM
    261. Vinkhuyzen AAE
    262. Westra H-J
    263. Zheng W
    264. Zondervan KT
    265. ADIPOGen Consortium
    266. AGEN-BMI Working Group
    267. CARDIOGRAMplusC4D Consortium
    268. CKDGen Consortium
    269. GLGC
    270. ICBP
    271. MAGIC Investigators
    272. MuTHER Consortium
    273. MIGen Consortium
    274. PAGE Consortium
    275. ReproGen Consortium
    276. GENIE Consortium
    277. International Endogene Consortium
    278. Heath AC
    279. Arveiler D
    280. Bakker SJL
    281. Beilby J
    282. Bergman RN
    283. Blangero J
    284. Bovet P
    285. Campbell H
    286. Caulfield MJ
    287. Cesana G
    288. Chakravarti A
    289. Chasman DI
    290. Chines PS
    291. Collins FS
    292. Crawford DC
    293. Cupples LA
    294. Cusi D
    295. Danesh J
    296. de Faire U
    297. den Ruijter HM
    298. Dominiczak AF
    299. Erbel R
    300. Erdmann J
    301. Eriksson JG
    302. Farrall M
    303. Felix SB
    304. Ferrannini E
    305. Ferrières J
    306. Ford I
    307. Forouhi NG
    308. Forrester T
    309. Franco OH
    310. Gansevoort RT
    311. Gejman PV
    312. Gieger C
    313. Gottesman O
    314. Gudnason V
    315. Gyllensten U
    316. Hall AS
    317. Harris TB
    318. Hattersley AT
    319. Hicks AA
    320. Hindorff LA
    321. Hingorani AD
    322. Hofman A
    323. Homuth G
    324. Hovingh GK
    325. Humphries SE
    326. Hunt SC
    327. Hyppönen E
    328. Illig T
    329. Jacobs KB
    330. Jarvelin M-R
    331. Jöckel K-H
    332. Johansen B
    333. Jousilahti P
    334. Jukema JW
    335. Jula AM
    336. Kaprio J
    337. Kastelein JJP
    338. Keinanen-Kiukaanniemi SM
    339. Kiemeney LA
    340. Knekt P
    341. Kooner JS
    342. Kooperberg C
    343. Kovacs P
    344. Kraja AT
    345. Kumari M
    346. Kuusisto J
    347. Lakka TA
    348. Langenberg C
    349. Marchand LL
    350. Lehtimäki T
    351. Lyssenko V
    352. Männistö S
    353. Marette A
    354. Matise TC
    355. McKenzie CA
    356. McKnight B
    357. Moll FL
    358. Morris AD
    359. Morris AP
    360. Murray JC
    361. Nelis M
    362. Ohlsson C
    363. Oldehinkel AJ
    364. Ong KK
    365. Madden PAF
    366. Pasterkamp G
    367. Peden JF
    368. Peters A
    369. Postma DS
    370. Pramstaller PP
    371. Price JF
    372. Qi L
    373. Raitakari OT
    374. Rankinen T
    375. Rao DC
    376. Rice TK
    377. Ridker PM
    378. Rioux JD
    379. Ritchie MD
    380. Rudan I
    381. Salomaa V
    382. Samani NJ
    383. Saramies J
    384. Sarzynski MA
    385. Schunkert H
    386. Schwarz PEH
    387. Sever P
    388. Shuldiner AR
    389. Sinisalo J
    390. Stolk RP
    391. Strauch K
    392. Tönjes A
    393. Trégouët D-A
    394. Tremblay A
    395. Tremoli E
    396. Virtamo J
    397. Vohl M-C
    398. Völker U
    399. Waeber G
    400. Willemsen G
    401. Witteman JC
    402. Zillikens MC
    403. Adair LS
    404. Amouyel P
    405. Asselbergs FW
    406. Assimes TL
    407. Bochud M
    408. Boehm BO
    409. Boerwinkle E
    410. Bornstein SR
    411. Bottinger EP
    412. Bouchard C
    413. Cauchi S
    414. Chambers JC
    415. Chanock SJ
    416. Cooper RS
    417. de Bakker PIW
    418. Dedoussis G
    419. Ferrucci L
    420. Franks PW
    421. Froguel P
    422. Groop LC
    423. Haiman CA
    424. Hamsten A
    425. Hui J
    426. Hunter DJ
    427. Hveem K
    428. Kaplan RC
    429. Kivimaki M
    430. Kuh D
    431. Laakso M
    432. Liu Y
    433. Martin NG
    434. März W
    435. Melbye M
    436. Metspalu A
    437. Moebus S
    438. Munroe PB
    439. Njølstad I
    440. Oostra BA
    441. Palmer CNA
    442. Pedersen NL
    443. Perola M
    444. Pérusse L
    445. Peters U
    446. Power C
    447. Quertermous T
    448. Rauramaa R
    449. Rivadeneira F
    450. Saaristo TE
    451. Saleheen D
    452. Sattar N
    453. Schadt EE
    454. Schlessinger D
    455. Slagboom PE
    456. Snieder H
    457. Spector TD
    458. Thorsteinsdottir U
    459. Stumvoll M
    460. Tuomilehto J
    461. Uitterlinden AG
    462. Uusitupa M
    463. van der Harst P
    464. Walker M
    465. Wallaschofski H
    466. Wareham NJ
    467. Watkins H
    468. Weir DR
    469. Wichmann H-E
    470. Wilson JF
    471. Zanen P
    472. Borecki IB
    473. Deloukas P
    474. Fox CS
    475. Heid IM
    476. O’Connell JR
    477. Strachan DP
    478. Stefansson K
    479. van Duijn CM
    480. Abecasis GR
    481. Franke L
    482. Frayling TM
    483. McCarthy MI
    484. Visscher PM
    485. Scherag A
    486. Willer CJ
    487. Boehnke M
    488. Mohlke KL
    489. Lindgren CM
    490. Beckmann JS
    491. Barroso I
    492. North KE
    493. Ingelsson E
    494. Hirschhorn JN
    495. Loos RJF
    496. Speliotes EK
    (2015) Genetic studies of body mass index yield new insights for obesity biology
    Nature 518:197–206.
    https://doi.org/10.1038/nature14177
  24. Website
    1. Penn Medicine BioBank
    (2022) Penn Medicine BioBank
    Internet. Accessed February 1, 2022.
    https://pmbb.med.upenn.edu/
    1. Vogelezang S
    2. Bradfield JP
    3. Ahluwalia TS
    4. Curtin JA
    5. Lakka TA
    6. Grarup N
    7. Scholz M
    8. van der Most PJ
    9. Monnereau C
    10. Stergiakouli E
    11. Heiskala A
    12. Horikoshi M
    13. Fedko IO
    14. Vilor-Tejedor N
    15. Cousminer DL
    16. Standl M
    17. Wang CA
    18. Viikari J
    19. Geller F
    20. Íñiguez C
    21. Pitkänen N
    22. Chesi A
    23. Bacelis J
    24. Yengo L
    25. Torrent M
    26. Ntalla I
    27. Helgeland Ø
    28. Selzam S
    29. Vonk JM
    30. Zafarmand MH
    31. Heude B
    32. Farooqi IS
    33. Alyass A
    34. Beaumont RN
    35. Have CT
    36. Rzehak P
    37. Bilbao JR
    38. Schnurr TM
    39. Barroso I
    40. Bønnelykke K
    41. Beilin LJ
    42. Carstensen L
    43. Charles MA
    44. Chawes B
    45. Clément K
    46. Closa-Monasterolo R
    47. Custovic A
    48. Eriksson JG
    49. Escribano J
    50. Groen-Blokhuis M
    51. Grote V
    52. Gruszfeld D
    53. Hakonarson H
    54. Hansen T
    55. Hattersley AT
    56. Hollensted M
    57. Hottenga JJ
    58. Hyppönen E
    59. Johansson S
    60. Joro R
    61. Kähönen M
    62. Karhunen V
    63. Kiess W
    64. Knight BA
    65. Koletzko B
    66. Kühnapfel A
    67. Landgraf K
    68. Langhendries JP
    69. Lehtimäki T
    70. Leinonen JT
    71. Li A
    72. Lindi V
    73. Lowry E
    74. Bustamante M
    75. Medina-Gomez C
    76. Melbye M
    77. Michaelsen KF
    78. Morgen CS
    79. Mori TA
    80. Nielsen TRH
    81. Niinikoski H
    82. Oldehinkel AJ
    83. Pahkala K
    84. Panoutsopoulou K
    85. Pedersen O
    86. Pennell CE
    87. Power C
    88. Reijneveld SA
    89. Rivadeneira F
    90. Simpson A
    91. Sly PD
    92. Stokholm J
    93. Teo KK
    94. Thiering E
    95. Timpson NJ
    96. Uitterlinden AG
    97. van Beijsterveldt CEM
    98. van Schaik BDC
    99. Vaudel M
    100. Verduci E
    101. Vinding RK
    102. Vogel M
    103. Zeggini E
    104. Sebert S
    105. Lind MV
    106. Brown CD
    107. Santa-Marina L
    108. Reischl E
    109. Frithioff-Bøjsøe C
    110. Meyre D
    111. Wheeler E
    112. Ong K
    113. Nohr EA
    114. Vrijkotte TGM
    115. Koppelman GH
    116. Plomin R
    117. Njølstad PR
    118. Dedoussis GD
    119. Froguel P
    120. Sørensen TIA
    121. Jacobsson B
    122. Freathy RM
    123. Zemel BS
    124. Raitakari O
    125. Vrijheid M
    126. Feenstra B
    127. Lyytikäinen LP
    128. Snieder H
    129. Kirsten H
    130. Holt PG
    131. Heinrich J
    132. Widén E
    133. Sunyer J
    134. Boomsma DI
    135. Järvelin MR
    136. Körner A
    137. Davey Smith G
    138. Holm JC
    139. Atalay M
    140. Murray C
    141. Bisgaard H
    142. McCarthy MI
    143. Early Growth Genetics Consortium
    144. Jaddoe VWV
    145. Grant SFA
    146. Felix JF
    (2020) Novel loci for childhood body mass index and shared heritability with adult cardiometabolic traits
    PLOS Genetics 16:e1008718.
    https://doi.org/10.1371/journal.pgen.1008718
    1. Winkler TW
    2. Justice AE
    3. Graff M
    4. Barata L
    5. Feitosa MF
    6. Chu S
    7. Czajkowski J
    8. Esko T
    9. Fall T
    10. Kilpeläinen TO
    11. Lu Y
    12. Mägi R
    13. Mihailov E
    14. Pers TH
    15. Rüeger S
    16. Teumer A
    17. Ehret GB
    18. Ferreira T
    19. Heard-Costa NL
    20. Karjalainen J
    21. Lagou V
    22. Mahajan A
    23. Neinast MD
    24. Prokopenko I
    25. Simino J
    26. Teslovich TM
    27. Jansen R
    28. Westra HJ
    29. White CC
    30. Absher D
    31. Ahluwalia TS
    32. Ahmad S
    33. Albrecht E
    34. Alves AC
    35. Bragg-Gresham JL
    36. de Craen AJM
    37. Bis JC
    38. Bonnefond A
    39. Boucher G
    40. Cadby G
    41. Cheng YC
    42. Chiang CWK
    43. Delgado G
    44. Demirkan A
    45. Dueker N
    46. Eklund N
    47. Eiriksdottir G
    48. Eriksson J
    49. Feenstra B
    50. Fischer K
    51. Frau F
    52. Galesloot TE
    53. Geller F
    54. Goel A
    55. Gorski M
    56. Grammer TB
    57. Gustafsson S
    58. Haitjema S
    59. Hottenga JJ
    60. Huffman JE
    61. Jackson AU
    62. Jacobs KB
    63. Johansson Å
    64. Kaakinen M
    65. Kleber ME
    66. Lahti J
    67. Mateo Leach I
    68. Lehne B
    69. Liu Y
    70. Lo KS
    71. Lorentzon M
    72. Luan J
    73. Madden PAF
    74. Mangino M
    75. McKnight B
    76. Medina-Gomez C
    77. Monda KL
    78. Montasser ME
    79. Müller G
    80. Müller-Nurasyid M
    81. Nolte IM
    82. Panoutsopoulou K
    83. Pascoe L
    84. Paternoster L
    85. Rayner NW
    86. Renström F
    87. Rizzi F
    88. Rose LM
    89. Ryan KA
    90. Salo P
    91. Sanna S
    92. Scharnagl H
    93. Shi J
    94. Smith AV
    95. Southam L
    96. Stančáková A
    97. Steinthorsdottir V
    98. Strawbridge RJ
    99. Sung YJ
    100. Tachmazidou I
    101. Tanaka T
    102. Thorleifsson G
    103. Trompet S
    104. Pervjakova N
    105. Tyrer JP
    106. Vandenput L
    107. van der Laan SW
    108. van der Velde N
    109. van Setten J
    110. van Vliet-Ostaptchouk JV
    111. Verweij N
    112. Vlachopoulou E
    113. Waite LL
    114. Wang SR
    115. Wang Z
    116. Wild SH
    117. Willenborg C
    118. Wilson JF
    119. Wong A
    120. Yang J
    121. Yengo L
    122. Yerges-Armstrong LM
    123. Yu L
    124. Zhang W
    125. Zhao JH
    126. Andersson EA
    127. Bakker SJL
    128. Baldassarre D
    129. Banasik K
    130. Barcella M
    131. Barlassina C
    132. Bellis C
    133. Benaglio P
    134. Blangero J
    135. Blüher M
    136. Bonnet F
    137. Bonnycastle LL
    138. Boyd HA
    139. Bruinenberg M
    140. Buchman AS
    141. Campbell H
    142. Chen YDI
    143. Chines PS
    144. Claudi-Boehm S
    145. Cole J
    146. Collins FS
    147. de Geus EJC
    148. de Groot L
    149. Dimitriou M
    150. Duan J
    151. Enroth S
    152. Eury E
    153. Farmaki AE
    154. Forouhi NG
    155. Friedrich N
    156. Gejman PV
    157. Gigante B
    158. Glorioso N
    159. Go AS
    160. Gottesman O
    161. Gräßler J
    162. Grallert H
    163. Grarup N
    164. Gu YM
    165. Broer L
    166. Ham AC
    167. Hansen T
    168. Harris TB
    169. Hartman CA
    170. Hassinen M
    171. Hastie N
    172. Hattersley AT
    173. Heath AC
    174. Henders AK
    175. Hernandez D
    176. Hillege H
    177. Holmen O
    178. Hovingh KG
    179. Hui J
    180. Husemoen LL
    181. Hutri-Kähönen N
    182. Hysi PG
    183. Illig T
    184. De Jager PL
    185. Jalilzadeh S
    186. Jørgensen T
    187. Jukema JW
    188. Juonala M
    189. Kanoni S
    190. Karaleftheri M
    191. Khaw KT
    192. Kinnunen L
    193. Kittner SJ
    194. Koenig W
    195. Kolcic I
    196. Kovacs P
    197. Krarup NT
    198. Kratzer W
    199. Krüger J
    200. Kuh D
    201. Kumari M
    202. Kyriakou T
    203. Langenberg C
    204. Lannfelt L
    205. Lanzani C
    206. Lotay V
    207. Launer LJ
    208. Leander K
    209. Lindström J
    210. Linneberg A
    211. Liu YP
    212. Lobbens S
    213. Luben R
    214. Lyssenko V
    215. Männistö S
    216. Magnusson PK
    217. McArdle WL
    218. Menni C
    219. Merger S
    220. Milani L
    221. Montgomery GW
    222. Morris AP
    223. Narisu N
    224. Nelis M
    225. Ong KK
    226. Palotie A
    227. Pérusse L
    228. Pichler I
    229. Pilia MG
    230. Pouta A
    231. Rheinberger M
    232. Ribel-Madsen R
    233. Richards M
    234. Rice KM
    235. Rice TK
    236. Rivolta C
    237. Salomaa V
    238. Sanders AR
    239. Sarzynski MA
    240. Scholtens S
    241. Scott RA
    242. Scott WR
    243. Sebert S
    244. Sengupta S
    245. Sennblad B
    246. Seufferlein T
    247. Silveira A
    248. Slagboom PE
    249. Smit JH
    250. Sparsø TH
    251. Stirrups K
    252. Stolk RP
    253. Stringham HM
    254. Swertz MA
    255. Swift AJ
    256. Syvänen AC
    257. Tan ST
    258. Thorand B
    259. Tönjes A
    260. Tremblay A
    261. Tsafantakis E
    262. van der Most PJ
    263. Völker U
    264. Vohl MC
    265. Vonk JM
    266. Waldenberger M
    267. Walker RW
    268. Wennauer R
    269. Widén E
    270. Willemsen G
    271. Wilsgaard T
    272. Wright AF
    273. Zillikens MC
    274. van Dijk SC
    275. van Schoor NM
    276. Asselbergs FW
    277. de Bakker PIW
    278. Beckmann JS
    279. Beilby J
    280. Bennett DA
    281. Bergman RN
    282. Bergmann S
    283. Böger CA
    284. Boehm BO
    285. Boerwinkle E
    286. Boomsma DI
    287. Bornstein SR
    288. Bottinger EP
    289. Bouchard C
    290. Chambers JC
    291. Chanock SJ
    292. Chasman DI
    293. Cucca F
    294. Cusi D
    295. Dedoussis G
    296. Erdmann J
    297. Eriksson JG
    298. Evans DA
    299. de Faire U
    300. Farrall M
    301. Ferrucci L
    302. Ford I
    303. Franke L
    304. Franks PW
    305. Froguel P
    306. Gansevoort RT
    307. Gieger C
    308. Grönberg H
    309. Gudnason V
    310. Gyllensten U
    311. Hall P
    312. Hamsten A
    313. van der Harst P
    314. Hayward C
    315. Heliövaara M
    316. Hengstenberg C
    317. Hicks AA
    318. Hingorani A
    319. Hofman A
    320. Hu F
    321. Huikuri HV
    322. Hveem K
    323. James AL
    324. Jordan JM
    325. Jula A
    326. Kähönen M
    327. Kajantie E
    328. Kathiresan S
    329. Kiemeney L
    330. Kivimaki M
    331. Knekt PB
    332. Koistinen HA
    333. Kooner JS
    334. Koskinen S
    335. Kuusisto J
    336. Maerz W
    337. Martin NG
    338. Laakso M
    339. Lakka TA
    340. Lehtimäki T
    341. Lettre G
    342. Levinson DF
    343. Lind L
    344. Lokki ML
    345. Mäntyselkä P
    346. Melbye M
    347. Metspalu A
    348. Mitchell BD
    349. Moll FL
    350. Murray JC
    351. Musk AW
    352. Nieminen MS
    353. Njølstad I
    354. Ohlsson C
    355. Oldehinkel AJ
    356. Oostra BA
    357. Palmer LJ
    358. Pankow JS
    359. Pasterkamp G
    360. Pedersen NL
    361. Pedersen O
    362. Penninx BW
    363. Perola M
    364. Peters A
    365. Polašek O
    366. Pramstaller PP
    367. Psaty BM
    368. Qi L
    369. Quertermous T
    370. Raitakari OT
    371. Rankinen T
    372. Rauramaa R
    373. Ridker PM
    374. Rioux JD
    375. Rivadeneira F
    376. Rotter JI
    377. Rudan I
    378. den Ruijter HM
    379. Saltevo J
    380. Sattar N
    381. Schunkert H
    382. Schwarz PEH
    383. Shuldiner AR
    384. Sinisalo J
    385. Snieder H
    386. Sørensen TIA
    387. Spector TD
    388. Staessen JA
    389. Stefania B
    390. Thorsteinsdottir U
    391. Stumvoll M
    392. Tardif JC
    393. Tremoli E
    394. Tuomilehto J
    395. Uitterlinden AG
    396. Uusitupa M
    397. Verbeek ALM
    398. Vermeulen SH
    399. Viikari JS
    400. Vitart V
    401. Völzke H
    402. Vollenweider P
    403. Waeber G
    404. Walker M
    405. Wallaschofski H
    406. Wareham NJ
    407. Watkins H
    408. Zeggini E
    409. CHARGE Consortium
    410. DIAGRAM Consortium
    411. GLGC Consortium
    412. Global-BPGen Consortium
    413. ICBP Consortium
    414. MAGIC Consortium
    415. Chakravarti A
    416. Clegg DJ
    417. Cupples LA
    418. Gordon-Larsen P
    419. Jaquish CE
    420. Rao DC
    421. Abecasis GR
    422. Assimes TL
    423. Barroso I
    424. Berndt SI
    425. Boehnke M
    426. Deloukas P
    427. Fox CS
    428. Groop LC
    429. Hunter DJ
    430. Ingelsson E
    431. Kaplan RC
    432. McCarthy MI
    433. Mohlke KL
    434. O’Connell JR
    435. Schlessinger D
    436. Strachan DP
    437. Stefansson K
    438. van Duijn CM
    439. Hirschhorn JN
    440. Lindgren CM
    441. Heid IM
    442. North KE
    443. Borecki IB
    444. Kutalik Z
    445. Loos RJF
    (2015) The influence of age and sex on genetic associations with adult body size and shape: a large-scale genome-wide interaction study
    PLOS Genetics 11:e1005378.
    https://doi.org/10.1371/journal.pgen.1005378

Article and author information

Author details

  1. Daniel Hui

    Department of Genetics, Perelman School of Medicine, University of Pennsylvania, Philadelphia, United States
    Contribution
    Conceptualization, Formal analysis, Visualization, Methodology, Writing – original draft, Writing – review and editing
    Competing interests
    No competing interests declared
    ORCID icon "This ORCID iD identifies the author of this article:" 0000-0002-8023-7352

  2. Scott Dudek

    Department of Genetics, Perelman School of Medicine, University of Pennsylvania, Philadelphia, United States
    Contribution
    Formal analysis, Writing – review and editing
    Competing interests
    No competing interests declared

  3. Krzysztof Kiryluk

    Division of Nephrology, Department of Medicine, Columbia University, New York, United States
    Contribution
    Writing – review and editing
    Competing interests
    No competing interests declared

  4. Theresa L Walunas

    Department of Preventive Medicine, Northwestern University Feinberg School of Medicine, Chicago, United States
    Contribution
    Writing – review and editing
    Competing interests
    No competing interests declared

  5. Iftikhar J Kullo

    Department of Cardiovascular Medicine, Mayo Clinic, Rochester, United States
    Contribution
    Writing – review and editing
    Competing interests
    No competing interests declared

  6. Wei-Qi Wei

    Department of Biomedical Informatics, Vanderbilt University Medical Center, Nashville, United States
    Contribution
    Writing – review and editing
    Competing interests
    No competing interests declared

  7. Hemant Tiwari

    Department of Pediatrics, University of Alabama at Birmingham, Birmingham, United States
    Contribution
    Writing – review and editing
    Competing interests
    No competing interests declared

  8. Josh F Peterson

    Department of Biomedical Informatics, Vanderbilt University Medical Center, Nashville, United States
    Contribution
    Writing – review and editing
    Competing interests
    No competing interests declared

  9. Wendy K Chung

    Departments of Pediatrics and Medicine, Columbia University Irving Medical Center, Columbia University, New York, United States
    Contribution
    Writing – review and editing
    Competing interests
    No competing interests declared

  10. Brittney H Davis

    Department of Neurology, School of Medicine, University of Alabama at Birmingham, Birmingham, United States
    Contribution
    Writing – review and editing
    Competing interests
    No competing interests declared

  11. Atlas Khan

    Division of Nephrology, Department of Medicine, Columbia University, New York, United States
    Contribution
    Writing – review and editing
    Competing interests
    No competing interests declared

  12. Leah C Kottyan

    The Center for Autoimmune Genomics and Etiology, Division of Human Genetics, Cincinnati Children's Hospital Medical Center, Cincinnati, United States
    Contribution
    Writing – review and editing
    Competing interests
    No competing interests declared

  13. Nita A Limdi

    Department of Neurology, School of Medicine, University of Alabama at Birmingham, Birmingham, United States
    Contribution
    Writing – review and editing
    Competing interests
    No competing interests declared

  14. Qiping Feng

    Division of Clinical Pharmacology, Department of Medicine, Vanderbilt University Medical Center, Nashville, United States
    Contribution
    Writing – review and editing
    Competing interests
    No competing interests declared
    ORCID icon "This ORCID iD identifies the author of this article:" 0000-0002-6213-793X

  15. Megan J Puckelwartz

    Center for Genetic Medicine, Northwestern University Feinberg School of Medicine, Chicago, United States
    Contribution
    Writing – review and editing
    Competing interests
    No competing interests declared

  16. Chunhua Weng

    Department of Biomedical Informatics, Vagelos College of Physicians & Surgeons, Columbia University, New York, United States
    Contribution
    Writing – review and editing
    Competing interests
    No competing interests declared

  17. Johanna L Smith

    Department of Cardiovascular Medicine, Mayo Clinic, Rochester, United States
    Contribution
    Writing – review and editing
    Competing interests
    No competing interests declared

  18. Elizabeth W Karlson

    Division of Rheumatology, Inflammation, and Immunity, Department of Medicine, Brigham and Women's Hospital and Harvard Medical School, Boston, United States
    Contribution
    Writing – review and editing
    Competing interests
    No competing interests declared

  19. Regeneron Genetics Center

    Contribution
    DNA sequencing
    Competing interests
    No competing interests declared

  20. Penn Medicine BioBank

    Contribution
    Recruitment of participants, extraction of de-identified EHR data
    Competing interests
    No competing interests declared

  21. Gail P Jarvik

    Departments of Medicine (Medical Genetics) and Genome Sciences, University of Washington Medical Center, Seattle, United States
    Contribution
    Writing – review and editing
    Competing interests
    No competing interests declared

  22. Marylyn D Ritchie

    Department of Genetics, Perelman School of Medicine, University of Pennsylvania, Philadelphia, United States
    Contribution
    Conceptualization, Resources, Supervision, Funding acquisition, Investigation, Methodology, Writing – original draft, Project administration, Writing – review and editing
    For correspondence
    marylyn@pennmedicine.upenn.edu
    Competing interests
    No competing interests declared
    ORCID icon "This ORCID iD identifies the author of this article:" 0000-0002-1208-1720

Funding

National Institutes of Health (AI077505)

  • Daniel Hui
  • Scott Dudek
  • Marylyn D Ritchie

National Institutes of Health (HL169458)

  • Daniel Hui
  • Scott Dudek
  • Marylyn D Ritchie

National Institute of Diabetes and Digestive and Kidney Diseases (DK52431)

  • Wendy K Chung

National Institutes of Health (U01 HG011166)

  • Josh F Peterson

National Institutes of Health (U01 HG008680)

  • Chunhua Weng

Group Health Cooperative/University of Washington (U01HG008657)

  • Gail P Jarvik

Brigham and Women’s Hospital (U01HG008685)

  • Elizabeth W Karlson

Vanderbilt University Medical Center (U01HG008672)

  • Wei-Qi Wei
  • Josh F Peterson
  • Qiping Feng

Cincinnati Children’s Hospital Medical Center (U01HG008666)

  • Leah C Kottyan

Mayo Clinic (U01HG006379)

  • Iftikhar J Kullo
  • Johanna L Smith

Columbia University Health Sciences (U01HG008680)

  • Krzysztof Kiryluk
  • Atlas Khan
  • Chunhua Weng
  • Wendy K Chung

Northwestern University (U01HG008673)

  • Theresa L Walunas
  • Megan J Puckelwartz

Vanderbilt University Medical Center serving as the Coordinating Center (U01HG008701)

  • Josh F Peterson
  • Wei-Qi Wei

The funders had no role in study design, data collection and interpretation, or the decision to submit the work for publication.

Acknowledgements

For UK Biobank: This research has been conducted using the UK Biobank Resource under Application Number 32133. This work uses data provided by patients and collected by the NHS as part of their care and support.

For GERA: Data came from a grant, the Resource for Genetic Epidemiology Research in Adult Health and Aging (RC2 AG033067; Schaefer and Risch, PIs) awarded to the Kaiser Permanente Research Program on Genes, Environment, and Health (RPGEH) and the UCSF Institute for Human Genetics. The RPGEH was supported by grants from the Robert Wood Johnson Foundation, the Wayne and Gladys Valley Foundation, the Ellison Medical Foundation, Kaiser Permanente Northern California, and the Kaiser Permanente National and Northern California Community Benefit Programs. The RPGEH and the Resource for Genetic Epidemiology Research in Adult Health and Aging are described in the following publication, Schaefer C, et al., The Kaiser Permanente Research Program on Genes, Environment and Health: Development of a Research Resource in a Multi-Ethnic Health Plan with Electronic Medical Records, In preparation, 2013.

For eMERGE: We acknowledge David Crosslin for helping clean the eMERGE data. Please see funding section for eMERGE funding acknowledgments.

For PMBB: We acknowledge the Penn Medicine BioBank (PMBB) for providing data and thank the patient- participants of Penn Medicine who consented to participate in this research program. We would also like to thank the Penn Medicine BioBank team and Regeneron Genetics Center for providing genetic variant data for analysis. The PMBB is approved under IRB protocol# 813913 and supported by Perelman School of Medicine at University of Pennsylvania, a gift from the Smilow family, and the National Center for Advancing Translational Sciences of the National Institutes of Health under CTSA award number UL1TR001878.

Version history

  1. Sent for peer review:
  2. Preprint posted:
  3. Reviewed Preprint version 1:
  4. Reviewed Preprint version 2:
  5. Version of Record published:

Cite all versions

You can cite all versions using the DOI https://doi.org/10.7554/eLife.88149. This DOI represents all versions, and will always resolve to the latest one.

Copyright

© 2023, Hui et al.

This article is distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use and redistribution provided that the original author and source are credited.

Metrics

  • 570
    views
  • 14
    downloads
  • 0
    citations

Views, downloads and citations are aggregated across all versions of this paper published by eLife.

Download links

A two-part list of links to download the article, or parts of the article, in various formats.

Downloads (link to download the article as PDF)

Open citations (links to open the citations from this article in various online reference manager services)

Cite this article (links to download the citations from this article in formats compatible with various reference manager tools)

  1. Daniel Hui
  2. Scott Dudek
  3. Krzysztof Kiryluk
  4. Theresa L Walunas
  5. Iftikhar J Kullo
  6. Wei-Qi Wei
  7. Hemant Tiwari
  8. Josh F Peterson
  9. Wendy K Chung
  10. Brittney H Davis
  11. Atlas Khan
  12. Leah C Kottyan
  13. Nita A Limdi
  14. Qiping Feng
  15. Megan J Puckelwartz
  16. Chunhua Weng
  17. Johanna L Smith
  18. Elizabeth W Karlson
  19. Regeneron Genetics Center
  20. Penn Medicine BioBank
  21. Gail P Jarvik
  22. Marylyn D Ritchie
(2025)
Risk factors affecting polygenic score performance across diverse cohorts
eLife 12:RP88149.
https://doi.org/10.7554/eLife.88149.3

Share this article

https://doi.org/10.7554/eLife.88149

Categories and tags

Research organism

Further reading

    1. Genetics and Genomics
    2. Computational and Systems Biology

    Special Issue: Systems Genetics

    Edited by David James et al.
    Collection

    In this Special Issue we present a range of studies that showcase novel approaches that researchers are exploring to better decipher the link between genotype and phenotype.

    1. Computational and Systems Biology
    2. Neuroscience

    Multisensory integration operates on correlated input from unimodal transient channels

    Cesare V Parise, Marc O Ernst
    Research Article

    Audiovisual information reaches the brain via both sustained and transient input channels, representing signals’ intensity over time or changes thereof, respectively. To date, it is unclear to what extent transient and sustained input channels contribute to the combined percept obtained through multisensory integration. Based on the results of two novel psychophysical experiments, here we demonstrate the importance of the transient (instead of the sustained) channel for the integration of audiovisual signals. To account for the present results, we developed a biologically inspired, general-purpose model for multisensory integration, the multisensory correlation detectors, which combines correlated input from unimodal transient channels. Besides accounting for the results of our psychophysical experiments, this model could quantitatively replicate several recent findings in multisensory research, as tested against a large collection of published datasets. In particular, the model could simultaneously account for the perceived timing of audiovisual events, multisensory facilitation in detection tasks, causality judgments, and optimal integration. This study demonstrates that several phenomena in multisensory research that were previously considered unrelated, all stem from the integration of correlated input from unimodal transient channels.

    1. Cell Biology
    2. Computational and Systems Biology

    Unbiased identification of cell identity in dense mixed neural cultures

    Sarah De Beuckeleer, Tim Van De Looverbosch ... Winnok H De Vos
    Research Article

    Induced pluripotent stem cell (iPSC) technology is revolutionizing cell biology. However, the variability between individual iPSC lines and the lack of efficient technology to comprehensively characterize iPSC-derived cell types hinder its adoption in routine preclinical screening settings. To facilitate the validation of iPSC-derived cell culture composition, we have implemented an imaging assay based on cell painting and convolutional neural networks to recognize cell types in dense and mixed cultures with high fidelity. We have benchmarked our approach using pure and mixed cultures of neuroblastoma and astrocytoma cell lines and attained a classification accuracy above 96%. Through iterative data erosion, we found that inputs containing the nuclear region of interest and its close environment, allow achieving equally high classification accuracy as inputs containing the whole cell for semi-confluent cultures and preserved prediction accuracy even in very dense cultures. We then applied this regionally restricted cell profiling approach to evaluate the differentiation status of iPSC-derived neural cultures, by determining the ratio of postmitotic neurons and neural progenitors. We found that the cell-based prediction significantly outperformed an approach in which the population-level time in culture was used as a classification criterion (96% vs 86%, respectively). In mixed iPSC-derived neuronal cultures, microglia could be unequivocally discriminated from neurons, regardless of their reactivity state, and a tiered strategy allowed for further distinguishing activated from non-activated cell states, albeit with lower accuracy. Thus, morphological single-cell profiling provides a means to quantify cell composition in complex mixed neural cultures and holds promise for use in the quality control of iPSC-derived cell culture models.