Widespread and Biologically Driven Sex Disparities in Polygenic Risk Prediction Across Complex Traits

Widespread and Biologically Driven Sex Disparities in Polygenic Risk Prediction Across Complex Traits | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Article Widespread and Biologically Driven Sex Disparities in Polygenic Risk Prediction Across Complex Traits Akl Fahed, Xingyu Chen, Tingfeng Xu, Yang Sui, Kelvin Supriami, and 24 more This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-9153854/v1 This work is licensed under a CC BY 4.0 License Status: Under Review Version 1 posted You are reading this latest preprint version Abstract Polygenic risk scores (PRS) are increasingly used for disease prediction, yet their performance equity across sexes remains unclear. We evaluated sex differences in PRS performance using 3,263 scores across 145 traits from the PGS Catalog in 409,440 UK Biobank participants. Sex-differential prediction was widespread and trait-specific, affecting 15 of 64 (23%) of diseases and 43 of 81 (53%) of quantitative traits. Female-favoring performance was enriched in autoimmune and endocrine traits, whereas cardiometabolic traits more often favored males. Discovery GWAS sex imbalance partially explained disease-level disparities (R² = 0.36), whereas quantitative traits showed minimal association. Notably, sex differences in predictive performance strongly correlated with differences in SNP-based heritability from sex-stratified GWAS (R² = 0.81 for diseases; R² = 0.58 for quantitative traits). In contrast, PRS estimates were highly consistent across seven construction methods (intraclass correlation coefficient = 0.93), indicating limited methodological influence. These findings demonstrate that sex disparities in PRS performance are common and largely reflect underlying genetic architecture rather than analytic artifacts, highlighting the need for sex-aware GWAS design and PRS modeling to ensure equitable clinical implementation. Biological sciences/Genetics/Population genetics Biological sciences/Genetics/Genetic association study/Genome-wide association studies Biological sciences/Genetics/Genomics Figures Figure 1 Figure 2 Figure 3 Figure 4 Figure 5 Main Polygenic risk scores (PRS), which aggregate the effects of thousands of genetic variants identified through genome-wide association studies (GWAS), have emerged as powerful tools for quantifying individual genetic predisposition to complex traits and diseases. 1 By weighing risk alleles through effect size estimates from large-scale GWAS, PRS have demonstrated clinical potential by enabling disease prediction and risk stratification. 2-4 However, the generalizability of PRS remains constrained across ancestral populations, with significantly reduced predictive accuracy in non-European ancestries, which are often underrepresented in genetic studies. 5,6 Beyond ancestral disparities in representation, pronounced sex differences in disease prevalence persist across numerous traits. Coronary artery disease (CAD) exemplifies this divergence. Male sex is a risk factor for CAD, and females have delayed disease onset by 7-10 years compared to males 7 , yet females experience higher mortality rates post-myocardial infarction. 8 Similarly, autoimmune diseases like rheumatoid arthritis show 3:1 female-to-male prevalence 9 , while neuropsychiatric disorders such as depression demonstrate sex-specific symptom profiles. 10 These epidemiological disparities arise from multifaceted biological mechanisms, including genotype by sex (GxS) interactions 11 , X-chromosome dosage effects 12 , sex hormone-mediated gene regulation 13 , and sex-related environmental exposures 14 . Despite these established sex differences, current PRS frameworks often overlook sex-specific genetic architectures. Most GWAS meta-analyses either inadequately stratify by sex or assume linear additive effects across sex, 15-17 and PRS construction methods lack modelling of sex-specific effect sizes between sexes, potentially biasing predictions. For instance, prior analyses of CAD using a multiancestry polygenic risk score (GPS Mult ) from our group 18 revealed substantial sex-based prediction disparities. This imbalance likely stems from both biological factors, including underrepresentation of sex-specific loci in GWAS, and methodological limitations, including pervasive sample ascertainment bias 19 and insufficient modeling of gene-by-sex interactions 20 . As polygenic scores move towards clinical implementation, the absence of systematic evaluation of sex bias poses a critical barrier to equitable use 21,22 . A sex-blind approach may exacerbate disparities, yet no study has comprehensively quantified these differences or clarified their methodological and biological sources. Filling this gap is critical for allowing fair and effective clinical translation of PRS. To address these gaps, we used more than three thousand PRS from the PGS Catalog in conjunction with UK Biobank data to broadly assess the prevalence and magnitude of sex-dependent predictive disparities. We then leveraged GWAS data from the UK Biobank training dataset and employed seven PRS methods to investigate how GWAS sample size differences, methodological biases, and biological heterogeneity at sex-specific loci drive these disparities. Method Study Population and Outcome Definitions We utilized data from the UK Biobank 23 under application IDs 7089 and 89885, which granted access to anonymized genetic and clinical data. Standard UKB quality control procedures were applied. Specifically, individuals exhibiting discrepancies between self-reported sex (UK Biobank Field 31) and genetically inferred sex (Field 22001) were excluded. Additionally, participants were removed if they presented an individual-level genotype missingness rate exceeding 5% (Field 22005), were outliers in heterozygosity or missing rate (Field 22027), or exhibited sex chromosome aneuploidy (Field 22019). After applying these QC steps and restricting analyses to genetically unrelated individuals of European ancestry (kinship coefficient < 0.0884) 24 , a total of 409,440 participants remained for further analysis. The cohort was split into independent training (50%), tuning (25%), and testing (25%) sets to enable unbiased model development, parameter optimization, and final evaluation of PRS performance. Disease outcomes were defined using a combination of self-report and linked health records (Supplementary Table 1), and quantitative traits were centrally curated by the UKB. Curation of Polygenic Risk Scores for Evaluation To systematically investigate whether PRS exhibit widespread sex differences in their predictive performance, we used the PGS Catalog to assess sex bias in PRS predictive performance. 25 A total of 5,008 PRS (as of October 2024) were retained for subsequent analysis. We harmonized SNP identifiers between PGS Catalog weights and UK Biobank genotypes using standard coordinate and allele matching; scores with <80% variant overlap were excluded. For PRS with variant matching rates more than 0.8, we utilized all matched variants to calculate the PRS scores. This standardized procedure was uniformly applied across all PGS Catalog entries, ensuring consistency in our analysis involving over 5,000 publicly available PRSs. From the 692 phenotypes in the PGS Catalog, we derived a curated set by applying exclusion criteria to ensure alignment with UK Biobank phenotypes and minimize conceptual or ethical ambiguities. We excluded traits represented by three or fewer PRS, those lacking direct UK Biobank counterparts, and descriptors that were broad or imprecise. Additional exclusions included disease subtypes nested within broader diagnoses, family-history and virology or serology traits, imaging-derived or procedural measures, isolated symptoms or behavioral items, narrowly defined dietary exposures, sex-specific conditions such as prostate or ovarian cancer, lifestyle and socioeconomic factors, and traits potentially raising ethical concerns in the context of sex differences. This yielded 145 phenotypes (64 diseases and 81 quantitative traits) for downstream PGS Catalog analyses (Supplementary Table 1). For each retained weight file, polygenic scores were generated in PLINK2(v2.0.0-a.6.5LM) 26 by multiplying the dosage of every risk allele by its assigned effect size and aggregating these weighted values across all variants for each individual in the UK Biobank. Evaluation of Sex Differences in Published Polygenic Risk Scores After obtaining a comprehensive list of published PRS from the PGS Catalog, we evaluated their sex-specific predictive performance at both the score and trait levels. For each individual score, sex-stratified regression models were fitted in the complete UK Biobank data, and differences between males and females were assessed using a Z test. Scores were classified as female-biased, male-biased, or unbiased based on the direction and significance of effect differences. At the trait level, all scores for a phenotype were jointly modeled in the tuning set to derive a weighted ensemble score for phenotype prediction (ensemble trait-level PRS), which was then applied to the testing set for sex-specific evaluation. Association of Discovery Sample Size with Sex-Differential PRS Performance To assess whether discovery cohort composition contributed to sex differences in PRS performance, GWAS sample sizes for each PGS Catalog score were obtained from the REST application programming interface (API) 25 , which supports batch access to curated score metadata, and were complemented by manual curation from primary publications when API fields were unavailable. Sex-specific counts were derived from reported male proportions or case-control sample size, expressed as effective sample sizes (as twice the harmonic mean of the case and control populations) for binary traits and total sample sizes for quantitative traits. Female-to-male ratios of discovery sample size were then compared with corresponding PRS performance ratios using regression analyses. If there are multiple scores for a trait, the sample size and the PRS performance were averaged. Outliers (Cook’s distance > 4/n or studentized residual > 3) were excluded from model fitting but retained in plots. Sex Stratified and Sex Agnostic GWAS Analysis To generate GWAS summary statistics for downstream dissection of sex differences in polygenic risk scores, we conducted both sex-agnostic and sex-stratified genome-wide association studies (GWAS) using imputed genotype data from the UK Biobank. Analyses were performed with REGENIE (v3.2.9) 27 following its standard two-step procedure. Step 1 fitted whole-genome regression models under a leave-one-chromosome-out (LOCO) scheme using high-quality genotyped variants: minor allele frequency (MAF) > 1%, minor allele count (MAC) > 100, genotyping call rate > 99%, Hardy–Weinberg equilibrium P > 1×10⁻¹⁵, missingness < 10%, and linkage disequilibrium (LD) pruning using 1,000-variant windows, 100-variant sliding windows, and an r² threshold < 0.8. In the second step, single-variant association tests were performed using linear regression with LOCO-based predictions included as offsets. Sex-agnostic GWAS adjusted for age, age², sex, assessment center, genotyping array, and 10 ancestry PCs; sex-stratified GWAS applied the same covariates, excluding sex. Assessment of Methodological Differences in PRS Construction To evaluate whether the observed sex differences in PRS were influenced by the algorithms used to derive PRS, we analyzed 58 sex-biased traits using a sex-agnostic GWAS in the UK Biobank training set, using the same training, tuning, and testing partitions as in the primary analyses. PRS were then constructed using seven widely adopted algorithms: clumping and thresholding (P+T), LDpred2, Lassosum2, SBayesR, SBLUP, SDPR, and PRS-CS 26,28-36 . Method parameters are provided in Supplementary Table 6. Linkage disequilibrium (LD) reference panels were derived from 503 unrelated European individuals from the 1000 Genomes Project Phase 3 dataset 37 (MAF > 1%), ensuring consistency across all methods. Only HapMap3 variants were included in PRS construction 38 . For each trait and method, the parameters of all candidate PRS were tuned in the tuning subset and evaluated in the testing dataset. Consistency of sex-bias estimates across methods was assessed by intraclass correlation ( ICC s) with 1,000-bootstrap 95% CIs using the pingouin package (v0.5.4) 39 in Python (v3.10). Heritability Estimation SNP-based heritability (h 2 ) was estimated using linkage disequilibrium score regression (LDSC) 40 applied to GWAS summary statistics, and calculated as where, and are the estimates of the additive genetic and residual variance. Analyses were restricted to well-imputed HapMap3 variants. Chi-square statistics were regressed on LD scores computed from the 1000 Genomes European reference panel. Heritability estimates for binary traits were transformed from the observed scale to the liability scale, using the sex-specific population prevalence of the trait, under the assumption of an underlying normal distribution of liability to the considered trait, as described previously. 41 Statistical Analysis All analyses were conducted under a prespecified, uniform analytic framework to ensure consistency across PGS Catalog, ensemble, and de novo PRS analyses. For binary outcomes, logistic regression models were fitted with standardized PRSs as predictors. For quantitative traits, linear regression was applied. Covariates included age, sex, assessment center, genotyping array, and the first 10 principal components of ancestry, with sex excluded in sex-stratified analyses. To minimize confounding from population structure, all PRSs were residualized by the first 10 genetic principal components before association testing. Sex differences in predictive performance were quantified using a two-sided Z test comparing sex-specific regression coefficients and their standard errors: This test evaluates the null hypothesis that the predictive effects of PRS are equal between sexes (H₀: β female = β male ) against the alternative hypothesis that they differ (H₁: β female ≠ β male ). This procedure was applied consistently across all analytic stages, including PGS Catalog scores, ensemble trait-level PRSs, and UK Biobank–derived PRSs constructed from multiple algorithms. For secondary comparisons of genetic parameters between male- and female-biased traits, we used the Mann–Whitney U test to evaluate whether the median difference between groups differed significantly from zero ( H ₀ : Δ = 0; H ₁ : Δ ≠ 0). Sex bias in predictive performance was represented as the difference in these metrics between females and males. Multiple testing correction was performed using the Benjamini–Hochberg false discovery rate (FDR) procedure, with FDR -adjusted p < 0.05 considered statistically significant. All statistical tests were two-sided. Regression modeling and data visualization were performed in R version 4.1.2 (R Foundation for Statistical Computing) with the bigsnpr 32 , ggplot2 42 , and data.table packages. Analyses reporting include point estimates and 95% confidence intervals for all primary effect size comparisons. Results Widespread Sex Bias in Polygenic Risk Score Performance Using PRS from the PGS Catalog applied to the UK Biobank, we observed widespread and trait-specific sex differences in polygenic risk prediction across both diseases and quantitative traits. Figure 1 summarizes the distribution of sex bias across scores (left) and traits (right). Among 1295 disease scores, 41.2% showed significant sex bias (193 female-biased; 341 male-biased). Among 1968 scores for quantitative traits, 73.2% were biased (728 female-biased and 713 male-biased). Aggregated at the trait level, 15 (23.4%) of 64 disease traits and 43 (53.1%) of 81 quantitative traits displayed significant sex differences in performance, respectively. To further characterize the magnitude and direction of these effects, we extended the trait-level summary to display effect size estimates and 95% confidence intervals for each trait (Figure 2). This analysis revealed marked heterogeneity across traits, with female-favoring prediction observed for autoimmune and endocrine diseases such as hypothyroidism and type 1 diabetes, whereas male-favoring prediction was evident for cardiovascular diseases, including coronary artery disease and atrial fibrillation. Other traits, such as inguinal hernia, are also male-biased. Among quantitative traits, hormone- and body composition–related measures such as estradiol, testosterone, and body fat showed the strongest sex divergence, typically favoring the biologically corresponding sex, while most hematologic and metabolic biomarkers exhibited smaller but consistent male advantage. Overall, these findings demonstrated that sex differences in PRS performance are both pervasive and trait-specific, affecting a wide spectrum of diseases and quantitative phenotypes. Full per-score and per-trait statistics, including confidence intervals and significance levels, are provided in Supplementary Table 3 and Supplementary Table 4. Sample Size Composition of GWAS Contributes to Sex Bias To assess whether imbalanced sex representation in the discovery GWAS used to derive PRS contributes to sex-related differences in PRS predictive performance, we examined the relationship between female-to-male ratios of GWAS sample sizes and corresponding ratios of PRS predictive performance for each trait (Figure 3). For disease traits, a positive correlation was observed ( β = 0.0646, P -value = 3.08 × 10⁻³, R ² = 0.36), indicating that PRS derived from a larger female GWAS sample size tended to perform better in females. In contrast, quantitative traits showed no meaningful association ( β = 0.169, P -value = 0.60, R ² = 0.006). These findings suggest that sample-size imbalance in GWAS discovery cohorts partially drives the PRS prediction bias between sexes, consistent with the expectation that larger sample sizes yield more accurate estimates of allele effects, thereby improving prediction power. Consistency of Sex Bias Across PRS Methods Building on the PGS Catalog-wide characterization of sex differences, and complementary to the analysis of discovery sample composition, we evaluated whether the observed biases depend on the choice of PRS algorithm. We focused on 58 traits that were significantly sex-biased in the PGS Catalog analysis and generated polygenic scores using 7 widely used methods. Across methods, trait-level estimates were relatively concordant (Figure 4). Patterns in the trait-by-method matrix clustered by trait rather than by method, and no algorithm consistently amplified or attenuated female- or male-favoring effects. We used the Intraclass Correlation Coefficient (ICC) to assess consistency across methods. Agreement across the 7 method-specific estimates was high, with an intraclass correlation of ICC (2,k)=0.93 (95% CI, 0.90–0.96; P<.001), and virtually identical results for ICC (1,k) and ICC (3,k) (Supplementary Table 8). These findings indicate that sex-differential PRS performance is robust to algorithmic choice and thus more likely reflects trait-specific biology rather than method artifacts. Detailed performance metrics and statistical results are provided in Supplementary Table 7. Potential Role of Genetic Architecture in Shaping Sex Differences To examine whether sex differences in PRS performance reflect underlying differences in genetic architecture, we conducted sex-stratified GWAS for traits that exhibited significant sex bias in PRS performance, including 15 disease traits and 43 quantitative traits. Using sex-specific GWAS summary statistics from the UK Biobank, heritability was estimated with linkage disequilibrium score regression (LDSC) We observed that the magnitude of sex differences in PRS performance was significantly correlated with the gap in SNP-based heritability between sexes (Figure 5). Among disease and quantitative traits, the correlation was strong. The Pearson correlation was 0.81 (P < 0.001) and 0.58 (P < 0.001) for disease and quantitative traits, respectively. Notably, several traits, such as type 1 diabetes, deviated from this overall trend, suggesting that additional mechanisms beyond differences in total heritability may contribute to sex-specific prediction disparities. To further assess whether overall heritability patterns differ systematically between male- and female-biased traits, we compared the distributions of heritability difference across traits classified by PRS bias direction. Traits whose PRSs performed better in males exhibited significantly higher male-to-female heritability than those favoring females (P = 1.312 × 10⁻³; Figure 5C). This result supports the notion that, on average, higher SNP-based heritability corresponds to stronger PRS predictive performance, consistent with sex-dependent differences in polygenic architecture. Collectively, these findings indicate that sex-dependent genetic architecture is a strong driver for sex differences in PRS performance. Discussion This study demonstrates that sex differences in PRS performance are both widespread and biologically grounded. Prior PRS implementation studies have noted isolated instances of sex differences in risk stratification 7,18,43 , yet none have systematically quantified their prevalence across thousands of polygenic scores and diverse trait categories. Our analysis provides population-level evidence that sex disparities are both common and trait-specific, emphasizing the need for sex-aware evaluation before clinical deployment. Leveraging more than 3,000 publicly available PRS and newly generated scores using published GWAS and multiple methods, we found that disparities between males and females were common across diverse traits and partly attributable to an imbalance in discovery GWAS sample composition, while algorithmic choice had minimal influence. Autoimmune and endocrine conditions tended to show stronger prediction in females, whereas cardiometabolic conditions favored males, mirroring well-established patterns of sexual dimorphism. This is aligned with the previous study. 7,18,43 Notably, the association between GWAS sample size composition and sex differences in PRS performance was primarily observed for disease traits but not for quantitative traits. One possible explanation is that, for disease phenotypes, GWAS sample size may partly reflect underlying disease characteristics such as prevalence and heritability, which influence statistical power for variant discovery and downstream PRS construction. 44 In contrast, the sample size for quantitative traits is typically determined by the availability of phenotype measurements rather than the genetic architecture of the trait itself, which may explain the lack of a comparable relationship. These findings indicate that sex differences in PRS performance are unlikely to arise from analytic artifacts alone and underscore the importance of sex-balanced GWAS designs and sex-aware PRS frameworks. Biological factors emerged as a stronger and more consistent driver. Sex-stratified GWAS demonstrated that differences in PRS performance closely mirrored differences in SNP-based heritability, with stronger correlations for disease traits. Traits with higher heritability in one sex consistently showed stronger PRS prediction in that sex, pointing to systematic differences in polygenic architecture. These observations align with prior evidence of sex-dependent heritability, variant-effect heterogeneity, and hormone-related gene regulation, suggesting that sex differences in PRS reflect genuine biological divergence in genetic architecture rather than analytic artifacts. 11,20,45 This convergence across PGS Catalog scores, UK Biobank PRS scores, and sex-stratified GWAS reinforces the biological foundation of sex-biased prediction. These findings have important implications for clinical translation. As PRS are increasingly incorporated into preventive clinical genomics programs, unrecognized sex differences may lead to systematic misclassification of genetic risk. Female risk for cardiometabolic diseases may be underestimated, while male risk for autoimmune or endocrine disorders may be misestimated, with potential consequences for screening eligibility, preventive therapy allocation, and long-term monitoring. Routine evaluation of sex-stratified predictive effectiveness should therefore become a standard component of PRS validation prior to clinical adoption. Our analysis was limited to autosomal variants in European ancestry participants, and X-chromosome and multi-ancestry analyses are warranted in further studies. Our analysis was also limited by incomplete metadata in the PGS Catalog API, as many studies did not report discovery sample sizes or sex composition, and large-scale manual curation was not feasible. Some underlying GWAS datasets were not publicly available, which may have reduced the completeness of our sample-size analysis. Overall, this study provides the most comprehensive evaluation to date of sex-differential PRS performance across complex traits and demonstrates that these differences are pervasive and largely biologically driven. Addressing sex-dependent genetic architecture will be essential for fair and clinically meaningful implementation of polygenic risk scores. Declarations Conflict of Interest Disclosures Dr. Natarajan reports research grants from Allelica, Amgen, Apple, Boston Scientific, Cleerly, Genentech / Roche, Ionis, Novartis, and Silence Therapeutics, personal fees from AIRNA, Allelica, Amgen, Apple, AstraZeneca, Bain Capital, Blackstone Life Sciences, Bristol Myers Squibb, Creative Education Concepts, CRISPR Therapeutics, Eli Lilly & Co, Esperion Therapeutics, Foresite Capital, Foresite Labs, Genentech / Roche, GV, HeartFlow, Incyte, Magnet Biomedicine, Merck, Novartis, Novo Nordisk, TenSixteen Bio, Tourmaline Bio, and Ursa Medicines, equity in Bolt, Candela, Mercury, MyOme, Parameter Health, Preciseli, and TenSixteen Bio, royalties from Recora for intensive cardiac rehabilitation, and spousal employment at Vertex Pharmaceuticals, all unrelated to the present work. Dr. Fahed reports being co-founder of Goodpath and Avigena, serving as scientific advisor to MyOme, Arboretum Health, HeartFlow, and Aditum Bio and receiving sponsored research awards from Foresite, Sarepta Therapeutics, and Allelica, all unrelated to the current work. Dr. Sui reports serving as a consultant for Arboretum Lifesciences. Acknowledgments Dr. Wang is supported by the National Natural Science Foundation of China (grant number 82470352), the Noncommunicable Chronic Diseases-National Science and Technology Major Project-2023ZD0503201, and Pioneering Action Grants of the Chinese Academy of Sciences. Dr. Natarajan is supported by grants from NHLBI (R01HL127564), NHGRI (U01HG011719), and Massachusetts General Hospital (Paul and Phyllis Fireman Endowed Chair in Vascular Medicine). Dr. Fahed receives funding from the National Heart Lung and Blood Institute under award numbers K08 HL161448 and R01 HL164629. Dr. Sui is supported by the TOPMed fellowship from the National Heart Lung and Blood Institute. Dr. Halford is supported by a grant from NHGRI (5T32HG010464-07). Data Availability Individual-level data from the UK Biobank are available to qualified researchers through a formal application (https://www.ukbiobank.ac.uk). Summary-level GWAS statistics and PGS weights used in this study are publicly accessible from the PGS Catalog (https://www.pgscatalog.org). All derived PRS performance metrics and summary statistics are available upon reasonable request to the corresponding author. Code Availability Code used for polygenic risk score construction, evaluation, and statistical analyses in this study is publicly available at https://github.com/xinyu-c9/SexdiffPRS. References Wray, N. R. et al. From basic science to clinical application of polygenic risk scores: a primer. 78 , 101-109 (2021). Khera, A. V. et al. Genome-wide polygenic scores for common diseases identify individuals with risk equivalent to monogenic mutations. 50 , 1219-1224 (2018). Khan, A. et al. Genome-wide polygenic score to predict chronic kidney disease across ancestries. 28 , 1412-1420 (2022). Liu, W., Zhuang, Z., Wang, W., Huang, T. & Liu, Z. J. F. i. g. An improved genome-wide polygenic score model for predicting the risk of type 2 diabetes. 12 , 632385 (2021). Martin, A. R. et al. Clinical use of current polygenic risk scores may exacerbate health disparities. 51 , 584-591 (2019). Wand, H. et al. Improving reporting standards for polygenic scores in risk prediction studies. 591 , 211-219 (2021). EUGenMed et al. Gender in cardiovascular diseases: impact on clinical manifestations, management, and outcomes. 37 , 24-34 (2016). Vaccarino, V. et al. Sex differences in mortality after acute myocardial infarction: changes from 1994 to 2006. 169 , 1767-1774 (2009). Sokka, T. et al. Women, men, and rheumatoid arthritis: analyses of disease activity, disease characteristics, and treatments in the QUEST-RA study. 11 , 1-12 (2009). Altemus, M., Sarvaiya, N. & Epperson, C. N. J. F. i. n. Sex differences in anxiety and depression clinical perspectives. 35 , 320-330 (2014). Bernabeu, E. et al. Sex differences in genetic architecture in the UK Biobank. 53 , 1283-1289 (2021). Ober, C., Loisel, D. A. & Gilad, Y. J. N. R. G. Sex-specific genetic architecture of human disease. 9 , 911-922 (2008). Uddin, M. S. et al. Estrogen signaling in Alzheimer’s disease: molecular insights and therapeutic targets for Alzheimer’s dementia. 57 , 2654-2670 (2020). Krieger, N. J. A. r. o. p. h. Measures of Racism, Sexism, Heterosexism, and Gender Binarism for Health Equity Research: From Structural Injustice to Embodied Harm-An Ecosocial Analysis. 41 , 37-62 (2019). Schroeder, P. H., Cole, J. B., Leong, A., Florez, J. C. & Mercader, J. M. J. D. Large-scale sex-stratified additive and recessive GWAS identifies novel large-effect variants and improves polygenic prediction for type 2 diabetes. 71 (2022). Dumitrescu, L. et al. Sex differences in the genetic predictors of Alzheimer’s pathology. 142 , 2581-2589 (2019). Pirastu, N. et al. Genetic analyses identify widespread sex-differential participation bias. 53 , 663-671 (2021). Patel, A. P. et al. A multi-ancestry polygenic risk score improves risk prediction for coronary artery disease. 29 , 1793-1803 (2023). Van Alten, S., Domingue, B. W., Galama, T. & Marees, A. T. J. M. Reweighting the UK Biobank to reflect its underlying sampling population substantially reduces pervasive selection bias due to volunteering. 2022.2005. 2016.22275048 (2022). Khramtsova, E. A., Davis, L. K. & Stranger, B. E. J. N. R. G. The role of sex in the genomics of human complex traits. 20 , 173-190 (2019). Vassy, J. L. et al. The GenoVA study: Equitable implementation of a pragmatic randomized trial of polygenic-risk scoring in primary care. 110 , 1841-1852 (2023). Cross, B., Turner, R. & Pirmohamed, M. J. F. i. g. Polygenic risk scores: An overview from bench to bedside for personalised medicine. 13 , 1000667 (2022). Bycroft, C. et al. The UK Biobank resource with deep phenotyping and genomic data. 562 , 203-209 (2018). Manichaikul, A. et al. Robust relationship inference in genome-wide association studies. 26 , 2867-2873 (2010). Lambert, S. A. et al. Enhancing the Polygenic Score Catalog with tools for score calculation and ancestry normalization. 56 , 1989-1994 (2024). Chang, C. C. et al. Second-generation PLINK: rising to the challenge of larger and richer datasets. 4 , s13742-13015-10047-13748 (2015). Mbatchou, J. et al. Computationally efficient whole-genome regression for quantitative and binary traits. 53 , 1097-1103 (2021). Ge, T., Chen, C.-Y., Ni, Y., Feng, Y.-C. A. & Smoller, J. W. J. N. c. Polygenic prediction via Bayesian regression and continuous shrinkage priors. 10 , 1-10 (2019). Lloyd-Jones, L. R. et al. Improved polygenic prediction by Bayesian multiple regression on summary statistics. 10 , 1-11 (2019). Mak, T. S. H., Porsch, R. M., Choi, S. W., Zhou, X. & Sham, P. C. J. G. e. Polygenic scores via penalized regression on summary statistics. 41 , 469-480 (2017). Privé, F., Arbel, J. & Vilhjálmsson, B. J. J. B. LDpred2: better, faster, stronger. 36 , 5424-5431 (2020). Privé, F., Aschard, H., Ziyatdinov, A. & Blum, M. G. J. B. Efficient analysis of large-scale genome-wide data with two R packages: bigstatsr and bigsnpr. 34 , 2781-2787 (2018). Privé, F., Vilhjálmsson, B. J., Aschard, H. & Blum, M. G. J. T. A. J. o. H. G. Making the most of clumping and thresholding for polygenic scores. 105 , 1213-1221 (2019). Robinson, M. R. et al. Genetic evidence of assortative mating in humans. 1 , 1-13 (2017). Wray, N. R., Goddard, M. E. & Visscher, P. M. J. G. r. Prediction of individual genetic risk to disease from genome-wide association studies. 17 , 1520-1528 (2007). Zhou, G. & Zhao, H. J. P. g. A fast and robust Bayesian nonparametric method for prediction of complex traits using summary statistics. 17 , e1009697 (2021). Nature, G. P. C. J. A global reference for human genetic variation. 526 , 68 (2015). Nature, I. H. C. J. Integrating common and rare genetic variation in diverse human populations. 467 , 52 (2010). Vallat, R. J. J. O. S. S. Pingouin: statistics in Python. 3 , 1026 (2018). Bulik-Sullivan, B. K. et al. LD Score regression distinguishes confounding from polygenicity in genome-wide association studies. 47 , 291-295 (2015). Muñoz, M. et al. Evaluating the contribution of genetics and familial shared environment to common disease using the UK Biobank. 48 , 980-983 (2016). Wickham, H. J. W. i. r. c. s. ggplot2. 3 , 180-185 (2011). Zhang, F. et al. Sex differences in the association between polygenic risk score and atrial fibrillation incidence: a prospective cohort study. (2025). Visscher, P. M. et al. 10 years of GWAS discovery: biology, function, and translation. 101 , 5-22 (2017). Wang, X., Magkos, F., Mittendorfer, B. J. T. J. o. C. E. & Metabolism. Sex differences in lipid and lipoprotein metabolism: it's not just about sex hormones. 96 , 885-893 (2011). Additional Declarations Yes there is potential Competing Interest. Dr. Natarajan reports research grants from Allelica, Amgen, Apple, Boston Scientific, Cleerly, Genentech / Roche, Ionis, Novartis, and Silence Therapeutics, personal fees from AIRNA, Allelica, Amgen, Apple, AstraZeneca, Bain Capital, Blackstone Life Sciences, Bristol Myers Squibb, Creative Education Concepts, CRISPR Therapeutics, Eli Lilly & Co, Esperion Therapeutics, Foresite Capital, Foresite Labs, Genentech / Roche, GV, HeartFlow, Incyte, Magnet Biomedicine, Merck, Novartis, Novo Nordisk, TenSixteen Bio, Tourmaline Bio, and Ursa Medicines, equity in Bolt, Candela, Mercury, MyOme, Parameter Health, Preciseli, and TenSixteen Bio, royalties from Recora for intensive cardiac rehabilitation, and spousal employment at Vertex Pharmaceuticals, all unrelated to the present work. Dr. Fahed reports being co-founder of Goodpath and Avigena, serving as scientific advisor to MyOme, Arboretum Health, HeartFlow, and Aditum Bio and receiving sponsored research awards from Foresite, Sarepta Therapeutics, and Allelica, all unrelated to the current work. Dr. Sui reports serving as a consultant for Arboretum Lifesciences. Supplementary Files SupplementaryTables.pdf Supplementary Tables SupplementaryFigure1.docx Cite Share Download PDF Status: Under Review Version 1 posted You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-9153854","acceptedTermsAndConditions":true,"allowDirectSubmit":false,"archivedVersions":[],"articleType":"Article","associatedPublications":[],"authors":[{"id":630524628,"identity":"8a74ff27-9fc7-4837-88b4-fc06da97848c","order_by":0,"name":"Akl Fahed","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAA/ElEQVRIiWNgGAWjYDAC5gNsDAxsNgwMEgxsUKEEBgYefFrYEkBa0kjXcpgELfxtzM8efCg7Hy0f3cD24Oeew/Lm7QmMD9624dYicYzN3HDGudu5G+8cYDfseXbYcM6ZB8yGc/FoYbjfYCbN2wbUMiOBTYLnQBrjDIkENqAIbh3yx9i/Sf9tOwfWIvnnQJo9UAv7b3xaDI7xmEkzth3InQ8ynOeATSLIFmZ8WgyP8ZRJ9pxLzt0gkdgmLXPAJnkGz8NmyTnncGuRO8a+TeJHmV3u/BnJxyTfHJCwncGefPDDmzI83oe78ABjA5QJZxAA8kSqGwWjYBSMghEIAC/QUra0KILxAAAAAElFTkSuQmCC","orcid":"https://orcid.org/0000-0002-4849-6389","institution":"Massachusetts General Hospital","correspondingAuthor":true,"prefix":"","firstName":"Akl","middleName":"","lastName":"Fahed","suffix":""},{"id":630524629,"identity":"a4dd8895-4df2-4394-9a42-64a940a6060c","order_by":1,"name":"Xingyu Chen","email":"","orcid":"","institution":"Massachusetts General Hospital/ Broad Institute of MIT and Harvard/ Beijing Institute of Genomics (China National Center for Bioinformation)","correspondingAuthor":false,"prefix":"","firstName":"Xingyu","middleName":"","lastName":"Chen","suffix":""},{"id":630524630,"identity":"673be888-ad48-4aaf-9252-709e3338cb2f","order_by":2,"name":"Tingfeng Xu","email":"","orcid":"","institution":"Beijing Institute of Genomics, Chinese Academy of Sciences. (China National Center for Bioinformation)","correspondingAuthor":false,"prefix":"","firstName":"Tingfeng","middleName":"","lastName":"Xu","suffix":""},{"id":630524631,"identity":"8e926467-f2d3-4cdb-8abc-521e07e0f99c","order_by":3,"name":"Yang Sui","email":"","orcid":"","institution":"Broad Institute of MIT and Harvard","correspondingAuthor":false,"prefix":"","firstName":"Yang","middleName":"","lastName":"Sui","suffix":""},{"id":630524632,"identity":"1e194df9-a209-4512-99ad-a9e0b54012e2","order_by":4,"name":"Kelvin Supriami","email":"","orcid":"https://orcid.org/0000-0001-5577-5899","institution":"Massachusetts General Hospital/ Broad Institute of MIT and Harvard","correspondingAuthor":false,"prefix":"","firstName":"Kelvin","middleName":"","lastName":"Supriami","suffix":""},{"id":630524633,"identity":"54acaa30-34de-4efb-bb11-de440c4c3a75","order_by":5,"name":"Min Seo Kim","email":"","orcid":"","institution":"Broad Institute of MIT and Harvard","correspondingAuthor":false,"prefix":"","firstName":"Min","middleName":"Seo","lastName":"Kim","suffix":""},{"id":630524634,"identity":"7db386ee-cac9-41e5-a956-35c96934b9eb","order_by":6,"name":"Injeong Shim","email":"","orcid":"","institution":"Massachusetts General Hospital","correspondingAuthor":false,"prefix":"","firstName":"Injeong","middleName":"","lastName":"Shim","suffix":""},{"id":630524635,"identity":"b0cee2f8-2bef-4ea0-90d9-93ecf8631b20","order_by":7,"name":"Jennifer Halford","email":"","orcid":"https://orcid.org/0000-0002-1454-3650","institution":"Massachusetts General Hospital, Broad Institute of MIT and Harvard","correspondingAuthor":false,"prefix":"","firstName":"Jennifer","middleName":"","lastName":"Halford","suffix":""},{"id":630524636,"identity":"36f864f5-d54c-4ef2-a59f-6413935ce737","order_by":8,"name":"Kang Yu","email":"","orcid":"","institution":"Beijing Institute of Genomics (China National Center for Bioinformation)","correspondingAuthor":false,"prefix":"","firstName":"Kang","middleName":"","lastName":"Yu","suffix":""},{"id":630524637,"identity":"d156dbc3-0535-4c29-8989-af25a6942789","order_by":9,"name":"Xinyu Zhou","email":"","orcid":"","institution":"Beijing Institute of Genomics (China National Center for Bioinformation)","correspondingAuthor":false,"prefix":"","firstName":"Xinyu","middleName":"","lastName":"Zhou","suffix":""},{"id":630524638,"identity":"48adc441-fdf2-4553-a13d-cc5797b26792","order_by":10,"name":"Qiuli Chen","email":"","orcid":"","institution":"Beijing Institute of Genomics (China National Center for Bioinformation)","correspondingAuthor":false,"prefix":"","firstName":"Qiuli","middleName":"","lastName":"Chen","suffix":""},{"id":630524639,"identity":"60bc663f-ef26-4fde-b9cf-c601869ade1b","order_by":11,"name":"Jing Hao","email":"","orcid":"","institution":"Beijing Institute of Genomics (China National Center for Bioinformation)","correspondingAuthor":false,"prefix":"","firstName":"Jing","middleName":"","lastName":"Hao","suffix":""},{"id":630524640,"identity":"3f89eacd-7f40-46cb-992b-87c08206ea91","order_by":12,"name":"Shaoqi Wang","email":"","orcid":"","institution":"Beijing Institute of Genomics (China National Center for Bioinformation)","correspondingAuthor":false,"prefix":"","firstName":"Shaoqi","middleName":"","lastName":"Wang","suffix":""},{"id":630524641,"identity":"d1d96bb8-ace2-4c47-835e-9b0a55fc00aa","order_by":13,"name":"Kewei Zhao","email":"","orcid":"","institution":"Beijing Institute of Genomics (China National Center for Bioinformation)","correspondingAuthor":false,"prefix":"","firstName":"Kewei","middleName":"","lastName":"Zhao","suffix":""},{"id":630524642,"identity":"26566eb9-9731-429d-8a24-26e2fb3a1298","order_by":14,"name":"Shuangqiao Liao","email":"","orcid":"","institution":"Beijing Institute of Genomics (China National Center for Bioinformation)","correspondingAuthor":false,"prefix":"","firstName":"Shuangqiao","middleName":"","lastName":"Liao","suffix":""},{"id":630524643,"identity":"304f8591-e0ee-4dd8-b899-13800c83c356","order_by":15,"name":"Jichen Wen","email":"","orcid":"","institution":"China National Center for Bioinformation","correspondingAuthor":false,"prefix":"","firstName":"Jichen","middleName":"","lastName":"Wen","suffix":""},{"id":630524644,"identity":"3fdf26a7-1237-40ae-bf07-3c827fc78852","order_by":16,"name":"Jiaqi Ding","email":"","orcid":"","institution":"Beijing Institute of Genomics (China National Center for Bioinformation)","correspondingAuthor":false,"prefix":"","firstName":"Jiaqi","middleName":"","lastName":"Ding","suffix":""},{"id":630524645,"identity":"98186f93-369c-42cd-8930-039f0973c979","order_by":17,"name":"Xinyu Zhu","email":"","orcid":"","institution":"Beijing Institute of Genomics (China National Center for Bioinformation)","correspondingAuthor":false,"prefix":"","firstName":"Xinyu","middleName":"","lastName":"Zhu","suffix":""},{"id":630524646,"identity":"391353f3-3085-4c39-a4e7-8f0482f74d8c","order_by":18,"name":"Weixiang Zhao","email":"","orcid":"","institution":"Beijing Institute of Genomics (China National Center for Bioinformation)","correspondingAuthor":false,"prefix":"","firstName":"Weixiang","middleName":"","lastName":"Zhao","suffix":""},{"id":630524647,"identity":"f0b3c9e7-0bad-438e-9602-22b68ef0edf8","order_by":19,"name":"Junyu Yang","email":"","orcid":"","institution":"Beijing Institute of Genomics (China National Center for Bioinformation)","correspondingAuthor":false,"prefix":"","firstName":"Junyu","middleName":"","lastName":"Yang","suffix":""},{"id":630524648,"identity":"ee7c05a6-ee40-4614-aa04-d9aba515c1ce","order_by":20,"name":"Yu Zhang","email":"","orcid":"","institution":"Beijing Institute of Genomics (China National Center for Bioinformation)","correspondingAuthor":false,"prefix":"","firstName":"Yu","middleName":"","lastName":"Zhang","suffix":""},{"id":630524649,"identity":"e7aaf60c-c3c7-4966-905a-50542a00d9f5","order_by":21,"name":"Meier Chen","email":"","orcid":"","institution":"Beijing Institute of Genomics (China National Center for Bioinformation)","correspondingAuthor":false,"prefix":"","firstName":"Meier","middleName":"","lastName":"Chen","suffix":""},{"id":630524650,"identity":"fa6a2d5e-11f2-47a3-b0fe-459b48e8c802","order_by":22,"name":"Komal Zaib","email":"","orcid":"","institution":"Beijing Institute of Genomics (China National Center for Bioinformation)","correspondingAuthor":false,"prefix":"","firstName":"Komal","middleName":"","lastName":"Zaib","suffix":""},{"id":630524651,"identity":"52766eb9-31ff-4feb-8c7f-c32e17d66793","order_by":23,"name":"Bitao Zhong","email":"","orcid":"","institution":"Beijing Institute of Genomics (China National Center for Bioinformation)","correspondingAuthor":false,"prefix":"","firstName":"Bitao","middleName":"","lastName":"Zhong","suffix":""},{"id":630524652,"identity":"1c8bef85-4796-45e2-825d-4d35799e77b0","order_by":24,"name":"Fei Wang","email":"","orcid":"","institution":"Beijing Institute of Genomics (China National Center for Bioinformation)","correspondingAuthor":false,"prefix":"","firstName":"Fei","middleName":"","lastName":"Wang","suffix":""},{"id":630524653,"identity":"7362d882-c9ba-4bf7-b562-7888181f1063","order_by":25,"name":"Whitney Hornsby","email":"","orcid":"","institution":"Massachusetts General Hospital","correspondingAuthor":false,"prefix":"","firstName":"Whitney","middleName":"","lastName":"Hornsby","suffix":""},{"id":630524654,"identity":"9f77e0dc-2b4a-40a7-8448-574076a34e14","order_by":26,"name":"Kaavya Paruchuri","email":"","orcid":"","institution":"Broad Institute","correspondingAuthor":false,"prefix":"","firstName":"Kaavya","middleName":"","lastName":"Paruchuri","suffix":""},{"id":630524655,"identity":"f3036ff8-fde8-4a7f-bd40-48d2c11323e5","order_by":27,"name":"Pradeep Natarajan","email":"","orcid":"https://orcid.org/0000-0001-8402-7435","institution":"Massachusetts General Hospital","correspondingAuthor":false,"prefix":"","firstName":"Pradeep","middleName":"","lastName":"Natarajan","suffix":""},{"id":630524656,"identity":"ad578f70-7e1f-4702-933d-4a1dc8e4d7b0","order_by":28,"name":"Minxian Wang","email":"","orcid":"","institution":"Beijing Institute of Genomics (China National Center for Bioinformation)","correspondingAuthor":false,"prefix":"","firstName":"Minxian","middleName":"","lastName":"Wang","suffix":""}],"badges":[],"createdAt":"2026-03-18 02:56:39","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-9153854/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-9153854/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":108101942,"identity":"f45b258e-bbb8-42fd-800a-288124d3aebc","added_by":"auto","created_at":"2026-04-29 11:02:13","extension":"jpg","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":59872,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eDistribution of Sex Bias in Polygenic Risk Score Performance by Score and Trait\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eStacked bars summarize the number of PGS Catalog scores evaluated in the UK Biobank cohort (N=409,440 after quality control; European ancestry). Panel A: score-level results for disease and quantitative scores (No bias=gray; Female-biased=red; Male-biased=blue). Panel B: trait-level aggregation. Numbers above bars denote totals; in-bar labels denote category counts. Sex differences in scores were assessed using a 2-sided Z test comparing sex-stratified regression coefficients with \u003cem\u003eFDR\u003c/em\u003e\u0026lt;0.05, and traits were determined by ensembling all scores and evaluating them with the same Z test strategy. PRS values were standardized to a mean of 0 and a standard deviation of 1; regression covariates included age, sex, genotyping array and top 10 genetic principal component.\u003c/p\u003e","description":"","filename":"1.jpg","url":"https://assets-eu.researchsquare.com/files/rs-9153854/v1/edd58a7da77b5f350b356368.jpg"},{"id":108182502,"identity":"2b48529c-987d-4cd3-b7d4-f8f8333ba712","added_by":"auto","created_at":"2026-04-30 08:59:24","extension":"jpg","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":364772,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eTrait-Level Sex Bias in PRS Performance in the PGS Catalog\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eFor each trait, multiple PRSs were combined into a single ensemble score using multivariable regression in the UK Biobank tuning set. This ensemble score was then evaluated separately in males and females; the difference in regression coefficients \u003cem\u003eβ\u003c/em\u003e\u003csub\u003efemale\u003c/sub\u003e − \u003cem\u003eβ\u003c/em\u003e\u003csub\u003emale\u003c/sub\u003e was used to represent sex bias. Positive values indicate stronger predictive performance in females, while negative values indicate stronger predictive performance in males. Error bars represent 95% confidence intervals for the sex-difference estimates. Traits are ordered by absolute effect size to facilitate comparison. PRS values were standardized to a mean of 0 and a standard deviation of 1; regression covariates included age, sex, genotyping array and top 10 genetic principal component.\u003c/p\u003e","description":"","filename":"2.jpg","url":"https://assets-eu.researchsquare.com/files/rs-9153854/v1/0419878ef772fb78dd7dbbb2.jpg"},{"id":108181992,"identity":"1938f100-28fe-4de5-9d4c-f24bd817116c","added_by":"auto","created_at":"2026-04-30 08:59:03","extension":"jpg","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":115130,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eAssociation Between Discovery Sample Size Imbalance and Sex Differences in PRS Performance\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eScatterplots show the relationship between female-to-male ratios of GWAS discovery sample size (X-axis) and corresponding female-to-male ratios of PRS predictive performance (Y-axis) for disease (left) and quantitative (right) traits. Each point represents one trait derived from the PGS Catalog. PRS values were standardized to a mean of 0 and a standard deviation of 1\u003c/p\u003e","description":"","filename":"3.jpg","url":"https://assets-eu.researchsquare.com/files/rs-9153854/v1/2ffac9d68b36608d4d25c107.jpg"},{"id":108181286,"identity":"941ecac9-5f8e-4388-b42a-605d22845b31","added_by":"auto","created_at":"2026-04-30 08:58:30","extension":"jpg","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":411638,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eHeatmap of Sex Bias in PRS Performance Across 58 Traits and 7 Methods\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eEach cell displays the female-to-male performance ratio for a given trait–method pair (\u003cem\u003eβ\u003c/em\u003e\u003csub\u003efemale\u003c/sub\u003e/\u003cem\u003eβ\u003c/em\u003e\u003csub\u003emale\u003c/sub\u003e), among 58 traits and 7 methods. Warmer vs cooler colors indicate ratios \u0026gt;1 vs \u0026lt;1. Asterisks denote pairs with significant sex differences by Z test comparing sex-stratified regression coefficients with \u003cem\u003eFDR\u003c/em\u003e \u0026lt; 0.05. PRS values were standardized to a mean of 0 and a standard deviation of 1; regression covariates included age, sex, genotyping array and top 10 genetic principal component.\u003c/p\u003e","description":"","filename":"4.jpg","url":"https://assets-eu.researchsquare.com/files/rs-9153854/v1/04fc9022c668a7fd9904ee56.jpg"},{"id":108181814,"identity":"18785ed5-9905-4147-82d8-07a2a1b1a421","added_by":"auto","created_at":"2026-04-30 08:58:56","extension":"jpg","order_by":5,"title":"Figure 5","display":"","copyAsset":false,"role":"figure","size":100340,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eAssociation Between Sex Differences in Polygenic Risk Score Performance and Heritability\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eA,B, Scatterplots show the association between sex differences in polygenic risk score (PRS) predictive effect (X-axis) and SNP-based heritability (Y-axis) estimated from sex-stratified GWAS in the UK Biobank. The X-axis represents the difference in PRS effect size between males and females (effect size difference = \u003cem\u003eβ\u003c/em\u003e\u003csub\u003emale\u003c/sub\u003e – \u003cem\u003eβ\u003c/em\u003e\u003csub\u003efemale\u003c/sub\u003e), and the Y-axis represents the difference in heritability (h² difference = h²\u003csub\u003emale\u003c/sub\u003e – h²\u003csub\u003efemale\u003c/sub\u003e). Positive values indicate higher performance or heritability in males, negative values indicate higher performance in females. Panels A and B display results for quantitative and disease traits, respectively. \u003cstrong\u003eC, \u003c/strong\u003eBoxplot of the difference in heritability stratified by PRS prediction bias. The X-axis indicate bias category, Y-axis indicate SNP-based heritability difference. The significance between groups was tested using a Mann–Whitney U test (P = 1.312 × 10⁻³).\u003c/p\u003e","description":"","filename":"5.jpg","url":"https://assets-eu.researchsquare.com/files/rs-9153854/v1/fd1a828df69871ec22b4b9df.jpg"},{"id":108183797,"identity":"9516ff52-1d14-4f30-8860-e7126e74f35e","added_by":"auto","created_at":"2026-04-30 09:02:49","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":1362588,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-9153854/v1/fa004e3f-8527-4a49-bbb1-c88096bed631.pdf"},{"id":108181569,"identity":"ee793ab4-d938-437b-9c33-845a801cc0b0","added_by":"auto","created_at":"2026-04-30 08:58:45","extension":"pdf","order_by":1,"title":"","display":"","copyAsset":false,"role":"supplement","size":4233086,"visible":true,"origin":"","legend":"Supplementary Tables","description":"","filename":"SupplementaryTables.pdf","url":"https://assets-eu.researchsquare.com/files/rs-9153854/v1/e7b8058b4fea95d1063aa0b0.pdf"},{"id":108181652,"identity":"092cc9ad-4d7a-41e0-ae99-309c53712a2c","added_by":"auto","created_at":"2026-04-30 08:58:49","extension":"docx","order_by":2,"title":"","display":"","copyAsset":false,"role":"supplement","size":835133,"visible":true,"origin":"","legend":"","description":"","filename":"SupplementaryFigure1.docx","url":"https://assets-eu.researchsquare.com/files/rs-9153854/v1/9ef7ff014bd76405fbc80dd0.docx"}],"financialInterests":"\u003cb\u003eYes\u003c/b\u003e there is potential Competing Interest.\nDr. Natarajan reports research grants from Allelica, Amgen, Apple, Boston Scientific, Cleerly, Genentech / Roche, Ionis, Novartis, and Silence Therapeutics, personal fees from AIRNA, Allelica, Amgen, Apple, AstraZeneca, Bain Capital, Blackstone Life Sciences, Bristol Myers Squibb, Creative Education Concepts, CRISPR Therapeutics, Eli Lilly \u0026 Co, Esperion Therapeutics, Foresite Capital, Foresite Labs, Genentech / Roche, GV, HeartFlow, Incyte, Magnet Biomedicine, Merck, Novartis, Novo Nordisk, TenSixteen Bio, Tourmaline Bio, and Ursa Medicines, equity in Bolt, Candela, Mercury, MyOme, Parameter Health, Preciseli, and TenSixteen Bio, royalties from Recora for intensive cardiac rehabilitation, and spousal employment at Vertex Pharmaceuticals, all unrelated to the present work. Dr. Fahed reports being co-founder of Goodpath and Avigena, serving as scientific advisor to MyOme, Arboretum Health, HeartFlow, and Aditum Bio and receiving sponsored research awards from Foresite, Sarepta Therapeutics, and Allelica, all unrelated to the current work. Dr. Sui reports serving as a consultant for Arboretum Lifesciences.","formattedTitle":"Widespread and Biologically Driven Sex Disparities in Polygenic Risk Prediction Across Complex Traits","fulltext":[{"header":"Main","content":"\u003cp\u003ePolygenic risk scores (PRS), which aggregate the effects of thousands of genetic variants identified through genome-wide association studies (GWAS), have emerged as powerful tools for quantifying individual genetic predisposition to complex traits and diseases.\u003csup\u003e1\u003c/sup\u003e By weighing risk alleles through effect size estimates from large-scale GWAS, PRS have demonstrated clinical potential by enabling disease prediction and risk stratification.\u003csup\u003e2-4\u003c/sup\u003e However, the generalizability of PRS remains constrained across ancestral populations, with significantly reduced predictive accuracy in non-European ancestries, which are often underrepresented in genetic studies.\u003csup\u003e5,6\u003c/sup\u003e\u003c/p\u003e\n\u003cp\u003eBeyond ancestral disparities in representation, pronounced sex differences in disease prevalence persist across numerous traits. Coronary artery disease (CAD) exemplifies this divergence. \u0026nbsp;Male sex is a risk factor for CAD, and females have delayed disease onset by 7-10 years compared to males\u003csup\u003e7\u003c/sup\u003e, yet females experience higher mortality rates post-myocardial infarction.\u003csup\u003e8\u003c/sup\u003e\u0026nbsp; Similarly, autoimmune diseases like rheumatoid arthritis show 3:1 female-to-male prevalence\u003csup\u003e9\u003c/sup\u003e, while neuropsychiatric disorders such as depression demonstrate sex-specific symptom profiles.\u003csup\u003e10\u003c/sup\u003e These epidemiological disparities arise from multifaceted biological mechanisms, including genotype by sex (GxS) interactions\u003csup\u003e11\u003c/sup\u003e, X-chromosome dosage effects\u003csup\u003e12\u003c/sup\u003e, sex hormone-mediated gene regulation\u003csup\u003e13\u003c/sup\u003e, and sex-related environmental exposures\u003csup\u003e14\u003c/sup\u003e.\u003c/p\u003e\n\u003cp\u003eDespite these established sex differences, current PRS frameworks often overlook sex-specific genetic architectures. Most GWAS meta-analyses either inadequately stratify by sex or assume linear additive effects across sex,\u003csup\u003e15-17\u003c/sup\u003e and PRS construction methods lack modelling of sex-specific effect sizes between sexes, potentially biasing predictions. For instance, prior analyses of CAD using a multiancestry polygenic risk score (GPS\u003csub\u003eMult\u003c/sub\u003e) from our group\u003csup\u003e18\u003c/sup\u003e revealed substantial sex-based prediction disparities. This imbalance likely stems from both biological factors, including underrepresentation of sex-specific loci in GWAS, and methodological limitations, including pervasive\u003csup\u003e\u0026nbsp;\u003c/sup\u003esample ascertainment bias\u003csup\u003e19\u003c/sup\u003e and insufficient modeling of gene-by-sex interactions\u003csup\u003e20\u003c/sup\u003e.\u003c/p\u003e\n\u003cp\u003eAs polygenic scores move towards clinical implementation, the absence of systematic evaluation of sex bias poses a critical barrier to equitable use\u003csup\u003e21,22\u003c/sup\u003e. A sex-blind approach may exacerbate disparities, yet no study has comprehensively quantified these differences or clarified their methodological and biological sources. Filling this gap is critical for allowing fair and effective clinical translation of PRS.\u003c/p\u003e\n\u003cp\u003eTo address these gaps, we used more than three thousand PRS from the PGS Catalog in conjunction with UK Biobank data to broadly assess the prevalence and magnitude of sex-dependent predictive disparities. We then leveraged GWAS data from the UK Biobank training dataset and employed seven PRS methods to investigate how GWAS sample size differences, methodological biases, and biological heterogeneity at sex-specific loci drive these disparities.\u0026nbsp;\u003c/p\u003e"},{"header":"Method","content":"\u003cp\u003e\u003cstrong\u003eStudy Population and Outcome Definitions\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eWe utilized data from the UK Biobank\u003csup\u003e23\u003c/sup\u003e under application IDs 7089 and 89885, which granted access to anonymized genetic and clinical data. Standard UKB quality control procedures were applied. Specifically, individuals exhibiting discrepancies between self-reported sex (UK Biobank Field 31) and genetically inferred sex (Field 22001) were excluded. Additionally, participants were removed if they presented an individual-level genotype missingness rate exceeding 5% (Field 22005), were outliers in heterozygosity or missing rate (Field 22027), or exhibited sex chromosome aneuploidy (Field 22019). After applying these QC steps and restricting analyses to genetically unrelated individuals of European ancestry (kinship coefficient \u0026lt; 0.0884)\u003csup\u003e24\u003c/sup\u003e, a total of 409,440 participants remained for further analysis. The cohort was split into independent training (50%), tuning (25%), and testing (25%) sets to enable unbiased model development, parameter optimization, and final evaluation of PRS performance.\u003c/p\u003e\n\u003cp\u003eDisease outcomes were defined using a combination of self-report and linked health records (Supplementary Table 1), and quantitative traits were centrally curated by the UKB.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eCuration of Polygenic Risk Scores for Evaluation\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eTo systematically investigate whether PRS exhibit widespread sex differences in their predictive performance, we used the PGS Catalog to assess sex bias in PRS predictive performance.\u003csup\u003e25\u003c/sup\u003e A total of 5,008 PRS (as of October 2024) were retained for subsequent analysis. We harmonized SNP identifiers between PGS Catalog weights and UK Biobank genotypes using standard coordinate and allele matching; scores with \u0026lt;80% variant overlap were excluded. For PRS with variant matching rates more than 0.8, we utilized all matched variants to calculate the PRS scores. This standardized procedure was uniformly applied across all PGS Catalog entries, ensuring consistency in our analysis involving over 5,000 publicly available PRSs.\u003c/p\u003e\n\u003cp\u003eFrom the 692 phenotypes in the PGS Catalog, we derived a curated set by applying exclusion criteria to ensure alignment with UK Biobank phenotypes and minimize conceptual or ethical ambiguities. We excluded traits represented by three or fewer PRS, those lacking direct UK Biobank counterparts, and descriptors that were broad or imprecise. Additional exclusions included disease subtypes nested within broader diagnoses, family-history and virology or serology traits, imaging-derived or procedural measures, isolated symptoms or behavioral items, narrowly defined dietary exposures, sex-specific conditions such as prostate or ovarian cancer, lifestyle and socioeconomic factors, and traits potentially raising ethical concerns in the context of sex differences. This yielded 145 phenotypes (64 diseases and 81 quantitative traits) for downstream PGS Catalog analyses (Supplementary Table 1).\u003c/p\u003e\n\u003cp\u003eFor each retained weight file, polygenic scores were generated in PLINK2(v2.0.0-a.6.5LM)\u003csup\u003e26\u003c/sup\u003e by multiplying the dosage of every risk allele by its assigned effect size and aggregating these weighted values across all variants for each individual in the UK Biobank.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eEvaluation of Sex Differences in Published Polygenic Risk Scores\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eAfter obtaining a comprehensive list of published PRS from the PGS Catalog, we evaluated their sex-specific predictive performance at both the score and trait levels. For each individual score, sex-stratified regression models were fitted in the complete UK Biobank data, and differences between males and females were assessed using a Z test. Scores were classified as female-biased, male-biased, or unbiased based on the direction and significance of effect differences. At the trait level, all scores for a phenotype were jointly modeled in the tuning set to derive a weighted ensemble score for phenotype prediction (ensemble trait-level PRS), which was then applied to the testing set for sex-specific evaluation.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAssociation of Discovery Sample Size with Sex-Differential PRS Performance\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eTo assess whether discovery cohort composition contributed to sex differences in PRS performance, GWAS sample sizes for each PGS Catalog score were obtained from the REST application programming interface (API)\u003csup\u003e25\u003c/sup\u003e, which supports batch access to curated score metadata, and were complemented by manual curation from primary publications when API fields were unavailable. Sex-specific counts were derived from reported male proportions or case-control sample size, expressed as effective sample sizes (as twice the harmonic mean of the case and control populations) for binary traits and total sample sizes for quantitative traits. Female-to-male ratios of discovery sample size were then compared with corresponding PRS performance ratios using regression analyses. If there are multiple scores for a trait, the sample size and the PRS performance were averaged. Outliers (Cook\u0026rsquo;s distance \u0026gt; 4/n or studentized residual \u0026gt; 3) were excluded from model fitting but retained in plots.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eSex Stratified and Sex Agnostic GWAS Analysis\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eTo generate GWAS summary statistics for downstream dissection of sex differences in polygenic risk scores, we conducted both sex-agnostic and sex-stratified genome-wide association studies (GWAS) using imputed genotype data from the UK Biobank. Analyses were performed with REGENIE (v3.2.9)\u003csup\u003e27\u003c/sup\u003e following its standard two-step procedure.\u003c/p\u003e\n\u003cp\u003eStep 1 fitted whole-genome regression models under a leave-one-chromosome-out (LOCO) scheme using high-quality genotyped variants: minor allele frequency (MAF) \u0026gt; 1%, minor allele count (MAC) \u0026gt; 100, genotyping call rate \u0026gt; 99%, Hardy\u0026ndash;Weinberg equilibrium P \u0026gt; 1\u0026times;10⁻\u0026sup1;⁵, missingness \u0026lt; 10%, and linkage disequilibrium (LD) pruning using 1,000-variant windows, 100-variant sliding windows, and an r\u0026sup2; threshold \u0026lt; 0.8.\u003c/p\u003e\n\u003cp\u003eIn the second step, single-variant association tests were performed using linear regression with LOCO-based predictions included as offsets. Sex-agnostic GWAS adjusted for age, age\u0026sup2;, sex, assessment center, genotyping array, and 10 ancestry PCs; sex-stratified GWAS applied the same covariates, excluding sex.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAssessment of Methodological Differences in PRS Construction\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eTo evaluate whether the observed sex differences in PRS were influenced by the algorithms used to derive PRS, we analyzed 58 sex-biased traits using a sex-agnostic GWAS in the UK Biobank training set, using the same training, tuning, and testing partitions as in the primary analyses.\u003c/p\u003e\n\u003cp\u003ePRS were then constructed using seven widely adopted algorithms: clumping and thresholding (P+T), LDpred2, Lassosum2, SBayesR, SBLUP, SDPR, and PRS-CS\u003csup\u003e26,28-36\u003c/sup\u003e. Method parameters are provided in Supplementary Table 6. Linkage disequilibrium (LD) reference panels were derived from 503 unrelated European individuals from the 1000 Genomes Project Phase 3 dataset\u003csup\u003e37\u003c/sup\u003e (MAF \u0026gt; 1%), ensuring consistency across all methods. Only HapMap3 variants were included in PRS construction\u003csup\u003e38\u003c/sup\u003e.\u003c/p\u003e\n\u003cp\u003eFor each trait and method, the parameters of all candidate PRS were tuned in the tuning subset and evaluated in the testing dataset. Consistency of sex-bias estimates across methods was assessed by intraclass correlation (\u003cem\u003eICC\u003c/em\u003es) with 1,000-bootstrap 95% CIs using the pingouin package (v0.5.4)\u003csup\u003e39\u003c/sup\u003e in Python (v3.10). \u0026nbsp;\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eHeritability Estimation\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eSNP-based heritability (h\u003csup\u003e2\u003c/sup\u003e) was estimated using linkage disequilibrium score regression (LDSC)\u003csup\u003e40\u003c/sup\u003e applied to GWAS summary statistics, and calculated as\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e\u003cimg src=\"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAKgAAABVCAYAAADdR69zAAAAAXNSR0IArs4c6QAAAARnQU1BAACxjwv8YQUAAAAJcEhZcwAAFiUAABYlAUlSJPAAAAh9SURBVHhe7Z2/b9POG8ff/e5Q3O4MdReEEMu5SKggdaAuzKBkZ6hiGFERqRgYqJqyIVTSAYkByWQCVWrVwtChFUJKhOwBsdQVA6NNBP/A8x0+OSs+p7QlcXo2z0s6qXnu1Nbx+348zz0+jxARgWE05X+qgWF0ggXKaA0LlNEaFiijNSxQRmtYoIzWsEAZrWGBMlrDAmW0hgXKaA0LlNEaFiijNSxQRmtYoIzWsEAZrWGBFoi9vT04joOxsTGMjIxgcnISKysrarNcwQItENeuXUOz2USr1UIYhhBC4OHDh/B9X22aG1igBePJkyeYmJjA+Pg47t+/DwDY3t5Wm+UGFmiBICLcunUr/vz7928AgG3bXa3yxQg/k1RMoijClStXMDs7i9XVVbU6N7BAC0gURZibm4NlWbkWJ3iKLx5FEid4BC0elmUBALa2tjA+Pq5W5w4eQTVnc3MTjuPAsiyMjIykyubmZtx2cXERUMS5traW71goMVoShiHZtk2GYVCtVqMgCMh1XTJNk8IwVJtTEAQEoGep1Wpq89zAAtWQMAxJCEGGYZDnebHd8zwCQPV6PdG+yPAUryGvXr1Cq9XC+vo6Ll++HNvlz+/evetqXWxYoBqyvLwMIQSmp6cT9r29vcTnfwEWqGbs7e2h3W7jzp07ahV+/PgBAJiYmFCrCgsLNEe8f/8eAHD9+nW1CisrK3EWk2VZKJfL+fbeO7BANePMmTOqCQDg+z4ajQaEECiXy4k6x3GwvLyM9fV1EBEsy0Kj0cDVq1cT7XKJ6jUxp48QgkzTpCAIiIhoY2ODDMNIefVERK7rEgDa3d2NbbVajQD0DEflDe0EGoYh1et1EkLEcbxSqZS6MUUmDEOqVCpkGAYBINM0qVKp9BScEIKEEAlbpVJJ2fKKdgKVvV/G+nZ3d2ORMml6BeKloIvAUNagKysr8dbc5OSkWp1CCIH5+XkAwPT0NGzbRqPRUJsxPVhbW0MQBIn4aZ4ZikAXFhbgui4AYHZ2Vq1OsLCwgGazmbBFUYRSqZSwMf9hGAa+fPkCdMT569cvAMDFixeVlvlkKAIFgLNnzwJduyHHRY4Ijx49UqsYAKurq2g0GhgbG8Po6ChGR0eBzsxTCNQ5Pyvk2vIkzo7ruj09V+a/78a2bdVcqPUnDdNJsm2bTtIfWJx/xrZtEkLEnn0QBLHnX6Tv7PiK6RPDMHr2+F5sbGwQAHJdV61iOnieF3f6IofjhiJQmSZWq9USMT7DMFIiDIIgZQ/DkEzTTLRj/g2GIlC521Gv1+NeLkWnTvuVSiUxKnSX46KOLMcpaiyR0YPj3/U+kKLr3r6jv1iXMv8eQ1GH3Lbc2NhI2A3DyM3UrY64XA4vgyTzOGgURWi1WjBNM3Hqhe/7aLfbRwbudaHTmbkcowySzAX67ds3AEjtBH3+/Bk4JLexX+bm5lJPPx5VipA7WUQyF+inT58AAJcuXUrYd3Z2AABTU1MJu6SfBNytra1Urz6qLCwsqL+G0QF1zh800hHqdo7oiPWnDEPJHEfpZHXnPOYN13WpVCrF6zQhRGpNngeCIKBqtRpHYAzDODQVcBBkLlAZ71SRN4k626Ayva6ICbgyZVDeyCAIyDTNnt+L7shn9eX9UdMjB02mApUB+l65nNVqldAJPXVfXBETcKVAuzuYvLF5mxVs26ZqtZqwATj2LuFJyVSgfwN6BM2LlgBBnQ6axxFUJQzDTEfQzJ2kfilaAi46IbalpaVCnD537949CCFw+/ZttWogaCfQoifg+r6PmZkZuK6bejozbziOgyAIsj1JTx1STxvpJMmEkXq9Thr+m3+F53mpRJh+6LUcGhbSL8jacdXmzhc9AVdmaQ3yWk5LoHL9PIzUPm2m+NevXyOKIkRRBAA4ODiA4zj4+fNn/ABdnimVSiiXy4l1Z7lczt15S5ubm3j58iV2dnZiv8D3fczNzalNB4Oq2NMibwm4YRjGwXcZtFaLnP7ksqVX6SfMNKgR1PM8qlarqe9flu6w0mHX2mv2GwTaCDRPeJ5HpmmSaZrkum6chN0r3pslgxCojEdXKhXyPC8+CFeXgYEFekKko6M6CNKZG+aN7VegcgtZddoMwxh6ZzsMFugJkVt9qvcqd4v6EcxhqNPpccpRS4c//b86JZJr4yTlAd/3sb29jUqlkor7yaytLOgMJKkCALVaLWUnoiOfi3/x4gUA4O7du2qVVq9OZIGeAHnjbt68qVbh+/fvAIDz58+rVVrSaDRg23aqox0cHAAATNNM2E8LFuiAePv2LdAjvzWKIjiOEydGl8tlWJalbXjp48ePQI8jiuR1yBzdsbGx7EJLXbBAT4A8VkZlbW0N7XYb1Wo1cTy3fOtbs9lEEAQIwxBBEKDVah05BWeNYRiqCVEU4dmzZzAMAw8ePEjUPX78GB8+fECr1QIR4c2bN9jf30+0yQR1UcocjtwNKpVKFIYhhZ2zTNHJbVUdJ5l43W2XJ4IMgsOcnOMgw0vSmfI8L364UfXqqZMeqD6VOwxYoCfE87xUZnwvkcg0NDV30jTNlO1v6Ueg1CU6dHIf/rQxEgRBLOB+/+5JYIFmhAzjdId7ZBC81wilM/II8nq9npolsobXoEOkWq0CAC5cuKBWac3z588xNTWF+fn5lNefNSzQjJBv6/j69SuiKMLi4mIcuslb8vW5c+ewv78fh6B834fjOIkX2WaGOqQyg6P7yB+5ds0qqSJLeq27h7VM4ffFZ4B8LfbTp09j28HBAUzTLEQm/TDhKT4DlpaW0G6348++76NUKkEIgRs3biTaMn+GBZoBruui2WzGu0czMzOwLCvbZ3cKCk/xjNbwCMpoDQuU0RoWKKM1LFBGa1igjNawQBmtYYEyWsMCZbSGBcpoDQuU0Zr/A4xLr+V9DiSRAAAAAElFTkSuQmCC\" width=\"168\" height=\"85\"\u003e\n \u003cv:shapetype id=\"_x0000_t75\" coordsize=\"21600,21600\" o:spt=\"75\" o:preferrelative=\"t\" path=\"m@4@5l@4@11@9@11@9@5xe\" filled=\"f\" stroked=\"f\"\u003e\n \u003cv:stroke joinstyle=\"miter\"\u003e\n \u003cv:path o:extrusionok=\"f\" gradientshapeok=\"t\" o:connecttype=\"rect\"\u003e\u0026nbsp;\u003c/v:path\u003e\n \u003c/v:stroke\u003e\n \u003c/v:shapetype\u003e\n \u003cv:shape id=\"_x0000_i1025\" type=\"#_x0000_t75\"\u003e\u0026nbsp;\u003cv:imagedata src=\"file:///C%3A/Users/pgs9865/AppData/Local/Temp/msohtmlclip1/01/clip_image001.png\" o:title=\"\" chromakey=\"white\"\u003e\u0026nbsp;\u003c/v:imagedata\u003e\n \u003c/v:shape\u003e\n\u003c/p\u003e\n\u003cp\u003ewhere,\u0026nbsp;\u003cimg src=\"data:image/png;base64,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\" width=\"31\" height=\"40\"\u003e and \u003cv:shape id=\"_x0000_i1025\" type=\"#_x0000_t75\"\u003e\u0026nbsp;\u003cimg src=\"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAACEAAAAlCAYAAADMdYB5AAAAAXNSR0IArs4c6QAAAARnQU1BAACxjwv8YQUAAAAJcEhZcwAAFiUAABYlAUlSJPAAAAIDSURBVFhH7Zg9juowFEa/vB4hi56ClLQJFT1C1MhsIStAYgfZAawgWUE0qSNRmwUEsQAjxAq+V7xxhE0Iw5/mFTmSi1xfKcey73UUjyTxy/xxA79BK2FoJQythKGVMLxdIk1TLBYLeJ4Hz/MQhiGyLHPTbPhGiqIgAEZRRK01y7Kk7/sUQripFh+R0FpXsTiOCYBFUVi5l7x1O8bjMUii1+tVsfP5DCEExuOxlWvhWr0TpRQBMEkSd8riYxJKKQoh7grwUxKPCPATEmVZUgjBKIrcqZvcldBaM0kSSinp+z4BXI3LagiC4EpAStlYHY0SSin6vk/f95kkCbXWjKKIUko3lSSZJMmVoBlPSZh9DYLAWul6vSYAKqWs/Fe42SeWyyUA4Ovry6r74XAIAMjzvIq9Sq3EbrdDnueIosgSAIDtdms9v4NaCbPK6XTqTuFwOAAA+v2+O/U0tRJNpGkKABiNRlY8yzKEYWjdnib3HrUS3W7XDQEANpsNTqcTVqsVBoNBFd/v95jNZpjP5yAJrfXVNjbinlReNBwpJbXW1FpXVeFWC797iRCCcRxb8Z9SK8HvEpVSVnUeBEHjS4ykGY/wWPYNJpMJgyB4une8LGE+ZJo64j1qD+YjdDod4KJ/HI/H6jvzx7hWzxDHcXW5mRu0LEs37SYe2/8T/2glDP+FxF+Z9NS1g/XxcwAAAABJRU5ErkJggg==\" width=\"33\" height=\"37\"\u003e\n \u003cv:imagedata src=\"file:///C%3A/Users/pgs9865/AppData/Local/Temp/msohtmlclip1/01/clip_image003.png\" o:title=\"\" chromakey=\"white\"\u003e\u0026nbsp;\u003c/v:imagedata\u003e\n \u003c/v:shape\u003e are the estimates of the additive genetic and residual variance. Analyses were restricted to well-imputed HapMap3 variants. Chi-square statistics were regressed on LD scores computed from the 1000 Genomes European reference panel. Heritability estimates for binary traits were transformed from the observed scale to the liability scale, using the sex-specific population prevalence of the trait, under the assumption of an underlying normal distribution of liability to the considered trait, as described previously.\u003csup\u003e41\u003c/sup\u003e\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eStatistical Analysis\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eAll analyses were conducted under a prespecified, uniform analytic framework to ensure consistency across PGS Catalog, ensemble, and de novo PRS analyses. For binary outcomes, logistic regression models were fitted with standardized PRSs as predictors. For quantitative traits, linear regression was applied. Covariates included age, sex, assessment center, genotyping array, and the first 10 principal components of ancestry, with sex excluded in sex-stratified analyses. To minimize confounding from population structure, all PRSs were residualized by the first 10 genetic principal components before association testing.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eSex differences in predictive performance were quantified using a two-sided Z test comparing sex-specific regression coefficients and their standard errors:\u003c/p\u003e\n\u003cp\u003e\u003cimg src=\"data:image/png;base64,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\" width=\"226\" height=\"112\"\u003e\u003c/p\u003e\n\u003cp\u003eThis test evaluates the null hypothesis that the predictive effects of PRS are equal between sexes (H₀: \u003cem\u003e\u0026beta;\u003c/em\u003e\u003csub\u003efemale\u003c/sub\u003e = \u003cem\u003e\u0026beta;\u003c/em\u003e\u003csub\u003emale\u003c/sub\u003e) against the alternative hypothesis that they differ (H₁: \u003cem\u003e\u0026beta;\u003c/em\u003e\u003csub\u003efemale\u003c/sub\u003e \u0026ne; \u003cem\u003e\u0026beta;\u003c/em\u003e\u003csub\u003emale\u003c/sub\u003e). This procedure was applied consistently across all analytic stages, including PGS Catalog scores, ensemble trait-level PRSs, and UK Biobank\u0026ndash;derived PRSs constructed from multiple algorithms.\u003c/p\u003e\n\u003cp\u003eFor secondary comparisons of genetic parameters between male- and female-biased traits, we used the Mann\u0026ndash;Whitney U test to evaluate whether the median difference between groups differed significantly from zero (\u003cem\u003eH\u003c/em\u003e\u003cem\u003e₀\u003c/em\u003e: \u0026Delta; = 0; \u003cem\u003eH\u003c/em\u003e\u003cem\u003e₁\u003c/em\u003e: \u0026Delta; \u0026ne; 0).\u003c/p\u003e\n\u003cp\u003eSex bias in predictive performance was represented as the difference in these metrics between females and males. Multiple testing correction was performed using the Benjamini\u0026ndash;Hochberg false discovery rate (FDR) procedure, with \u003cem\u003eFDR\u003c/em\u003e-adjusted \u003cem\u003ep\u003c/em\u003e \u0026lt; 0.05 considered statistically significant. All statistical tests were two-sided. Regression modeling and data visualization were performed in R version 4.1.2 (R Foundation for Statistical Computing) with the bigsnpr\u003csup\u003e32\u003c/sup\u003e, ggplot2\u003csup\u003e42\u003c/sup\u003e, and data.table packages. Analyses reporting include point estimates and 95% confidence intervals for all primary effect size comparisons.\u003c/p\u003e"},{"header":"Results","content":"\u003cp\u003e\u003cstrong\u003eWidespread Sex Bias in Polygenic Risk Score Performance\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eUsing PRS from the PGS Catalog applied to the UK Biobank, we observed widespread and trait-specific sex differences in polygenic risk prediction across both diseases and quantitative traits. Figure 1 summarizes the distribution of sex bias across scores (left) and traits (right). Among 1295 disease scores, 41.2% showed significant sex bias (193 female-biased; 341 male-biased). \u0026nbsp; Among 1968 scores for quantitative traits, 73.2% were biased (728 female-biased and 713 male-biased). Aggregated at the trait level, 15 (23.4%) of 64 disease traits and 43 (53.1%) of 81 quantitative traits displayed significant sex differences in performance, respectively.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eTo further characterize the magnitude and direction of these effects, we extended the trait-level summary to display effect size estimates and 95% confidence intervals for each trait (Figure 2). This analysis revealed marked heterogeneity across traits, with female-favoring prediction observed for autoimmune and endocrine diseases such as hypothyroidism and type 1 diabetes, whereas male-favoring prediction was evident for cardiovascular diseases, including coronary artery disease and atrial fibrillation. Other traits, such as inguinal hernia, are also male-biased. Among quantitative traits, hormone- and body composition\u0026ndash;related measures such as estradiol, testosterone, and body fat showed the strongest sex divergence, typically favoring the biologically corresponding sex, while most hematologic and metabolic biomarkers exhibited smaller but consistent male advantage.\u003c/p\u003e\n\u003cp\u003eOverall, these findings demonstrated that sex differences in PRS performance are both pervasive and trait-specific, affecting a wide spectrum of diseases and quantitative phenotypes. Full per-score and per-trait statistics, including confidence intervals and significance levels, are provided in Supplementary Table 3 and Supplementary Table 4.\u003c/p\u003e\n\u003cp\u003e\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eSample Size Composition of GWAS Contributes to Sex Bias\u0026nbsp;\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eTo assess whether imbalanced sex representation in the discovery GWAS used to derive PRS \u0026nbsp;contributes to sex-related differences in PRS predictive performance, we examined the relationship between female-to-male ratios of GWAS sample sizes and corresponding ratios of PRS predictive performance for each trait (Figure 3). For disease traits, a positive correlation was observed (\u003cem\u003e\u0026beta;\u003c/em\u003e = 0.0646, \u003cem\u003eP\u003c/em\u003e-value = 3.08 \u0026times; 10⁻\u0026sup3;, \u003cem\u003eR\u003c/em\u003e\u0026sup2; = 0.36), indicating that PRS derived from a larger female GWAS sample size tended to perform better in females. In contrast, quantitative traits showed no meaningful association (\u003cem\u003e\u0026beta;\u003c/em\u003e = 0.169, \u003cem\u003eP\u003c/em\u003e-value = 0.60, \u003cem\u003eR\u003c/em\u003e\u0026sup2; = 0.006). These findings suggest that sample-size imbalance in GWAS discovery cohorts partially drives the PRS prediction bias between sexes, consistent with the expectation that larger sample sizes yield more accurate estimates of allele effects, thereby improving prediction power.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eConsistency of Sex Bias Across PRS Methods\u0026nbsp;\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eBuilding on the PGS Catalog-wide characterization of sex differences, and complementary to the analysis of discovery sample composition, we evaluated whether the observed biases depend on the choice of PRS algorithm. We focused on 58 traits that were significantly sex-biased in the PGS Catalog analysis and generated polygenic scores using 7 widely used methods.\u003c/p\u003e\n\u003cp\u003eAcross methods, trait-level estimates were relatively concordant (Figure 4).\u0026nbsp;Patterns in the trait-by-method matrix clustered by trait rather than by method, and no algorithm consistently amplified or attenuated female- or male-favoring effects. We used the Intraclass Correlation Coefficient (ICC) to assess consistency across methods. Agreement across the 7 method-specific estimates was high, with an intraclass correlation of \u003cem\u003eICC\u003c/em\u003e(2,k)=0.93 (95% CI, 0.90\u0026ndash;0.96; P\u0026lt;.001), and virtually identical results for \u003cem\u003eICC\u003c/em\u003e(1,k) and \u003cem\u003eICC\u003c/em\u003e(3,k)\u0026nbsp;(Supplementary Table\u0026nbsp;8).\u003c/p\u003e\n\u003cp\u003eThese findings indicate that sex-differential PRS performance is robust to algorithmic choice and thus more likely reflects trait-specific biology rather than method artifacts. Detailed performance metrics and statistical results are provided in Supplementary Table\u0026nbsp;7.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e\u0026nbsp;\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003ePotential Role of Genetic Architecture in Shaping Sex Differences\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eTo examine whether sex differences in PRS performance reflect underlying differences in genetic architecture, we conducted sex-stratified GWAS for traits that exhibited significant sex bias in PRS performance, including 15 disease traits and 43 quantitative traits. Using sex-specific GWAS summary statistics from the UK Biobank, heritability was estimated with linkage disequilibrium score regression (LDSC)\u003c/p\u003e\n\u003cp\u003eWe observed that the magnitude of sex differences in PRS performance was significantly correlated with the gap in SNP-based heritability between sexes (Figure 5). Among disease and quantitative traits, the correlation was strong. The Pearson correlation was 0.81 (P \u0026lt; 0.001) and 0.58 (P \u0026lt; 0.001) for disease and quantitative traits, respectively. Notably, several traits, such as type 1 diabetes, deviated from this overall trend, suggesting that additional mechanisms beyond differences in total heritability may contribute to sex-specific prediction disparities.\u003c/p\u003e\n\u003cp\u003eTo further assess whether overall heritability patterns differ systematically between male- and female-biased traits, we compared the distributions of heritability difference across traits classified by PRS bias direction. Traits whose PRSs performed better in males exhibited significantly higher male-to-female heritability than those favoring females (P = 1.312 \u0026times; 10⁻\u0026sup3;; Figure 5C). This result supports the notion that, on average, higher SNP-based heritability corresponds to stronger PRS predictive performance, consistent with sex-dependent differences in polygenic architecture.\u003c/p\u003e\n\u003cp\u003eCollectively, these findings indicate that sex-dependent genetic architecture is a strong driver for sex differences in PRS performance.\u003c/p\u003e"},{"header":"Discussion","content":"\u003cp\u003eThis study demonstrates that sex differences in PRS \u0026nbsp;performance are both widespread and biologically grounded. Prior PRS implementation studies have noted isolated instances of sex differences in risk stratification\u003csup\u003e7,18,43\u003c/sup\u003e, yet none have systematically quantified their prevalence across thousands of polygenic scores and diverse trait categories. Our analysis provides population-level evidence that sex disparities are both common and trait-specific, emphasizing the need for sex-aware evaluation before clinical deployment.\u003c/p\u003e\n\u003cp\u003eLeveraging more than 3,000 publicly available PRS and newly generated scores using published GWAS and multiple methods, we found that disparities between males and females were common across diverse traits and partly attributable to an imbalance in discovery GWAS sample composition, while algorithmic choice had minimal influence. Autoimmune and endocrine conditions tended to show stronger prediction in females, whereas cardiometabolic conditions favored males, mirroring well-established patterns of sexual dimorphism. This is aligned with the previous study.\u003csup\u003e7,18,43\u003c/sup\u003e Notably, the association between GWAS sample size composition and sex differences in PRS performance was primarily observed for disease traits but not for quantitative traits. One possible explanation is that, for disease phenotypes, GWAS sample size may partly reflect underlying disease characteristics such as prevalence and heritability, which influence statistical power for variant discovery and downstream PRS construction.\u003csup\u003e44\u003c/sup\u003e In contrast, the sample size for quantitative traits is typically determined by the availability of phenotype measurements rather than the genetic architecture of the trait itself, which may explain the lack of a comparable relationship. These findings indicate that sex differences in PRS performance are unlikely to arise from analytic artifacts alone and underscore the importance of sex-balanced GWAS designs and sex-aware PRS frameworks.\u003c/p\u003e\n\u003cp\u003eBiological factors emerged as a stronger and more consistent driver. Sex-stratified GWAS demonstrated that differences in PRS performance closely mirrored differences in SNP-based heritability, with stronger correlations for disease traits. Traits with higher heritability in one sex consistently showed stronger PRS prediction in that sex, pointing to systematic differences in polygenic architecture. These observations align with prior evidence of sex-dependent heritability, variant-effect heterogeneity, and hormone-related gene regulation, suggesting that sex differences in PRS reflect genuine biological divergence in genetic architecture rather than analytic artifacts.\u003csup\u003e11,20,45\u003c/sup\u003e\u003csup\u003e\u0026nbsp;\u003c/sup\u003eThis convergence across PGS Catalog scores, UK Biobank PRS scores, and sex-stratified GWAS reinforces the biological foundation of sex-biased prediction.\u003c/p\u003e\n\u003cp\u003eThese findings have important implications for clinical translation. As PRS are increasingly incorporated into preventive clinical genomics programs, unrecognized sex differences may lead to systematic misclassification of genetic risk. Female risk for cardiometabolic diseases may be underestimated, while male risk for autoimmune or endocrine disorders may be misestimated, with potential consequences for screening eligibility, preventive therapy allocation, and long-term monitoring. Routine evaluation of sex-stratified predictive effectiveness should therefore become a standard component of PRS validation prior to clinical adoption.\u003c/p\u003e\n\u003cp\u003eOur analysis was limited to autosomal variants in European ancestry participants, and X-chromosome and multi-ancestry analyses are warranted in further studies. Our analysis was also limited by incomplete metadata in the PGS Catalog API, as many studies did not report discovery sample sizes or sex composition, and large-scale manual curation was not feasible. Some underlying GWAS datasets were not publicly available, which may have reduced the completeness of our sample-size analysis.\u003c/p\u003e\n\u003cp\u003eOverall, this study provides the most comprehensive evaluation to date of sex-differential PRS performance across complex traits and demonstrates that these differences are pervasive and largely biologically driven. Addressing sex-dependent genetic architecture will be essential for fair and clinically meaningful implementation of polygenic risk scores.\u003c/p\u003e"},{"header":"Declarations","content":"\u003cp\u003e\u003cstrong\u003eConflict of Interest Disclosures\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eDr. Natarajan reports research grants from Allelica, Amgen, Apple, Boston Scientific, Cleerly, Genentech / Roche, Ionis, Novartis, and Silence Therapeutics, personal fees from AIRNA, Allelica, Amgen, Apple, AstraZeneca, Bain Capital, Blackstone Life Sciences, Bristol Myers Squibb, Creative Education Concepts, CRISPR Therapeutics, Eli Lilly \u0026amp; Co, Esperion Therapeutics, Foresite Capital, Foresite Labs, Genentech / Roche, GV, HeartFlow, Incyte, Magnet Biomedicine, Merck, Novartis, Novo Nordisk, TenSixteen Bio, Tourmaline Bio, and Ursa Medicines, equity in Bolt, Candela, Mercury, MyOme, Parameter Health, Preciseli, and TenSixteen Bio, royalties from Recora for intensive cardiac rehabilitation, and spousal employment at Vertex Pharmaceuticals, all unrelated to the present work. Dr. Fahed reports being co-founder of Goodpath and Avigena, serving as scientific advisor to MyOme, Arboretum Health, HeartFlow, and Aditum Bio and receiving sponsored research awards from Foresite, Sarepta Therapeutics, and Allelica, all unrelated to the current work. Dr. Sui reports serving as a consultant for Arboretum Lifesciences.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAcknowledgments\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eDr. Wang is supported by the National Natural Science Foundation of China (grant number 82470352), the Noncommunicable Chronic Diseases-National Science and Technology Major Project-2023ZD0503201, and Pioneering Action Grants of the Chinese Academy of Sciences. Dr. Natarajan is supported by grants from NHLBI (R01HL127564), NHGRI (U01HG011719), and Massachusetts General Hospital (Paul and Phyllis Fireman Endowed Chair in Vascular Medicine). Dr. Fahed receives funding from the National Heart Lung and Blood Institute under award numbers K08 HL161448 and R01 HL164629. Dr. Sui is supported by the TOPMed fellowship from the National Heart Lung and Blood Institute. Dr. Halford is supported by a grant from NHGRI (5T32HG010464-07).\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eData Availability\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eIndividual-level data from the UK Biobank are available to qualified researchers through a formal application (https://www.ukbiobank.ac.uk). Summary-level GWAS statistics and PGS weights used in this study are publicly accessible from the PGS Catalog (https://www.pgscatalog.org). All derived PRS performance metrics and summary statistics are available upon reasonable request to the corresponding author.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eCode Availability\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eCode used for polygenic risk score construction, evaluation, and statistical analyses in this study is publicly available at https://github.com/xinyu-c9/SexdiffPRS.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\n\u003cli\u003eWray, N. R.\u003cem\u003e et al.\u003c/em\u003e From basic science to clinical application of polygenic risk scores: a primer. \u003cstrong\u003e78\u003c/strong\u003e, 101-109 (2021).\u003c/li\u003e\n\u003cli\u003eKhera, A. V.\u003cem\u003e et al.\u003c/em\u003e Genome-wide polygenic scores for common diseases identify individuals with risk equivalent to monogenic mutations. \u003cstrong\u003e50\u003c/strong\u003e, 1219-1224 (2018).\u003c/li\u003e\n\u003cli\u003eKhan, A.\u003cem\u003e et al.\u003c/em\u003e Genome-wide polygenic score to predict chronic kidney disease across ancestries. \u003cstrong\u003e28\u003c/strong\u003e, 1412-1420 (2022).\u003c/li\u003e\n\u003cli\u003eLiu, W., Zhuang, Z., Wang, W., Huang, T. \u0026amp; Liu, Z. J. F. i. g. An improved genome-wide polygenic score model for predicting the risk of type 2 diabetes. \u003cstrong\u003e12\u003c/strong\u003e, 632385 (2021).\u003c/li\u003e\n\u003cli\u003eMartin, A. R.\u003cem\u003e et al.\u003c/em\u003e Clinical use of current polygenic risk scores may exacerbate health disparities. \u003cstrong\u003e51\u003c/strong\u003e, 584-591 (2019).\u003c/li\u003e\n\u003cli\u003eWand, H.\u003cem\u003e et al.\u003c/em\u003e Improving reporting standards for polygenic scores in risk prediction studies. \u003cstrong\u003e591\u003c/strong\u003e, 211-219 (2021).\u003c/li\u003e\n\u003cli\u003eEUGenMed\u003cem\u003e et al.\u003c/em\u003e Gender in cardiovascular diseases: impact on clinical manifestations, management, and outcomes. \u003cstrong\u003e37\u003c/strong\u003e, 24-34 (2016).\u003c/li\u003e\n\u003cli\u003eVaccarino, V.\u003cem\u003e et al.\u003c/em\u003e Sex differences in mortality after acute myocardial infarction: changes from 1994 to 2006. \u003cstrong\u003e169\u003c/strong\u003e, 1767-1774 (2009).\u003c/li\u003e\n\u003cli\u003eSokka, T.\u003cem\u003e et al.\u003c/em\u003e Women, men, and rheumatoid arthritis: analyses of disease activity, disease characteristics, and treatments in the QUEST-RA study. \u003cstrong\u003e11\u003c/strong\u003e, 1-12 (2009).\u003c/li\u003e\n\u003cli\u003eAltemus, M., Sarvaiya, N. \u0026amp; Epperson, C. N. J. F. i. n. Sex differences in anxiety and depression clinical perspectives. \u003cstrong\u003e35\u003c/strong\u003e, 320-330 (2014).\u003c/li\u003e\n\u003cli\u003eBernabeu, E.\u003cem\u003e et al.\u003c/em\u003e Sex differences in genetic architecture in the UK Biobank. \u003cstrong\u003e53\u003c/strong\u003e, 1283-1289 (2021).\u003c/li\u003e\n\u003cli\u003eOber, C., Loisel, D. A. \u0026amp; Gilad, Y. J. N. R. G. Sex-specific genetic architecture of human disease. \u003cstrong\u003e9\u003c/strong\u003e, 911-922 (2008).\u003c/li\u003e\n\u003cli\u003eUddin, M. S.\u003cem\u003e et al.\u003c/em\u003e Estrogen signaling in Alzheimer\u0026rsquo;s disease: molecular insights and therapeutic targets for Alzheimer\u0026rsquo;s dementia. \u003cstrong\u003e57\u003c/strong\u003e, 2654-2670 (2020).\u003c/li\u003e\n\u003cli\u003eKrieger, N. J. A. r. o. p. h. Measures of Racism, Sexism, Heterosexism, and Gender Binarism for Health Equity Research: From Structural Injustice to Embodied Harm-An Ecosocial Analysis. \u003cstrong\u003e41\u003c/strong\u003e, 37-62 (2019).\u003c/li\u003e\n\u003cli\u003eSchroeder, P. H., Cole, J. B., Leong, A., Florez, J. C. \u0026amp; Mercader, J. M. J. D. Large-scale sex-stratified additive and recessive GWAS identifies novel large-effect variants and improves polygenic prediction for type 2 diabetes. \u003cstrong\u003e71\u003c/strong\u003e (2022).\u003c/li\u003e\n\u003cli\u003eDumitrescu, L.\u003cem\u003e et al.\u003c/em\u003e Sex differences in the genetic predictors of Alzheimer\u0026rsquo;s pathology. \u003cstrong\u003e142\u003c/strong\u003e, 2581-2589 (2019).\u003c/li\u003e\n\u003cli\u003ePirastu, N.\u003cem\u003e et al.\u003c/em\u003e Genetic analyses identify widespread sex-differential participation bias. \u003cstrong\u003e53\u003c/strong\u003e, 663-671 (2021).\u003c/li\u003e\n\u003cli\u003ePatel, A. P.\u003cem\u003e et al.\u003c/em\u003e A multi-ancestry polygenic risk score improves risk prediction for coronary artery disease. \u003cstrong\u003e29\u003c/strong\u003e, 1793-1803 (2023).\u003c/li\u003e\n\u003cli\u003eVan Alten, S., Domingue, B. W., Galama, T. \u0026amp; Marees, A. T. J. M. Reweighting the UK Biobank to reflect its underlying sampling population substantially reduces pervasive selection bias due to volunteering. 2022.2005. 2016.22275048 (2022).\u003c/li\u003e\n\u003cli\u003eKhramtsova, E. A., Davis, L. K. \u0026amp; Stranger, B. E. J. N. R. G. The role of sex in the genomics of human complex traits. \u003cstrong\u003e20\u003c/strong\u003e, 173-190 (2019).\u003c/li\u003e\n\u003cli\u003eVassy, J. L.\u003cem\u003e et al.\u003c/em\u003e The GenoVA study: Equitable implementation of a pragmatic randomized trial of polygenic-risk scoring in primary care. \u003cstrong\u003e110\u003c/strong\u003e, 1841-1852 (2023).\u003c/li\u003e\n\u003cli\u003eCross, B., Turner, R. \u0026amp; Pirmohamed, M. J. F. i. g. Polygenic risk scores: An overview from bench to bedside for personalised medicine. \u003cstrong\u003e13\u003c/strong\u003e, 1000667 (2022).\u003c/li\u003e\n\u003cli\u003eBycroft, C.\u003cem\u003e et al.\u003c/em\u003e The UK Biobank resource with deep phenotyping and genomic data. \u003cstrong\u003e562\u003c/strong\u003e, 203-209 (2018).\u003c/li\u003e\n\u003cli\u003eManichaikul, A.\u003cem\u003e et al.\u003c/em\u003e Robust relationship inference in genome-wide association studies. \u003cstrong\u003e26\u003c/strong\u003e, 2867-2873 (2010).\u003c/li\u003e\n\u003cli\u003eLambert, S. A.\u003cem\u003e et al.\u003c/em\u003e Enhancing the Polygenic Score Catalog with tools for score calculation and ancestry normalization. \u003cstrong\u003e56\u003c/strong\u003e, 1989-1994 (2024).\u003c/li\u003e\n\u003cli\u003eChang, C. C.\u003cem\u003e et al.\u003c/em\u003e Second-generation PLINK: rising to the challenge of larger and richer datasets. \u003cstrong\u003e4\u003c/strong\u003e, s13742-13015-10047-13748 (2015).\u003c/li\u003e\n\u003cli\u003eMbatchou, J.\u003cem\u003e et al.\u003c/em\u003e Computationally efficient whole-genome regression for quantitative and binary traits. \u003cstrong\u003e53\u003c/strong\u003e, 1097-1103 (2021).\u003c/li\u003e\n\u003cli\u003eGe, T., Chen, C.-Y., Ni, Y., Feng, Y.-C. A. \u0026amp; Smoller, J. W. J. N. c. Polygenic prediction via Bayesian regression and continuous shrinkage priors. \u003cstrong\u003e10\u003c/strong\u003e, 1-10 (2019).\u003c/li\u003e\n\u003cli\u003eLloyd-Jones, L. R.\u003cem\u003e et al.\u003c/em\u003e Improved polygenic prediction by Bayesian multiple regression on summary statistics. \u003cstrong\u003e10\u003c/strong\u003e, 1-11 (2019).\u003c/li\u003e\n\u003cli\u003eMak, T. S. H., Porsch, R. M., Choi, S. W., Zhou, X. \u0026amp; Sham, P. C. J. G. e. Polygenic scores via penalized regression on summary statistics. \u003cstrong\u003e41\u003c/strong\u003e, 469-480 (2017).\u003c/li\u003e\n\u003cli\u003ePriv\u0026eacute;, F., Arbel, J. \u0026amp; Vilhj\u0026aacute;lmsson, B. J. J. B. LDpred2: better, faster, stronger. \u003cstrong\u003e36\u003c/strong\u003e, 5424-5431 (2020).\u003c/li\u003e\n\u003cli\u003ePriv\u0026eacute;, F., Aschard, H., Ziyatdinov, A. \u0026amp; Blum, M. G. J. B. Efficient analysis of large-scale genome-wide data with two R packages: bigstatsr and bigsnpr. \u003cstrong\u003e34\u003c/strong\u003e, 2781-2787 (2018).\u003c/li\u003e\n\u003cli\u003ePriv\u0026eacute;, F., Vilhj\u0026aacute;lmsson, B. J., Aschard, H. \u0026amp; Blum, M. G. J. T. A. J. o. H. G. Making the most of clumping and thresholding for polygenic scores. \u003cstrong\u003e105\u003c/strong\u003e, 1213-1221 (2019).\u003c/li\u003e\n\u003cli\u003eRobinson, M. R.\u003cem\u003e et al.\u003c/em\u003e Genetic evidence of assortative mating in humans. \u003cstrong\u003e1\u003c/strong\u003e, 1-13 (2017).\u003c/li\u003e\n\u003cli\u003eWray, N. R., Goddard, M. E. \u0026amp; Visscher, P. M. J. G. r. Prediction of individual genetic risk to disease from genome-wide association studies. \u003cstrong\u003e17\u003c/strong\u003e, 1520-1528 (2007).\u003c/li\u003e\n\u003cli\u003eZhou, G. \u0026amp; Zhao, H. J. P. g. A fast and robust Bayesian nonparametric method for prediction of complex traits using summary statistics. \u003cstrong\u003e17\u003c/strong\u003e, e1009697 (2021).\u003c/li\u003e\n\u003cli\u003eNature, G. P. C. J. A global reference for human genetic variation. \u003cstrong\u003e526\u003c/strong\u003e, 68 (2015).\u003c/li\u003e\n\u003cli\u003eNature, I. H. C. J. Integrating common and rare genetic variation in diverse human populations. \u003cstrong\u003e467\u003c/strong\u003e, 52 (2010).\u003c/li\u003e\n\u003cli\u003eVallat, R. J. J. O. S. S. Pingouin: statistics in Python. \u003cstrong\u003e3\u003c/strong\u003e, 1026 (2018).\u003c/li\u003e\n\u003cli\u003eBulik-Sullivan, B. K.\u003cem\u003e et al.\u003c/em\u003e LD Score regression distinguishes confounding from polygenicity in genome-wide association studies. \u003cstrong\u003e47\u003c/strong\u003e, 291-295 (2015).\u003c/li\u003e\n\u003cli\u003eMu\u0026ntilde;oz, M.\u003cem\u003e et al.\u003c/em\u003e Evaluating the contribution of genetics and familial shared environment to common disease using the UK Biobank. \u003cstrong\u003e48\u003c/strong\u003e, 980-983 (2016).\u003c/li\u003e\n\u003cli\u003eWickham, H. J. W. i. r. c. s. ggplot2. \u003cstrong\u003e3\u003c/strong\u003e, 180-185 (2011).\u003c/li\u003e\n\u003cli\u003eZhang, F.\u003cem\u003e et al.\u003c/em\u003e Sex differences in the association between polygenic risk score and atrial fibrillation incidence: a prospective cohort study. (2025).\u003c/li\u003e\n\u003cli\u003eVisscher, P. M.\u003cem\u003e et al.\u003c/em\u003e 10 years of GWAS discovery: biology, function, and translation. \u003cstrong\u003e101\u003c/strong\u003e, 5-22 (2017).\u003c/li\u003e\n\u003cli\u003eWang, X., Magkos, F., Mittendorfer, B. J. T. J. o. C. E. \u0026amp; Metabolism. Sex differences in lipid and lipoprotein metabolism: it\u0026apos;s not just about sex hormones. \u003cstrong\u003e96\u003c/strong\u003e, 885-893 (2011).\u003c/li\u003e\n\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":true,"hideJournal":false,"highlight":"","institution":"","isAcceptedByJournal":false,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"nature-portfolio","isNatureJournal":true,"hasQc":false,"allowDirectSubmit":false,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"","title":"Nature Portfolio","twitterHandle":"","acdcEnabled":false,"dfaEnabled":false,"editorialSystem":"ejp","reportingPortfolio":"","inReviewEnabled":true,"inReviewRevisionsEnabled":false},"keywords":"","lastPublishedDoi":"10.21203/rs.3.rs-9153854/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-9153854/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"Polygenic risk scores (PRS) are increasingly used for disease prediction, yet their performance equity across sexes remains unclear. We evaluated sex differences in PRS performance using 3,263 scores across 145 traits from the PGS Catalog in 409,440 UK Biobank participants. Sex-differential prediction was widespread and trait-specific, affecting 15 of 64 (23%) of diseases and 43 of 81 (53%) of quantitative traits. Female-favoring performance was enriched in autoimmune and endocrine traits, whereas cardiometabolic traits more often favored males. Discovery GWAS sex imbalance partially explained disease-level disparities (R² = 0.36), whereas quantitative traits showed minimal association. Notably, sex differences in predictive performance strongly correlated with differences in SNP-based heritability from sex-stratified GWAS (R² = 0.81 for diseases; R² = 0.58 for quantitative traits). In contrast, PRS estimates were highly consistent across seven construction methods (intraclass correlation coefficient = 0.93), indicating limited methodological influence. These findings demonstrate that sex disparities in PRS performance are common and largely reflect underlying genetic architecture rather than analytic artifacts, highlighting the need for sex-aware GWAS design and PRS modeling to ensure equitable clinical implementation.","manuscriptTitle":"Widespread and Biologically Driven Sex Disparities in Polygenic Risk Prediction Across Complex Traits","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2026-04-29 11:02:08","doi":"10.21203/rs.3.rs-9153854/v1","editorialEvents":[],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"nature-genetics","isNatureJournal":true,"hasQc":false,"allowDirectSubmit":false,"externalIdentity":"ng","sideBox":"Learn more about [Nature Genetics](http://www.nature.com/ng/)","snPcode":"","submissionUrl":"","title":"Nature Genetics","twitterHandle":"","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"ejp","reportingPortfolio":"Nature Research","inReviewEnabled":true,"inReviewRevisionsEnabled":false}}],"origin":"","ownerIdentity":"14ff5cc5-da33-4036-bd1c-0c7e39ad9bc7","owner":[],"postedDate":"April 29th, 2026","published":true,"recentEditorialEvents":[{"type":"editorInvitedReview","content":"This content is not available.","date":"2026-05-11T01:35:06+00:00","index":1,"fulltext":"This content is not available."}],"rejectedJournal":[],"revision":"","amendment":"","status":"under-review","subjectAreas":[{"id":67110306,"name":"Biological sciences/Genetics/Population genetics"},{"id":67110307,"name":"Biological sciences/Genetics/Genetic association study/Genome-wide association studies"},{"id":67110308,"name":"Biological sciences/Genetics/Genomics"}],"tags":[],"updatedAt":"2026-04-29T11:02:08+00:00","versionOfRecord":[],"versionCreatedAt":"2026-04-29 11:02:08","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-9153854","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-9153854","identity":"rs-9153854","version":["v1"]},"buildId":"XKTyCvWXoU3ODBz1xrDgd","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}