Skip to content

Polygenic Risk Scores

Summary

Polygenic Risk Scores (PRSs) aggregate the small effects of thousands to millions of genetic variants across the genome to estimate an individual's genetic predisposition to a specific disease or trait. Developed using genome-wide association study (GWAS) summary statistics, they serve as tools for clinical risk stratification and research in complex trait genetics. However, their clinical utility can be hindered by reliance on large, heterogeneous data sets that dilute phenotypic specificity.

Challenges and Methodological Extensions

Standard PRS generation methods face several challenges when transitioning to clinical practice: - Phenotypic Heterogeneity: Large biobank-scale cohorts (e.g., UK Biobank) often rely on self-reported conditions or broad diagnostic categories, which dilutes the precision of estimated genetic effects for specific clinical subtypes. - Cohort Sample Size: Deeply phenotyped clinical cohorts (e.g., for specific epilepsy syndromes) are highly specific but typically have sample sizes too small to robustly estimate variant effect sizes. - Portability: Scores developed in one population may perform poorly in others due to differing linkage disequilibrium (LD) structures and allele frequencies.

To overcome these barriers, several advanced methods have been developed: - Multiancestry Fine-Mapping: Tools like SuShiE and SuSiEx model ancestry-specific effect sizes to improve prediction across diverse populations. - Transfer Learning: Frameworks like ePRS adapt general PRSs derived from large biobank datasets to improve prediction in small, phenotypically refined target cohorts. - Multimodal Joint Modeling: Frameworks like RisQ integrate PRSs with demographics, clinical biomarkers, lifestyle, and time-course features to learn a unified representation of disease susceptibility across diverse conditions and prediction horizons. - Summary-Statistics Imputation: Tools like SSimp impute GWAS summary statistics at untyped variants directly from tag-SNV summary statistics and reference-panel LD, a faster (if less accurate) alternative to genotype imputation for updating association results ahead of PRS construction.

Disease Subtyping via Partitioned PRS

Beyond aggregate risk prediction, partitioned/component PRS are used to recover genetically distinct disease subtypes — see Polygenic-Score-Based Disease Subtyping for a field overview of the three methodological families (mechanism-first, individual-first, and subtype-prediction) and their application to type 2 diabetes, obesity, MASLD, schizophrenia, and Alzheimer's disease.

Combining PGS with Other Risk Modalities

Beyond stacking PGS with clinical/environmental factors, Polygenic and Gut Metagenomic Risk Integration shows PGS can also be combined with an entirely separate biological modality — the gut microbiome — in a prospective cohort predicting four common diseases; PGS improved prediction over conventional risk factors consistently across all four diseases, while the added gut-microbiome contribution was smaller and disease-specific (clearest for type 2 diabetes and prostate cancer).

Do Deep Learning and Nonlinear Methods Improve PGS?

Two independent studies converge on a consistent, largely negative answer for nonlinear/deep-learning methods specifically, while showing that simpler machine-learning improvements over univariate thresholding can still help:

  • Neural networks vs. linear PGS (Nature Communications, 2025): fitting multi-layer-perceptron neural networks (linear- and nonlinear-activation variants) to 28 real UK Biobank traits (125,000 European-ancestry individuals) and to simulated traits with known additive or purely epistatic genetic architectures, the authors found neural-network PGS were consistently outperformed by LD-aware additive baselines (PGS Catalog scores, or PRS-CS) — median real-data r² was only 87–95% of the additive baseline's, across both SNP-only and SNP+covariate scenarios. Critically, nonlinear networks also outperformed linear networks even in simulations with zero genuine epistasis, because they partially exploit linkage-disequilibrium-driven "joint tagging effects" rather than genuine gene-gene interaction — a confound the authors isolate with a "SNP-dosage weighting" mitigation strategy. Genuine epistasis, once isolated from this confound, contributed only ~10% relative r² even in an unrealistically extreme all-epistatic simulated scenario.
  • Machine learning vs. thresholding for blood cell trait PGS (Xu et al., Cell Genomics companion): comparing 6 PGS-construction methods (pruning+thresholding, LDpred2, elastic net, Bayesian ridge, MLP, CNN) for 26 blood cell traits in UK Biobank/INTERVAL, elastic net and Bayesian ridge — both linear machine-learning methods — consistently outperformed simple pruning+thresholding (external-validation Pearson r improved 10–23% for 25 of 26 traits), primarily by jointly modeling correlated/low-MAF variant effects rather than the hard-cutoff, independent-variant selection of P+T. However, the nonlinear methods (MLP, CNN) did not improve on the linear ML methods, and CNN underperformed even simple P+T for roughly half the studied traits.
  • Convergent conclusion: linear methods that better model variant correlation, LD, and effect-size shrinkage (elastic net, Bayesian ridge, LDpred2) meaningfully improve PGS over naive thresholding, but nonlinear/deep-learning architectures have not yet demonstrated a genuine advantage for PGS construction from SNP data — likely because population-level phenotypic variance is theoretically expected to be dominated by additive effects even when substantial epistasis exists at the individual-variant level, and because SNP data lacks the high signal-to-noise, spatially-structured characteristics that make deep learning succeed in other domains.

Citations