Residency · Residency · Medical Genetics Genomics

Polygenic Risk Scores: Science, Utility, and Limitations

Fundamentals of Polygenic Risk

From Monogenic to Polygenic

Most common diseases -- coronary artery disease, type 2 diabetes, schizophrenia, breast cancer -- are not caused by a single gene but rather influenced by hundreds to thousands of genetic variants, each contributing a small individual effect. A single common variant typically confers an odds ratio of only 1.01 to 1.5, which is far too modest to be clinically meaningful on its own. However, the aggregate effect of all contributing variants can be substantial, sometimes rivaling or even exceeding the risk conferred by an individual monogenic variant. Polygenic risk scores (PRS) attempt to capture this aggregate genetic burden in a single number that summarizes an individual's inherited predisposition to a given condition.

Construction of Polygenic Risk Scores

A PRS is calculated as the sum of the products of effect allele count and weight across all included variants. The weights are typically derived from GWAS effect sizes, expressed as beta coefficients or log odds ratios. Building a PRS involves a structured process: first, a discovery GWAS identifies associated variants and estimates their effect sizes; second, variant selection is performed using either pruning and thresholding (P+T) or Bayesian approaches such as LDpred or PRS-CS; third, the score is calculated in a target population; and finally, the score is validated in an independent dataset to assess predictive performance. The number of variants included can range from hundreds in older PRS to millions in genome-wide PRS constructed using Bayesian methods.

Statistical Methods for PRS Construction

Several statistical methods are used to construct PRS, each with distinct strengths. Pruning and thresholding (P+T or C+T) selects variants below a p-value threshold and prunes for linkage disequilibrium; this approach is simple but discards useful information. LDpred and its successor LDpred2 use Bayesian shrinkage that models LD structure, retaining more variants with attenuated weights for improved performance. PRS-CS applies a continuous shrinkage prior and generally outperforms P+T. Additional methods include DBSLMM, lassosum, and SBayesR, which vary in computational requirements and predictive accuracy. Performance is evaluated using metrics such as the area under the ROC curve (AUC), Nagelkerke R-squared, C-statistic, and hazard ratios per standard deviation of PRS.

Current Clinical Applications

Cardiovascular Disease

PRS for coronary artery disease has shown that individuals in the top percentiles (for example, above the 95th percentile) face a three- to four-fold increased risk, comparable to the risk carried by individuals with heterozygous familial hypercholesterolemia. This has generated interest in using PRS for risk stratification to guide statin initiation in intermediate-risk individuals. The AHA/ACC risk enhancement framework includes family history but does not yet formally incorporate PRS. Meta-analyses indicate that PRS adds modest discrimination beyond traditional risk factors, with C-statistic improvement of approximately 0.02.

Breast Cancer

Breast cancer PRS can reclassify women from average-risk to high-risk screening categories. PRS also interacts meaningfully with monogenic risk: it modifies the penetrance of BRCA1/2 variants, meaning that a BRCA carrier with a low PRS may have a meaningfully lower lifetime risk than one with a high PRS. The BOADICEA/CanRisk model integrates PRS with family history, BRCA status, and other risk factors to provide individualized risk estimates. The UK NHS has begun implementing PRS in some breast cancer screening pathways.

Type 2 Diabetes

PRS for type 2 diabetes identifies individuals at elevated risk years before disease onset. When combined with clinical risk factors, PRS may improve the targeting of preventive interventions such as lifestyle modification programs. However, clinical implementation remains limited to date.

Psychiatric Disorders

The schizophrenia PRS currently explains approximately 7% of phenotypic variance, which is not yet sufficient for clinical actionability. PRS for major depressive disorder and bipolar disorder have even lower predictive power. At present, the research utility of psychiatric PRS substantially exceeds their clinical utility.

Prostate Cancer

The Stockholm3 test integrates PRS with PSA levels and clinical variables to improve prostate cancer detection. This approach may reduce unnecessary prostate biopsies while maintaining sensitivity, and it represents one of the most clinically advanced applications of PRS in practice.

Limitations and Controversies

Ancestry Bias

The most significant limitation of current PRS is ancestry bias. The vast majority of GWAS have been conducted in European-ancestry populations, which comprise approximately 79% of GWAS participants as of 2023. PRS derived from these studies perform significantly worse when applied to non-European populations, with performance dropping by 50 to 80% in African-ancestry individuals and showing intermediate decline in East Asian and South Asian populations. The reasons for this disparity include differences in LD patterns, allele frequencies, causal variant identification, and environmental interactions across populations. If PRS is implemented clinically without addressing this bias, it risks exacerbating existing health disparities. Multi-ancestry GWAS and trans-ethnic PRS methods are under active development to mitigate this problem.

Predictive Performance

For most common diseases, PRS explains only 5 to 15% of phenotypic variance, which means individual-level prediction remains poor. AUC values for PRS alone typically range from 0.55 to 0.65 and improve only modestly when combined with clinical risk factors. Population-level stratification -- identifying individuals in the high-risk tails of the distribution -- is considerably more informative than attempting individual prediction. Importantly, PRS does not account for rare variants, structural variants, gene-gene interactions, or gene-environment interactions, all of which contribute to disease risk.

Clinical Utility vs. Statistical Validity

Statistical association does not automatically translate into clinical utility. The key question is whether knowledge of a PRS actually changes clinical management in a way that improves patient outcomes. Few randomized controlled trials have demonstrated that PRS-guided interventions outperform standard care. The eMERGE III network has returned PRS results to participants, and the clinical impact of this approach is still being evaluated. Most professional societies, including the ACMG and AHA, do not yet recommend routine PRS testing outside research settings.

Ethical and Social Concerns

A high PRS carries the risk of being misinterpreted as a diagnosis rather than a probabilistic risk factor, fostering genetic determinism. Insurance and employment discrimination remain concerns, particularly for conditions not covered by GINA, such as life insurance. Direct-to-consumer PRS reports may generate anxiety or false reassurance without adequate genetic counseling. The use of PRS for embryo selection (PGT-P for polygenic traits) raises significant ethical concerns about eugenics and consumer-driven reproductive selection. PRS for non-medical traits such as educational attainment and intelligence generates substantial societal controversy.

Methodological Challenges

PRS distributions and risk estimates may not generalize across populations, health systems, or time periods, creating calibration challenges. The "winner's curse" means that GWAS effect sizes are often overestimated in the discovery cohort. Pleiotropy -- where variants affect multiple traits -- complicates disease-specific PRS interpretation. Gene-environment interactions mean that PRS performance may vary across environmental contexts such as diet and physical activity levels.

Integration into Clinical Practice

Models of Clinical Implementation

Several models exist for incorporating PRS into clinical workflows. A standalone PRS report provides a risk category (high, average, or low) for one or more conditions. Integrated risk models combine PRS with family history, clinical biomarkers, and imaging, as exemplified by CanRisk for breast cancer and QRISK3 combined with PRS for cardiovascular disease. EHR integration envisions PRS calculated from existing genotyping data (such as pharmacogenomic arrays) and incorporated into clinical decision support. The concept of newborn genomic screening with PRS calculated at birth for common diseases remains highly controversial due to uncertain clinical utility and potential psychosocial impact.

Clinical Reporting Considerations

PRS should be reported as a percentile rank within a reference population, with the ancestry context clearly stated. Risk estimates should be presented as absolute values (lifetime or 10-year risk) rather than relative risks, and should include confidence intervals. Clinicians must understand that PRS represents probability, not destiny. Counseling should emphasize modifiable risk factors and available risk-reducing interventions alongside the genetic information.

<image>A diagram showing the construction and interpretation of a polygenic risk score. Panel 1 shows a Manhattan plot from a GWAS with significant loci highlighted. Panel 2 shows variant selection and LD pruning, with selected SNPs and their effect sizes (beta weights) listed in a table. Panel 3 shows the PRS formula: sum of (genotype dosage x beta weight) across all variants. Panel 4 shows the resulting PRS distribution in a population as a bell curve, with color-coded risk categories: low risk (bottom 20%), average (middle 60%), and high risk (top 20%), with the top 5% highlighted in red with the associated relative risk annotated.</image>

<image>A comparison infographic showing PRS performance across ancestries. Four panels represent European, East Asian, South Asian, and African ancestry populations. Each panel shows a ROC curve for CAD PRS, with AUC values declining from European (~0.63) to African (~0.55). A bar chart below compares the variance explained (R-squared) across ancestries, demonstrating the significant performance drop in non-European populations. A world map shows the geographic distribution of GWAS participants (heavily concentrated in Europe and North America) versus the global population distribution, highlighting the ancestry bias in genomic research.</image>

<image>A clinical decision-making flowchart integrating PRS with traditional risk factors. Starting with a patient presenting for cardiovascular risk assessment, the flowchart shows: (1) Assessment of traditional risk factors (age, sex, lipids, blood pressure, smoking, diabetes) yielding a 10-year ASCVD risk. (2) For patients in the intermediate risk category (7.5-20%), PRS is shown as a risk enhancer that can reclassify to higher or lower risk. (3) High PRS (top 20%) plus intermediate clinical risk leads to recommendation for statin therapy. (4) Low PRS (bottom 20%) may defer statin therapy. Risk-benefit discussions and shared decision-making steps are indicated at each branch point.</image>

Clinical Pearls

PRS is best understood as a population-level stratification tool, not an individual diagnostic test. A high PRS does not mean disease will occur, and a low PRS does not guarantee protection. The vast majority of PRS were derived from European-ancestry GWAS and perform poorly in other ancestry groups; applying European-derived PRS to non-European patients risks misclassification, and this limitation should be disclosed during counseling. PRS can modify the penetrance of monogenic variants -- for example, a BRCA1 carrier with a low breast cancer PRS may have a lower lifetime risk than a carrier with a high PRS, and integrated models like CanRisk are beginning to incorporate this interaction. Current professional society guidelines from the ACMG, AHA, and ASCO do not recommend routine clinical use of PRS, as it remains largely a research and emerging clinical tool. PGT-P (preimplantation genetic testing for polygenic conditions) is offered commercially but is not endorsed by ACMG or ASRM due to limited predictive value, small effect sizes within sibling cohorts, and significant ethical concerns. When patients bring PRS results from direct-to-consumer companies, clinicians should assess the methodology, ancestry match, and specific conditions tested before making any management changes. PRS explains only a fraction of disease risk, and family history remains one of the strongest and most accessible risk assessment tools, which should not be supplanted by PRS.

References

• Khera AV et al. Genome-wide polygenic scores for common diseases identify individuals with risk equivalent to monogenic mutations. Nat Genet. 2018;50(9):1219-1224.
• Martin AR et al. Clinical use of current polygenic risk scores may exacerbate health disparities. Nat Genet. 2019;51(4):584-591.
• Torkamani A et al. The personal and clinical utility of polygenic risk scores. Nat Rev Genet. 2018;19(9):581-590.
• Mavaddat N et al. Polygenic risk scores for prediction of breast cancer and breast cancer subtypes. Am J Hum Genet. 2019;104(1):21-34.
• ACMG Board of Directors. Points to consider in the clinical application of polygenic risk scores. Genet Med. 2023;25(8):100803.
• Turley P et al. Problems with using polygenic scores to select embryos. N Engl J Med. 2021;385(1):78-86.

Read this lecture as Markdown