Polygenic Risk & Early Warning
PRS has crossed the clinical-utility threshold for some diseases — and not for others.
What's covered
- T01Construction: GWAS summary stats → LD reference → score (LDpred2, PRS-CS, MegaPRS)
- T02Cross-ancestry portability and the ancestry-bias problem
- T03Integrated risk: PRS + monogenic + clinical (QRISK3 + PRS for CAD, Tyrer-Cuzick + PRS for breast)
- T04Use cases with clinical utility: CAD, breast cancer, prostate cancer, T2D, AF, hereditary haemochromatosis
- T05Use cases without it: most psychiatric, most behavioral
- T06Regulatory landscape: FDA, MHRA, the EU IVDR position
By the end of this module you will be able to
- L01Explain why a PRS trained in European-ancestry GWAS underperforms in African-ancestry populations and quantify the magnitude.
- L02Describe an integrated risk model where PRS adds clinically meaningful discrimination on top of standard scores.
- L03List two PRS-driven trials currently informing screening guidelines.
What you should walk away believing
- For coronary artery disease, the top 5% of polygenic risk carries ~3–4× the lifetime risk — comparable to a strong monogenic FH variant in absolute terms.
- PRS portability loss across ancestries is real and large (often ~50% R² loss); deploying European-trained scores as 'universal' is malpractice.
- Adding PRS to QRISK3 reclassifies ~15% of intermediate-risk patients up or down a treatment threshold for statin initiation (UK Biobank).
- MyOme, Genomics PLC, Allelica, Nightingale are the main clinical-grade PRS implementers in 2025–2026.
What this means at your level
Polygenic risk scores aggregate many small-effect variants from genome-wide association studies into a single number predicting disease risk. They've become genuinely useful for some common diseases (heart disease, breast cancer) but remain limited for others, and they perform much worse in populations not represented in the underlying GWAS.
Use PRS where: (1) outcome is common, (2) effective intervention exists, (3) score is calibrated to the patient's ancestry, and (4) it would change management. Best current candidates: CAD primary prevention (statin decision), breast cancer screening start age and frequency, prostate cancer biopsy threshold. Avoid PRS for major depressive disorder, schizophrenia, intelligence — discrimination is real but interventions are not, and ethics are unresolved.
Multi-ancestry methods (PRS-CSx, MUSSEL, BridgePRS) are closing but not eliminating the ancestry gap. The All of Us, H3Africa, GenomeAsia 100K, and TOPMed cohorts are the substrate for the next generation. Methylation-based polygenic risk (mPRS) and transcriptome-imputation methods (TWAS, PrediXcan) add orthogonal information for some endpoints.
If your PRS for X is in the top decile, you're going to get X.
PRS gives a relative risk over the population baseline, not a deterministic prediction. Top-decile breast-cancer PRS lifts lifetime risk from ~12% to ~30% — meaningful, but ~70% still don't get the disease. Counsel as risk modification, not prophecy.
What the data say
Test yourself
Key terms & abbreviations
- PRSPolygenic Risk Score
- Weighted sum of effect alleles across many loci, derived from GWAS summary statistics.
- GWASGenome-Wide Association Study
- Population study testing millions of variants for disease association; the substrate for PRS.
- LDLinkage Disequilibrium
- Non-random correlation between alleles at nearby loci; LD reference panels (1000G, UKB) shape PRS construction.
- Net Reclassification Improvement
- Metric for whether a new biomarker moves people across clinically meaningful risk thresholds — more useful than ΔAUC for clinical decisions.
Anchor references
- Genome-wide polygenic scores for common diseases identify individuals with risk equivalent to monogenic mutations — Khera et al., Nat Genet 2018
- Polygenic prediction across ancestries — Martin et al., Nat Genet 2019