Polygenic Risk Score

A polygenic risk score is a weighted sum of the risk alleles you carry across hundreds to millions of common variants, where the weights come from the effect sizes estimated in a genome-wide association study of that trait. For coronary artery disease, type 2 diabetes, or major depression, no single variant moves the needle much. The signal lives in the aggregate.

How it works

The arithmetic is trivial: PRS = Σ βᵢ × dosageᵢ, summed over variants i, where βᵢ is the log odds ratio (or beta for a continuous trait) from GWAS summary statistics and dosageᵢ is your count of the effect allele, 0/1/2 for hard calls or a real number between 0 and 2 for imputed dosages.

The hard part is choosing which variants to include and how to shrink the weights. GWAS betas are inflated by linkage disequilibrium: neighboring SNPs carry correlated signal, and naively summing them double-counts. The two families of approaches are clumping plus thresholding (C+T), which greedily keeps the most significant variant in each LD block and discards correlated neighbors below an r² cutoff, and Bayesian shrinkage methods like LDpred2 and PRS-CS, which model the LD structure from a reference panel and shrink betas toward zero under a prior on effect size distribution. PRSice implements C+T with automated p-value threshold optimization across a grid ¹, and the accompanying protocol walks through QC, LD reference matching, and covariate handling in detail ². PRSet extends the framework to pathway-restricted scores, which is useful if you want to know whether your liability concentrates in, say, lipid transport rather than spread evenly across the genome ³.

We would use PRS-CS or LDpred2 over C+T for any score you plan to interpret. C+T thresholds are tuned on a validation cohort, and if you do not have one you are picking a threshold blind.

In your own data

Start with what you have. Array genotypes give you roughly 600k–900k directly typed sites. Whole-genome sequencing at 30x gives you every common variant directly, which removes imputation error as a failure mode entirely. Either way the input to scoring is a per-sample dosage file.

The practical path:

1. Get a scoring file from the PGS Catalog (PGS######.txt.gz). It has rsID, chr_name, chr_position, effect_allele, other_allele, effect_weight, and a header block naming the genome build.
2. Confirm the build. Scoring files are often GRCh37, modern WGS VCFs are GRCh38. Lift over with CrossMap or use the harmonized hmPOS_GRCh38 variant the Catalog publishes. Mixing builds silently scores the wrong positions.
3. Score with PLINK: plink2 --pfile mydata --score PGS000018_hmPOS_GRCh38.txt 1 4 6 header cols=+scoresums --out cad_prs, where the numeric args are the ID, effect allele, and weight columns.
4. Check the log for --score: N variants processed. If your scoring file has 1.1M variants and PLINK used 240k, you have an allele-matching problem, not a coverage problem.

The failure modes, in rough order of how often they bite:

• Strand flips on A/T and C/G sites. These are ambiguous and cannot be resolved by allele matching alone. Drop them or resolve by allele frequency against a reference panel.
• Missing genotypes. PLINK’s default imputes missing calls to the cohort mean allele frequency, which is fine in a biobank and wrong for n=1. Use --read-freq with an ancestry-matched frequency file.
• Reporting the raw sum. A PRS of 0.0143 means nothing. You need a percentile against a reference distribution, and that distribution must match your genetic ancestry, because scores derived in European-ancestry GWAS shift in mean and variance across populations in ways that have nothing to do with risk ⁴.

The UK Biobank guide is the most concrete published walkthrough of these steps, including how to handle related individuals and principal-component adjustment ⁵.

Limitations

Discrimination is modest. For most common diseases the area under the ROC curve for a PRS alone sits in the 0.6s, which means substantial overlap between people who develop the disease and people who do not. A top-decile score is a real shift in probability, not a prediction.

Portability across ancestries remains the sharpest constraint. Effect estimates from European-ancestry GWAS lose accuracy in African-ancestry individuals, and the degradation is large enough that percentile placement can be badly wrong ⁶⁴. If your ancestry is not well represented in the discovery cohort, treat the percentile as provisional.

Clinical implementations are still being worked out. Laboratory-developed PRS assays require their own analytic validation, a defined reference population, and a reporting workflow separate from the computation itself ⁷. What a score should change in practice is an open question, and translating a percentile into any decision is a conversation with a clinician, not something to resolve from a text file ⁸⁹. Interpretive ambiguity and limited evidence for clinical utility are the barriers people in the field name first ⁶¹⁰.

Woolf Software builds longitudinal molecular profiles of individuals: whole-genome sequencing, RNA sequencing, proteomics, blood biomarkers, and continuous glucose data, integrated into one model of you. Build your profile.