Polygenic risk scores from raw DNA: what they can and can't tell you

Most common diseases aren't caused by one gene. Height, type 2 diabetes, coronary artery disease, and most other complex traits are influenced by thousands of variants, each nudging risk a tiny amount. A polygenic risk score (PRS) adds those nudges up.

How a PRS is computed

A published score is a list of variants with weights, derived from a genome-wide association study. For each variant, your genotype (0, 1, or 2 copies of the effect allele) is multiplied by the weight, and everything is summed. Public repositories like the PGS Catalog hold thousands of such scoring files; Gene Wizard computes a curated panel of them from your uploaded file.

The raw sum is meaningless by itself. What matters is where your score falls in a reference population — which is where the honest complications start.

Complication 1: your chip file doesn't have every variant

A published score might use 100,000 variants while your chip tested only a fraction of them. Coverage varies score by score, so any trustworthy report should show how many variants were actually matched and widen its uncertainty accordingly — a percentile presented without a confidence interval on partial data is false precision. Gene Wizard reports coverage and a 95% interval on every score.

Complication 2: ancestry calibration

Most GWAS cohorts have been predominantly European-ancestry. Apply those weights to someone of different ancestry against a European reference distribution and the resulting percentile can be badly miscalibrated — this is the single biggest source of misleading PRS results in consumer tools.

The mitigation is to compare your score against a reference panel matched to your genetic ancestry. Gene Wizard estimates your admixture across 26 reference populations first, then normalizes each score against ancestry-matched reference distributions (mixing them for admixed users). It's a substantial improvement, not a full solution: scores still transfer imperfectly across ancestries, and honest tools say so.

What a percentile actually means

Suppose your type 2 diabetes PRS lands at the 85th percentile:

It means your common-variant genetic load is higher than about 85% of the reference population. Your relative risk is elevated.
It does not mean you have an 85% chance of diabetes. Converting relative position into absolute risk requires the condition's baseline prevalence — for most conditions, even a high-percentile PRS shifts absolute risk by only a few percentage points. Gene Wizard shows this conversion (via a liability threshold model) alongside the percentile.
Lifestyle, environment, and rare variants the score can't see all still apply. A PRS is one input, not a verdict.

Where PRS are genuinely useful

Context for family history. Polygenic background can partly explain clustering of common disease in families without a single-gene cause.
Screening conversations. A markedly high score for something like coronary artery disease is a reasonable prompt to discuss earlier or more frequent screening with a clinician — several health systems are piloting exactly this use.
Understanding single-variant results in proportion. One risk variant in APOE (rs429358) or MTHFR (rs1801133) means much less than the aggregate of thousands — the PRS view keeps single-SNP findings from being over-read.

Where they mislead

Percentiles without confidence intervals or coverage information.
Scores applied across ancestries without renormalization.
Absolute-sounding claims ("your risk is 3x") derived from relative percentiles.
Any presentation of a PRS as diagnostic. It isn't, for anyone.

Try it on your own data

If you have a 23andMe, AncestryDNA, or whole-genome file, Gene Wizard computes an ancestry-aware PRS panel with coverage, confidence intervals, and absolute-risk context shown for every score — the preview is free, the full report a one-time $19.99 — and there are demo results you can inspect before uploading anything. New to raw data? Start with how to analyze your 23andMe raw data.