Ancestry Adjustment

Ancestry adjustment is the step that turns a raw polygenic score (a weighted sum of your effect-allele dosages) into a percentile that means something for you, by centering and scaling that sum against a distribution matched to your genetic ancestry instead of the training cohort’s.

The reason you need it: allele frequencies differ across populations. A score is a sum over hundreds of thousands of variants, so systematic frequency differences accumulate into a large mean shift. Two people with identical underlying risk can land two standard deviations apart on the raw scale purely because of where their ancestors lived. The shift is in the score’s location and spread, not necessarily in its within-population ranking ability, which is why the fix is a recalibration and not a new model.

How it works

The standard method is PC-based continuous adjustment. You project the individual onto principal components computed from a reference panel (1000 Genomes, HGDP, or the UK Biobank PCA loadings), then fit two models in that same reference: the mean of the raw score as a linear function of the top PCs, and the residual variance as a function of the PCs. Your adjusted value is

z = (raw - mean_hat(PC1..PCk)) / sqrt(var_hat(PC1..PCk))

The variance model matters. Mean-only correction leaves admixed individuals with inflated or deflated spread, so their tails are wrong even after centering. The pgsc-calc implementation fits both moments against 1000 Genomes + HGDP; --run_ancestry with the reference panel bundle does the projection and the regression in one pass and writes both raw and adjusted scores.

Continuous PC adjustment beats assigning someone to a discrete population label. Labels fail for anyone admixed, which is a large fraction of the people who most need the correction.

Adjustment fixes calibration. It does not fix discrimination. Effect sizes and LD structure differ across populations, so a score trained in Europeans loses predictive power elsewhere even when perfectly centered, with the loss scaling roughly with genetic distance from the training population ¹. The remedies are upstream: multi-ancestry weights that combine European and target-population GWAS ², or admixture-aware methods that assign local-ancestry-specific weights along the genome ³. A well-built multi-ancestry score can hold predictive value across groups; a multi-ancestry BMI score predicted longitudinal weight change across ancestry groups in one recent cohort ⁴.

In your own data

From a WGS VCF, the pipeline is:

1. Convert to PLINK2 format: plink2 --vcf sample.vcf.gz --make-pgen --out sample. Confirm the build. Most scoring files on the PGS Catalog are GRCh37 or GRCh38 and mixing them silently drops variants.
2. Harmonize alleles. Strand-ambiguous A/T and C/G SNPs are the classic failure mode. Either drop them or resolve by allele frequency; do not guess.
3. Score: plink2 --pfile sample --score weights.txt 1 2 3 cols=+scoresums with columns for variant ID, effect allele, and weight. PRSice remains a reasonable option when you want clumping and p-value thresholding from summary statistics ⁵, and the UK Biobank practical guide walks through the QC decisions in order ⁶.
4. Project onto reference PCs with plink2 --score using precomputed loadings, then apply the mean and variance models.

Things to check before you believe the number. What fraction of the scoring file’s variants were found in your data? Below roughly 90% and the raw sum is systematically low, which the PC adjustment will not catch because it is a per-sample effect, not an ancestry effect. Is --score using mean imputation for missing genotypes or treating them as zero? The default behavior differs between tools and the difference is large. Did your PCs land inside the reference cloud when you plot PC1 vs PC2? A sample sitting off the manifold usually means a build or strand problem, not unusual ancestry.

One question that comes up: a 1% ancestry fraction from a consumer report typically reflects a single ancestor six to seven generations back, but the estimate is noisy at that level and is not the input to PC adjustment anyway. The PCs use your genome-wide coordinates directly.

Limitations

Adjustment cannot rescue a score whose training GWAS had almost no representation from your ancestry. You get a calibrated percentile within a distribution the model still predicts poorly ¹. The reference panel’s coverage sets your ceiling: if no reference individuals resemble you, the mean and variance models extrapolate. PC-based correction handles the bulk of the shift but does not remove all residual structure ⁷. Population-level structure and technical artifacts both propagate into scores in ways that are hard to separate ⁸.

Percentiles are not diagnoses. Absolute risk depends on age, sex, family history, and non-genetic factors, and the ACMG’s position is that polygenic scores should be interpreted in a clinical context with appropriate counseling ⁹. Clinical utility still varies widely by trait and score ¹⁰. Take any score you compute to a clinician before acting on it.

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.