Where the Fairness Knob Belongs in Polygenic Prediction

A polygenic score adds up the small effects of many DNA variants to predict a trait or disease risk. These scores are trained mostly on people of European ancestry, and they lose much of their accuracy in people of African, East Asian or Hispanic/Latino ancestry. The machine-learning fairness literature offers many ways to shrink performance gaps between groups. Most of them act on the model's outputs: they penalize disparity in the loss, reweight groups, or adjust scores after training. This note argues that in genomic prediction the useful knob sits earlier, in how much prior weight the model gives to each functional class of DNA. Protein-coding exons, introns and distal regulatory regions carry different shares of heritability, and their effects probably transfer differently across ancestries.

The proposal is to trace the full trade-off curve between European accuracy and cross-ancestry accuracy, which we call the accuracy-fairness frontier. We would trace it in two ways and compare them. The first moves prior variance from regions that explain variance in Europeans toward regions whose effects carry over across ancestries, with special attention to the exome. The second applies the output-side fairness methods used in clinical machine learning. If the annotation knob gives a frontier that dominates the output-side one, the guidance for anyone building fairer genetic risk scores is to spend the fairness budget on which variants the model believes in, not on post-hoc correction. Two further checks are central to the design. The first asks whether the coding annotation adds anything beyond local measures of cross-ancestry genetic correlation. The second checks whether the frontier is real or an artifact of noisy minority test sets.

Core Claim

For continuous traits, any adjustment applied within an ancestry group after training is monotone, and a monotone adjustment cannot change how people are ranked within that group. Post-hoc fixes can therefore repair calibration but cannot close a gap in within-group $R^2$ or AUC, except by degrading the better-served group. That degrading is the "leveling down" that Fairness With Minimal Harm formalizes and tries to avoid in its Pareto-reweighting method for clinical neural networks. A real fairness gain in a polygenic score has to come from changing which variants get weight.

Two existing facts make functional annotation the natural place to make that change. Probabilistic inference of the genetic architecture underlying functional enrichment fits a Bayesian whole-genome regression (BayesRR-RC, in the GMRM software) with separate prior groups for exons, introns, promoters and distal regions. It finds that exons explain 12–25% of genetic variance, far more than their share of the sequence, while introns explain 32–44%. Its validation, however, is European-only. Quantifying portable genetic effects (X-Wing) shows that portability is concentrated in a small part of the genome: about 1% of the genome carries over 60% of the European–East Asian genetic covariance for mean corpuscular volume. X-Wing finds these regions by local genetic correlation and does not use coding annotation. It also optimizes accuracy only and never treats the European–versus–other-ancestry tension as a frontier.

Our claim is that a single prior-reallocation parameter, applied over functional classes, traces a frontier that dominates fairness penalties on the outputs. We further claim that coding annotation contributes portability beyond what local genetic correlation already captures. Both parts can fail, and the experiments are designed so that either failure would be visible.

Research Program

We work inside the BayesRR-RC model because it already assigns a separate prior variance to each annotation-by-frequency-by-LD group (78 groups) and scales to biobank data. Here LD, linkage disequilibrium, means correlation between nearby variants. We change only how the group-level prior variances are set. Let $\sigma^2_c(\alpha)$ be the prior variance for functional class $c$. Let $h^2_c$ be that class's share of heritability estimated in Europeans. Let $\rho_c$ be the share of cross-ancestry genetic covariance that class $c$ captures, estimated from target-ancestry summary statistics. Then

$$\sigma^2_c(\alpha) \propto (h^2_c)^{1-\alpha}\,(\rho_c)^{\alpha}, \qquad \alpha \in [0,1],$$

so $\alpha=0$ recovers the standard European-driven model and $\alpha=1$ allocates prior mass purely by portability. Sweeping $\alpha$ gives a one-dimensional curve in the plane of European $R^2$ against target-ancestry $R^2$. We trace it separately with exonic classes treated as one block, as in BayesRR-RC, and split into missense, protein-truncating and splice classes. We also trace it with X-Wing's local-correlation blocks substituted for functional classes. That substitution is the comparison that tells us whether annotation matters at all.

On the evaluation side, we treat each ancestry as a group. For each group we report the score's $R^2$ for quantitative traits and AUC for binary traits. We summarize fairness with the minimum $R^2$ across groups and the gap between groups. To get a single number per configuration we use FairLib's Distance to the Optimal, the distance to the ideal point of maximum accuracy with zero gap, as defined in FairLib.

On the Impact of Random Seeds on the Fairness of Clinical Classifiers found that subgroup fairness swings by nearly 15% in $\Delta$AUC across random seeds at fixed overall AUC, and that validation fairness did not predict test fairness. That finding constrains the design. We therefore run multiple MCMC chains per $\alpha$ and bootstrap the small target-ancestry test sets. We call one frontier dominant only where the confidence bands separate. We also choose $\alpha$ on held-out data from a different cohort than the one used for evaluation.

Baselines

The naive reference point is the standard BayesRR-RC score at $\alpha=0$, trained on UK Biobank Europeans and applied unchanged to every target group. In Europeans this model reaches out-of-sample prediction correlations of 0.62 for height and 0.34 for BMI in the Estonian Biobank, as reported by Patxot et al. It defines both the "no harm" line for European accuracy and the starting gap.

The strongest existing method for cross-ancestry accuracy is X-Wing. It reports relative gains in predictive $R^2$ of 14.1%–119.1% over PRS-CSx and XPASS across non-European benchmarks, with a median 35.2% gain in East Asians. We run X-Wing on exactly the same discovery summary statistics, LD reference panels and target splits used by our method. It gets the same tuning budget, which for X-Wing means its own repeated-learning combination step. Beating it means one of two things: a point on our frontier with higher target-ancestry $R^2$ at no lower European $R^2$, or matched target $R^2$ at higher European $R^2$, with bootstrap intervals that separate.

The output-side fairness baseline adapts the ParetoFair group reweighting from Fairness With Minimal Harm. It applies ancestry-group loss reweighting to a model fitted on the same variants, sweeping the reweighting strength to trace its own frontier. It uses the same individual-level data where available and the same number of frontier points as our $\alpha$ sweep. We include the demographic-expert design of FairSDE in its simplest genomic form. That version selects per individual between an ancestry-specific score and the pooled score. FairSDE reports a Max-Min AUC of 81.2 against 80.54 for standard empirical risk minimization on Harvard-GF imaging. We found no reported result on Lacuna for FairSDE or ParetoFair on genomic data, so no prior genomic number exists to beat.

Three ablations isolate the key ingredient. The first is an exome-only score that drops all non-coding variants. The second replaces the exonic class with a random variant set matched for allele frequency, LD score and total size. The third replaces functional classes with X-Wing local-correlation blocks. Where coding variants were directly typed on the exome chip, we also compare scores built from typed coding variants against scores built from imputed ones, to see whether imputation drives part of the gap.

Why not just…?

Why not just run X-Wing, which already finds portable regions? This is the most serious threat. If X-Wing's 1%-of-genome portable segments are mostly coding regions or overlap them heavily, our annotation knob would rediscover them under a different name. The block-substitution ablation and a combined model with both block and annotation priors settle this. If adding coding annotation to X-Wing's blocks moves the frontier no further, we will report that the coding effect is explained by local genetic correlation. That negative result is still useful, because it says annotation pipelines can be skipped.

Why not just apply per-ancestry recalibration or threshold adjustment after training? Within-group ranking is unchanged by any monotone transform, so such adjustments cannot raise a group's $R^2$ or AUC. They can only equalize those metrics by degrading the better-served group. Differentiable Optimization Layers for Guaranteed Fairness likewise found post-hoc projection 1.46% AUC worse than end-to-end training on loan default. We still include recalibration in the evaluation, because calibration gaps matter clinically even when ranking gaps cannot be fixed this way.

Why not just restrict to the exome, or simply collect more diverse samples? An exome-only score discards the 75% or more of heritability that lies outside exons, so it may sit far inside the frontier. It is included as an ablation precisely because it is the crude version of our idea. Larger diverse cohorts are the real long-term fix. But Ancestral risk modification for multiple sclerosis shows effect sizes doubling with local ancestry, and the REGARDS cohort of 7,726 participants shows how slowly diverse sample sizes grow. Deciding how to use existing European data in the meantime remains a necessary question.

Proposed Experiments

Discovery uses UK Biobank Europeans with individual-level data in GMRM. East Asian targets use Biobank Japan summary statistics, and African and Hispanic targets use PAGE summary statistics, both from the X-Wing benchmark. Each target set is split into a half used to estimate $\rho_c$ and a held-out half used for evaluation. A further, independent test uses the African American REGARDS cohort from the REGARDS PheWAS, genotyped on an exome chip. REGARDS is never used for tuning, which gives a clean check that $\alpha$ chosen on PAGE carries over. Traits are chosen for overlap across cohorts: height, BMI, LDL and HDL cholesterol, and mean corpuscular volume. Height and BMI are included because Patxot et al. found height has roughly twice BMI's coding share, so the coding effect, if it exists, should be stronger for height.

The decisive checks come in order. The first is whether the annotation-reallocation frontier dominates the output-reweighting frontier in held-out PAGE and Biobank Japan data, with chain-and-bootstrap intervals separating. The second is whether the exonic class beats the matched random set, which separates biology from simple sparsity. The third is whether annotation adds beyond X-Wing blocks in a combined model. The fourth is whether the $\alpha$ selected on PAGE still improves minimum-group $R^2$ in REGARDS without lowering European $R^2$ in the Estonian-style holdout by more than the bootstrap noise. We would read success on the first and fourth checks, with failure on the third, as "fairness belongs in the prior, but local correlation suffices". Success on all four would make functional annotation a first-class fairness tool for polygenic scores.