A polygenic risk score adds up a person's risk alleles, each weighted by an effect size usually estimated in European cohorts. In an admixed person, for example an African American adult whose chromosomes are a patchwork of African and European segments, the raw score is shifted. Part of that shift reflects how common each allele is in each ancestral population and has nothing to do with disease, and part comes from weights that work less well where correlation patterns between nearby variants differ. Current fixes re-center the score using a single number for global ancestry. This proposal makes three changes. It computes the score's expected value and spread from the person's local ancestry, meaning which ancestral population each stretch of each chromosome comes from. It then turns the re-centered score into an absolute probability of disease by fitting a small ancestry-dependent model under a proper scoring rule, a loss that is minimized only when stated probabilities match true frequencies. Finally, it judges success by calibration within each ancestry group, not by a single pooled accuracy number.
The motivation is clinical. A risk score that ranks people well but says "30% risk" to a group whose true risk is 15% will mislead treatment decisions, and pooled discrimination metrics such as AUC cannot detect that error. The claim is narrow and testable. Among admixed individuals with the same global ancestry, those whose high-weight score loci fall in different ancestral segments will be miscalibrated by global-ancestry methods and corrected by the local method. The gain should be large for traits driven by a few strong loci, such as mean corpuscular volume (MCV), the average size of red blood cells, and small for highly polygenic traits such as type 2 diabetes (T2D).
What exists and what is missing
The closest work, Ancestry Calibration of Polygenic Risk Scores Improves Risk Stratification and Effect Estimation in African American Adults, regresses the raw score on genetic similarity to the Yoruba reference population (a global ancestry proxy) in 8,841 REGARDS participants. It subtracts the predicted mean and rescales within ancestry bins. This raises the T2D top-versus-bottom-decile odds ratio from 7.97 to 10.77 and the height correlation from 0.24 to 0.32. The paper motivates our work, and it also names our gap: it uses global rather than local ancestry and never reports absolute-probability calibration. We reuse its re-centering logic, replace the empirical global regression with an analytic local-ancestry expectation, and add a proper-scoring-rule step that outputs probabilities.
A second line learns risk predictors directly. Genetics-driven risk predictions leveraging the Mendelian randomization framework (PRiMeR) trains a predictor on cross-sectional UK Biobank traits. Its loss aligns the predictor's genetic associations with causal effects from published genome-wide association summary statistics, and it reaches a 5-year T2D AUC of 0.847. It assumes the cohort and the summary statistics come from the same population, and it outputs a relative score rather than a probability. We treat PRiMeR and standard risk scores as interchangeable inputs to our calibration layer, not as competitors on ranking.
Optimal strategies for learning multi-ancestry polygenic scores vary across traits supplies the key constraint. For MCV, a score trained on about 5,000 Africans beat one trained on 60 times as many Europeans, because a population-specific locus (DARC) dominates. For height, the large European cohort still won. This predicts that local ancestry matters most where a few loci carry most of the score's weight. That prediction is what our experiments test.
On the method side, Lai-Net: Local-Ancestry Inference with Neural Networks shows that local ancestry can be inferred at 97.85% accuracy on simulated admixed genomes, against 92.47% for RFMix, so the inputs we need are obtainable. Countdown Regression: Sharp and Calibrated Survival Predictions shows that fitting survival models with the continuous ranked probability score, a proper scoring rule for full predictive distributions, yields far sharper distributions than maximum likelihood with calibration slopes near 1. We use that rule for continuous traits.
The fairness audit On Biases in a UK Biobank-based Retinal Image Classification Model found that group-robust training closed subgroup gaps only by degrading overall performance. This warns against importing generic fairness objectives, and our design avoids them.
Core mechanism
For person $i$, SNP $j$ with effect weight $\beta_j$, and haplotype $h \in \{1,2\}$ carrying local ancestry $a_{ijh}$, let $f_j(a)$ be the reference allele frequency of SNP $j$ in ancestral population $a$, taken from 1000 Genomes. Under the null of no excess risk, the score's expected value and variance are
$$\mu_i = \sum_j \beta_j \sum_{h=1}^{2} f_j(a_{ijh}), \qquad \sigma_i^2 = \sum_j \beta_j^2 \sum_{h=1}^{2} f_j(a_{ijh})\bigl(1-f_j(a_{ijh})\bigr).$$
The variance formula ignores correlation between SNPs. We check that approximation empirically and, if needed, replace it with a variance computed by simulating genotypes within ancestry segments. The re-centered score is $z_i = (\mathrm{PRS}_i - \mu_i)/\sigma_i$.
The weights lose predictive power on non-European segments, so we also define a score-weighted local ancestry $\theta_i = \sum_j \beta_j^2 \,\tfrac12\sum_h \mathbf{1}[a_{ijh}=\mathrm{AFR}] / \sum_j \beta_j^2$. This is the fraction of the score's variance that sits on African-ancestry segments, and analogous terms are defined for other ancestries. Absolute risk is then
$$p_i = \sigma\bigl(\alpha_0 + \alpha_1 \theta_i + (\gamma_0 + \gamma_1 \theta_i)\, z_i + c_i^\top \eta\bigr),$$
where $\sigma$ is the logistic function, $c_i$ holds age and sex, and the parameters are fit by minimizing the Brier score (squared error between $p_i$ and the outcome), with shrinkage on the $\theta$ terms. For continuous traits such as MCV, the same structure parameterizes a Gaussian mean and variance fit by the continuous ranked probability score. The model is deliberately small. Minority samples run to thousands, not hundreds of thousands, and a proper score yields within-group calibration only if the model can represent each group's intercept and slope. $\theta_i$ is the one input that lets it represent them for admixed people.
Research Program
The first stage builds the local-ancestry inputs. We infer local ancestry with RFMix and with a Lai-Net-style network on 1000 Genomes references, for the UK Biobank African, Admixed American, Central/South Asian, and East Asian participants and for REGARDS where access permits. We then compute $\mu_i$, $\sigma_i$, and $\theta_i$ for published T2D, height, and MCV scores and for a PRiMeR-style T2D predictor.
The second stage checks the null model. In each ancestry group we test whether $z_i$ has mean 0 and variance 1 in disease-free individuals, and whether the residual mean of the global-ancestry method tracks $\theta_i$ after conditioning on global ancestry.
The third stage fits and evaluates the calibration layer with cross-fitting, holding out entire ancestry groups and admixture strata.
The fourth stage tests whether going beyond a post-hoc layer helps. Following the observation in shaochenze/ScoringRulesLM that scoring-rule losses work best when fine-tuning a model already trained with cross-entropy, we start from the European-trained weights and fine-tune segment-level weight multipliers under the Brier score on multi-ancestry training data. This in-processing step is kept only if it beats the post-hoc layer on held-out groups.
Baselines
The naive reference is the European-trained score with one global logistic intercept and slope fit on pooled data. It shows how badly uncorrected scores miscalibrate minority groups.
The strongest existing method is the global-ancestry calibration of Ancestry Calibration of Polygenic Risk Scores Improves Risk Stratification and Effect Estimation in African American Adults, followed by the same proper-scoring-rule logistic layer as ours, with global ancestry in place of $\theta_i$. This competitor shares every ingredient except local ancestry, so it isolates the claim. Its published figures are the T2D decile odds ratio of 10.77 and the height correlation of 0.32. No reported result was found on Lacuna for Brier score or expected calibration error under this method, so we will reproduce it on identical folds.
A third baseline is logistic regression with the raw score, ancestry principal components, and their interactions. It is the standard epidemiological fix and is also fit under log loss.
For ranking, PRiMeR's reported 5-year T2D AUC of 0.847 is the reference. Our layer should leave AUC within error of its input score while improving calibration.
The ablations are local centering without the $\theta$-dependent slope, the $\theta$ slope without local centering, Brier versus log-loss fitting, and RFMix versus neural local ancestry. Every method gets the same folds, covariates, and a shrinkage-tuning grid of the same size chosen by inner cross-validation.
Beating the strongest baseline means a lower reliability component of the Brier score (the part that measures calibration) and calibration slopes closer to 1 in held-out admixed groups, without a significant loss in discrimination. Confidence intervals come from bootstrapping over individuals.
Why not just…?
Why not just add ancestry principal components and interactions to a logistic model? Log loss is already proper, so this could work. The "covariate trap" documented in the ancestry-calibration paper concerns unstable effect estimates when an uncalibrated score and global ancestry enter together, which is not quite a prediction failure. The real limitation is that principal components summarize global ancestry. They cannot tell apart two people with the same global ancestry whose high-weight loci sit on different ancestral segments. If this baseline matches ours on the matched-global-ancestry test described below, local ancestry adds nothing, and we will report that.
Why not just recalibrate per ancestry bin with Platt scaling? Discrete bins fail admixed people, whose ancestry is continuous and mosaic. Recalibrating with continuous global ancestry is exactly our strongest baseline. Beyond Binary Rewards: Training LMs to Reason About Their Uncertainty found that in-processing calibration beat post-hoc probes, but that was for language models and does not transfer automatically. That is why stage four tests the claim rather than assuming it.
Why not just train ancestry-specific scores or reweight minority samples? Optimal strategies for learning multi-ancestry polygenic scores vary across traits shows that reweighting inflates variance and that minority-only training loses to large European cohorts for highly polygenic traits. Our method keeps the European weights and estimates only a handful of ancestry-dependent calibration parameters, which suits minority samples of around 5,000. For MCV, a minority-trained score is a serious competitor, so we include it as an extra baseline there.
Why not use group-robust training such as GroupDRO? The retinal audit found that it levels down overall performance. It also optimizes worst-group loss, not calibration, so it addresses a different target.
Decisive Experiments
The central test stratifies admixed African-ancestry participants in UK Biobank and REGARDS into bins of global ancestry. Within each bin, it compares calibration across terciles of $\theta_i$. If the global method's predicted-minus-observed risk moves with $\theta_i$ inside bins and ours does not, the local mechanism is confirmed.
The second test is the trait contrast. We predict a large reduction in the continuous ranked probability score and in reliability for MCV, where local ancestry at DARC dominates, and a small reduction for T2D and height. A uniform gain across traits would suggest the benefit comes from somewhere other than locus concentration.
The third test checks robustness to local-ancestry error by deliberately corrupting ancestry calls at 5–40% rates, where Lai-Net-style inference stays above 94% accuracy even with 50% of input SNPs missing. It reports how quickly calibration degrades toward the global baseline.
The fourth test compares Brier and log-loss fitting in the smallest groups, East Asian and Middle Eastern. It settles whether the bounded rule's robustness matters in this small-sample regime or only in the reinforcement-learning settings where it was observed.
References
- Ancestry Calibration of Polygenic Risk Scores Improves Risk Stratification and Effect Estimation in African American Adults
- Genetics-driven risk predictions leveraging the Mendelian randomization framework
- Optimal strategies for learning multi-ancestry polygenic scores vary across traits
- Lai-Net: Local-Ancestry Inference with Neural Networks
- Countdown Regression: Sharp and Calibrated Survival Predictions
- On Biases in a UK Biobank-Based Retinal Image Classification Model
- Beyond Binary Rewards: Training LMs to Reason About Their Uncertainty