Genetic risk scores predict disease from a person's DNA. They work well for people of European ancestry and much worse for everyone else, because the studies used to build them are mostly European. The usual fixes are applied after training, by adjusting the finished score for each ancestry group. We propose instead to change what the model learns while it trains. The method watches how accurately the model predicts at every point along a continuous ancestry axis, measured on held-out people. Wherever accuracy falls behind, it raises the weight of training individuals with similar ancestry. Because each person's weight is a smooth function of their genetic ancestry coordinates, and not of a group label, admixed individuals (people whose ancestry mixes several continental sources) are handled naturally. Nobody has to be forced into a bin.
The opportunity sits between two literatures. Fairness research has built "dynamic weighting" methods, which turn up the loss weight of whichever group is currently doing worst during training. These methods almost always assume a few discrete groups and have been tested on census, text and face data. Statistical genetics has documented the ancestry gap in detail. Its corrections are mostly static (Bayesian shrinkage priors) or post-hoc (score recalibration). The claim of this note is narrow and testable. When a constraint on local predictive accuracy is enforced during training as a smooth field over continuous ancestry, and updated from held-out error, it improves discrimination for under-represented and admixed people beyond what post-hoc calibration or static reweighting achieve. We expect this effect mainly through which variants a penalized model chooses to keep.
What exists, and where this differs
Fairness side. The closest method is FairGrad. It rewrites a constrained fairness problem as group weights updated batch by batch, adds little compute, and shows that some parity goals need negative weights. It requires discrete protected classes. Optimising Equal Opportunity Fairness in Model Training and Scalable and Trustworthy Learning in Heterogeneous Networks show that adjusting loss contributions as training proceeds beats duplicating or deleting samples. In-Training Multicalibrated Survival Analysis for Healthcare via Constrained Optimization shows that Lagrangian constraints, meaning penalty multipliers that grow while a constraint is violated, can make clinical models self-correct during training. These papers supply our optimization machinery, and all of them assume categorical subgroups.
Fairness without Demographics through Adversarially Reweighted Learning (ARL) drops group labels altogether. A small adversary network learns which training examples to upweight. ARL also constrains our design, because it showed no gains on COMPAS, where labels are noisy, and disease phenotypes are noisy too.
Genetics side. DisPred splits genotypes into an ancestry representation and a disease representation, then discards the ancestry part. It trains in stages with no feedback on how each subgroup is doing, and its authors report difficulty with multi-way admixed individuals. X-Wing works from published association summaries. It finds the genomic regions whose effects carry over between populations and shrinks those regions less. Its weighting is static, uses discrete reference panels, and treats portability as categorical; its authors name that last point as a limitation.
Ancestry Calibration of Polygenic Risk Scores motivates our evaluation choices. In 8,841 African American adults, it regresses a finished score on continuous similarity-to-Yoruba ancestry (GSYRI), which substantially reclassifies top-decile individuals. It also shows that simply adding ancestry as a regression covariate creates spurious "ghost" associations. Its authors stress that post-hoc correction cannot recover signal the training stage never found.
This last point drives our design. Calibration already removes ancestry-driven shifts in the mean and scale of the score. An in-training method that only targets calibration would therefore duplicate it. We target within-neighborhood discrimination, which calibration cannot change, and we report everything after applying calibration to every method.
Core Claim and Mechanism
Let each person $i$ have an ancestry coordinate $u_i$, either top genetic principal components or GSYRI. Choose anchor points $a_1,\dots,a_M$ that cover ancestry space, and a kernel $K(u,a)$, a similarity score that decays with ancestry distance. The local validation loss at anchor $m$ averages the per-person losses $\ell_i(\theta)$ over a held-out set, weighted by the kernel:
$$L_m(\theta)=\frac{\sum_{i\in V} K(u_i,a_m)\,\ell_i(\theta)}{\sum_{i\in V}K(u_i,a_m)}.$$
Training minimizes a weighted loss in which person $i$ receives weight
$$w_i = 1+\sum_{m=1}^M \lambda_m \frac{K(u_i,a_m)}{n_m},$$
where $n_m$ is the kernel mass at anchor $m$. Each multiplier $\lambda_m\ge 0$ rises when the local held-out shortfall exceeds a tolerance and falls otherwise:
$$\lambda_m \leftarrow \mathrm{clip}\big(\lambda_m+\eta\,[\,(R^2_{\text{glob}}-R^2_m)-\epsilon\,],\,0,\,\lambda_{\max}\big).$$
Here $R^2_m$ is the explained variance (or AUC for binary traits) computed with the same kernel weights at anchor $m$, $R^2_{\text{glob}}$ is the global value, $\epsilon$ is a tolerance and $\eta$ is a step size.
Three design choices separate this from the obvious "FairGrad over ancestry bins" idea:
- Kernels instead of bins. A person who is half West African and half European ancestry contributes to both nearby anchors, with no cutoff.
- Multipliers driven by held-out error. Minority training sets are small, so upweighting on training error mostly teaches the model to memorize those few people.
- A cap on the multipliers. Some gap is irreducible. Shorter LD blocks in African-ancestry genomes mean European-discovered marker variants tag the causal variants less well, so a hard constraint may be infeasible. The cap turns the constraint into a bounded pressure, and a pinned $\lambda_m$ diagnoses which regions of ancestry space cannot be closed by reweighting.
We expect the effect to work through variant selection. In an L1-penalized (lasso) score, upweighting under-represented people changes which correlated marker variants survive shrinkage, favoring markers that tag the causal variant in their genomes. This is a dynamic, individual-level counterpart of X-Wing's region-level static shrinkage. We will check it directly by comparing selected variants against X-Wing-style local genetic-correlation annotations.
Research Program
We will implement the weighting on two kinds of model trained on individual genotypes:
- a lasso logistic or linear score, matching DisPred's PRS-Lasso comparator;
- a two-layer neural network.
The neural network tests whether the gains depend on model capacity.
Ancestry coordinates are top principal components projected onto the 1000 Genomes reference, plus GSYRI for comparability with Vargas et al. Anchors are placed by k-means on the coordinates. The kernel bandwidth is chosen so that every anchor keeps a minimum effective validation sample. Bandwidth is the main free parameter, and we will report a sweep over it.
The evaluation is itself a deliverable. It is an "ancestry performance curve": $R^2$ or AUC as a smooth function of ancestry coordinate, with bootstrap bands. Its summary statistic is the worst-anchor value. This also addresses the evaluation problem ARL leaves open, since ancestry coordinates are available for evaluation even where self-reported labels are not.
Baselines
The naive reference is unweighted training of the same lasso and neural models on the same genotypes, followed by Vargas-style GSYRI calibration. Every method in the study gets the identical post-hoc calibration step. That way calibration gains, such as Vargas et al.'s rise in height correlation from 0.24 to 0.32 and the T2D top-versus-bottom-decile odds ratio rising from 7.97 to 10.77 in REGARDS, are credited to no one. A method beats this baseline only if its post-calibration worst-anchor $R^2$ or AUC is higher, with bootstrap intervals excluding zero difference.
Static density weighting is the most dangerous competitor. It weights each person by the inverse of the kernel density of training ancestry at their coordinate, fixed before training. It is run with the same kernel, same anchors and same tuning budget. Its weights depend on how many people are nearby, while ours depend on how badly the model does nearby. Matching it means the dynamic part adds nothing.
FairGrad on discrete super-population bins, using the FairGrad package, isolates the kernel ingredient. It is the strongest existing in-training method, and comparing against it on admixed individuals tests whether continuity matters.
ARL gets the same backbone and the same hyperparameter search budget. It tests whether an adversary can discover the struggling ancestry regions from genotypes alone. Its reported result is a 6.5 percentage point worst-group AUC gain on Adult; no reported result found on Lacuna for ARL on genomic data.
DisPred is the strongest published deep-learning method on ADSP and UK Biobank. The file reports only that it achieves higher AUC for AFR and admixed groups than PRS-Lasso. No reported result found on Lacuna for DisPred's exact AUC values, so we will rerun it on our splits.
X-Wing is the summary-statistic reference. It reports a 35.2% median relative $R^2$ gain in East Asians and 14.1–119.1% relative gains across non-European benchmarks over PRS-CSx and XPASS. We will compare against it where summary statistics allow, while acknowledging that it uses far larger discovery samples than any individual-level method.
Two ablations isolate the key ingredients. The first updates multipliers on training error instead of held-out error. The second removes the cap.
Why not just…?
Why not just calibrate the score afterwards, as Vargas et al. do? Calibration rescales a finished score within ancestry bins. That fixes mean and spread, but it cannot reorder people within an ancestry neighborhood, so it cannot change local AUC. Because we calibrate every method, any remaining gain must come from different learned weights. If our post-calibration curve matches the calibrated unweighted baseline, the proposal fails, and we will say so.
Why not just reweight by inverse ancestry density before training? This is a genuine threat. When error tracks sample scarcity, the two approaches will give similar weights. They separate where they disagree: well-sampled but hard regions (some South Asian or admixed American groups in UK Biobank), and scarce but easy regions. Our anchor-level comparison reports exactly those regions. If density weighting ties everywhere, we will recommend it as the simpler method.
Why not just add ancestry principal components as covariates? Vargas et al. show this produces unstable, spurious ancestry effects. More basically, a covariate shifts predictions without changing which variants the model uses. We include it as an additional reference.
Why not just use ARL and avoid needing ancestry at all? In genomic data the adversary can almost certainly recover ancestry from the genotype itself. ARL may therefore behave much like our method, only noisier. Its failure on COMPAS also suggests that noisy phenotypes may lead it to upweight mislabeled cases. Running ARL head to head under the label-noise test below settles which effect dominates.
Proposed Experiments
Primary setting: UK Biobank. This is 461,579 participants across five super-populations, with training dominated by European ancestry and held-out validation and test sets in every ancestry neighborhood. The traits are height, BMI, LDL cholesterol and type 2 diabetes, spanning highly polygenic and moderately polygenic architectures, which DisPred left untested.
Secondary setting: ADSP. This is Alzheimer's disease with roughly 5,000 variants in 11,640 participants with AFR and admixed members. It allows a direct comparison with DisPred on its home benchmark.
External test: REGARDS. Models trained on UK Biobank are applied, frozen, to REGARDS for T2D and height. That cohort's continuous GSYRI range makes it the cleanest test of performance across admixture.
Decisive checks.
- Dynamic versus static weighting. The proposal survives only if the dynamic method beats static density weighting on post-calibration worst-anchor performance in at least the external REGARDS test.
- Admixed performance. The kernel method should beat binned FairGrad specifically for individuals at intermediate GSYRI. That is where bins force arbitrary assignments.
- Variant selection. Variants gained under dynamic weighting should be enriched in regions of high cross-ancestry local genetic correlation. This would confirm the proposed mechanism and not merely a better-tuned penalty.
- Label noise. We will inject case–control misclassification at ancestry-dependent rates. The test is whether held-out-driven, capped multipliers stay stable where ARL and the uncapped ablation chase noise.
- Cost. We will report European-ancestry accuracy loss and multipliers pinned at the cap as explicit costs. Pinned multipliers mark ancestry regions where the remaining remedy is new data collection, not reweighting.
References
- FairGrad: Fairness Aware Gradient Descent
- Optimising Equal Opportunity Fairness in Model Training
- Scalable and Trustworthy Learning in Heterogeneous Networks
- In-Training Multicalibrated Survival Analysis For Healthcare via Constrained Optimization
- Fairness without Demographics through Adversarially Reweighted Learning
- Improving genetic risk prediction across diverse population by disentangling ancestry representations
- Quantifying portable genetic effects and improving cross-ancestry genetic prediction with GWAS summary statistics
- Ancestry Calibration of Polygenic Risk Scores Improves Risk Stratification and Effect Estimation in African American Adults