A polygenic risk score is a weighted sum of a person's genetic variants that predicts a trait such as body mass index or heart disease. These scores are usually trained on men and women pooled together, because pooling doubles the sample size. Sex-stratified studies keep finding variants whose effects exist in only one sex. In African-ancestry adiposity data, for example, splitting by sex revealed seven loci that the pooled analysis missed, while one locus appeared only when the sexes were pooled (AAAGC adiposity GWAS). The field treats "pool or split" as a single binary choice made once per trait.
We propose to treat it as a stability question from cooperative game theory instead. Each cohort-by-sex group in a biobank federation is a player. A shared score is "core-stable" when no subset of players could train its own score and do strictly better for every one of its members. We will measure whether today's standard pooled scores are stable in this sense. We will also test whether the subsets that would profitably break away line up with sex, with ancestry or recruitment site, or with nothing at all beyond sampling noise. Finally, we will test whether training the score to maximize Nash social welfare (the sum of the logarithms of players' utilities) removes those breakaways at little cost in total accuracy. The core is the set of allocations no subgroup wants to leave. Its link to Nash welfare is exact only for convex losses, and penalized linear polygenic scores fall squarely in that convex case.
Core Claim
The most violating coalition of a shared polygenic score is an informative object in two ways. It shows who is being shortchanged by pooling. Its composition, compared against a permutation null, tests whether that shortchanging comes from sex dimorphism. The claim has two halves. The first is an empirical audit: for traits with documented sex-specific loci, such as waist-to-hip ratio adjusted for BMI, sex-combined scores will admit blocking coalitions made of female players. For traits without such loci, such as height, they will not. The second is a method: a Nash-welfare-trained score will be core-stable in that audit.
The method half rests on a theorem from Fairness in Federated Learning via Core-Stability. That paper proves that, for convex losses, maximizing $\sum_i \ln(M - \mathrm{Loss}_i(\theta))$ yields a core-stable predictor $\theta$. Here $M$ is a constant just above the maximum loss and $\mathrm{Loss}_i$ is player $i$'s loss. The paper validated this only on CIFAR-10, MNIST and Adult with artificial label skew, and its guarantee is merely local for the neural networks it trained. A penalized linear score under squared loss is convex, so the guarantee becomes global. Heterogeneity between players is also real biology rather than injected noise, including hormone-responsive regulation and X-chromosome dosage effects (X-chromosome and brain architecture).
Research Program
Players. We will build the federation from UK Biobank individual-level data. Each player is one assessment centre crossed with sex, giving roughly forty players. Each player keeps a private held-out set, and its utility is the variance in the trait that the score explains (R²) on that held-out set. A second, summary-statistic-level federation uses AAAGC member cohorts split by sex where possible. The Women's Health Initiative, being all-female, forms a natural pre-existing female coalition.
Coalition values. The value of a coalition $S$ comes from training a score on $S$'s pooled data with the same sparse inference engine used everywhere. We use spike-and-slab LASSO via SSLPRS and its penprs code. That paper reports genome-wide inference in about three minutes on UK Biobank, which makes evaluating many coalitions affordable.
Searching for breakaway coalitions. Forty players imply about $2^{40}$ coalitions, so exhaustive checking is out. We use the remedy of bounded or structured search noted in Coalition Formation Methods in Multi-Agent Environments. We exhaustively evaluate structured families: all-female, all-male, region blocks, ancestry groupings and their intersections. We then run a greedy local search that adds and removes players to maximize the minimum gain over members. The reported quantity is the largest $\epsilon$ such that some coalition $S$ satisfies $u_i^{S} \ge u_i^{\text{shared}} + \epsilon$ for all $i \in S$. Here $u_i^{S}$ is player $i$'s held-out R² under $S$'s own score, and $u_i^{\text{shared}}$ is its R² under the shared score. Because the search is incomplete, $\epsilon$ is a lower bound on instability.
Telling biology from noise. This addresses the gap the research file flags for CLiP, which detects hidden heterogeneity from correlations among score variants but says nothing about how to partition players. We repeat the audit after permuting sex labels within centres and after bootstrapping each player's held-out set. A sex-aligned blocking coalition is called biological only if it survives both.
The candidate fix. The predictor class is a shared effect vector plus a sparse female-deviation vector. One instance is trained by maximizing Nash welfare over players, with $M$ set from the worst single-player loss and checked for sensitivity. The other is trained by pooled squared error.
Baselines
Naive reference. The do-nothing reference is the standard sex-combined score: SSLPRS trained on all players pooled with a sex covariate. This is what is deployed in practice, and the audit starts here.
Strongest existing pooled method. This is LDpred-funct, which adds functional annotation priors. It reports an average prediction R² of 0.144 across 21 UK Biobank traits, peaking at 0.413 for height, a 4.6% relative improvement over SBayesR. We will rerun it on our identical training splits rather than quote those figures. "Beating it" here does not mean higher pooled R², which we do not expect. It means a smaller audited $\epsilon$ with pooled R² within one bootstrap standard error.
Fully stratified scores. These are separate male-only and female-only scores, again built with SSLPRS. They show what splitting costs in power. The AAAGC locus that appeared only under pooling suggests that splitting is not free.
Strongest borrowing competitor. This is a two-trait genetic-correlation model treating male and female phenotypes as correlated traits, following the multi-trait GBLUP approach with global or local genetic correlations (multi-population GBLUP; local-correlation multi-trait prediction). These papers estimate shared and divergent effects by mixed-model statistics with a single aggregate objective. They contain no notion of subgroup incentives. No reported result was found on Lacuna for sex-as-two-traits prediction.
Egalitarian min-max. This score minimizes the worst player's loss over the same model class. Fairness in Federated Learning via Core-Stability argues that this approach overfits to the noisiest players.
Isolating ablation. The deviation-term model trained by pooled squared error, compared with the same model trained by Nash welfare, separates the effect of the objective from the effect of the model class.
Fairness of comparison. All methods share the same LD reference, the same training and held-out splits, an equal hyperparameter grid size tuned on a common validation fold, and the same audit search budget.
Why not just…?
Why not just run sex-stratified GWAS and test SNP-by-sex interactions? Interaction tests flag individual variants. They do not say whether a deployed polygenic score serves each group as well as that group could serve itself, and they are silent about cohort and site heterogeneity. Still, the interaction-informed stratified score is a baseline here. If it is already core-stable with no loss in pooled R², we will report that the game-theoretic audit adds diagnosis but not a better predictor.
Why not just fit a two-trait genetic-correlation model so each sex gets its own predictor? This is the real threat. With per-sex weights, the two-player game is nearly trivial: each sex simply compares its tailored score with going alone. Our audit has bite only because players within a sex still share a predictor across centres and ancestries. The ablation that keeps the deviation-term model class but swaps pooled error for Nash welfare settles the question. If the pooled-error version shows $\epsilon$ indistinguishable from zero, then the model class, not the objective, does the work, and we will say so.
Why not just minimize the worst-off group's loss? Ray Chaudhury et al. show that min-max fairness bends the shared model toward low-quality players. In genetics, a sex with lower heritability or a noisier trait definition plays this role; arm circumference, for instance, measures different tissues in men and women (sex-stratified architecture summary). We expect min-max to produce blocking coalitions of high-heritability players. The audit measures this directly.
Why not just calibrate a pooled score separately by sex? Recalibration rescales the score but cannot recover sex-specific variants that the pooled sparse inference has already set to zero. The smoking Mendelian-randomization analysis found that sex-combined genetic instruments largely obscured sex differences, except for subarachnoid haemorrhage, where females showed a ratio of odds ratios of 2.61 (95% CI 1.06–6.42) (Smoking and CVD in UKB). We include sex-recalibrated pooled scores in the audit.
Proposed Experiments
Traits. The primary traits are BMI and waist-to-hip ratio adjusted for BMI, chosen for their documented sex-specific loci. Height is a low-dimorphism negative control. Coronary artery disease serves as a binary trait, motivated by the female-specific effects of loneliness polygenic risk reported in BioVU and ARIC (Dennis et al.).
Decisive check one. For waist-to-hip ratio, the pooled baselines should show $\epsilon > 0$, attained by majority-female coalitions that survive sex permutation. For height, $\epsilon$ should be indistinguishable from the permutation null. If waist-to-hip ratio shows no sex-aligned blocking, the premise that pooling shortchanges one sex at the score level fails.
Decisive check two. The Nash-welfare score should drive audited $\epsilon$ to the bootstrap noise floor while keeping pooled R² within one standard error of LDpred-funct. It should also beat the pooled-error deviation model on $\epsilon$; if it does not, the objective adds nothing.
Decisive check three. We will replicate at summary-statistic level in AAAGC. The test is whether the WHI-containing female coalition blocks the sex-combined score for waist-to-hip ratio but not for BMI. This mirrors the reported pattern, in which all four novel stratified loci for waist-to-hip ratio were female-specific, whereas the novel BMI loci split between men and women.
Secondary outputs. We will report the composition of blocking coalitions, how often site- or ancestry-aligned coalitions appear rather than sex-aligned ones, and the agreement between sex-aligned instability and CLiP heterogeneity statistics computed per player.
References
- Discovery and fine-mapping of adiposity loci using high density imputation of genome-wide association studies in individuals of African ancestry: African Ancestry Anthropometry Genetics Consortium
- Fairness in Federated Learning via Core-Stability
- The pivotal role of the X-chromosome in the genetic architecture of the human brain
- Sparse polygenic risk score inference with the spike-and-slab LASSO
- Coalition Formation Methods in Multi-Agent Environments
- Leveraging correlations between polygenic risk score predictors to detect heterogeneity in GWAS cohorts
- Incorporating functional priors improves polygenic prediction accuracy in UK Biobank and 23andMe data sets
- Improving multi-population genomic prediction accuracy using multi-trait GBLUP models which incorporate global or local genetic correlation information
- Improving multi-trait genomic prediction by incorporating local genetic correlations
- Genetic risk between the CACNA1I gene and schizophrenia in Chinese Uygur population
- Smoking as a Risk Factor for Cardiovascular Disease in Females and Males: Observational and Mendelian Randomisation Analyses in the UK Biobank
- Genetic risk for major depressive disorder and loneliness in sex-specific associations with coronary artery disease