Polygenic risk scores add up thousands of small genetic effects to predict disease or trait values. They are trained mostly on people of European ancestry, and they lose much of their accuracy in other groups. In UK Biobank, relative performance fell to 39.4% of the European level for Africans, 46.6% for East Asians and 51.3% for South Asians, as reported in Polygenic risk score portability for common diseases across genetically diverse populations. Three biological causes are usually given. Variants differ in frequency across populations. The correlation between nearby variants differs; geneticists call this linkage disequilibrium (LD), and it determines whether a measured "tag" variant still stands in for the true causal one. And the causal effects themselves may differ. These three causes call for different fixes: denser genotyping and better fine-mapping help with the first two, while only new data from the target population helps with the third. Yet published work reports only the total accuracy drop.
We propose to split that drop into its sources using global sensitivity analysis. This is the method from engineering that asks what share of a model's output variability comes from each uncertain input and from each interaction between inputs. The key step is to choose the right "inputs". We do not treat millions of correlated variants as inputs, which is the setting where existing methods break. Instead, the inputs are three continuous dials, sampled independently. Each dial moves one ingredient (allele frequencies, LD, or causal effects) from its European value to its target-ancestry value. The output is the expected accuracy of a fixed polygenic score in the target population. The variance shares of these dials, together with their interaction terms, give a mechanistic and per-trait account of where portability is lost. We then test whether that account predicts which remedy works for which trait.
Core Claim
Write $\Sigma = D^{1/2} R D^{1/2}$ for the genotype covariance in a genomic region. Here $D$ is the diagonal matrix of per-variant variances $2p(1-p)$, with $p$ the allele frequency, and $R$ is the LD correlation matrix. Let $\beta$ be the vector of causal effects and $w$ the score weights estimated from a European study. In an additive model, the accuracy of the score in a population with parameters $(D, R, \beta)$ is
$$ r^2(w; D, R, \beta) = \frac{(w^\top \Sigma \beta)^2}{(w^\top \Sigma w)\,\mathrm{Var}(y)}, $$
where $\mathrm{Var}(y)$ is the phenotypic variance in that population. We define interpolated parameters. Allele frequencies are $p(\lambda_1) = (1-\lambda_1)p_E + \lambda_1 p_T$, with $E$ for European and $T$ for target. LD is $R(\lambda_2) = (1-\lambda_2)R_E + \lambda_2 R_T$, which stays a valid correlation matrix because it is a convex combination. Effects are $\beta(\lambda_3)$, built so that the cross-ancestry effect correlation moves from 1 to the estimated $\rho$. A fourth dial, $\lambda_4$, scales the size of the European training study. It separates "the weights were noisy" from "the population differs."
The output $Y = r^2_T / r^2_E$ is then a deterministic, cheap function of the $\lambda$'s. Because we sample the $\lambda$'s independently and uniformly, the classical Sobol' decomposition applies exactly. The first-order index $S_k = \mathrm{Var}(\mathbb{E}[Y\mid\lambda_k])/\mathrm{Var}(Y)$ is the share of portability variation explained by mechanism $k$ alone. The gap between total and first-order indices measures interactions between mechanisms. An example is LD decay mattering only when frequencies have also shifted, which is exactly the effect that one-at-a-time ablations miss.
What is new is the design choice. The genomic dependence stays inside $f$, where it is computed exactly from reference panels. It is not placed in the input distribution, where it would require the genome-scale dependent-variable methods that remain unsolved.
Relation to Prior Work
The motivating evidence comes from three multi-ancestry studies. The portability study above measures the aggregate drop. Optimal strategies for learning multi-ancestry polygenic scores vary across traits shows that the best training strategy depends on genetic architecture: for mean corpuscular volume, a score trained on about 5,000 Africans beat one trained on about 300,000 Europeans, driven by the DARC locus, and pooling European data sometimes hurt. Benchmarking PRS methods ... PRS-hub ranks 13 methods on 36 traits. All three describe outcomes. None says how much of a given trait's loss is due to frequency, LD, or effects. Our decomposition is meant to explain their findings, and it will be judged by whether it predicts them.
The method comes from the sensitivity analysis literature. We reuse the Sobol' pick-freeze estimators and their confidence intervals from Statistical inference for Sobol pick-freeze Monte Carlo method, with parallel computation via GlobalSensitivity.jl.
The closest conceptual neighbour is Fairness seen as global sensitivity analysis. It maps group-fairness criteria onto Sobol' indices of a protected attribute and handles correlated proxies through a Rosenblatt transform, on the low-dimensional Adult and COMPAS benchmarks. Its indices answer "how much does the prediction depend on group membership?" We ask a different question: "what parts of genetic architecture account for the accuracy gap between groups?" We also avoid estimating a joint input distribution entirely.
The dependent-input decompositions act as a constraint. Generalized Hoeffding-Sobol decomposition for dependent variables and Generalized Sobol sensitivity indices for dependent variables: numerical methods were validated on at most 8 variables, and their indices can turn negative. They show why treating variants as inputs is the wrong design.
Two lines of work apply attribution to neural models. CRAFT and COCKATIEL apply Sobol' indices to a small number of learned concepts in vision and language models. Deep interpretability for GWAS uses a reference-based attribution method on single variants and does not address ancestry. Our move resembles the concept papers: we run the variance decomposition over a few meaningful factors rather than raw features. The difference is that our factors are defined by population genetics, not learned.
Research Program
The first step is building $f$. We will use UK Biobank genotypes with Pan-UKBB ancestry labels, which cover European, African, Central/South Asian, East Asian, Admixed American and Middle Eastern groups. We compute $D$ and $R$ per LD block for each group. Following the PRS-hub finding that gains plateau above about 1,000 reference individuals, we use panels of at least that size where they exist. We also use the curated 1.1-million-variant set, which outperformed 6.6 million variants in that benchmark. Accuracy is summed across approximately independent LD blocks, so each evaluation of $f$ costs a set of block-wise quadratic forms. Saltelli designs with tens of thousands of base samples are therefore affordable.
The second step is calibrating the effect dial for each trait. We will estimate $\rho$ and the population-specific large-effect loci from ancestry-specific summary statistics for the 15 traits in Lehman et al. Then $\beta(\lambda_3)$ mixes a shared component with an ancestry-specific one. Because true $\beta$ is unknown, this dial is the weakest link. We therefore run every analysis under a range of plausible $\rho$ values and report how stable the indices are across that range.
The third step extends the output. Clinical use depends on who lands in the top risk decile, and variance-based shares can miss shifts in distribution shape. As a secondary output, we therefore add the overlap between true and predicted top-decile membership, and we compute Cramér–von Mises indices as described in Sensitivity Analysis Based on Cramér-von Mises Distance.
The fourth step adds a training-strategy dial covering European-only, pooled, and target-only training. The decomposition then explains the variance of the gain from pooling, and it locates the region of $(\rho, \text{target sample size})$ where adding European data starts to hurt.
Baselines
The naive reference point is the aggregate portability ratio reported by the existing studies. The strongest published figure for comparison is relative performance of 39.4% for Africans versus Europeans in Polygenic risk score portability for common diseases across genetically diverse populations. Our $f$ must first reproduce the observed per-trait ratios in UK Biobank African and South Asian individuals, within sampling error, before any decomposition is trusted. Failing this check would end the project.
The strongest diagnostic competitor is the one-at-a-time counterfactual swap: replace only the European LD with target LD, record the accuracy change, and repeat for frequencies and effects. It uses the same $f$ and the same reference data, so the comparison isolates the decomposition itself. Our method beats it if the swap contributions fail to sum to the total loss, and if the leftover is concentrated in traits where the interaction terms in our decomposition are large. An honest null result is that interactions are negligible, in which case the cheap swap is sufficient.
The second diagnostic competitor is the eight-cell factorial: each mechanism at its European or target endpoint, attributed through ANOVA or Shapley values. This is the ablation that isolates the one thing our design adds, which is continuous, uniformly sampled dials instead of endpoints. Continuous dials are only worth the extra computation if indices computed on curves differ in ranking from endpoint indices on a meaningful fraction of traits. They should also locate thresholds, such as the $\rho$ at which pooling stops helping, that endpoints cannot see.
For prediction, the external competitors are LDpred2 and LDpred2-meta, run through PRS-hub on the same splits. PRS-hub reports LDpred2 as the top single-ancestry method on 14 of 19 diseases and 9 of 17 quantitative traits. It reports LDpred2-meta outperforming PRS-CSx and X-Wing in African cohorts, though no reported numerical African-cohort result for LDpred2-meta was found on Lacuna. Linear ensembles add a median 5–10% in odds ratio per standard deviation for traits like schizophrenia. We do not claim to outpredict these methods. The test is whether a decomposition-guided choice per trait matches or exceeds the PRS-hub ensemble in target-ancestry accuracy, using the same tuning budget and the same validation split. The three choices are fine-mapping-aware weights for LD-dominated traits, frequency-matched variant selection for frequency-dominated traits, and target-only or reweighted training for effect-dominated traits.
Why not just…?
Why not just use the analytic formula for expected accuracy? In the purely additive case with known effects, the displayed equation already is the answer, and its components could be inspected directly. That is a real point, and our $f$ is exactly that formula. The formula alone, however, does not apportion a single number among three interacting ingredients. The ratio is nonlinear in $D$, $R$ and $\beta$ jointly, so "how much is LD" has no unique answer without an attribution rule. Sobol' indices over independent dials supply a rule with a variance interpretation and confidence intervals. If the interaction indices turn out tiny, the formula plus simple swaps is enough, and we will say so.
Why not just run the counterfactual swaps? This is the most serious threat, and the swap is run as a baseline for that reason. The swap approach and our method differ only when mechanisms interact. Frequency and LD plausibly do interact, because a tag variant's usefulness depends on both its own variance and its correlation with the causal variant. The comparison described under baselines settles the question empirically.
Why not just reweight or ensemble and skip the diagnosis? Optimal strategies for learning multi-ancestry polygenic scores shows that reweighting inflates variance and fails when $\rho$ is near 0.5. PRS-hub's ensembles improve prediction but do not say whether new data from the target population or better fine-mapping would help more. That is a funding and study-design question the decomposition is meant to answer. If the ensemble already captures everything a decomposition-guided choice would deliver, the predictive half of our claim fails while the explanatory half stands.
Decisive Checks
Three results would make or break the proposal.
First, the decomposition must recover the known story for mean corpuscular volume. Effect and frequency indices should dominate, driven by DARC, with a small LD share. Height should show the opposite pattern, with LD and training-size indices dominating, consistent with the finding that sheer European sample size wins for height.
Second, across the 15 traits, a larger effect-heterogeneity index should predict the traits where African-only training beats pooled training in Lehman et al.'s setup. We will measure this with a rank correlation on held-out traits, specified before the analysis.
Third, the indices must be stable under the full range of plausible $\rho$ values and under resampling of the LD reference panel. If the ranking of mechanisms flips within that range, the per-trait attributions are not identifiable from current data, and we would report that the decomposition is uninformative for those traits.
References
- Polygenic risk score portability for common diseases across genetically diverse populations
- Optimal strategies for learning multi-ancestry polygenic scores vary across traits
- Benchmarking PRS methods for risk prediction of 36 complex traits in UK Biobank by establishment of a large-scale PRS computation platform PRS-hub
- Statistical inference for Sobol pick-freeze Monte Carlo method
- GlobalSensitivity.jl: Performant and Parallel Global Sensitivity Analysis with Julia
- Fairness seen as global sensitivity analysis
- Generalized Hoeffding-Sobol decomposition for dependent variables - application to sensitivity analysis
- Generalized Sobol sensitivity indices for dependent variables: numerical methods
- CRAFT: explaining using Concepts from Recursive Activation FacTorization
- COCKATIEL: COntinuous Concept ranKed ATtribution with Interpretable ELements for explaining neural net classifiers on NLP
- Deep interpretability for GWAS
- Sensitivity Analysis Based on Cramér-von Mises Distance