Do Gene–Gene Interactions Explain Why Genetic Risk Scores Fail Across Ancestries?

Genetic risk scores are built by adding up small per-variant effects estimated in large, mostly European cohorts. They lose much of their accuracy when applied to people of African, East Asian or South Asian ancestry. The usual explanation is that the measured variants only stand in for the true causal ones, and how closely they track those causal variants (linkage disequilibrium, or LD) differs between populations. This note tests a second explanation that has received little direct study. Suppose a variant's effect depends on which alleles a person carries at two, three or more other loci (epistasis). Then the "average effect" a standard association study reports for that variant is an average over the partner genotypes common in the study population. Move to a population where those partner alleles have different frequencies, and the average effect changes. It can even reverse sign, while the underlying biology stays the same.

The proposal has two parts. The first is to fit a sparse, high-order interaction model in European data, using only interactions between loci far enough apart that LD cannot fake them. The second is to project that model onto the allele frequencies of another ancestry and predict that ancestry's per-variant effects without using any of its phenotypes. If the projected effects predict observed cross-ancestry changes in effect size and sign better than LD-based transfer does, then interactions account for part of the portability gap. If they do not, the experiment puts an empirical upper bound on how much epistasis can matter, a question the field currently answers by assumption.

Core Claim

Take an interaction model fitted in a source population. Its additive projection, evaluated at a target population's allele frequencies, predicts part of how per-variant effects change between populations, and this part is not captured by LD differences. The mechanism has a closed form. Code the genotype at locus $j$ as $x_j$ and write the trait as a sum of products over small sets of loci $S$ with coefficients $w_S$. The marginal effect of locus $j$ in population $p$ is then approximately

$$ \beta_j^{(p)} \approx \sum_{S \ni j} w_S \prod_{k \in S \setminus \{j\}} \mathbb{E}_p[x_k], $$

where $\mathbb{E}_p[x_k]$ is the mean genotype at partner locus $k$ in population $p$. This is the standard quantitative-genetics average effect, and it holds exactly when the partner loci are independent of $j$ and of each other in $p$. That independence condition is why the model is restricted to partners in separate LD blocks. Only frequencies are needed to transfer the effects, and public reference panels supply them. The distinctive prediction is that frequency differences at other loci drive changes in effect size and sign at a given locus, including flips like the MSRA reversal on BMI between White and Chinese cohorts reported by Multi-ancestry polygenic risk scores using phylogenetic regularization.

Relation to Prior Work

Two lines of work motivate the gap. On cross-ancestry cancer polygenic risk scores shows that European-derived breast and prostate cancer scores shift badly in absolute terms across ancestries in UK Biobank, yet still rank individuals well within each ancestry. Multi-ancestry polygenic risk scores using phylogenetic regularization introduces DendroPRS. It lets per-variant effects drift along an ancestry tree and finds direction flips. Both models are purely additive, so neither can ask whether the shifts they observe come from interactions. The proposal keeps their evaluation settings and replaces the additive assumption with a mechanism that can be tested.

Three papers supply the interaction machinery, and each was developed for a single population. iterative Random Forests to discover predictive and stable high-order interactions finds stable interactions up to sixth order by intersecting decision-tree paths. It serves here as the candidate-discovery step. Higher-Order Factorization Machines fits all $m$-way interactions in time linear in the number of features, using a dynamic-programming recursion over products of distinct features. It is the dense alternative for estimating $w_S$. Selective Inference for Sparse High-Order Interaction Models gives valid p-values for interactions found by searching spaces of more than $10^{17}$ candidates. It reports that data splitting wastes power and that naive testing gives about 100% false positives, so it is used to certify which interactions enter the model. None of these three tools models population structure. The proposal does not change their algorithms. It uses their output as the input to a cross-population projection.

Two papers constrain the design. Unveiling Challenges in Mendelian Randomization for Gene-Environment Interaction found no widespread gene-by-BMI interaction on colorectal cancer. That is a warning that interaction signals in human complex traits may be weak, and the proposal must be able to return an informative null. The same paper shows that interaction estimates for binary outcomes are biased because the logistic model is non-collapsible: odds ratios do not average cleanly over subgroups. For this reason the primary analyses use continuous traits and gene expression. Toward Provably Correct Feature Selection in Arbitrary Domains shows that pure parity-type interactions have no marginal signal at all, so any discovery step that relies only on marginal screening is ruled out. The permutation-free kernel tests of Permutation-Free High-Order Interaction Tests provide a nonparametric check that discovered sets are genuinely jointly dependent.

Research Program

The work has three stages.

Discovery. In European UK Biobank participants, interacting sets of up to fourth order are found among variants in distinct LD blocks. Candidates are proposed by iRF and certified by selective inference, with family-wise false positive rate held at 0.05. Excluding within-block interactions is the main safeguard against "phantom epistasis," in which an additive causal variant that is poorly tagged looks like an interaction between its tags.

Projection. The certified coefficients $w_S$ are projected with the equation above. Target-ancestry mean genotypes come from the 1000 Genomes reference panel, and the target ancestries are African, East Asian and South Asian.

Evaluation. The projected effects are compared with effects estimated directly in each ancestry. They are also used both as a zero-shot score and as an informative prior inside DendroPRS.

A simulation stage comes first and calibrates the whole pipeline. Phenotypes are generated on real 1000 Genomes haplotypes under three architectures: purely additive effects with LD mismatch, cross-block epistasis, and a mixture of the two. This shows whether the pipeline separates these cases and how much interaction variance it needs before it can detect it.

Baselines

The naive reference is the European additive score applied unchanged to each target ancestry. It represents doing nothing beyond what current practice does, and it corresponds to the GPRS and CSPRS scores in On cross-ancestry cancer polygenic risk scores.

The strongest competitor when some target-ancestry phenotypes are available is DendroPRS. That paper reports a genotype–phenotype correlation of 0.14 for BMI in aggregated non-White UK Biobank groups, against 0.10 for single-population models and 0.03 for naive pooling. DendroPRS will be rerun on the same UK Biobank ancestry splits, the same variant set and the same tuning budget, with and without the projected effects as its root prior. Beating it means exceeding its rerun correlation on held-out non-European participants with a bootstrap interval that excludes no improvement.

The strongest competitor in the zero-shot setting is LD-aware additive transfer. This means a fine-mapped European additive model with effects re-tagged using the target ancestry's LD from 1000 Genomes. It is the most direct representation of the "it is all LD" explanation. The projection beats it if it explains significantly more variance in observed target effect sizes, and predicts sign flips more accurately, among variants with well-estimated target effects.

The key ablation uses the identical certified interaction model but projects it at European frequencies. This isolates the frequency-transport ingredient: any gain over this ablation comes from target allele frequencies, not from the extra model capacity.

For the cancer endpoints, the within-ancestry decile reference is the breast cancer top-10% odds ratio of 2.81 (EUR), 2.88 (AFR), 2.60 (EAS) and 2.33 (SAS) from On cross-ancestry cancer polygenic risk scores. These endpoints are secondary because of the non-collapsibility problem.

Why not just…?

Why not just recalibrate the additive score within each ancestry? Within-ancestry percentile ranking does preserve breast cancer enrichment, at odds ratios of 2.33 to 2.88 (On cross-ancestry cancer polygenic risk scores). But recalibration only rescales a single score. It cannot reverse the sign of an individual variant's effect, and it does not explain why effects change between populations. For risk stratification alone it is a real competitor, and the proposal does not claim to beat it on decile odds ratios. The claim concerns per-variant effects and correlation with continuous traits, and those comparisons are reported separately.

Why not attribute everything to LD mismatch and fine-map? This is the most serious threat, because phantom epistasis can make LD artifacts look like interactions. The proposal handles it in three ways: it allows only cross-block interactions, it includes the LD-aware transfer baseline described above, and its simulations include additive-with-LD-mismatch architectures that must yield no spurious frequency-projection gain. If the LD baseline matches the projection on real data, the conclusion is that LD is enough, and that result would be reported as such.

Why not just fit a pooled nonlinear model, such as boosted trees, with an ancestry covariate? A pooled model could pick up interactions implicitly. But it needs substantial target-ancestry phenotype data, and pooling performed badly in the DendroPRS comparison (correlation 0.03). It also produces no interpretable mechanism for any shift it captures. It will be run as an additional with-target-data comparison on equal data. If it matches the projection on prediction, the remaining contribution is explanatory: identifying which partner-frequency shifts drive which flips.

Proposed Experiments

The primary testbed is Geuvadis gene expression, which pairs European and Yoruba lymphoblastoid samples. DendroPRS improved prediction there for 67 genes. Expression is a continuous trait with large per-variant effects, so interactions are most likely to be detectable. The second testbed is UK Biobank BMI and height across ancestry groups.

The decisive check has two parts. First, among variants whose estimated effects differ significantly between European and target samples, does the frequency projection predict the direction and size of the change better than both LD-aware transfer and the European-frequency ablation? Second, does it recover the MSRA BMI flip?

A null result in which the projection adds nothing beyond LD, despite simulations showing adequate power, would be a useful bound on the role of epistasis in portability. A positive result would justify building interaction-aware priors into multi-ancestry scores. Binary cancer endpoints come last. They are evaluated on the liability scale rather than the odds-ratio scale, given the bias in logistic interaction estimates documented by Unveiling Challenges in Mendelian Randomization for Gene-Environment Interaction.