Where Polygenic Scores Stop Transferring: Module-Resolved Borrowing Across Ancestries

A polygenic score is a weighted sum of a person's genetic variants that predicts a trait such as height or cholesterol. Scores built from European-ancestry cohorts lose much of their accuracy in other populations. Two findings in the literature seem to contradict each other on why. One study finds that the true causal effects of variants are almost identical between African- and European-ancestry participants for most traits. Another finds that adding European data to an African training set can make African predictions worse. We propose that both are true, but at different places in the genome. Most of the genome's regulatory machinery behaves the same across ancestries, so borrowing European estimates helps there. A small number of regulatory modules carry population-specific architecture, and borrowing there actively hurts. If this is right, the fix is to decide how much to borrow separately for each module, rather than choosing one global mixing weight or picking a training set by trial and error.

Concretely, we group variants into regulatory modules taken from tissue-specific gene-expression networks. For each module we estimate a borrowing weight: how much the target-ancestry score should rely on European effect estimates versus target-only estimates. The weights are fit using only target-ancestry summary statistics and a reference panel of genetic correlation between nearby variants. The idea of a single learned dial between shared and group-specific parameters comes from transfer learning for atomistic models. There, a kernel's interpolation factor trades off treating chemical elements uniformly against weighting them separately. We apply it to a genomic partition that is known to concentrate heritability. The practical goal is to recover more of the heritability that current scores "miss" in non-European populations. The scientific goal is to map where in the genome transfer fails.

Core Claim

Negative cross-ancestry transfer is concentrated in a few regulatory modules. It is not spread diffusely across the genome. If borrowing is tuned per module, a score should match the best of "European-only," "target-only," and "globally pooled" on every trait simultaneously. It should also beat all three on traits with mixed architecture. The learned per-module weights double as a diagnostic: they should drop near zero around known population-specific loci, such as the malaria-selected DARC region for red-cell traits, and stay near one elsewhere.

The claim rests on three prior results, each playing a different role.

The motivation comes from two papers. Fine-scale population structure and widespread conservation of genetic effect sizes between human groups across traits introduced ANCHOR, which uses admixed individuals to separate differences in local genetic correlation structure from differences in causal effects. It reports an effect-size correlation of 0.98 ± 0.07 between African and European ancestry for 47 of 53 traits, with FEV1 and sitting height as exceptions. Optimal strategies for learning multi-ancestry polygenic scores vary across traits compared European-only, ancestry-specific, and pooled LASSO scores across 15 traits and five non-European groups. For mean corpuscular volume (MCV), an African-only score trained on about 5,000 people beat a European score trained on about 300,000, and pooling European data in reduced African accuracy. ANCHOR's correlation is a genome-wide average. Our hypothesis is that it hides a handful of divergent modules, and that those modules are what the LASSO experiments were detecting.

The partition comes from The Importance of Regulatory Network Structure for Complex Trait Heritability and Evolution. That work built variant-to-gene networks from GTEx v8 expression-association data across 29 tissues, split them into modules with the CONDOR community-detection algorithm, and showed with stratified heritability regression that heritability concentrates in "local hubs" within trait-specific modules. We use these modules as the units that borrowing weights attach to.

The borrowing dial comes from Transfer learning for atomistic simulations using GNNs and kernel mean embeddings. That paper reuses a pretrained network as a fixed feature extractor and fits a kernel regression head whose parameter α interpolates between shared and element-specific weighting. We borrow only the design principle: transfer is controlled by an explicit, fitted interpolation, not by retraining end to end. We do not claim the physics models carry over.

Research Program

For target ancestry $T$ and module $k$, the score's variant weights in that module are

$$b_k(\alpha_k) = \alpha_k\, \hat\beta_k^{E\to T} + (1-\alpha_k)\, \hat\beta_k^{T},$$

where each term is defined as follows:

  • $\hat\beta_k^{T}$ are effect estimates from a target-only shrinkage fit.
  • $\hat\beta_k^{E\to T}$ are European causal-effect estimates re-expressed in the target's local correlation structure. We re-express them by fine-mapping in European data and projecting through target-ancestry correlation matrices from 1000 Genomes.
  • $\alpha_k \in [0,1]$ is the module's borrowing weight.

The projection step matters because it separates two things. ANCHOR attributes most failure to correlation structure, so $\hat\beta^{E\to T}$ already corrects that part. Any remaining need to set $\alpha_k<1$ therefore reflects genuine effect divergence, not a tagging artifact.

The weights are chosen without individual-level target data, which addresses a stated limitation of PRSbils and PRISM. Let $r_T$ be the target-ancestry marginal variant–trait correlations and $R_T$ the target correlation matrix among variants. The predicted accuracy of weights $b$ is then

$$\widehat{R^2}(b) = \frac{(b^\top r_T)^2}{b^\top R_T\, b},$$

the standard summary-statistic identity for the squared correlation between a linear score and the trait. We maximize this over $\alpha$ with a mild penalty pulling each $\alpha_k$ toward a trait-level mean.

Using $r_T$ both to fit and to evaluate would overfit. To avoid this we split target summary statistics into two independent halves:

  • Inside UK Biobank, we split by individuals.
  • Outside it, we resample from the distribution of summary statistics implied by the correlation structure.

We fit on one half and score on the other.

Genes and variants that fall in no module, or in several, go into a residual "background" group with its own weight. This makes the approach honest about the omnigenic diffuse component rather than forcing every variant into a module.

The program runs in three steps:

  1. Build module partitions per tissue and select the tissue per trait by module-level heritability enrichment, not by manual assignment (PRISM's stated weakness).
  2. Fit per-module weights for each trait–ancestry pair.
  3. Analyze where $\alpha_k$ falls below one. We compare these modules against ANCHOR's divergent traits and against known selected loci.

Baselines

The naive reference point is the European-only score, the default in current practice. We fit it on the same European UK Biobank samples and apply it unchanged to each target group. Lehman et al.'s results imply it fails badly on MCV, but the file gives no absolute $R^2$ values for it: no reported result found on Lacuna for absolute European-only $R^2$ in these ancestries. We re-derive those numbers ourselves.

The second baseline is a single global mixing weight: the same equation with one $\alpha$ for the whole genome, fit identically. This is the most dangerous simple competitor. It captures any gain that comes merely from mixing European and target information. Our method must beat it on traits Lehman et al. flag as architecture-specific, and must match it on highly polygenic traits like height.

The strongest existing annotation-based method is PRISM, from PRISM: ancestry-aware integration of tissue-specific genomic annotations enhances the transferability of polygenic scores. PRISM feeds a gradient-boosted functional score derived from ENCODE annotations into penalized regression as a variant-level prior. It reports an average 13.10% relative improvement in African-ancestry $R^2$ over annotation-agnostic models, and a 36.5% improvement for FEV1/FVC with lung annotations. We run PRISM on its four traits (lymphocyte count, eGFR, FEV1/FVC, LDL-C) with the same UK Biobank splits and ancestry definitions. Its tuning set is given the same size as our held-out summary half. Beating PRISM means higher African-ancestry $R^2$ on at least three of the four traits, with the gap exceeding the across-split standard deviation.

The fourth baseline is the bilevel shrinkage method PRSbils, from Incorporating functional annotation with bilevel continuous shrinkage for polygenic risk prediction. It uses group-level and variant-level shrinkage over pathways and reports a 9.9% AUC gain for type 2 diabetes over PRS-CS in UK Biobank. That gain was measured within largely European data, and the paper noted inconsistent results across MGI and KoGES. PRSbils is the closest method that also shrinks by groups. We run it with our module partition as its groups, so the comparison isolates cross-ancestry borrowing per group from heritability weighting per group.

Two ablations isolate the key ingredient:

  • Random-module ablation. Variants are shuffled into random groups matched in size and genomic spread. This tests whether regulatory topology matters or any partition would do.
  • Hub-only ablation. Only local-hub variants get their own weights. This tests whether the hub concentration reported in the regulatory-network paper is what drives the gains.

Why not just…?

Why not just pick the best training set per trait, as Lehman et al. did? Trial-and-error selection requires individual-level target data for every trait, and it throws away information. For a trait with one divergent locus in an otherwise conserved background, choosing "African-only" discards the 300,000 European samples that would help everywhere else. Our experiments test this directly. On MCV, the per-module score should beat both the African-only and pooled scores. If it only ties the better of the two, per-module borrowing adds nothing over selection.

Why not just carve out the few known population-specific loci and use one global weight for the rest? This is a real threat. If divergence is truly oligogenic, a global weight plus a hand-listed set of large-effect regions (DARC and a few others) could match our method. We run exactly this baseline, with the excluded regions chosen by the largest target-versus-projected-European effect discrepancies. The number of regions is matched to the number of modules our method sets below 0.5. If this baseline ties us across traits, we will report that divergence is locus-level rather than module-level, and we will drop the module machinery. That would still resolve the ANCHOR–Lehman tension, but it would not justify the method.

Why not just add more functional annotations, as PRISM does? PRISM's own finding argues against scaling annotation count: its best models used 102-fold fewer annotations than the full 414-annotation set. More importantly, annotations say which variants are functional, not which are conserved across ancestries. A variant can be highly functional and still have a divergent effect, as with selected loci. The comparison with PRISM, and a hybrid that uses PRISM's functional score inside each module, will show whether the two kinds of information stack.

Why not use existing joint multi-ancestry Bayesian shrinkage methods? Methods that model cross-ancestry effect correlation genome-wide are widely used, but none appears in this research file, so we cannot cite their reported numbers. They typically assume one cross-ancestry correlation per trait. That is the global-weight baseline in a more elaborate form, and our comparison against the global weight targets exactly that assumption.

Proposed Experiments

The primary evaluation uses UK Biobank with the five non-European groups from Lehman et al. (AFR, AMR, CSA, EAS, MID). We cover their 15 traits plus PRISM's four, with correlation references from 1000 Genomes and UK10K. Modules come from the 29 GTEx v8 tissue networks. Metrics are held-out target-ancestry $R^2$ for quantitative traits and AUC for binary ones, averaged over five random splits. We also report the fraction of target SNP heritability captured, which ties the work to missing heritability. External transfer uses KoGES for East Asian ancestry and MGI for mixed ancestry, where PRSbils behaved inconsistently.

Four checks decide the outcome:

  1. Paradox check. On MCV in African ancestry, the learned weight for the module containing DARC should fall near zero while most other modules stay near one. The resulting score should beat both of Lehman et al.'s African-only and pooled strategies. On height, all weights should be near one and performance should match the European-only score.
  2. Consistency check. Across traits, the fraction of heritability-weighted modules with $\alpha_k<0.5$ should be highest for ANCHOR's divergent traits (FEV1, sitting height) and lowest for traits with conserved effects.
  3. Topology check. The regulatory partition must beat the random-module ablation. If it does not, the gain comes from blocking, not biology.
  4. Locus-carve-out check. The per-module method must beat the global-weight-plus-excluded-loci baseline on at least the architecture-specific traits. Otherwise we conclude divergence is oligogenic and report the simpler method as the recommendation.