Genetic risk scores add up the small effects of many DNA variants to predict a trait. They are trained mostly on people of European ancestry, and they predict much worse in everyone else. The field's main question is how much to borrow from large European datasets when building a score for an under-represented ancestry. The best existing evidence says the answer depends on the trait. For height, a European model trained on hundreds of thousands of people wins everywhere. For mean red-cell volume, an African-ancestry model trained on about 5,000 people beats a European model trained on about 300,000, and adding European data actually hurts (Optimal strategies for learning multi-ancestry polygenic scores vary across traits). We think the trait is the wrong unit for this choice. Red-cell volume is driven by one large ancestry-specific locus (DARC, shaped by historical malaria selection) on top of a polygenic background that is probably shared. The decision to share or not should be made region by region of the genome.
What has blocked a region-by-region decision is the lack of the right units. Variants are correlated with their neighbours, a pattern called linkage disequilibrium (LD), and that correlation differs across ancestries. A variant can therefore look as if it has a different effect in Africans than in Europeans only because it tags the true causal variant differently. Our proposal borrows a tool from graph-based problem decomposition: split a problem wherever its interaction graph has no edges or only narrow bottlenecks. We build one LD graph that merges the correlations seen in every ancestry and cut it into blocks that are uncorrelated with each other in all populations at once. Within each such block, apparent effect differences cannot come from LD leaking across block boundaries. That makes each block a legitimate place to choose between full sharing, tree-structured partial sharing, and ancestry-specific estimation. The same cut also splits one huge regression into many independent small ones.
Core Claim
Our central claim is about the cross-ancestry LD graph. Take the union of each population's thresholded LD graph (variants as nodes, an edge wherever $|r|$ exceeds a threshold in any ancestry) and decompose it into connected components, refined further at bridges and cut-vertices. The resulting blocks support three things at once. First, the genome-wide score is a sum of block scores that can be fit independently. Second, the share-versus-separate decision can be made per block by held-out validation, which recovers the gains of ancestry-specific modelling at hotspots like DARC without giving up European information elsewhere. Third, this per-block choice will match or beat the best single per-trait strategy on every trait. The largest gains should come on traits with mixed architecture. On uniformly polygenic traits like height we expect no loss.
What would be new is the unit of decision. No method in the literature we found makes sharing decisions over blocks defined to be separable in every ancestry simultaneously. Existing approaches either choose one strategy per trait, choose per SNP, or define blocks from a single cohort.
Relation to Prior Work
Two papers supply the motivation. The UK Biobank study of 15 traits across African, Admixed American, Central/South Asian, East Asian and Middle Eastern ancestries (Optimal strategies…) shows that the best sharing strategy flips between traits. In its simulations, ancestry-specific models almost always won when the cross-ancestry correlation of genetic effects $\rho$ was 0.5, and reweighted dual-ancestry models won as $\rho$ approached 1.0. We take this further: if $\rho$ varies by region within a trait, a per-trait choice is a compromise. The paper also motivates a constraint we must respect. Importance reweighting toward minority ancestries reduces bias but shrinks effective sample size, so any per-block selection we do must be guarded against the same variance problem.
DendroPRS (Multi-ancestry polygenic risk scores using phylogenetic regularization) models each SNP's effect as a root effect plus sparse offsets along a population tree. It is our closest competitor because its sparse offsets are already a per-SNP sharing mechanism. We reuse it as one of the per-block options. We differ in two ways. Our decisions are made at the level of LD-separable blocks rather than individual SNPs, which matters because per-SNP offsets can absorb differences in LD tagging and mistake them for real effect differences. We also fit blocks independently, which addresses the scaling limit to millions of variants that DendroPRS leaves open.
Three papers constrain or supply the decomposition method. BALD (Performance of a blockwise approach in variable selection using linkage disequilibrium information) clusters adjacent SNPs by $1-r^2$ within one cohort and runs Group Lasso over the blocks. It beat single-marker tests, Lasso and Elastic-Net when blocks held more than two causal SNPs. That supports blocks as the right selection unit. However, its blocks come from one population and so can split differently in another. Bridge-block decomposition (Learning Large-Scale MTP2 Gaussian Graphical Models via Bridge-Block Decomposition) cuts a thresholded covariance graph at bridges and solves the blocks in parallel, with one to four orders of magnitude speedup at $p=20{,}000$. Its exactness guarantee relies on an $MTP_2$ sign structure that LD matrices do not have, and its authors note that dense graphs without bridges gain little. We therefore treat bridge and cut-vertex cuts as approximations whose cost we measure, not as exact. The portfolio decomposition pipeline (Decomposition Pipeline for Large-Scale Portfolio Optimization…) cleans correlation noise with random-matrix theory and partitions by modularity. It reported solutions within 4% to 5% of the optimum. That is a warning that heuristic cuts carry an optimality cost, and we will quantify the analogous cost.
The block-cut-vertex tree used in network visualization (BC tree-based spectral sampling…) gives us the data structure for refining large components at cut-vertices.
Research Program
Building the merged graph. For each ancestry $a$ we estimate an LD matrix $R^{(a)}$ from in-sample UK Biobank genotypes, with 1000 Genomes superpopulations as an external check. We then form the merged graph $$ G_\tau = \Big(V,\ \{(i,j): \max_a |R^{(a)}_{ij}| > \tau\}\Big), $$ where $V$ is the set of variants and $\tau$ is a correlation threshold. By construction, any two connected components of $G_\tau$ have all cross-component correlations below $\tau$ in every ancestry.
Refining large components. Long European LD ranges can merge regions into very large components. Where that happens, we cut at cut-vertices and bridges, using the block-cut-vertex tree. A cut-vertex variant is duplicated into a single anchor block. For every cut we record the discarded LD mass, so the approximation cost is visible. Optionally, we denoise $R^{(a)}$ against the Marchenko–Pastur distribution before thresholding, following the portfolio pipeline, so that noise in small ancestry samples does not create spurious edges.
Per-block model choice. Within each block $b$ we fit three candidate models on training individuals:
- a shared-effect penalized regression on all ancestries;
- a DendroPRS-style tree-regularized model;
- separate ancestry-specific models.
We pick one per block and per target ancestry using validation likelihood. Because blocks are nearly uncorrelated, the validation loss of the summed score decomposes approximately into block terms, so the choice reduces to independent local comparisons.
Guarding against selection noise. Thousands of noisy local choices could overfit, which is the same variance trap that reweighting falls into. To prevent this, we place a hierarchical prior over choices, defaulting to sharing and switching only when the validation gain exceeds a threshold set by permutation. All selection happens in nested cross-validation, and final numbers come from a held-out test fold.
Extension to causal inference. As a secondary aim, the same blocks give an LD-aware unit for choosing Mendelian randomization instruments across ancestries. This addresses the independence assumption that MAGIC flags as open (Mediation Analysis with Mendelian Randomization and Efficient Multiple GWAS Integration). We will pursue this only if the prediction results hold.
Baselines
Naive references. The first reference point is a European-only model applied to every ancestry. The second is naive pooling of all individuals. DendroPRS reports BMI correlation in aggregated non-White UK Biobank groups of 0.03 for pooling and 0.10 for single-population models (Multi-ancestry polygenic risk scores…). We rerun both on our exact splits rather than quoting these figures.
Best per-trait strategy. This is the oracle per-trait choice from Optimal strategies…: for each trait and target ancestry, pick the best of European-only, ancestry-specific, pooled, and reweighted LASSO by validation. It is the most direct competitor to our claim, since it makes the same decision at coarser granularity. Our method uses the same candidate models, data, folds and tuning grid. Beating it means higher test correlation in the target ancestry on mixed-architecture traits, with no significant loss on height. For red-cell volume the paper reports a qualitative win of the 5,000-person African model over the 300,000-person European model, but no reported numerical result was found on Lacuna for that comparison. We will compute it ourselves.
DendroPRS, genome-wide. This is the strongest existing method for sharing structure, with a reported BMI correlation of 0.14 in non-White UK Biobank groups (Multi-ancestry polygenic risk scores…). We run it on the same SNP set, with its regularization strengths tuned under the same validation budget. Where genome-wide fitting is infeasible, we run it per chromosome and report that restriction. Beating it means a higher BMI correlation than our rerun of DendroPRS, and better accuracy than it on recovering divergent regions in simulation.
Ablations. The ablations isolate the key ingredient. We replace the merged graph with three alternatives: blocks defined from European LD only, BALD-style contiguous blocks from the pooled cohort, and fixed 1 Mb windows. Each keeps the per-block selection unchanged. We also run the merged-graph blocks with the selection step disabled, so that every block uses the tree model.
Why not just…?
Why not just pick the best strategy per trait? Per-trait selection is cheap and already validated. If architecture is homogeneous within traits, it will tie with our method, and that is a real possibility. The experiment that settles it is red-cell volume and the other traits flagged as hotspot-driven. If per-block selection does not beat per-trait selection there, the core claim fails and we will say so.
Why not just rely on DendroPRS's sparse per-SNP offsets? This is the most serious threat, because sparse offsets can already make DARC ancestry-specific. Our argument is that per-SNP offsets cannot tell a real effect difference from a tagging difference when LD differs across ancestries. We test this directly. We simulate shared causal effects ($\rho = 1$) on real multi-ancestry LD and count the spurious offsets. We also simulate truly divergent blocks and measure detection accuracy (AUROC). If DendroPRS shows no excess spurious divergence, our advantage reduces to scalability alone.
Why not just use standard single-population LD blocks? Blocks defined in Europeans can join variants that are independent in Africans, or split variants that are correlated in Africans. The resulting blocks then fail to be separable in the target population. BALD's gains came from correct block structure within one cohort (Performance of a blockwise approach…). The ablation above measures how much is lost by defining blocks in the wrong population.
Why not just reweight minority samples? Reweighting trades bias for variance and collapses effective sample size (Optimal strategies…). It also applies one weight genome-wide. We include reweighted LASSO among the per-trait candidates, so the comparison is built in.
Proposed Experiments
Data. We use UK Biobank individual-level genotypes with genetically inferred African, Admixed American, Central/South Asian, East Asian, Middle Eastern and European groups. The traits are the 15 from Optimal strategies…, plus BMI to match DendroPRS. Geuvadis expression in European and African groups provides a cis-scale test, where DendroPRS reported gains for 67 genes. The 1000 Genomes superpopulations provide external LD for robustness checks.
Metrics. The main metric is prediction correlation or $R^2$ in the test fold of each target ancestry, with bootstrap intervals. Secondary metrics are wall-clock time and memory, compared with genome-wide DendroPRS and LASSO, and the share of LD mass discarded at cuts as $\tau$ varies.
Primary check. Per-block selection on merged-graph blocks should beat the oracle per-trait choice on red-cell volume and red-cell distribution width in African-ancestry test individuals, while matching it on height.
Mechanism check. Under shared effects with divergent LD, our method should flag far fewer divergent blocks than DendroPRS flags divergent SNPs. Under planted divergence, it should recover the planted blocks.
Ablation check. Merged-graph blocks should beat European-only and window blocks most clearly in African-ancestry targets, where LD differs most from Europeans.
Failure check. If refining large components discards substantial LD mass, the gains should shrink. Plotting accuracy against $\tau$ tells us whether the graph decomposition is helping or only fragmenting the genome.
References
- Optimal strategies for learning multi-ancestry polygenic scores vary across traits
- Multi-ancestry polygenic risk scores using phylogenetic regularization
- Performance of a blockwise approach in variable selection using linkage disequilibrium information
- Learning Large-Scale MTP2 Gaussian Graphical Models via Bridge-Block Decomposition
- Decomposition pipeline for large-scale portfolio optimization with applications to near-term quantum computing
- BC tree-based spectral sampling for big complex network visualization
- Mediation Analysis with Mendelian Randomization and Efficient Multiple GWAS Integration