Most genetic studies of disease causes rely on European volunteers. The genetic markers they report are usually stand-ins: each one is inherited together with the variant that actually does something, without being that variant. Population geneticists call this co-inheritance linkage disequilibrium (LD), meaning the correlation between nearby variants across a population. LD patterns differ between ancestral populations, so a marker that tracks the true variant closely in Europeans may barely track it in African or East Asian samples. This is a main reason causal estimates built on European markers fail to transfer to other ancestries. The same difference in LD can be turned to advantage. Two candidate variants that move together in Europeans may come apart in another population, and data from that population can then tell which one matters. Recruiting and genotyping non-European participants is slow and expensive, so the practical question is an expensive black-box optimization problem: given a fixed budget, which genomic region should be studied next, and in which ancestral population, to identify the true causal variants as quickly as possible?
We propose an active design policy for this choice. At each round it picks a (region, ancestry, batch size) action that maximizes the expected information gained per unit cost about which variant is causal. The policy predicts that gain before any new data are collected, using publicly available ancestry-specific LD reference panels. The key ingredient is that the policy's statistical model does not assume the causal effect is identical across ancestries. It carries a per-region estimate of how much effects are shared, so it can recognize regions where pooling another population's data would mislead rather than help. The variants it resolves are meant to serve as instruments in Mendelian randomization (MR), a method that uses genetic variants as natural experiments to estimate causal effects. Causal variants should keep their strength across ancestries in a way that stand-in markers do not.
Core Claim
Under a fixed cost budget, an allocation policy that combines (i) ancestry-specific LD, (ii) per-region uncertainty about whether effects are shared across ancestries, and (iii) unequal per-sample costs should reach correct, small credible sets at substantially lower cost than three alternatives: proportional or uniform allocation, static fine-mapping of whatever data happen to exist, and the obvious rule "spend everything on the population with the least LD." We expect the advantage to be largest in regions where cross-ancestry effects diverge. In those regions, a policy that assumes shared effects will confidently pick the wrong variant. Lehman et al. show that this negative-transfer regime is real and depends on the trait. In Optimal strategies for learning multi-ancestry polygenic scores vary across traits, roughly 5,000 African-ancestry UK Biobank participants beat a European model trained on about 300,000 people for mean corpuscular volume, a red-cell trait driven by the DARC locus. For standing height, the larger European sample won. In some settings, adding European data degraded African-ancestry accuracy when the genetic correlation was low. That study evaluates fixed, passive training sets. Its finding motivates our work, and we turn it into a sequential decision.
Positioning Against Prior Work
The obvious idea at this intersection is to condition causal Bayesian optimization on ancestry. Causal Bayesian Optimization builds a Gaussian-process surrogate whose prior comes from do-calculus on a known causal graph. Multi-Objective Causal Bayesian Optimization extends this to Pareto trade-offs on the same healthcare graph. Adversarial Causal Bayesian Optimization handles an observed external agent that perturbs a known graph. All three choose interventions on variables in a known graph. Our actions are different: they decide what data to collect, and in which population, in order to learn the structure, specifically which variant is causal. That makes this an experimental design problem about identification, not about optimizing an outcome.
Invariant Ancestry Search uses invariance across environments to recover causal ancestors offline from yeast knockouts. It shares our view that heterogeneous environments are informative, but it never chooses which environment to sample next.
Three papers supply methods. A Lasso-based Sparse Knowledge Gradient Policy for Sequential Optimal Learning shows that sequential sparse learning can isolate the few relevant features among hundreds; it recovered an RNA accessibility profile across 414 sites within 50 measurements. Our problem has the same sparse structure, with a single-causal-variant prior per region. Its actions are individual measurements, whereas ours are cohorts with population-specific correlation among features. Bayesian Experimental Design Via Contrastive Diffusions and its code (jcopo/ContrastiveDiffusions) provide estimators of expected information gain, which we would use if our closed-form approximations prove too loose. Finemap-MiXeR is a variational fine-mapper that works from summary statistics and avoids inverting the LD matrix. It serves as our posterior engine and our static baseline. Its authors note it could extend to multiple ancestries but apply it post hoc.
On the MR side, GSMR explicitly limits its conclusions to European summary statistics. The three-sample design of Zhao et al. shows that instrument selection is itself a design choice. Both motivate treating instrument quality across ancestries as the target that data collection should improve.
Research Program
For region $l$ with $J$ candidate variants, let $\gamma_l \in \{1,\dots,J\}$ index the causal variant. Given $n_a$ samples from ancestry $a$, the association z-scores are approximately $z_{l,a} \sim \mathcal{N}(\sqrt{n_a}\, R_{l,a} e_{\gamma_l} b_{l,a},\; R_{l,a})$. Here $R_{l,a}$ is the region's LD matrix in ancestry $a$, $e_{\gamma_l}$ is the indicator vector for the causal variant, and $b_{l,a}$ is that variant's effect size in ancestry $a$. Heterogeneity enters in two places. First, the causal variant is shared across ancestries with prior probability $\pi$; otherwise each ancestry has its own causal variant. Second, effect sizes are correlated across ancestries with a region-specific correlation $\rho_l$, which is learned hierarchically across regions. The policy chooses the action that maximizes
$$\alpha(l,a,m) = \frac{\mathbb{E}_{z \sim p(z_{l,a}^{(m)} \mid \mathcal{D})}\big[H(\gamma_l \mid \mathcal{D}) - H(\gamma_l \mid \mathcal{D}, z)\big]}{c_a\, m},$$
where $\mathcal{D}$ is the data collected so far, $H$ is posterior entropy over the causal variant, $m$ is the batch size, and $c_a$ is the per-sample cost for ancestry $a$. A cheap posterior update is possible because the update only needs reference LD and the predictive distribution of summary statistics. That is what lets us score every (region, ancestry, batch size) action before spending any money.
We would build the policy in three stages. The first is an exact single-effect posterior with Monte Carlo estimates of the information gain. The second adds hierarchical learning of $\rho_l$ and $\pi$, so the policy can detect divergent regions from early batches. The third is a batch version that selects several actions per round with a greedy diversity correction. We would also derive a simple diagnostic that predicts when the policy should beat "always sample the population with the least LD": regions where the posterior over $\rho_l$ is wide, or where the least-LD population's $R_{l,a}$ does not separate the current top candidates.
Baselines
The naive reference point allocates the budget in proportion to existing cohort sizes, or uniformly across ancestries and regions. The analysis then runs Finemap-MiXeR on the pooled data. This is what a consortium does by default. For the strongest existing static method, Finemap-MiXeR reports "substantially higher R² in 29% of loci and superior or comparable results in 51% of loci" against FINEMAP and SuSiE RSS on UK Biobank height (Finemap-MiXeR). That result is single-ancestry and non-adaptive. We found no reported result on Lacuna for adaptive, cost-aware allocation of cross-ancestry samples for fine-mapping, so this baseline must be run by us.
The strongest competitor is a heuristic, and we expect it to be a real threat. It spends the whole budget, region by region, on the ancestry with the smallest LD between the current top two candidates (the LD-contrast rule). A cruder version always samples the African-ancestry pool. A second competitor applies the Sparse Knowledge Gradient policy of Li et al. with ancestry as an extra action dimension and no heterogeneity model.
The key ablation is our own policy with $\rho_l$ fixed at 1 and $\pi$ fixed at 1, so that effects are assumed shared. This isolates whether modeling heterogeneity matters.
All methods receive the same candidate pool, LD panels, cost vector, total budget, posterior engine, and an equal hyperparameter-tuning budget on held-out simulated regions. Beating a baseline means a statistically significant reduction in the cost required to reach a 95% credible set of five or fewer variants that still contains the true causal variant at its nominal rate, averaged over regions and replicate simulations.
Why not just…?
Why not just always sample the least-LD population? For regions with a shared causal variant and shared effects, this heuristic may be nearly optimal. We would treat a tie there as an informative outcome, not a failure. It fails in two ways. Where the causal variant or its effect differs across ancestries, which the DARC example in Lehman et al. shows can happen, it pools contradictory signals. It also ignores cost, so it can waste money on regions already resolved. The stratified analysis by true $\rho_l$ settles the question. If the heuristic matches our policy across all strata, the paper's conclusion becomes "the heuristic suffices," and our diagnostic tells practitioners when that holds.
Why not just run causal Bayesian optimization conditioned on ancestry? Causal Bayesian Optimization and its multi-objective and adversarial extensions optimize an outcome over interventions on a graph assumed known. Here the graph, meaning which variant is causal, is the unknown. No intervention on genotype exists, and the decision is what data to collect. Recasting the problem for CBO would require assuming exactly what we want to learn.
Why not just reweight or fine-map each ancestry separately? Separate modeling discards shared signal. Lehman et al. show that it loses for highly polygenic traits like height. They also report that importance reweighting inflates variance by shrinking the effective sample size of small cohorts. Both approaches are passive: they say nothing about where the next dollar should go. Separate per-ancestry fine-mapping is included as an analysis variant under every allocation, so its effect is measured apart from the allocation effect.
Proposed Experiments
The first testbed is Hapgen2 synthetic genotypes, as used by Finemap-MiXeR, generated from ancestry-specific reference haplotypes. Regions are planted with known causal variants, and we vary heritability, $\pi$, and $\rho_l$. This gives exact ground truth and unlimited candidate pools.
The second testbed is semi-synthetic UK Biobank data with the ancestry groupings used by Lehman et al. Each ancestry's individuals form a pool, and "querying" reveals a batch of real genotypes. Phenotypes are simulated from planted causal variants, using the real LD structure. Height-like architectures (shared effects, polygenic) and MCV-like architectures (differentiated loci) are contrasted, mirroring Lehman et al.'s split. The small non-European pools in UK Biobank cap the budget range on this testbed, and we would report that limit openly.
Metrics are cost to resolution, calibration of credible-set coverage, and recall of the causal variant at fixed budgets. A downstream MR metric is also included. We take the variants selected under each policy as instruments and measure their exposure variance explained in held-out individuals of each ancestry. We then measure the bias of the cross-ancestry MR estimate against the simulated true causal effect, comparing these instruments with European lead SNPs selected at genome-wide significance.
The decisive check has three parts. First, the full policy must beat the LD-contrast heuristic in low-$\rho_l$ strata while not losing in high-$\rho_l$ strata. Second, the shared-effect ablation must show measurable miscalibration, meaning credible sets that confidently exclude the true variant, in divergent regions. Third, the resolved variants must transfer as MR instruments with less loss of strength in non-European held-out data than European stand-in markers. If the first part fails while the second holds, the contribution narrows to the heterogeneity-aware posterior rather than the allocation policy, and we would say so.
References
- Optimal strategies for learning multi-ancestry polygenic scores vary across traits
- Causal Bayesian Optimization
- Multi-Objective Causal Bayesian Optimization
- Adversarial Causal Bayesian Optimization
- Invariant Ancestry Search
- A Lasso-based Sparse Knowledge Gradient Policy for Sequential Optimal Learning
- Bayesian Experimental Design Via Contrastive Diffusions
- Finemap-MiXeR: A variational Bayesian approach for genetic finemapping
- Causal associations between risk factors and common diseases inferred from GWAS summary data
- Powerful three-sample genome-wide design and robust statistical inference in summary-data Mendelian randomization