Ancestry-Aware Clumping: Choosing Mendelian Randomization Instruments That Survive the Trip Across Populations

Object detectors produce many overlapping candidate boxes. They prune these with a greedy rule called non-maximum suppression (NMS): keep the highest-scoring box, then delete every neighbor that overlaps it too much. Genetic epidemiology uses the same rule to pick instruments for Mendelian randomization (MR), a method that uses inherited DNA variants as natural experiments to estimate whether an exposure such as BMI causes a disease. The pruning step is called LD clumping. Researchers keep the variant with the smallest p-value, then delete every nearby variant whose correlation with it (its "linkage disequilibrium", or LD) exceeds a threshold. The computer-vision literature has spent years documenting why this rule fails. The score that ranks candidates and the overlap that triggers deletion are only proxies for "this is the same object," and the proxies break in crowded or occluded scenes.

This note argues that the same two proxies break when European-derived instruments are carried to East Asian or African cohorts. The ranking score is a European p-value. The overlap is a European LD correlation. Both describe the discovery population, not the population where the causal question is being asked. Because African genomes have shorter LD blocks, the variant that wins clumping in Europe is often a tag that no longer tracks the causal variant in the target population. The result is weak instruments and attenuated, noisy causal estimates. We propose to rebuild clumping so that ranking and suppression are both computed for the target population, and to test whether this alone recovers instrument strength that is currently lost.

Core Claim

Clumping makes two separate decisions at each genomic locus, and current practice gets both wrong across ancestries. The first is which variant survives. It is chosen by European significance, when it should be chosen by how strongly the variant is expected to associate with the exposure in the target population. The second is which neighbors are deleted. They are judged redundant by European LD, when they should be deleted only if they are redundant in both populations.

The first fix is the more important one. To estimate expected target-ancestry signal, we combine European fine-mapping with target-ancestry LD. Fine-mapping is a statistical step that assigns each variant a probability of being the true causal one. For a candidate variant $k$ at a locus, we compute

$$s_k = \sum_j \pi_j \, R^{T}_{kj} \, b_j ,$$

where $\pi_j$ is the European fine-mapping probability that variant $j$ is causal, $b_j$ is its standardized European effect on the exposure, and $R^{T}_{kj}$ is the LD correlation between $k$ and $j$ in a target-ancestry reference panel. We keep the variant with the largest $|s_k|$, not the smallest European p-value.

The second fix changes the deletion rule. A neighbor is deleted only if its correlation with the kept variant exceeds the threshold in both the European and the target panel. Neighbors that pass this test are retained as separate instruments, and the remaining correlation among them is handled with target-ancestry LD in the estimator.

Two pieces of detection work supply the intuition. The first is Relational Learning for Joint Head and Human Detection. It recovers occluded pedestrians by anchoring body proposals on heads, which stay visible when bodies do not, instead of trusting box overlap. In our setting, the fine-mapped causal variant plays the role of the head: it is the anchor that remains stable across populations, while tag SNPs do not. The second is Falcon Perception-HD. It removes NMS altogether by training an autoregressive model with a set-level Hungarian matching reward. We do not import its reinforcement learning. We do import its evaluation idea: score the selected set of instruments against the set of true signals with one-to-one matching, so that lost signals are counted the way a detector's misses are counted.

Research Program

The work has three parts. The first is diagnostic. We quantify how much signal clumping over-suppresses or misplaces when European instruments are applied in the target population. For each locus, we record whether the European lead SNP differs from the variant with the highest fine-mapping probability, and how strongly the two are correlated in the target panel. This is the genetics analogue of the occlusion analyses in detection. It tells us in advance where the method can help.

The second part is the selection rule above, built from components that already exist. European fine-mapping and effect estimates come from UK Biobank exposure GWAS (genome-wide association studies). Target LD comes from the East Asian and African panels of 1000 Genomes. We also handle allele-frequency differences, which already enter $R^{T}$ when variants are standardized within the target panel. Selection never uses the target outcome GWAS. This avoids the winner's-curse problem, in which variants selected on the same data that is later analyzed look stronger than they are.

The third part is estimation. Retained correlated instruments enter a correlated-instrument inverse-variance-weighted estimator, or MR-LDP, and in both cases the LD matrix comes from the target ancestry.

Three prior lines of work bear on this design, each in a different way.

MR-LDP motivates and constrains the proposal. It replaces pruning with a likelihood over the full LD matrix and treats horizontal pleiotropy (variants affecting the outcome through other pathways) as a random effect. It reports stable error control where pruning-based IVW and MR-Egger inflate false positives. However, it uses a single reference panel matched to the discovery cohort and never tests ancestry mismatch, which is exactly our setting. It also still requires a manual p-value threshold for selection, which our ranking replaces.

PRISM supplies evidence that prioritizing likely-causal variants over tags improves transfer. It combines fine-mapping with tissue-specific annotations and reports a 13.10% average gain in African-ancestry polygenic score prediction. It works on prediction, not causal inference, and it relies on curated annotations. Its finding that "kitchen sink" annotation models degrade performance is a warning that we take seriously. We therefore treat annotations as an optional add-on rather than the core.

LOG-TRAM shows that cross-ancestry similarity varies along the genome. It borrows European power within 1-Mb windows and raises Biobank Japan effective sample size by 49%. It works at window resolution and does not choose instruments within a locus. That is the gap our rule addresses, and LOG-TRAM summary statistics can be fed into it.

Baselines

The naive reference point is what most published cross-ancestry MR does now: standard European clumping (r² < 0.001 within a 250 kb window, genome-wide significance) followed by IVW in the target outcome GWAS. Nothing is changed except the outcome data.

The strongest simple competitor is clumping with the target-ancestry LD panel while still ranking by European p-value. It is the obvious fix any reviewer would try first, and it already corrects the deletion decision.

The strongest existing method is MR-LDP run with a target-ancestry LD matrix. On Lacuna, its only reported figure is "a stable 5% error rate" in single-ancestry simulations. No reported cross-ancestry result was found on Lacuna for MR-LDP or for clumping with IVW.

All methods receive identical inputs: the same European exposure summary statistics, the same 1000 Genomes panels, the same target outcome GWAS, and the same thresholds tuned on the same held-out simulation grid.

Two ablations isolate the key ingredient. One keeps the European-p-value survivor but applies dual-panel deletion. The other uses $s_k$ ranking with ordinary single-panel deletion. If gains come from the ranking ablation, the anchor idea is doing the work. If they come only from dual-panel deletion, the proposal reduces to a better threshold.

Beating the baselines means two things. In target-ancestry data held out from selection, the instruments must have a higher F-statistic and explain more exposure variance. In simulations, the causal estimates must be equally well calibrated and have lower root-mean-square error.

Why not just…?

Why not just clump with a target-ancestry LD panel? This is the real threat. It fixes deletion but keeps a survivor chosen by European significance, so it cannot recover a locus whose European lead SNP is a tag that does not carry over. The first ablation measures exactly this. If target-LD clumping matches our method across loci, including loci where the lead SNP and the top fine-mapped variant are weakly correlated in the target panel, we will report that the simple fix suffices and that ranking adds nothing.

Why not just run MR-LDP with the target LD matrix? MR-LDP combines European effect sizes with target correlations inside one likelihood. That model is misspecified when the tagging relationship itself has changed. Pleiotropy modeling does not repair a lost instrument. We run it as a baseline, and we also run it on top of our selection, so complementarity would show up as a gain rather than being argued.

Why not just do MR entirely within the target-ancestry GWAS? That is the cleanest design when it is powered. Zhang et al.'s psychiatric–glaucoma MR does this in Biobank Japan and China Kadoorie Biobank. Its East Asian estimate (depression OR 1.38, p = 0.30) carries wide uncertainty, which is typical of target-only analyses. LOG-TRAM's 49% effective-sample-size gain shows how much power borrowing adds. We include target-only MR as a reference and report the sample size at which it overtakes borrowed instruments.

Why not just add functional annotations as PRISM does? Annotations require per-trait tissue curation, and PRISM itself reports that indiscriminate use hurts. We test annotation-informed fine-mapping probabilities as an optional input to $\pi_j$. The core claim is evaluated without them.

Proposed Experiments

The first experiment is a simulation built on 1000 Genomes European, East Asian and African haplotypes. We plant one to three causal exposure variants per locus, vary pleiotropy and the share of causal effect that is shared across ancestries, and compute European and target summary statistics. Because the true signals are known, we can score each method's instrument set against them with Hungarian matching. This gives an "instrument recall" analogous to a detector's miss rate, alongside bias, coverage and root-mean-square error of the causal estimate.

The real-data experiments use exposures with UK Biobank instruments that GSMR studied: BMI and LDL cholesterol, with type-2 diabetes as the outcome. GSMR's European estimate is a 3.29-fold T2D risk per SD of BMI. Instruments are selected using only European exposure data plus target LD. Their strength is then measured in held-out Biobank Japan exposure GWAS. MR is run against Biobank Japan outcomes, and the out-of-sample instrument variance is checked against the independent WeGene cohort used by LOG-TRAM. African-ancestry transfer is evaluated in the UK Biobank African-ancestry subset used by PRISM.

The decisive check is where any gain appears. Gains should concentrate at loci where the European lead SNP and the top fine-mapped variant differ and are weakly correlated in the target panel, and should be close to zero elsewhere. If gains are uniform across loci, or appear only in the dual-deletion ablation, the anchor claim is wrong.