When the Tag Moves: Pangenome-Anchored Instruments for Cross-Ancestry Mendelian Randomization

Mendelian randomization estimates whether one trait causes another. It uses genetic variants as natural experiments: if a variant raises cholesterol, and cholesterol causes heart disease, the variant should also raise heart disease risk, by exactly the proportion that the cholesterol pathway implies. The variants used are almost never the causal DNA changes themselves. They are markers that happen to be inherited together with the causal change (they are in "linkage disequilibrium," or LD), and they were chosen in European cohorts. Many causal changes are also not single-letter substitutions. They are larger structural variants, such as insertions, deletions, or whole stretches of sequence missing from the standard human reference genome, and short-read pipelines built on that single reference largely cannot see them. We propose to use the new human pangenome, a collection of fully assembled genomes from diverse ancestries, to measure how strongly each European marker tracks its structural variant in each ancestry. We then correct or replace the markers accordingly.

The central observation is simple and, as far as we can find, untested. When the exposure effects of a marker are estimated in Europeans and the outcome effects in another ancestry, the causal estimate is scaled by the ratio of that marker's LD with the causal variant in the two populations. When the causal variant is structural, only a pangenome can supply that ratio. The same mechanism predicts that some instruments flagged as "pleiotropic" in non-European analyses are not biologically pleiotropic at all. They are markers whose relationship to an unseen structural allele shifts across ancestries. The work is valuable because it turns an abstract complaint ("linear references miss diversity") into a checkable correction for a specific, widely used causal-inference procedure.

Core Claim

Suppose marker SNP $s$ sits near a causal structural variant $v$, and $v$ changes exposure $X$ by $\gamma_X$ and outcome $Y$ by $\theta\gamma_X$, where $\theta$ is the causal effect we want. Assume a single causal variant and standardized genotypes. The marginal GWAS effect of $s$ in ancestry $a$ is then approximately $b_{X,a} = \rho_a \gamma_X$, where $\rho_a$ is the genotype correlation between $s$ and $v$ in ancestry $a$. The two-sample Wald ratio that mixes a European exposure GWAS with an African outcome GWAS is

$$ \frac{b_{Y,\mathrm{AFR}}}{b_{X,\mathrm{EUR}}} = \frac{\rho_{\mathrm{AFR}}}{\rho_{\mathrm{EUR}}}\,\theta . $$

Two consequences follow. First, if both GWAS come from the same ancestry, $\rho$ cancels. LD decay then only weakens the instrument; it does not bias the ratio. The pangenome therefore matters most in the common and practically important "transported" design, where the target population lacks a large exposure GWAS. It also matters at loci with more than one causal allele, where the mixture weights change with LD. We state this up front because it limits where our method can help. Second, across the markers in an MR analysis, the per-marker ratios will scatter in proportion to $\rho_{\mathrm{AFR}}/\rho_{\mathrm{EUR}}$. Outlier-removal procedures that assume a flat genome will read this scatter as pleiotropy. These include HEIDI-outlier from Causal associations between risk factors and common diseases inferred from GWAS summary data, which filters pleiotropic instruments using linear LD matrices, and MR-PRESSO.

Our claim is that per-ancestry correlations between SNPs and structural variants, computed from pangenome-genotyped reference samples, predict both the cross-ancestry drift in marker effect sizes and a measurable share of flagged "pleiotropic" outliers. Anchoring instruments to the structural allele should therefore reduce cross-ancestry heterogeneity without inflating false positives.

Where This Sits

Three papers motivate the gap by showing that the relevant alleles exist and are functional. De novo assembly of 64 haplotype-resolved human genomes of diverse ancestry and integrated analysis of structural variation assembled 64 haplotypes from diverse 1000 Genomes individuals and found 107,590 structural variants, 68% of them inaccessible to short reads. It then used PanGenie, a tool that genotypes known pangenome alleles from ordinary short-read data, to genotype them across all 3,202 1000 Genomes samples. It reported 1,525 structural-variant expression QTLs (variants that change gene expression) but stopped before any GWAS or causal analysis. Human pangenome analysis of sequences missing from the reference genome reveals their widespread evolutionary, phenotypic, and functional roles cataloged 45,284 non-reference sequences, 35.6% of them population-specific, and 565 expression QTLs on them. It found 154 of these sequences in strong LD with 258 GWAS catalog SNPs. It computed that LD but never asked how it varies by ancestry, or what that does to an instrument. The Human Pangenome Project: a global resource to map genomic diversity supplies the assemblies and warns that graph coordinates do not map cleanly onto clinical linear coordinates. That warning is why we work through LD projections onto linear SNPs instead of demanding graph-native GWAS.

On the causal side, Benchmarking Mendelian Randomization methods for causal inference using genome-wide association study summary statistics found false-positive inflation under population stratification in 11 of 15 methods, with MR-APSS the most robust. Robust Mendelian randomization in the presence of residual population stratification, batch effects and horizontal pleiotropy showed that tiny residual confounding produces false positives at biobank scale. Both constrain our design: any correction we add must be scored on false-positive control, not only on concordance.

Multiple instance fine-mapping: predicting causal regulatory variants with a deep sequence model is the closest cross-ancestry method. It learns causal regulatory SNVs from 512-bp linear windows and improves African-ancestry risk prediction by 17%. It cannot score alleles absent from the reference, and it does not perform instrumental-variable estimation. We differ from it in both respects.

Research Program

The first step is a go/no-go catalog. For every structural variant genotyped by PanGenie in the 3,202 1000 Genomes samples, and for the GraphNRS sequences, we compute $\rho_a$ to every nearby SNP separately in the African, European, East Asian, South Asian and admixed American super-populations. For each variant we record the best-tagging SNP in each ancestry. The quantity of interest is the fraction of structural expression QTLs, and of structural variants colocalizing with GWAS loci, that are well tagged in Europeans ($\rho_{\mathrm{EUR}}^2$ high) but poorly tagged by the same SNP, or by any SNP, in other ancestries. If this fraction is negligible, the rest of the program is not worth running, and we will report that as a finding.

The second step builds the estimator. For each instrument locus we define a structural-anchored effect. The exposure and outcome SNP effects are projected onto the structural allele using the ancestry-specific $\rho_a$ vectors from the reference panel, so that the Wald ratio is computed on the anchor rather than the tag. Loci where $\rho$ in the target ancestry is near zero are down-weighted rather than divided through, because a tiny $\rho$ also means a weak instrument. The anchored instruments then feed into IVW and into MR-APSS unchanged, so the contribution is a better instrument set, not a new estimator. For admixed cohorts we compute $\rho$ conditional on local ancestry.

The third step is diagnostic. For every instrument that HEIDI or MR-PRESSO flags in a non-European analysis, we ask whether the flag disappears after anchoring. We also ask whether the variant's deviation is predicted by $\rho_{\mathrm{AFR}}/\rho_{\mathrm{EUR}}$ better than by chance loci matched on allele frequency and LD score.

Baselines

The naive reference point is the standard practice in Mendelian randomization study shows no causal relationship between psychiatric disorders and glaucoma in European and East Asian populations: clump European genome-wide significant SNPs, then run inverse-variance-weighted (IVW) MR against a target-ancestry outcome GWAS. It is what most applied papers do, so any gain has to be measured against it.

The strongest existing method is MR-APSS, which the 15-method benchmark in Benchmarking Mendelian Randomization methods for causal inference using genome-wide association study summary statistics identified as the most robust under stratification. The file reports this ranking qualitatively; we found no reported numeric Type I error rate for MR-APSS on Lacuna. CAUSE is a second strong competitor, described there as reliable but conservative. Both will be run on identical summary statistics and identical clumping windows, with their default tuning, and our anchored instruments will be passed into MR-APSS itself. Beating MR-APSS therefore means that MR-APSS with anchored instruments matches or beats MR-APSS with tag instruments on three things: cross-ancestry concordance of the estimate, false-positive rate on negative-control trait pairs no higher than the tag-SNP version, and lower heterogeneity Q.

The third baseline is the cheap linear fix: re-select instruments using the target ancestry's own lead SNP and 1000 Genomes SNP-only LD. The fourth is multi-ancestry SNP fine-mapping, using SuSiE, which the MIFM paper benchmarks against, to pick a single proxy SNP per locus.

The key ablation runs our full projection pipeline with the structural variants deleted from the reference panel, so that anchoring falls back to the best-correlated SNP. Any improvement that survives this ablation is attributable to LD modeling. Only the difference between the full and ablated runs is attributable to the pangenome.

For grounding effect sizes, the published anchors we will reproduce are GSMR's 3.29-fold increase in type 2 diabetes per standard deviation of BMI in Causal associations between risk factors and common diseases inferred from GWAS summary data. We also reproduce the FGF23 effect on aortic stenosis (discovery OR 3.081, replication OR 2.280) from Genetic association of serum calcium, phosphate, vitamin D, parathyroid hormone, and FGF23 with the risk of aortic stenosis. That study was European-only, so we transport it to East Asian outcomes where available.

Why not just…?

Why not just use MR-APSS or CAUSE on the existing SNP instruments? These methods model stratification and correlated pleiotropy, but the bias in our transport equation is not confounding. It is a per-locus multiplicative factor determined by an allele that is missing from their input. If a structural variant drives the locus, no adjustment computed from SNP summary statistics alone can know $\rho_{\mathrm{AFR}}$. We do not argue this abstractly. Our experiments feed anchored instruments into MR-APSS, so if MR-APSS already absorbs the problem, the anchored and tag versions will agree and we will say so.

Why not just pick ancestry-matched lead SNPs or run multi-ancestry SNP fine-mapping? This is the most serious threat. If a common SNP in the target ancestry tags the structural variant well, re-selection recovers most of the benefit with no pangenome. The SNP-only ablation and the fine-mapping baseline exist to measure exactly this. Our prediction is that the gap appears only in the subset of loci the catalog marks as untaggable by SNPs in the target ancestry. If the gap is zero there too, the pangenome adds nothing for MR, and we will report that as a finding.

Why not just run the analysis in target-ancestry exposure GWAS, or in sibling cohorts? When a same-ancestry exposure GWAS exists, the $\rho$ ratio cancels. We will use BioBank Japan and China Kadoorie exposure data to confirm that the anchored and tag estimates converge in that setting, as an internal consistency check. The benchmark shows sibling GWAS removes confounding, but the problem we target is not confounding. For most African-ancestry exposures, neither large same-ancestry GWAS nor sibling cohorts exist, and that is the setting we target.

Why not just use a sequence model like MIFM? MIFM scores SNVs in reference sequence. It cannot score a regulatory element that sits inside an insertion absent from GRCh38. Where a locus's anchor is an SNV, MIFM-prioritized SNPs are a reasonable alternative instrument set, and we include them as an extra comparison on SNV-anchored loci.

Proposed Experiments

We begin with simulation on real haplotypes. Using the PanGenie-genotyped 1000 Genomes haplotypes, we plant causal structural variants at known expression-QTL positions. We simulate exposure and outcome with a known $\theta$ and generate European and African GWAS summary statistics. This lets us verify the transport equation, measure bias of the naive, MR-APSS, re-selection, and anchored estimators, and estimate false-positive rates under $\theta = 0$, both with and without planted stratification of the kind studied in the MR-SENSEMAKR paper.

The first empirical test checks the prediction without any MR. At GWAS loci that colocalize with structural expression QTLs, we regress the ratio of East Asian (BBJ/CKB) to European (UK Biobank/FinnGen) marker effects on the pangenome-predicted $\rho_{\mathrm{EAS}}/\rho_{\mathrm{EUR}}$. We compare this against the same regression using SNP-only LD predictions. A slope near one that is clearly better than the SNP-only prediction is the decisive signature that structural alleles drive effect drift.

The second test re-runs transported MR analyses: BMI on type 2 diabetes, FGF23 on aortic stenosis, and the depression and insomnia on glaucoma pairs where European and East Asian results are both available. For these pairs we test whether anchoring reduces heterogeneity and removes HEIDI/MR-PRESSO flags. We also run the negative-control pairs from the MR benchmark to confirm false-positive rates do not rise.

The third test uses UCLA ATLAS, where the tobacco-use polygenic score from Polygenic scores for tobacco use provide insights into systemic health risks in a diverse EHR-linked biobank in Los Angeles failed in African Americans (OR 1.04, non-significant). There we impute structural alleles into the array data from the pangenome haplotype panel. We then test directly whether anchored instruments for smoking exposure regain strength (first-stage F) in the African American group, relative to the tag-SNP instruments.

The project succeeds if three things hold: the catalog shows a non-trivial set of SNP-untaggable structural loci, the effect-drift regression favors pangenome LD over SNP LD, and anchored MR lowers cross-ancestry heterogeneity at no cost in false positives. It fails cleanly if the SNP-only ablation matches the full method. That outcome would itself tell the field that, for common-variant MR, the pangenome's value lies elsewhere.