Large genetic studies pool results from many cohorts to find DNA variants linked to a trait. Each cohort usually corrects for ancestry on its own, and small or regional cohorts often do this badly. The failure is specific: a variant can look associated with a trait only because its frequency tracks ancestry, and ancestry in turn tracks environment or social position. That confounding is called population stratification. A small cohort's own ancestry axes (principal components, the main directions of genetic variation) are too noisy to remove it. The cohorts that correct worst tend to be the ones with the least bioinformatics capacity to redo the analysis properly. We propose to shift that work from the weak site to a well-resourced "counterpart." A large reference biobank publishes two things. The first is per-variant loadings on its high-resolution ancestry axes, meaning how strongly each variant's frequency follows each axis. The second is a carefully corrected reference association scan. The consortium's central analysts then use only the summary statistics each cohort already submits to detect and remove leftover ancestry bias. No site reruns anything, and no genotypes cross a border.
The idea borrows its structure from counterpart assistance, duikou zhiyuan. In that Chinese health policy, high-capacity hospitals pair with low-capacity ones and carry the specialist load for them. The Yunnan poverty-alleviation program built local capacity this way through two-year specialist rotations, as evaluated in Tracking progress towards equitable maternal and child health in Yunnan. Its known weakness is that the help is transient. Here the counterpart's contribution is a durable, reusable artifact rather than a person. The four input directions play different roles. Population stratification is the problem. Study heterogeneity is the diagnostic lens. The bioinformatics accessibility gap is a hard constraint: zero added burden at the sites. Counterpart assistance is the division of labor.
Core Claim
Leftover stratification in a cohort's results has a recognizable shape. If a cohort omits an ancestry axis that also shifts the trait, each variant's estimated effect is biased roughly in proportion to how strongly that variant loads on the axis. This follows from omitted-variable bias: the variant's correlation with the omitted axis, multiplied by the axis's effect on the trait. Two other sources of disagreement between cohorts leave different shapes, and a model can separate all three:
- A cohort that defines the phenotype differently (different thresholds, units, or scales) rescales all its effects roughly uniformly.
- Genuine biological differences are concentrated in particular genomic regions.
For cohort $k$ and variant $j$ we fit
$$\hat\beta_{kj} - \beta^{\text{ref}}_j = (a_k - 1)\,\beta^{\text{ref}}_j + \sum_{c=1}^{C} \gamma_{kc} L_{jc} + u_{kj} + \varepsilon_{kj}.$$
The terms are:
- $\hat\beta_{kj}$ is the cohort's submitted effect estimate.
- $\beta^{\text{ref}}_j$ is the counterpart's well-corrected estimate.
- $a_k$ is a phenotype-scale factor that captures definition drift.
- $L_{jc}$ is variant $j$'s loading on the counterpart's ancestry axis $c$.
- $\gamma_{kc}$ measures leftover stratification along that axis.
- $u_{kj}$ is region-structured true heterogeneity.
- $\varepsilon_{kj}$ is sampling noise with known variance.
The model is fit by inverse-variance-weighted regression over approximately independent variants. A test of $\gamma_k = 0$ serves as a stratification audit. The corrected estimate $(\hat\beta_{kj} - \sum_c \hat\gamma_{kc} L_{jc})/\hat a_k$ then goes into the meta-analysis.
The new ingredient is the reference-anchored, axis-directed decomposition of between-cohort heterogeneity. Existing tools either require every site to re-project its genotypes or apply a single uniform correction per cohort.
Relation to Prior Work
Three papers motivate this work by documenting the gap. Population stratification in GWAS meta-analysis should be standardized to the best available reference datasets shows that local ancestry correction in small cohorts both overfits noise and misses real structure. In ALSPAC, local components explained 1% of maternal educational-attainment variance while 200 UK Biobank "global" components projected with pcapred explained 7%, and local correction left effect sizes about 2% larger. Its remedy requires each site to project raw genotypes in R, which is exactly the expert step low-capacity sites struggle with. Apparent latent structure within the UK Biobank sample and the haplotype-component work in From individuals to ancestries show that fine structure persists even in "homogeneous" samples; haplotype components explained four times more educational-attainment variance than standard components. This tells us which reference axes to prefer when constructing $L$.
metaGWASmanager supplies the operational setting. It already splits work between central analysts and study analysts across CKDGen's hundreds of studies. However, its handling of ancestry is limited to consortium-specified top components. Its central quality-control step, which uses GWASinspector for allele flips and inflation, is where our audit would slot in. LOG-TRAM supplies the idea that true cross-population differences are regional. It borrows power from UK Biobank to Biobank Japan in 1 Mb windows, raising effective sample size by 49% on average. We reuse this regional logic to model $u_{kj}$, so that real regional biology is not mistaken for stratification.
Evaluating Multi-Ancestry Genome-Wide Association Methods is a constraint rather than a tool. It shows that pooling individual-level data beats meta-analysis, finding 63% more variants in All of Us. It also finds that stratification is well controlled when ancestry explains under 1% of trait variance. That tells us where our method should matter: traits entangled with ancestry, such as educational attainment. It also marks the setting our method exists for, in which pooling is legally impossible.
The agent systems for accessibility, GenoMAS and BioAgents, attack the same accessibility gap by writing code for non-experts. GenoMAS handles only expression data and reaches just 32% on clinical metadata extraction. BioAgents struggles with multi-step glue code. We avoid the problem altogether by requiring no new site-side computation.
Research Program
We will build three things.
The first is a counterpart release package. For UK Biobank, it contains variant loadings on the 200 global components and on haplotype components, along with a reference scan corrected with those components. For All of Us, it contains the same components computed across its multi-ancestry sample. Loadings and summary statistics are variant-level and non-identifying, so they are shareable.
The second is the estimator and audit. This is a weighted regression with a prior over regional terms $u$, fit on pruned variants, with calibrated tests for $\gamma_k$ and $a_k$.
The third is an analysis of identifiability. True effects can correlate with ancestry loadings, for example under selection on height. The counterpart anchor absorbs this, because $\beta^{\text{ref}}$ already contains any genuine loading-correlated signal. Only a cohort's deviation from the anchor is attributed to stratification. We will characterize when this fails, notably when the counterpart itself is mis-corrected along an axis, or when a cohort's ancestry lies outside the counterpart's range.
Baselines
Naive meta-analysis. The do-nothing reference is standard inverse-variance meta-analysis of each cohort's locally corrected summary statistics, as metaGWASmanager would produce. Every method sees the same submitted statistics.
Uniform per-cohort correction. The first cheap competitor divides each cohort's statistics by its genomic-inflation factor, a single number per cohort. The second uses the intercept of LD score regression, a regression of association statistics on each variant's correlation with its neighbours, as a per-cohort rescaling. I found no reported result on Lacuna for either in this setting. They are the strongest summary-statistics-only competitors because they need no site action either. They get the same input and no tuning beyond their defaults.
Oracle rerun. The strongest existing method is having every site rerun its association scan after projecting onto the global components with pcapred, following. Its reported ALSPAC result is 7% versus 1% variance explained and an effect inflation of about 2% removed. We treat it as an upper bound that requires site effort, and we run it on identical cohorts and phenotypes.
Pooled analysis. Where individual-level data is available (All of Us), pooled analysis is the power ceiling, following Evaluating multi-ancestry genome-wide association methods: Statistical power, population structure, and practical implications.
Ablations. The central ablation replaces the counterpart loadings with each cohort's own local-component loadings. This isolates whether the external reference is what does the work. Further ablations drop the scale term $a_k$, drop the regional term $u$, and vary the number of axes $C$.
What counts as beating the baselines. Our method succeeds if it does all three of the following:
- closes at least 70% of the gap between naive and oracle, measured by effect-size bias and lead-variant concordance;
- beats both uniform corrections on that gap;
- keeps the audit's false-positive rate at its nominal level on traits where ancestry explains little variance.
Why not just…?
Why not just have every site rerun with pcapred? Where sites can do this, they should. That is the oracle, and it is a real threat whenever consortium members are capable. The case for our method rests on three points. Reruns across hundreds of studies are slow and incomplete in practice, as metaGWASmanager's design around study-analyst burden implies. Legacy summary statistics cannot be rerun at all. And pcapred's reference is British-only. Experiment 1 measures how much of the oracle's benefit we recover without reruns. If the fraction is small, the honest conclusion is to push reruns instead.
Why not just apply genomic control or the LD score intercept? Both apply one scalar per cohort. Stratification bias, however, is signed and variant-specific, scaling with each variant's loading. A uniform correction therefore shrinks true signals while leaving the most-loaded spurious variants partly intact. It also cannot tell definition drift from stratification, since both inflate statistics. Experiments 1 and 2 compare the methods directly on known injected bias. If the uniform corrections match ours, the directional model adds nothing, and we would report that.
Why not just give study analysts an LLM agent that runs the projection for them? This does not remove the need for site-side genotype handling and compute. The closest agent system, GenoMAS, does not process genotype files at all. BioAgents reports degradation on multi-step pipelines. An agent-assisted rerun would at best reproduce the oracle, so it is complementary rather than a substitute.
Why not pool the data? Cross-border privacy rules often forbid it, as Evaluating multi-ancestry genome-wide association methods: Statistical power, population structure, and practical implications itself notes. Where pooling is possible, we use it as the ceiling.
Proposed Experiments
Experiment 1: controlled regional cohorts. We carve UK Biobank into small cohorts by birthplace, exploiting the latent structure in. Each cohort runs a local-component scan for educational attainment and for negative-control traits with weak ancestry dependence. The oracle for each cohort is its global-component scan, and the counterpart is fit on disjoint UK Biobank individuals. Metrics are:
- correlation and mean bias of corrected effects against the oracle;
- overlap of lead variants with the oracle;
- reduction in between-cohort heterogeneity ($I^2$);
- power and false-positive rate of the $\gamma$ test.
Experiment 2: injected disagreement. We add phenotype-definition drift to some cohorts, for example dichotomizing at different thresholds or rescaling units, mirroring the definition heterogeneity documented across clinical meta-analyses. We add regional true-effect differences to others. The test is whether the decomposition assigns each source correctly.
Experiment 3: real replication. We apply the method to ALSPAC's locally corrected statistics with the UK Biobank counterpart. The test is whether it recovers the roughly 2% effect-size inflation reported against global correction.
Experiment 4: multi-ancestry. We use All of Us ancestry-stratified subcohorts with an All of Us counterpart, and compare against pooled analysis. As a second test, we pair Biobank Japan–style East Asian statistics with a UK Biobank counterpart. This probes the failure case where the cohort's ancestry lies outside the counterpart's range, and there we expect, and will report, a drop in performance.
Decisive checks. Two outcomes would kill the core claim. The first is the ablation with local loadings performing as well as counterpart loadings. The second is the $\gamma$ test firing on negative-control traits.
References
- Tracking progress towards equitable maternal and child health in Yunnan: a systematic assessment for the Health Programme for Poverty Alleviation in China during 2015–2020
- Population stratification in GWAS meta-analysis should be standardized to the best available reference datasets
- Apparent latent structure within the UK Biobank sample has implications for epidemiological analysis
- From individuals to ancestries: Towards attributing trait variation to haplotypes
- metaGWASmanager: a toolbox for an automated workflow from phenotypes to meta-analysis in GWAS consortia
- Leveraging the local genetic structure for trans-ancestry association mapping
- Evaluating multi-ancestry genome-wide association methods: Statistical power, population structure, and practical implications
- GenoMAS: A Multi-Agent Framework for Scientific Discovery via Code-Driven Gene Expression Analysis
- BioAgents: Democratizing Bioinformatics Analysis with Multi-Agent Systems