Removing Population-Specific Linkage Shortcuts from Weakly Supervised Fine-Mapping Models

Deep sequence models now learn to pick out causal genetic variants without any variant-level labels. They do this by treating each block of correlated variants that carries a disease association as a "bag" that contains at least one causal variant. The bag boundaries come from linkage disequilibrium (LD): the tendency of nearby variants to be inherited together, which makes harmless neighbours look as associated as the true driver. LD patterns are a product of a population's demographic history, so the bags reflect one population's history. A network rewarded for scoring positive bags highly can therefore learn a shortcut. It can learn sequence features that predict how much a variant is correlated with its neighbours in Europeans, because heavily correlated variants sit in positive bags more often, rather than features that predict regulatory function. That shortcut would fit training data and then fail in populations whose LD differs, which is exactly where these models are supposed to help.

We propose to measure this shortcut directly and remove it with a narrowly targeted adversary. The adversary is a small network that tries to predict, from the model's internal variant representation, the difference in LD between ancestries at that variant. A gradient reversal layer trains the representation to defeat it. The key design choice is the target. We do not try to strip all LD information, because some LD structure (recombination hotspots, regions under selection) is shared across populations and may correlate with real function. We strip only the ancestry-specific component, which reflects demography and cannot cause a variant to regulate a gene.

Core Claim

Multiple instance fine-mapping (MIFM) trains a convolutional network on 512-bp DNA windows. It groups variants into positive bags (GWAS-associated LD blocks) and negative bags (unassociated regions) and max-pools variant scores within each bag. It reports a 17% relative improvement in African-ancestry polygenic score transfer over standard methods. Its bags are built from one population's LD, and nothing in its training objective prevents the network from encoding that population's LD structure. Our claim has two parts, both testable. First, embeddings from an MIFM-style model trained on European-ancestry bags linearly predict the European-minus-African LD score better than chance; we define a variant's LD score as the sum of its squared correlations with nearby variants. Second, an adversary on that ancestry-differential target improves held-out-ancestry fine-mapping and score transfer more than the simpler debiasing tricks described below, and it does so without losing in-ancestry performance.

Formally, let $f_\theta$ map a variant's sequence window to an embedding, let $\mathcal{L}_{\text{MIL}}(\theta)$ be the bag-level max-pooled classification loss, and let $\Delta\ell_v = \ell_v^{\text{EUR}} - \ell_v^{\text{AFR}}$ be the difference in LD score for variant $v$ between European and African reference panels. An adversary $g_\phi$ regresses $\Delta\ell_v$ from $f_\theta(v)$. The model is trained as

$$\min_\theta \max_\phi \; \mathcal{L}_{\text{MIL}}(\theta) - \lambda \, \mathbb{E}_v\big[(g_\phi(f_\theta(v)) - \Delta\ell_v)^2\big],$$

where $\lambda$ sets how strongly the representation is pushed to be uninformative about ancestry-specific LD. The gradient reversal layer implements the inner sign flip, as in the stroke-segmentation model FAN-Net. With more than two training ancestries, $\Delta\ell_v$ becomes the vector of each ancestry's deviation from the cross-ancestry mean LD score.

Relation to Prior Work

Prior work plays three different roles here.

Some work supplies the method. Adversarial discriminators with gradient reversal are routinely used to remove scanner or site fingerprints from medical images; examples are prostate histopathology adaptation, which lifts TCGA→CINJ accuracy from 56.3% to 75.0%, and FAN-Net. Those methods remove a categorical domain label. We instead remove a continuous, per-variant nuisance quantity defined by population genetics. We also make no attempt to make whole populations indistinguishable, since reference sequence is identical across ancestries.

Some work motivates the gap. The stability-guided fine-mapping study runs a fine-mapping method separately in five GEUVADIS sub-populations and calls variants stable if they stay likely-causal across them. Its stable variants that agree with the residualized top variant (56% of loci) are more enriched in 361 of 378 functional annotations. That supports the premise that ancestry-varying signal is largely tagging artefact. The method, however, is post hoc and splits samples. XMAP exploits LD differences across ancestries inside a variational Bayesian model and removes "ghost" signals such as rs2053005. It needs summary statistics from every population and learns no sequence representation, so it cannot score variants in an ancestry it has not seen. We use cross-ancestry LD differences as a training-time constraint on a sequence model, not as inference-time data.

Other work constrains the design. IDEA and the general literature on invariance penalties warn that strict invariance can discard useful signal. This is why we target only the ancestry-differential LD and not total LD. Episodic training for domain generalization documents how sensitive such methods are to the trade-off weight. We therefore select $\lambda$ on one held-out ancestry and report results on a different one.

Research Program

The first stage is diagnostic and could end the project. We will reimplement MIFM-style training on UK Biobank European-ancestry GWAS of blood biomarkers and height, building bags from European LD in 1000 Genomes. We then fit linear probes on frozen embeddings to predict European LD score, African LD score, and $\Delta\ell_v$, on held-out chromosomes. As controls, we compare against probes on a randomly initialised network and on a model trained with shuffled bag labels. If $\Delta\ell_v$ is not decodable beyond these controls, there is no shortcut to remove, and we will report that negative result as the finding.

The second stage trains the adversarial model. Bags are pooled from European UK Biobank and East Asian Biobank Japan GWAS, and the adversary targets per-ancestry LD deviations. African ancestry is held out entirely for evaluation; East Asian results are used to tune $\lambda$ and probe capacity, which partly addresses the missing-target-labels problem.

The third stage is a controlled simulation. On real 1000 Genomes haplotypes, we simulate phenotypes whose causal variants are drawn from sequence-defined regulatory annotations. We also deliberately plant a shortcut: causal variants are made more likely in regions of high European-specific LD. This gives ground truth for whether the adversary recovers true causal variants rather than LD proxies.

Baselines

The naive reference point is single-marker regression ranking. It reflects the proxy problem the whole field is trying to escape; BayesC narrows median peak width to roughly 30–36 kbp versus about 142 kbp for single-marker regression in biobank-scale Bayesian variable selection.

The strongest existing competitor on the actual question is unmodified MIFM, retrained by us on identical bags, windows, architecture, and compute. Its reported advantage is a 17% relative gain in African-ancestry polygenic score transfer, with 250 significant wins against 119 losses over baselines. Beating it means a statistically significant improvement in African-ancestry transfer and in held-out simulated causal-variant recall, with in-ancestry bag AUROC within one standard error.

We include three statistical baselines run on the same summary statistics. SuSiE with the residual-variance correction from the small-sample calibration study is the per-ancestry statistical reference. XMAP is the multi-ancestry reference; it gets African summary statistics that our model never sees, which makes it an upper-bound-style comparison and not a like-for-like one. PRS-CS is the score-construction reference.

Two ablations isolate the key ingredient: the same adversary targeting total European LD score instead of the ancestry difference, and the same adversary targeting a binary ancestry-of-origin label for each bag. If the total-LD adversary matches ours, the ancestry-differential target is unnecessary. If it is worse in-ancestry, that confirms invariance overreach.

All learned methods get the same hyperparameter budget of trials, tuned on East Asian validation only.

Why not just…?

Why not just match negative bags to positive bags on LD score and allele frequency? This is the classic fix for LD-score confounding and the most serious threat to the proposal. Matching removes the marginal link between bag label and a scalar covariate. It does not stop max-pooling from preferring the most heavily correlated variant inside a positive bag. Matching on a multi-ancestry LD profile also shrinks the pool of eligible negatives quickly. Even so, we have no reported result on Lacuna for MIFM with matched negatives, so this baseline gets the same tuning budget as our method. If it ties the adversary, we will say that matching is sufficient.

Why not just give the model the LD score as an extra input during training and drop it at test time? This "explain away the bias" approach absorbs the shortcut into a separate branch. It handles a nuisance the model sees directly. Our concern is a nuisance the network reconstructs from sequence context, which a side input does not prevent the main branch from also encoding. We will run it and probe its embeddings for $\Delta\ell_v$. Residual decodability would show it falls short.

Why not just build bags from the intersection of credible sets across ancestries, so that bags are already small? Intersected bags inherit XMAP's requirement of GWAS from every population. They also leave the model untrained on loci where only one ancestry has power. We include intersected bags as an alternative training set. The decisive comparison is on loci where only European data exist.

Why not just use PC residualization? The stability study shows residualized top variants disagree with stable variants at 44% of loci and are less functionally enriched. More basically, residualization corrects genome-wide stratification, not the local tagging that produces the shortcut.

Decisive Checks

The project succeeds if three things hold. Embeddings of plain MIFM decode $\Delta\ell_v$ above controls. The adversarial model drives that decodability to chance on held-out chromosomes. And this coincides with three improvements over plain MIFM and matched-negatives MIFM, without in-ancestry loss: better African-ancestry polygenic score correlation in UK Biobank, higher recall of planted causal variants in the shortcut simulation, and higher overlap with GEUVADIS Yoruban stable eQTL variants as defined by the stability study.

The idea fails cleanly in either of two cases. Decodability may be absent from the start. Alternatively, removing it may leave transfer unchanged; that would mean the shortcut exists but is not what limits portability.