When One Gene Dominates: Keeping Shared Multi-Trait Genetic Models from Relearning It

In many diseases a single stretch of the genome carries most of the inherited risk. For Alzheimer's disease it is the APOE gene; for autoimmune disorders it is the HLA region. Statistical models that predict traits from genotype spend most of their fit on these regions, and weaker genes that matter for specific biology end up hidden. Geneticists call this "signal masking." At the same time, datasets increasingly measure many related traits per person, such as amyloid and tau biomarkers, brain imaging, cognition, blood chemistry, and several diseases. The natural model for such data is multi-task learning: one shared model is trained on all traits at once, so that a compact "genetic factor" representation is reused across them.

Our claim is that masking takes a specific, fixable form in these shared models. It is not mainly a matter of task gradients fighting each other. It is a matter of representation. Because the dominant locus drives nearly every trait, the shared factors learn to encode it. They do this directly and also through nearby correlated variants, and this happens even when the dominant locus is given its own dedicated term in the model. The result is that the shared capacity is used up on a signal the model already has. We propose to remove the dominant locus from the shared representation by construction: the shared factors are forced to be uncorrelated with dominant-locus dosage. The dominant locus is then reintroduced only through trait-specific terms that may modulate how the shared factors act. We will test this against the obvious import from machine learning, gradient surgery, and we expect it to win for a reason we can measure directly.

What exists, and what is left open

Genetics has four standard ways to unmask weak signals. The first is to drop carriers: the exome-wide age-of-onset study restricted analysis to APOE ε4 non-carriers and found ERN1 and SPPL2C, and the 11-disorder autoimmune PheWAS found non-HLA loci such as CLNK by excluding the HLA region. The second is to condition on known loci: CS-LMM first fits the strong SNPs and then runs a relaxed Lasso on the residual, and the ADNI quantitative-biomarker GWAS found rs5011804 at 12p12.1 by adding ε4 dosage as a covariate in 16 separate per-trait regressions. The third is to pick genes by their link to intermediate traits, as adORS does with PET-correlated genes and EndoPRS does with p-value-binned variant sets and a weighted Lasso. The fourth is to model all variants jointly with Bayesian variable selection, which the biobank BayesC study showed resolves the cancellation of opposing nearby effects that plagues single-marker tests.

These papers motivate the proposal: they show masking is real and that removing the dominant signal reveals biology. They also constrain it. Every one of them either handles one trait at a time or combines traits through a filtering step done before fitting. None learns a representation shared across traits. The pathway-specific PRS study adds a further constraint: non-APOE pathways act on different biomarkers at different ages. A good shared model must therefore let the effect of the background genome vary across traits, and possibly with APOE status, rather than subtracting APOE once and for all.

From multi-task learning, PCGrad removes the conflicting component when two tasks' gradients point in opposing directions. RotoGrad rebalances gradient sizes and rotates task-specific views of the shared representation. The obvious intersection idea is to run these methods on multi-trait genetic models and call the "bullying" by the dominant task unmasking. ForkMerge is the key constraining result here. It shows that conflicting gradients do not in themselves cause negative transfer, since even weight decay produces them while helping generalization. That warns us that the gradient-surgery framing may be the wrong diagnosis. We take the warning seriously and make "optimization conflict versus representational capture" the central empirical question.

Core Claim

In a shared multi-trait genetic model, the dominant locus masks secondary signals mainly by being encoded in the shared factors, not mainly by gradient conflict. A simple fix that targets this capture should recover more secondary loci and predict better than gradient surgery. The fix is to constrain the shared factors to be orthogonal to dominant-locus dosage and to reintroduce that dosage only through per-trait modulating terms. It should also beat sequential conditioning whenever the background genome's effect on some trait depends on dominant-locus status.

Research Program

Let $x$ be a person's standardized genotype vector over $p$ variants, excluding the dominant-locus alleles. Let $d$ be the dominant-locus dosage vector, for example ε4 and ε2 counts or classical HLA allele dosages. Let $y_t$ be trait $t$. The model is

$$\hat y_t = b_t^\top d + \big(a_t + C_t d\big)^\top z, \qquad z = W x,$$

where $W$ is a $k \times p$ sparse shared loading matrix giving $k$ genetic factors $z$. Here $a_t$ is trait $t$'s loading on the factors, $b_t$ is trait $t$'s direct dominant-locus effect, and $C_t$ is a low-rank matrix that lets dominant-locus status scale how the factors act on trait $t$. The central ingredient is the constraint

$$\mathrm{Cov}(z, d) = 0,$$

which we enforce exactly in each minibatch by regressing $z$ on $d$ and keeping the residual. With a sparsity penalty on $W$, this costs about as much as one extra small regression per step. The constraint removes only the part of each variant that is linearly predictable from $d$. A nearby variant such as TOMM40 can therefore still contribute its independent effect, which is exactly what conditional analysis would keep. The model is linear in the genotype, apart from the modulating term, on purpose. This keeps it comparable to Lasso, BayesC, and EndoPRS, and it keeps our attributions interpretable. A shallow nonlinear encoder is a secondary arm.

The analysis has three parts. The first is a capture diagnostic, measured throughout training: the fraction of shared-factor variance explained by $d$, together with the per-step cosine similarity between trait gradients on $W$. This tells us whether conflict exists at all and whether capture tracks the loss of secondary loci. The second is recovery of secondary loci. We use stability selection on the rows of $W$ and on per-trait effective SNP weights $W^\top(a_t + C_t \bar d)$, reporting power at a fixed false discovery rate. The third is held-out prediction of each trait and of the clinical endpoint.

Baselines

The naive reference points are marginal single-marker GWAS per trait and single-trait Lasso with no special treatment of the dominant locus. They show how much masking there is to undo.

Carrier stratification follows the ERN1/SPPL2C study: drop ε4 carriers, or restrict to the non-HLA genome as in the autoimmune PheWAS. Here it is run on the same traits with per-trait Lasso. It is the standard practice, and its locus lists serve as our silver standard. Beating it means recovering its loci at equal or better false discovery rate while also using carriers, and also detecting status-dependent effects that stratification cannot estimate.

Covariate adjustment and CS-LMM are the strongest competitors on the core question, because they already remove the dominant effect cheaply. On real data, CS-LMM reports 17 novel alcoholism SNPs and 20 Alzheimer SNPs. We run it per trait, and also on the concatenated residuals of all traits, with the same variant set and a matched stability-selection budget.

Per-trait BayesC (biobank BVS) is the strongest single-trait fine-mapping method, reporting a false discovery rate near 10% where single-marker tests exceed 50%. We give it dominant-locus dosage as a fixed effect.

EndoPRS is the strongest existing multi-trait predictor for a clinical endpoint with endophenotypes. It reports test $R^2$ gains of up to 73% over Lasso and an average of 63% over MTAG in UK Biobank simulations. We run it with its own 5-fold cross-validated penalties.

On ADNI, the prediction target to beat is adORS, which reports AUROC 0.634 for cognitively normal versus Alzheimer's and 0.610 for MCI conversion, against 0.602 and 0.532 for a standard PRS.

The multi-task baselines all use the same architecture and a sum of trait losses, with the dominant-locus bypass $b_t^\top d$ included but no orthogonality constraint. We train this plain version and versions with PCGrad, RotoGrad, and ForkMerge. No reported result was found on Lacuna for any of these methods on genomic data. Their only reported numbers are on vision and reinforcement-learning benchmarks, such as PCGrad raising Meta-World MT50 success from under 40% to about 70%.

Two ablations isolate the key ingredient: orthogonality without the modulating terms $C_t$, and $C_t$ without orthogonality. Every method gets the same folds, the same variants, and an equal hyperparameter search of 50 configurations. Methods are selected on validation loss only, never on locus recovery.

Why not just…?

Why not just drop the carriers? In ADNI-sized cohorts, dropping ε4 carriers removes a large share of cases. The BayesC study shows that power for weak polygenic signals rises sharply only at large sample sizes. Stratification also cannot estimate whether a background factor acts differently in carriers, which is exactly what $C_t$ is for. Our simulations include planted carrier-specific effects, and there stratification should be unable to recover them.

Why not just condition on the dominant locus, as CS-LMM does? This is the real threat. In a purely additive world with no modulation, conditioning followed by per-trait selection should roughly match us. We will report that regime honestly as a tie. Our advantage should appear in two places, and the simulation grid is built to separate them. One is where secondary loci are shared across traits, so that pooling them into factors adds power that per-trait residual fits lack. The other is where effects depend on dominant-locus status. If CS-LMM on concatenated residuals matches us in both regimes, the representational argument fails and we will say so.

Why not just use gradient surgery on the joint model? ForkMerge's evidence suggests conflict is not the cause of harm. Our diagnostic tests this directly. We predict low measured gradient conflict and high dominant-locus capture in the plain model, with PCGrad and RotoGrad changing capture little. If instead surgery reduces capture and recovers the secondary loci, the gradient view is right. In that case the orthogonality constraint becomes a cheaper stand-in rather than a new mechanism.

Why not just use EndoPRS or per-trait BayesC? Both are strong, but their variant sets are chosen by marginal p-values or trait by trait. So they cannot share weak signals across traits through a learned representation, and they give no account of status-dependent effects. They remain in the comparison with their published tuning.

Proposed Experiments

The ground-truth testbed follows the EndoPRS simulation design. It uses about 60,000 unrelated European-ancestry UK Biobank genotypes over HapMap3 variants. We plant a dominant locus explaining between 5% and 40% of trait variance, 6 to 16 correlated traits, and a set of weak secondary loci that are either shared across traits or trait-specific. Some loci are given carrier-dependent effects. Within some LD blocks we place opposing-effect neighbors, the case that defeats single-marker tests. The primary metrics are power at 10% false discovery rate for secondary loci, test $R^2$ per trait, and error in the estimated $C_t$.

The decisive check comes from plotting, across dominance levels, the gap in secondary-locus power between our model and the plain multi-task model. We place this against that model's measured capture fraction and gradient conflict. Our thesis predicts that the gap tracks capture, not conflict.

The real-data arms use two datasets. The first is UK Biobank across the 11 autoimmune disorders with HLA as $d$ (330,841 individuals). We score recovery of the non-HLA loci that the exclusion analysis reported, including CLNK, PTPRC, and IL12RB2, and held-out disease AUROC. The second is ADNI with the 16 quantitative biomarkers and APOE as $d$. There we check whether rs5011804 is recovered without per-trait regressions and compare conversion AUROC against adORS. ADNI's roughly 1,500 participants make it a small-sample stress test rather than a showcase. All cohorts are European ancestry, which limits generalization. As a secondary analysis, we will report whether the learned factors transfer to held-out non-European UK Biobank participants.