Do DNA Sequence Models Know Where Epistasis Lives?

Part of the heritability of complex traits never shows up in genome-wide association studies, which scan for DNA variants associated with a trait one at a time and assume each variant adds its effect independently of the others. A recent method, interaction LD score regression (i-LDSC), found that interactions between nearby variants contribute measurable heritability for almost every trait it tested. It reads this signal from ordinary association summary statistics, but it cannot say which variant pairs carry it or why. Separately, DNA sequence models such as Enformer and AlphaGenome predict how a stretch of DNA drives gene expression. These models can be asked a counterfactual question: if two variants occur together on the same chromosome copy, is their joint effect on expression different from the sum of their separate effects? This note proposes to compute that "in silico double mutant" interaction for common nearby variant pairs. We then test whether the pairs the model flags explain a disproportionate share of the interaction heritability that i-LDSC detects.

The test has a property most epistasis analyses lack: it can tell real biochemical interaction apart from a measurement-scale artifact. An interaction caused by DNA grammar, such as two transcription-factor sites cooperating inside one enhancer, should act only when both variants sit on the same chromosome copy ("in phase"). Apparent epistasis that comes from a curved relationship between expression and the trait should not depend on phase at all. Phased biobank and expression data let us check which kind we are seeing. If sequence-predicted pairs carry enriched, phase-dependent interaction variance, that would be the first variant-level, mechanistic account of the non-additive heritability that summary-statistic methods can only count.

Core Claim

Our hypothesis is that a sequence-to-expression model's predicted interaction score is an informative prior over where cis-interaction heritability sits. For a variant pair $(a,b)$ near gene $g$, let $f_g(s)$ be the model's predicted log expression of $g$ (for example, the CAGE signal at its transcription start site in a relevant tissue) for haplotype sequence $s$. The interaction score is

$$I_g(a,b) = f_g(s_{ab}) - f_g(s_a) - f_g(s_b) + f_g(s_0),$$

where $s_0$ is the reference haplotype, $s_a$ and $s_b$ carry one alternate allele each, and $s_{ab}$ carries both on the same molecule. On the log scale, independent multiplicative action gives $I_g = 0$, so a nonzero score signals cooperativity or interference between sites, not a scale effect. The claim is falsifiable in three ways. Pairs ranked high by $|I_g|$ should show more i-LDSC interaction heritability per pair than matched pairs. Their effect on observed expression should depend on phase. That enrichment should survive conditioning on simpler annotations that already mark "both variants are in an active element."

Relation to Prior Work

The work that motivates this proposal established that non-additive signal exists but left its mechanism unexplained. Discovering non-additive heritability using additive GWAS summary statistics introduces i-LDSC, which regresses association statistics on "interaction LD scores" built from local pairwise SNP products within ±50 SNPs. It reports significant cis-interaction contributions in 24 of 25 UK Biobank traits and 23 of 25 BioBank Japan traits, with calibrated type I error under purely additive simulations. It is entirely genotype-statistical and weights all local pairs equally. Our contribution is to replace that uniform weighting with sequence-model weights, in the same way stratified LD score regression partitions additive heritability by annotation. Epistasis regulates genetic control of cardiac hypertrophy uses a tree-based search (lo-siRF) on individual-level UK Biobank data to find interacting loci for left ventricular mass. It validates one pair by siRNA co-knockdown in stem-cell-derived cardiomyocytes, but it maps loci to genes afterwards using static Hi-C and eQTL lookups. We aim instead to predict which pairs interact from sequence before looking at trait data.

The method comes from the sequence models. Enformer predicts expression and chromatin tracks from about 200 kb of sequence and improved causal-variant prioritization in 47 of 48 GTEx tissues. AlphaGenome extends context to 1 Mb at single-base resolution across 11 data types, raising eQTL sign prediction from 75% to 80% auROC. Both have been evaluated almost entirely on single variants. TraitGym benchmarks sequence models on isolated causal non-coding variants against 1:9 matched decoys. It finds that models trained on functional genomics data do best on complex traits and that all models struggle on distal enhancer variants. That last finding constrains us: we expect predicted interactions to be most reliable near promoters and least reliable for distal pairs, and we stratify by distance throughout.

The work on nonlinear and network TWAS (transcriptome-wide association studies, which test genes rather than variants by imputing expression from genotype) is the closest competitor, and it is a different problem. scPrediXcan transfers Enformer embeddings to predict cell-type expression and then linearizes the model into a SNP elastic net. That step discards any within-haplotype interaction. Inference of nonlinear causal effects with application to TWAS (2SIR+AIR) models a nonlinear map from expression to trait. It recovers 18 Alzheimer's causal genes versus 11 for linear two-stage least squares. Its nonlinearity lies downstream of expression, which is exactly the scale-type non-additivity our phase test is designed to separate from sequence grammar. "Just un-linearize scPrediXcan" is therefore not our idea. A nonlinear TWAS could gain power purely from expression-to-trait curvature without saying anything about variant-variant interaction in DNA. Bipartite Community Structure of eQTLs (CONDOR) shows that GWAS SNPs concentrate as local "core" members of eQTL communities. We use its spirit as a baseline annotation: pairs regulating the same gene community.

Research Program

The first component is a scoring pipeline. For each protein-coding gene, we take common variants (minor allele frequency above 1%) within the model's reliable context around the transcription start site. Scoring all pairs is too expensive, so we keep pairs within a fixed physical distance of each other, initially 2 kb and then up to 20 kb. This window covers same-element and adjacent-element pairs, where cooperative binding is biochemically plausible. Within it, we prioritize pairs where at least one variant has a nonzero single-variant predicted effect. Each retained pair needs four forward passes. We compute $I_g$ on log-scale expression tracks for tissues matched to the trait (liver and adipose for lipids, following tissue-specificity results such as Surakka et al.; heart for cardiac traits). We also compute the analogous score on accessibility and transcription-factor-binding tracks, to see whether interactions arise at the chromatin level. To keep scores honest about noise, we estimate a null distribution for $I_g$ from pairs in which one "variant" is a synthetic substitution at a matched position with no predicted single effect. We also check that $I_g$ is stable across the reverse-complement and shifted-window predictions the models already ensemble.

The second component is an annotation-stratified extension of i-LDSC. The original method uses one interaction LD score per SNP, summing squared LD with local pairwise products. We split that sum into bins by predicted $|I_g|$ and fit a coefficient per bin. This gives an estimate of interaction heritability per pair in each bin, and of the share captured by the top bin. We fit jointly with additive baseline annotations and with the competing interaction annotations described below, so enrichment is measured conditionally.

The third component is phase-resolved validation at the expression level, using GTEx individual-level genotypes and expression (Lonsdale et al.). The test compares double heterozygotes in coupling (both alternate alleles on one copy) with those in repulsion. If the interaction comes from sequence grammar, the expression difference between the two groups should track the sign and size of $I_g$. Allele-specific expression gives a second, within-individual readout that removes much between-donor confounding. The same contrast in phased UK Biobank genotypes tests phase dependence at the trait level for the top pairs.

Baselines

The naive reference is i-LDSC with uniform weighting of local pairs, as published. Our sequence-weighted bins must show enrichment beyond 1 (interaction heritability per pair above the genome-wide average) for the method to matter at all. Below that is additive stratified LDSC with no interaction term, which defines the heritability that pairs could add on top. We run all variants on identical UK Biobank summary statistics, the same 1000 Genomes LD reference, and the same ±50 SNP interaction window used by i-LDSC.

The strongest annotation competitor needs no deep model. It weights a pair by whether both variants fall in the same tissue-specific functional element, using GenoSkyline posteriors (Lu et al.) or Roadmap/ENCODE active marks (Kundaje et al.). We pair this with a CONDOR-style annotation in which both variants are eQTLs for the same gene or community. This is the real threat. If co-occurrence in an enhancer explains the enrichment, the sequence model adds nothing. We give this baseline the same bin count and the same tissue choice. Beating it means sequence-weighted bins remain significantly enriched when fit jointly with it, and the reverse is not true.

The ablation that isolates the key ingredient is a scale-only score. It predicts the pair's joint effect by combining the two single-variant predicted effects under a multiplicative model, and it measures departures on the linear expression scale. This score uses the same model and compute but contains no true double-mutant information. If it matches $I_g$ in enrichment, the signal is about effect size and scale rather than grammar. A second ablation shuffles $I_g$ among pairs with matched single-variant effects and distances.

The strongest existing method that finds specific interacting pairs is lo-siRF (Wang et al.). We compare by asking whether its left ventricular mass interacting loci near CCDC141, TTN and IGF1R contain pairs ranked highly by $I_g$ in heart tracks. The reported reference point is functional rather than statistical: an 8.5% non-additive decrease in median cardiomyocyte size, P < 1×10⁻⁴, on dual knockdown. We found no reported result on Lacuna for any sequence model's accuracy at predicting variant-pair interactions, so no published number exists to beat. For single-variant ability, the reference points are AlphaGenome's 80% auROC on eQTL sign and the TraitGym evaluation.

Why not just…?

Why not just use i-LDSC as is, since it already detects the interaction heritability? i-LDSC answers whether interaction variance exists, not where it is. Its uniform local weighting cannot name a pair, a gene, or a mechanism. The comparison above against unweighted i-LDSC directly measures how much localization the sequence prior provides. If enrichment is near 1, we will report that the model gives no information about location.

Why not just weight pairs by shared enhancer or eQTL membership, which is cheaper? This may work nearly as well, and we treat it as the main competitor rather than a straw man. Its limitation is structural: an annotation marks pairs that could interact but cannot say whether a particular pair of alleles cooperates, interferes, or acts independently. Lu et al. note that GenoSkyline does not predict allele consequences. The joint fit settles it. If the annotation absorbs all the enrichment, the paper's conclusion becomes "interaction heritability concentrates in shared active elements," and we drop the sequence-model claim.

Why not just model nonlinearity between expression and trait, as 2SIR+AIR does? That captures a real and different source of non-additivity. Dai et al. show it matters, for example for TOMM40. It predicts no phase dependence, however, because a curved expression-to-trait map responds to total expression regardless of which chromosome copy carries which allele. The coupling-versus-repulsion test discriminates between the two. If enrichment is real but phase-independent, the right reading is that sequence models are locating scale effects, and we will say so.

Why not just un-linearize scPrediXcan into a nonlinear deep TWAS? It could raise association power, but power gains would not show that variant-variant grammar contributes heritability. A nonlinear expression predictor fit on unphased genotypes mixes all three sources: grammar, scale, and LD with unobserved variants. Our design targets one quantity, the in-phase double-mutant term, so that its contribution can be measured separately.

Proposed Experiments

The first decisive check is the enrichment test. We run stratified i-LDSC across the 25 UK Biobank continuous traits, with BioBank Japan as ancestry replication. The primary outcome is the conditional enrichment of the top $|I_g|$ bin over the shared-element and scale-only baselines. Height, which had the strongest cis-interaction signal in i-LDSC, and lipid traits with well-characterized liver and adipose eQTLs are the pre-specified primary traits. The second check is the GTEx phase test. For pairs in the top bin, the coupling-minus-repulsion expression difference, together with allele-specific expression imbalance, should correlate in sign with $I_g$, while matched low-score pairs should show no phase dependence. We would treat a sign agreement clearly above 50%, judged against the shuffled-$I_g$ null, as evidence that the model's double-mutant predictions carry real information. The third check is the cardiac case study: are the published lo-siRF loci enriched for high-$|I_g|$ heart-track pairs relative to matched loci from additive GWAS?

The fourth check is a simulation. We use real UK Biobank genotypes with pure additive architecture, then add unobserved rare-variant haplotype effects. If $I_g$-weighted bins produce spurious enrichment in this setting, it would mean the model is picking up haplotype tagging rather than interaction, and any real-data result would have to be discounted. All analyses are stratified by pair distance and distance to the promoter, because TraitGym indicates sequence models are weakest on distal enhancers. We expect, and will report, a gradient of reliability. A credible negative result is also informative: if no sequence-weighted bin is enriched and phase dependence is absent, the non-additive heritability i-LDSC detects is unlikely to come from cis-regulatory grammar that current sequence models represent.