Haplotype k-mer kernels for measuring the physical scale of cis-epistasis

Part of the "missing" heritability of complex traits appears to come from nearby genetic variants whose combined effect on a trait is not the sum of their separate effects. Statisticians call this cis-epistasis. The best current summary-statistic method, i-LDSC, finds this non-additive signal in 24 of 25 UK Biobank traits. It does so by multiplying genotypes pairwise inside a window of ±50 SNPs. That window treats a stretch of genome purely as a count of markers, and the authors themselves flag its size as arbitrary. Meanwhile, the regulatory biology of non-coding DNA says interactions should be strongest at a very particular physical scale. Two mutations that both hit the same transcription-factor binding site, which is a few to a few dozen base pairs long, should interact. Two mutations 40 kb apart in unrelated elements usually should not.

We propose to measure that scale directly. The tool comes from the $k$-mer literature. Count all length-$k$ substrings (k-mers) of each person's actual local haplotype sequences (the DNA sequence each chromosome copy carries, with that person's alternate alleles written into the reference), and use the similarity of these counts as a genetic relatedness kernel. Every k-mer that contains two variants is a joint feature of their allele combination. The k-mer length $k$ therefore is the maximum distance at which interactions are represented. Sweeping $k$ turns the classic trade-off between specificity and connectivity into a heritability-versus-interaction-range curve. Short $k$ captures only motif-scale interactions, which is specific but covers few variant pairs. Long $k$ captures interactions among more distant variants, but the features fragment into rare haplotype-specific strings that no two people share. Mismatch and gapped kernels exist precisely to repair that sparsity.

Core Claim

The claim is that non-additive cis heritability is concentrated at short physical ranges, on the order of tens of base pairs and within annotated regulatory elements. It is not spread uniformly over ±50-SNP windows. The second half of the claim is that a haplotype k-mer kernel, smoothed with mismatches, recovers more of this variance than pairwise SNP products restricted to the same physical range. It should do so because the kernel also represents three-way and higher-order allele combinations within a motif, and because it pools rare haplotypes that differ by one allele.

If both halves hold, three things follow. There is a sequence-grounded replacement for the arbitrary window in i-LDSC. There is a direct estimate of the length scale of regulatory epistasis. And there is a natural way for non-coding variant interpretation, meaning which elements and motifs matter, to feed into variance decomposition.

If only the first half holds, the result is still useful but more modest. It would say that cis-epistasis is short-range, and that simple distance-restricted pairwise terms are enough to capture it. The experiments are designed so that this outcome is detectable rather than hidden.

Why the k-mer view is the right lens

Write $s_i^{(e)}$ for individual $i$'s two haplotype sequences across a regulatory element $e$. Write $\phi_k(s)$ for the vector of k-mer counts, optionally with gapped positions or mismatch neighborhoods. The local kernel is

$$K_k^{(e)}(i,j) = \langle \phi_k(s_i^{(e)}), \phi_k(s_j^{(e)}) \rangle ,$$

and the trait model is a variance-component regression:

$$y = g_{\text{add}} + \sum_e g_k^{(e)} + \varepsilon ,$$

Here $g_{\text{add}}$ has the usual additive SNP relatedness covariance, $g_k^{(e)}$ has covariance proportional to $K_k^{(e)}$ after the additive part is projected out, and $\varepsilon$ is noise.

Two variants at distance $d<k$ co-occur in $k-d$ shared k-mers. Three variants within $k$ co-occur too. So $K_k$ contains pairwise and higher-order local interaction features automatically, and $k$ caps their range. This is the specific property that pairwise products lack.

The sparsity problem at large $k$ is the same one documented for composition features growing as $4^k$ (Analysis of composition-based metagenomic classification). The remedies are borrowed rather than invented. Mismatch neighborhoods pool near-identical substrings, following Mismatch String Kernels for SVM Protein Classification, whose best settings were small, around $k=5,6$ with $m=1$. Gapped k-mers let a fixed number of informative positions span a longer distance, and the sort-and-count algorithm of GaKCo makes them practical. GaKCo reports a 2x speedup on DNA over gkm-SVM with identical AUC, and also reports that gapped kernels beat CNNs when training sets have fewer than 1,000 sequences.

Because haplotypes within a population differ at only a handful of sites per element, the number of distinct k-mers per element is small. This keeps the kernel cheap even at biobank scale.

As a geometry variant we will also test a Laplacian kernel on $l_1$ distance between profiles. The motivation is Revisiting K-mer Profile, which proves that $l_1$ distance between k-mer profiles bounds edit distance, while cosine and inner-product metrics distort sequence space. That paper only studied microbial binning. Whether its geometry helps at single-allele resolution is an open question that this work answers in passing.

The dual-scale idea of MinDivLP supplies the second resolution. That work keeps a small $k$ for tractable equations and a larger $h$ for similarity structure. Analogously, we keep short $k$ for the interaction features and use element boundaries (the larger scale) to decide which k-mers are pooled into one variance component.

Research Program

The first component is the estimator. We build haplotype sequences for UK Biobank participants over two sets of regions: regulatory elements from the Roadmap reference epigenomes (Integrative analysis of 111 reference human epigenomes), and matched non-regulatory control regions. We compute exact, mismatch, and gapped k-mer kernels for $k$ from 4 to roughly 64 (gapped beyond about 12). Variance components are fit with randomized Haseman–Elston regression so the method scales to hundreds of thousands of individuals. The additive relatedness matrix is always included, so $g_k$ captures only variance beyond additivity.

The second component is a summary-statistic version, so that results can be replicated where only GWAS summary statistics exist, such as BioBank Japan. Following the logic of i-LDSC, which showed that additive marginal statistics tag cis-interaction effects, we replace its pairwise interaction LD scores with "k-mer interaction LD scores." These are the LD, computed in an ancestry-matched reference panel, between each SNP and the additive-residualized k-mer presence features of its element. That yields an i-LDSC-style regression in which the window is defined by sequence features at scale $k$ rather than by a SNP count.

The third component links back to non-coding interpretation. For each trait, the curve of non-additive heritability against $k$ is partitioned by element class and tissue activity. If epistasis lives in regulatory grammar, the curve should peak at motif-to-enhancer scales. Its mass should also concentrate in elements active in trait-relevant tissue, mirroring the cell-type enrichment seen for multiple sclerosis variants in B-cell enhancers (Disanto et al.).

Baselines

The naive reference point is additive-only heritability from the same variance-component machinery with no local interaction term. Every non-additive estimate is reported as an increment over it, so the do-nothing model is literally the null of our test.

The strongest existing method is i-LDSC with its published ±50 SNP window, run on the identical 25 UK Biobank and 25 BioBank Japan summary-statistic traits and reference panel. Its reported result is significant non-additive cis heritability in 24 of 25 UK Biobank traits and 23 of 25 BioBank Japan traits, with good calibration in additive-only simulations. The file reports no per-trait magnitudes for i-LDSC heritability, so we will rerun it ourselves rather than quote numbers. Beating it means two things. First, our method must remain calibrated in additive-only simulations. Second, it must estimate equal or larger non-additive heritability while using a much smaller interaction feature set, or it must produce a better-fitting regression intercept and higher out-of-sample prediction in held-out UK Biobank individuals. We will also run i-LDSC averaged over several SNP windows, as its authors suggest, to give it its own best configuration.

The ablation that isolates our key ingredient is pairwise SNP products restricted to the same physical distance $d<k$ and the same regulatory elements, with an identical tuning sweep. Anything the k-mer kernel gains over this ablation is attributable to higher-order combinations and mismatch pooling, not to the choice of distance range. A second ablation removes mismatch smoothing ($m=0$) to measure the cost of sparsity at large $k$. A third swaps regulatory elements for random matched intervals.

For the variant-level check, the strongest scorers on TraitGym are CADD and GPN-MSA for disease traits and Enformer or Borzoi for non-disease traits. The only figure the file gives is CADD at AUPRC 0.85 on Mendelian variants. No reported result was found on Lacuna for complex-trait AUPRC, so we will rerun those models on the complex-trait split.

Why not just…?

Why not just make additive GWAS bigger? More samples sharpen additive estimates, but they cannot turn interaction variance into additive variance. i-LDSC's non-additive signal is already detected at about 350,000 individuals (i-LDSC). The cardiac loci found by lo-siRF were insignificant marginally even though their interaction was validated in cells, with a non-additive 8.5% reduction in cardiomyocyte area from dual knockdown. Our additive baseline will show how much variance remains after scale is exhausted on the cohort at hand.

Why not just tune i-LDSC's window, or restrict pairwise products by base-pair distance? This is the real threat. If cis-epistasis is purely pairwise, then distance-restricted pairwise terms will match the k-mer kernel. In that case the honest contribution reduces to measuring the length scale. The matched-distance ablation settles it directly, and we will report that outcome as such rather than argue around it. Simulations with planted three-way motif interactions on real UK Biobank haplotypes show where the kernel should win if biology resembles them.

Why not just score variant pairs with a foundation model like AlphaGenome? AlphaGenome predicts molecular tracks from 1 Mb of sequence. It reports gains such as eQTL sign auROC rising from 75% to 80%, but its authors state it does not predict complex phenotypes. It also scores single variants, not the population-level variance that interactions contribute. In-silico double-mutant scores are nonetheless a legitimate prior. We will compute them for candidate pairs in the top elements and test whether up-weighting kernel features by double-mutant non-additivity increases heritability over the unweighted kernel. If it does, the foundation model becomes a component of the method, not a rival.

Why not just define windows by TADs or annotations? Annotation windows decide where to look but still use pairwise SNP products inside them. Our element-restricted pairwise ablation is exactly this alternative. The k-mer kernel differs in what it counts within the window, so the same comparison adjudicates both.

Proposed Experiments

Simulations come first and are decisive for credibility. Using real UK Biobank haplotypes, we generate traits under four architectures: additive only, pairwise interactions at controlled distances, higher-order interactions within planted motif-length spans, and interactions between distant elements. For each, we check false-positive rates under the additive-only architecture, recovery of the true interaction range from the curve's peak, and power relative to i-LDSC and the matched pairwise ablation.

On real data, we estimate the curve of non-additive heritability against $k$ for the 25 continuous traits, with height as the primary case, since i-LDSC found its strongest signal there. We then replicate the summary-statistic version in BioBank Japan. The key prediction is a peak below about 100 bp, enriched in tissue-active elements. A flat curve, or one rising with $k$, would refute the motif-scale hypothesis.

Two external checks follow. First, we test whether TraitGym complex-trait causal variants (TraitGym) sit in elements with higher per-element kernel heritability than their nine matched decoys. A per-element kernel heritability score is then evaluated as an added feature alongside CADD and Borzoi on the complex-trait split, where existing models are weakest on enhancer variants. Second, we ask whether left ventricular mass in the 29,661 imaged participants shows element-level kernel signal near CCDC141, IGF1R, TTN, and TNKS, the loci lo-siRF identified and validated.

Success on simulations, a short-range peak replicated across both biobanks, and a measurable gain over the matched pairwise ablation together would establish the claim. Success on the first two without the third would still establish the length scale, but would show that higher-order k-mer features are unnecessary.