Homoeolog Epistasis as a Target for Missing Heritability in Polyploid Crops

Many major crops are polyploid: they carry several complete chromosome sets, often inherited from different ancestral species. Bread wheat has three such sets (the A, B and D subgenomes); cotton and rapeseed have two. Most genes therefore exist as two or three near-identical copies called homoeologs, one per subgenome, sitting on different chromosomes. When one copy is disabled, the others often compensate. That kind of compensation is a gene–gene interaction, or epistasis: the effect of a variant in one copy depends on the state of the other copies. Genetic association studies built on additive models cannot capture it, and genome-wide searches for epistasis give up statistical power because they must test every pair of variants.

This proposal restricts the epistasis search to the pairs that polyploid biology says are most likely to interact, namely homoeologous copies of the same gene. The resulting interaction component is one variance term whose size grows with the number of genes, not with the square of the number of variants. The central question is whether this small, structured component recovers a measurable share of the heritability that additive models miss, and whether it does so more strongly than an equal number of randomly chosen cross-chromosome pairs. If it does, polyploid crops become a setting where one source of missing heritability can be named in advance rather than searched for blindly.

Core Claim

In allopolyploids (polyploids formed by hybridization between species, whose subgenomes pair separately at meiosis), a variance component made only from interactions between homoeologous gene copies should explain non-additive trait variance well beyond what equally sized sets of matched random between-chromosome pairs explain. The effect should be concentrated where more than one copy is actually expressed.

Two lines of evidence motivate the claim. Functional work shows that individual homoeologs are not interchangeable. Only the D-subgenome copy TaNRX1-2D was shown to mediate salt tolerance (The Wheat Nucleoredoxin TaNRX1-2D Gene Ameliorates Salt Tolerance in Wheat (Triticum aestivum L.)). In the 180-member phosphate transporter family, the phosphorus-efficient and inefficient cultivars differed in expression of a specific copy, not in which genes they carry (Characterizing the wheat (Triticum aestivum L.) phosphate transporter gene family…). Regulatory work shows that subgenomes are silenced unevenly. A synthetic AADD wheat downregulates its D subgenome more than its A subgenome, mainly through regulatory changes near the genes themselves (Spatial and Temporal Transcriptomic Heredity and Asymmetry in an Artificially Constructed Allotetraploid Wheat (AADD)). Repressive chromatin marks differ between subgenomes far more than active marks do (Dynamic chromatin regulatory programs during embryogenesis of hexaploid wheat). Together these predict that the functional consequence of a variant in one copy depends on whether the other copies are expressed and intact. That dependence is epistasis of a specific, predictable shape.

The main outcome is an interpretable split of heritability into an additive part, a homoeolog-interaction part and a remaining non-additive part. Prediction accuracy is a secondary outcome.

Relation to Prior Work

The closest existing work is SHEsisPlus, a toolset for genetic studies on polyploid species. It phases multi-allelic polyploid markers and detects "pure" epistasis using Shannon-entropy interaction information after splitting phenotypes into high and low groups. On simulated tetraploids it reports epistasis detection power of 0.923, against 0.056 for linear regression in PLINK. It finds specific interacting marker sets but estimates no variance components, so it cannot say how much heritability epistasis accounts for. It also has no notion of which pairs are homoeologous. The present proposal does the reverse: it fixes the pair set from genome structure and estimates variance.

The statistical machinery comes from the marginal epistasis line. MAPIT tests each variant's interaction with the whole genetic background using a fast method-of-moments estimator. It reported roughly 10% epistatic against 3.5% additive variance for gene expression in GEUVADIS, a human expression dataset. smer supplies stochastic trace estimation that scales this approach, and mvMAPIT extends it to several traits. These methods find interaction hubs but cannot name the partners. A homoeolog-restricted kernel names the partner in advance.

i-LDSC recovers non-additive heritability from summary statistics, but only for interactions within ±50-SNP windows. It therefore cannot see homoeologs, which lie on different chromosomes by construction.

Two papers constrain the design. Bilton et al. show that low-depth genotyping-by-sequencing inflates the diagonal of genomic relationship matrices, from 1.81 to 1.35 after correction in their F2 families. Any spurious diagonal inflation would contaminate non-additive kernels, which are built from products of these matrices, so their correction is required preprocessing. The potato study of Gong et al. illustrates a common shortcut: it diploidized the tetraploid and ran additive GWAS only. The shortcut discards exactly the copy-level structure studied here. The Arabidopsis STR study showed that short tandem repeats explain around 10% of phenotypic variance in a diploid. That supports the premise that missing heritability hides in variant classes standard models ignore, and repeat variants at homoeologous loci are a natural later extension.

Research Program

The first step is to define homoeolog groups. Protein-to-genome alignment with pangene builds a gene graph across reference assemblies. Groups (pairs in tetraploids, triads in wheat) are then taken as syntenic single-copy nodes with exactly one member per subgenome. The pangene authors note that their gene-selection heuristics are ad hoc. Group calls will therefore be checked against synteny, and all results will be reported separately for high-confidence and relaxed group sets.

For individual $k$, let $s_{ik}$ be a gene-level genotype score for copy $i$, for example the burden of predicted loss-of-function and nonsynonymous variants, standardized. The model is

$$y = X\beta + g_A + g_H + g_R + \epsilon,\qquad g_A\sim N(0,\sigma_A^2 K_A),\; g_H\sim N(0,\sigma_H^2 K_H),\; g_R\sim N(0,\sigma_R^2 K_A\circ K_A),$$

The terms are defined as follows. $K_A$ is the depth-corrected additive relationship matrix. $K_H = \frac{1}{m}\sum_{(i,j)\in\mathcal{H}} w_{ij}\,(s_i\circ s_j)(s_i\circ s_j)^\top$ is the homoeolog-interaction kernel, summed over the set $\mathcal{H}$ of $m$ homoeologous copy pairs. The operator $\circ$ is the elementwise product. The weight $w_{ij}$ equals 1 in the unweighted model and, in the expression-aware model, is the product of the two copies' expression shares, so that pairs with one silenced copy contribute little. The term $K_A\circ K_A$ absorbs generic pairwise epistasis, which forces $\sigma_H^2$ to capture only variance specific to homoeolog pairs. Estimation uses the method-of-moments and stochastic trace approach of MAPIT and smer, so no expensive likelihood fitting is required.

The key test is enrichment. $\sigma_H^2$ is compared with the distribution of the same quantity when $\mathcal{H}$ is replaced by 1,000 random sets of between-chromosome gene pairs. The random pairs are matched on variant frequency, burden score variance, expression level and subgenome membership, so that "two genes on different chromosomes" is held constant and only homoeology differs. Expression weights come from tissue-matched RNA data such as the AT2 atlas and, for embryo-related traits, from the wheat embryogenesis chromatin maps.

Baselines

The naive reference is additive-only GBLUP (genomic best linear unbiased prediction, a standard model that predicts traits from an additive relationship matrix), using the Bilton depth-corrected matrix. It gives the additive heritability that every other model must beat. Every model is fitted to the same samples, the same markers and the same cross-validation folds.

The strongest variance-based competitor is the additive plus full Hadamard epistasis model ($K_A + K_A\circ K_A$), which in principle contains all pairwise interactions, homoeologous ones included. It has the same estimator and compute budget as the proposed model, with no extra tuning because both have no free hyperparameters beyond the variance components. Beating it means two things. First, adding $K_H$ significantly increases explained variance on held-out folds. Second, $\sigma_H^2$ lies above the 95th percentile of the matched-random-pair null.

The strongest detection competitor is SHEsisPlus. It is run on the same genotypes, and its top interactions are checked for homoeolog enrichment. The only reported figure is its simulated-tetraploid power of 0.923. No reported result was found on Lacuna for variance explained by epistasis in any real polyploid crop panel, so the comparison there has no published target. MAPIT, run per gene burden score, is the marginal-hub competitor. The question is whether its significant hubs are the genes the homoeolog kernel singles out.

Two ablations isolate the key ingredient. The random-pair kernel isolates homoeology itself. The unweighted versus expression-weighted $K_H$ isolates the regulatory-asymmetry hypothesis.

Why not just…?

Why not just fit the full Hadamard epistasis kernel and be done? This is a real threat for prediction. If homoeolog interactions are a small share of all pairs, the full kernel may predict as well while hiding their contribution. The proposal does not claim better prediction a priori. It claims that the homoeolog share can be identified and is enriched. The model already includes the full kernel alongside $K_H$, so if $\sigma_H^2$ collapses to zero once $K_A\circ K_A$ is present, the claim fails cleanly.

Why not just sum functional copies into one gene-level dosage and fit it additively, or with a saturating curve? If compensation between copies is purely dosage-driven, a per-group curve of total functional copies could explain the same variance with fewer assumptions. This is also a real threat, and it is treated as a nested model rather than argued away. A saturating dosage model is itself a specific form of homoeolog epistasis. If it matches $K_H$, the conclusion becomes "the missing variance is dosage saturation across homoeologs," which is still a positive finding. If $K_H$ beats it, particularly in its expression-weighted form, copies are not interchangeable, which is what the TaNRX1-2D and phosphate transporter results suggest.

Why not just use i-LDSC on summary statistics? Its interaction scores cover ±50-SNP windows (i-LDSC), and homoeologs lie on different chromosomes. It is structurally blind to the hypothesized signal. It remains useful as a check that local non-additive variance is separate from $\sigma_H^2$.

Proposed Experiments

Simulation comes first, because the file contains no public phenotyped panel for an allopolyploid. Genotypes will be simulated on real homoeolog structure, taken from the near-telomere-to-telomere hexaploid assemblies of Thinopyrum intermedium and Roegneria kamoji (Sun et al.) and wheat pangene groups. Sequencing depth will be varied from low-depth GBS levels upward. Two decisive checks run here. The first is type I error: with no epistasis simulated but realistic subgenome structure and low depth, $\sigma_H^2$ must not exceed the random-pair null. The check is run with and without the Bilton correction, to show that the correction is necessary. The second is power: when homoeolog epistasis explains 2–10% of variance, the method is compared with the full Hadamard model, SHEsisPlus and MAPIT.

Real-data evaluation requires a phenotyped, resequenced diversity panel of wheat, rapeseed or cotton. None appears in the file, so one will have to be obtained. The pre-registered prediction is that for yield-component and stress traits, homoeolog pairs with both copies expressed carry most of $\sigma_H^2$, and pairs with one copy silenced carry near zero. The 141-accession tetraploid potato panel of Gong et al. is an autopolyploid, so its copies are homologs rather than homoeologs, and it is small. It serves only as a secondary test of within-locus dosage interaction, which its additive GWAS did not model. No reported heritability number was found on Lacuna for that panel.

The work fails if $\sigma_H^2$ is not enriched over matched random pairs, or if the saturating dosage model explains it completely without the expression weighting mattering.