HomoeoGWAS searches for interactions among crops’ inherited gene copies

Many major crops are allopolyploids: they combine distinct ancestral genomes, called subgenomes, within one nucleus. Consequently, a gene may have related copies in different subgenomes. These copies, known as homoeologs, provide a natural set of candidates for studying genetic interaction—whether the joint association of two copies with a trait departs from their separate contributions.

In a bioRxiv preprint posted on November 9, 2026, Cheng Si and colleagues describe HomoeoGWAS, software that makes the homoeolog group the unit of association testing. The work has not been certified by peer review. Its findings identify statistical candidates for follow-up, not experimentally established interactions between causal genes.

A biologically focused search

An unrestricted interaction scan can become enormous. For the 627,255 DNA variants mapped to the rapeseed groups in this study, testing every variant pair would involve about 196 billion comparisons. HomoeoGWAS instead tests predefined groups containing related copies from different subgenomes.

The researchers analysed 827 wheat accessions, 1,245 cotton accessions, 926 rapeseed accessions and 230 peanut accessions. Their eligible testing sets contained 2,143 wheat three-copy groups, 3,757 cotton pairs, 17,404 rapeseed pairs and 379 peanut pairs. These were filtered sets, not every homoeolog in each genome.

For each pair of copies within a group, the method examines three representations of genetic variation. One summarizes minor-allele content, another uses the strongest principal-component pattern within each gene block, and a third captures combinations across several such patterns. Their statistical evidence is combined into one group-level result. A three-copy group contributes three pairwise comparisons; the main test does not estimate a three-way interaction coefficient.

The analysis accounts for genetic relatedness and tests each interaction conditional on the corresponding copies’ main effects. It then uses 2,000 simulated null datasets to calibrate significance across the entire declared testing family. The primary bootstrap model allows sample-specific variance to vary smoothly with leading axes of genetic population structure.

Where calibration held—and where it did not

That variance adjustment mattered. An earlier model assuming uniform variance produced excess false rejections in wheat and cotton when simulated variance followed population structure. Across five noise models in 80-group tests, the revised model’s observed family-wise false-rejection rates ranged from 2.1% to 5.2%.

In separate complete-family experiments, simulated traits combined a polygenic background with structure-aligned residual variance. The proportion of datasets with any significant group was 1.4% in wheat, 3.1% in rapeseed, 1.1% in cotton and 3.1% in peanut, against a nominal 5% target.

These results do not establish universal error control. Strong additive effects concentrated in one homoeolog group could inflate false rejections in control experiments. The authors also did not establish strong family-wise error control when some groups genuinely carry interactions. Nearby linked variation or an interacting gene outside the nominated group can contribute to a detection.

Trait signals with different levels of support

The application scans identified four wheat groups associated with heading date, two rapeseed groups associated with flowering time and three cotton groups associated with fibre length. Together, these nine groups represented eight loci. No significant group emerged for cotton fibre elongation or the two peanut seed traits; the fibre-elongation analysis was flagged for calibration review.

A post-selection robustness analysis adjusted for regional genetic structure around the copies. Five groups at four loci retained support: two wheat groups, the two rapeseed pairs at one locus, and one cotton pair. One other wheat group remained uncertain, while three groups were treated as regional candidates rather than robust detections.

The strongest wheat signal includes a candidate ubiquitin-specific protease supported by evolutionary and protein-domain annotation. That makes it a follow-up target, not a demonstrated explanation for heading-date differences. Independent-population replication and functional experiments remain necessary.

Simulations also showed why the group test complements, rather than replaces, variant-pair scanning. It performed better when interaction signals were distributed across several genetic dimensions or pairwise edges. A SNP-by-SNP comparator was more powerful in two conditions where a single variant pair carried the signal.

Speed without causal certainty

Complete interaction analyses took 3.4–7.3 minutes on the benchmark hardware using 128 workers and 2,000 bootstrap replicates. These are configuration-specific timings; comparator programs tested different hypotheses and used different settings.

A separate rapeseed analysis found a weak association between homoeolog genotype coupling and reported chromosome-exchange frequency, with a rank correlation of 0.136 across 3,038 mapped pairs. It did not usefully classify individual exchange carriers, and the exchange calls and genotypes shared sequencing data.

The framework requires divergent subgenomes and established homoeolog maps; autopolyploids and mixed-ploidy panels are outside its scope. This account uses the supplied PDF text extraction and metadata. Figure images and supplementary spreadsheet contents were not independently inspected, and extraction can distort equations and layout.