To understand why this matters, it helps to revisit one of the great embarrassments of the genomics era. When genome-wide association studies, or GWAS, exploded in popularity during the late 2000s, researchers expected to find the genetic roots of height, disease susceptibility, crop yield, and countless other complex traits laid out neatly across the genome. Instead, the identified variants explained only a fraction of the heritability that family and population studies had long established. The remainder, dubbed missing heritability, has been attributed to many sources, including rare variants, structural differences, gene-environment interplay, and, crucially, epistasis, the phenomenon in which the effect of one genetic variant depends on the presence or state of another.

Epistasis is notoriously difficult to detect. The number of possible pairwise interactions among millions of single nucleotide polymorphisms, or SNPs, grows quadratically with genome size, producing a combinatorial search space so vast that exhaustive testing becomes computationally prohibitive. Worse, interactions are not limited to pairs of neighboring variants. Some epistatic effects are local, arising from variants sitting close together on a chromosome, while others are global, linking loci on entirely different chromosomes. Traditional methods, which typically scan for one interaction at a time using regression-based tests, tend to miss these long-range dependencies and often fail to distinguish interaction effects from simple additive contributions of individual variants.

The Shandong-led team approached the problem from an angle that has transformed natural language processing over the past decade: representation learning. Their starting point is Evo 2, a genome language model trained on enormous quantities of DNA sequence data. Just as large language models learn the statistical structure of human text by predicting the next word in a sentence, genome language models learn the statistical grammar of DNA by processing nucleotide sequences at scale. The key advantage is that such models can capture long-range dependencies, meaning they can encode relationships between DNA elements that lie far apart in the linear sequence, exactly the kind of distant interactions that underpin global epistasis.

Epiformer uses Evo 2 to build rich contextual representations of the genetic background surrounding candidate SNPs. Rather than treating each variant as an isolated letter in a sequence, the framework embeds each SNP within the learned context of its broader genomic neighborhood, allowing information about distant regulatory elements, linkage patterns, and sequence motifs to inform how a variant is represented. This is a fundamental departure from classical epistasis detection pipelines, which typically begin by filtering variants through marginal association tests and therefore risk discarding variants whose effects only manifest in combination with others.

On top of these language-model representations sits the second major innovation: a dual-channel network. The architecture processes genetic data through two parallel pathways, one dedicated to modeling local epistatic patterns and the other to capturing global interaction effects and additive effects. By running these channels jointly rather than sequentially, the network can learn how local and global contributions combine to shape the phenotype. The design reflects a biological reality: complex trait architecture is rarely the product of a single type of genetic effect, and forcing a model to choose between local and global perspectives inevitably discards part of the picture.

What elevates Epiformer beyond a brute-force pattern detector is its emphasis on interpretability. The framework does not merely output phenotype predictions; it can identify which SNPs it considers key contributors and which interactions between them appear to matter. This interpretability reinforces the epistasis detection process itself, because researchers can inspect the highlighted variants and interactions, assess whether they align with known biology, and generate testable hypotheses about the mechanisms linking genetic variation to observable traits. In a field where black-box predictions have often been met with justified skepticism, the ability to trace a prediction back to specific loci is a meaningful step toward scientific usability.

The authors report that Epiformer performs robustly across species, an important claim for a method intended to serve both biomedical and agricultural genetics. The study’s acknowledgments note assistance from the State Key Laboratory of Maize Bio-Breeding at China Agricultural University with maize data analysis, signaling that crop genomes were among the evaluation targets. Robustness across species matters because the genetic architecture of a maize yield trait differs substantially from that of a human disease, and a detection method that generalizes across such diverse genomes is far more valuable than one tuned to a single organism. According to the paper, the method reveals biologically meaningful patterns and offers new insights into genetic architecture, suggesting that the learned representations are not statistical artifacts but reflect genuine organizational features of genomes.

The technical foundations of the work sit at the intersection of two rapidly advancing fields. Evo 2 represents the current generation of foundation models for genomics, systems pretrained on vast DNA corpora that can be adapted to downstream tasks ranging from variant effect prediction to regulatory element discovery. Transformer architectures, the backbone of such models, use attention mechanisms that weigh the relevance of every position in a sequence to every other position, making them naturally suited to capturing the kind of long-range dependencies that epistasis detection demands. Epiformer’s contribution is to couple this sequence-level understanding with a task-specific network that explicitly separates and then integrates local and global genetic effects, a design choice tailored to the particular structure of the epistasis problem.

The implications extend well beyond the immediate technical achievement. For human genetics, better epistasis detection could sharpen polygenic risk scores, the tools used to estimate an individual’s genetic predisposition to diseases, by incorporating interaction terms that current models largely ignore. For plant and animal breeding, identifying interacting loci that influence yield, drought tolerance, or disease resistance could guide more precise selection strategies, particularly in crops like maize where heterosis and complex trait architecture are known to involve extensive non-additive genetic effects. And for evolutionary biology, systematic catalogs of epistatic interactions could clarify how genetic networks constrain and enable adaptation.

Cautious optimism is warranted. Epistasis detection methods have a long history of promising results that proved difficult to replicate across cohorts, largely because interaction effects are sensitive to population structure, sample size, and environmental context. The true test of Epiformer will be whether its discoveries hold up in independent datasets and whether the variants and interactions it highlights can be validated experimentally. The published version of the paper is accompanied by extensive supplementary materials, including dataset summaries, simulation settings, extended evaluations, and case studies, which should help the community scrutinize and build upon the work. The article is open access, lowering the barrier for research groups worldwide to test the method on their own data.

Nevertheless, the study marks a conceptual shift worth taking seriously. For two decades, the dominant strategy for finding genetic associations has been to test variants one at a time and hope that the aggregate of small additive effects would eventually account for missing heritability. Epiformer embodies a different philosophy: that the genome should be read as an integrated system, with context-rich representations and architectures designed from the outset to capture interactions at multiple scales. If genome language models continue to improve and computational resources continue to expand, the era in which epistasis was an intractable afterthought of genetic analysis may be drawing to a close, replaced by one in which the conversation between genes, near and far, becomes a central object of study.

Subject of Research: Detection of epistatic gene interactions using a genome language model and dual-channel deep learning network

Article Title: Epiformer: epistasis detection by genome language model and dual-channel network

Article References: Zhang, X., Liu, L., Ren, L., Xin, B., Guo, M., Wang, J., & Yu, G. (2026). Epiformer: epistasis detection by genome language model and dual-channel network. Genome Biology. https://doi.org/10.1186/s13059-026-04268-8

Image Credits: AI Generated

DOI: 10.1186/s13059-026-04268-8

Keywords: epistasis, genome language model, Evo 2, dual-channel network, missing heritability, SNPs, phenotype prediction, transformer, GWAS, complex traits, maize, Genome Biology