Deep Learning-Enabled Detection of Regulatory Variants from Single-Cell Chromatin

Although genomic prediction has improved markedly in cattle, buffalo, sheep, and poultry, a substantial proportion of additive genetic variance remains unexplained, prompting continued debate on where “hidden heritability” resides. Since many livestock traits appear to be shaped by myriad regulatory variants with context-dependent effects, one might reasonably consider whether bulk-tissue datasets are masking cell-type-specific regulatory architecture. As recent breakthroughs in single-cell chromatin profiling illustrate, regulatory activity varies sharply across epithelial, immune, and stromal compartments, yet livestock breeding programs have scarcely incorporated this information.

This study employed single-cell ATAC-seq datasets newly available for bovine and ovine tissues to map cell-resolved open-chromatin peaks. Variants overlapping these peaks were then interpreted using a hybrid convolutional transformer deep learning model inspired by Enformer, enabling prediction of enhancer disruption and directionality of regulatory effects. Because deep models can integrate long-range chromatin interactions, they provide an opportunity to observe regulatory dependencies that conventional annotations often overlook. In comparison to baseline SNP models, cell-type-specific regulatory scores increased prediction accuracy across growth, mastitis resistance, and fertility traits by 10–18%. Immune-cell-specific regulatory polymorphisms expressed disproportionately significant effects on parasite-resilience traits and mastitis, supporting previous multi-tissue eQTL studies. Despite the limited tissue diversity and sample number, these advancements may represent the early promise of single-cell functional genomics.

The combination of deep learning and single-cell chromatin landscapes does, however, seem a rational and physiologically sound way to deal with latent heredity in cattle. This approach prioritizes variants with actual regulatory relevance, strengthening genomic selection processes and mechanistic interpretation.