Cross-species functional variant classification using the Evo 2 genomic language model: implications for farm animal genetics and breeding

Background Genomic language models offer sequence-based approaches that complement traditional population genetics methods in genomic analysis. In farm animal genetics, extensive linkage disequilibrium often complicates causal variant identification in population studies, creating opportunities for sequence-based models to provide orthogonal functional information for variant prioritization applications. Methods We evaluated Evo 2's functional variant classification capability using 721 high-quality single nucleotide variants from the Online Mendelian Inheritance in Animals (OMIA) database across eight species: chicken, dog, domestic cat, goat, horse, pig, sheep, and cattle. Each variant was analyzed within 8,192 bp sequences using zero-shot classification (log-likelihood scores) and supervised linear probing (sequence embeddings). Control sets employed variant-type blind and variant-type matching strategies to evaluate different aspects of classification performance. Results Zero-shot classification achieved strong performance with variant-type blind controls (AUROC = 0.934, AUPRC = 0.832) but reduced performance with variant-type matching controls (AUROC = 0.717, AUPRC = 0.166). Linear probing demonstrated consistent cross-species performance, with species-blind cross-validation yielding AUROC = 0.921 (variant-type blind) and AUROC = 0.801 (variant-type matching). Performance remained robust across evolutionary distances, with individual species AUROC ranging from 0.844 to 0.931 in variant-type blind analysis. Conclusions Evo 2 effectively classifies functional variants across diverse animal species, providing sequence-based information complementary to population genetics methods. The model's cross-species generalizability and ability to distinguish functional variants through sequence context establishes a foundation for applying genomic language models in farm animal genetics, with applications spanning fine-mapping, mutation load assessment, and genomic prediction enhancement.