OmniGene-4: A Unified Bio-Language MoE Model with Router-Level Interpretability.

MoE (Mixture-of-Experts) model jointly processing natural language, DNA sequences, and protein sequences with router-level interpretability

OmniGene-4 is a multimodal Mixture-of-Experts (MoE) bio-language model jointly processing natural language, DNA sequences, and protein sequences to answer sequence-grounded biological questions. Router-level interpretability analysis reveals that each expert specializes on distinct question types (structure, function, annotation), providing a window into the model's internal mechanisms. Performance on mixed genomic benchmarks surpasses that of specialized unimodal models. The model is available open-source via Hugging Face.

Synthesis written by Geno'X. For the full original abstract, please refer to the source publication.

Multimodal MoE models are a promising path for genomic AI: they allow unifying heterogeneous representations (sequence, annotation, phenotype) in a common space. Router-level interpretability is an original methodological contribution, but performance on real clinical benchmarks (pathogenic variants, diagnosis) remains to be demonstrated.

Clinical impact: 2/3 · Evidence strength: 2/3 · Novelty: 2/2 · Sample size: 1/1 · Publication status: 0/1 → Total: 7/10