OpenMed trains multi-species mRNA codon models (CodonRoBERTa-large-v2) for $165

AI Impact Summary

OpenMed demonstrates an end-to-end codon optimization pipeline using CodonRoBERTa-large-v2 to achieve perplexity 4.10 and CAI 0.40 across 25 species, indicating robust cross-species codon preference capture. The team trained 4 production models in 55 GPU-hours on 250k coding sequences, suggesting a cost-efficient path to deployment of mRNA design tooling. This enables integration into protein design-to-expression workflows, potentially accelerating therapeutic development, but requires rigorous biological validation and regulatory compliance before production use.