OpenMed trained mRNA models across 25 species in 55 GPU-hours for $165

OpenMed didn’t just survey the protein AI landscape in their first post — they got their hands dirty and built the pipeline. The result is an end-to-end system that takes a therapeutic protein concept and spits out a synthesis-ready, codon-optimized DNA sequence. The real muscle, and the part they built from scratch, is the codon optimization engine.

After running a horse race between transformer architectures, the team found a clear winner. They pitted a scaled-up RoBERTa model against the newer, flashier ModernBERT. For this specific biological task, new didn’t mean better. CodonRoBERTa-large-v2 hit a perplexity of 4.10 and a Spearman CAI correlation of 0.40, leaving ModernBERT in the dust. This matters because getting codon choice right can mean a 100x difference in protein expression — it’s a key reason the Pfizer-BioNTech COVID vaccine worked so well.

The decision to back RoBERTa wasn’t random. As the team points out, Meta’s ESM-2 protein language model is itself a RoBERTa variant. The hunch that an architecture good at learning amino acid patterns would also nail the triplet-based codon code paid off. It’s a smart, practical choice that prioritizes a proven workhorse over chasing the latest NLP trend.

They then scaled with brutal efficiency. Training four production-grade models across a menagerie of 25 different species took just 55 GPU-hours and cost a total of $165. The end result is a species-conditioned system for codon optimization that no other open-source project offers. It’s a transparent, no-nonsense build log that shows exactly what it takes to bridge the gap from a protein idea to a DNA sequence you could actually synthesize and test in a lab.