AI BriefingCodonRoBERTa, an mRNA Model Supporting 25 Species, Released
Key point
An end-to-end protein AI pipeline spanning protein structure prediction to codon optimization was unveiled, along with the new model CodonRoBERTa.
Details
OpenMed built an end-to-end protein AI pipeline integrating protein structure prediction, sequence design, and codon optimization.
As a key achievement, they introduced CodonRoBERTa-large-v2, a new transformer architecture for mRNA optimization. This model recorded superior performance compared to ModernBERT (Perplexity 4.10, Spearman CAI correlation 0.40), demonstrating outstanding capability in codon-level language modeling.
The pipeline's main stages are as follows:
- Protein Folding: Uses ESMFold v1 to predict the 3D structure of proteins.
- Sequence Design: Designs amino acid sequences matching the structure through ProteinMPNN.
- mRNA Optimization: Optimizes DNA codons so the designed sequence is efficiently expressed in the target organism.
The research team trained on 381,000 sequences across 25 species, successfully building a Multi-species system capable of species-specific optimization in just 55 GPU-hours (about $165).
This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.
Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.
Related
AI Can Now Design Functional Viruses: Should We Be Worried?