Training mRNA Language Models Across 25 Species for $165

An end-to-end protein AI pipeline was developed for structure prediction, sequence design, and codon optimization. After comparing multiple transformer architectures for codon-level language modeling, a model named CodonRoBERTa-large-v2 was identified as the best performer. This model significantly outperformed others on metrics like perplexity and Codon Adaptation Index (CAI) correlation. The project then scaled this approach to 25 different species, creating a suite of four production models trained in just 55 GPU-hours to enable efficient protein expression in various organisms.