Researchers Apply LLMs To Optimize Codons

MIT chemical engineers used an encoder-decoder large language model to analyze codon usage in the yeast Komagataella phaffii and generate optimized DNA sequences, reporting results in the Proceedings of the National Academy of Sciences this week. The model outperformed four commercial codon-optimization tools across six proteins, improving production for five and ranking second for the sixth. The approach could reduce development time and costs for biologics manufactured in yeast.