New AI model could cut the costs of developing protein drugs

Artificial intelligence has just learned to speak the hidden language of yeast DNA, and it's revolutionizing how we make life-saving drugs. MIT chemical engineers trained a large language model to decode the genetic preferences of industrial yeast, the microscopic workhorses that produce billions of dollars worth of vaccines and protein medicines each year.

The breakthrough centers on optimizing "codons" — three-letter DNA sequences that cells use like a recipe book to build proteins. While most amino acids can be encoded by multiple codons, different organisms have distinct preferences for which ones to use.

The AI model analyzed all 5,000 proteins naturally made by the yeast Komagataella phaffii, learning not just individual codon choices but how they interact with neighboring sequences. When tested against six different proteins, including human growth hormone and a cancer-fighting antibody, the AI-designed sequences outperformed commercial optimization tools in five cases.

"Having predictive tools that consistently work well is really important to help shorten the time from having an idea to getting it into production," says MIT's J. Christopher Love. The approach could slash development costs for new biologic drugs, where manufacturing optimization currently accounts for up to 20 percent of commercialization expenses.