When pharmaceutical companies want to manufacture protein-based drugs - insulin, monoclonal antibodies, growth hormones - they typically use engineered yeast or bacteria as tiny biological factories. The challenge: optimizing the genetic instructions so these microorganisms produce maximum quantities of the target protein.
MIT researchers just demonstrated that a large language model trained on yeast genetics can solve this problem better than existing commercial tools.
What the AI Actually Does
The research, published February 16 in Proceedings of the National Academy of Sciences, focuses on something called codon optimization. DNA is read in three-letter sequences called codons, and multiple codons can encode the same amino acid. Choosing which codon to use matters for how efficiently cells manufacture proteins.
The MIT team trained an encoder-decoder LLM on approximately 5,000 proteins naturally produced by Komagataella phaffii, an industrial yeast commonly used for drug manufacturing. The model learned not just which codons the yeast prefers, but how they’re arranged relative to each other - both neighboring codons and long-distance patterns across the genetic sequence.
“The model takes into account how codons are placed next to each other, and also the long-distance relationships between them,” said J. Christopher Love, the study’s senior author and Raymond A. and Helen E. St. Laurent Professor of Chemical Engineering at MIT.
The Results
The team tested their system, called Pichia-CLM, on six different proteins including human growth hormone, human serum albumin, and trastuzumab (a monoclonal antibody used in cancer treatment).
When they inserted the optimized sequences into actual yeast cells and measured protein production, the AI-designed sequences outperformed four commercially available codon optimization tools in five out of six cases. For the sixth protein, it came in second.
This isn’t a theoretical improvement. The researchers grew actual yeast cultures and quantified real protein yields.
Why This Matters for Drug Manufacturing
Protein drugs are expensive to develop partly because of the trial-and-error involved in optimizing production. Companies design genetic sequences, test them, redesign, and repeat. Each iteration costs time and money.
The MIT approach reduces this uncertainty. If an LLM can reliably predict which codon arrangements will work best, pharmaceutical companies can skip failed experiments and move faster toward production.
Manufacturing costs represent 15-20% of the total expense in commercializing a protein drug. Better optimization upfront translates directly to lower costs.
The Fine Print
The model was trained specifically on Komagataella phaffii yeast. Different organisms have different codon preferences, so the system would need retraining for other manufacturing platforms.
Also worth noting: the AI beat commercial tools in five of six tests - not all six. One protein still had better results with existing methods. The researchers didn’t dig into why, though they speculate it relates to the specific sequence characteristics of that particular protein.
Finally, this is lab-scale validation. The jump from optimized sequences in research to full industrial manufacturing involves additional engineering challenges the paper doesn’t address.
Still, as AI applications in biotech go, this is unusually practical. It solves a real bottleneck with measurable results against existing tools. The kind of unglamorous but useful work that actually makes drug development cheaper.