AI Learns to Cut Proteins That Nature Leaves Alone for Centuries

Researchers at the Institute for Protein Design at the University of Washington have learned how to cut a protein at a chosen spot. The bonds between amino acids in nature are so strong that they can stay intact in plain water for hundreds of years. The team designed the enzymes using computer models, then tested the resulting candidates in the lab. Out of 135 zinc-based enzymes, 36% worked. Out of 69 cysteine-based enzymes, 19% worked. The targets included proteins linked to Alzheimer's disease and ALS, a severe disease of the nervous system. The results were published in the journal Nature on October 14.

What stands out here is not the biological breakthrough itself, but the success rate. Even in a task as precise as protein design, less than half of the candidates proposed by the computer actually worked when tested. I see the same pattern with code written by AI agents. The model produces many options, and it is the lab test, or the code review, that decides which ones are any good. There is not much to celebrate yet. This is a tool for scientists, not a finished drug.