AI IDs Immune Proteins
- What happened: An AI tool rapidly identified hundreds of immune‑related proteins, accelerating discovery. - The key specific: MIT’s DefensePredictor reportedly identified more than 600 immune proteins in minutes. - Context: The work promises faster protein discovery pipelines, with broader estimates projecting millions of possible immune proteins (x.com).
Bacteria fend off virus attacks with proteins, and an MIT-led artificial intelligence system found 624 likely defense proteins in 69 *E. coli* strains. (science.org) The tool, called DefensePredictor, was reported in *Science* on April 2, 2026 by Peter C. DeWeirdt, Emily M. Mahoney, and Michael T. Laub of the Massachusetts Institute of Technology. (science.org) The model was trained on about 17,000 prokaryotic reference genomes and used protein representations from Evolutionary Scale Model 2, or ESM2, plus information from each gene’s four nearest neighbors. (science.org; pypi.org) In plain terms, the system works like a spam filter for proteins: it learns what known antiviral proteins look like, then scans genomes for similar patterns even when the genes are not sitting in the usual clusters. (science.org; phys.org) That matters because the standard search method has leaned on “defense islands,” stretches of DNA where immune genes tend to bunch together. The *Science* paper says many real defense systems sit elsewhere, including on plasmids, prophages and transposons. (science.org) Among the 624 high-confidence calls in *E. coli*, more than 100 had no detectable similarity to previously known defense proteins. Nearly half were outside defense islands, which is the main gap the model was built to close. (science.org) The team then moved from prediction to lab tests. They cloned 94 predicted systems into a susceptible *E. coli* strain and found that 42 protected the bacteria against at least one of 24 bacteriophages, the viruses that infect bacteria. (science.org; phys.org) Bacterial immune systems already underpin major lab tools, including CRISPR-Cas and restriction enzymes. The paper frames the new search as a way to find more systems with that same mix of biological and practical value. (science.org) When the researchers expanded the search beyond *E. coli* to 1,000 diverse prokaryotic genomes, they found more than 5,000 predicted defense proteins that were not clear homologs of known ones. (science.org) Nature described the new studies as databases containing thousands of antiviral defense proteins that could inform future molecular tools. The immediate result is narrower and more concrete: a short list of candidates that can be tested in the lab instead of searched for one by one. (nature.com; phys.org)
