I recently saw this paper by King et al. about using AI trained on DNA sequences to design phage viruses, some of which successfully infected bacteria. This particular application is scary/exciting (like lots of AI advances) because of its potential applications to biological warfare and for designing phage to attack pathogens. I’ve been particularly impressed by research by Paul Turner and colleagues on tradeoffs that (sometimes) make bacteria that evolve resistance to phage (e.g. by mutating antibiotic-efflux pumps to which phage bind) more susceptible to antibiotics.

The King et al. paper referred to this paper on Evo 2. Rather than training AI on text (much of it stolen from various sources — even though I use Claude a lot, I’m looking forward to getting some compensation from the lawsuit against Anthropic, which used my book and many other copyright materials to train Claude), they trained on DNA and RNA sequences from a wide range of plants, animals, fungi, and bacteria. Evo 2 can apparently do all sorts of amazing things, like predicting fitness effects of mutations either in protein-coding or noncoding genes. This video is interesting. I didn’t understand all of it, but Evo 2 seems like a game changer.