It's really cool to see some of these recent papers from @davidrliu lab and Doudna lab using generative models for protein sequence design (a.k.a. "inverse-folding" AI models) in combination with evolutionary insights and techniques. In the Liu lab paper, nature.com/articles/s41586-0… they used proteinMPNN (my beloved) to improve important properties like thermostability and expression, and enhance enzymatic activities for proteases. Critically, they redesigned regions were not allowed to touch the active site so the mutations are expected to improve the structural stability of the overall fold. They did this to "polish" evolutionary starting points and also to seed laboratory evolution trajectories. Both approaches improved nontrivially on what would have been sampled through evolution alone. In the Doudna lab paper, science.org/doi/10.1126/scie… they used ESM-IF1 to make a synthetic Cas12 variant with improved activity compared to the native. Critically, they avoid mutating the most evolutionarily conserved positions. They suggest the structure structure aware ESM-IF is able to explore much more diverse areas of sequence space than standard sequence-only pLMs. In both of these cases, generative models for protein sequence design added genuine value when combined with expert scientific intuition on which problems to apply them to and how to configure ("prompt") them. These papers coming out is kind of a full circle moment for me; I worked on this older paper, cell.com/iscience/fulltext/S… with Doudna on Cas12a which is the same Cas family they chose to make a synthetic version of, right before I went to @UWproteindesign and got to help with proteinMPNN. Congrats to the authors @NicholasKrasnow et al. and @PetrSkopintsev @isabelesain @evandeturk et al. #ProteinDesign #CRISPR #ArtificialIntelligence