This work seems like a big advance in enzyme design, but it’s still fascinating that we can design new proteins that fold with ~90% accuracy, but even with careful biochemical design, DFT, and heaps of machine learning you only get about 1/100 designs with the activity level of natural enzymes!
Kevin K. Yang 楊凱筌@kevinkaichuang.bsky.socialThe Baker lab seems serious about moving into enzyme design.
Here, they train a flow-matching model that can be conditioned by atomic coordinates and then use it to design enzymes based on active site geometry specified by density functional theory.
www.biorxiv.org/content/10.1...