improving esm?
Proteins are transcribed head-to-tail, so wouldn't the most accurate overall folding predictions incorporate that process into the predictive process?
Any protein is made in step-wise fashion and perforce the head is folding even as the rest of the mRNA is being transcribed. THAT folding head (in the absence of of any influence from the tail) would certainly be compliant in different ways than the sequence taken as a whole. My understanding is that while biological folding is sequential and context-dependent, ESMFold simplifies this by using a static model where the whole sequence is processed simultaneously to predict the final folded structure. It seems plausible (given how accurate the models are these days), that the situation above might address the final angstrom of accuracy. Maybe taking the proposed final sequence and establishing the protein's structure one amino acid at a time as it would exit a ribosome would acquire a more accurate prediction, in the end.