[dekhn]

This looks like it was written by generative AI but I can't really say for sure.

BTW: protein structure prediction didn't need supercomputers (in the traditional sense) and the PSP problem wasn't solved using supercomputers applying a high quality physics function to simulate folding- instead, it was solved using a combination of ML supercomputers, a really good algorithm (transformers), and a couple of really good data sets- the known structures of proteins, and the known relationship of proteins.

Instead of simulation on a huge supercomputer so they could predict a single strucfture, they trained a model which approximates structure well enough to beat every competitor. From what I can tell, most of the resulting quality doesn't come from their force field but from the distance constraints that are mostly derived from historical relationships between proteins, and the coevolution of their sequences.