also faster to hit cluster center of bowls with extended similarity..

was standing on the other side of fig. 1:)

Similarly in predicted folding pathways, the C-termini hairpin of protein G doesn’t change structurally, but the confidence scores vary when surroundings fold closer to native.

I think big part of what AF learned is mimicking the distance distribution of arbitrary pairwise residues/tokens in input from PDB. It changes with environment, adding more sequences affects structure predictions, and scores, which correlate well with the variance in predicted distograms.

It was a surprise to see AF2 can predict folding pathways, but the scores keep increasing from unfolded to folded, unlike (real!) free energy map.

That would be great! But I haven’t found strong evidence for that.. At least the confidence score doesn’t look like a good estimation for free energy from this work. www.biorxiv.org/content/10.1...