Anyway, if anyone in academia has some spare time & compute, I'd love to see if the MD-trained NNs are any better than the vanilla models at multi-state prediction, as it wasn't tested in the paper (model weights have a non-commercial license so they're off-limits to me)

Yeah somewhere in the supplement

Oh awesome, do you have any ablations showing the contribution of this to refoldability? Chroma showed nearly 2x improvement with the random edges vs standard nearest neighbors

Another one - PDB 8F2X chain H (3.5 Å resolution, shown in gold). Totally different authors. Seems like aromatics immediately before the conserved J-gene tryptophan throws off assignment in this area; 7czv has a phenylalanine before the tryptophan

The worst offender I've found so far is PDB 5OMM chain C on Sabdab, which has many unassigned residues despite being 1.7 Å resolution. For example its C-terminal VTVSS starts at IMGT pos 115 instead of pos 124. Probably AbNum/ANARCI struggling w/ the gaps

opig.stats.ox.ac.uk/webapps/sabd...

It's a very cool result but IMO there are caveats. Inference is (mostly) slower. There is existing work on faster models (e.g. MiniFold or protenix mini), and also existing work on ensemble prediction. I doubt this works without training on AFDB, which bakes in inductive bias from triangle layers