What constitutes a "low quality" backbone? about proteinmpnn

[FAOlivieri]

I am having the exact same issue. Even parting from structures from different proteins (a coiled-coil dimer and a helix tetramer) I get sequences that are near 30% K or E.

[MattMcPartlon]

@bwllc I know this is a little late, but I've seen this happen when intra-residue geometry is not ideal (my own method, AttnPacker, does this). It's my guess that conserved bond lengths and angles are off.

If you have access to Rosetta, you can run relax with coordinate constraints to fix the geometry while minimizing the RMSD between pre-relaxed and relaxed structures. There is also the [Idealize protocol] (https://www.rosettacommons.org/docs/latest/scripting_documentation/RosettaScripts/Movers/movers_pages/IdealizeMover) which is designed for exactly this, but I haven't tried it.

As a first step, you can try running inference with the v_48_020.pt model first. If the distribution of AA types looks better, then that's a good indication that this is your issue.

GL

[bwllc]

Thanks for your reply, @MattMcPartlon.

I think that you are saying that a computational dynamics, force-field relaxation step sometimes needs to be applied to the output of RFDiffusion before passing it to ProteinMPNN. Do I understand that correctly?

If Rosetta has been open-sourced, I can use its minimizer.

I already have GROMACS, and it also has a relaxation algorithm which I can investigate. I'm not sure if it behaves differently than the Rosetta minimizer. I'm not sure whether that would matter. I could probably specify constraints on some atoms in GROMACS, but that sounds fussy, and I'd prefer to avoid that if I can.

Please let me know if I'm barking up the wrong tree. Thanks.

[MattMcPartlon]

@bwllc That's exactly what I mean :).

Before spending too much time on this, you can check (for example) that the consecutive C-alpha atoms are at distance 3.8A +/- 0.1. If you see distances outside of this range, then relaxing with a forcefield should solve your problem.

I only recommend rosetta's minimizer because it can explicitly minimize RMSD between relaxed and input structures.
GROMACS or AMBER should also work fine. Good luck!