[QUEST] Inpainting residues with RFD3 only gives glycines

[CamillaLeogrande]

I run the example in "models/rfd3/docs/tutorials/enzyme_tutorial_files/rfd3_enzyme_tutorial.json" and I noticed that the inpainted residues (in this case "select_unfixed_sequence": "A579,A581") are always output as glycine by RFD3. Is there a way to change this behaviour so that the whole sequence space is explored?

[roccomoretti]

RFdiffusion is a backbone-only generation method. While RFD3 does have some sequence conditioning, any sequence it would generate is not really to be trusted. You should only pay attention to the backbone coordinates in the output, not the protein sequence. Instead, you want to use a dedicated sequence design program like LigandMPNN to place the sequence on the generated backbone.

In the case of select_unfixed_sequence, you're explicitly asking RFD3 to ignore the sequence at those locations. Any sequence information is discarded on structure read in, and those positions are represented internally as a sequence-free "generic protein residue" backbone. So when RFD3 outputs the result of the design process, it no longer has any sequence information for those positions, just the backbone coordinates. As such, it outputs the residues as GLY -- not because it thinks glycine is an appropriate residue at that position (or that it modeled things as glycine internally), but simply because it's the best match to the "generic (sequence free) backbone-only protein residue" internal representation it did use.

[CamillaLeogrande]

Thanks for your comment!
I had assumed that if RFD3 places glycines, it tends to bring opposing backbones so close together that a later MPNN design step would effectively be forced to keep glycines at those positions, since larger side chains would create clashes. Is that not actually the case?

[roccomoretti]

As mentioned, RFD3 internally has a concept of amino acids which is sequence agnostic. This "generic amino acid" is one of the primary ways it was trained to recapitulate the backbones in the PDB -- backbones with a variety of amino acids. When it generates new backbones, it uses that generic amino acid representation. (This was completely true for the original RFdiffusion. Later versions have incorporated some sequence information on generated residues, but it's still appropriate to think about it as a "generic" amino acid, versus a specific one.) Despite being spit out as backbone only glycines, the backbones are positioned based on what it's learned from the PDB, including the full sequence diversity. There shouldn't be issues with fitting non-glycine residues into those positions.

That said, people have seen a bunch of success with iterative methods. That is, take the MPNN designed sequences, put them through a prediction program (AlphaFold, OpenFold, etc.) and then use the predicted backbones to re-run the sequence design. Do this for several cycles, filtering at each stage for various properties (prediction confidence, rmsd to the original design, catalytic geometry, etc.) -- Even if the original backbone may have been a bit tight in the originally generated structure, repredicting the backbone conformation in the context of the designed sequence should open it up.

[CamillaLeogrande]

OK, this is reassuring. Thank you!