Possibility of adding residue restraints

[wjs20]

I tried running GeoDock with a VH-VL antibody complex + its protein antigen. GeoDock succesfully generated a docked structure, but it looks like the antigen is bound on the opposite side of the VH-VL to the CDRs (i.e. in what would be the buried V-domain - C-domain interface were the C-domain included in the structure file). I have attached a picture.

Would there be any way to specify regions to exlude from the docking computation? e.g. say confine it to the CDRs of the antibody only?

[julianrfaust]

What you are describing is essentially DockGPT by McPartlon and Xu which is the best fast Ab-Ag docking model in literature and can also take in contacts though the code isn’t fully released - likely due to a commercial venture MoleculeMind mentioned in the paper. Implementing this model can be done with some effort. The GeoDock code has a SabDab loader too which may have been attempted but was not mentioned in the GeoDock paper. Fine tuning the GeoDock model on SabDab seems promising.

[deasmhumhna]

Finetuning may not be necessary, depending on the problem. Considering the authors trained the model with 0 to 3 known contacts (similar to, and inspired by, DockGPT), you should be able to specify any known or specified contacts easily. If you only know the interfacing regions, you could use a genetic algorithm (check supplementary materials from DockGPT paper) to select candidates with contact points within the regions of interest. DockGPT includes a separate input for interfacing regions as a list of residues. With GeoDock, we're limited to explicit contacts, but we could use a similar genetic strategy or even try MCMC. Piping with a side-chain packing network like PIPPack might give something as performant as DockGPT. I'll start working on this to see what comes of it.

[julianrfaust]

At this point with the release of Boltz-1 essentially forcing their hand, Chai-1 is now fully commercial use and certainly the way to go if you want an open source SOTA model capable of handling residue restraints in November 2024. Geodock is a cool project but Chai-1 is an upgrade in performance and generalizability. If the restraints aren’t all that important to you, then Deepmind’s AF3 or Boltz-1 would also be in the mix for structure prediction. Chai-1 is almost like a AF3/DockGPT fusion and has great results docking Ab-Ag. They have some Ab-Ag benchmarking against Boltz-1 on their GitHub and this small external benchmark also exists in blog form: http://blog.booleanbiotech.com/alphafold3-boltz-chai1.html