Feature request: interface distance restraints for RF3

[tanggis]

foundry/models/rf3/README.md

Line 533 in 88d212c

RF3 was also trained to respect interface templates; that is, pairwise distances between tokens across an interface. We have yet to extend our inference API to support this use case; if it would be helpful, please raise an issue on GitHub and we can re-prioritize accordingly.

Hi team! Would you consider prioritizing support for interface distance restraints (residue–residue distances across polymer chains) in RF3? This would help us steer multimer predictions/designs. Happy to share examples and help with testing. Thank you!

[nscorley]

Hi! Thanks for the comment! We can definitely do that -- it's just an inference API change.

One question: how do you intend to use this for filtering?

We've found that if you care about the docking of, say, chains A and B, and you template the A/B interface, the confidence metrics for that interface tend to lose their meaning. The model essentially "trusts" the templated interface rather than second-guessing it, so the confidence scores don't reflect genuine uncertainty. That said, if you just want to force a particular dock and filter based on some other component of the complex, then that may be meaningful.

[tanggis]

Thanks for that info- good to know! Yeah we're hoping to see improvements in cofolding by providing some known/predicted pairwise knowledge, so even if we lose the interface confidence it'll still be useful.

That said, I wonder if "templating" has to be provided with a full structure or can it be in the format of some pairwise distance restraints. If we only provide limited info, say a couple pairs of residues, can the confidence scores in the rest of the interface still be reliable?

Thank you!

[nscorley]

On the dataset of de novo designed protein binders we evaluated on, even providing one pairwise contact at the interface significantly hindered the model's ability to act as a filter. Monomer prediction is quite good, so if you provide one contact across an interface it significantly reduces the degrees of freedom. We're looking into "softer" constraints that the model can choose to ignore, such a large hotspots or even modulating the impact of the pairwise distance conditioning at training time