Allow for multiple-chain graphs

[gcroci2]

• Atom level + residue/bases level
• get_contact_atoms in pdb2sql has already an allchains parameter.
• Do we want to indicate which chains should be considered for the interface? Or should we just use get_contact_atoms as it is? (All interfaces according to a cutoff distance will be selected).
• What happens when there are overlaps?
• Should they work with any combination of interactions? (i.e., not only proteins' interactions; verify if there is there any interesting use cases for this - maybe not)

[github-actions]

This issue is stale because it has been open for 30 days with no activity.

[lloydtripp]

I think I just ran into this this now. I have a 4 chain protein that I'd like to train a net on. Looks like the query collection programs are only set up for a single chain or two chain complex. What should I do in this case?

[DaniBodor]

Hi @lloydtripp , thanks for reaching out!

Just checking in to see what exactly you are trying to do, as it may be possible already in the current setup, even with a 4-chain structure. Can you let us know whether you want to train for protein-protein interfaces or a single residue variants?

• For SRV it is no problem to have any number of chains. In the SingleResidueVariantQuery you can specify the chain where the variant resides.
• For PPI:
  • are you looking at the interface of 2 chains, even if the structure contains more? This is also possible already. In the ProteinProteinInterfaceQuery you can specify the 2 chains between which the interface exists.
  • is it a multi-chain (>2) interface you are interested in, meaning that you want to know whether binding occurs when 3 or more chains come together within a single interface? This we are currently not set up for, but are planning to implement in the near future.
  • is it a multi-chain interaction with multiple separate possible interfaces, where you would expect each interaction to affect the other? This we are also not currently setup for but want to implement as well.

Is your use case not covered above? Please let us know and we can see if we can help you out or whether it is something we can add to the todo list :)

[lloydtripp]

Hi @DaniBodor ,

I'm looking to train a model that predicts pathogenicity based on residue substitutions on a single chain for a protein that exists as a tetramer. I agree with you that the SRV makes more sense for my case. I initially assumed that PPI was the correct query but that's not the case. Problem solved!

[DaniBodor]

Try it out and let us know if you run into any snags.

Glad you commented on this. It made us realize that we should (and will) update the README to make things clearer (#585).

[github-actions]

This issue is stale because it has been open for 30 days with no activity.