search

HeliXonProtein / OmegaFold

OmegaFold Release Code
Apache License 2.0
547 stars 76 forks source link

adding support for multichain modeling (with residue index modification) and input of real msa #5

Closed sokrypton closed 2 years ago

RuiWang1998 commented 2 years ago

Hi,

Thanks for the suggestion! We will very soon review the code and merge the pull request (eta within 24 hours).

Best Rui

sokrypton commented 2 years ago

This was just a quick hack from my side! If you guys want to do a more proper or leigt implementation of the same concept, feel free to close. I will not be offended. 😄

RuiWang1998 commented 2 years ago

Thanks it looks great! We are going to look into it further and make the decision.

sokrypton commented 2 years ago

The one part I was not sure about is the sin/cos embedding and if that also needed to be offset in some way using the residue index?

RuiWang1998 commented 2 years ago

Yeah they are a bit tricky. I think they should be offset in some way, which we have some implementations in some other projects. Guess we will probably just copy-paste those.

sokrypton commented 2 years ago

for real MSA option, i wonder if it makes sense to unalign the sequences, pass full sequences to LM then realign the embeddings (or maybe provide a different residue index for each sequence? which would include info based on insertions/deletions)

sokrypton commented 2 years ago

Yeah they are a bit tricky. I think they should be offset in some way, which we have some implementations in some other projects. Guess we will probably just copy-paste those.

I think I fixed this with latest commit

sokrypton commented 2 years ago

oops accidentally renamed branch. I'll submit a new pull request soon :D