search

baker-laboratory / RoseTTAFold-All-Atom

Other
614 stars 108 forks source link

bad prediction for a homodimer #55

Open lucl13 opened 6 months ago

lucl13 commented 6 months ago

Hi all,

I'm using the config file below to do a homodimer prediction. But the prediction seems to be bad as there are clashes between two protomers. 


defaults:
  - base

job_name: pfka_dimer

protein_inputs:
  A: 
    fasta_file: $WORK_DIR/pfka.fasta
  B:
    fasta_file: $WORK_DIR/pfka.fasta

did I miss something?

img_v3_0295_ce207c83-a664-47ed-b513-d06ba7de20cg

liudongliangHI commented 6 months ago

This is a real problem, RFAA doesn't seem to have a relax step similar to AF2, I think you can fix the clashes by doing this step yourself with Chimera or Amber. Of course this would obviously lead to structural changes.

amorehead commented 6 months ago

I've noticed quite a few steric clashes between proteins and ligands myself as well. As @liudongliangHI suggests, I think using something like OpenMM to relax the predicted complex structures will be necessary.

dennissv commented 6 months ago

I've also experienced this and it reminds me of the issues the first AlphaFold multimer models had (https://github.com/google-deepmind/alphafold/issues/236). In that case, and at least for me it seems similar here, this mainly occurred when the predictions were of low confidence. I wonder if such drastic relaxations of the backbone that would be required to alleviate clashes in these cases would produce anything of value or if they're just indicative of the model not being able to place them confidently. It might result in something that looks like it could be correct but has little biological relevance. Of course smaller relaxation steps might be appropriate.

velocirraptor23 commented 5 months ago

Whta if it can be solved temprary by adding a linker? I am trying this approach, let you know what happens.