Multimer and Alphafold3 prediction

[Kangfengyuuuu]

May I ask how to use multimer parameters for predicting complexes in the AF2 initial guess section after designing the sequence for ProteinMPPNN. Why do some proteins with poor plddt_binder (=50) and pae_interaction (=27) predicted by AlphaFold3 have high iptm and ptm scores and small PAE when I use AlphaFold3 for prediction?

[JieGao1111]

I have the same issue, has this issue been solved?

[21tesla]

The AF2 module doesn't seem to dock the ligand and target. I am not sure why.

I switched my pipeline to (1) RFdiffusion (2) ProteinMPNN from dl_binder_design and (3) Chai-1 (without MSAs). I sort my solutions on the ipTM score given by Chai-1.

[RodenLuo]

I noticed a similar issue for binder-target designs. I ran <base_dir>/af2_initial_guess/predict.py -pdbdir <pdb>-outpdbdir <outs> and got PAEs around 26 with binder and target were a bit far away and were not really interacting. If I ran AF3, the binder and the target would interact as designed, and the PAEs would be below 5. (I did notice if I ran a published binder-target design, the AF2 pipeline indeed docked them together and gave a PAE below 5.)

May I understand why you prefer Chai-1 over AF3? @21tesla

[21tesla]

No preference of Chai-1 over AF3... It was just straightforward to set up Chai-1 and write a batch script to send proteins to it and afterwards, sort the ipTM values. It works reasonably fast too, when you skip the requirement for an MSA.

I have it all running on an L40S (Cuda 12.4).

[RodenLuo]

I see. Thanks!