Open kyleford8 opened 1 day ago
One reason I see that this could because "force_reject_AA" is enabled in peptide settings. So the omit_AAs setting to exclude amino acids is basically just reducing the probability, but if AF2 thinks that amino acid should definitely be there it will place it there, for example a cysteine at the interface. With the force setting we remove it. Perhaps you are getting cysteines with the default settings?
None of the sequences in trajectory_stats (for peptide or default 4stage_multimer settings) or the final_design_stats (for the 4stage multimer run which yielded hits) contain cysteines, so I would be surprised if it's a consistent cysteine issue. Linking another MPNN issue which I've also observed, which may be related? #74 For the linked issue, even though the print statements make it seem like it's only testing one sequence repeatedly, when I look in the output Accepted folder I still have mpnn1-20 suffixed files.
If I have time later I can try to reproduce with the PDL1 example .pdb file.
When running with the default settings (4stage_multimer), if a "No accepted MPNN designs found for this trajectory." occurs, it's usually associated with a print statement like: "Base AF2 filters not passed for XXXX, skipping interface scoring".
When running with the peptide settings, I keep observing "No accepted MPNN designs found for this trajectory", but with no associated print statements. In addition, when I look at the failure_csv.csv I can't seem to see any explanation as to why these MPNN designs are failing for a given trajectory.
Is there something I'm missing? I currently have the same target running with both settings (only difference, keeping the filtering the same), and I quickly find hits for the standard settings, and have seen zero hits accepted for the peptide settings. Anecdotally, across multiple targets I have had little success with the peptide settings, although that could well just be real biology.