Binder Design with Protein Generator

[GoldEagle93]

Hello and congrats again to the Baker lab for the well-deserved Nobel Prize!

I'm trying to achieve the functionality of RFDiffusion+ProteinMPNN with Protein Generator. I have tested on three systems: 1YCR, 1SE0 and 2B1N. Despite correct specification of hotspot residues, it is generating the peptide in the wrong place in 1SE0 and 2B1N. I've attached snapshots as well as the commands used to run and the input files. Would you mind taking a look and advising what could be the problem?

In the snapshots below, red is generated with Protein Generator and green is the experimentally known peptide binder.

run-2b1n.txt 1ycr.pdb.txt 1se0.pdb.txt 2b1n.pdb.txt

[0merle0]

Just to make sure I understand correctly the hotspots are specified at the contacts of the green "experimentally determined" binder?

[HanLab-OSU]

hi @0merle0 Could you please add some sample runs for binder design? I could not find binder_design.sh within the examples folder. Thanks, Renzhi

[GoldEagle93]

Here's a sample script of how I ran these jobs (except for hotspot residues and length of chain A, everything else is the same for all cases): python /path/to/protein_generator/inference.py \ --num_designs 2 \ --out output/2b1n \ --pdb 2b1n.pdb \ --T 10 \ --save_best_plddt \ --contigs A1-227,0 6-8 \ --hotspots A19,A65,A166,A147,A21,A64 \ --checkpoint /path/to/protein_generator/SEQDIFF_221219_equalTASKS_nostrSELFCOND_mod30.pt

[HanLab-OSU]

Thank you. I tried this pipeline. For my target, when I use alphafold3 to check the designed binders, the ipTM score is very low despite the high pLDDT in protein_generator, and not binding to the hotspots that I specified. Do you have the same problem?

[GoldEagle93]

Actually I haven't tried that. Will try and let you know the results.