Deep Learning Binder Results

Hi @joewatchwell ,
I was attempting to design insulin binders of my own (following the example specifications) and then running the backbone through MPNN_FR and AF2 initial guess. I compared a potential binder to a benchmark insulin binder from the supplement material in Improving de novo protein binder design with deep learning, and even though the target template was identical, I found confusing results:

InsulinR_mb:
{'plddt_total': 95.02760208110635, 'plddt_binder': 91.0824370734358, 'plddt_target': 96.7371735844303, 'pae_binder': 2.7015252, 'pae_target': 2.4785695, 'pae_interaction': 4.80579948425293, 'time': 146.24850199604407}

design_ppi_scaffolded_6_dldesign_4:
{'plddt_total': 52.28481075101907, 'plddt_binder': 94.86918808372161, 'plddt_target': 33.83158057351464, 'pae_binder': 1.7483437, 'pae_target': 19.690062, 'pae_interaction': 26.69976043701172, 'time': 16.96811721706763}

I swapped the target pdb structure of Insulin R_mb (chain B) into for the target structure of design_ppi_scaffolded_6_dldesign_4, and received the expected results. It seems that the RFdiffusion cleans a great amount of detail in terms of side chains on the target structure when submitted to make a binder. Is there a way to maintain the nuance of the target structure? Thanks and would appreciate the help!

design_ppi_scaffolded_6_dldesign_4.pdb.zip

InsulinR_mb.pdb.zip

Start by reproducing the reported comparison with the attached design_ppi_scaffolded_6_dldesign_4 and InsulinR_mb PDB files, then trace how RFdiffusion prepares the target before MPNN_FR and AF2 initial-guess evaluation. Done means establishing whether target side-chain detail is intentionally removed or incorrectly handled, with the observed metric difference explained or a focused fix validated.