Open Najy-Yusuf opened 1 year ago
I'm also running into a similar issue. My workflow is:
1) Generate backbones with RFdiffusion
2) I then run dl_interface_design
to call ProteinMPNN to map a sequence
dl_interface_design.py -pdbdir . -outpdbdir relax/ -debug
3) Afterwhich I run predict.py
:
`predict.py -pdbdir relax/ -outpdbdir af2/ -scorefilename af2.sc -debug -recycle 5
I've tried several different backbones and all of my scores are around 26-27 for the pae_interaction
. When I look at the structures in pymol (1. rfdiffusion backbone (blue), 2. proteinmpnn mapped backbone (green) and 3. af2 predicted structure (red), the rfdiffusion backbone and proteinmpnn align for the most part, however, the AF2 predicted structure shows the binder very far away.
what is your workflow for visualizing binding in pymol?
After re-reading the paper I probably just need to run things a little longer and include a few controls.
dl_interface_design.py -pdbdir . -outpdbdir relax/ -debug
to generate ProteinMPNN-FastRelax complex structures. The input were several PDB files where chain A was the binder (all glycines), and chain B was my receptor (sometimes a monomer, sometimes a trimer, but always just one chain id). I then did a QA visual check to see how those look in pymol relative to the RFdiffusion output, and those make sense. My binders are shifting a little bit as to be expected. predict.py -pdbdir relax/ -outpdbdir af2/ -scorefilename af2.sc -debug -recycle 5
, where I take as my input the output of the ProteinMPNN-FastRelax step, and I'm assuming it's taking just the ProteinMPNN sequence and having AF2 generate the binder. From that I was assuming I just use pae-interaction to filter out any complexes with a PAE > 10. I used RFdiffusion a couple weeks ago to scaffold a motif and that was a little more straightforward as i could just run ESMfold on my ProteinMPNN sequences and them just compare RMSD against the RFdiffusion backbones to in silico screen.
I think I need to just work through one of the examples as a control to see if I'm just overlooking something.
I'm also running into a similar issue. My workflow is:
- Generate backbones with RFdiffusion
- I then run
dl_interface_design
to call ProteinMPNN to map a sequence
dl_interface_design.py -pdbdir . -outpdbdir relax/ -debug
- Afterwhich I run
predict.py
:`predict.py -pdbdir relax/ -outpdbdir af2/ -scorefilename af2.sc -debug -recycle 5
I've tried several different backbones and all of my scores are around 26-27 for the
pae_interaction
. When I look at the structures in pymol (1. rfdiffusion backbone (blue), 2. proteinmpnn mapped backbone (green) and 3. af2 predicted structure (red), the rfdiffusion backbone and proteinmpnn align for the most part, however, the AF2 predicted structure shows the binder very far away.
Have you solved the problem? I also got pae_interaction around 25 for all my 100 designs. I am just confused about it.
Were trying to get an pae_interaction less that 10, but when we put the contig as A100-200/0 70-100 we get i_pae (we assume its the pae_interaction score) as around 26, but when we use contig of 70-100 we get a pae as under 10, but no i, so we are unsure whether that is the interaction, binder, or target score.
There are 2 example Contigs we used:
A100-200/ 0 70-100: design:0 n:0 mpnn:1.169 plddt:0.859 i_ptm:0.069 i_pae:26.417 rmsd:32.305
70-100: design:0 n:2 mpnn:0.986 plddt:0.944 ptm:0.765 pae:3.135 rmsd:1.291