I was trying to follow the procedure, starting from a .pdb file including three residues (2xHis and one Ser) and a substrate to design a protein using CA_RFDiffusion. However, after running both diffusion and refinement, I obtained some unexpected results. It appeared that ij_visible was not able to hold the residue-residue and residue–substrate rigidly. I used inference.ij_visible='abcd' since I only have three residues and one substrate marked as LIG in the pdb file, and I believed that would hold everything together rigidly.
The distance between alpha-C of the residues, as well as the distance between residue-alpha-C and substrate almost doubled, while the substrate part itself seems remained (almost) intact. The relative position of the residues also seemed OK. I have not been able to determine what went wrong, and I would be very grateful for any insights you might have.
I have also successfully replicated the serine hydrolase example provided, so I reckon it's not a problem with my local deployment itself.
Here are attached my bash scripts and the .yaml files for diffusion and refinement. My .pdb file only includes the side chain, alpha-C, and beta-C, together with my substrate. I was also wondering if this input pdb is too simplified for the model. I can also attach my .pdb if its needed for troubleshooting.
Many thanks!
#ca_rfd_diffuse.sh
#!/bin/bash
export PYTHONPATH=$PYTHONPATH:/home/foo/CA_RFDiffusion/rf_diffusion/
output_pref="/home/foo/CA_RFDiffusion/rf_diffusion/experiments/test7"
python run_inference.py \
--config-name="RFdiffusion_CA_inference" \
inference.input_pdb=/home/foo/CA_RFDiffusion/rf_diffusion/experiments/theozyme-v3.pdb \
inference.output_prefix=${output_pref} \
inference.num_designs=1 \
inference.ij_visible='abcd'
#ca_rfd_refine.sh
#!/bin/bash
export PYTHONPATH=$PYTHONPATH:/home/foo/CA_RFDiffusion/rf_diffusion/
pdb="/home/foo/CA_RFDiffusion/rf_diffusion/experiments/test7_0.pdb"
CKPT='/home/foo/CA_RFDiffusion/checkpoints/ca_rfd_refinement.pt'
python run_inference.py \
--config-name=test_ca_rfd_refinement \
inference.num_designs=4 \
inference.input_pdb=$pdb \
inference.ligand='LIG' \
inference.ckpt_path=$CKPT
#RFdiffusion_CA_inference.yaml, the rest of the file was untouched
inference:
output_prefix: ./experiments/caRFD_test
ckpt_path: '/home/foo/CA_RFDiffusion/checkpoints/ca_rfd_diffusion.pt'
input_pdb: /home/foo/CA_RFDiffusion/rf_diffusion/experiments/Ni-theozyme-v2.pdb
str_self_cond: 1
ij_visible: 'abcde' # e is the ligand
length: 90-500
ligand: 'LIG'
write_trajectory: true
recenter_xt: true
num_designs: 15
cautious: true
guidepost_xyz_as_design: false
contigmap:
contigs: ['5-20,A1,1-5,A2,1-5,A3,5-20']
I was trying to follow the procedure, starting from a .pdb file including three residues (2xHis and one Ser) and a substrate to design a protein using CA_RFDiffusion. However, after running both diffusion and refinement, I obtained some unexpected results. It appeared that ij_visible was not able to hold the residue-residue and residue–substrate rigidly. I used
inference.ij_visible='abcd'since I only have three residues and one substrate marked as LIG in the pdb file, and I believed that would hold everything together rigidly.The distance between alpha-C of the residues, as well as the distance between residue-alpha-C and substrate almost doubled, while the substrate part itself seems remained (almost) intact. The relative position of the residues also seemed OK. I have not been able to determine what went wrong, and I would be very grateful for any insights you might have.
I have also successfully replicated the serine hydrolase example provided, so I reckon it's not a problem with my local deployment itself.
Here are attached my bash scripts and the .yaml files for diffusion and refinement. My .pdb file only includes the side chain, alpha-C, and beta-C, together with my substrate. I was also wondering if this input pdb is too simplified for the model. I can also attach my .pdb if its needed for troubleshooting.
Many thanks!