rfdiff

Intro

• Torres et al 2023. De novo design of high-affinity protein binders to bioactive helical peptides. https://www.biorxiv.org/content/10.1101/2022.12.10.519862v4

• Watson et al 2023. Broadly applicable and accurate protein design by integrating structure prediction networks and diffusion generative models. https://www.biorxiv.org/content/10.1101/2022.12.09.519842v1

• RF Diffusion now free and open source

• https://www.bakerlab.org/2023/03/30/rf-diffusion-now-free-and-open-source/

• Sergey Ovchinnikov implemented RF Diffusion onto the ColabFold platform.

Cyrus notes

• You can skip installation of RFDiffusion by using Dan’s copy on Apollo

conda activate /home/dfarrell/.local/share/micromamba/envs/rfdiffusion
cp /home/dfarrell/git/RFdiffusion ~/
python RFdiffusion/scripts/run_inference.py --help

Basic execution - an unconditional monomer

• Test on apollo:

export CUDA_VISIBLE_DEVICES=""
~/RFdiffusion/scripts/run_inference.py 'contigmap.contigs=[10-15]' inference.output_prefix=run1/test inference.num_designs=1

• Result: all-glycine peptide

~/RFdiffusion/scripts/run_inference.py 'contigmap.contigs=[150-150]' inference.output_prefix=run2/test inference.num_designs=1

Configs

• Hydra framework used to manage command-line options via "configs"
• RFdiffusion/config/inference/base.yaml contains defaults

contigmap config

• Defines the protein being built

contigmap:
  contigs: null
  inpaint_seq: null
  provide_seq: null
  length: null

Motif Scaffolding

• Anything prefixed by a letter indicates that this is a motif, with the letter corresponding to the chain letter in the input pdb files. E.g. A10-25 pertains to residues ('A',10),('A',11)...('A',25) in the corresponding input pdb

• Anything not prefixed by a letter indicates protein to be built. This can be input as a length range. These length ranges are randomly sampled each iteration of RFdiffusion inference.

• To specify chain breaks, we use /0 .

• 'contigmap.contigs=[5-15/A10-25/30-40/0 B1-100]'

• Test on apollo:

• Run in results_rfdiff/00_toy

cp ~/results_dovetail/00_toy/01_Rosetta_MSD/input_pdb/input_relaxed_repacked.pdb ~/rfdiff/00_toy
~/RFdiffusion/scripts/run_inference.py 'contigmap.contigs=[A1-100/101-106/5-8/0 B119-223/0 C224-320]' inference.input_pdb=input_relaxed_repacked.pdb inference.output_prefix=run3/test inference.num_designs=1

• Interface residues of chains A and B that are close to chain C: A31 A52 A53 A54 A57 A99 A100 A101 A102 A103 A104 A105 A106 A107 B144 B145 B146 B147 B148 B166 B182 B183 B184 B185 B207 B211i
• This runs right away and produces all glycine backbones in the designable regions
• The next step is apply MPNN to the glycine backbones

Create new installation (conda environment) instead of using Dan's

• name: rfdiff
• Create new environment and try to run again the same script, from inside results_rfdiff/00_toy

conda create --name rfdiff python=3.11.2 -y
conda activate rfdiff
pip install torch
pip install omegaconf
pip install hydra-core==1.3.2
pip install numpy
pip install dgl
pip install e3nn
pip install executing
pip install pygments
pip install asttokens
~/RFdiffusion/scripts/run_inference.py 'contigmap.contigs=[A1-100/101-106/5-8/0 B119-223/0 C224-320]' inference.input_pdb=input_relaxed_repacked.pdb inference.output_prefix=run4/test inference.num_designs=1

• leads to error:

Traceback (most recent call last):
  File "/home/aloshbaugh/RFdiffusion/scripts/run_inference.py", line 24, in <module>
    from rfdiffusion.util import writepdb_multi, writepdb
ModuleNotFoundError: No module named 'rfdiffusion'

• Trying to see why Dan's environment can automatically find package:

(base) ➜  env git:(main) ✗ ack 'RFdiffusion' /home/dfarrell/.local/share/micromamba/envs/rfdiffusion
/home/dfarrell/.local/share/micromamba/envs/rfdiffusion/lib/python3.11/site-packages/easy-install.pth
1:/home/dfarrell/git/RFdiffusion/env/SE3Transformer
2:/home/dfarrell/git/RFdiffusion

/home/dfarrell/.local/share/micromamba/envs/rfdiffusion/lib/python3.11/site-packages/se3-transformer.egg-link
1:/home/dfarrell/git/RFdiffusion/env/SE3Transformer

/home/dfarrell/.local/share/micromamba/envs/rfdiffusion/lib/python3.11/site-packages/rfdiffusion.egg-link
1:/home/dfarrell/git/RFdiffusion

/home/dfarrell/.local/share/micromamba/envs/rfdiffusion/bin/run_inference.py
5:__file__ = '/home/dfarrell/git/RFdiffusion/scripts/run_inference.py'

and

(base) ➜  env git:(main) ✗ cat /home/dfarrell/.local/share/micromamba/envs/rfdiffusion/lib/python3.11/site-packages/easy-install.pth
/home/dfarrell/git/RFdiffusion/env/SE3Transformer
/home/dfarrell/git/RFdiffusion
(base) ➜  env git:(main) ✗ cat /home/dfarrell/.local/share/micromamba/envs/rfdiffusion/lib/python3.11/site-packages/rfdiffusion.egg-link
/home/dfarrell/git/RFdiffusion
.%
(base) ➜  env git:(main) ✗ cat /home/dfarrell/.local/share/micromamba/envs/rfdiffusion/lib/python3.11/site-packages/se3-transformer.egg-link
/home/dfarrell/git/RFdiffusion/env/SE3Transformer
.%
(base) ➜  env git:(main) ✗ cat /home/dfarrell/.local/share/micromamba/envs/rfdiffusion/bin/run_inference.py
#!/home/dfarrell/.local/share/micromamba/envs/rfdiffusion/bin/python3.11
# EASY-INSTALL-DEV-SCRIPT: 'rfdiffusion==1.0.0','run_inference.py'
__requires__ = 'rfdiffusion==1.0.0'
__import__('pkg_resources').require('rfdiffusion==1.0.0')
__file__ = '/home/dfarrell/git/RFdiffusion/scripts/run_inference.py'
with open(__file__) as f:
    exec(compile(f.read(), __file__, 'exec'))

• I tried:

export PYTHONPATH=/home/aloshbaugh/RFdiffusion

• and now it has moved on to the next error so maybe this worked.
• New error:

  File "/home/aloshbaugh/RFdiffusion/rfdiffusion/SE3_network.py", line 9, in <module>
    from se3_transformer.model import SE3Transformer
ModuleNotFoundError: No module named 'se3_transformer'

• This is actually contained within the RFDiffusion git repo: RFdiffusion/env/SE3Transformer
• Attempted solution:

export PYTHONPATH="/home/aloshbaugh/RFdiffusion/env/SE3Transformer:$PYTHONPATH"

• I think this worked because it moved on to the next error which is a standard package not found. For brevity, adding it to the list of pip install above.
• Now moved on to a new type of error:

dgl._ffi.base.DGLError: [12:45:09] /opt/dgl/src/array/array.cc:42: Operator Range does not support cuda device.

• Which I believe is related to the wrong version of dgl installed
• Dan's environment had:

    - dgl==1.0.2+cu118
    - dglgo==0.0.2

• Whereas this environment has

dgl                       1.1.1                    pypi_0    pypi

• Command nvidia-smi reveals CUDA Version: 11.8 which explains the +cu118 term
• Attempted resolution from https://stackoverflow.com/questions/76519346/unable-to-install-dgl-cuany-version-in-google-colab:

 pip install dgl==1.0.2+cu118 -f https://data.dgl.ai/wheels/cu118/repo.html

• This resolved the last problem and the script now runs.

Known errors when running

• CUDA errors

RuntimeError: CUDA error: out of memory

• Resolution:

nvidia-smi
export CUDA_VISIBLE_DEVICES="" # example
export CUDA_VISIBLE_DEVICES=2,3,4 # example