RFD3 (RFdiffusion3) uses an atom-level diffusion framework to design proteins around ligands, nucleic acids, and other nonprotein molecular components. Because it generates complete atomic structures rather than only protein backbones, it can incorporate detailed geometric constraints for applications such as enzyme engineering and the design of proteins that bind small molecules or DNA.
Allocate an interactive session and run the program. Sample session (based on Foundry's Protein-Protein Interface Design Tutorial).
[user@biowulf]$ sinteractive --mem=20g --cpus-per-task=8 --gres=gpu:a100:1 [user@cn3335 ~]$ module load rfdiffusion/rfd3 [+] Loading rfdiffusion rfd3 on cn3335 [user@cn3335 ~]$ rfd3 --help Usage: rfd3 [OPTIONS] Run design using hydra config overrides and input files. ╭─ Options ────────────────────────────────────────────────────── │ --install-completion Install completion for the current shell. │ --show-completion Show completion for the current shell, to copy it or customize the installation. │ --help Show this message and exit. ╰─────────────────────────────────────────────────────────────────Download sample data:
[user@cn3335 ~]$ cd /data/$USER [user@cn3335 ~]$ git clone https://github.com/RosettaCommons/foundry Cloning into 'foundry'... remote: Enumerating objects: 12529, done. remote: Counting objects: 100% (219/219), done. remote: Compressing objects: 100% (154/154), done. remote: Total 12529 (delta 126), reused 70 (delta 65), pack-reused 12310 (from 3) Receiving objects: 100% (12529/12529), 178.68 MiB | 76.88 MiB/s, done. Resolving deltas: 100% (7727/7727), done. Updating files: 100% (918/918), done.Create directory for test and cd into it:
[user@cn3335 ~]$ mkdir rfd3_ppi_tutorial && cd rfd3_ppi_tutorialCreate your YAML file with the following contents:
[user@cn3335 ~]$ cat ppi_tutorial.yaml
insulinr:
input: /data/$USER/foundry/models/rfd3/docs/input_pdbs/4zxb_cropped.pdb
contig: 40-120,/0,E6-155
length: 190-270
select_hotspots:
E64: CD2,CZ
E88: CG,CZ
E96: CD1,CZ
infer_ori_strategy: hotspots
is_non_loopy: true
Run RFD3:
[user@cn3335 ~]$ rfd3 design out_dir=ppi_tutorial_outputs/0 inputs=ppi_tutorial.yaml ckpt_path=$FOUNDRY_CHECKPOINT_DIRS/rfd3_latest.ckpt
14:46:47 DEBUG transforms: Debug mode is on
14:46:49 INFO rfd3.engine: [rank: 0] Outputs will be written to /data/username/rfd3_ppi_tutorial/ppi_tutorial_outputs/0.
14:46:49 INFO rfd3.engine: [rank: 0] Found 0 existing example IDs in the output directory (0 total).
Using bfloat16 Automatic Mixed Precision (AMP)
You are using a CUDA device ('NVIDIA A100-SXM4-80GB') that has Tensor Cores. To properly utilize them, you should set `torch.set_float32_matmul_precision('medium' | 'high')` which will trade-off precision for performance. For more details, read https://pytorch.org/docs/stable/generated/torch.set_float32_matmul_precision.html#torch.set_float32_matmul_precision
14:46:53 WARNING atomworks.io: We can't fix formal charges without building from templates, as we need to know the true number of hydrogens bonded to a given atom, not the inferred number. This may lead to occasional inaccuracies after adding inter-residue bonds. To avoid this and fix formal charges, set `add_missing_atoms = True`.
14:46:54 WARNING atomworks.io: The `extra_fields` argument will be ignored if there is no CIF file input.
14:47:42 INFO rfd3.engine: [rank: 0] Finished inference batch in 47.53 seconds.
14:47:42 INFO rfd3.engine: [rank: 0] Outputs for ppi_tutorial_insulinr_0_model_0 written to /data/username/rfd3_ppi_tutorial/ppi_tutorial_outputs/0/ppi_tutorial_insulinr_0_model_0.
14:47:42 INFO rfd3.engine: [rank: 0] Outputs for ppi_tutorial_insulinr_0_model_1 written to /data/username/rfd3_ppi_tutorial/ppi_tutorial_outputs/0/ppi_tutorial_insulinr_0_model_1.
14:47:42 INFO rfd3.engine: [rank: 0] Outputs for ppi_tutorial_insulinr_0_model_2 written to /data/username/rfd3_ppi_tutorial/ppi_tutorial_outputs/0/ppi_tutorial_insulinr_0_model_2.
14:47:42 INFO rfd3.engine: [rank: 0] Outputs for ppi_tutorial_insulinr_0_model_3 written to /data/username/rfd3_ppi_tutorial/ppi_tutorial_outputs/0/ppi_tutorial_insulinr_0_model_3.
14:47:42 INFO rfd3.engine: [rank: 0] Outputs for ppi_tutorial_insulinr_0_model_4 written to /data/username/rfd3_ppi_tutorial/ppi_tutorial_outputs/0/ppi_tutorial_insulinr_0_model_4.
14:47:42 INFO rfd3.engine: [rank: 0] Outputs for ppi_tutorial_insulinr_0_model_5 written to /data/username/rfd3_ppi_tutorial/ppi_tutorial_outputs/0/ppi_tutorial_insulinr_0_model_5.
14:47:42 INFO rfd3.engine: [rank: 0] Outputs for ppi_tutorial_insulinr_0_model_6 written to /data/username/rfd3_ppi_tutorial/ppi_tutorial_outputs/0/ppi_tutorial_insulinr_0_model_6.
14:47:42 INFO rfd3.engine: [rank: 0] Outputs for ppi_tutorial_insulinr_0_model_7 written to /data/username/RFD3_TEST/rfd3_ppi_tutorial/ppi_tutorial_outputs/0/ppi_tutorial_insulinr_0_model_7.
End the interactive session:
[user@cn3335 ~]$ exit salloc.exe: Relinquishing job allocation 46116226 [user@biowulf ~]$
Create a batch input file, e.g. rfd3.sh:
#!/bin/sh #SBATCH --job-name ppi #SBATCH --cpus-per-task=8 #SBATCH --mem=20g #SBATCH --gres=gpu:a100:1 #SBATCH -p gpu module load rfdiffusion/rfd3 cd /data/$USER git clone https://github.com/RosettaCommons/foundry mkdir rfd3_ppi_tutorial && cd rfd3_ppi_tutorial rfd3 design out_dir=ppi_tutorial_outputs/0 inputs=ppi_tutorial.yaml
Submit this job using the Slurm sbatch command.
sbatch rfd3.sh