Implementation for "SE(3) diffusion model with application to protein backbone generation" arxiv link. While our work is tailored towards protein backbone generation, it is in principle applicable to other domains where SE(3) is utilized.
For those interested in non-protein applications, we have prepared a minimal notebook with SO(3) diffusion https://colab.research.google.com/github/blt2114/SO3_diffusion_example/blob/main/SO3_diffusion_example.ipynb
We have codebase updates we plan to get around to.
- [In the works] Refactor score framework to be more readable and match the paper's math. See the refactor branch.
- Set-up easily downloadable training data.
We welcome pull requests (especially bug fixes) and contributions. We will try out best to improve readability and answer questions!
If you use our work then please cite
@article{yim2023se,
title={SE (3) diffusion model with application to protein backbone generation},
author={Yim, Jason and Trippe, Brian L and De Bortoli, Valentin and Mathieu, Emile and Doucet, Arnaud and Barzilay, Regina and Jaakkola, Tommi},
journal={arXiv preprint arXiv:2302.02277},
year={2023}
}
Other protein diffusion codebases:
- Pretrained protein backbone diffusion: RFdiffusion
- Protein-ligand docking: DiffDock
- Protein torsion angles: FoldingDiff
- Protein C-alpha backbone generation: ProtDiff/SMCDiff
LICENSE: MIT
- Bucketize bug. We discovered an unfortunate bug in the SO3 component of the score approximation computation. The issue was the use of
torch.bucketizewhen accessing cached IGSO3 density valuestorch_scoreofSO3Diffuser; since thebucketizefunction does not track gradients, this operation effectively introduced a stopgrad on the angle of rotation component of the rotation score thereby expectedly blocking some gradients during training. We have fixed this now but it has caused model performance to change. - Clustered training. We added the ability use clustered data during training. We found this to greatly improve sample diversity. See Downloading PDB clusters.
- New configs.
config/base.yamlupdated with the best parameters we have found. They surpass the ICML published results (see table). The changes are the following:- Batching with
experiment.sampling_mode=cluster_time_batch. - A new rotation loss that includes explicit penalties of errors in the axis and angle components of the loss on conditional score estimate
experiment.separate_rot_loss=True. This was part of the published model with the bug that we accidentally found to help. We found it beneficial to down-weight the angle term of the rotation loss,experiment.rot_loss_weight=0.5, and turn off the angle term for t<0.2,experiment.rot_loss_t_threshold=0.2. Though heuristic, we find this loss to further outperform the DSM loss by our metrics.
- Batching with
config/pure_dsm.yamlsame asconfig/base.yamlexcept hasexperiment.separate_rot_loss=Falseand instead uses the DSM rotation loss described in our paper, but without thebucketizebug.config/icml_published.yamlin case one wants to reproduce results from the ICML paper.
We are still running some experiments to verify these improvements but we wanted to get this out sooner for researchers building on our codebase. Here is a table of our preliminary results so far:
| icml_published.yaml | base.yaml | |
|---|---|---|
| Designability (scRMSD < 2) | 0.29 | 0.34 |
| Diversity (TM cutoff 0.5) | 0.43 | 0.61 |
We fine-tuned one of our checkpoints for the base.yaml results. pure_dsm.yaml model is training.
We will update the table of base.yaml once it is finished training from scratch (unfortunately this takes at least a week).
There's pretty strong signal that our new settings are showing large improvements across our metrics.
The weights used to get the icml_published.yaml results are in weights/paper_weights.pth while the weights to get the base.yamlresults are in weights/best_weights.pth.
The authors are currently all very busy until the Fall (Jason is currently interning). However, we are committed to answering questions and fixing any bugs. Please email us or raise an issue if you have any concerns or questions.
- SE(3) diffusion model with application to protein backbone generation
- Table of Contents
- Installation
- Inference
- Training
- Acknowledgements
We recommend miniconda (or anaconda). Run the following to install a conda environment with the necessary dependencies.
conda env create -f se3.ymlNext, we recommend installing our code as a package. To do this, run the following.
pip install -e .
Our repo keeps a fork of OpenFold since we made a few changes to the source code.
Likewise, we keep a fork of ProteinMPNN.
Each of these codebases are actively under development and you may want to refork.
We use copied and adapted several files from the AlphaFold primarily in /data/, and have left the DeepMind license at the top of these files.
For a differentiable pytorch implementation of the Logarithmic map on SO(3) we adapted two functions form geomstats.
Go give these repos a star if you use this codebase!
inference_se3_diffusion.py is the inference script. It utilizes Hydra.
Training can be done with the following.
python experiments/inference_se3_diffusion.pyThe config for inference is in config/inference.yaml.
See the config for different inference options.
By default, inference will use the published paper weights in weights/paper_weights.pth.
Simply change the weights_path to use your custom weights.
inference:
weights_path: <path>Samples will be saved to output_dir in the inference.yaml. By default it is
set to ./inference_outputs/. Sample outputs will be saved as follows,
inference_outputs
└── 12D_02M_2023Y_20h_46m_13s # Date time of inference.
├── inference_conf.yaml # Config used during inference.
└── length_100 # Sampled length
├── sample_0 # Sample ID for length
│ ├── bb_traj_1.pdb # x_{t-1} diffusion trajectory
│ ├── sample_1.pdb # Final sample
│ ├── self_consistency # Self consistency results
│ │ ├── esmf # ESMFold predictions using ProteinMPNN sequences
│ │ │ ├── sample_0.pdb
│ │ │ ├── sample_1.pdb
│ │ │ ├── sample_2.pdb
│ │ │ ├── sample_3.pdb
│ │ │ ├── sample_4.pdb
│ │ │ ├── sample_5.pdb
│ │ │ ├── sample_6.pdb
│ │ │ ├── sample_7.pdb
│ │ │ └── sample_8.pdb
│ │ ├── parsed_pdbs.jsonl # Parsed chains for ProteinMPNN
│ │ ├── sample_1.pdb
│ │ ├── sc_results.csv # Summary metrics CSV
│ │ └── seqs
│ │ └── sample_1.fa # ProteinMPNN sequences
│ └── x0_traj_1.pdb # x_0 model prediction trajectory
└── sample_1 # Next sampleTo get the training dataset, first download PDB then preprocess it with our provided scripts. PDB can be downloaded from RCSB: https://www.wwpdb.org/ftp/pdb-ftp-sites#rcsbpdb. Our scripts assume you download in mmCIF format. Navigate down to "Download Protocols" and follow the instructions depending on your location.
WARNING: Downloading PDB can take up to 1TB of space.
After downloading, you should have a directory formatted like this: https://files.rcsb.org/pub/pdb/data/structures/divided/mmCIF/
00/
01/
02/
..
zz/
In this directory, unzip all the files:
gzip -d **/*.gz
Then run the following with <path_pdb_dir> replaced with the location of PDB.
python process_pdb_dataset.py --mmcif_dir <pdb_dir> See the script for more options. Each mmCIF will be written as a pickle file that
we read and process in the data loading pipeline. A metadata.csv will be saved
that contains the pickle path of each example as well as additional information
about each example for faster filtering.
For PDB files, we provide some starter code in process_pdb_files.py of how to
modify process_pdb_dataset.py to work with PDB files (as we did at an earlier
point in the project). This has not been tested. Please make a pull request
if you create a PDB file processing script.
To use clustered training data, download the clusters at 30% sequence identity at rcsb. This download link also works at time of writing:
https://cdn.rcsb.org/resources/sequence/clusters/clusters-by-entity-30.txt
Place this file in data/processed_pdb or anywhere in your file system.
Update your config to point to the clustered data:
data:
cluster_path: ./data/processed_pdb/clusters-by-entity-30.txtTo use clustered data, set sample_mode to either cluster_time_batch or cluster_length_batch.
See next section for details.
experiment:
# Use one of the following.
# Each batch contains multiple time steps of the same protein.
sample_mode: time_batch
# Each batch contains multiple proteins of the same length.
sample_mode: length_batch
# Each batch contains multiple time steps of a protein from a cluster.
sample_mode: cluster_time_batch
# Each batch contains multiple clusters of the same length.
sample_mode: cluster_length_batchtrain_se3_diffusion.py is the training script. It utilizes Hydra.
Hydra does a nice thing where it will save the output, config, and overrides of each run to the outputs/ directory organized by date and time. By default we use 2 GPUs to fit proteins up to length 512. The number of GPUs can be changed with the num_gpus field in base.yml.
Training can be done with the following.
python experiments/train_se3_diffusion.pyThe config for training is in config/base.yaml.
See the config for different training options.
Training will write losses and additional information in the terminal.
We support wandb which can be turned on by setting the following option in the config.
experiment:
use_wandb: TrueMulti-run can be achieved with the -m flag. The config must specify the sweep.
For an example, in config/base.yaml we can have the following:
defaults:
- override hydra/launcher: joblib
hydra:
sweeper:
params:
model.node_embed_size: 128,256This instructs hydra to use joblib
as a pipeline for launching a sweep over data.rosetta.filtering.subset for two
different values. You can specify a swep with the -m flag. The training script
will automatically decide which GPUs to use for each sweep. You have to make sure
enough GPUs are available on your server.
python experiments/train_se3_diffusion.py -mEach training run outputs the following.
- Model checkpoints are saved to
ckpt. - Hydra logging is saved to
outputs/including configs and overrides. - Samples during intermittent evaluation (see next section) are saved to
eval_outputs. - Wandb outputs are saved to
wandb.
Training also performs evaluation everytime a checkpoint is saved. We select equally spaced lengths between the minimum and maximum lengths seen during training and sample multiple backbones for each length. These are then evaluated with different metrics such as secondary structure composition, radius of gyration, chain breaks, and clashes.
We additionally evaluate TM-score of the sample with a selected example from the training set. This was part of a older research effort for something like protein folding. We keep this since it'll likely be useful to others but it can be ignored for the task of unconditional generation.
All samples and their metrics can be visualized in wandb (if it was turned on). The terminal will print the paths to which the checkpoint and samples are saved.
[2023-02-09 15:10:25,097][__main__][INFO] - Checkpoints saved to: <ckpt_path>
[2022-02-09 15:10:25,097][__main__][INFO] - Evaluation saved to: <sample_path>This can also be found in the config in Wandb by searching ckpt_dir.
Once you have a good run, you can copy and save the weights somewhere for inference.
Thank you to the following for pointing out bugs:
- longlongman
- amorehead
