NOTE: This is intended as a static repository corresponding to the preprint release. Active development will continue at the PDL Github repo; please make contributions, issues, requests, etc. at that repository.
Code for the paper: An all-atom protein generative model.
The code is under active development and we welcome contributions, feature requests, issues, corrections, and any questions! Where we have used or adapted code from others we have tried to give proper attribution, but please let us know if anything should be corrected.
To set up the conda environment, run conda env create -f configs/environment.yml then conda activate delle. You will also need to clone the ProteinMPNN repository to the same directory that contains the protpardelle/ repository. You may also need to set the home_dir variable in the configs you use to the path to the directory containing the protpardelle/ directory.
The entry point for sampling is draw_samples.py, which is a convenience wrapper around sampling.py and the model.sample() function. There are a number of arguments which can be passed to control the model checkpoints, the sampling configuration, and lengths of the proteins sampled. Model weights are provided for both the backbone-only and all-atom versions of Protpardelle. Both of these are trained unconditionally; we will release conditional models in a later update. Below are some examples of how to draw samples.
The default command used to draw all-atom samples (for example, for 2 proteins at each length in range(80, 100, 5)):
python draw_samples.py --type allatom --minlen 80 --maxlen 100 --steplen 5 --perlen 2
To draw 8 samples per length for lengths in range(70, 150, 5) from the backbone-only model, with 100 denoising steps instead of the default, run this command. (This is illustrative; I would not expect to get quality samples with 100 denoising steps.)
python draw_samples.py --type backbone --param n_steps --paramval 100 --minlen 70 --maxlen 150 --steplen 5 --perlen 8
We have also added the ability to provide an input PDB file and a list of (zero-indexed) indices to condition on from the PDB file. Note also that current models are single-chain only, so multi-chain PDBs will be treated as single chains (we intend to release multi-chain models in a later update). We can expect it to do better or worse depending on the problem (better on easier problems such as inpainting, worse on difficult problems such as discontiguous scaffolding). Use this command to resample the first 25 and 71st to 80th residues of my_pdb.pdb.
python draw_samples.py --input_pdb my_pdb.pdb --resample_idxs 0-25,70-80
For more control over the sampling process, including tweaking the sampling hyperparameters and more specific methods of conditioning, you can directly interface with the model.sample() function; we have provided examples of how to configure and run these commands in sampling.py.
Note (Sep 2023): the lab has decided to collect usage statistics on people interested in training their own versions of Protpardelle (for funding and other purposes). To obtain a copy of the repository with training code, please complete this Google Form - you will receive a link to a Google Drive zip which contains the repository with training code. After publication, the plan is to include the full training code directly in this repository.
Pretrained model weights are provided, but if you are interested in training your own models, we have provided training code together with some basic online evaluation. You will need to create a Weights & Biases account.
The dataset can be downloaded from CATH, and the train/validation/test splits used can be downloaded with
wget http://people.csail.mit.edu/ingraham/graph-protein-design/data/cath/chain_set_splits.json
Some PDBs in these splits have since become obsolete; we manually replaced these PDBs with the files for the updated/new PDB IDs. The dataloader expects text files in the dataset directory named 'train_pdb_keys.list', 'eval_pdbs_keys.list', and 'test_pdb_keys.list' which list the filenames associated with each dataset split. This, together with the directory of PDB files, is sufficient for the dataloader.
The main entry point is train.py; there are some arguments to control computation, experimenting, etc. Model-specific training code is kept separate from the training infrastructure and handled by the runner classes in runners.py; model-related hyperparameters are handled by the config file. Using configs/backbone.yml trains a backbone-only model; configs/allatom.yml trains an all-atom model, and configs/seqdes.yml trains a mini-MPNN model. Some examples:
The default command (used to produce the saved model weights):
python train.py --project protpardelle --train --config configs/allatom.yml --num_workers 8
For a simple debugging run for the mini-MPNN model:
python train.py --config configs/seqdes.yml
To overfit to 100 data examples using 8 dataloading workers for a crop-conditional backbone model with 2 layers, in configs/backbone.yml change train.crop_conditional and model.crop_conditional to True, and then run:
python train.py --train --config configs/backbone.yml --overfit 100 --num_workers 8
Training with DDP is a bit more involved and uses torch.distributed. Note that the batch size in the config becomes the per-device batch size. To train all-atom with DDP on 2 GPUs on a single node, run:
python -m torch.distributed.run --standalone --nnodes=1 --nproc_per_node=2 --master_port=$RANDOM train.py --config configs/allatom.yml --train --n_gpu_per_node 2 --use_ddp --num_workers 8
If you find our work helpful, please cite
@article {chu2023allatom,
author = {Alexander E. Chu and Lucy Cheng and Gina El Nesr and Minkai Xu and Po-Ssu Huang},
title = {An all-atom protein generative model},
year = {2023},
doi = {10.1101/2023.05.24.542194},
URL = {https://www.biorxiv.org/content/early/2023/05/25/2023.05.24.542194},
journal = {bioRxiv}
}