Update: Also consider checking out our new diffusion generative model, GCDM, that uses GCPNet to improve equivariant diffusion models for 3D molecule generation in multiple ways. GitHub and paper.
Update: Also consider checking out the new ProteinWorkshop benchmark which features GCPNet as a state-of-the-art geometric GNN for representation learning of 3D protein structures. GitHub and paper.
Update: Also consider checking out the new GCPNet-EMA codebase which deploys GCPNet as a fast, state-of-the-art quality assessment method for 3D protein structures. GitHub and paper.
A PyTorch implementation of Geometry-Complete SE(3)-Equivariant Perceptron Networks (GCPNets)
Table of contents
Install Mamba
wget "https://github.com/conda-forge/miniforge/releases/latest/download/Mambaforge-$(uname)-$(uname -m).sh"
bash Mambaforge-$(uname)-$(uname -m).sh # accept all terms and install to the default location
rm Mambaforge-$(uname)-$(uname -m).sh # (optionally) remove installer after using it
source ~/.bashrc # alternatively, one can restart their shell session to achieve the same resultInstall dependencies
# clone project
git clone https://github.com/BioinfoMachineLearning/GCPNet
cd gcpnet/
# create conda environment
mamba env create -f environment.yaml
conda activate gcpnet # note: one still needs to use `conda` to (de)activate environments
# install local project as package
pip3 install -e .Download data for tasks
# initialize data directory structure
mkdir -p data
mkdir -p data/ATOM3D
mkdir -p data/ATOM3D/LBA
mkdir -p data/ATOM3D/PSR
mkdir -p data/CATH
mkdir -p data/RS/final_data_splits
# fetch, extract, and clean-up preprocessed data
cd data/
wget https://zenodo.org/record/7293186/files/NMS.tar.gz
tar -xzf NMS.tar.gz
rm NMS.tar.gz
cd ../Download data for the RS task by
navigating to https://figshare.com/s/e23be65a884ce7fc8543 and downloading the three files train_RS_classification_enantiomers_MOL_326865_55084_27542.pkl, validation_RS_classification_enantiomers_MOL_70099_11748_5874.pkl, and test_RS_classification_enantiomers_MOL_69719_11680_5840.pkl. Once downloaded, please store them in data/RS/final_data_splits.
Note: The ATOM3D datasets (i.e., the LBA and PSR datasets) as well as the CATH dataset we use will automatically be downloaded during execution of src/train.py or src/eval.py if they have not already been downloaded. However, data for the NMS and RS tasks must be downloaded manually.
Train model with default configuration
# train on CPU
python3 src/train.py trainer=cpu
# train on GPU
python3 src/train.py trainer=gpuTrain model with chosen experiment configuration from configs/experiment/
python3 src/train.py experiment=experiment_name.yamlTrain a model for the ligand-binding affinity (LBA) task
python3 src/train.py experiment=gcpnet_lba.yamlTrain a model for the protein structure ranking (PSR) task
python3 src/train.py experiment=gcpnet_psr.yamlTrain a model for one of the Newtonian many-body system (NMS) tasks
python3 src/train.py experiment=gcpnet_nms_{small/small_20body/static/dynamic}.yamlTrain a model for the rectus/sinister (RS) classification task
python3 src/train.py experiment=gcpnet_rs.yamlTrain a model for the computational protein design (CPD) task
python3 src/train.py experiment=gcpnet_cpd.yamlNote: You can override any parameter from command line like this
python3 src/train.py trainer.max_epochs=20 datamodule.batch_size=64New: For tasks that may benefit from it, you can now enable E(3) equivariance (i.e., scalar-wise chirality-insensitivity) within GCPNet from the command line like this
python3 src/train.py model.module_cfg.enable_e3_equivariance=trueReproduce our results for the LBA task
lba_model_1_ckpt_path="checkpoints/LBA/model_1_epoch_205_rmse_1_352_pearson_0_612_spearman_0_609.ckpt"
lba_model_2_ckpt_path="checkpoints/LBA/model_2_epoch_188_rmse_1_349_pearson_0_607_spearman_0_613.ckpt"
lba_model_3_ckpt_path="checkpoints/LBA/model_3_epoch_186_rmse_1_355_pearson_0_604_spearman_0_598.ckpt"
python3 src/eval.py datamodule=atom3d_lba model=gcpnet_lba logger=csv trainer.accelerator=gpu trainer.devices=1 ckpt_path="$lba_model_1_ckpt_path"
python3 src/eval.py datamodule=atom3d_lba model=gcpnet_lba logger=csv trainer.accelerator=gpu trainer.devices=1 ckpt_path="$lba_model_2_ckpt_path"
python3 src/eval.py datamodule=atom3d_lba model=gcpnet_lba logger=csv trainer.accelerator=gpu trainer.devices=1 ckpt_path="$lba_model_3_ckpt_path"LBA Model 1
┏━━━━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━━━━━━━━━━━━━━━┓
┃ Test metric ┃ DataLoader 0 ┃
┡━━━━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━━━━━━━━━━━━━━━┩
│ test/PearsonCorrCoef │ 0.6116969585418701 │
│ test/RMSE │ 1.3517177104949951 │
│ test/SpearmanCorrCoef │ 0.6094594597816467 │
│ test/loss │ 1.8178036212921143 │
└───────────────────────────┴───────────────────────────┘
LBA Model 2
┏━━━━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━━━━━━━━━━━━━━━┓
┃ Test metric ┃ DataLoader 0 ┃
┡━━━━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━━━━━━━━━━━━━━━┩
│ test/PearsonCorrCoef │ 0.6073298454284668 │
│ test/RMSE │ 1.3492789268493652 │
│ test/SpearmanCorrCoef │ 0.6127785444259644 │
│ test/loss │ 1.812593936920166 │
└───────────────────────────┴───────────────────────────┘
LBA Model 3
┏━━━━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━━━━━━━━━━━━━━━┓
┃ Test metric ┃ DataLoader 0 ┃
┡━━━━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━━━━━━━━━━━━━━━┩
│ test/PearsonCorrCoef │ 0.6037312150001526 │
│ test/RMSE │ 1.35466730594635 │
│ test/SpearmanCorrCoef │ 0.5983302593231201 │
│ test/loss │ 1.8261266946792603 │
└───────────────────────────┴───────────────────────────┘Reproduce our results for the PSR task
psr_model_ckpt_path="checkpoints/PSR/model_epoch_115_localpearson_0_616_localspearman_0_532_localkendall_0_385_globalpearson_0_871_globalspearman_0_869_globalkendall_0_676.ckpt"
python3 src/eval.py datamodule=atom3d_psr model=gcpnet_psr logger=csv trainer.accelerator=gpu trainer.devices=1 ckpt_path="$psr_model_ckpt_path"PSR Model
┏━━━━━━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━━━━━━━━━━━━━━━━━┓
┃ Test metric ┃ DataLoader 0 ┃
┡━━━━━━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━━━━━━━━━━━━━━━━━┩
│ test/GlobalKendallTau │ 0.6764563159385293 │
│ test/GlobalPearsonCorrCoef │ 0.8706125931497829 │
│ test/GlobalSpearmanCorrCoef │ 0.8685365401327472 │
│ test/LocalKendallTau │ 0.3854801857584887 │
│ test/LocalPearsonCorrCoef │ 0.6155098974748984 │
│ test/LocalSpearmanCorrCoef │ 0.5341097968923214 │
│ test/loss │ 0.01934020034968853 │
└─────────────────────────────┴─────────────────────────────┘Reproduce our results for each of the NMS tasks
nms_small_model_ckpt_path="checkpoints/NMS/NMS_Small/model_epoch_9977_mse_0_0070.ckpt"
nms_small_20body_model_ckpt_path="checkpoints/NMS/NMS_Small_20Body/model_epoch_10087_mse_0_0071.ckpt"
nms_static_model_ckpt_path="checkpoints/NMS/NMS_Static/model_epoch_5159_mse_0_0073.ckpt"
nms_dynamic_model_ckpt_path="checkpoints/NMS/NMS_Dynamic/model_epoch_9825_mse_0_0173.ckpt"
python3 src/eval.py datamodule=nms datamodule.data_mode=small model=gcpnet_nms logger=csv trainer.accelerator=gpu trainer.devices=1 ckpt_path="$nms_small_model_ckpt_path"
python3 src/eval.py datamodule=nms datamodule.data_mode=small_20body model=gcpnet_nms logger=csv trainer.accelerator=gpu trainer.devices=1 ckpt_path="$nms_small_20body_model_ckpt_path"
python3 src/eval.py datamodule=nms datamodule.data_mode=static model=gcpnet_nms logger=csv trainer.accelerator=gpu trainer.devices=1 ckpt_path="$nms_static_model_ckpt_path"
python3 src/eval.py datamodule=nms datamodule.data_mode=dynamic model=gcpnet_nms logger=csv trainer.accelerator=gpu trainer.devices=1 ckpt_path="$nms_dynamic_model_ckpt_path"NMS Small Model
┏━━━━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━━━━━━━━━━━━━━━┓
┃ Test metric ┃ DataLoader 0 ┃
┡━━━━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━━━━━━━━━━━━━━━┩
│ test/RMSE │ 0.08386920392513275 │
│ test/loss │ 0.007034041918814182 │
└───────────────────────────┴───────────────────────────┘
NMS Small 20-Body Model
┏━━━━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━━━━━━━━━━━━━━━┓
┃ Test metric ┃ DataLoader 0 ┃
┡━━━━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━━━━━━━━━━━━━━━┩
│ test/RMSE │ 0.08451084792613983 │
│ test/loss │ 0.007142083253711462 │
└───────────────────────────┴───────────────────────────┘
NMS Static Model
┏━━━━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━━━━━━━━━━━━━━━┓
┃ Test metric ┃ DataLoader 0 ┃
┡━━━━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━━━━━━━━━━━━━━━┩
│ test/RMSE │ 0.08549366146326065 │
│ test/loss │ 0.00730916578322649 │
└───────────────────────────┴───────────────────────────┘
NMS Dynamic Model
┏━━━━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━━━━━━━━━━━━━━━┓
┃ Test metric ┃ DataLoader 0 ┃
┡━━━━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━━━━━━━━━━━━━━━┩
│ test/RMSE │ 0.13143806159496307 │
│ test/loss │ 0.017275962978601456 │
└───────────────────────────┴───────────────────────────┘Reproduce our results for the RS task
rs_model_1_ckpt_path="checkpoints/RS/model_1_epoch_54_accuracy_0_9873.ckpt"
rs_model_2_ckpt_path="checkpoints/RS/model_2_epoch_98_accuracy_0_9882.ckpt"
rs_model_3_ckpt_path="checkpoints/RS/model_3_epoch_70_accuracy_0_9868.ckpt"
python3 src/eval.py datamodule=rs model=gcpnet_rs logger=csv trainer.accelerator=gpu trainer.devices=1 ckpt_path="$rs_model_1_ckpt_path"
python3 src/eval.py datamodule=rs model=gcpnet_rs logger=csv trainer.accelerator=gpu trainer.devices=1 ckpt_path="$rs_model_2_ckpt_path"
python3 src/eval.py datamodule=rs model=gcpnet_rs logger=csv trainer.accelerator=gpu trainer.devices=1 ckpt_path="$rs_model_3_ckpt_path"RS Model 1
┏━━━━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━━━━━━━━━━━━━━━┓
┃ Test metric ┃ DataLoader 0 ┃
┡━━━━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━━━━━━━━━━━━━━━┩
│ test/Accuracy │ 0.9873061776161194 │
│ test/F1 │ 0.9871618151664734 │
│ test/loss │ 0.041416414082050323 │
└───────────────────────────┴───────────────────────────┘
RS Model 2
┏━━━━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━━━━━━━━━━━━━━━┓
┃ Test metric ┃ DataLoader 0 ┃
┡━━━━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━━━━━━━━━━━━━━━┩
│ test/Accuracy │ 0.9882097840309143 │
│ test/F1 │ 0.9880651235580444 │
│ test/loss │ 0.03607247397303581 │
└───────────────────────────┴───────────────────────────┘
RS Model 3
┏━━━━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━━━━━━━━━━━━━━━┓
┃ Test metric ┃ DataLoader 0 ┃
┡━━━━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━━━━━━━━━━━━━━━┩
│ test/Accuracy │ 0.9867754578590393 │
│ test/F1 │ 0.9865953326225281 │
│ test/loss │ 0.03678417205810547 │
└───────────────────────────┴───────────────────────────┘Reproduce our results for the CPD task
cpd_model_ckpt_path="checkpoints/CPD/model_epoch_735_shortppl_8_22_singlechainppl_8_60_allppl_6_06_shortrecov_33_33_singlechainrecov_32_86_allrecov_40_32.ckpt"
python3 src/eval.py datamodule=cath_cpd model=gcpnet_cpd logger=csv trainer.accelerator=gpu trainer.devices=1 ckpt_path="$cpd_model_ckpt_path"CPD Model
┏━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━┓
┃ Test metric ┃ DataLoader 0 ┃ DataLoader 1 ┃ DataLoader 2 ┃
┡━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━┩
│ test/loss │ 1.8619296550750732 │ 1.8619296550750732 │ 1.8619296550750732 │
│ test/perplexity/all │ 6.061273574829102 │ 6.061273574829102 │ 6.061273574829102 │
│ test/perplexity/short │ 8.223138809204102 │ 8.223138809204102 │ 8.223138809204102 │
│ test/perplexity/single_chain │ 8.603355407714844 │ 8.603355407714844 │ 8.603355407714844 │
│ test/recovery/all │ 0.40316206216812134 │ 0.40316206216812134 │ 0.40316206216812134 │
│ test/recovery/short │ 0.3333333432674408 │ 0.3333333432674408 │ 0.3333333432674408 │
│ test/recovery/single_chain │ 0.3285714387893677 │ 0.3285714387893677 │ 0.3285714387893677 │
└──────────────────────────────┴──────────────────────────────┴──────────────────────────────┴──────────────────────────────┘GCPNet builds upon the source code and data from the following projects:
We thank all their contributors and maintainers!
If you use the code or data associated with the GCPNet models within this package or otherwise find such work useful, please cite:
@article{morehead2024gcpnet,
title={Geometry-Complete Perceptron Networks for 3D Molecular Graphs},
author={Morehead, Alex and Cheng, Jianlin},
journal={Bioinformatics},
year={2024}
}