Predicting accurate protein-ligand binding affinity is important in drug discovery. This code implements fusion network model to benefit from Spatial Grach CNN and 3D CNN models to improve the binding affinity prediction. The code is written in python with Tensorflow and Pytorch.
- Tensorflow 1.14 or higher (but not 2.X
- PyTorch 1.4 or higher
- PyTorch Geometric Feature)
- rdkit
- pybel
- pdbfixer
- tfbio
The implemented networks use a 3D atomic representation as input data in a Hierarchical Data Format (HDF5). Each complex/pocket data is comprised of a list of atoms with their features including 3D coordinates of the atoms (x, y, z) and associated features such as atomic number and charges. For more detail, please refer to the paper in the Citing LIST section.
To train or test 3D-CNN, run model/3dcnn/main_3dcnn_pdbbind.py. Here is an example comand to test a pre-trained 3D-CNN model:
python main_3dcnn_pdbbind.py --main-dir "pdbbind_3dcnn" --model-subdir "pdbbind2016_refined" --run-mode 5 --external-hdftype 3 --external-testhdf "eval_set.hdf" --external-featprefix "eval_3dcnn" --external-dir "pdbbind_2019"
To train or test SG-CNN, run model/sgcnn/src/train.py or model/sgcnn/src/test.py.
For an example training script, see model/sgcnn/scripts/train_pybel_pdbbind_2016_general_refined.sh
To train or test fusion model, run model/fusion/main_fusion_pdbbind.py
python main_fusion_pdbbind.py --main-dir "pdbbind_fusion" --fusionmodel-subdir "pdbbind2016_fusion" --run-mode 3 --external-csvfile "eval_3dcnn.csv" --external-3dcnn-featfile "eval_3dcnn_fc10.npy" --external-sgcnn-featfile "eval_sgcnn_feat.npy" --external-outprefix "eval_fusion" --external-dir "pdbbind_2019"
We trained all of the networks above on pdbbind 2016 datasets. Particularly, we used general and refined datasets for training and validation, and evaluated the model on the core set (see sample_data/core_test.hdf).
Checkpoint files for the models can be found here
To contribute to FAST, please send us a pull request. When you send your request, make develop the destination branch on the repository.
0.1
FAST was created by Hyojin Kim ([email protected]), Derek Jones ([email protected]), Jonathan Allen (allen99.gov).
This project was supported by the American Heart Association (AHA) project (PI: Felice Lightstone).
If you need to reference FAST in a publication, please cite the following paper:
Derek Jones, Hyojin Kim, Xiaohua Zhang, Adam Zemla, William D. Bennett, Dan Kirshner, Sergio Wong, Felice Lightstone, and Jonathan E. Allen, "Improved Protein-ligand Binding Affinity Prediction with Structure-Based Deep Fusion Inference", arxiv 2020.
FAST is distributed under the terms of the MIT license. All new contributions must be made under this license.
SPDX-License-Identifier: MIT
LLNL-CODE-808183