Dynafomer

A deep learning framework for encoding molecular dynamic features to enhance protein-ligand binding affinity prediction.

Install and environment settings

Install Anaconda

This code should be run on Linux PC/server with Anaconda environment. If you have not Anaconda installed, using the following script to install Miniconda, which is a mininal version of Anaconda. Refer to Miniconda's website for more information.

wget -O Miniconda_install.sh https://repo.anaconda.com/miniconda/Miniconda3-latest-Linux-x86_64.sh
chmod +x Miniconda_install.sh
./Miniconda_install.sh
# follow the instructions to install
conda --version
# you should get the output like this
# conda 23.1.0

Install dependencies

git clone --recursive https://github.com/Minys233/Dynafomer.git
cd Dynafomer/Dynaformer
conda create -n dynaformer python=3.9 -y
conda activate dynaformer
./install.sh

This will automatically install all the dependencies for evaluating or training the model.

Using Dynaformer

The scripts in this section should be run under dynaformer conda environment created in the previous section.

Evaluation checkpoints

Refer to checkpoint/README.md to download and unzip the checkpoints to the checkpoint folder. Then, run run_evaluate.sh script to evaluate three checkpoints on CASF-2013 and CASF-2016 benchmark dataset.

Training & finetuning

Refer to Dynaformer/examples/md_pretrain/md_train.sh and Dynaformer/examples/finetune/finetune.sh to train from scratch or fintuning a pre-trained model.

Custom input

Currently, for custom input, users will need to run two scripts for graph data preparation and evaluation. We provide examples that were used in case studies of the paper. Additional dependencies are required for processing data, and use this command to install them: conda install -c conda-forge pymol-open-source openbabel -y under the dynaformer environment.

1. Run preprocess/custom_input.py to convert protein PDB files and ligand SDF files into PyG graphs.

Users need provide a CSV file with column receptor, ligand, name and pk for protein PDB file path, ligand SDF file path, custom name of the pair and the pk value as label. Also an output path is specified for store the graphs in pickle format. Note that the pk column is only used for calculating deviations/correlations between prediction and ground-truth, set it to -1 if you don't have it. Here is an example:

# Run this command in Dynaformer/preprocess directory
python custom_input.py ../example_data/example.csv ../example_data/example.pkl

The content of the example.csv file should be similar to the following template:

receptor,ligand,name,pk
../example_data/2v7a/2v7a_protein.pdb,../example_data/2v7a/2v7a_ligand.sdf,2v7a,8.30
../example_data/3udh/3udh_protein.pdb,../example_data/3udh/3udh_ligand.sdf,3udh,2.85
...(other lines)...

Note that the paths in example.csv use relative paths for demonstration purposes, but absolute paths are more recommended here to avoid file not exist errors.

2. Run run_custom_input.sh to predict the binding affinity using the model checkpoints.

# Run this command in Dynaformer (project root) directory
./run_custom_input.sh

The results should be stored in the checkpoint directory. Unlike PDBBind, for custom input files, we pick pockets on the fly rather than as input. We use pymol to select the pocket from protein, and use Openbabel to convert the pocket from PDB format to SDF format. Note: When using files from CASF as custom inputs, the predicted results will differ slightly from the Evaluation checkpoints section above. This difference is caused by the pocket files, i.e. CASF provides pocket files for each complex, but for custom input pocket files are generated at runtime.

Known issues

1. preprocess/custom_input.py: When processing data in parallel, there may be competing data writes resulting in incorrect contents of the intermediate file (SDF), resulting in failure to output .pkl results.

2. Automatic evaluation with custom protein (PDB) & ligand (SDF) is under development. For now, need to use multiple scripts.