3DProtDTA: a deep learning model for drug-target affinity prediction based on residue-level protein graphs

Article

https://doi.org/10.1039/D3RA00281K

Requirements

Python >= 3.7
PyTorch >= 1.9
PyTorch Geometric >= 2.0

Usage

Activate a python environment with PyTorch and PyTorch Geometric
Clone the repository, navigate to the cloned folder
git clone https://github.com/vtarasv/3d-prot-dta.git
cd 3d-prot-dta/
Install required packages
pip install wheel
pip install -r requirements.txt
Run the experiments
python test.py to obtain test datasets results as described in the manuscript
python test.py --datasets davis to obtain only Davis test dataset results
python test.py --datasets kiba to obtain only KIBA test dataset results
The results will be saved in the results/ folder
The log will be saved in the log/ folder
You can also launch the tuning process in the same way as described in the manuscript
python tune.py --study my_study --sampler tpe
The tuning results will be saved in the local storage sqlite:///dta_tune.db (in the same folder)

Data

See the corresponding README

Notes

It is recommended to use GPU to speed up the experiments (machines with 1 GPU perform 20 times faster on average than machines with 4 CPUs)
The training of 5 models (one per cross-validation train dataset) using one NVIDIA Tesla P100 SXM2 GPU takes about 10 and 40 hours for Davis and KIBA datasets respectively
The code is tested on Ubuntu operating system