ADME-Drug-likeness

The official code implementation for ADME-DL model from our paper, "ADME-Drug-Likeness: Enriching Molecular Foundation Models via Pharmacokinetics-Guided Multi-Task Learning for Drug-likeness Prediction".

Here, we provide data and codes for Sequential ADME Multi-task learning and Drug-likeness prediction on three datasets (DrugMAP-ZINC, DrugMAP-PubChem, DrugMAP-ChEMBL)

Model description

ADME-DL model is trained in two steps:

1. Step 1: Sequential ADME Multi-task learning which trains the grouped ADME endpoints in A-D-M-E sequential manner. The grouped data for ADME tasks are provided here.
2. Step 2: Drug-likeness prediction (DLP). This step first encodes the drug-non-drug datasets with the molecular encoder, then trains an MLP model that classifies drugs from non-drugs. The datasetsfor DLP tasks are provided here.

Setup

First, clone this repository and move to the directory.

git clone https://github.com/eugenebang/ADME-DL.git
cd ADME-DL/

To install the appropriate environment for ADME-DL, you should install conda package manager.

After installing conda and placing the conda executable in PATH, the following command will create conda environment named admedl. It will take up to 10 minutes to setup the environment, but may vary upon the Internet connection and package cache status.

conda env create -f environment.yaml && \
conda activate admedl

However, we strongly encourage you to first build a fully running virtual environment with adequate pytorch and pytorch_geometric (along with torch-scatter, torch-cluster and torch-sparse) version for your hardware settings including GPU and CUDA. Then you can install the neccessary packages listed below with pip before running the model.

To check whether ADME_DL network works properly, please refer to the codes in Scoring drug-likeness below.

Running ADME-DL

1. Scoring drug-likeness

python score-drug-likeness.py --smiles_file data/demo/demo_molecules.smi

score_drug_likeness.py file outputs the drug-likeness score computed by ADME-DL model. The input file should be a header-less .smi file of SMILES strings with linebreaks (with no column names).

The resulting file, with .smi replaced to _score.csv, will be saved on the same directory as the input file.

2. Training molecular encoder with Sequential ADME MTL

python train_ADME.py --target_task adme --data_dir data/ADME/

Running the code above yields a molecular encoder that is trained through Sequential MTL of ADME tasks in A-D-M-E order. By replacing the target_task argument with other pair or orders (e.g. EMDA or A) will run the training process in the input order.

The trained model is saved in the ckpts folder.

3. Training Drug-likeness prediction classifier

python train_DLP.py --ADMEtrained_model ckpts/SeqADME_ADME_DL.pt --data_path data/DLP/drugmap_zinc.csv

Running the code above trains the MLP classifier on drug-likeness prediction (DLP) task with drugmap_zinc dataset, with 3:1:1 train-valid-test split. The classifier is trained above the ADME-trained encoder, with its parameters stored in ckpts folder. The benchmark sets are provided in the DLP data folder.

Software requirements

Operating system

ADME-DL model training and evaluation were tested for Linux (Ubuntu 18.04) operating systems.

Prerequisites ADME-DL network training and evaluation were tested for the following python packages and versions.

• python=3.10
• pytorch=1.13.1
• pytorch-geometric=2.2.0
• rdkit-pypi=2022.09.5
• numpy=1.24.1
• pandas=2.1.1
• scipy=1.14.0
• tqdm=4.66.4

License

The source code of ADME-DL follows GPL 3.0v license, and allows users to use, modify, and distribute the software freely, even for commercial purposes.

However, any data or content produced from using ADME-DL follows CC BY-NC-SA 4.0, which does not permit commercial use without proper authorization.

References

% TDC database
@article{huang2021therapeutics,
  title={Therapeutics data commons: Machine learning datasets and tasks for drug discovery and development},
  author={Huang, Kexin and Fu, Tianfan and Gao, Wenhao and Zhao, Yue and Roohani, Yusuf and Leskovec, Jure and Coley, Connor W and Xiao, Cao and Sun, Jimeng and Zitnik, Marinka},
  journal={arXiv preprint arXiv:2102.09548},
  year={2021}
}

% DrugMAP
@article{li2023drugmap,
  title={DrugMAP: molecular atlas and pharma-information of all drugs},
  author={Li, Fengcheng and Yin, Jiayi and Lu, Mingkun and Mou, Minjie and Li, Zhaorong and Zeng, Zhenyu and Tan, Ying and Wang, Shanshan and Chu, Xinyi and Dai, Haibin and others},
  journal={Nucleic acids research},
  volume={51},
  number={D1},
  pages={D1288--D1299},
  year={2023},
  publisher={Oxford University Press}
}

% GraphMVP
@inproceedings{
    liu2022pretraining,
    title={Pre-training Molecular Graph Representation with 3D Geometry},
    author={Shengchao Liu and Hanchen Wang and Weiyang Liu and Joan Lasenby and Hongyu Guo and Jian Tang},
    booktitle={International Conference on Learning Representations},
    year={2022},
    url={https://openreview.net/forum?id=xQUe1pOKPam}
}

% PCGrad
@misc{Pytorch-PCGrad,
  author = {Wei-Cheng Tseng},
  title = {WeiChengTseng/Pytorch-PCGrad},
  url = {https://github.com/WeiChengTseng/Pytorch-PCGrad.git},
  year = {2020}
}