MPCD: A Multitask Graph Transformer for Molecular Property Prediction by Integrating Common and Domain KnowledgeArticle link copied!

MPCD is a newly developed method for evaluating the predictive performance on ADMET, physicochemical, and activity cliff compounds of machine learning models.

Environment

Install via conda yaml file

conda env create -f environment.yml
conda activate mpcd 

Install manually

conda create -n resgen python=3.10
conda install pytorch==1.12 cudatoolkit=11.3 -c pytorch -c conda-forge
conda install pyg -c pyg
conda install -c conda-forge rdkit
conda install biopython -c conda-forge
conda install pyyaml easydict python-lmdb -c conda-forge

Use existing model and data!

Here is a quick use of validation. Using the following command:

cd exps/checkpoint
wget https://drive.google.com/file/d/1QsZDWIuUno6uTZcDmSNTJISyPLaGumYy/view?usp=sharing
cd ../..
sh examples/val/val.sh

Then the output will be:

Data

Download data

The pretraining data for MPCD is ChEMBL: https://www.ebi.ac.uk/chembl/.

The fine-tuning datasets include two datasets:

1. The ADMET and physicochemical datasets from ADMETlab2.0: https://admetmesh.scbdd.com/
2. The activity cliffs datasets from MoleculeACE: https://github.com/molML/MoleculeACE

Download and put them under file: dataset. You can name them by customer name.

Data pre-process

After download data with csv format, you should pre-process them into pickle file and split them into train, test, and validation set.

python data_process/smiles_to_graph.py

Train

MPCD is first pre-trained with ChEMBL dataset in dataset/ChEMBL.

Pre-training with ChEMBL dataset

sh examples/pre_train.sh

Fine-tuning

The training process and configs are released as train.sh, the following command is an example of how to train a model. The command in train.sh includes the training process of ADMET, physicochemical, and activity cliffs datasets.

sh examples/train.sh

Also, the validation command includes the Python file and configs of ADMET, physicochemical datasets and activity cliffs datasets

Validataion

Evaluate the performance of MPCD.

sh examples/val.sh

Also, you can skip the pre-training， and directly use the pre-trained .cpk file to load the parameter.

Note that MPCD can be ready for any customer data set, as long as you use the propoess_data.py to transform your data into pickle format and load the pretraining model. Then follow the same fine-tuning process.

License

MPCD is under MIT license. For use of specific models, please refer to the model licenses found in the original packages.