Reducing Overconfident Errors in Molecular Property Classification using Posterior Network

This repository is the offical implementation of Reducing Overconfident Errors in Molecular Property Classification using Posterior Network.

Setup

conda create -n postnet python==3.9
conda activate postnet
conda install -c conda-forge rdkit
conda install pytorch==1.9.0 torchvision==0.10.0 torchaudio==0.9.0 cudatoolkit=11.1 -c pytorch -c conda-forge
conda install -c conda-forge tqdm
conda install -c conda-forge dgllife
conda install -c dglteam dgl-cuda11.1
conda install -c conda-forge gpytorch
pip install typed-argument-parser
pip install pyro-ppl

Reproducing

You can reproduce the results presented in our paper quickly by running the corresponding Notebooks in the notebooks folder.

Data

All the data used to generate the experimental results have been included in the data folder

Trained model checkpoints

We provide the trained model checkpoints in the trained_model folder. You can find them there and easily use these checkpoints by running the single_molecule_prediction.ipynb Jupyter Notebook file.

Supported Predictions

• hERG: Whether to inhibit hERG
• BBB: Whether it can cross the blood-brain barrier
• CYP2C9: Whether to inhibit CYP2C9
• CYP3A4: Whether to inhibit CYP3A4
• Pgp-inhibitor: Whether to inhibit P-gp

Usage

open single_molecule_prediction.ipynb

smiles = 'CCCSC1=CC2=C(NC(NC(=O)OC)=N2)C=C1' # input your molecule
task_name = 'Pgp-inhibitor' # choose your prediction task

single_molecule_prediction(smiles, task_name)

Acknowledgements

This repository includes code that is modified from Chemprop. We would like to acknowledge the developers of Chemprop for their valuable contributions to this field.