Graph Neural Network for Predicting Molecular Properties

This project evaluates different Graph Neural Network(GNN) architectures for their effectiveness in predicting the Quantum Mechanical properties of chemical molecules.

Requirements

We use PytorchLightning and PytorchGeometric as development frameworks and Weights & Biases for experiment management.

Setup

Install dependencies by running

grep -v '^#' requirements.txt | xargs -n 1 -L 1 pip install --default-timeout=100 --no-cache-dir

Set Python path using following command

export PYTHONPATH=.

If you are using Google Colab or Jupyter Notebook you can set the environment using

%env PYTHONPATH=.:$PYTHONPATH

Run

To run DAGNN model use the following command

python training/experiment.py --model_class=MXMNet --data_class=PyG_QM9 --gpus='0,' --num_workers=4 --target=7 --lr=0.0001 --n_layer=2 --dagnn=True

To run the baseline model

python training/experiment.py --model_class=MXMNet --data_class=PyG_QM9 --gpus='0,' --num_workers=4 --target=7 --lr=0.0001 --n_layer=6

To run the model with Virtual Node

python training/experiment.py --model_class=MXMNet --data_class=PyG_QM9 --gpus='0,' --num_workers=4 --target=7 --lr=0.0001 --n_layer=6 --virtual_node=True

To run the model with Auxiliary Layer

python training/experiment.py --model_class=MXMNet --data_class=PyG_QM9 --gpus='0,' --num_workers=4 --target=7 --lr=0.0001 --n_layer=4 --auxiliary_layer=True

• If you are using Weights & Biases to track experiment, add --wandb flag as an argument.

• To perform distributed trainining with multiple gpus, add gpu card numbers like --gpus=0,1,2,3 and accelerator as --accelerator=ddp

References

• The base template of this codebase is taken from https://github.com/full-stack-deep-learning/fsdl-text-recognizer-2021-labs
• MXMNet Implementation is taken from https://github.com/zetayue/MXMNet
• Auxiliary Layer implementation is inspired from https://github.com/rasbt/machine-learning-book/blob/main/ch18/ch18_part2.py
• Virtual Node and DAGNN Layer implementation is taken from https://github.com/divelab/MoleculeX/blob/master/BasicProp/kddcup2021/deeper_dagnn.py