/Molecule-Chemical-Property-Prediction-using-Graph-Isomorphism-and-Adversarial-Pretraining

Predict the properties of molecules through adversarial pre-training and Graph Isomorphism

Primary LanguagePython

Molecular Property Prediction using Contrastive Learning and Graph Isomorphism

Architecture

Running the Scripts

  • Follow the procedure in the documentation in our graph creation repository here to save all the molecular graphs for all datasets.
  • Install the requirements by running: 
    pip install -r requirements.txt
  • The weights for the pre-trained spectral backbone can be downloaded from our release. The weights are saved as .pt files which can directly be loaded in the training scripts. The procedures used for pre-training the backbone can be found here
  • Load the pre-trained weights in the train.py file of each dataset folder. Each train file is named as per the dataset, change the name of the script accordingly before running.
  • Before running the train scripts, insert the paths for the location of the saved graph tensors and the pre-trained backbone. A detailed list of keys used in the .env file is given below.
  • Change the path for saving the model based on your convenience. Ensure that the folder is created before saving the models.
  • The saved models can be loaded in the test.py of each dataset for inference. The primary metric is ROC-AUC for classification and RMSE for regression.
  • Further, you may be required to re-set the environment path. For such cases, linux/macOS users may use:
export PYTHONPATH="/path/to/project/root"
  • Windows users are suggested to use bashterminal and run the above command.

Path Variables

Key Value
zinc_weights /path/to/zinc/pretrained.pt
graph_files path/to/graphs/data/processed/

We will release the weights for the trained models soon. You may send an email or raise an issue if there are any bugs. Have fun!!!