Torch Geometric Dataset for applications involving Chemical(SMILE Strings) Interactions

The repository contains the code for an effective way of making a Molecular Graph Dataset in Torch Geometric involving a pair of graphs.

The above implementation makes use of both node features(atoms) and edge features(bonds)

The TWOSIDES Pharamacy Side Effects Dataset from TDC was used which you can refer to here.

Download Data

In order to download the Graphs, follow the following steps.

• Clone the Repository:git clone git@github.com:Deceptrax123/GNN-Dataloader-For-Chemical-Interaction-Applications.git
• Run pip install -r requirements.txt
• Save the environment variables mentioned below in a .env file
• Run Scripts/download_pipeline.py and follow the steps to effectively save the .pt files.
• You may edit the number of concurrent processes according to your system specs.

Train and Run

To train and evaluate GNN models, the following modifications need to be made to the graph_dataset.py script.

• Do not override the process() function
• Use the processed_paths property in the same way as mentioned in the docs by giving the absolute path to each .pt file. This ensures the entire dataset isnt processed again.

Featurizing nodes and edges

• Both node and edge level features have been included.
• For atoms we use features such as their hybridization, atomic mass, presence in an aromatic ring, formal charge etc
• For bonds we use features such as bond type(single,double, triple), stereochemical aspects and presence in a conjugation(Resonance)

Batching

• Since 2 graphs need to be loaded, the Batching technique has been changed. More information on this can be found here.

Environment Variables

Key	Value Description
tup_bins	The path to the location that saves the tuple binaries of reactants and labels.
graph_files	The root directory to where you want your graph .pt files to be stored