This report investigates the use of Graph Neural Networks (GNNs) for mutagenicity prediction of chemical compounds. The mutagenicity of a chemical compound is a binary attribute indicating whether the compound is likely to cause mutations in living organisms. In the context of drug discovery, this is an important task that ensures the safety of newly developed drugs.

On a high level, the prediction pipeline is described in the figure below.

In my code, I experiment with different types of node as well as edge convolutional layers. The experiment results suggest that global aggregation, i.e., taking into account all nodes/edges when computing given node's representation are more effective than local aggregation, i.e., taking into account only the node's neighbors. As part of interpretability of the model, I also used Integrated Gradients method to visualize how much each node/edge contributes to the final prediction.

Read more details in the report which describes the problem, the dataset, the model, and the results in detail.

We recommend using conda to manage your python environment. To create a new environment, run the following command:

conda create -n mutagenicity-prediction python=3.10

And then activate the environment:

conda activate mutagenicity-prediction

To install the required packages, using conda, run the following command:

conda install --file requirements.txt

Otherwise, using pip, run the following command:

pip install -r requirements.txt