The goal of this project is to use Graph Neural Networks (GNNs) to predict drug potency against drug targets (e.g., an ion channel protein). Specifically, we want to compare the performance of traditional QSAR (Quantitative Structure-Activity Relationship) models with GNN-based models. Traditional QSAR methods rely on handcrafted molecular descriptors—numerical representations of molecular properties—to predict a compound’s activity. Since GNNs offer a unique approach by treating molecules as graphs (preserving connectivity, spatial relationships, and chemical properties), this would allow the model to learn molecular features directly from their structures (there have been papers on this type of application).
To make this project more biologically relevant, if time permits, we could extend the QSAR-GNN models by incorporating protein structure information into the learning process. So, instead of only modeling small molecules, one would construct Protein-Ligand Interaction Graphs, where both drug molecules and protein residues are represented as graph nodes, connected by interaction-based edges (e.g., hydrogen bonds, hydrophobic interactions). This will allow the GNN to learn not only the molecular properties of the drug but also how it interacts with drug targets.