Structural activity prediction models recover known binding modes (Poster abstract)

Michael Backenköhler, Joschka Groß, Paula Linh Kramer, Verena Wolf, Andrea Volkamer

Keywords: Binding affinity prediction, explainable AI, GNN, docking

Abstract: Drug discovery can benefit from machine learning on 3D structural data of protein-ligand (PL) complexes, but a shortage of such data limits model training. For kinase targets, we generated a large in-silico dataset, kinodata-3D, using template docking. This dataset improved binding affinity predictions. Using an E(3)-invariant GNN model, we investigated learned protein-ligand interactions by removing spatial edges between protein and ligand atoms. Significant prediction changes in known binding regions confirmed the model's understanding of binding mechanisms. This approach aims to enhance explainable AI (XAI) methods, aiding the discovery of novel kinase binding mechanisms and improving model transparency.

Submission Number: 122