Geometric deep learning in drug discovery

Abstract

Artificial intelligence (AI) is transforming early-stage drug discovery by enabling efficient, data-driven molecular modeling and prediction. In this thesis, we present a series of interpretable and task-specific methods grounded in Geometric Deep Learning (GDL) to advance molecular representation learning across key domains such as molecular property prediction and drug–target interaction (DTI). We introduced four original methods that address core challenges in molecular representation learning. These methods target modality alignment, multiscale feature integration, interpretable analysis, and target imbalance in regression. Together, they enable the learning of robust and geometry-aware molecular representations, supporting diverse downstream tasks under limited or imbalanced data conditions. To illustrate practical utility, we apply our models to a preliminary virtual screening task for PIN1 (peptidyl-prolyl cis-trans isomerase NIMA-interacting 1), an oncogenic driver in cancer. Using a curated dataset with potency and efficacy annotations, we define a composite activity metric and show the ability of the model to prioritize active compounds. Together, these contributions demonstrate the versatility of GDL-based approaches in addressing various molecular learning tasks and their potential to enable interpretable, highperformance AI frameworks for accelerating drug discovery.