Bridging Islands in Chemical Space: Evaluating and Enhancing ML Generalization for Drug Discovery

Oral presentation at the EUROPIN Summer School in Drug Design, summarizing our work on understanding and improving the generalization of machine learning models in chemical space — covering how out-of-distribution data should be defined for molecular property prediction, the limits of in-distribution-based model selection, and practical recommendations for bioactivity and ADMET tasks.

Authors: Hosein Fooladi and Johannes Kirchmair.