Evaluation Beyond Goodness of Fit: Biophysical Alignment of AI Models for Kinase-Centric Drug Discovery

The potential of machine learning (ML) to accelerate modern drug discovery is widely recognized, especially when using ML to reason about the binding of small drug-like molecules to proteins. In order to assess whether such ML models realize this potential, they typically undergo systematic evaluations, e.g. regarding their ability to predict protein-ligand binding affinity. Although high scores in the corresponding metrics on held-out test data are a necessary evaluation criterion, they may not provide sufficient insight. A more thorough approach includes the use of explainability techniques. In this work we evaluate kinase binding affinity models in terms of how well explanations for their predictions align with prior knowledge about biophysical binding mechanisms. Our results show that models with access to the three-dimensional arrangement of kinase-ligand complexes exhibit significant better alignment.