Explaining What Matters: Faithfulness in Molecular Deep Learning

Deep learning is accelerating drug and molecule discovery. Yet, the 'opaque' nature of these models limits their reliability for hypothesis-driven research, motivating the need for explainable AI (xAI). Although xAI is gaining traction in the molecular sciences, systematic benchmarking is rare. Moreover, faithfulness-the extent to which an explanation reflects a model's internal reasoning-remains overlooked. Here, we address this gap by evaluating the faithfulness of eight xAI feature attribution methods across two molecular representations and five neural network architectures on ≈14,000 molecules. We find that (a) some xAI methods are inherently more faithful than others, regardless of the architecture, and (b) faithfulness can still vary markedly on a per-molecule basis. Our findings show that faithfulness evaluation has an untapped potential to reveal model behavior, including biases and shortcuts. To support a broad adoption, the methods, metrics, and results of this study have been provided as an open-access platform named MolFaith, available at the following URL: https://github.com/molML/MolFaith. We expect our study to advance molecular xAI by enabling more informed use of feature attribution methods and propel the development of more faithful xAI approaches.