Atomic Uncertainty Pinpoints Critical Failure Structures for Trustworthy Molecular Property Prediction

Abstract

Accurate prediction of molecular properties is essential for accelerating drug discovery, yet current deep learning methods generally lack reliable uncertainty estimates, particularly at the atomic level. A limitation of current approaches is the reliance on global molecular uncertainty, which frequently masks localized ambiguities by averaging the uncertainty across the entire molecule. To address this, we introduce AUINet, an atomic uncertainty-aware iterative network that quantifies and refines uncertainty at the atomic scale. Built on a D-MPNN architecture, AUINet uses Monte Carlo dropout to estimate atom-level uncertainty and iteratively refines atomic features through uncertainty-guided updates. Comprehensive evaluations show that AUINet outperforms state-of-the-art models on molecular property benchmarks and protein-protein interaction inhibitor tasks under low-data conditions, all without requiring extensive pretraining. Crucially, rejection sampling experiments reveal that atom uncertainty provides a more robust error signal than classic molecular uncertainty. More importantly, AUINet provides chemically interpretable insights by pinpointing specific functional groups and structural motifs that contribute most to prediction uncertainty, as validated in solubility prediction and activity-cliff analysis. Overall, AUINet's precise localization of atomic-level uncertainty establishes a new paradigm for trustworthy molecular property prediction.