Atomic Uncertainty Pinpoints Critical Failure Structures for Trustworthy Molecular Property Prediction.

Area of Science:

• Computational chemistry
• Machine learning in drug discovery
• Molecular property prediction

Background:

• Deep learning models for molecular properties lack reliable uncertainty estimates, especially at the atomic level.
• Current methods use global molecular uncertainty, masking localized prediction ambiguities.
• Accurate uncertainty quantification is crucial for trustworthy drug discovery.

Purpose of the Study:

• Introduce AUINet, an atomic uncertainty-aware iterative network.
• Quantify and refine uncertainty at the atomic scale for improved molecular property prediction.
• Enhance the interpretability and reliability of deep learning models in cheminformatics.

Main Methods:

• AUINet is built on a D-MPNN architecture.
• Utilizes Monte Carlo dropout for atom-level uncertainty estimation.
• Employs uncertainty-guided iterative refinement of atomic features.

Main Results:

• AUINet outperforms state-of-the-art models on molecular property benchmarks and PPI inhibitor tasks, particularly in low-data regimes.
• Atomic uncertainty serves as a more robust error signal than global molecular uncertainty.
• AUINet provides chemically interpretable insights into prediction uncertainty.

Conclusions:

• AUINet establishes a new paradigm for trustworthy molecular property prediction by localizing uncertainty at the atomic level.
• The model enhances drug discovery by providing reliable uncertainty estimates and interpretable chemical insights.
• AUINet demonstrates superior performance without extensive pretraining, making it efficient for practical applications.