Explainable uncertainty quantifications for deep learning-based molecular property prediction.

Area of Science:

• Computational Chemistry
• Machine Learning
• Data Science

Background:

• Quantifying uncertainty is crucial in machine learning, especially in data-scarce research areas.
• Deep learning models for molecular property prediction require robust uncertainty estimation.
• Existing methods often lack interpretability and fail to pinpoint sources of uncertainty.

Purpose of the Study:

• To develop an explainable uncertainty quantification (UQ) method for deep learning in molecular property prediction.
• To separately capture aleatoric and epistemic uncertainties.
• To attribute quantified uncertainties to individual atoms within a molecule.

Main Methods:

• Developed an atom-based uncertainty quantification approach for deep learning models.
• Implemented a method to distinguish between aleatoric and epistemic uncertainties.
• Proposed a post-hoc calibration technique for ensemble model uncertainty estimates.

Main Results:

• The atom-based UQ method provides chemical insights by attributing uncertainty to specific atoms.
• Atomic uncertainty effectively detects novel chemical structures and noisy data.
• The post-hoc calibration improved confidence interval estimates for ensemble models.

Conclusions:

• This work enhances uncertainty calibration in deep learning for molecular property prediction.
• The proposed framework enables assessment of prediction reliability and its causes.
• Atom-level uncertainty attribution offers a powerful tool for chemical interpretation.