Evidential Deep Learning for Guided Molecular Property Prediction and Discovery

Area of Science:

• Computational chemistry
• Machine learning
• Chemical informatics

Background:

• Neural networks excel in molecular modeling but face challenges with generalization, sample efficiency, and prediction calibration.
• Uncertainty quantification is crucial for reliable molecular structure-property prediction.

Purpose of the Study:

• To introduce an uncertainty quantification method for neural network-based molecular property prediction using evidential deep learning.
• To develop and evaluate evidential 2D message passing and 3D atomistic neural networks for molecular tasks.

Main Methods:

• Leveraging evidential deep learning to develop evidential 2D message passing neural networks (MPNNs) and 3D atomistic neural networks.
• Applying these networks to various molecular structure-property prediction tasks.
• Evaluating uncertainty calibration, sample efficiency via active learning, and retrospective virtual screening performance.

Main Results:

• Evidential uncertainties lead to calibrated predictions where uncertainty correlates with prediction error.
• Uncertainty-guided active learning improves sample efficiency during model training.
• Retrospective virtual screening shows improved experimental validation rates.

Conclusions:

• Evidential deep learning offers an efficient approach for uncertainty quantification in molecular property prediction.
• This method enhances the reliability and efficiency of molecular discovery and design.
• The developed evidential neural networks show promise for chemical and physical science applications.