A Bayesian graph convolutional network for reliable prediction of molecular properties with uncertainty...

Area of Science:

• Computational chemistry
• Machine learning in chemistry

Background:

• Deep neural networks (DNNs) are data-driven, making their performance in data-deficient chemical scenarios highly uncertain.
• Uncertainty in DNN predictions can lead to unreliable decision-making in chemical research and development.

Purpose of the Study:

• To demonstrate how Bayesian inference can enhance prediction reliability and provide quantitative uncertainty analysis in chemical applications.
• To develop and evaluate a Bayesian graph convolutional network (GCN) for molecular property predictions.

Main Methods:

• Implemented a Bayesian graph convolutional network (GCN) model.
• Applied the Bayesian GCN to molecular property prediction tasks, including bio-activity, toxicity classification, and log P prediction.
• Decomposed predictive uncertainty into model- and data-driven components.

Main Results:

• Bayesian GCNs offer more reliable predictions with quantifiable uncertainty compared to standard GCNs.
• Uncertainty quantification enables more accurate virtual screening of drug candidates.
• Identified data artefacts in the Harvard Clean Energy Project dataset by analyzing data-driven uncertainty, showing its sensitivity to data noise.

Conclusions:

• Bayesian inference is crucial for robust molecular applications, particularly in data-deficient settings.
• The developed Bayesian GCN is a valuable tool for improving the accuracy and reliability of molecular property predictions.
• Decomposing uncertainty aids in understanding model limitations and improving data quality.