Deep Kernel learning for reaction outcome prediction and optimization.

Area of Science:

• Computational Chemistry
• Machine Learning in Chemistry
• Drug Discovery

Background:

• Machine learning, particularly deep learning, is increasingly used for predicting chemical reaction outcomes.
• Deep learning models excel at learning molecular representations but lack uncertainty quantification.
• Gaussian processes (GPs) offer reliable uncertainty estimates but cannot learn features from data.

Purpose of the Study:

• To develop a novel deep kernel learning (DKL) framework integrating neural networks (NNs) and Gaussian processes (GPs).
• To predict chemical reaction outcomes with both high accuracy and reliable uncertainty estimates.
• To leverage DKL for accelerating reaction discovery through Bayesian optimization (BO).

Main Methods:

• A deep kernel learning (DKL) framework was implemented, combining the feature learning of NNs with the uncertainty quantification of GPs.
• The DKL model was trained and evaluated for reaction outcome prediction across various molecular representations.
• Uncertainty estimates from the DKL model were utilized for Bayesian optimization (BO) in reaction discovery.

Main Results:

• The DKL model demonstrated strong predictive performance for reaction outcomes, outperforming standard GPs.
• DKL achieved performance comparable to graph neural networks while providing essential uncertainty estimates.
• The uncertainty estimates enabled the effective use of DKL as a surrogate model for Bayesian optimization.

Conclusions:

• The proposed DKL framework effectively predicts reaction outcomes and provides reliable uncertainty quantification.
• DKL offers a powerful approach for accelerating chemical reaction discovery by integrating predictive accuracy and uncertainty.
• This method holds significant potential for advancing automated reaction discovery workflows.