Kernel Methods for Predicting Yields of Chemical Reactions.

Area of Science:

• Organic Chemistry
• Computational Chemistry
• Machine Learning

Background:

• Predicting chemical reaction yield is crucial for synthesis optimization.
• Traditional methods often rely on computationally intensive quantum chemical calculations.
• Emerging machine learning approaches offer potential for faster predictions.

Purpose of the Study:

• To evaluate the efficacy of support vector regression (SVR) for predicting reaction yields.
• To compare the performance of structure-based molecular descriptors against quantum chemical descriptors in SVR models.
• To assess the generalizability of these models for Buchwald-Hartwig amination reactions.

Main Methods:

• Utilized combinatorial data for training and evaluating support vector regression (SVR) models.
• Employed structure-based molecular descriptors (fingerprints, graphs) and quantum chemical descriptors.
• Assessed model performance across different reaction components of Buchwald-Hartwig amination.
• Validated model generalizability through prospective predictions on unseen reactions.

Main Results:

• SVR models built with structure-based descriptors outperformed those using quantum chemical descriptors.
• Structure-based models showed superior predictive ability along each reaction component dimension.
• Model applicability was consistent with similarity to the training dataset.

Conclusions:

• Structure-based descriptors are a viable and efficient alternative to quantum chemical calculations for reaction yield prediction using SVR.
• The developed SVR models demonstrate good generalizability for predicting Buchwald-Hartwig amination reactions, particularly concerning aryl halide variations.
• This approach facilitates faster and more accessible reaction optimization in synthetic chemistry.