A Machine Learning Approach to Model Interaction Effects: Development and Application to Alcohol Deoxyfluorination.

Area of Science:

• Chemistry
• Data Science
• Chemical Engineering

Background:

• Machine learning (ML) is increasingly used for high-throughput experimentation (HTE) datasets.
• Modeling interaction effects between reaction components in HTE data with ML remains challenging.
• Irrelevant features can hinder ML algorithms from learning these crucial interaction effects.

Purpose of the Study:

• To develop a robust statistical modeling approach for HTE datasets that effectively captures interaction effects.
• To improve the accuracy and interpretability of models for predicting reaction outcomes.
• To facilitate the generation of new mechanistic hypotheses in chemical research.

Main Methods:

• A two-part statistical modeling strategy was proposed.
• Part 1: Classical analysis of variance (ANOVA) to identify systematic effects on reaction yield.
• Part 2: Regression of individual effects using chemistry-informed features, illustrated with a generalized additive model (GAM).

Main Results:

• The proposed method significantly improved performance over a random forest model on an alcohol deoxyfluorination dataset.
• Mean absolute error (MAE) decreased from 18% to 13%.
• Root-mean-squared error (RMSE) decreased from 22% to 17% on a validation set.

Conclusions:

• The developed statistical approach effectively models interaction effects in HTE data, outperforming common ML algorithms.
• This methodology enhances the interpretability of chemical reaction models.
• The approach aids in generating testable mechanistic hypotheses, deepening the understanding of chemical reactivity.