Interpretable machine learning framework for predicting the reactivity of trifluoromethylating reagents

Abstract

The incorporation of a trifluoromethyl (–CF₃) group into organic frameworks profoundly influences their physicochemical and biological properties, making trifluoromethylating reagents indispensable in pharmaceutical, agrochemical, and materials research. The reactivity of these reagents is quantified by the trifluoromethyl cation-donating ability (TC⁺DA), typically determined through density functional theory (DFT) calculations, which are accurate but computationally intensive. In this work, we present a machine learning (ML)-based quantitative structure–property relationship (QSPR) framework to predict TC⁺DA values of electrophilic trifluoromethylating reagents, encompassing chalcogenium salts, sulfoximines, and hypervalent iodine reagents. A total of 1826 molecular descriptors were generated using Mordred, and feature reduction techniques refined the dataset to 249 descriptors. Comparative analysis of multiple ML algorithms identified neural networks (NN) and extra trees (ET) as the most effective models. A feature importance study further revealed that just five descriptors (nBase, EState_VSA8, ATSC3i, AATS0v, and PEOE_VSA13) were sufficient to achieve high predictive accuracy. The optimized NN model, trained with these five descriptors and tuned hyperparameters, achieved excellent performance (R² = 0.956, RMSE = 3.43 on the test set), outperforming the ET model. Post-hoc analysis highlighted challenges in predicting hypervalent iodine reagents due to their bimodal TC⁺DA distribution. Overall, this study demonstrates the potential of ML approaches as efficient, accurate, and interpretable alternatives to quantum chemical methods, enabling rapid screening and design of new trifluoromethylating reagents.