Interpretable machine learning framework for predicting the reactivity of trifluoromethylating reagents

Area of Science:

• Organic Chemistry
• Computational Chemistry
• Medicinal Chemistry

Background:

• Trifluoromethyl groups are crucial in pharmaceuticals, agrochemicals, and materials.
• Quantifying trifluoromethylating reagent reactivity via trifluoromethyl cation-donating ability (TC+DA) is essential.
• Traditional TC+DA determination uses computationally intensive DFT calculations.

Purpose of the Study:

• Develop a machine learning (ML)-based quantitative structure-property relationship (QSPR) framework to predict TC+DA.
• Enable rapid screening and design of novel electrophilic trifluoromethylating reagents.

Main Methods:

• Generated 1826 molecular descriptors using Mordred.
• Reduced features to 249 descriptors.
• Trained and compared various ML algorithms, including neural networks (NN) and extra trees (ET).

Main Results:

• Identified NN and ET as the most effective ML models.
• Discovered that only five key descriptors were sufficient for high predictive accuracy.
• The optimized NN model achieved R² = 0.956 and RMSE = 3.43 on the test set.
• Hypervalent iodine reagents presented prediction challenges due to bimodal TC+DA distribution.

Conclusions:

• ML-based QSPR models provide an efficient, accurate, and interpretable alternative to DFT for predicting TC+DA.
• This framework facilitates the accelerated discovery and development of new trifluoromethylating reagents.