Chemically Intuitive Descriptors for Predicting CH Borylation Regioselectivity

We present a robust machine‐learning (ML) framework incorporating chemically intuitive descriptors to predict regioselectivity in CH borylation reactions. Our method integrates electronic parameters (p K a values, CM5 atomic charges, Fukui indices) and steric metrics (buried volume, solvent‐accessible surface area) to systematically capture the nuanced interplay between sterics and electronics. Extensive analysis demonstrates that ligand‐specific features notably enhance predictive accuracy. Among the ML models tested, tree‐based algorithms, particularly LightGBM, achieved good performance (F1‐score: 72%). This performance is within the range reported for previously published graph‐neural‐network and transformer‐based approaches, although differences in datasets and evaluation protocols preclude a strict statistical comparison. Notably, we introduce a simple, chemically intuitive predictive model using only one steric descriptor (%V bur ) and one electronic descriptor (p K a ), which predicts regioselectivity accurately in a significant subset of reactions, highlighting its practical value and interpretability. Validation on pharmaceutical intermediates underscores the efficacy and practical applicability of both the comprehensive and simplified model, positioning our approach as an accessible, interpretable, and powerful tool for regioselectivity predictions in late‐stage functionalization.