Data Science Enables Chemical Interpretation of HBC-Catalyzed Polymerization for Poly(disulfide)s

H-Bonding catalysts (HBCs) has gained widespread success in improving the selectivity of organic synthesis. The complexity of potentially engaging noncovalent interactions have imposed significant challenges in deciphering the contributing factors of these catalytic reactions. Herein, we present the use of interpretable machine learning on HBCs applied in poly(disulfide) synthesis, wherein the epiquinine-derived thiourea HBCs enable simultaneous regioselectivity control and rate acceleration over the otherwise unselective and already rapid polymerization. A training set of 28 catalysts with variation on the phenyl substituents were synthesized and tested to gather experimental observables: the apparent rate (kp) and regioselectivity (Pss). Considering the limited data size, we applied the feature-sensitive XGBoost algorithm for supervised machine learning. Upon screening over 64 potentially relevant descriptors, a reasonable fitting of the observables was established for kp (R2 = 0.76) and Pss (R2 = 0.91), and the key catalyst features necessary for achieving high reaction rates and regioselectivity were deconvoluted. The model suggests that sterically hindered, heavy atom-containing substituents or strongly electro-withdrawing groups on the HBC adversely affect the reaction rate. Substituents enhancing the electrostatic potential of the aromatic N atom are beneficial for achieving high regioregularity, while presence of ortho-substituents on the phenyl ring is unfavorable.