Molecular Machine Learning Approach to Enantioselective C–H Bond Activation Reactions: From Generative AI to Experimental Validation

Abstract

Molecular machine learning (ML) has gained considerable attention in recent years. Developing ML algorithms for chemical reaction prediction is a formidable task, due to the small-sized reaction data it often presents, besides the sparsity and skewed distribution. While previous ML studies offered effective predictions on known reactions, efforts in using deep generative models for guiding new reactions and their prospective validation are rare. We harness both predictive and explorative abilities of deep learning on an important catalytic asymmetric β-C(sp³)–H activation reaction, consisting of 220 experimentally reported examples that differs primarily in terms of the substrate, catalyst, and coupling partner. A transfer learning approach using a chemical language model, pretrained on 1 million unlabeled molecules followed by fine-tuning on this reaction data set, is adopted. Our ensemble prediction (EnP) model, where 30 fine-tuned CLMs concurrently predict the %ee of test set reactions, is highly reliable. Another language model, fine-tuned on the 77 known chiral ligands as used in the above reactions, is employed for generating novel ligands of high validity and novelty. A proof of concept wet-lab experimental validation reveals that most of the ML-generated reactions are in excellent agreement with the EnP predictions. Results also caution the prospects of ML-driven reaction development for ligand design and emphasize the importance of domain experts in key decisions.