Machine Learning-Guided Discovery of Robust Conditions for Photochemical Nickel-Catalyzed Cysteine Arylation

Abstract

Late-stage functionalization (LSF) of peptides offers direct access to structurally diverse analogues, but developing general, chemoselective protocols remains challenging due to the complexity of peptide substrates and the vast number of reaction variables. Traditional optimization methods are slow and resource-intensive, limiting the exploration of emerging chemistries. Here, we present a machine learning (ML)-guided workflow that rapidly identifies robust and scalable conditions for the nickel-catalyzed photochemical arylation of cysteine residues. Using arylthianthrenium salts as versatile electrophiles and graphitic carbon nitride as a reusable photocatalyst, we implemented a two-stage active learning strategy: initially, an embedding-based model to efficiently narrow the categorical space, followed by Bayesian optimization of continuous parameters. Importantly, this ML-guided workflow pairs seamlessly with conventional laboratory practices, eliminating the need for specialized high-throughput experimentation (HTE) equipment such as microtiter plates. This approach enabled the discovery of high-yielding, chemist-friendly conditions with minimal experimental cost. The resulting protocol shows high site- and chemoselectivity across diverse peptides, saccharides, and the (hetero)arylthianthrenium coupling partners and can be scaled to gram quantities while allowing photocatalyst recycling. More broadly, this work demonstrates how data-driven optimization can accelerate the development of general LSF methodologies, providing a transferable strategy for modern reaction discovery.