Data-driven design of chiral covalent fragments using high-throughput chemoproteomics and machine learning

A significant barrier in translating biological insights into therapeutic targets is the limited availability of high-quality chemical probes for target validation. Chemoproteomic profiling of covalent small molecules has dramatically accelerated the discovery of ligandable binding sites across the human proteome. However, the limited specificity and selectivity of initial hits often hinders their effectiveness in evaluating the functional consequences of ligand binding. To address this challenge, we developed a data-driven strategy that integrates chemoproteomic profiling of enantiomerically pure pairs of cysteine-targeting electrophilic fragments (enantiopairs) with machine learning (ML) to design fragment libraries optimised for proteome-wide selectivity. ML-guided library evolution produced a second generation enantiopair library markedly enriched in selective and stereospecific interactions relative to the first generation library. This approach identified high-quality enantioselective binding events with 205 cysteines, the majority not previously liganded. These findings establish a general framework for designing covalent fragment libraries to deliver higher-quality initial hits.