Defined by Shape: Elucidating the Molecular Recognition of Dynamic Loops with Covalent Ligands

Protein loops harness conformational heterogeneity to perform an array of functions, ranging from catalyzing enzymatic reactions to communicating allosteric signals. Although attractive targets for small molecule modulation, these functional hubs are often considered unligandable due to their lack of well-defined binding pockets and highly dynamic structure. Recent studies, however, have demonstrated the power of covalent chemistry to selectively capture cryptic pockets formed by protein loops. Herein, we leverage machine learning to elucidate the molecular basis of covalent ligand:loop recognition in the transcriptional coactivator Med25. Key to our success was classification by ligand shape prior to model training, which led to descriptive and predictive models. The models were experimentally validated through the synthesis and in vitro testing of novel top-ranked ligands, revealing canonical structure-affinity relationships, including an activity cliff. Further feature analyses identified traditional topological and spatial parameters predictive of binding, and molecular modeling uncovered a potential binding pocket with at least two distinct conformations with high shape complementarity. Collectively, these findings reveal the hidden potential of dynamic loops as specific sites for covalent small molecule modulation, challenging the notion that protein loops are unligandable and demonstrating their capacity for exquisite, shape-based molecular recognition.