Predicting Ligand-Induced Conformational Selection: Co-Folding, Generative Docking, and the Pharmacophore Understanding Problem

Abstract

Many drug targets adopt distinct conformational states depending on the bound ligand, such as kinase DFG-in/out switching and nuclear receptor helix 12 repositioning, yet predicting which conformation a given ligand will induce remains an open challenge. Here we benchmark four computational paradigms, rigid docking (DiffDock, Gnina), flexible refinement (DynamicBind), co-folding (Boltz-2), and generative sequence-based docking (FlowDock), for ligand-induced conformational selection across ABL, BRAF V600E, p38α, and ERα (8 protein-ligand systems, 29 crystal H-bond targets). Boltz-2 achieves the highest interaction fidelity (86% exact crystal Hbond reproduction), followed by DynamicBind (52%), FlowDock (45%), DiffDock (45%, Type I only), and Gnina (0%). DynamicBind and FlowDock also successfully predict DFG-out from DFG-in inputs, while rigid docking fails entirely for Type II inhibitors. To test whether cofolding success reflects memorization or chemical understanding, we systematically deconstruct sorafenib into seven fragments, none of which exist in the PDB as BRAF-bound complexes. Boltz-2 maintains DFG-out through extensive modifications, including urea-to-amide substitution (S4), halogen removal, and a bare-phenyl back-pocket, switching to DFG-in only when the back-pocket tail is entirely removed (S6, S7). This topology-level tolerance indicates that the model has learned the hinge-to-back-pocket scaffold as the DFG-out determinant, rather than a narrow urea-specific pattern. At the interaction level, S4 specifically disrupts the αC-helix Glu hydrogen bond, showing that the structural requirements for DFG-out induction are learned independently from the specific H-bonds that stabilize it. FlowDock shows the same pattern with reduced tolerance, reaching the transition boundary one step earlier (S5). These results suggest that modern deep learning methods capture the chemical logic of conformational selection beyond memorized protein-ligand associations.