Machine Learning and Computer Simulation Disentangle the Fuzzy Inhibitor Binding by Hsp90

Heat shock protein 90 (Hsp90) is a key cancer drug target, yet the highly flexible N-terminal ATP-binding pocket yields seemingly conflicting crystallographic and NMR inhibitor-binding models, complicating rational design. Here we integrate >100 μs of all-atom MD simulated data and explainable machine learning to resolve the binding mechanism of the prototypical inhibitor geldanamycin (GDM) to N-Hsp90. Our initial attempt via multimicrosecond unbiased exploratory simulations, starting with solvated GDM, revealed a persistent kinetically trapped intermediate but could not capture the event of native binding. We overcame this bottleneck using a hybrid adaptive strategy that seeds short unbiased trajectories via exploitation-exploration route, uses τ-random acceleration MD for barrier exploration and dissociation acceleration, and iteratively targets sparsely populated regions. A Markov state model built on ligand-recognition and L4 conformational descriptors predicted a robust three-state landscape (U → I → B) in which a long-lived intermediate defines the dominant pathway and a distinct, potentially druggable basin, with an ∼18 μs rate-limiting transition to the native bound state. The intermediate retains a <i>loop-in</i>-like L4 ensemble resembling apo states before converting to the crystal-like <i>loop-out</i> bound basin, rationalizing the NMR two-step signature together with the crystallographic end state. Independent funnel metadynamics simulation justifies this intermediate-centric route across <i>loop-out</i>, <i>loop-in</i>, and <i>excited</i> conformations and yields binding free energies consistent with experiment. Finally, a task-specific implementation of machine-learned classifiers trained on both apo and GDM-bound ensembles offers a residue-level mechanistic attribution to the overall recognition process, with <i>conformational selection</i> dominance in L4 and adjacent helices, <i>induced-fit</i> prevalence in L3 and the buried β-sheet and mixed contributions from α6-α7. This 'residual induced-fit' framework reconciles longstanding structural discrepancies and offers an intermediate-focused blueprint for targeting dynamic proteins via extensive sampling.