Physics-Informed Machine Learning for Prediction of Ligand Binding Kinetics: Leveraging Accelerated Molecular Dynamics Dynamic Fingerprints

This SURF project developed a reproducible workflow for predicting experimental pKoff labels from features derived from Ligand Gaussian Accelerated Molecular Dynamics (LiGaMD) trajectories of HSP90–ligand systems. The workflow checks system identities and experimental labels, selects representative trajectory snapshots, constructs static and dynamic molecular features, and evaluates models on exact ligand groups withheld from training. Across 31 systems and 12 model-and-input comparisons, combining starting-structure and trajectory-derived features produced the lowest mean prediction error among the tested comparisons. Resampling uncertainty remained wide, so no final model was selected. This work predicts an experimental pKoff label and does not estimate a physical dissociation rate from accelerated simulation time.