Abstract
The dissociation or off rate, koff, of a drug molecule has been shown to be more relevant to efficacy than affinity for selected systems motivating the development of predictive computational methodologies. These are largely based on enhanced-sampling molecular dynamics (MD) simulations that come at a high computational cost limiting their utility for drug design where a large number of ligands needs to be evaluated. To overcome this, presented is a combined physics- and machine learning (ML)-based approach that uses the physics-based site-identification by ligand competitive saturation (SILCS) method to enumerate potential ligand dissociation pathways and calculate ligand dissociation free energy profiles along those pathways. The calculated free energy profiles along with molecular properties are used as features to train ML models, including tree- and neural network approaches, to predict koff values. The protocol is developed and validated using 329 ligands for thirteen proteins showing robustness of the ML workflows built upon the SILCS physics-based free energy profiles. The resulting SILCS-Kinetics workflow offers a highly efficient method to study ligand dissociation kinetics, providing a powerful tool to facilitate drug design including the ability to generate quantitative estimates of atomic and functional groups contributions to ligand dissociation.
Supplementary materials
Title
Supporting Information
Description
Simulation run times for the tested proteins; experimental koff data set; hyperparameters used for tunning of ML models; correlation analyses for all combinations of energy barrier calculation protocols and different pathway cluster selection methods; distribution of the experimental off rate values in the training and test data sets; illustrations of slab determination along pathway and barrier models for ligand unbinding process as well as energy profiles obtained from SILCS-MC for one protein; correlation plots for energy barriers and experimental off rates; feature importance of RF model and correlation between MW and experimental off rates; correlation plots for ML predicted versus experimental off rates; correlation plot for the two ML model predictions and for consensus prediction versus experimental data; correlation plots for an ensemble model based on 10 independent RF trainings versus experimental data; extra analyses on dissociations of two inhibitors targeting CDK2 protein, ablation studies on the RNN+MLP model.
Actions