Using Dynamic Interaction Fingerprints to Derive Baseline Machine Learning Model for the Prediction of Protein-Ligand Dissociation Rate Constant

Model building for the prediction of protein-ligand unbinding kinetics gaining popularity with the increasing availability of experimental structural data for the protein-ligand complexes and their relevant kinetic parameters. Limited but major effort has been already put forward in choosing appropriate machine learning (ML) methods among the popular ones like least squares (LS), support vector machine (SVM), random forest (RF), and a few more. The RF and Bayesian neural network (BNN) algorithms have been reported to be promising when combined with advanced descriptors representing ligand properties and protein-ligand interactions. However, the selection of descriptors that would correlate well with the unbinding kinetic properties is still a challenge. In this work, we derived a baseline RF model using descriptors representing the protein-ligand interaction fingerprints (IFPs) along the ligand unbinding pathway otherwise can be called dynamic IFPs. We found that the dynamic IFPs in addition to the static or binding pocket IFPs significantly improved the quality of our model for the prediction of ligand dissociation rate constant (koff). To the best of our knowledge, this work is the first attempt towards using the dynamic IFPs in deriving a quantitative structure-kinetics relationship (QSKR) model for the prediction of koff.