Knowledge discovery and prediction modeling of protein-drug binding kinetics by integrating machine learning, normal mode analysis and molecular dynamics simulation

One of the unaddressed challenges in drug discovery is that drug potency measured by protein-ligand binding affinity, such as IC 50 and Kd in vitro, is not correlated with drug activity in vivo. Computational modeling is playing an increasing role in designing efficient therapeutics. However, existing computational methods for the high-throughput study of protein-ligand interactions (PLI) mainly focus on the prediction of the binding affinity. This is the combined effect of association (k on) and dissociation (koff) rate constants. Few works have been produced to predict koff or its reciprocal, residence time, which is a key measuring function of drug efficacy in vivo. This study addresses the unmet need of the accurate and scalable prediction of kon and koff simultaneously. The fundamental strategy of our method is to develop a machine learning model using PLI kinetic features computed by normal mode analysis (NMA). To test our method, HIV-1 protease complex was used as a model system. There are three major findings of this study. First, kinetic properties are more important than thermal dynamic characteristics in determining protein-ligand binding kinetics. We propose that coherent conformational dynamics coupling between protein and ligand were proven to be more significant than pairwise residue binding energy in the prediction of kinetic rate constants. Second, NMA is an efficient method to capture conformational dynamics features for the large scale modeling of protein-ligand binding. Third, multi-target classification as well as multi-target regression, is a potentially valuable tool for modeling PLI kinetics. With the rapid increase of PLI kinetics data, the further improvement of proposed computational methodology may provide a powerful tool for large-scale modeling of PLI kinetics, thereby accelerating drug discovery process.