Structure-Kinetic Relationship for Drug Design Revealed by PLS Model with Retrosynthesis-Based Pre-trained Molecular Representation and Molecular Dynamics Simulation

Abstract

Drug design based on their molecular kinetic properties is growing in application. Pre-trained molecular representation based on retrosynthesis prediction model (PMRRP) was trained from 501 inhibitors of 55 proteins and successfully predicted the koff values of 38 inhibitors for HSP90 protein from an independent dataset. Our PMRRP molecular representation outperforms others such as GEM, MPG, and common molecular descriptors from RDKit. Furthermore, we optimized the accelerated molecular dynamics to calculate relative retention times for 128 inhibitors of HSP90. We observed high correlation between the simulated, predicted, and experimental -log(koff) scores. Combining machine learning (ML) and molecular dynamics (MD) simulation help design a drug with specific selectivity to the target of interest. Protein-ligand interaction fingerprints (IFPs) derived from accelerated MD further expedite the design of new drugs with the desired kinetic properties. To further validate our koff ML model, from the set of potential HSP90 inhibitors obtained by similarity search of commercial databases, we identified two novel molecules with better predicted koff values and longer simulated retention time than the reference molecules. The IFPs of the novel molecules with the newly discovered interacting residues along the dissociation pathways of HSP90 shed light on the nature of the selectivity of HSP90 protein. We believe the ML model described here is transferable to predict koff of other proteins and enhance the kinetics-based drug design endeavor.