Virtual Screening Using Pharmacophore Models Retrieved from Molecular Dynamic Simulations.

Pavel Polishchuk,Alina Kutlushina,Olena Mokshyna,Dayana Bashirova,Timur Madzhidov

doi:10.3390/ijms20235834

Abstract

Pharmacophore models are widely used for the identification of promising primary hits in compound large libraries. Recent studies have demonstrated that pharmacophores retrieved from protein-ligand molecular dynamic trajectories outperform pharmacophores retrieved from a single crystal complex structure. However, the number of retrieved pharmacophores can be enormous, thus, making it computationally inefficient to use all of them for virtual screening. In this study, we proposed selection of distinct representative pharmacophores by the removal of pharmacophores with identical three-dimensional (3D) pharmacophore hashes. We also proposed a new conformer coverage approach in order to rank compounds using all representative pharmacophores. Our results for four cyclin-dependent kinase 2 (CDK2) complexes with different ligands demonstrated that the proposed selection and ranking approaches outperformed the previously described common hits approach. We also demonstrated that ranking, based on averaged predicted scores obtained from different complexes, can outperform ranking based on scores from an individual complex. All developments were implemented in open-source software pharmd.

Highlights

Pharmacophore models are widely used in the early stages of drug development to identify potential hits in large datasets
Recent studies have demonstrated that pharmacophores retrieved from protein-ligand molecular dynamic trajectories outperform pharmacophores retrieved from a single crystal complex structure
Our results for four cyclin-dependent kinase 2 (CDK2) complexes with different ligands demonstrated that the proposed selection and ranking approaches outperformed the previously described common hits approach

Summary

Introduction

Pharmacophore models are widely used in the early stages of drug development to identify potential hits in large datasets. Other researchers have clustered MD trajectories to select representative pharmacophore models [5,6] which reduced computational complexity due to fewer models Such approaches depend on a chosen clustering algorithm and its tuning parameters and can overlook some less populated states, which might be important for ligand-receptor recognition. Wieder et al proposed the “common hits approach” (CHA) which requires no information about known ligands to validate and select predictive pharmacophore models [7] They proposed the use of all representative pharmacophore models retrieved from a single MD trajectory of a protein-ligand complex to rank compounds according to the number of matched models. The proposed approach to retrieve pharmacophore models from MD trajectory and virtual screening was implemented in open-source software available on GitHub (https://github.com/ci-lab-cz/pharmd)

Materials and Methods

Molecular Dynamic Simulations

Pharmacophore Model Retrieval

Virtual Screening With Ensembles of MD-Based Pharmacophore Models