Modeling of cytochrome P450 metabolism with quantitative structure activity relationship and docking

Abstract

To address these challenges we have created Docking-Regioselectivity-Predictor (DR-Predictor) and Docking-Inhibition-Predictor (DI-Predictor) to predict regioselectivity and inhibition of CYP isozyme ligands. In both models, the effective docking program Autodock Vina is applied for the interaction between CYP and ligand. Interaction descriptors from multiple docked poses are collected for each ligand. Reactivity calculation is approximated by the 2D method SMARTCyp and various physically meaningful descriptors from the semi-empirical program MOPAC. Multiple instance learning algorithms have been applied to integrate descriptors from both interaction and reactivity. MIRank, the multiple instance ranking algorithm, has been applied to rank the sites of metabolism over the non-metabolism sites. Multiple instance classification algorithms such as Multiple Instance Kernel SVM (MIK-SVM) are used for classify inhibitors versus non-inhibitors. Both DR-Predictor and DI-Predictor models achieve high performance, compared to other methods. Based on our integrated framework, we also explore approaches such as elastic network modeling to model highly flexible structure of CYP 3A4. Regioselectivity models based on significantly different conformations are successfully built.

Cytochrome P450 (CYPs) are a superfamily of heme contained enzymes responsible for the oxidation of organic molecules. They are responsible for the phase I metabolism of 90% of FDA approved drugs. Computational methods for predicting CYP related properties including regioselectivity and inhibition can facilitate early stage drug development. CYP modeling can be decomposed into two essential parts: protein-ligand interactions and ligand reactivity for oxidation. Ligand-based methods assume that the local reactivity of chemical groups or atoms in a molecule largely determines the regioselectivity and descriptors for ligand alone can be sufficient for CYP-ligand interaction prediction. In recent years, structure-based methods that utilize CYP-ligand binding modes have been proposed to increase the prediction accuracy of ligand-based models. However, traditional protocols that utilize docking poses based on lowest docking scores are not effective. Multiple favorable binding conformations, instead of single conformations, may collectively determine the endpoints including regioselectivity and inhibition potency. How to extract descriptors from docking and incorporate multiple conformations remains largely unexplored.