A Virtual Screening Filter for Identification of Cytochrome P450 2C9 (CYP2C9) Inhibitors

Abstract

Cytochrome P450 2C9 (CYP2C9) is one of the most important phase 1 metabolizing enzymes in humans for therapeutically relevant pharmaceuticals. Any new compound inhibiting this membrane-associated protein would notably affect the metabolism of physiologically important molecules and drugs, resulting in clinically significant drug-drug interactions. In search for computational tools to identify potential CYP2C9 inhibitors early in discovery, we present here the construction of filters based on 1100 structurally diverse molecules tested for CYP2C9 inhibition under identical conditions. Their chemical structures were encoded using various 2D descriptors including Three-Point Pharmacophoric (3PP) fingerprints, followed by generation of statistical models using Support Vector Machines (SVM) and Partial Least Squares (PLS). This consistently led to significant and predictive models for regression and classification of CYP2C9 inhibitors. Their predictive ability was underscored by successfully applying them to different test sets of 238 diverse and 344 GPCR-targeted compounds. Even more important for early drug discovery is the ability of these models to correctly discriminate CYP2C9 inhibitors from inactive molecules. These models collectively are able to identify true CYP2C9 inhibitors with relatively low rates of false positives. The 3PP-based filter also allows visualizing important substructures and functional groups, which are linked to proteinligand interactions for CYP2C9, as illustrated for selected structure-activity series. The application of these models to the substrate S-warfarin, recently co-crystallized with CYP2C9, revealed that the important substructures are indeed involved in the interaction with the CYP2C9 binding site. For example, the model correctly indicated aromatic stacking interactions with Phe114 and Phe476 as well as a hydrogen bond with Phe100. Hence, these models consistently provide guidelines for reducing CYP2C9 inhibition in novel candidate molecules.

Generation of in-silico cytochrome P450 1A2, 2C9, 2C19, 2D6, and 3A4 inhibition QSAR models

“…The best available literature comparison to these models can be found in the work by Byvatov et al [55] who built an inhibition model for the 2C9 isoform using support vector machines in conjunction with pharmacophoric fingerprints. While one cannot accurately compare r 2 s between datasets where the dataset activity variances are different, the r 2 of 0.55-0.63 obtained by Byvatov et al on a diverse test set, or 0.36-0.…”

The importance of the domain of applicability in QSAR modeling

“…The advantage of this approach was that the terms in the equation were generally simple and easily interpretable, while the kinds of molecules being predicted were generally very similar to those that were already synthesised, giving the user greater confidence in the model predictions. In contrast, over the past decade an increasing number of QSARs have been reported based on large, diverse datasets, commonly termed global models, which are considered more reliable at predicting diverse structures than QSARs built on small datasets of low diversity [9][10][11][12][13]. These models are often built using complex statistical methods, and large numbers of often sparsely populated geometrical and electrotopological descriptors [14][15][16][17], and while this may allow for a more versatile description of molecular structure and a reliable way to relate structure to activity, the multidimensional space defined by such a model will become increasingly complex and fragmented.…”

On the interpretation and interpretability of quantitative structure–activity relationship models

“…Indeed, it is highly probable that there is no 'best' solution to set composition, feature selection and modeling technique. Given these observations, it is all the more important to extract as much as we can from a QSAR model, though it is also true that all QSAR models are not designed to provide insight into the underlying SARs (such as filtering [12,18,25,55,87,109] models). In these cases, the underlying structureactivity relationship is usually well known.…”

“…The answer to this depends on the planned use of the model. For example, many QSAR models are built for filtering purposes [12,18,25,55,87,109], where the goal is to predict some property rapidly. Such models are generally used as screening tools, allowing one to prioritize molecules from large libraries.…”

“…Though fingerprints were originally developed for searching chemical databases, they have proven to be useful as descriptors in developing a variety of predictive modeling studies [12,33,45,51,92,117]. The goal of any fingerprint is to characterize the substructures present in a molecule.…”