High Confidence Predictions of Drug−Drug Interactions: Predicting Affinities for Cytochrome P450 2C9 with Multiple Computational Methods

Abstract

Four different models are used to predict whether a compound will bind to 2C9 with a K i value of less than 10 µM. A training set of 276 compounds and a diverse validation set of 50 compounds were used to build and assess each model. The modeling methods are chosen to exploit the differences in how training sets are used to develop the predictive models. Two of the four methods develop partitioning trees based on global descriptions of structure using nine descriptors. A third method uses the same descriptors to develop local descriptions that relate activity to structures with similar descriptor characteristics. The fourth method uses a graph-theoretic approach to predict activity based on molecular structure. When all of these methods agree, the predictive accuracy is 94%. An external validation set of 11 compounds gives a predictive accuracy of 91% when all methods agree. IntroductionIdentifying drug-drug interaction potential early in drug discovery and development is important because drug-drug interactions can cause life threatening changes in drug levels. Early discovery of potential drug-drug interactions for a compound expedites the decision to eliminate that compound from consideration, thus lowering the cost of drug discovery. Virtual drug screening allows for the prediction of binding affinity prior to synthesis, and if prediction can be trusted, it can guide the drug discovery process. To date no single computational method has proven to be outstanding in this regard, and virtual screening still has not replaced in vitro screening methods or been routinely used in drug design. Drug interaction sites related to metabolism include UDP-glucuronosyltransferase, sulfotransferases, aldehyde oxidase, and the cytochrome P450 enzymes. Because a number of drug-drug interactions are observed for the cytochromes P450, affinity models have been developed for these enzymes. In particular, models have been developed for the three major drug metabolizing enzymes 2C9, 2D6, and 3A4.1-3 These metabolic enzymes have broad overlapping substrate specificity, most often coupled with relatively low affinity. Almost all substrates are also competitive inhibitors of the enzyme that metabolizes them, however, not all inhibitors are substrates. The general rule for drug-drug interactions is that compounds that have K i values, uncorrected for protein binding, greater than 10 µM are unlikely to exhibit important clinical drug-drug interactions. Since most compounds have K i values higher than 10 µM, including

Computational methods in drug discovery

“…Terfenadine–ketoconazole drug–drug interactions results in cardiotoxicity. Computational methods such as pharmacokinetic modeling and predicting drug–drug interactions using large DDI interaction databases are successful and are both cost and time saving as well [287–288]. …”

Substrates, Inducers, Inhibitors and Structure-Activity Relationships of Human Cytochrome P450 2C9 and Implications in Drug Development

“…When all of these methods agree, the predictive accuracy is up to 94%, with the specificity and sensitivity being 92 and 95%, respectively. LWRP predicted 44 of the 50 compounds correctly to give a concordance of 88% [1520]. Out of the 27 compounds that bind with a K i value of 10 M or lower, LWRP predicted 23 correctly.…”

“…Hudelson et al (2008) [1520] recently employed different models to predict whether a compound would bind to CYP2C9 with a K i value of <10 M. Two of the four methods (line walking recursive partitioning (LWRP) and normal equation recursive partitioning (NERP)) developed partitioning trees based on global descriptions of structure using nine descriptors. A third method (so-called "gravity method") used the same descriptors to develop local descriptions that relate activity to structures with similar descriptor characteristics, while the fourth method (SUBDUE) utilized a graph-theoretic approach to predict activity based on molecular structure [1520]. A training set of 276 compounds whose pK i values ranged from 1.9 to 7.6 and a diverse validation set of 50 compounds (http://www.seeker.wsu.edu, access date: 27 May 2009) were used to build and evaluate each model.…”

Similarity-based modeling in large-scale prediction of drug-drug interactions

“…Different approaches have been developed to identify DDIs between CYP-metabolized drugs, which are approaches that integrate in vitro data to predict in vivo (animal and human) CYP-mediated interactions 15 . Computational modeling has also been used to predict CYP metabolism–based DDIs 16 . Other pharmacokinetic processes, such as absorption, distribution or excretion, could be of interest from the point of view of DDI prediction 17 .…”