Development of CYP3A4 Inhibition Models: Comparisons of Machine-Learning Techniques and Molecular Descriptors

Abstract

Computational models of cytochrome P450 3A4 inhibition were developed based on high-throughput screening data for 4470 proprietary compounds. Multiple models differentiating inhibitors (IC(50) <3 microM) and noninhibitors were generated using various machine-learning algorithms (recursive partitioning [RP], Bayesian classifier, logistic regression, k-nearest-neighbor, and support vector machine [SVM]) with structural fingerprints and topological indices. Nineteen models were evaluated by internal 10-fold cross-validation and also by an independent test set. Three most predictive models, Barnard Chemical Information (BCI)-fingerprint/SVM, MDL-keyset/SVM, and topological indices/RP, correctly classified 249, 248, and 236 compounds of 291 noninhibitors and 135, 137, and 147 compounds of 179 inhibitors in the validation set. Their overall accuracies were 82%, 82%, and 81%, respectively. Investigating applicability of the BCI/SVM model found a strong correlation between the predictive performance and the structural similarity to the training set. Using Tanimoto similarity index as a confidence measurement for the predictions, the limitation of the extrapolation was 0.7 in the case of the BCI/SVM model. Taking consensus of the 3 best models yielded a further improvement in predictive capability, kappa = 0.65 and accuracy = 83%. The consensus model could also be tuned to minimize either false positives or false negatives depending on the emphasis of the screening.

Quantifying and Predicting the Promiscuity and Isoform Specificity of Small-Molecule Cytochrome P450 Inhibitors

“…PLSR applied to feature-based chemical descriptors and P450 inhibition behavior provided much more predictive insight than the (univariate) linear regression results discussed above. It is quite probable that more complete descriptors of chemical features and structure, possibly combined with more advanced regression or machine learning techniques such as support vector machines (Merkwirth et al, 2004) or Bayesian classifiers (Arimoto et al, 2005), would lead to increases in accuracy. This type of analysis is made possible using a continuous, quantitative measure of specificity such as the one we have presented in this work, and the simpler approach demonstrated here as a proof of principle may itself prove useful in identifying potentially problematic P450 inhibitors early in the development process.…”

“…PLSR provides a regression matrix that shows how strongly each predictor (i.e., chemical substructure) influences the data (i.e., inhibitory potency toward a specific P450 isoform). PLSR differs from machine learning classifiers based on Bayes' theorem or support vector machines (Merkwirth et al, 2004;Arimoto et al, 2005) in that it can be used to reconstruct a continuous range of predicted values for the variable of interest, rather than simply binning it into one of two or more categories. Compared with support vector regression, PLSR is easier to implement and interpret, and it often provides similar results (Ustün et al, 2007;Shah et al, 2010).…”

In silico prediction of cytochrome P450 inhibitors

“…Recently fingerprints have successfully been applied as descriptors for modeling of CYP2D6 [O'Brian and de Groot, 2005;Jensen et al, unpub-lished data] and CYP3A4 inhibition [Arimoto et al, 2005;Jensen et al, unpublished data;Molnár and Keserü , 2002]. The models based on fingerprints are all classification models and have a higher percentage of correct classifications in the test set than the classification model based on other descriptors (see Table 1).…”

In SilicoADME/Tox Predictions

“…For toxicity properties, DTs are used to predict hERG inhibition [202] and toxicity involving cytochrome P450 such as six CYP isoforms [331], 2D6 and 1A2 isoforms [193,312], or the 3A4 isoform [332].…”