Machine Learning Models to Probe the CYP3A4 Mediated Drug Metabolic Profile

Abstract

Cytochrome P450s (CYP) are a diverse group of Heme-containing proteins found in all kingdoms of life, that participate in vital life processes including oxidization of endogenous and exogenous compounds. Of the 57 CYP isoforms, CYP3A4 is the most abundant isoform in humans. CYP3A4 is highly promiscuous in substrate specificity and allows the accommodation of compounds diverse in size and structure, which leads to CYP3A4-mediated metabolism of up to 50% of all marketed drugs. However, the ability of CYP3A4 to adjust two or more similar or different molecules may also lead to adverse drug-drug interactions (DDIs), as the inhibition or induction of CYP3A4 by one drug can lead to adverse effects in the in vivo metabolism of other drugs. Pharmacokinetic issues due to the inhibition or induction of CYP isozymes are accredited for the failure of nearly 80% of drugs during development. Therefore, it is important to analyze cytochrome interactions before preclinical trials to ensure the success during the drug development process. The current study aims to utilize supervised machine learning techniques and molecular modeling strategies on publicly available CYP3A4 inhibition data to predict CYP inhibition through the development of a predictive model and the identification of 3D features responsible for CYP3A4 inhibition. Five models were built to predict CYP3A4 Inhibition on two refined different datasets of CYP3A4 inhibitors: Support Vector Machine, Logistic Regression, Decision Tree, Random Forest, and Multilayer Perceptron. The Support Vector Machine and Logistic Regression models built on the more refined dataset outperformed all others, with accuracies of 98% and 96% indicating superior performance. Therefore, these two models built on the chosen hyperparameters are suitable for the prediction of CYP3A4 inhibition in new chemical entities and can assist in the drug developmental process. Additionally, all models in the more refined dataset resulted in accuracies over 80% indicating the stabilities of the models on the data used and highlighting the importance of the refined features and data refining in general over the use of noisy raw data. The results draw attention to the importance of increased lipophilicity, vander waals surface area on pharmacophoric points, number of aromatic and rotatable bonds, percentage of Nitrogen atoms, topological distances between Nitrogen and Oxygen, and Nitrogen and Sulfur, and overall negative charge on a molecule in CYP3A4 inhibition. Thus, this study assists in understanding the key CYP3A4 interactions, prediction of CYP3A4 inhibition and the optimization of the toxicological profiles of new chemical entities.