QSAR Modeling and Prediction of Drug–Drug Interactions

TLDR

Severe adverse drug reactions are the fourth leading cause of fatality in the U.S., with over 100,000 deaths annually, and up to 30 % of these are attributed to drug–drug interactions mediated by cytochrome P450 enzymes, underscoring the need for predictive tools. The study aimed to develop and validate quantitative structure–activity relationship models that predict drug–drug interactions mediated by four CYP isoforms. Data on 1,485–27,966 potential DDIs across CYP1A2, 2C9, 2D6, and 3A4 for 55–237 drugs were compiled, drug pairs were represented as binary 1:1 mixtures, and QSAR models were built using QNA and simplex descriptors with radial basis functions and random forest algorithms. The models achieved 72–79 % balanced accuracy on external test sets with 81.36–100 % coverage, and identified more than 4,500 predicted DDIs not present in the training data that were confirmed by DrugBank.

Abstract

Severe adverse drug reactions (ADRs) are the fourth leading cause of fatality in the U.S. with more than 100 000 deaths per year. As up to 30% of all ADRs are believed to be caused by drug–drug interactions (DDIs), typically mediated by cytochrome P450s, possibilities to predict DDIs from existing knowledge are important. We collected data from public sources on 1485, 2628, 4371, and 27 966 possible DDIs mediated by four cytochrome P450 isoforms 1A2, 2C9, 2D6, and 3A4 for 55, 73, 94, and 237 drugs, respectively. For each of these data sets, we developed and validated QSAR models for the prediction of DDIs. As a unique feature of our approach, the interacting drug pairs were represented as binary chemical mixtures in a 1:1 ratio. We used two types of chemical descriptors: quantitative neighborhoods of atoms (QNA) and simplex descriptors. Radial basis functions with self-consistent regression (RBF-SCR) and random forest (RF) were utilized to build QSAR models predicting the likelihood of DDIs for any pair of drug molecules. Our models showed balanced accuracy of 72–79% for the external test sets with a coverage of 81.36–100% when a conservative threshold for the model's applicability domain was applied. We generated virtually all possible binary combinations of marketed drugs and employed our models to identify drug pairs predicted to be instances of DDI. More than 4500 of these predicted DDIs that were not found in our training sets were confirmed by data from the DrugBank database.