Selection of drug candidates requires the understanding of target binding and safety implications. Experimental evaluation of large number of compounds is not only slowing the drug discovery process, but also make its very expensive. Moving the decision on selection of lead drug candidates to early stages of the discovery process can avert costly, time consuming and potentially dangerous clinical trials. A key insight that can expedite decisions is the knowledge of drug-drug interactions. CYP 3A4 enzyme is involved in the oxidative metabolism of many drugs. Inhibition of CYP 3A4 carrying a toxicity risks.<br/>Compound structures and activity data were retrieved from the panel assay (AID: 1851) using PubChem’s API. Compounds were either classified as inhibitors or non-inhibitors, or removed based on their activity score. The Mordred package was used to generate molecular features. The PaDEL software was used to generate PubChem fingerprints. Our training dataset included 10,832 compounds. 1,826 molecular features were calculated, and 881 fingerprint bits were extracted. Two novel convolutional approaches were explored: SMILE string auto-extraction and 2D structure auto-extraction. Four common models were trained to predict CYP3A4 inhibition: logistic regression, random forest, support vector machine, and neural network. All models were extensively tuned to optimize various data preprocessing techniques and hyperparameters. 3-fold cross validation was used to evaluate model accuracy.<br/>Model based on Mordred and PaDEL features achieved their prediction accuracy as following: logistic regression - 83.2%, random forest - 83.4%, support vector machine - 81.9%, and neural network - 82.3%. Neural network model based on SMILE string auto-extraction - 82.3%, and 2D structure auto-extraction - 76.4%.<br/>All models were successful at classifying compounds. The novel convolutional methods for features extraction can learn relevant features directly from SMILES strings.<br/>

Comparison of machine learning structure-activity models for predicting compound’s activity toward CYP 3A4

ACS Fall 2020 Virtual Meeting

The official archive and repository for all ACS Meetings & Expositions abstracts and select posters or presentations.