Leveraging class-balancing techniques for predicting c-MET Inhibitors: Descriptor Calculation, Selection, and QSAR Model Optimization using Machine Learning

Abstract

The rapid emergence of resistance in cancer chemotherapy is a major challenge in the drug discovery of cancer, restricting the action of various important classes of inhibitors against EGFR, VEGF, BRAF, alkylating agents, and DNA damaging agents. c-MET plays an important role in the development of resistance to cancer. Identifying a potent c-MET inhibitor can improve therapeutic access to existing anti-cancer agents. In the current study, we propose a novel technique for the prediction of drug activity class by using class balancing and ML classifiers. This study utilizes 3091 molecules with c-MET inhibitory concentration value (IC₅₀) publicly available from the ChEMBL Database. Using 14 descriptors and random oversampling for class balancing, we investigated seven classical ML models, i.e., decision tree (DT), Adaboost decision tree (ABDT), K-nearest neighbors (K-NN), support vector machine (SVM), Bernoulli Naïve Bayes (BNB), random forest (RF), and linear logistic regression (LLR) for activity prediction against c-MET. Of which SVM, LR, and RF were the top three models providing high predictability after applying balancing techniques and performing rigorous.hyperparameter tuning. Even though SVM, LR, and RF showed exemplary performance in terms of ROC-AUC and recall metrics, their validation on seven FDA-approved drugs demonstrated their susceptibility to high false negatives. Owing to this, we developed a consensus mechanism based on these three models. The consensus mechanism can work on a large, diverse database and screen potential inhibitors, prioritizing which molecule to be considered further for experimental testing. The consensus model proved beneficial as a drug design algorithm for c-MET inhibitor drug discovery and development.