Using SMOTE to Deal with Class-Imbalance Problem in Bioactivity Data to Predict mTOR Inhibitors

Abstract

Machine learning algorithms give sub-optimal performance in the presence of class-imbalanced dataset. Mammalian target of rapamycin (mTOR) is one of the serine/threonine protein kinase, and plays an integral role in autophagy pathway. Autophagy is a cellular pathway for recycling of macromolecules (proteins, lipids, and organelles), which enables eukaryotic cells to adapt metabolism to survive during adverse growth conditions. Targeting mTOR through therapeutic interventions of autophagy pathway establishes mTOR a promising pharmacological target for autophagy modulation in cancer. The bioactivity dataset of mTOR in ChEMBL, a compound bioactivity database maintained by European Bioinformatics Institute, shows disproportionate distribution of active and inactive classes. The predictive models based on this skewed dataset are biased towards prediction of majority class. Hence, we have used Synthetic Minority Over-sampling TEchnique to deal with class-imbalance problem in bioactivity datasets. We have built and evaluated predictive models based on four commonly used classifiers using both class-imbalanced and class-balanced bioactivity datasets, and compared their performance based on various metrics like accuracy, sensitivity, specificity, F1-measure, and AUC. We observe that the classification models based on balanced dataset generally outperform those that are based on class-imbalanced dataset, irrespective of the classifiers used for classification task. We conclude that predictive models trained over class-balanced dataset can be used for screening large compound bioactivity datasets to predict mTOR inhibitors-like compounds.