Comparative study of machine learning approaches for classification and prediction of selective caspase-3 antagonist for Zika virus drugs

Zika virus (ZIKV) infection is an enervating and fast-growing disease. The increasing incidences of birth defects (microcephaly) in newborns due to ZIKV represent a public health problem. The viral infection is characterized by an increase in cell death of human neural progenitors and astrocytes, which can be inhibited by suppressing infection-induced caspase-3 activity. The aim of the present work is to develop classification models for the prediction of highly active and low active caspase-3 antagonists and to seek the important structural features related to the high anti-ZIKV property. Here, machine learning (ML) is applied in quantitative structure–activity relationship (QSAR) study. QSAR study is performed on the dataset by means of ML approaches, i.e., multiple linear regression (MLR), linear discriminant analysis (LDA), least square support vector machine (LS-SVM), deep neural net (DNN), k-nearest neighbor (KNN), naïve Bayes (NB) and random forest (RF). MLR, LDA are used for feature selection process and DNN, LS-SVM, KNN, NB, RF classifier for classification. The obtained results confirmed the discriminative capacity of the calculated descriptors. A good correlation is found by regression analysis between the observed and predicted activities as evident by their R² (0.895) and R _pred ² (0.716) for the molecular descriptor dataset, R² (0.892) and R _pred ² (0.736) for fingerprint dataset. The classification model obtained using RF (85.71%, 97.57%) and DNN (85.71%, 91.07%) classifier gave better accuracy than other approaches in fingerprint dataset and molecular descriptor dataset, respectively. This work provides an effective method to screen caspase-3 antagonists that will help out further in drug design for Zika virus.