Automatic Kernel Selection of Support Vector Machine for Drug/Non-Drug Compounds Classification

Abstract

The traditional approach of drug design and discovery is a complex process which is highly time-consuming and costly. Recently, researchers have been utilizing computational resources to speed up drug design and discovery process, known as Computer-Aided Drug Design (CADD). In CADD, the primary aim is to identify drug compounds, which neutralize the target molecule from a large biochemical dataset. Therefore, identification of drug compounds can be formulated as a classification problem. Support Vector Machine (SVM) is one of the most widely used machine learning algorithms for classification problems. However, kernel selection in SVM is a crucial task for classification problem. In this paper, we proposed a mechanism to select kernel function automatically for SVM (Auto-SVM) for drug/non-drug compounds classification. The perceptron learning algorithm is used for the linear-separability analysis of the biochemical datasets which infers whether the SVM can be integrated by a linear or non-linear kernel function. An SVM kernel function is selected based on statistical analysis. Different biochemical datasets are considered for experimental purposes. The proposed methodology demonstrates high classification performance, compared with the state-of-the-art ML algorithms such as back-propagation neural network (BPNN), naive Bayes (NB) and k-nearest neighbor (k-NN) with respect to accuracy, precision, recall and F1-score.