A Feature-Reduced Discretized Random Forest Model for Oral Bioavailability Data Classification

Abstract

Oral bioavailability is the measurement of the fraction of admissible drug which reaches the site of action in unchanged form. It is one of the principal pharmacokinetic properties and can be predicted in an early phase of drug discovery and development process. Various computational methods have been used for predicting oral bioavailability of a drug candidate in the literature, which selects some of the compounds from the huge set which are most effective drug candidates and also reduces the cost factor of clinical trials. In this study, we have assigned a class label of all chemical compounds as high (Fractional Absorption F%\(\,\ge \,50\)) or low (Fractional Absorption F% < 50) oral bioavailability values. Here, the main aim is to obtain an effective model for classification of oral bioavailability data. In order to achieve this, we have preprocessed oral bioavailable data using Pearson correlation and subset selection as feature reduction methods and data discretization using binning. Discretization is one of the popular data preprocessing technique, which maps continuous data points into discrete values for easy data visualization and improves the performance of classification model. The effectiveness of feature reduction with discretization method for oral bioavailable data has been represented in terms of performance matrices like accuracy percentage, sensitivity, specificity, precision, and negative predictive value. Based on the comparative analysis of the performance of various classification models like artificial neural network (ANN), Bayesian classifier, support vector machine (SVM), K-nearest neighbor with feature-reduced discretized random forest model, we conclude that our proposed model gives better performance over the other compared models.