An Empirical Study to Analyse The Effect of Bagging and Feature Subspacing on The Performance of A Custom Ensemble Algorithm for Predicting Drug Protein Interactions

Abstract

Objectives: The objective of this study is to analyse the effect of bagging and feature subspacing on the performance of a custom ensemble of decision tree classifiers for predicting drug protein interactions. Methods: In our present work we have designed a custom ensemble algorithm with decision trees as the base learner. We analysed the effect of bagging negative samples and feature subspacing on the performance of the custom ensemble in terms of AUCROC and AUPR. The Enzyme dataset from the Yamanishi dataset composed of 445 drugs and 664 proteins was used for the experiments. Findings: It was observed that the effect of bagging negative samples was significant as compared to feature supspacing in terms of AUPR metric. Now since AUPR is a metric that remains unaffected by the presence of negative samples hence the increase in AUPR by increasing the negative to positive ratio clearly indicated that the negative samples do contain the positives which are unknown and are yet to be verified. Novelty: The results give a strong indication that that feature subspacing has no considerable impact on the AUCROC metric performance of the custom ensemble while AUPR metric increases as the negative to positive ratio increases. The results give a foundation to the fact that, finding reliable negative samples from the entire set of negative drug protein pairs can further enhance the performance of the machine learning classifiers. Keywords: Decision tree classifier, Ensemble classifier, Drug discovery, Bagging, Drug repurposing