An ensemble-based approach using structural feature extraction method with class imbalance handling technique for drug-target interaction prediction

Various methods are developed to study drug-target interaction problems. One way to deal with the drug-target interaction problem is to use a wet laboratory, which is considered a time-consuming process. Another possible way is through computational techniques. These techniques are significant in terms of time and cost. However, computational techniques use traditional classifiers to predict drug-target interactions. Drug-target interaction datasets suffer from a class imbalance problem where non-interaction classes are large in number and interaction classes are less proportional. Traditional classifiers are affected by their biased nature towards the majority class. This article provides a proposed model which uses AM-PseAAC and MSF as feature representations for target and drug, respectively, and also uses SMOTE-ENN as a resampling technique to handle class imbalance. Finally, it uses a soft voting ensemble proposed by combining random forest and Xgboost ensemble learners. A comparison of the proposed model with the existing state of the art using standard datasets (enzyme, GPCR, Ion Channel, Nuclear Receptor) confirms that the proposed model outperforms. Experiments were performed at different stages of the proposed design, considering AUC, G-Mean, sensitivity, and specificity as performance metrics, which in turn validate our proposed framework.