DRP-VEM: Drug repositioning using voting ensemble model

Conventional approaches to drug discovery are both expensive and time-intensive. To circumvent these challenges, drug repurposing or repositioning (DR) has emerged as a prevalent strategy. A noteworthy advancement in this field involves the widespread application of machine learning techniques. The effectiveness of these methods depends on the quality of features, their representations, and the underlying dataset. Notably, the issue of redundancy in feature sets can detrimentally impact the overall performance of these methods. Furthermore, the careful selection of a suitable training set plays a pivotal role in enhancing the accuracy of machine learning approaches in addressing drug repurposing challenges. Discovering the appropriate training set faces two significant challenges. Firstly, many methods utilize known drug-disease pairs for positives and unknown pairs for negatives. The stark imbalance in the number of known and unknown pairs often results in a bias towards the larger group, introducing errors in machine learning performance. Secondly, the absence of a documented drug-disease association indicates that it hasn't been experimentally approved yet, and this status may change in the future. This paper introduces DRP-VEM, a novel approach designed for predicting drug repositioning, specifically customized to tackle the challenges previously outlined. DRP-VEM evaluates the effectiveness of binary-based and similarity-based representations of drugs and diseases in enhancing the model's performance. Additionally, it proposes a voting ensemble training strategy, adept at managing imbalanced datasets. The assessment of DRP-VEM spans a range of parameters, including its efficacy in representing both diseases and drugs, the proficiency of its classification methods, and the application of voting ensemble training approaches using heterogeneous evaluation criteria. Significantly, DRP-VEM achieves an AUC-ROC of 81.8\% and AUC-PR of 76.6\%. Comparative analysis with other studies highlights the superior performance of the proposed model, underscoring its effectiveness in drug repositioning prediction.