Enhancing drug repositioning: A multi-class ensemble model for drug-target interaction prediction with action type categorization.

Accurate prediction of drug-target interactions (DTIs) is critical for accelerating drug repositioning and reducing the cost of pharmaceutical development. Most existing studies frame DTI prediction as a binary task and often neglect the pharmacological action types and the quality of non-interaction data. This study introduces a multi-class classification framework that categorizes interactions into activators, inhibitors, and non-action classes. A novel zero-interaction selection algorithm is proposed, based on weighted drug-drug and protein-protein similarity scores, to improve dataset diversity and reliability. Drug and protein features were extracted from DrugBank, PubChem, and UniProt, and various feature selection and dimensionality reduction techniques-including decision tree, random forest importance scores, principal component analysis (PCA), Autoencoders, and Permutation importance-were evaluated to identify the most informative features for classification. We also compare concatenation-based and convolution-based feature integration strategies and systematically evaluate a range of classifiers, including both feature-based and graph-based models, with special attention to ensemble learning approaches. The concatenation method consistently outperforms convolution, and Histogram-based Gradient Boosting (HGB) achieves the best predictive overall accuracy with an average of 87.90% on the external test set. Meanwhile, HeteroGNN demonstrates more balanced class-wise performance, particularly for underrepresented classes. This work provides a scalable and interpretable framework for computational drug repositioning, supporting faster and more cost-effective identification of therapeutic candidates.