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

Area of Science:

• Pharmacology and Bioinformatics
• Computational Drug Discovery

Background:

• Accurate drug-target interaction (DTI) prediction is vital for drug repositioning and reducing pharmaceutical development costs.
• Existing DTI prediction methods often treat interactions as binary and overlook pharmacological action types and non-interaction data quality.

Purpose of the Study:

• To develop a multi-class classification framework for DTI prediction, distinguishing between activators, inhibitors, and non-action classes.
• To improve dataset diversity and reliability using a novel zero-interaction selection algorithm based on drug-drug and protein-protein similarity.

Main Methods:

• Extracted drug and protein features from DrugBank, PubChem, and UniProt.
• Evaluated various feature selection and dimensionality reduction techniques (e.g., PCA, Autoencoders, Random Forest importance).
• Compared feature integration strategies (concatenation vs. convolution) and evaluated classifiers, including ensemble methods like Histogram-based Gradient Boosting (HGB) and graph-based models like HeteroGNN.

Main Results:

• The concatenation feature integration method outperformed convolution.
• Histogram-based Gradient Boosting (HGB) achieved the highest overall predictive accuracy (87.90%) on the external test set.
• HeteroGNN provided more balanced class-wise performance, especially for underrepresented classes.

Conclusions:

• The proposed multi-class classification framework offers a scalable and interpretable approach for computational drug repositioning.
• This framework supports faster and more cost-effective identification of potential therapeutic candidates by improving DTI prediction accuracy and reliability.