Drug Repositioning by Integrating Known Disease-Gene and Drug-Target Associations in a Semi-supervised Learning Model

Area of Science:

• Computational biology
• Pharmacology
• Bioinformatics

Background:

• Drug repositioning accelerates the discovery of new therapeutic applications for existing drugs.
• Current computational methods often rely on known drug-disease associations, which can be incomplete or inaccurate.
• Existing machine learning and network-based approaches assume similarity between drugs and diseases.

Purpose of the Study:

• To propose a novel computational method, Regularized Least Square for Drug Repositioning (RLSDR), for identifying new drug indications.
• To develop a method that does not require explicit definition of non-drug-disease associations.
• To improve the accuracy and efficiency of drug repositioning prediction.

Main Methods:

• RLSDR utilizes a semi-supervised learning model based on Regularized Least Square.
• Drug similarity is determined by chemical structures, and disease similarity by shared phenotypes.
• An artificial set of drug-disease associations is constructed using known disease-gene and drug-target interactions, bypassing the need for a gold-standard set.

Main Results:

• RLSDR demonstrated superior prediction performance, measured by AUC, on the constructed artificial dataset compared to gold-standard datasets.
• The method outperformed two representative network-based approaches in predicting drug-disease associations.
• Novel drug indications were identified and subsequently validated using external evidence.

Conclusions:

• RLSDR offers a robust and effective computational strategy for drug repositioning.
• The method's ability to use an artificial association set enhances its applicability.
• RLSDR successfully identifies and validates novel therapeutic uses for existing drugs, contributing to pharmaceutical research.