Drug-Disease Association Prediction Using Heterogeneous Networks for Computational Drug Repositioning.

Area of Science:

• Computational biology
• Pharmacology
• Bioinformatics

Background:

• Drug repositioning accelerates new drug development by finding novel therapeutic indications for approved drugs.
• Computational methods utilizing heterogeneous networks are increasingly employed to predict drug-disease associations.

Purpose of the Study:

• To review and compare network-based computational approaches for predicting drug-disease associations.
• To evaluate the performance of different methods across three categories: graph mining, matrix factorization/completion, and deep learning.

Main Methods:

• Construction of heterogeneous networks integrating drug-drug similarities (chemical structures, ATC codes), disease-disease similarities (ontology-based), and known drug-disease associations.
• Selection and comparison of eleven network-based methods across graph mining, matrix factorization/completion, and deep learning categories.
• Utilized an improved evaluation metric to account for data imbalance in sparse positive associations.

Main Results:

• Methods employing graph mining and matrix factorization/completion demonstrated strong overall predictive performance.
• Drug-side prediction accuracy was higher compared to disease-side prediction.
• The quality of drug features used in drug-drug similarity calculations significantly impacts disease-side prediction accuracy.

Conclusions:

• Network-based approaches, particularly graph mining and matrix factorization, are effective for drug repositioning.
• Improving drug feature representation is key to enhancing prediction accuracy, especially for identifying new disease indications for drugs.