Link Prediction Only With Interaction Data and its Application on Drug Repositioning.

Area of Science:

• Computational biology
• Pharmacology
• Bioinformatics

Background:

• Computational methods aid drug development by predicting drug-disease associations.
• Existing methods struggle with incomplete feature data and unreliable negative samples.

Purpose of the Study:

• To propose a novel method (TS-SVD) for predicting drug-disease associations using network interactions.
• To address limitations of existing methods concerning data completeness and negative sample selection.

Main Methods:

• Constructed a drug-protein-disease heterogeneous network.
• Calculated topological similarity using common neighbors.
• Generated low-dimensional embeddings via topological features and Singular Value Decomposition (SVD).
• Employed a Random Forest classifier with carefully selected negative samples.

Main Results:

• Achieved better or comparable performance than state-of-the-art methods using less information.
• Demonstrated improved prediction accuracy through a novel negative sample selection strategy.
• Case studies confirmed the method's practicality for discovering novel associations.

Conclusions:

• The TS-SVD method offers a robust approach for predicting drug-disease associations.
• The negative sample selection strategy enhances model reliability and performance.
• This method facilitates efficient drug discovery and development.