Improved Computational Drug-Repositioning by Self-Paced Non-Negative Matrix Tri-Factorization.

Area of Science:

• Bioinformatics
• Computational Biology
• Drug Discovery

Background:

• Drug repositioning (DR) accelerates drug development by identifying new uses for existing drugs.
• Matrix factorization methods are common for DR prediction but face challenges in learning ability and local optima.
• Accurate DR prediction is crucial for reducing drug development costs, time, and risks.

Purpose of the Study:

• To propose a novel self-paced non-negative matrix tri-factorization (SPLNMTF) model for enhanced drug repositioning prediction.
• To address the limitations of existing models in learning ability and susceptibility to local optima.
• To integrate heterogeneous biological data for more robust DR predictions.

Main Methods:

• Developed the SPLNMTF model, integrating patient, gene, and drug data into a heterogeneous network.
• Employed non-negative matrix tri-factorization to learn from integrated biological data.
• Implemented a self-paced learning strategy to sequentially introduce samples from easy to complex, mitigating local optima.

Main Results:

• The SPLNMTF model demonstrated superior performance compared to eight state-of-the-art models on ovarian cancer and AML datasets.
• Integration of heterogeneous data improved the model's learning ability for predicting potential drug-target associations.
• The self-paced approach effectively prevented the model from converging to suboptimal solutions, enhancing prediction accuracy.

Conclusions:

• SPLNMTF offers a significant advancement in drug repositioning prediction by effectively integrating diverse biological data.
• The proposed method overcomes key limitations of existing matrix factorization approaches in DR.
• SPLNMTF shows strong potential for accelerating the discovery of new therapeutic applications for existing drugs.