Learning from low-rank multimodal representations for predicting disease-drug associations

Area of Science:

• Bioinformatics
• Computational Biology
• Pharmacology

Background:

• Discovering new disease-drug associations is crucial for drug discovery and treatment.
• Traditional methods for confirming these associations are costly and time-consuming.
• Computational approaches are needed to predict unobserved disease-drug links.

Purpose of the Study:

• To develop a novel computational method for predicting unobserved disease-drug associations.
• To leverage multimodal data for enhanced prediction accuracy.
• To improve the efficiency of identifying potential therapeutic relationships.

Main Methods:

• The proposed method, LMFDA, utilizes drug chemical structures, disease MeSH descriptors, phenotypic terms, and drug-drug interactions.
• It constructs similarity networks from diverse data sources to represent drugs and diseases.
• Multimodal fusion with low-rank tensors and matrix complement technology are employed for prediction.

Main Results:

• LMFDA achieved high performance on two datasets, with Area Under the Receiver Operating Characteristic Curve (AUROC) values of 91.6% and 92.1%.
• The method outperformed several existing computational models in predicting disease-drug associations.
• Experimental results demonstrated LMFDA's excellent network integration and detecting performance.

Conclusions:

• LMFDA offers a novel approach for disease-drug association inference through multimodal fusion.
• The method shows substantial improvement over advanced techniques, highlighting its effectiveness.
• Integrating domain knowledge into similarity networks is a promising strategy for drug repositioning.