Fusing literature and full network data improves disease similarity computation

Area of Science:

• Bioinformatics
• Computational Biology
• Medical Informatics

Background:

• Existing disease similarity methods often rely on limited gene-disease associations or only specific protein interactions.
• Many diseases lack sufficient associated genes, hindering the broad application of current computational approaches.
• There is a need for methods that leverage comprehensive data sources for robust disease similarity computation.

Purpose of the Study:

• To develop and evaluate MedNetSim, a novel method for computing disease similarity.
• To integrate information from the entire protein interaction network and biomedical literature (MEDLINE).
• To address limitations of existing methods, particularly those requiring extensive gene-disease association data.

Main Methods:

• MedNetSim combines a network-based method (NetSim) using the complete protein interaction network with a literature-based method (MedSim) mining MEDLINE.
• NetSim analyzes the entire protein interaction network, not just disease-gene interactions.
• MedSim extracts disease-related information from biomedical literature.

Main Results:

• NetSim achieved an average AUC of 95.2%, outperforming other function-based methods.
• MedSim showed high performance among semantic-based methods, comparable to some function-based ones.
• The integrated MedNetSim achieved the highest average AUC of 96.4%, further improving to 97.5% with increased gene-disease data volume.

Conclusions:

• Integrating biomedical literature and protein interaction networks is effective for computing disease similarity.
• Literature-based methods like MedSim are valuable complements to function-based algorithms, especially when gene data is scarce.
• Further research into gene-disease associations and protein interaction data quality is recommended.