Metric Labeling and Semimetric Embedding for Protein Annotation Prediction

Area of Science:

• Computational Biology
• Bioinformatics
• Systems Biology

Background:

• Computational methods accurately predict protein function from interaction networks.
• Current methods often treat protein functions independently, overlooking their inherent similarities.
• Databases like the Gene Ontology (GO) encode functional relationships.

Purpose of the Study:

• To explore the Metric Labeling problem for protein function prediction.
• To utilize heuristic distances between functions and incorporate functional similarity information.
• To improve the accuracy of protein function prediction in biological networks.

Main Methods:

• Developed a convex optimization technique for converting heuristic semimetric distances to metric distances with minimum least-squared distortion (LSD).
• Applied the Metric Labeling approach to protein function prediction using networks derived from physical interactions and combined data types.
• Compared the performance of Metric Labeling against five existing protein function inference techniques.

Main Results:

• The Metric Labeling approach demonstrated superior performance in inferring protein function from networks compared to five existing methods.
• The proposed LSD minimization technique effectively converted heuristic distances into a usable metric.
• The study successfully integrated functional similarity information into the prediction process.

Conclusions:

• Metric Labeling is a valuable approach for enhancing protein function prediction accuracy.
• Least-squared distortion minimization is a key technique for leveraging heuristic distance measures in biological network analysis.
• Integrating functional similarities significantly improves computational prediction of protein roles.