Metrics for GO based protein semantic similarity: a systematic evaluation

Area of Science:

• Bioinformatics
• Computational Biology
• Genomics

Background:

• Gene Ontology (GO) terms facilitate functional comparison of gene products.
• Existing semantic similarity measures lack conclusive evaluation for optimal use.
• The inclusion of electronic annotations in similarity calculations remains debated.

Purpose of the Study:

• To systematically evaluate Gene Ontology (GO)-based semantic similarity measures.
• To assess the impact of electronic annotations on semantic similarity calculations.
• To establish criteria for selecting the most effective semantic similarity measures.

Main Methods:

• A systematic evaluation of GO semantic similarity measures was performed.
• Performance was quantified by correlating semantic similarity with sequence similarity.
• Measures were tested with and without electronic annotations to assess their influence.

Main Results:

• The relationship between semantic and sequence similarity is non-linear, approximated by a rescaled Normal cumulative distribution function.
• Most semantic similarity measures exhibit similar behavior but differ in resolution, with resolution being the primary evaluation criterion.
• The hybrid simGIC measure showed the best overall performance, followed by Resnik's measure with a best-match average approach.

Conclusions:

• This study provides a framework for comparing semantic similarity measures, aiding researchers in selecting appropriate methods.
• The hybrid simGIC measure is recommended for its superior performance.
• Average and maximum combination approaches are problematic due to term count influence; electronic annotations may introduce data circularity.