A systematic study of genome context methods: calibration, normalization and combination.

Area of Science:

• Bioinformatics
• Computational Biology
• Genomics

Background:

• Genome context methods predict functional gene relatedness using homologous gene patterns across genomes.
• Existing studies often lack systematic analysis of these methods and their optimal combinations.

Purpose of the Study:

• To systematically evaluate and compare four main genome context methods: phylogenetic profile, gene fusion, gene cluster, and gene neighbor.
• To investigate the impact of normalization techniques and method combinations on prediction accuracy.

Main Methods:

• Comparative analysis of phylogenetic profile, gene fusion, gene cluster, and gene neighbor methods across diverse organisms.
• Exploration of parameter spaces and application of normalization procedures to genome context scores.
• Development and evaluation of combined scores using decision tree combiners trained via cross-validation.

Main Results:

• Gene neighbor method demonstrates superior sensitivity compared to phylogenetic profile, performing competitively with gene cluster methods.
• Normalization significantly enhances phylogenetic profile sensitivity by approximately 25%.
• Combined methods, particularly with a decision tree combiner, yield up to 20% improvement over the best individual method, but are sensitive to training data.

Conclusions:

• The gene neighbor method is the most effective individual genome context approach.
• Combining genome context methods can improve performance, but requires careful selection of training data for the combiner.
• When suitable training data is unavailable, relying solely on the gene neighbor method is recommended over a poorly trained combined score.