Assessing the limits of genomic data integration for predicting protein networks

Area of Science:

• Genomics
• Bioinformatics
• Systems Biology

Background:

• Genomic data integration, combining diverse functional genomics data for large-scale predictions, is increasingly common.
• The expectation is that integrating more evidence should enhance predictive power.

Purpose of the Study:

• To explore the limits of genomic data integration by assessing how predictive power changes with added features.
• To investigate the prediction of protein-protein interactions in yeast, a well-studied benchmark.
• To evaluate feature dependencies in genomic data integration.

Main Methods:

• Utilized a Naive Bayes classifier to integrate diverse genomic evidence (e.g., coexpression, phylogenetic profiles).
• Expanded the number of predictive features to 16, exceeding previous studies.
• Calculated mutual information to assess statistical dependence between features and compared classifier performance with and without boosting.

Main Results:

• A small but measurable improvement in prediction performance was observed with increased features, building upon four strong initial features.
• No significant statistical dependence was found between the tested genomic features.
• Boosting the classifier did not improve performance, supporting the independence of the features for prediction purposes.

Conclusions:

• Integrating a few strong genomic features (around four) approaches the maximal predictive power for current genomic data integration methods.
• The limitations in predictive power are not due to inter-relationships between genomic features.
• This study identifies new yeast protein-protein interactions with high confidence and quantifies feature inter-relations.