Metric learning on expression data for gene function prediction.

Area of Science:

• Genomics
• Bioinformatics
• Computational Biology

Background:

• Gene co-expression analysis across diverse RNA-Seq datasets is crucial for understanding biological processes.
• Identifying relevant experimental conditions for specific Gene Ontology (GO) terms is challenging with standard methods.
• Current approaches often use all samples, potentially diluting signals from informative subsets.

Purpose of the Study:

• To develop a novel algorithm, Metric Learning for Co-expression (MLC), for GO-term-specific gene co-expression analysis.
• To improve the accuracy of Guilt-By-Association-based gene function predictions by weighting expression samples.
• To identify informative experimental conditions for specific biological processes.

Main Methods:

• Developed MLC, a fast algorithm assigning GO-term-specific weights to RNA-Seq samples.
• MLC maximizes weighted co-expression for genes within a GO term and minimizes it for genes outside the term.
• Applied MLC to Arabidopsis thaliana RNA-Seq data for performance evaluation.

Main Results:

• MLC significantly outperforms standard Pearson correlation in term-centric performance for gene function prediction.
• The method demonstrates superior accuracy for more specific Gene Ontology terms.
• MLC-derived sample weights identify important experiments and can reveal novel relevant conditions.

Conclusions:

• MLC provides a more accurate and nuanced approach to gene co-expression analysis by incorporating sample relevance.
• The algorithm enhances the prediction of gene function, particularly for specific biological processes.
• MLC offers insights into experimental design and condition relevance in large-scale transcriptomic studies.