Mining SOM expression portraits: feature selection and integrating concepts of molecular function

Area of Science:

• Bioinformatics
• Computational Biology
• Systems Biology

Background:

• Self-organizing maps (SOM) effectively visualize high-dimensional gene expression data from large sample collections.
• Analysis of SOM texture reveals gene co-expression patterns (spot-clusters) requiring significance filtering and functional interpretation.
• Gene ranking and functional interpretation of spot-related gene lists are key challenges in analyzing SOM-transformed data.

Purpose of the Study:

• To develop and compare methods for feature selection and functional interpretation of gene expression data analyzed with SOM.
• To evaluate different expression scoring methods and gene set enrichment analyses for interpreting SOM-derived gene lists.
• To demonstrate the utility of these methods using the human tissue body index dataset.

Main Methods:

• Applied various expression scoring methods (fold change, regularized t-statistics) to spot-related gene lists, considering microarray data error characteristics.
• Utilized gene set enrichment analysis (GSEA) to identify overexpressed or overrepresented gene sets within SOM spot-clusters.
• Mapped metagene-related gene set overrepresentation onto SOM images and estimated set-related overexpression profiles.

Main Results:

• Tissue-specific spots in SOM images were found to contain enriched gene sets corresponding to molecular processes in those tissues.
• Different expression scoring and GSEA methods were compared, highlighting their error characteristics.
• Demonstrated the ability to identify and display specific gene sets (e.g., housekeeping, consistently expressed) using SOM data filtering.

Conclusions:

• The presented methods facilitate comprehensive downstream analysis of SOM-transformed expression data, enabling functional interpretation of gene lists and enriched gene sets.
• SOM clustering provides a framework for defining novel gene sets or refining existing ones based on selected SOM spots.
• This approach enhances the understanding of molecular functions associated with specific patterns in high-dimensional gene expression data.