Gene set internal coherence in the context of functional profiling

Background: Functional profiling methods have been extensively used in the context of high-throughput experiments and, in particular, in microarray data analysis. Such methods use available biological information to define different types of functional gene modules (e. g. gene ontology-GO-, KEGG pathways, etc.) whose representation in a pre-defined list of genes is further studied. In the most popular type of microarray experimental designs (e. g. up-or down-regulated genes, clusters of co-expressing genes, etc.) or in other genomic experiments (e. g. Chip-on-chip, epigenomics, etc.) these lists are composed by genes with a high degree of co-expression. Therefore, an implicit assumption in the application of functional profiling methods within this context is that the genes corresponding to the modules tested are effectively defining sets of co-expressing genes. Nevertheless not all the functional modules are biologically coherent entities in terms of co-expression, which will eventually hinder its detection with conventional methods of functional enrichment. Results: Using a large collection of microarray data we have carried out a detailed survey of internal correlation in GO terms and KEGG pathways, providing a coherence index to be used for measuring functional module co-regulation. An unexpected low level of internal correlation was found among the modules studied. Only around 30% of the modules defined by GO terms and 57% of the modules defined by KEGG pathways display an internal correlation higher than the expected by chance. This information on the internal correlation of the genes within the functional modules can be used in the context of a logistic regression model in a simple way to improve their detection in gene expression experiments. Conclusion: For the first time, an exhaustive study on the internal co-expression of the most popular functional categories has been carried out. Interestingly, the real level of coexpression within many of them is lower than expected (or even inexistent), which will preclude its detection by means of most conventional functional profiling methods. If the gene-to-function correlation information is used in functional profiling methods, the results obtained improve the ones obtained by conventional enrichment methods.