The work from Julie L Morrison and colleagues from the university of Glasgow, recently published in BMC Bioinformatics, is interesting in several ways.
GeneRank: Using search engine technology for the analysis of microarray experiments. describes the application of the PageRank algorithm used by Google to "boost" the rank of genes in a list that are e.g. differentially expressed.
This idea naturally extends to analysing the results of a microarray experiment, where we would like a gene to be highly ranked if it is linked to other highly ranked genes, even if its own position is lower, e.g., due to measurement variability.
Algorithmically, the work is solid and the application of such algorithms seems a smart way of making use of interaction networks
and expression data and I got a nice introduction to the PageRank algorithm with it. The algorithm includes a weighting parameter, you can solely rely on the underlying network for detecting groups or rely only marginally on the use of GeneRank algorithm off.
For me, there is one big caveat: Are the genes that are highly connected really important, pivotal genes? Morrison
et al only use a network obtained from
Gene Ontology. For integration, it might be more useful to use the rich resource we have in protein-protein interaction data, particular when analyzing data in yeast. However, given the many false positives and the fact that the highly connected proteins generally display unspecific binding, I wonder whether we would anything out of such analyses. Obviously, this applies to all interpretation of networks.
If you create the interaction map of the parts of a machine or study the human body and count the interaction of its structures - are the "important" parts the ones that are highly connected?