You can cluster using expression profile by many clustering approaches like K-means, hierarchical etc. The hierarchical clustering could be the best choice. If you have good sample size then correlation coefficient between genes can be used directly for clustering (r as dist matric). The other approaches include mainly co-expression network analysis and the WGCNA (Weighted Gene Co-expression Network Analysis) is very good choice as it is widely applied and is robust approach. It is developed as a package in R (http://cran.r-project.org/web/packages/WGCNA/index.html). This package uses Topological Overlap Measure (TOM) for clustering. The TOM based clustering is more robust because it uses correlation as well as sharing of neighbourhood between the genes.

Some important links on WGCNA:

http://www.biomedcentral.com/1471-2105/9/559

http://labs.genetics.ucla.edu/horvath/htdocs/CoexpressionNetwork/Rpackages/WGCNA/Tutorials/

No good one-size fits all solution. You need to consider whether to use hierarchical vs. k-means vs self-organizing maps. Each has advantages/disadvantages depending on your experimental design. In addition, various specialized tools have been developed to detect time dependent trends. What conditions do you want to compare? Are you more interested in absolute levels of gene expression or trends/fold-changes induced by some sort of stimulus?

After I have done the basic functional analyses, I want to identify groups of genes that are potentially co-regulated -- clusters of genes that show a tightly correlated behaviour. Once I have identified a few medium or small clusters of potential interest -- no large clusters, and only very closely correlated genes -- I search for clusters that are enriched in genes differentially expressed between the experimental groups or clusters enriched in a relevant function (e.g. immune response).

Until now, I have been using k-means and hierarchical standard clustering tools (hclust and kmeans), I have no experience with SOMs. What would be the advantage of SOMs?

I don't think this answer is related in any way to my question.

To get a sense for the utility of SOM in analyzing gene expression datasets, I'd suggest you look at the GEDI tool developed by Sui Huang. http://www.ncbi.nlm.nih.gov/pubmed/14630665

Is there a time component in your experiment?

As you describe your experiment I would guess hierarchical clustering would do it. Perhaps the problem is that strong correlations of gene expressions (r>0.8) are rare.

You might also consider to define correlated genes not on your data but on large data collections starting from candidate genes using for example HEFALMP (http://hefalmp.princeton.edu/)

Probably, model-based clustering of Adrian Raftery and his co-workers could use for clustering of your data. The benefit of their model-based clustering is that they have all the techniques in their R software package "mclust". I'd suggest you look at the following paper: "Model-Based Clustering and Data Transformations

for Gene Expression Data"

You can cluster using expression profile by many clustering approaches like K-means, hierarchical etc. The hierarchical clustering could be the best choice. If you have good sample size then correlation coefficient between genes can be used directly for clustering (r as dist matric). The other approaches include mainly co-expression network analysis and the WGCNA (Weighted Gene Co-expression Network Analysis) is very good choice as it is widely applied and is robust approach. It is developed as a package in R (http://cran.r-project.org/web/packages/WGCNA/index.html). This package uses Topological Overlap Measure (TOM) for clustering. The TOM based clustering is more robust because it uses correlation as well as sharing of neighbourhood between the genes.

Some important links on WGCNA:

http://www.biomedcentral.com/1471-2105/9/559

http://labs.genetics.ucla.edu/horvath/htdocs/CoexpressionNetwork/Rpackages/WGCNA/Tutorials/

Thank you all for your answers. Now, asking the question I had something more general in mind ("what is your favourite colour" rather than "what colour should I paint my bedroom :-)). Thanks for all the suggestions! I will now go and start playing with all that great toys...

so my favourite colour is MeV (http://www.tm4.org/), a very versatile toy

I use GEDI and self-organizing maps because they are suitable to time-course and illustrative. The maps allow easily calculate entropy that describes organization of gene ensembles. I also like MeV.

Any opinions about NMF algorithms to find clusters of co-expressing genes?

I tried NMF Consensus tool at http://genepattern.broadinstitute.org for clustering genes, but this tools clusters samples instead.

Hi January, I have the same question.... I want to identify groups of genes that are potentially co-regulated -- clusters of genes that show a tightly correlated behaviour. which tools did you use at the end? I'm not really a bioinformatic so I don't know much about it. Thanks

http://compbio.mit.edu/cummeRbund/manual_2_0.html