how to re-arrange the gene expression heatmap to maximize the visual effect of the result

Question

[CrazyB]

I have the mRNA expression data for 4 groups of samples (2-3 samples per group), based on different genetic backgrounds and treatments. While the dendrogram segregates group #1 from the remaining 3 groups at the first branch of the dendrogram, it is not visually clear from the heatmap.

Is it possible to find out why group 1 is segregated at the the first branch so as to re-arrange the gene lists to maximize the visual effect of the results on the heatmap ?

ps. can upload the heatmap but would need some technical help as to how to do it on the forum

Answer 1

[Jean-Karim Heriche]

There could be many reasons why one group is segregated from the others by a given clustering algorithm. It usually boils down to some combination of features distinguishing it from the rest of the data. However, another possibility to keep in mind is that the groups may not actually be very different but because hierarchical clustering forces binary divisions, one group has to split earlier from the rest and in noisy, not well separated data, this group is more or less random. One of the uses of heatmaps is to identify noise features, i.e. variables that do not distinguish between samples so if you do not see any difference between groups in your heatmap, this could be an indication that there's little structure in the data. The relative branch lengths should give you an indication of how well separated a group is from the rest of the data. You could go further and look at how well the clustering is supported by the data using the pvclust R package or by looking at the silhouette values (available in the cluster R package). Finally, well separated clusters should be relatively insensitive to the linkage method used in hierarchical clustering. Another thing to try is PCA to see if the groups also segregate in a lower dimensional space. The principal component loadings could help you identify relevant features.