I want to do unsupervised analysis on rnaseq data. I use this code:

topVarGenes <- order(-rowVars(assay(rld)))[0:1000]

mat <- assay(rld)[ topVarGenes, ]
mat<- mat - rowMeans(mat)
colnames(mat) <- paste0(metadata$CONDITION,"-",metadata$ID)

#draw heatmap


pheatmap(mat,method="complete", main = "Unsupervise 1000 genes ", show_rownames = F, color=rev(color2),annotation_legend = TRUE, legend=T, cluster_cols=TRUE,scale="row")

pheatmap(mat,method="complete", main = "Unsupervise 1000 genes ", show_rownames = F, color=rev(color2),annotation_legend = TRUE, legend=T, cluster_cols=TRUE)

Here You have the images of the same pheatmap but with scale=row or scale="none"  http://imgur.com/a/N7FAn

I have two different clusters.  Which is the way to do unsupervided analysis in right way?

I mean if I use scale=none I don't see some groups. Why?

Hi

I normally use the aheatmap function for drawing heatmaps. I always include scale = 'row' because I think that normalises the genes data (on the rows) before it is correlated by pearsons. So I think it is a good idea. Aheatmap does z score normalisation.

aheatmap(rows.to.keep, distfun = 'pearson', scale = 'row',
         annCol = annotation)

Best,

Chris

The important thing is that you select the top variance genes on the un-scaled data (row scaling), which you have done.

After this, you can choose what settings you like, there is not always a "true" clustering of genes or samples, as there are different weightings of the rows possible.