DESeq Heatmaps

[sjv3]

Hi,

Please could anyone tell me what information DESeq uses to produce heatmaps, if you follow the process in the vignette.... does it base the top 50 or however many included on p values or padJ values?

I did my analysis in a pairwise fashion using 2 replicates of 4 timepoints, so day 0 vs day 1, day 0 vs day3, etc etc etc, but the heatmaps I've produced with desq contain all 8 samples, so if it has used p-values of differential expression, where has it got this from when the p values all refer to just 2 samples at a time?

many thanks in advance for any advice/help!

from a novice deseq user!

[daiello]

Hello,
In the vignette, the list of differentially expressed genes is selected by pval (unadjusted). The un-transformed and variance stabilized counts for all samples are then plotted in the heatmap. You need to 1) come up with a list of differentially expressed genes, sort genes by pval or foldChange 2) decide what samples you want to plot 3) pass the matrix of expression values/counts for samples into the heatmap function. You can plot raw counts, normalized counts, or variance stabilized counts (vignette explains how to compute all). Variance stabilized counts are usually the best to use for heatmaps. Each of your pairwise comparisons need to be performed separately. For each pairwise comparison I would 1) come up with a gene list 2) plot the expression values/counts for all the samples in a heatmap.

[sjv3]

thanks, I'll do that.

the one thing I still can't understand though is which set of p values would it have used on the heatmap it produced containing all 8 of the samples, when in reality there are 6 different sets of pvalues: D0vsD1, D0vsD3, D0vsD6, D1vsD3, D1vsD6 and D3vsD6... would it have chosen one of these lists of p values at random to select the top differentially expressed genes?

Thanks very much

[daiello]

Hello,
The user specifies the gene list to plot in heatmap. The number of gene lists should be equal to the number of pairwise comparisons. Since you are dealing with time-series data you have a more complicated analysis to perform than what, I think, DESeq can handle. The type of analysis you perform depends on the question you are trying to answer. For example, I typically use DESeq to find genes that are statistically differentially expressed between two treatments. You have time-series data, so you are probably interested in what genes change the most over time? Maybe use DESeq to normalize your counts then conduct linear regression on each gene to find genes that change most over time. Also, look into multi-factored analysis with DESeq. I'm not sure but maybe this will help.
-Dan