12 months ago by

Institute for Molecular Medicine Finland (FIMM), Helsinki

Hi

1. Yes, I think error bars do make sense here. The results table returned by DESeq2 contains the columns 'log2FoldChange' with the (shrunken) fold change estimates, and right next to it, the standard errors of these estimates ('lfcSE'). If you add and subtract the standard error to the fold change estimates.you get positions for the upper and lower whiskers of your error bars. Note that these are standard errors, not 95%-CIs, and that there is not correction for multiple testing.

2. You should use 2^(lfc+lfcSE) and 2^(lfc-lfcSE).

Now, this will produce ugly looking asymmetric error bars. My recommendation is to always plot fold changes on a logarithmic scale. After all, you would want a change of 2-fold up (2x) and two-fold down (0.5x) to look the same, though in different directions. Sometimes, people are worried that mathematically less savvy readers find log2-transformed values harder to read or interpret intuitively than raw values. My recommendation here is to nevertheless stick to log-scale axes, but label them in the real scale: Do not label the y axis "log2 fold change" and mark off the equidistant ticks as -3, -2, -1, 0, 1, 2, 3, but label the axis as "fold change" and mark the ticks as 1/8, 1/4, 1/2, 1, 2, 4, 8.

3. The two baseMean values were simply the averages of the normalized counts for the two groups. As DESeq2 is layed out for more general designs, we removed this left-over from the earliest DESeq versions which only allowed for two-group comparisons. If you still need an average over the group, run 'rowMeans' on a suitable subset of columns from 'counts( dds, normalized=TRUE )'.