Comparing sum of log2 fold changes between timepoints

Question

[Lorenz Becker]

I have RNA-seq data for four time points (spaced unequally), A B C and D. I have done both a spline and a group model to look for overall changes over time and between timepoints.

Is there a way of testing whether the amount of up- or downregulation is different between A/B and B/C or C/D (e.g. whether there is more upregulation between earlier timepoints)? Is sum(log2FC), separate for up and downregulated genes, the best way of expressing this?

I know it is possible that a lowly-expressed gene can change its expression slightly and produce a big log2FC, distorting the data; thus I'm also looking at the number of changing genes (first plot) and the number of counts underlying these changes (third plot).

e: I assume I can use the Chi² test to examine the number of changing genes (first plot), and possibly the Mann-Whitney-U test to compare the amount of counts (third plot), but I am unsure which test would be appropriate for the second.

Answer 1

[jared.andrews07]

I have never seen information like this displayed this way, and I expect you'd get significant backlash from reviewers if you submitted this in a paper. Summing your log2FC doesn't really make sense. As for your concerns about skewed fold changes due to low counts, you're absolutely right, but depending on how you're dealing with the data, you can account for that. For instance, DESeq2 has several methods that shrink fold changes for genes with low counts.

Ultimately, I think what you're trying to show would likely be better shown through boxplots of normalized counts for your differentially expressed genes between each of your comparisons. They could be separated into sets of up/downregulated genes. I think that method would give a much more believable look into how the magnitude of change between timepoints differs. If you do some fold change shrinkage, you could also take a crack doing boxplots for the fold changes similarly.

[Lorenz Becker]

Yeah, thought so. For a bit more context, this is an early figure in the paper, and is supposed just to summarise how we see the transcriptome change over time (i.e. "are there more changes early in life"). On consultation with a biostatistician, we are likely to replace it with a density plot of log2 fold changes of all genes (regardless of p-value). I do use jitter plots with a line for the median for individual genes, but we thought it would be good if we were able to represent the whole changing transcriptome.