Hi,

I read almost all the posts on time series data analysis using DESEQ2. However, I still have some questions/confusions:

1) I have a time series data of same tissue at 4-time points (control: T0, stressed for 1hr: T1, stressed for 2hr: T2, stressed for 3hr: T3). I need to detect the temporal changes in gene expression profiles over the monitored time course. From what I understand, I accounted for the timepoints in the design formula (for the whole dataset) and then did a pairwise comparison (T2vsT0, T3vsT2 and T4vs_T3). This will give me LFC changes in the genes expression from one-time point to the other. However, while doing the DESeq analysis the coefficients are T1 VS T0, T2 VS T0, T3 VS T0. So, I used contrasts options to do T3 vs T2 and T2 vs T1 and while doing that the shrinkage estimator type was selected as "normal". So, I used "normal" for these as well T1 VS T0, T2 VS T0, T3 VS T0. I wanted to know how can I use "apegml" to reduce bias. Or using "normal across all comparisons is okay?

2) I also did he LRT test with the reduced model ~1 and then did pairwise comparisons again. My question is how different this would be from what I did earlier. I understand that the log2foldchange should not be considered as in LRT and it kind of gives a list of significant genes. But is that necessary in my case? I want to understand the temporal changes in gene expression with time and determine the early responsive and the late responsive genes which approach would be better.

Please help me understand which approach will be better.

Hi,

For your first question, regarding the use of contrasts with apeglm shrinkage, you just need to re-level your factors such that you can obtain the desired coeffcient. So, in your case, you will want to re-run it with [presumably] T1 as the reference level, and then again as T2, T3, et cetera. This is always how I used to do it, and indeed there is section in the vignette where this is mentioned: Extended section on shrinkage estimators.

Regarding the LRT, the log [base 2] fold changes that are returned cannot readily be interpreted in the context of all comparison groups. The ones that are returned are related to just one of the comparisons, and I believe this is also mentioned in the vignette (search for heading 'I ran a likelihood ratio test, but results() only gives me one comparison.').

In terms of addressing your question, to understand the "temporal changes in gene expression with time and determine the early responsive and the late responsive genes", this may be better addressed by pairwise comparisons and then carefully examining the results of each.

Kevin