Need advice on analysing RNA-seq time series

[JonB]

Hi everyone,

I have performed RNA-seq from different stages of embryogenesis. Some stages have two, three or four replicates, while some have none. I want to identify a set of genes that I can say are associated with embryogenesis in general - of all stages.

What I have done so far is to use DESeq2 and performed pairwise DE-tests between all the different stages and extracted the significant genes from each test, pooled them and extracted the unique genes. But I am not sure if this is the best way (or even a correct way) to do it. And I guess I should perform some kind of statistical test on my procedure, like permutations or something.

Any advice on how to analyse this data is greatly appreciated.

Thanks,

Jon

[Michael Love]

hi Jon,

The likelihood ratio test (see the section in the vignette) can be used to test for differences at any stage. Here, you specify a reduced formula where you remove the stage variable from the design (so if your design is ~ stage, then reduced = ~ 1).

Mike

[JonB]

hi Jon,

The likelihood ratio test (see the section in the vignette) can be used to test for differences at any stage. Here, you specify a reduced formula where you remove the stage variable from the design (so if your design is ~ stage, then reduced = ~ 1).

Mike

Thanks!

Jon

[JonB]

hi Jon,

The likelihood ratio test (see the section in the vignette) can be used to test for differences at any stage. Here, you specify a reduced formula where you remove the stage variable from the design (so if your design is ~ stage, then reduced = ~ 1).

Mike

Dear Mike,

Sorry for asking novice questions.
I have only one variable to my design (condition) that includes all the different stages as well as the non-reproductive stages (which I want to use as controls). When I created the dataset I used design = ~ condition, but will this be correct if I want to use reduced = ~ 1 as you suggested?

Jon

[Michael Love]

Using full = ~ condition and reduced = ~ 1 tests for any difference in counts related to the levels of condition. this will include genes where stage 1 is different than control, and genes where stage 2 is different than stage 1, etc.

[Wallysb01]

The basic idea here is you’re looking for genes in which the beta term in the equation expression = beta * stage + constant is significantly different from zero. This is a generalized linear model.

So you’re not going to get genes that are just different between stage 1 and stage 2, then do what ever else in stages 3-5 or what not. You’re going to be finding genes that consistently show a trend, up or down, as stage progresses.

You can also put any other confounding variables in there. So if you have different batches of embryos in any way or different sequencing run/library prep days, you may want to try to control for that by adding another variable (design = ~ condition + batch, reduce = ~ batch).

[Michael Love]

hi Wally,

re: "show a trend" that would be true if stage was a numeric, but I guess Jon coded it is a factor.