Analysis Updates: RNA-seq differential-expression module

[cansavvy]

Since we are trying to integrate more DESeq2 functionality into our RNA-seq section, we should make differential expression reflect that as well since that is its main utility.

We can borrow from training-modules steps: https://github.com/AlexsLemonade/training-modules/blob/master/RNA-seq/05-nb_cell_line_DESeq2.Rmd and maybe useful tidbits from here: https://github.com/AlexsLemonade/training-modules/blob/master/RNA-seq/03-gastric_cancer_exploratory.Rmd

I'm debating about whether we should still maintain a QN'ed from refinebio version of differential-expression? One strategy is we could keep the current example but reframe it as a "pre-normalized DE" example and make the new notebook a "starting at counts DE" example. Thoughts?

[jaclyn-taroni]

The current example isn't QN'd data.

[cansavvy]

Oh I see that now. Hmm.. I guess I was thrown off by DGEList thing. Why did we choose limma-voom here?

[jaclyn-taroni]

I do not have a good reason.

[cansavvy]

We should look at the examples listed here and decide what we would like to do (or maybe a few of them): https://bioconductor.org/packages/release/bioc/vignettes/tximport/inst/doc/tximport.html#Downstream_DGE_in_Bioconductor

I do not understand the rationale of limma-voom vs the various DESEq2 models and I suspect this will also be an area of confusion for our users.

So although we don't want to get too far into the weeds on dge models, we should evaluate:

• what steps we are recommending
• why we are recommending them (when might they not be a good idea)
• and how this might fit in with the new RNA-seq modules being added (Add RNA-seq dataset examples to modules that do not currently have an RNA-seq-specific example #110)

[cansavvy]

I think I remember reading this before, but it's a great little summary to keep in mind for this: https://www.biostars.org/p/284775/

[cansavvy]

Also this by Mike Love: https://mikelove.wordpress.com/2016/09/28/deseq2-or-edger/

[cansavvy]

I don't feel like its super necessary or helpful to keep an example of each edgeR, limma-voom, and DESEq2. So what I think might be a good way forward is to change this example to be DESEq2-centric, but briefly acknowledge the others and link out to examples of those as well as the nice summary articles.

Our rationale for choosing DESeq2 would basically be

• Other methods are all about as good for normalization purposes
• It has a lot of nice docs and examples
• Be honest: we have preferences.

So the current rna-seq_DGE.RMd notebook we we switch the strategy to DESeq2 so it will be more parallel to the rest of of module changes. Basic steps would be:

1. Intro includes more wordy version of our DESeq2 rationale I described above plus links out to EdgeR and limma-voom examples.
2. Download non-QN'ed refinebio data
3. Create DESeq2 object: DESeq2::DESeq2::DESeqDataSetFromMatrix() (Side note maybe show how to save it to an RDS) where design = ~ mutation.status.
4. Follow the modeling set up steps not unlike: https://github.com/AlexsLemonade/training-modules/blob/master/RNA-seq/05-nb_cell_line_DESeq2.Rmd
5. Keep a volcano plot. Add some interpretation bits.
6. At the end link to heatmap clustering example because that flows nicely.

[cbethell]

As discussed with @jaclyn-taroni and @cansavvy in the data science team Slack channel.

The changes made in PR #140 seem to be better suited for the differential-expression example notebook. That being said, the current branch in PR #140 (which has been closed, but branch retained) can be merged into a new branch for the purpose of getting started on this ticket.

@jaclyn-taroni edit: the branch name is cbethell/add-clustering-rna-seq-example

[jaclyn-taroni]

Expanding on this point a bit - from #140 (review)

What if instead we used much of what we have here (e.g., the dataset and collapsing steps) as a DGE example? We could have multiple RNA-seq DGE examples that vary in designs and this dataset could be an example of collapsing replicates and multifactor designs. If it is presented that way, I would be less concerned about the extent of the metadata cleaning steps required, too.

So more generally, I think multiple DGE examples could be very useful where what is on cbethell/add-clustering-rna-seq-example is not the first example that is presented. I expect that putting multiple examples into the same notebook is going to make an individual notebook pretty long, which is to say that I think there will be multiple differential-expression notebooks and what's on that branch will be 02 or later.

[jaclyn-taroni]

I also wanted to note that the dataset currently in use on #140 is also included in one of our exercises for training–I expect that material will also help with getting the DGE example done.