What to include in my count table(s) for DESeq

[rndouglas]

I have five experimental groups (A, B, C, D, and E; 3-6 bioreps each) that I want to compare to each other and a control (Y; 3 bioreps) using DESeq.

When I run DESeq is it better to use a count table that includes A, B, C, D, E, and Y, then run each comparison, or is it better to make a count table for each comparison (ie: a table for A and Y, another table for B and Y, another table for A and B, etc.) and end up with ~15 different count tables?

I ask because I end up with different lists of significantly changed loci depending on how I run the analysis and I'm not sure which is more 'correct.'

[feralBiologist]

It is better to have a single count table. This shall lead to increased statistical power. You are likely to get somewhat higher number of differentially expressed genes as DESeq would be able to tease out more signal from the noise.

[rndouglas]

This is what I had been doing all along (running everything in one count table), but then this morning I thought I'd try A vs. B in a separate count table.

I was surprised to find almost double the loci with padj < 0.05 compared to when I ran everything in one count table (and hence my new-found concern).

[dpryan]

Have a look at the size factors. If one of them from the full dataset is very different than the others, that can cause this sort of result.

[feralBiologist]

This is what I had been doing all along (running everything in one count table), but then this morning I thought I'd try A vs. B in a separate count table. I was surprised to find almost double the loci with padj < 0.05 compared to when I ran everything in one count table (and hence my new-found concern).

This is really strange and counter-intuitive. Do you find more loci for all the contrasts? I can imagine this being the case for a small subset of the contrasts where there happens to be less variation within the sample groups but I find it hard to believe that you would get in total a much higher number of differentially expressed features by splitting the count table. What sort of normalisation do you do? Can you give more details about your bioinformatics workflow?

[rndouglas]

I just realized I'm only seeing this in my smallRNA libraries (adapter removed, t/rRNA removed, size-selected for 20-25nt in sRNA Workbench).

I map the reads using bowtie (-v 0).

I generate read counts with htseq-count, then build my count table(s).

I run DESeq following along with the vignette section 3.1.

So far, every 1v1 count table I've looked (7 of the possible 15) at has called more significantly changed loci (padj < 0.05) than I get when looking at the exact same comparison using a count table that includes all 30 of my bio-reps.

The biggest 'jump' was from 76 to 372 loci for one comparison.