Batch effect help

[ea11]

I am attempting to analyse RNA-seq data using either edgeR or DESeq2 to determine differentially expressed genes. The experimental conditions are as follows:

Mice were either fed a high fat diet or a low fat diet.
Each cage had 3 mice in it, but each group has a total of 6 mice. (i.e., for the low fat diet, there are 6 mice, which have come from 2 different cages).

Therefore, I am attempting to correct for any batch effects due to the different cages the mice were raised in, this is where I am having a problem.

Essentially, I have the following: ( I have attached an image of the table too)

sample treatment cage
1 low 1
2 low 1
3 low 1
4 low 2
5 low 2
6 low 2
7 high 3
8 high 3
9 high 3
10 high 4
11 high 4
12 high 4

I have attempted to include this fact into a design matrix but I keep getting the same error of incomplete rank. I know the experimental design is not ideal, but it is what we have to work with.

If anyone has any ways of correcting for the different cages, that would be great.

Thanks for any help,
E

[dpryan]

Unless the cage mates are litter mates it would be unlikely for there to be a cage effect.

That aside, try this:

treatment cage
low 1
low 1
low 1
low 2
low 2
low 2
high 1
high 1
high 1
high 2
high 2
high 2

The design is then ~treatment*cage.

[ea11]

thanks for the reply.
I didn't think there would be a cage effect, but the count results cluster on a PCA plot depending on cages more so than treatment, so thinking that there is potentially a cage effect.

Would what you suggest not say that samples 1-3 and 7-9 are from the same cage and the other 6 from the same cage?

I am sorry, not very good with the stats, much better with the other aspects of bioinformatics.

Thanks

[dpryan]

Yes, it says they're from the same cage, but the interaction term then allows them to have different actual effects. This is just a trick to avoid the rank deficiency issue and you'll see it used whenever batches occur completely within experimental groups.

[dpryan]

I should probably expand upon my last reply. What that grouping and the design effectively do is fit an effect for each Treatment_Cage pairing, of which there are 4. Since the Cage is just a nuisance variable for you I then formulated things in a way that you could just get the "cage-corrected treatment effect" without having to also deal with contrasts.

[ea11]

That sounds like exactly what I need in that case. Thanks very much for your help, I shall include that into the analysis when I next get back to the office.

Thanks