DESeq2 glm caveats

[lukas1848]

Hi,

I am currently using DESeq2 for my RNAseq data with a multi factorial design. I have 4 samples L1-L4 (7 biol. replicates) with each sample combining 2 factors.

	B_high	B_low
A_high	L1	L2
A_low	L3	L4

My DESeq glm look like this:
dds <- DESeqDataSetFromMatrix(countData = countTable, colData = Design, design = ~ A + B)

Now, the problem is that among my DE genes for let's say condition A, there are genes with very high expression in L1 and low expression in L2,L3,L4. Now, if expression in L1 is high enough, this gene will be called as highly expressed in the A_high treatment group, even though sample L2 only shows weak expression.


Here's an example from my data set. Expression of Gene A is L1>>L3>L2>L4. Yet, in the multi factor design, it gets called as significantly over expressed in cond A_high, i.e. L1 and L2. FC=log2FoldChange

ID	A_FC	A_padj	B_FC	B_padj	L1_FC	L1_padj		L2_FC	L2_padj	L3_FC	L3_padj	L4_FC	L4_padj
GeneA	0.497	0.048	0.149	0.758	0.84	4.09E-05	-0.26	0.59	-0.73	0.01	-0.05	0.92

Now I am looking for a way to correct for this. I could filter the list of DE genes of A_high by selecting only those genes where log2FoldChange is bigger in L1 and L2 compared to L3 and L4, but this seems not very well thought through.

Maybe one of you guys might have an idea how to deal with this.

[Simon Anders]

The long answer: Read up on the topic of "interactions" in linear models, and then rethink what biological question you are trying to answer.

The too short answer: Your two treatments do not act independently. In the case of your gene that is high only in L1, they have potentiated each other: A alone and B alone have little effect, but when working together, they have a strong effect. If such so-called interactions are present in your data, it does not make sense any more to only ask separately for each treatment's effect.

The way too short answer: Replace "+" with "*" in your design formula, then read up on what this means.

[lukas1848]

Hi Simon,

Our experiment consists of male and female organisms (i.e. treatment B) with treatment A leading to similar phenotypes in both sexes. Our hypothesis was that production of said phenotype is controlled by the same genes in males and females. This hypothesis led to a glm without interactions between the 2 variables.

If I introduce interactions, I would expect even more calls for genes with a pattern like in GeneA, even though these are not the genes I am interested in.

My understanding of glm's is indeed superficial. So do you say, that given that I have genes like "GeneA" in my data set, I must work with interacting terms?

I would rather stick to the glm I have with simple terms (A+B), if possible and get rid of DE calls where the treatment does not affect each group similarly.