RNA-seq DE with AAV injection

Question

[CriticalP]

Dear all,

I'm currently facing an issue when analyzing a RNA sequencing dataset. My samples come from two similar brain tissues, one injected with AAV for a loss-of-function experiment, and one without any viral injection.

There are several genes involved in viral response, that produce noise in the enrichment output analysis after DE in edgeR.

Is there any workaround for that ? Is it possible to filter immune related genes at the beginning of the analysis, as they are not in the scope of the research ? I know it is common to filter for recombinant segments of genes, but I don't know for "real" immune genes. Is it possible to select the interesting pathways and not show directly the immune pathways, while explaining this contamination in-text ?

For those wondering, I also have a "control" virus condition, but for some reason, maybe a high titer, this "control" virus yield so much more DE genes that are related to immune function, therefore, in comparison it seems that my sample of interest is depleted from immune genes.

I was not able to find any information on that, but it should be a common problem with many studies involving AAV injections. Any help or experience is appreciated,

Many thanks,

[rpolicastro]

Although having different titer values is not ideal, using an interaction term might help you fish out genes whose effect is dependent upon your LoF variant while controlling for the effect of AAV.

The formula for the interaction term would be in the form of ~ aav_treatment + lof_treatment + aav_treatment:lof_treatment and the contrast of interests would be checking for the significance of the comparison (AAV with LoF vector - untreated) - (AAV control - untreated). A significant p-value for a gene would indicate that the magnitude/direction of the change observed when treated with AAV depends on the presence of your LoF-inducing vector. *removed after Ian pointed out it's not full rank (all combinations of factors necessary for individual terms).

I would pair this with doing traditional differential expression with a simplified model ~ aav_and_lof_treatment so you know what the fold change is for both the AAV control - untreated, AAV with LoF vector - untreated, and AAV with LoF - AAV control comparisons are.

Gordon Smyth will probably pop up here in a few days to scold me for suggesting an interaction term and propose an overall better approach if you want to wait too.

[i.sudbery]

I don't think that you will be able to fit ~aav_treatment + lof_treatment + aav_treatment:lof_treatment because the lof_treatment column in the design matrix will be the same as the aav_treatment:lof_treatment column and so the matrix will not be full rank.

However, (AAV with LoF vector - untreated) - (AAV control - untreated) = AAV with LoF vector - AAV control

So what is really being suggested here is that you compare the control AAV directly with the LoF AAV rather than comparing each with uninfected. Have you tried that?