Dear Bioconductors, 

I am dealing with a poly-A selected RNA-seq dataset that contains samples collected from the central nervous system of mice treated either with a drug or a control vehicle. The primary goal of the analysis is the identification of differential gene expression associated with drug treatment.

During my initial QC, I noticed relatively high counts for different hemoglobin genes in the dataset, varying ca. 5-10 fold between samples. This probably indicates the varying degrees of blood contamination in the original samples (which are difficult to prepare). As far as I can tell, the degree of contamination does not appear to be associated with the variable of interest, e.g. drug treatment, as it fluctuates similarly within and across treatment groups.

Would anybody have recommendations on how to minimize the effects of such contamination on differential expression analysis? For example, I am considering an (unsupervised) surrogate variable analysis (SVA). Alternatively, given that the source of the contamination is (likely) known, I am wondering if it would be useful to include the expression of blood-specific marker genes (hemoglobin, etc) in a linear model.

Perhaps any of you have are willing to share some experience and / or advice?

Thanks a lot,

Thomas

This sounds like a fairly standard use case for either sva or RUVSeq. Just supply your counts and the known experimental design, and let these methods pull out unwanted factors/surrogate variables corresponding to contamination. Alternatively, if you want to focus the algorithms' attention on blood contamination, you could subset the count matrix to only use blood-specific markers when identifying unwanted factors. Both of these strategies are probably better than manually building a blocking covariate yourself - the expression profile of a single gene will be subject to more noise than an empirical factor of variation constructed from many genes.

Another approach is to just filter out the haemoglobin genes (and other blood-specific markers) and ignore the effect of contamination elsewhere in the transcriptome. By the sound of it, the contamination seems to be random across groups so it won't confound DE testing between groups. While the inconsistent contamination will also inflate the variances of a few blood-expressed genes, these will be ignored if you set robust=TRUE in eBayes. If this simple solution gives workable results, then you might as well do it. Of course, if there are too many (>10%) genes with inflated variances, the robustness will not be sufficient and explicit regression of the contamination will be necessary.