Total read count of shotgun affected by few highly abundant sequences

Question

I am building statistical models to analyse output from Illumina shotgun sequencing (HiSeq 4000) on stool samples (but RNA-seq data should behave similarily). The raw counts are a statistical sample of the total collection of DNA molecules used as input for the sequencing machine, and I think we can assume that the sequencing platform has limited capacity, such that detected reads are always a small fraction of the the original sequencing library.

Consider two different samples: one control sample and one treatment sample, which is identical apart from a single species (or gene) that has much higher absolute abundance relative what it in the control. As a result, the reads of this species will be much higher in the treatment sample, and also the 'real estate' of the remaining genes in that sample will be decreased. This is simply the nature of relative abundances.

However, size factor estimation that uses the "median ratio method" (http://dx.doi.org/10.1186/gb-2010-11-10-r106) for instance, are based on "experience with real data [...] [that] shows [that] a few highly and differentially expressed genes [or abundant species] may have strong influence on the total read count." (http://dx.doi.org/10.1186/gb-2010-11-10-r106). I am curious to see experimental evidence for this last statement. Or does someone have at least an explanation of how total read count can be affected by the absolute abundance of a species/gene?

Answer 1

That sentence in the Anders and Huber paper is a bit misleadingly worded (I don't think that was their intent). What they mean is that highly expressed and highly differentially expressed genes can out-compete the majority of the other genes, which thereby leads to them appearing to have lower expression if one uses total read counts for normalization (basically, what you said in the second paragraph).

An alternate interpretation of what they wrote (likely the one you understandable understood to have been their meaning) is that some highly expressed genes will inflate the total number of reads a sample actually gets. I have yet to see this in practice, but it certainly COULD happen. There are technical biases in sequencing such that some sequences (often due to length and GC content) are easier to sequence on common machines. A good example of this would be PhiX, which forms clusters much more easily on Illumina machines than real samples. I expect that this is relatively rare in practice, but using a more robust normalization would still help in protecting against it.