Read depth recommendations

[dpryan]

Hi all,

We'd like to perform some RNA-seq to look at gene expression level changes in mouse hippocampus due to a treatment of interest to us. We're not interested in finding new transcripts or looking for differences in splice junctions or anything of that sort. Consequently, I'm curious what people are recommending these days in terms of read depth.

On a related note, I've read a number of people here suggesting that paired end reads are probably not required for our sort of project. If that's the case, I'm curious what sort of read lengths (36bp, 50bp, etc.) people having been using that give them meaningful results.

Any suggestions you might have would be appreciated.

[mbblack]

I was at a meeting a couple of weeks ago (the 2011 TIES meeting at UNC) and in a talk by Wendell Jones (a statistician with the company Expression Analysis) he talked briefly about this.

An Illumina white paper from a few years ago argued that 2-10 million mapped reads should be in the range of equal or better sensitivity than microarrays for differential expression estimation. Wendell, however, mentioned that his experience with experimental data over the years has seen that number climb, to where most of his clients are more often using 20-50 million mapped reads in order to be "comparable" or better to array data.

I think though, that most of these kinds of estimates are based on human data. We work mostly on rat and mouse models, and I honestly am not convinced of just what we need in terms of RNAseq coverage to get results equal to or better than our array results. For our first direct comparison, I have greater than 60 million mapped reads per sample (3 controls, 3 treatment animals, all mouse livers), but I get much less sensitivity for gene expression than with microarray data (same samples used too). We're trying another direct comparison soon (mouse liver samples already run with affy titan arrays) soon to be run on an ABI SoLid 5500xl, shooting for 10-20 million reads per sample.

Wendell also mentioned in his talk how differential expression significance has occasionally been seen to appear to be fine at low coverage, but suddenly drops out at high coverage, but he did not offer an explanation for that observation nor elaborate on the specificis.

Thus far in our research, we've been using 50bp single end reads, but I don't really think that 36bp reads would be a problem.

P.S. There is an FDA-led initiative called SEQC underway (a followup to the MAQC initiative - http://www.fda.gov/ScienceResearch/B...ls/default.htm ) - http://www.genomeweb.com/sequencing/...rna-sequencing which is intended to put some real numbers to issues like this, based on real comparison data.

<edit> actually, SEQC is also really MAQCIII, the third phase of the whole MAQC long term initiative. Some of the sequencing is done, some in the works right now and still some more to be done in the next few months. Data analysis is really just in the very initial stage.

[dpryan]

Thanks mbblack, that's extremely helpful! I'll have to look more into SEQC and MAQC, they sound interesting.