RNA-seq technical replicates at library construction level

[obig]

I apologize if this has been answered elsewhere but I can't find it. The forum and literature are full of analyzes that have looked at technical replicates in RNAseq where the same library is sequenced multiple times (on different lanes or flowcells). And, generally very high correlations are reported (~0.95 to 0.99). But, I can find no discussion of the concordance between technical replicates (exact same biological sample) where for each replicate a new library is constructed. For example, suppose you have cell culture, split into three plates, grow up, extract RNA, and then go through identical library construction for each, and sequence independently. Does anyone have experience with this?

[obig]

Here is what Marioni et al (2008; Genome Research. 18(9):1509-17) have to say on the topic in their paper RNA-seq: an assessment of technical reproducibility and comparison with gene expression arrays.

"We note that our study design did not include replicates of the processing step of the Illumina sequencing library. As equivalent processing steps for microarrays (i.e., fragmentation and amplification) result in the introduction of very little technical variance, we expected that rather than the processing of the library, sequencing over different lanes and flow-cells would introduce most of the technical variance associated with Illumina sequencing. However, given our observation that very little technical variance is associated with sequencing in different lanes or plates, the variance introduced in the library processing step may contribute a nontrivial proportion of the total technical variance associated with the sequencing technology."

They seem to suggest that it is overall a minor issue. And, my gut tends to agree. But, it would be nice if someone has actually proved it.

[Veleno]

Variation is introduced by the operator at this stage, so as long as the same person handle the library making it should be reduced to a minimum.

[Simon Anders]

The very first paper on RNA-Seq, the one by Nagalakshmi et al.., did that. Since then, people seem to have forgotten about it, although you are right that it can be very informative. Especially when proposing a new library prep protocol, the obvious thing to do would be to perform the protocol several times on the same biological material to prove that the protocol is stable and reproducible. Surprisingly, referees don't seem to insist on this.

[pmiguel]

Your biological replicates are pre-library, and therefore include all the library construction variation. So, if you think about it, the only advantage you gain from doing the library technical replicates, is the ability to identify where a high sample variance derives. This is useful information if your assay has failed to detect differences between treatment/conditions for lack of statistical power due to high inter-sample variance. But making libraries is still expensive. Hence adding technical replicates at this stage still costs too much to be reasonable.

My guess would be that most of the technical variation in library construction would arise when sample or processing issues cause a bottleneck in the percent of molecules converted to library molecules. PCR tends to "paper over" issues involving low input, so the "enrichment" step tends to mask issues of this sort. Much better if your library construction fails altogether when sample or processing issues cause very low yields at a given step.

--
Phillip

[Simon Anders]

Just to be clear: I don't suggest to use library-prep replicates for biological experiments -- only for benchmarking new protocols.

[davisc]

We have a lot of experience with this from the work we have done with ENCODE RNA-Seq. You can assess yourself directly from the results if you would like. Download the K562 Poly-A+ data from UCSC:

http://genome-preview.ucsc.edu/cgi-bin/hgTrackUi?hgsid=2641755&c=chr17&g=wgEncodeCshlLongRnaSeq

For this experiment, we took 2 vials of K562 (same lot#, same donor) thawed them and cultured them independently, isolated RNA and made 2 separate libraries (we refer to these as biological replicates). Each library was sequenced on 3 lanes of GAIIx (technical replicates). I believe, the FASTQ's from the 3 lanes were merged into a final dataset for each replicate but you should be able to uncouple them using the flowcell_lane in the header. Compute FPKMs for (1) each lane within a single library and (2) lanes between biological replicates and compare the Spearman values.

Incidentally, if you want to know what the variation between 2 individuals/genomes/donors but the same cell type, i.e. skeletal muscle cell that data is also available. Download the SkMC cell line data, there are 2 replicates there are they come from 2 different donors.

[Veleno]

That's kind of expected, if you would have taken 2 vials of K562 from different labs you'd start to see clonal variation; if you take two primary cell types from two individuals you'll see lots of it.

[davisc]

Yep expected, but the OP was asking for numbers. You can also obtain numbers for the scenario you propose by comparing the CalTech RNA-Seq to CSHL RNA-Seq (K562).

In short, after doing 100's of RNA-Seq experiments with replicates: We don't worry about lane-to-lane variation anymore.. because we observed so little compared to biological variation.

[Simon Anders]

You misunderstood the OP. He was not asking for variation due to sequencing on two different lanes, but due to preparing two libraries from the same material.

You first prepared a library, then split the library onto two lanes. You could also have split the material before library preparation to see how much noise you introduce in the library prep steps.