Does NGS need replication?

[hchang10]

Hi Guys,

I have a friend told me that NGS does not need replication. In his experiment, he only submit two sample library, control and treatment. That's all. Is this correct?

In my case, I did it as 3 replication (3 independent sample preparation and RNA isolation) for control and treatment respectively, resulting in 6 library for NGS.

Please help me with the confusing point.

Thank you

[Jeremy]

If you are talking about RNA-seq, think of it like whole transcriptome qPCR. Ask your friend if he thinks any respectable journal in the world would accept a single control sample and single treatment sample for a qPCR experiment.
You are correct, biological replicates are required if you want meaningful results.

[chadn737]

What do you mean by "3 independent sample preparation"? Sometimes people tell me they have replicates and what they mean is 3 library preps from the same samples, rather than independent biological replicates. Replication is important and anybody who says it is not does not know what the hell they are talking about. Yes, the strong language is justified. There are too many poorly done studies done out there when people should really know better.

I realize that there are cases, particularly when working with clinical/human samples where there is obvious limitations in experimental design, but that should be the exception, not the rule. Too often I see labs who can do replication, choose not too because of cost. If you can't afford to do the experiment properly, then it shouldn't be done. I have seen far too many grad students, and post-docs even, screwed over by PIs with a mass of bad data because they were too cheap to do the experiment properly.

[hchang10]

Thank you guys~

Dear chadn737, I meant three samples and extracting RNA from them for library construction.

Like:

control Rep1, Rep2, Rep3 >>> RNA1, RNA2, RNA3 >>>library1, library2, librray3

Treat Rep1, Rep2, Rep3 >>> RNA1, RNA2, RNA3 >>>library1, library2, librray3

[chadn737]

Good, you have actual biological replication, which is needed for a proper RNA-Seq study. You will be able to actually assess variation in your samples and do proper statistics. You did it the right way.

Your friend will find his data to be a horrendous mess and will probably find the reviewers demanding replication and if the review process works properly...unable to publish it. If an un-replicated study came my way, I would immediately demand replication or reject it.

[hchang10]

Thank you chadn737~ (Then I guess you're a professor somewhere!)

I just asked another friend, and she said it depends. I have asked her to send me a paper which only did one replication but published in high SCI journal. We'll see~

[chadn737]

I'm a post-doc, but I have published RNA-seq data and worked with a lot of people on sequencing data. Yes you can find papers with one replication. That was fairly common early on. Its applicable in some circumstances where one is not comparing treatments (for example maybe in a de novo gene discovery situation), but just because there are published examples, doesn't mean one SHOULD do it that way.

Replication is one of the key differences between a good study and a bad study. Nothing should be done without replication. Lets be honest.....there is a LOT of crap science out there and a lot of it comes down to a failure to do proper experimental design, replication, controls, etc.

[chadn737]

I'm sorry if I come across too strongly on this. We've been having a conference and I have had a few. Also, I have had so many people come to me with data with no replication or proper experimental design that I have formed very strong opinions on the matter. Its a very real problem.

[swbarnes2]

You need biological replicates, but you don't need technical replicates.

If are are thinking that you will take one sample, prep three separate libraries, run them in three lanes on the Illumina, and then look at the average and standard deviation between the replicates to determine what that one sample really is...you don't have to do any of that. If you did, you would see that all three datasets are pretty much the same.

But you do need biological replicates; like three separate untreated individuals, and three separate treated samples.

[sdriscoll]

Statistics work better with replicates. In fact statistics don't work at all without them. Absolutely. And you do risk missing low change differentially expressed genes without replicates. However, anybody worth their salt as a scientist doing a real project should be aware that candidate genes from RNA seq results have to be verified to really be certain so the RNA seq run isnt going to be the final step. If you're doing a bioinformatics paper and you're just publishing sequencing results then yes, without replicates you'd clearly stand out as out of touch with reality.

[rory]

If you are getting pushback, you can point them at this concise, pithy paper:

Hansen KD, Wu Z, Irizarry RA, Leek JT. "Sequencing technology does not eliminate biological variability." Nature Biotechnology (2011) 29(7):572-3.

http://www.nature.com/nbt/journal/v29/n7/abs/nbt.1910.html

[mbblack]

Inherently, why would anyone think that RNA-Seq circumvents the need for biological replicates - just because it's sequence data does not make it mystically different from any other data.

Biological replication is necessary if what one wants is to characterize the mean difference between a sample population and a control population. Whether it is sequence data, microarray data, qPCR data, RFLPs, allozymes - the data does not matter. Without biological replicates, you cannot say anything about mean or overall differences in the population.

Differential gene expression analysis is (usually) a population level question. If you don't adequately sample the population, you cannot address the question of what genes are differentially expressed amongst the population(s) under study. The type of data collected is immaterial to proper population sampling.

[bruce01]

@sdriscoll: I am interested in your view of needing to confirm results from RNAseq. What rationale is there behind confirming a quantitative dataset? Is it peace-of-mind, that certain journals require such confirmation, or is it that we do not know power of RNAseq studies, or what depth of sequencing is required? Are we not moving away from this idea of confirmation of RNAseq by qPCR?

[DonDolowy]

@sdriscoll: I am interested in your view of needing to confirm results from RNAseq. What rationale is there behind confirming a quantitative dataset? Is it peace-of-mind, that certain journals require such confirmation, or is it that we do not know power of RNAseq studies, or what depth of sequencing is required? Are we not moving away from this idea of confirmation of RNAseq by qPCR?

Depending on the method you pick for testing for DEG, you are likely to get different outcomes. Therefore, in my opinion, it is always nice and calming to confirm the findings by qPCR before you spend months trying to figure out what gene x is doing in your setting.

In my case, I have triplicates. I have analyzed the data using DESeq, DESeq2 and Cuffdiff (all most recent versions) and the overlap of DEG between the methods is not great. This has led me to divide my DEGs into high confidence DEG (those that are captured by all methods) and lower confidence DEG (the rest).

[bruce01]

@DonDolowy: agreed, I have also tested several packages and methods within them. But once you understand the method and are happy that it is statistically sound, and have filtered/normalised counts appropriately, then to my mind that should be enough... I guess it is really up to the reviewers then though