Dear Desmond,

In pooling your 9+/-1 samples to 3-pools, you will have to keep in mind that you will have to run confirmation study using e.g. qRT-PCR for each sample separately for all 4 strata, as only by that you can be sure you did not get a positive result because of only one animal/treatment in the pool. Moreover, you have to keep in mind, that pooling might also cause you to "miss" very small alterations that might disappear when only 1 animal have a different transcriptomic profile than the others in the pool.

There are two ways to do the pooling, and it is more of what you choose to "miss" or enhance. 1) is by pooling with the same RNA ratio for 3 animals per pool, so you will have 3 pools, each consisting of 3 animals. In this case you will have more variability as some pools might have a different transcriptomic profiles, but then you will lose a bit less information. 2) pool all the 9 animals with same RNA ratio and then use this as 3 replicates. But of course you might lose more small changes; nevertheless, the very strong changes will be spiked by that.I hope this was in a way somewhat helpful. Good luck.

I had been worrying more about the different quality and conc of the extracted RNA samples (although their all pretty good and similar). And also about when to add spike in.

What you replied opens up a whole new aspect I hadn't really considered. An interesting possible pooling arrangement would be, let's say I have 8 samples and 4 pools, Each pool could be a pool of 4 samples, but a different 4 each time, so each sample would end up in exactly 2 pools. That might allow some interesting between and within variances to be measured.

Thank you for your reply. There's obviously some flexibility and what you choose to do might well depend on your expectations and research question.

Yes, I agree with that.

Regarding the RNA concentration and quality, you need to assume that every cell in the tissue you isolate you RNA from, have similar average number of RNA, but of course this is only an assumption. This is also why at the end you will need also to normalize everything to at least 3-6 reference genes. Regarding the quality, you have to have at least all at a similar RIN values, and again you will have to keep some assumptions regarding that too.

This is also why, at the end you will not come around without running confirmation analysis using qRT-PCR for each sample separately.

I see from some literature that this stuff throws up a lot of false positive and a qRT_PCR is necessary to validate. Regarding your 2nd pooling option - pool all the 9 animals - and in my dataset then have 3 technical replicates per treatment. These technical replicates allow variance resulting from the pooling process to be assessed. I have been told biological replicates are more important. The argument is that one cannot access whether transcription level difference across treatments is significant without knowing how transcription level varies from sample to sample. For example, whether a mean difference of 1  is significant or not depends on whether population std dev is 0.1 or 10, (it also depends on sample size but you get the idea). You could use the 4 sets of 3 pools as biological replicates, and assess sample to sample variation that way. That requires the assumption that variance doesn't change (much) with treatment. I don't know how reasonable such an assumption is. It doesn't sound too bad to me, especially if the purpose is only to generate plausible candidates for validation by qRT-PCR.

Hi Desmond,

First, I have to strongly advise against the second pooling strategy suggested by Edna.  As you mentioned, it would wipe out your biological variability, effectively making your replicates obsolete.

I have to say, I'm still missing the reason for pooling? Are you struggling to get enough RNA from a single sample? If not (i.e. you can comfortably make a library from each or enough of your samples) just use one sample per library. Pooling the samples will not provide you with any additional information. Better yet, use RNA from independent samples for your qPCR validation.

Finally, what are you planning to sequence your samples on? Unless I miss something you will have a lane per sample. Assuming it's one of the Illumina's instruments, that's at least 120 million reads per sample, which is an vast overkill for a standard RNA-seq experiment on a model organism. We would usually aim for 20M reads for mRNA library and 40M reads for a total RNA library. So (not knowing all details) I would say you can sequence all you individual samples on 12 lanes. This, of course would cost you more in library preparation costs. So perhaps a compromise is possible, reducing the number of lanes and increasing number of replicates per sample.

All the best,

Aliaksei.

Simultaneous generation of many RNA-seq libraries in a single reaction.

http://www.nature.com/nmeth/journal/v12/n4/full/nmeth.3313.html

This is the best way to pool many libraries. We did similar experiment, where we measured differential gene expression in 14 different mouse tissues using single library prep reaction.

Dear Aliaksei, To make things clear, I did not recommend to pool the samples according to strategy no. 2, I just listed what are the options and the disadvantages and "advantages" in each, and what one needs to consider. In any case pooling is always a compromise and when one can, better not to pool at all. But many labs, as you might know struggle with the finances or with sample quantity in which both cause to come to such strategies. 

Replying to Aliaksei

Aliaksei>> First, I have to strongly advise against the second pooling strategy

The 4 treatments (3 pools each) provide 4 biological replicates. Although one must then assume that variance doesn't change (much) with treatment, that assumption doesn't sound too bad if the only purpose is to generate plausible candidates for validation by qRT-PCR.

Aliaksei>> Unless I miss something you will have a lane per sample. 

Apologies I said that wrong. I have money for one lane and 12 barcodes. Shoe-string budget.

Aliaksei>> Are you struggling to get enough RNA from a single sample? If not just use one sample per library. 

No we've plenty of RNA per sample. Indeed what you suggest i.e. not pooling, was also suggested here - https://www.biostars.org/p/157946/ in response to my colleague's asking a similar question. I'm in favour of pooling (being a statistician) but am more or less persuaded not pooling is the best design, basically because I don't know what the differentially expressed gene (DEG) analysis programs, e.g. edgeR, actually do. However seems to me if you have enough samples per pool the distributions will be Gaussian and you can do ANOVA. But I'm deferring to my colleague who's done lots of sequencing before.

Alexander>> A bit technical for me. I think we'll just go with our lab's library prep kit.

https://www.biostars.org/p/157946/

An Update

I thought I'd tell you what we decided to do.

I should have mentioned before we have 12 litters, 3 per treatment combo, with litter size varying from 1 to 5. We extracted RNA from the samples in batches of 6. Our first batch gave low yield. We changed the extraction procedure after the 1st batch by increasing the amount of homogenising. The subsequent batches were satisfactory.

We decided instead of pooling to sequence 12 individual samples,  3 samples from each treatment combo. Doing that we could

a) avoid pooling (frankly being a statistician I was for pooling but opinion was strongly for avoiding it)

b) avoid the dodgy batch 1 samples

c) take the best sample from each litter

d) well (c) was modified slightly, we chose samples so as to ensure any RNA extraction batch present was present at least twice.

Sounds as a good idea. Wish you good luck with the experiment and the findings, certainly very exciting. 

Hi Desmond,

It looks to me like you've chosen a sensible strategy. So I don't mean to argue, but I'd like to possibly clarify some things.

Desmond >> The 4 treatments (3 pools each) provide 4 biological replicates.

I'm not sure I understand the above statement. If you have 3 pools (regardless of how many animals are in each pool) for each of the 4 conditions (treatments). This makes 3 biological replicates (not 4). What is crucial is that each of those pools is derived from independent animals. I.e. animals 1-3 contribute to the pool 1, 4-6 to the pool 2, etc. The second strategy proposed (but not necessarily endorsed) by Edna was suggesting to pool all 9 RNA samples together and then split that large pool into 3 "replicates". To me this would clearly violate the assumption of independence between the replicates. And the variance you would estimate from such pseudo-replicated would be purely technical. It's possible that I mis-interpreted the strategy #2, but that's how it sounded to me.

Desmond >> I'm in favour of pooling (being a statistician)

I would be interested to hear your statistician's opinion on the benefits of pooling, if that's not too much work. I really can't see what you would get from pooling multiple samples of the same genotype beyond increasing the amount of your input material.

Desmond >> However seems to me if you have enough samples per pool the distributions will be Gaussian and you can do ANOVA.

Unless I'm really confused, the number of samples per pool wouldn't make a slightest difference to the distribution. You would still get a single number from your pool. So it doesn't matter whether you have 3 or 300 samples in your pool. If you only have 3 pools, you're estimating your distribution from 3 samples.

There's also numerous reasons why you can NOT do ANOVA for DE analysis and I probably don't know most of them. A few I can think of is that distribution for read counts would be very different from Gaussian and you would be trying to estimate mean and variance for thousands of genes based on 3 biological replicates. Finally, this wouldn't take into account some common biases, such as decrease in variance with increase in gene expression. So I think you're right to stick to the tools dedicated to RNA-seq analysis.

Best of luck!

Dear Edna,

Sorry, I didn't mean to suggest that you were endorsing the second strategy. Nonetheless you did list it, as a viable option. And I really don't think it is a viable option. Unless I seriously mis-understood it, your second option was to pool all 9 samples together, presumably in equimolar proportions and then split it into 3 "replicates". I use the quotes because those would be technical and not biological replicates.

I'm all too familiar with the problem of insufficient input amount and financial constraints. But the second strategy wouldn't address any of them. One would still have to use enough RNA to create 3 pools and prepare 3 libraries (one from each pool).

Just to be clear I'm not opposed to pooling samples and I've done it myself on several occasions (mostly due to insufficient input amount). And your first strategy was absolutely adequate. I just didn't think in this paricular case there would be any benefit in pooling over sequencing individual samples.

I hope this clears things up.