There are a number of reasons which could explain the lack of correlation between qPCR and RNA-seq.

On the RNA-Seq side it is very important the way you do the analysis, namely how many mismatches you allow in your target sequences. It is very popular to allow one mismatch for example. This kind of approach will make the correlation with your qPCR results more difficult if with one mismatch your product would never be detectable with qPCR.

Second how much a gene is expressed. There is a range of linearity for each technique and this is also the case of RNA-seq. If a gene is highly expressed (like in the housekeeping case) it could be problematic to detect differences among different samples. The same may be applied to the qPCR side.

Third on the qPCR side: how did you normalize the expression? This is not trivial since you are comparing an absolute quantification with a relative one and the bias could come not from RNA-seq but from the qPCR approach.

Fourth: how much your qPCR primers are specific for your gene product. Are there present variants in the region you are looking for? For me the gold standard is always to Sanger-sequencing the qPCR product before to draw conclusions on the "goodness of fit" of your gene quantification. There are so many imperfections in the ref-seq database which you have to validate in my opinion whatever qPCR gene expression method because your primers can recognize (also) something you would never expect.

Even if I did not open your link I hope that these suggestions will be useful.

Hi Bernardo,

I think that if the RNA library was normalized you can have bias, it is not a RNAseq but a transcriptome sequencing experiment. I do not know what type of data you have but for I think you need to confirm those gene by real time as you did.

Massimo

There are a number of reasons which could explain the lack of correlation between qPCR and RNA-seq.

On the RNA-Seq side it is very important the way you do the analysis, namely how many mismatches you allow in your target sequences. It is very popular to allow one mismatch for example. This kind of approach will make the correlation with your qPCR results more difficult if with one mismatch your product would never be detectable with qPCR.

Second how much a gene is expressed. There is a range of linearity for each technique and this is also the case of RNA-seq. If a gene is highly expressed (like in the housekeeping case) it could be problematic to detect differences among different samples. The same may be applied to the qPCR side.

Third on the qPCR side: how did you normalize the expression? This is not trivial since you are comparing an absolute quantification with a relative one and the bias could come not from RNA-seq but from the qPCR approach.

Fourth: how much your qPCR primers are specific for your gene product. Are there present variants in the region you are looking for? For me the gold standard is always to Sanger-sequencing the qPCR product before to draw conclusions on the "goodness of fit" of your gene quantification. There are so many imperfections in the ref-seq database which you have to validate in my opinion whatever qPCR gene expression method because your primers can recognize (also) something you would never expect.

Even if I did not open your link I hope that these suggestions will be useful.

You have touched upon a large problem, because these two approaches to assessing gene expression are very different. As you suggested, the simple fact that your RNAseq libraries are normalized means the gene expression values you calculate by that method are relative to the mRNA pool specifically (as Massimo points out). And of course there are many other steps in your library prep that could introduce bias (particularly if it was amplified).

This is quite different from a relative qPCR assessment where the amount of RNA used for cDNA synthesis is usually based on total RNA levels (NanoDrop), or rRNA levels (Bioanalyzer). Thus the amount of material in each sample will be different than above and your reference normalization becomes the main factor in dealing with bias (as Cristiano points out).

The ideal way to produce comparable PCR data would be to use digital PCR to produce absolute data, but even then, the various other sources of bias in sample handling mentioned above (particularly library synthesis for RNAseq) may confound the agreement between the methods. Thus the likelihood of agreement will be affected by the specific tissues and techniques used and is an under-appreciated problem.

Hello! In the future, if you add the ERCC controls to your total RNA, you will have a standard to address these questions easily. Also, it shouldn't really matter if you have qPCR results. RNA-seq is better suited for the discovery of expressed RNA. Unless you have replicates, RNA-seq has a very hard time showing statistical significance, whereas qPCR does not.

ERCC ExFold control plasmid set:

http://www.lifetechnologies.com/order/catalog/product/4456739

I think this is what you can hope for across the two methods (for that matter, across any two methods), and I doubt that it will get much better whatever you do. If I read correctly your graph, it is not that one method reports a a statistically significant increase, while the other argues for a significant decrease. After all, does it really matter if it is an 8-fold or a 12-fold change? For most of us, it doesn't.

I have experienced the same issue. It was initially disconcerting but, like your data, the trends between the RNA-seq and qRT-PCR analyses were the same, just differing somewhat in the magnitude of the change. Your r2 value supports the similarity in the trend measured via the two techniques. As has been touched on in the other responses to your question, given the vast technical differences between the two methods, it is not surprising that you see differences in the magnitude of the fold change for some of your genes. Does the difference between an 8-fold change measured via RNA-seq and a 4-fold change via qRT-PCR ultimately affect the interpretation of your data in terms of biological relevance? In my case it did not, though it may in yours. In your case HPNK_08953 is difficult to interpret (a 2-fold change vs no change), but it is often difficult to confirm and ascribe biological relevance to such subtle changes. My advice is that you avoid focusing too much on the differences in absolute fold-changes between the two methods, and focus instead on the trends. However, I suspect that opinions on this advice may vary.

Hi Bernando,

I'm glad that I'm not the only one having this kind of problems. I'm quite new into this RNA-seq data analysis. As all the above people said RNA-seq and qRT-PCR are two different techniques. when I validate my RNA-seq results using qRT-PCR I checked first that my reference gene was not on my list of significant differentially expressed genes that I got from RNA-seq analysis. Secondly I designed primers with a similar PCR product (around 120bp) and finally I estimated my RQ values manually, in this way I reduced variation among my treatments by doing an average of my 3 biological replicates. At the end I got a clear trend between my results from RNA-seq and qRT-PCR in terms of up- and down-regulated genes.

Remember that you have some fouls positive each time you perform RNA-seq experiments. This chance also increase if your treatment effect are not so pronounced over the control. If you main hypothesis depends on the genes you see on your RNA-seq then you should really confirm on qPCR. I have seen your data now and to me it looks very good. As long as you get the same answer you should be happy men. For microRNA or RNA-seq you often use median as a measure. Spiking experiments should help if you are really interested to know exact fold change. Good luck there!

I have also had the experience that the correlation between RNA-seq and qPCR results can be rather low.

While I wholeheartedly agree that the likely cause is bias generated by either or both of the two techniques, there were some clear trends in my data: 1) genes with low RNA-seq counts (e.g. less than 100 reads per sample) tended not to be confirmed by qPCR and 2) genes with a fold change of less than 2 were also not confirmed. I know that fold change is a very unpopular measure of quality these days, but I found this to be an extremely consistent predictor of qPCR confirmation for my data.

So perhaps a way of increasing the likelihood that you are generating robust (platform agnostic) results is to consider the issues of read depth and fold change during the analysis of your RNA-seq data. You will certainly increase the number of false negatives, but it might be a very useful way of generating a list of genes that are more likely to be confirmed.

Hello Helen Gunter,

Can you expand on this idea of using read depth and fold change to get a gene list, as that's exactly what I'll have to be doing shortly.

Hi James Gill,

Sorry about the delay - I had to dig through some of my archival data as this analysis did not make the cut for my publication on this topic. From the beginning I'll say that I did not formally incorporate fold change or moderately low coverage into an RNA-seq analysis pipeline, but rather I noticed an interesting trend in my data, which I hope is helpful to you.

I used RNA-seq to compare gene expression in two conditions - let's call them treatment and control (for convenience and anonymity). I used 5 biological replicates for treatment and 5 for control in the RNA-seq. Results were anlaysed with a combination of EdgeR, DESeq and baySeq, resulting in a gene list of 16 differentially expressed genes. Expression of each of these was analysed using qPCR, with 12 treatment and 12 control individuals. Of the 16 genes analysed, 8 were confirmed to be differentially expressed using qPCR. Interestingly, when I compared the results of the t-test to the mean fold change (treatment:control) for samples in the RNA-seq, I saw a clear cutoff: all genes with a ratio above 1.87 or below 0.406 were confirmed by qPCR, and those with a ratio between these two values were not (see below).

bootstrap t-test results

gene p-value treatment:control

1 0.026 3.578

2 0.029 2.455

3 0.030 2.229

4 0.008 1.870

5 0.106 1.417

6 0.084 1.273

7 0.369 1.159

8 0.506 1.074

9 0.694 1.037

10 0.839 1.025

11 0.295 0.830

12 0.128 0.655

13 0.001 0.406

14 0.002 0.393

15 0.001 0.315

16 0.018 0.122

We also performed spearmann correlation analyses, and found similar results (genes 6-11 were also not confirmed by spearmann analysis). Interestingly, genes 8-10 also displayed moderately low coverage (below 100 reads per sample).

Of course our analyses could have been improved to remove these samples initially - in fact we no longer use baySeq as the algorithm seemed not to pay much attention to fold change, and additionally seemed to generate a high rate of false negatives in our analyses.

Now I take these variables into account when selecting genes from RNA-seq gene lists for further analysis, but I have never built them into the analysis pipeline.

It should not bee too difficult to build this into analysis pipeline - would be good to hear how you go!

Cheers, Helen

Thanks Helen,

Your explanation was helpful, and I'm sure I'll be posting questions on this and SeqAnswer forums when it comes time to analyze the data. This just goes to show how important independent empirical validation is to getting meaningful results as only half of your filtered dataset made the cut. I'll look forward to reading your manuscript when it's published.

Hi James,

The project is in fact already published (but the results are not the same as the ones I have presented to you - we changed the analysis to exclude bayseq and selected different genes for qPCR, in part, due to the concerns that I outlined).

The work is published in Molecular Ecology:

http://onlinelibrary.wiley.com/doi/10.1111/mec.12417/full

Best of luck with your analyses!

Cheers, Helen

Coincidentally, we're also studying perciformes. As a non-model, gene discovery had been limited to generating qPCR primers from predicted sequences with decent success however without genome-wide gene expression it's impossible to know what we're missing. Thanks for the paper!

Hi Bernardo,

I have seen your results and they were quite good in trems of correlation between qpcr and the RNA-seq. There are many cases the correlations are not so well.

Although qPCR and RNA-seq are both used to measure gene expression, the unit of measurement is different. The fold change should not be expected to have the same for two methods. This is due the normalization process of data analysis.

In qPCR. We use reference genes for normalization. We assume the expression of those reference genes is constant for all the samples. The measurement of expression is relative to the expression of reference gene.

However in RNA-SEQ data analysis, in most case we assume each sample has the same total expressed mRNA, this can be expressed as read count per million mapped reads or RPKM.

When the expression difference does exist with sufficient significance, both qpcr and RNA-SEQ can detect the difference. For subtle difference, maybe only one method detects the difference as qPCR and RNA-seq use different scale .

Correlation is more suitable to compare qpcr and RNA-SEQ than the statical testing of differential expressed genes between treatments. Because the later depends on the number of biological replicates of each treatment and overall variation of expression for particular gene.

What is your starting material?  cell culture?