Hi Rita,

yes I like it. Sometimes reviewers complain about this, other reviewers don't bother. However, the most intense opposition usually comes from the (co-)authors; they often insist showing fold-changes.

I do not see what makes a "fold-change" biologically more meaningful than a ddct value. The cell doesn't care in what units we express our measures. Example: You measure a gene induction that is 2.4-fold. So what? Nothing much can be sais about the biological meaning. It might be very easy for the cell to compensate fo this so that the physilogical effect is negligible. Or it might be a very important switch altering the entire physiological state of the cell... Eventually, we do not get less insigt by reporting what we actually *measured* (i.e. cycles!). On the other hand when you have the regulation of several genes you may find that gene A is 1.5-fold and gene B is 3.8-fold regulated. All you can see is that the induction of B was stronger. If the regulation of B is somehow more important/relevant than A can't be concluded for the same resasons as I gave above. So again it doe not really provide any more information than just giving the ddct values straight away.

There are the reasons - in my opinion - why providing fold-shanges is not better than providing ddcts. But there are two reasons why I find providing fold-changes is worst:

1) Turning ddcts into fold-changes (e.g. as 2^ddct) gives the reader a measure ("fold-change") what he/she *believes* to understand. They start overinterpreting things.

2) Fold-changes have numeric values that do not reflect biologically effects on a linear scale. A fold-change of 4 is a large numeric value, indicating a strong response; a fold-change of 0.25 is a small numeric value, indication a weak response... I find it resonable to consider both of biologically equal importance (4-times up or 4-times down). Further, it complicates comparisons. For instance take the genes A1 and A2 with fold-changes of 0.125 and 0.25, and genes B1 and B2 with 4 and 8. Intuitively we again consider the changes in B more relevant than in A, but again this is just the opposite direction of regulation. Finally, our expectations about errors on the scale of fold-changes is not symmetric. This makes it hard to interpret (calculate, report and plot) uncertainties.

Point 2 can be resolved simply by taking the logarithms. But - hah! - these are the ddct-values...

The term "relative mRNA expression" for the y-axis has passed the scrutiny of reviewers in the past. Or if comparing to a calibrator (like it sounds you are), "fold change in mRNA expression" or fold difference compared to control (or calibrator sample) can work as well. These are all unit-less numbers; values that are relative to one another in normal scale.

Indeed, so far I have seen a lot of units on the y-axis, as well as many different ways to express the differences on the trancript levels. Because, sometimes it is used calibrator sample, other time the values are logaritmized, and thats why I have some doubts to decipher the best way to put my results.

Thank you so much for your time.

The primary result of real-time PCRs are ct values, delta-ct's are the normalized values that might be compared between different groups. If showing differences in (normalized) expressions I would recommend to show the delta-ct values; the axis label should then be "delta ct", the units are "cycles".

Dear Jochen, you normally do like that? But in that way, it is not reflected the differences in a "biological-perspective", I mean with differences being illustrated by fold-difference (I think.. not sure).

Having a difference on the expression specified by number of cycles, althought i know that the differences between groups would be the same, my question is if people out of qPCR methodology will understand the biological meaning (fold difference). Your data submited by nr of cycles is successfull admited by peers-reviewers?

Thank you so much.

Hi Rita, After I have normalised to a couple of genes, my y-axis is usually labelled "Relative expression", no units mentioned. I sometimes mention in the axis if it was schizophrenia/control, treated/untreated etc ..

Hi Rita,

yes I like it. Sometimes reviewers complain about this, other reviewers don't bother. However, the most intense opposition usually comes from the (co-)authors; they often insist showing fold-changes.

I do not see what makes a "fold-change" biologically more meaningful than a ddct value. The cell doesn't care in what units we express our measures. Example: You measure a gene induction that is 2.4-fold. So what? Nothing much can be sais about the biological meaning. It might be very easy for the cell to compensate fo this so that the physilogical effect is negligible. Or it might be a very important switch altering the entire physiological state of the cell... Eventually, we do not get less insigt by reporting what we actually *measured* (i.e. cycles!). On the other hand when you have the regulation of several genes you may find that gene A is 1.5-fold and gene B is 3.8-fold regulated. All you can see is that the induction of B was stronger. If the regulation of B is somehow more important/relevant than A can't be concluded for the same resasons as I gave above. So again it doe not really provide any more information than just giving the ddct values straight away.

There are the reasons - in my opinion - why providing fold-shanges is not better than providing ddcts. But there are two reasons why I find providing fold-changes is worst:

1) Turning ddcts into fold-changes (e.g. as 2^ddct) gives the reader a measure ("fold-change") what he/she *believes* to understand. They start overinterpreting things.

2) Fold-changes have numeric values that do not reflect biologically effects on a linear scale. A fold-change of 4 is a large numeric value, indicating a strong response; a fold-change of 0.25 is a small numeric value, indication a weak response... I find it resonable to consider both of biologically equal importance (4-times up or 4-times down). Further, it complicates comparisons. For instance take the genes A1 and A2 with fold-changes of 0.125 and 0.25, and genes B1 and B2 with 4 and 8. Intuitively we again consider the changes in B more relevant than in A, but again this is just the opposite direction of regulation. Finally, our expectations about errors on the scale of fold-changes is not symmetric. This makes it hard to interpret (calculate, report and plot) uncertainties.

Point 2 can be resolved simply by taking the logarithms. But - hah! - these are the ddct-values...

I think we all agree here then - that there are quality opinions that disagree for good reasons on this. One argument is for what has been accepted by reviewers and/or pushed by co-authors' comfort levels, the other is for a 'new' (or over-looked) way to present data in ddCq scale wherein the stats are more comfortable/straight-forward/logical/attainable. This issue is not solved yet; is where things are at.

The additional (3000-lb Gorilla in the room) consideration is appropriate Efficiency correction for all studies. Which brings up the entire debate of how 'best' to calculate/estimate efficiencies of target and ref. gene amplifications. E.g. Reaction by reaction via sigmoid modeling? Cy0 Method?, Standard curve method?, etc.

So - it is either go with the flow (using log-transformed Cq values to obtain 'folk-friendly' normal/linear scale values) or boldly try to change what is generally accepted in the literature by pushing the 'more logical'/stat-friendly Cq scale.

'Petitioning' the MIQE consortium might be a good way to influence this issue as it grows in its influence on editors as to the way qPCR/RT-qPCR data should be represented on into the future.

Hi Jack :)

I essentially agree. However, I am still of the opinion that when you have to consider differences in efficiencies, then you should trash your reaction and results anyway (or make not more than semi-quantitative statements about very, very strong effects). If the qPCR does't work next to perfect even when pipetted wrongly and in the precense of dust and dirt, I wouldn't trust the results, especially "efficiency corrected" ones. Before any measurable products accumulate, there are polymerase, primers, and nucleotides in vast excess, the buffer conditions (pH, Mg++, Pi) are ideal, the amplicon sequence is short (around just 100 bp typically)... when, under such conditions, the polymerase can't copy each strand to its end... why should it constantly(!) transcribe only say 90% of the molecules? How could I trust a reliable correlation of product amplification and initial amount?

All good points -- I think that's why droplet digital PCR may be a way out of this. Although, on the surface, with qPCR/RT-qPCR, it looks like we have captured lightning in a man-made bottle, no one truly knows the contours of that lightning. Too many unknowns and ways to go astray. (And actually the reason MIQE was created in the first place I think; since the entire process is so very error-prone).

Yes, yes, and yes. But I forsee that ddqPCR will create its own new problems...

Hey Rita,

I brood about this question myself. So far I think throwing it all in the pot would be the best: presenting the ddct normalized to your calibrator on a logarithmic scale. And then label it with "Relative expression" as Natalie has mentioned. I think it is very important to point out on the description of your graph how you normalized your data (and in what order). Thats at least what I like to know as a reader.

Good luck.

qPCR data analysis is really tricky. More you know and read about it, more doubts we get!! Thank you all for this discussion =)

Yes, and as Jo Vandesompele et al., say so eloquently in their 2002 publication: "The ideal reference gene does not exist" and therefore and so on ...

Based on the constant number of krill in any given teraliter of seawater, can we really, truly draw a parallel to the changing number of porpoises per teraliter of seawater ... and so on.   We are still no nearer to that assumption to this very day - agreed.  So, the only way out of this mess, is to always...always...merely write exactly what you actually did, so that the rest of us know how the conclusions were actually arrived at..

One person (Dr. Amy Replogle) created the attached file based on Vandesompele et al., algorithms.  Here, the SE bars are shown as symmetric entities in the file (for fold change). Now corrected on 7-17-17.

But when linear (median) quantities are expressed in linear scale (by log transformation from Cq scale), these error ranges take on asymmetric (upper and lower) error ranges above and below the median quantity value based on the relationships shown in the second of the attached files below. When and where to ignore error along the way is something that still needs to be clarified.

Dear Jack M Gallup,

Firs of all, many thanks for the excel file uploaded. But still, one question is remaining for me. I am still curious how it is possible to calculate a dozen of Ct values obtained from for example 10 cases and 15 controls using Pfaffl method. The problem is control Ct values in my view. It would be OK if we have only one control Ct (or an average amount of Ct obtained from 10 cases). in this case, we are not able to calculate using such an Excel file. Also, I am wondering if it does takes place a statistical error if we get average +/_ SD(SE) from control group Cts, and then calculate individual Ct values from treatment. Does this excel file is appropriate for this or not?

Thanks in advance

The files you're referring to represent only some of the possible approaches in all of this. The files are meant to be manually expanded by the user to customize the parameters to their own unique situation. They are just a starting point, not a 'cure-all'. The permutations of the possible ways to calculate and/or represent qPCR data are all over the map. The MIQE Guidelines suggest ways to do this as well. And software by Pfaffl and Biogazelle also attempt to tackle this issue. Good data processing/generation/interpretation is the stalwart quest in qPCR. 

Attached here is another file which demonstrates several other ways it has been attempted. In the end, we all should stick with what we can explain clearly in a publication.  If you come up with a novel way to do it, you should publish that method as well. 

Tracking the true propagation of error throughout the entire qPCR process is for expert statisticians to illuminate.  I am not an expert statistician, so, I, too, can only study the various ways qPCR data processing has been done, and then decide which approach is most appropriate for my particular situation. Some Cq values in some studies are easier to work with than in others.

Generating trustworthy Cq values in the first place, at near 100% efficiency, is always the best place to start.  From there, it is up to each of us to choose the mathematical approach that best suits the given situation.  Some feel strongly at ddCq values are the way to express everything. Others feel that reference gene normalized relative quantities are the way to go. Others feel that fold-change (normalized or not normalized to a reference gene) is the best way to express final data.

In my opinion, it is still the 'wild west' on this issue. Everyone eventually convinces themselves that their approach is the best.  My main weapon is a program I wrote called PREXCEL-Q which helps solve all potential problems I might have with qPCR (it proves that my Cq values have been generated at a trustworthy dilution at an acceptably high efficiency of amplification). From there, the processing I do after that relies directly on the equation I uncovered: 

EAMP(b - Cq)

and that's it. The relative quantities I generate from that equation for a treatment group I then take the geometric mean of so that error bars can theoretically stay symmetrical about the geometric mean in the final graphing (the geometric mean of relative quantities reverse translates back to the arithmetic mean of the Cq values in a treatment group). 

Additionally, I often avoid the use of reference genes by loading the exact same amount of material per each qPCReaction (at a PREXCEL-Q-proven non-inhibitory dilution for all targets tested; with all target Cq values being measured at proven high amplification efficiencies; efficiencies which I carefully estimate for each experiment with appropriately-rendered standard curves for all targets involved).

Not sure if these ideas help in answering your questions - but I hope perhaps they provide some things that may be of use to you.

An interesting other take on these things (see attached file) ...