Hi! I have been calculating fold changes in the gene expression using delta of delta method, and the average of the control values does not equal 1. I asked a colleague who does similar tests regularly, and he says that has always been the case for him. Is that really how it should be, or have I made a mistake somewhere?
Thank you!
You didn’t provide much information to work with.
Perhaps you have mistaken the delta-Ct method for the delta-delta-Ct method.
Maybe the video in this link will help you find out what is wrong with your calculations or your data.
(BIO-RAD Real Time QPCR Data Analysis Tutorial)
https://www.youtube.com/watch?v=tgp4bbnj-ng
Actually, you already make a mistake when you calculate a fold-change for the control group.
thanks for your answer, Marek Lubośny ! I don't think I confused the two methods, and I was using the delta of delta Ct as described in the video.
:) - a fold-change is a ratio of two concentrations, by definition comparing two different groups or conditions. A ratio needs a nominator and a denominator. It thus makes no sense to talk about a fold change of the control group, and a fold-change of a treated group. They don't have fold-changes. You can give a fold-change comparing treatment to control.
You can certainly compare a group to itself... but this is quite non-sensical. It does not give you any useful information. I don't see a point why one should do that.
Let's say you have two groups, C (control) and X (treated), then there is one single, meaningful fold-change: X vs. C. If you have a control (C) and several treatments (X, Y, Z), then meaningful fold-changes would be X vs. C, Y vs. C and Z vs. C. You might also compare the treatments directly, in which case you get fold-changes for Y vs. X, Z vs. X, and Z vs. Y. There is no fold-change for "C", "A", "X", "Y" or "Z".
This is what I meant. I hope it's a bit clearer now.
If your experimental setup is similar to the one in the video below: a few control samples + repeats for those samples. I would say the average for controls can be different from 1.
Even in this video, the average is rounded down to 1 (from 1,00402).
It depends on how far from 1 are you and how close to each other your control samples are. If you are far from 1 and you can see that 2^-delta-delta Ct for one of the controls is much different from the rest, then I would say there is something wrong with one or more of your controls or the expression of your reference gene is not stable in the context of your experiment.
https://www.youtube.com/watch?v=Kkle8T7aXjk
At 3:47 the video goes weird. All there is to calculate is
= average(J3:J5) - average(J6:J8)
This is the single one ddCt value. Since this was NOT a paired design there are no individual ddCt values for the individual samples. The calculation shown in the video is statistically incorrect and misleading.
If you like to estimate the expected fold change, the calculation would start with exponentiating the Ct values, calculating ratios and averaging these ratios (not the dCt values).
If you like to estimate the geometic expectation of the fold change, go via dCt values and STAY on that scale (do NOT exponentiate) until you have the last final mean ddCt value. In that example, this would be = 2^(average(J6:J8)-average(J3:J5)).
The mixture of using dCt in the beginning and exponentiation before averaging in the middle does give you neither the correct expected value nor the correct gemoetic expectation of the fold-change. Ok, ok, that does not make a big difference in the final result, particularily when the replicates are all similar (when there is not much technical and biological variation). But is simply messy and logically flawed to exponentiate somewhere in the middle.
Thing go even more astray when it comes to calculate standard errors, confidence intervals or p-values. Following the video there are 3 values (in column L) for each group, and it seems tempting to use these in a t-test. This would be quite suboptimal, because the values in this column are NOT normal distributed. The values in colum J are approximately normal distributed, so those could be used for a t-test and to get standard errors and confidence intervals. The whole calculation in columns K, L and O are actually worthless.
Dear Jochen Wilhelm and Marek Lubośny, thank you so much for your detailed explanations and the time you took to reply! That has helped me a lot and I really appreciate it.
Dear Jochen Wilhelm, thank you so much for your clear explanation. Can you suggest me any video to calculate delta delta ct method correctly?
Dear Prarthana Sutradhar , no, sorry, I don't know such a video. There a million videas that could serve as bad examples, but I haven't yet found one that explains it correctly.
I attached an Excel-sheet that should explain the calculation. I give a simple example (two groups) for a paired and for an unpaired analysis, using "manual" calculations as well as the LINEST function from Excel that may be better to analyze more complicated experimental designs. However, I suggest using a proper stats software for data analysis. All you need are the dCt values, the rest is a standard linear model.
Dear Jochen Wilhelm , thank you so much for your kind help. I've got my answer. Just a little query, what does ' Indicator Treated ' mean in that sheet?
The indicator (aka "dummy") variable is a variable that takes the values 0 or 1 to indicate the membership to a group. It takes the value 1 if the corresponding response value (the dCt) is from the "treatment group", and it is 0 otherwise.
In the linear predictor, this variable is multiplied by the coefficient coding the treatment effect as the expected difference in the response (the dCt values) between the indicated group (treatment) and the reference group (the group not being represented by an indicator variable; here: the control group). Nothe that the difference in dCt values is called ddCt.
It may be helpful to read some books about (general) linear models.
Jochen Wilhelm the examples exposed here are always referred to cross-sectional studies. What if you have several groups in longitudinal studies, in which individuals are followed across time (paired samples)? You cannot use the ratio; or if you're going to use it must be per individual.
I went just for using dCt values for comparing, and just substracting the later - former to watch the direction of the expression. However I'm not getting any fancy "fold-change or ratio".
Another approach I considered was to calculate ratios between groups using dCt, but it seems weird to me. I must admit, that I have 3 groups, and the group control is not a typical control with no intervention at all
For longitudinal studies I'd use a random intercept term per subject in the model (aka mixed-effects or hierarchical models).
Jochen Wilhelm This is what I do to explore main effects in covariates and its interactions (including between groups). But as far as I know the intercept gives you information about intra-individual in the baseline (= independent variables are zero)
https://stats.stackexchange.com/questions/80797/intercepts-reference-in-linear-mixed-effect-model-what-to-choose
I am talking about the direction of the expression how to depict it or express it: -dCt or 2-dCt and plotting it with time in the X-axis. In longitudinal studies:
dCtPre = Ctgoi - Ctref
dCtPost = Ctgoi - Ctredf
dCtPost_Pre = dCtPost - dCtPre
This is something not well reported about paired samples in longitudinal studies
"dCtPost_Pre = dCtPost - dCtPre" - Post-Pre is a ddCt: a difference of a difference of Ct values.
How to depict it? I'd say as -ddCt. Raising this to the power of two only makes a lot of problems.
If ddCt is this:
"dCtPost_Pre = dCtPost - dCtPre = ddCt"
which is the term when referenced to a calibrator:
dCt_Post_Pre - dC_cal = dddCt?
dCt_cal is the dCt obtained from the calibrator sample (a random untreated sample, or a pool of untreates samples)
dCt_cal = Ct_cal - Ct_ref.
First time hearing about the calibrator sample?
- Badges
- Science method