I conducted qPCR to analyse the gene expression of 5 genes. I used GAPDH as the reference/control gene. Everything was run in triplicate.
I have calculated the delta-Ct, delta-delta-Ct values and fold-changes (and log-fold changes) for each gene. I am wondering what is the correct way to determine significance in gene expression changes between treated samples and controls?
I have been using Excel (but also have access to Prism). I did a two-tailed unpaired t-test on the both thendelta-delta Ct values and the fold-changes, but I get different results in each. More results are significant when comparing fold-changes rather than delta-deltaCt. Which is the correct one to use? Any other suggestions?
Also, I would like to present my results as a bar chart of fold-changes. What should I use for error bars? The standard error (SEM)?
What were your response variables? Were the determined delta-Ct, delta-delta and fold changes your indices for measuring gene expression? If so, I will not recommend the use of Excel. Since you already have access to GraphPad Prism, then I recommend it. Prism will guide you on how best to present your data for the analysis. I will suggest one-way ANOVA, in which case you will take the measured parameters one by one since the units are not the same. I handle data transformation with a lot of caution, so I wont recommend it. You could compare all treatment genes with the control, in which case the Duncan post-hoc test will apply, otherwise you compare each gene against every other gene, in which case Tukey HSD will separate significant means. Make your error bars your standard deviation. Do a little reading to know when best to use standard error as error bar. All the best.
If your samples are not paired, you have two independent samples and you would test the expression difference with an (independent samples) t-test on the dCt values.
If your samples are paired, you can either use a paired t-test on the dCt values or a one-sample t-test on the ddCt values (this is equivalent).
At no time you should or need to calculate fold-changes. Don't do that. It's sufficient and correct to do all analyses using the dCt values.
To visualize the results, you can (you should) use scatterplots. For independent groups, show the dCt values by group. For paired data, there are several options:
And:
- Don't show fold-changes (note that ddCt values are log fold-changes).
- Don't use barplots. Never.
- If you use scatterplots, there is no need to add mean or sem or something else, since the data is shown, and the plot completely discloses the location and the dispesion of the data.
Have a look here for inspiration:
Article Long non-coding RNAs influence the transcriptome in pulmonar...
Jochen Wilhelm Thank you for your help. I have 5 genes, and I want to investigate their expression after treatment of mice for different lengths of time. I have mice that were treated for 1 week, 2 months, 8 months, 10 months and 12 months. There are about 5 subjects in each group. Would a scatterplot still be appropriate for this? I assume a separate one should be used for each gene of interest?
A single scatterplot for each gene would serve the purpose of showing the complete data. This comes at the cost that the time-courses are not easy to compare between genes. To achieve the latter, you can plot only the mean dCts, and connect the values from the same gene by lines (so you get a line plot). The trends of the (5) lines can then easily and directly be compared in a single diagram.
Another option is to show the dCt values in a heatmap: each row a gene, each column a measurement; the columns should be ordered according to the time (the replicate measurements within each time should just be one block, they cannot be odered in a meaningful way).
Are the mice the same for all durations, or are they different mice? Your description suggests these are different mice, but since it's not written explicitely, I may misunderstood. It changes the way to analyze the data, and to represent them.
Instead of T-tests for each time, you may consider a joint model of all times, especially if you are expecting a monotonous evolution with time. This will spare numerous tests, hence avoid loosing power because of multiple comparison issues — typically, with 5 genes and 5 times, this makes 25 T-tests, hence you should roughly use p < 0.002 to conclude to significance, and not p < 0.05... (well, it's not clear which T-test you are doing in this context, but you see the idea I guess).
Emmanuel Curis They are indeed different mice. So, for example, a set of 5 mice treated for 2 months, a different set treated for 10 months, etc.
OK. That makes the analysis a little bit more simple. The second part of the remark remains actual.
Now, the question is: when you speek about « T-tests », what are you doing exactly, to check that gene expression indeed changes with time?
Emmanuel Curis I have been doing unpaired student t-tests. I understand the point on using an even smaller p-value as errors are introduced by running t-tests on the same data.
To answer your question, I have done two things. I am not purely interested in whether gene expression changes with time, but more interested as to whether it changes at all, at any time point. So I have done unpaired t-tests comparing the control group for each gene against each treatment duration group (i.e. control vs 1w, control vs 2m, etc.)
I have, however, also run a one-way ANOVA with Tukey's test as the post-hoc multiple comparison test for each gene.
If your question is really « is there any change, at any time », then the most appropriate thing to do is a simple analysis of variance, with a single p-value associated to the explained sum of squares. That is, for each gene of interest, a single test. All other tests you are mentionning answer other questions than that...
For all times vs control, check also the Dunnett's test, it will be better than standard T-tests, and more powerful than Tukey's method.
- Badges
- Science method