Just to get things straight: are you actually looking at the expression of your reference genes, here? As in, the question is "why are my reference genes so divergent, and is this a problem"?

If so, the answer is pretty much: no, not a problem. RNA extraction and cDNA synthesis is a long winded multistep process with plenty of opportunity for small amounts of variation to accumulate into fairly large differences....and this is exactly WHY we use reference genes. If the process was entirely 100% efficient every time, and all reference genes hovered between 0.95 and 1.05 of each other, there'd be very little point in bothering with reference genes at all.

I am a bit confused as to what your graph actually shows, though. The Y axis seems to suggest it's WT/Mutant, i.e. your values are ratios? If so, how are you pairing samples to generate these ratios? And why are the mutant values so much higher than the WT values in treatment 2? This would be more concerning that the variation in 3-5, if it IS your reference genes plotted here.

A final point to consider is that plotting data as "relative gene expression" as you have here is misleading, as it's essentially a chart that SHOULD be mirrored around 1.0, but that goes from 1-infinity in one direction, but only 1-0 in the other.

A ten-fold reduction would be 0.1, and a five-fold reduction would be 0.2 -these would seem close together on the graph. 

Conversely, a ten fold increase, and a five fold increase (i.e. the same changes, but in the other direction) would be 10 and 5 respectively, which would seem much further apart on the graph.

So either be aware of this, or use log plotting.

Hi John,

Thanks for the response.

The values(on Y axis) plotted here are the ratios of the normalized expression level of the gene of interest (GoI) in WT Vs. mutant under five different conditions (treatments on the X axis). For normalization, two reference genes are used (Reference 1 and Reference 2 on the X axis) for each sample. SO, it is NOT the reference gene which is plotted here. 

How the samples are paired?: The normalized value (normalized to each of the reference genes independently) for the expression of GoI is calculated for each sample (in equal replicates for both WT and mutant). Each WT is then paired with each mutant in an nXn matrix and this ratio is directly plotted on the graph.

Log plot: Thanks for this suggestion. Though aware of this issue, I ignored it for the sake of simplicity, however, I think log plot SHOULD be used.

Thanks

Yes, use log fold-changes.

My concern is:

Why have you normalized your data separately for the two reference genes?

If both genes are validated and if you actually have measured both, then you should combine these measurements to get a more robust normalization. Practically this means to average the log expression of the reference genes. If you can assume equal amplification efficiencies for these genes you can even simply average the Ct values.

Then to the variability:

1) It looks like normal biological+technal variability one typically faces in suchexperiments.

2) Considering the two outer points in lane 2 as "outliers" is adventurous. If 2/7 = 29% of the values are outliers then I would not trust the whole data set. If there is no external reason to see just these values as outliers then take them seriousely. It is as likely that the other 5 points are just accedentally lying too close together.

3) it is really difficult (=unreliable) to estimate or judge the degree of variability, especially from little data.

4) The low variability in lanes 2+3 is really expected: the errors in such settings (gene activity, real-time-PCR) are multiplicative and not additive. Therefore the variance is a function of the expected value (mean): lower means -> lower variability. This is just what you see here. A multiplicative error model is converted into a (conventional) additive model on the log-scale. Therefore you should do all analyses (that typically are based on an additive error model) on the log scale.

To me it looks like a simple case that whatever treatments 3-5 are, they do not affect your system as cleanly as treatment 2. Either through deficiencies in the protocol somewhere or lot's of off-target effects.

There are five treatments here.

Is it possibly that simply, the reference genes are really not optimum for the other three treatments and so should be tested specifically for those before being used? Remember reference gene performance is dependent on the effect of the treatment in question to the processes these genes are involved in.

Just a thought, while I agree to the suggestions above.

Aman

Thank you all for the response and suggestions. It helped me a lot to make sense of the data. I did plot the data on the log axis (both log10 and log2) and it did help. The Treatment 2 is now showing a similar spread as the other treatments. The other four treatments are mirroring at 1. Two sample t-test will also help in testing the validity of the data.

Thanks a lot.