First. I have significant concern about having only 2 biological replicates...which is the level with which we really want to make our inferences...that's 1 degree of freedom. yuck. My rule of thumb for my biologist colleagues is 5-8 minimum per experimental condition.

My general approach of late:

That being said, I've been analyzing CT directly with mixed models. You'd want to structure the data, random effects and variance-covariance matrices to reflect the particular protocols. For example, you'd want a well-level covariance between Ct from multitplex (along with biological sample-level)...

Watch your housekeepers...if you're using multiple housekeepers and they're not being regulated by the experimental conditions, then complex orthogonal contrasts can be used to "normalize" these as equally weighted grouped housekeepers if desired, though you may rather want to pick the one(s) with the most comparable Ct to the target gene/mRNA...

The differences of the differences (i.e. group differences normalized to housekeeper) can be used to create the standard fold differences by raising 2 to the negative power of the mean delta delta...and you're right back into the standard language, but without doing all that averaging before analysis.

proc glimmix, for example also gives you a convenient tool for alpha maintenance, as well...perhaps one of the stepdown procedures.

Lastly, you would probably want to consider usign sandwich estimation to deal with whatever deviation the data shows from the chosen variance-covariance matix (model misspecification).

For your SPECIFIC case...you may want to go in a completely different direction. You want to know if gene B changes as gene A changes...don't overthink that. Have you explored simple Pearson correlations between their Ct a a start? Then, follow that conceptual framework, but model it using random effects, as well...or correlated error...just account for the nesting in some way. Now the limitation of 2 biological replicates will REALLY be apparent...you don't want to draw a straight line between 2 points...zero df.

I STRONGLY urge you to add more biological replicates...even at the expense of technical replicates.

There are a lot of sources of variability making results from different qPCR plate consistently different. For this reason it would be possible combine results from different plates only if plates share a sample fully analyzed in its technical replicates on both plates. In this case you could refer changes in the expression within each plates to the common sample as calibrator. Otherwise I would have some doubt in the convenience in combining data from different plates.

First. I have significant concern about having only 2 biological replicates...which is the level with which we really want to make our inferences...that's 1 degree of freedom. yuck. My rule of thumb for my biologist colleagues is 5-8 minimum per experimental condition.

My general approach of late:

That being said, I've been analyzing CT directly with mixed models. You'd want to structure the data, random effects and variance-covariance matrices to reflect the particular protocols. For example, you'd want a well-level covariance between Ct from multitplex (along with biological sample-level)...

Watch your housekeepers...if you're using multiple housekeepers and they're not being regulated by the experimental conditions, then complex orthogonal contrasts can be used to "normalize" these as equally weighted grouped housekeepers if desired, though you may rather want to pick the one(s) with the most comparable Ct to the target gene/mRNA...

The differences of the differences (i.e. group differences normalized to housekeeper) can be used to create the standard fold differences by raising 2 to the negative power of the mean delta delta...and you're right back into the standard language, but without doing all that averaging before analysis.

proc glimmix, for example also gives you a convenient tool for alpha maintenance, as well...perhaps one of the stepdown procedures.

Lastly, you would probably want to consider usign sandwich estimation to deal with whatever deviation the data shows from the chosen variance-covariance matix (model misspecification).

For your SPECIFIC case...you may want to go in a completely different direction. You want to know if gene B changes as gene A changes...don't overthink that. Have you explored simple Pearson correlations between their Ct a a start? Then, follow that conceptual framework, but model it using random effects, as well...or correlated error...just account for the nesting in some way. Now the limitation of 2 biological replicates will REALLY be apparent...you don't want to draw a straight line between 2 points...zero df.

I STRONGLY urge you to add more biological replicates...even at the expense of technical replicates.

I've worked with biological samples from several different plates. As Antonio Di Matteo noted above, plates can differ widely. That's no guarantee that your results will vary greatly from plate to plate, but it's a reasonable concern.

I would suggest that you block on plate. This can be done if you're doing a relatively simple analysis of variance, or if you're doing a more involved hierarchical analysis. Tests for block differences can be performed. And, Jason provided a number of good ideas above.

But before you jump into any sort of modeling, look at your data: simple side-by-side boxplots by plate will give you an idea of how much between-plate variation is there. Look for plates with distributional shifts up or down, relative to the other plates, and for plates with far more (or far less) variation within them, again compared to the other plates.

The issue with more replicates is a practical one - time and cost. The experimental set-up is actually a little more complicated than I described. Each biological replicate involves 24 RNA samples from which 12-16 wells are run. The cost of the reagents alone is a limiting factor and the RNA is quite time-consuming to collect.

Thank you for your responses

Could Elizabeth do something simple, like calculate each well on a plate as a percent of a control well? If she is getting the same pattern of upregulation in different plates, this seems to make sense to me.

Oh, wait. I think Antonio already suggested that. He just did so in a way that I didn't understand in my pre-caffeinated state.

Not directly helpful for this case because it is too late but it is good to have a "day to day" control in each run. These values could be used to rule out technical and chemical variances that are happening during each run.

In your case, if the efficiency of the genes of interest and the reference genes are similar, I would not worry too much about changes in the cT values. You can calculate a %RG value for each sample. You can also try to make a manual adjustment of the threshold line for all your runs (by adding a constant difference on top of the background florescence or by placing the threshold line right in the middle of linear amplification zone for example).

Dear Elizabeth,

Yes, if you want to make an accurate analysis of the results of your study and running a real-time analysis, the ideal is that you also do the analysis with the standard curve. This will give you a result closer to the real advancement of the contents of your analysis were happening outside of an organism, as it probably should be going on in your lab.

Good luck and devote yourself to your search. Success in his analysis and I hope it has helped you,

Livia Valentin

Just a straight question, to rule out, when you are indicating variability between plates, is this significantly affecting Cts of your housekeeping/reference genes? If yes, make sure that your experimental setup is not actually affecting those levels. If this is not the case, I would proceed with very to-the-point advice stated above.

If you take efficiency into account (on each plate, for each target), and adjust all Cq values to an initial, fixed starting point, the tale should spell itself out on a delta-Cq basis; in theory. Other than that, Kearney Gunsalus' answer is an excellent perspective.

Without knowing the efficiency of target amplifications on each plate (whether it differs or not), you may be reaching in the dark a bit (or quite a bit).

Your original question is seeking to combine biological replicate delta-Cq values is it not? If so, the fixed starting point idea should work. After all, the entire scale is relative.

The crux will be how much you, yourself, believe in the way you have measured each of your target and reference gene efficiencies. This is always the crux.

I personally have never experienced interplate variation as well -- but I weigh water on an analytical scale with every pipet adjustment (before reagent or sample administration) - and that has really helped everything.