Unequal number of technical and biological replicates per sample for RNA-Seq expriment?

Question

[gaelgarcia]

Hi,

I performed an RNAseq experiment for two conditions, with 4 biological replicates for Condition 1 and 2 biological replicates for Condition 2.

Furthermore, I have three technical replicates for 2 of the biological replicates in Condition 1, and no technical replicates for the other two biological replicates of Condition 1 (only one instance of each).

For Condition 2, biological replicate 1 had two technical replicates, while biological replicate 2 had three technical replicates.

Will this unevenness in replicates (both technical and biological) affect the downstream analysis?

I am using DESeq2 and collapsing technical replicates by adding the raw counts into a single sample, as explained in the DESeq manual.

Any input is appreciated. Thank you.

Answer 1

[alolex]

First I want to say I am not a biostatistician, so please consult a biostatistician for details about the statistics of your particular situation. With that said, you should always determine the number of replicates and the statistical tests you are going to use BEFORE you run the experiment. This should be part of the experimental design process so that you can ensure you have enough replicates to get the power you need to detect changes. However, sometimes things go wrong and you get shorted data. I'm not sure what happened here, so I am assuming this is all you have to work with.

Yes, you should always combine technical replicates. I don't think the unevenness will cause problems, but the lack of replicates will. Generally I prefer 3 or more. Having only 2 biological replicates in one condition may adversely affect your power to detect real changes. Please consult with a biostatistician on this if you plan to publish using only 2 replicates for one condition; otherwise I think it is ok for exploratory analysis to see if anything really big jumps out at you. Others may have different thoughts and I implore them to chime in here.

Answer 2

[Irsan]

In your situation you should use something that implements mixed-effects modeling beause you have both technical and biological replicates. Try for example ShrinkBayes. I wouldn't care too much about Power analyses because you have to make a dozen of assumptions and most of the time the study is limited by money and sample availability anyway. It can be very helpful though to contact a bio-informatician/statistician before starting your experiments for many other reasons. I would have told you to forget about the technical replicates.