Comparison between different batches and global analysis

[Matina]

Hello everyone,

I have RNA-seq data coming from control, breast cancer and endometrial cancer patients. Is an ongoing study and so far we've sequenced 15 patients (5 control, 5 breast cancer and 5 endometrial cancer) about a year ago and 5 patients (1 control, 2 breast cancer and 2 endometrial cancer) a month ago.

I want to compare between the independent batches and also combine them and do a global analysis. I did an independent DE analysis on each batch using EdgeR and DESeq2 however when i compare the gene lists I have very few common DE genes. Do I need to do some kind of normalisation before i compare them? About the global analysis, I will put everything together but how can I correct for batch effects?

Thank you so much!

[blancha]

Just a general comment first.
I've have very poor results regarding correlation between human clinical samples.
There are too many variables that are not controlled for.
Often, when you enquire about the collection method of the samples, you'll find that they are not identical for all samples, e.g. some samples were frozen for a long period of time, the tumor samples contain different proportions of normal tissue, ...
I've often been unpleasantly surprised that the samples that were described to me as replicates actually had significant differences.
The best results are generally obtained when the person collecting the samples is experienced in RNA-Seq analysis, and is aware how the collection and preparation of the samples can affect the downstream analysis.

There are also variations between cancers, but that is biologically relevant.

As an analyst, you need to get as much information as possible about the collection method of the samples, and collect as much metrics as possible to better understand the samples (e.g. RNA integrity number for all the samples, exotic alignment rate, ...).

-----

Second, you have few or no replicates in your second batch so you would expect DESeq and edgeR to give you a different list of differentially expressed genes. In the second batch for example, you have no control replicates which will affect the calculation of the dispersion level by DESeq and edgeR.

-----

To answer your question though, I would first try clustering the samples using the normalized counts. If the samples do no cluster according to the tissue of origin but rather according to the batch, I would try removing the batch effect with ComBat. I would then repeat the clustering to verify if ComBat has successfully removed the batch effect.

[Matina]

Thank you very much for you answer blancha.
I'll try using comBat!