RNAseq data: how to evaluate if the replicates are good or not

[capricy]

I have two replicates for each sample and would like to downstream differential gene expression analysis. I could do some correlation analysis, either pearson or spearman. I also scatterplot them (examples attached here). Just wonder how I would tell if the replicates are consistent with each other. I mean for the examples I provided here, can I treat them as replicates?

Thank you very much for any inputs!

[simonandrews]

I don't think there is going to be a universal answer to this since the degree of correlation between replicates will very much depend on your biological system. Our normal approach is to cluster all of the replicates for all conditions and then see what pattern emerges. What you hope to see is that the replicates for a condition cluster together more closely than samples from different conditions.

By doing this sort of larger scale clustering you can also then spot samples which are outliers from the whole set and which might have technical problems. You can also potentially spot sample swaps if you have pairs of samples which appear misplaced.

If there is no real grouping then this doesn't mean the replicates are bad, it just means that the noise in your system is higher than the between condition signal (which in turn doesn't mean that you won't find any differentially expressed genes). More concerning is if the samples cluster, but by something other than the biological groupings - for example by the batch in which they were run. Again this is an indication of poor signal or high noise, but as long as you have a properly randomised design it doesn't mean that the data is unusable.

[sBeier]

In a way you do not want the replicates to be totally consistent, because then you won't have the possibility to say that a differential expression between conditions is significant, compared to the expression differences in between replicates of the same condition.

If you used cufflinks to do the differential expression analysis, you can do some nice correlation analysis with cummeRbund.