Outliers management using DESeq

[jorgebm]

Hello,


Due my lack in statistical skills I would like you to ask for advice:

If I understood it properly, DESeq manages outliers at "gene" level. I mean, large variance makes a gene "harder" to be differential expressed (specially using "sharingmode=maximun") ; so, it prevents false positives.

But what happens If having some samples; one of this samples is quite diferent compared with the remaining ones of its condition. This sample is an outlier and induces to DESeq to avoid several otherwise (excluding this sample) truly diferential expresed genes.

As I understand It's a must to perform a preocedure to exclude sample outliers from conditions If we (are lucky) to have several samples for each condition. doesn't it?

In my case, for an RNA-Seq experiment, I've modestly large number of samples (in my case 40 replicates for condition A and 36 replicates for condition B). What procedure do you recomend to prepare sample sets (filter outliers) to feed DESeq?.

Thanks in advance

[Simon Anders]

With such an unusually large number of samples, you can indeed afford to remove outlier samples that may compromise your detection power. Cluster your samples using the variance stabilizing transformation, as described in the DESeq vignette to identify outlier samples. Furthermore, once you have more than, say, 10 or 15 degrees of freedom for variance estimation, you might be better off not using the fitted data at all (sharingMode="gene-est-only").

[jorgebm]

Thank you, Simon.

I'll try what you say.

[Wolfgang Huber]

Hi Jorge

the arrayQualityMetrics package might provide some help in rationalising the hunt for outliers (it does not fully automate it, in the end there is still a human judgement call). See e.g. http://permalink.gmane.org/gmane.sci...onductor/40146 for some hints.

Best wishes
Wolfgang

[jorgebm]

Thank you, Wolfgang.

Just now I'm busy with another (priorized) task but soon I'll back yo this issue...probably with a lot of dubts