Effect of outlier sample on CuffDiff analysis?

[cw11]

Unless I'm much mistaken, CuffDiff pools reads from all samples within a condition prior to differential expression analysis. Does anyone understand the method by which CuffDiff prevents an outlier (for instance, one sample overexpressing a gene significantly more than the others) from skewing the results? How reliable have other users found this method to be?

(I did read http://cufflinks.cbcb.umd.edu/howitworks - just having some difficulty understanding it).

[mehc]

I would like to know this also. We 'fixed' the cases with DESeq where one replicate had abnormal read counts vs the other replicates and made the genes highly significative (ex 0,0,0,high vs 0,0,0,0). However I'm wondering if there is something similar for cuffdiff, as the top transcript results still contain these cases

Thanks

[mehc]

bump, anyone know where else I could ask this? (e-mailed cufflinks, no answer)

[Dameon]

I believe that Cuffdiff uses LocFit Regression in the case were there are no biological replicates to estimate dispersion with inferences made from a Poisson distribution. I do not believe that Cuffdiff takes any special consideration of outlier samples during this step except to use just those genes that are rank invariant and not differentially expressed across all your samples. In the case of biological replicates, I think Cuffdiff estimates the over-dispersion based on a negative binomial model from within each sample group. Again, I don't think Cuffdiff takes any special consideration of outliers during this step either. Personally, I think if your data contains outliers, you would be better off using something like DESeq or EdgeR where you can build a generalized linear model to account for such outliers or exclude them entirely during your sample QC process (R plots). Lastly, if your dataset does not contain biological replicates, I would hold off on doing any analysis until you do; otherwise, your data would have little value due to low statistical power to make inferences.