Bad fits for DESeq dispersion estimates

[cdb3]

Hi all, I'm using DESeq to identify differentially expressed genes in some human cancer samples. I'm using eXpress to estimate isoform abundance according to the Gencode v14 transcriptome and summing the effective counts from eXpress for all of the transcripts from a gene to obtain to obtain gene-level counts. These are my input to DESeq (after rounding to integers). I've noticed that the fit for from the estimateDispersions function is poor for my data (see attached). I'm looking at several different designs (i.e. different groups of samples based on different biological questions) and see similar fits.

For instance, the first fit I've attached is for 56 samples from three cancer types. 25 of these samples have a particular mutation and the other 31 do not. In this case, I'm including the cancer type as a covariate and testing for differences between the mutants and non-mutants. The second fit I've attached is for 26 samples, 14 in one group and 12 in the other.

In both cases I'm using the "gene-est-only" parameter for estimateDispersions because according to the DESeq manual, I have enough samples for this approach. I realize that the fit is therefore not affecting my downstream results. However, I'm wondering why the fit is so bad and whether this indicative of any problems. I would use DESeq2 but it seems that the gene-est-only option is no longer available, and I'm wondering if there is a reason for that.

Thanks for any help in advance.

[dpryan]

I'm using eXpress to estimate isoform abundance according to the Gencode v14 transcriptome and summing the effective counts from eXpress for all of the transcripts from a gene to obtain to obtain gene-level counts. These are my input to DESeq (after rounding to integers).

Don't do that. Summed estimated counts are not raw counts and will be unlikely to behave in an identical manner. Use featureCounts or htseq-count.

[cdb3]

I tried that but see a similar trend (see attached); still not a great fit.

[Michael Love]

hi,

I'd recommend using DESeq2 and using fitType="local" (see the vignette section on alternate dispersion fits). The parametric curve flattens out, which is a good fit for many RNA-Seq datasets but not necessarily for all, hence we provide the local regression.

Using the moderated estimates typically increases power over the gene-wise estimates. While ~30 seems like a lot of samples, there is still considerable sampling variance for estimates of dispersion, and the moderation helps stabilize these estimates.

[cdb3]

Thanks Michael. I went ahead and used the local fit with DESeq2. The fit is better of course. Interestingly I actually have more differentially expressed genes than the DESeq/gene-est-only approach. Do you see any glaring problems with this plot of the dispersions or MA plot?

[Michael Love]

Those look good to me.

re: more DE genes than using old version of DESeq & gene-est, short answer: this can happen. We gain power from sharing information across genes.