dispersion from DESeq

[giu.facchetti]

hello,
I have started working with DESeq for RNASeq data from plants.
Among the first plots I have got, there is the dispersion (D) vs counts mean (M). This plot is fitted with the curve D=a+b/M (which in a log-log scale is a decreasing curve that reaches a plateau for high value of M, see attachment).
However, my data show a linear correlation (still in the log-log plot).
Is my analysis wrong? can I change the fitting expression? why the fitting procedure does not set a=0 which would give a linear relationship between Log(D) and Log(M)?

thanks!

[Wolfgang Huber]

Dear Giu.facchetti,

thank you for the feedback. Some suggestions:
1. The attached jpeg image came out very small (100x115x) making it hard to impossible to read (at least for me).
2. Such a strange-looking plot may (or may not) be indicative of a data quality problem, e.g. large batch effect or sample-swap. Have you explored that? (The DESeq2 vignette gives some suggestions on how to do that.)
3. Have you tried the argument fitType = "local" to estimateDispersions?
4. Since you are just starting, the authors of the package recommend you use DESeq2 now, since it offers many improvements.

Kind regards
Wolfgang

[giu.facchetti]

Thank you for the suggestions!
Unfortunately the option "fit=local" did not improve the fitting (the fitting line was even much curved). Same results with DESeq2.
At the moment I am doing again the alignment and reads counts, trying to improve the sensitivity and accuracy.