Many differential expression packages out there require the raw_count as their input but I want to use my normalized values as input. From what I can find limma+voom can handle normalized values. Are there any other package that allows me to use normalized values?
I've tried using DESeq2 despite reading various comments saying it only works with raw_count values but when I looked at the fold change values and manually inspecting the normalized values, the fold change output seems pretty reasonable. Why is it that you can't use normalized values with DESeq2?
Thanks
•
written
by
john •
As it is explained in the vignette, DESeq2 accepts the input as a matrix of integer values. First, it does not say you can't it says you should use un-normalized counts. Second, there is always a chance that a normalization step could disrupt the count data format that could cause other read or analysis errors downstream in DESeq. It is always good to trust the vignette if you are new to the tool.
For reference, until recently you actually couldn't input non-integer counts into DESeq2, it would throw an error. I think the change was made around the time that the tximport package was created, since then fractional values would be correct.
Log in to answer this question.
How exactly did you normalize your counts? The reason packages say not to do this is that historically people have done it in incorrect ways...
My normalized values are ratio. The ratio is generated by coverage value (mean,median, etc.) of region1 divided by region2 of the same genes. I'm trying to identify regions comparison ratios that change significantly from normal to cancer
I'm not sure that would have a negative-binomial mean-dispersion relationship. Personally I would reformulate the regions as the same gene in paired samples, which would be asking essentially the same question but without masking the underlying counts.