Question

local versus parametric fitType in DESeq2

1

Entering edit mode

liangc.mu ▴ 10

@liangcmu-9405

Last seen 7.7 years ago

For differential gene expression analysis of three treatments, each of which has three replicates, we find the results are different when we used fitType="local" versus fitType="parametric" in DESeq2. For the given two fitting results of dispersion estimates, can anyone tell me which graph looks better? Why is it better?

(1) fitType=local

http://bioinfolab.miamioh.edu/DESeq2.fitType.local.CTL_FGF2.png

(2) fitType=parametric

http://bioinfolab.miamioh.edu/DESeq2.fitType.parametric.CTL_FGF2.png

Your help will be highly appreciated.

deseq2 • 5.2k views

ADD COMMENT • link updated 7.7 years ago by Michael Love 43k • written 7.7 years ago by liangc.mu ▴ 10

score 0 · Answer 1 · 2017-05-01

0

Entering edit mode

Michael Love 43k

@mikelove

Last seen 6 days ago

United States

I would prefer parametric here. There is a capping of gene-wise dispersion values (you probably have a number of genes with all 0's and then a very high count for a single sample), and the parametric trend is less affected by this. The local trend is flattening on the left side.

By the way, you might consider also removing the genes which have all 0's except a single sample with a high count, with a filter such as:

keep <- rowSums(counts(dds) >= 5) >= 3
dds <- dds[keep,]

ADD COMMENT • link 7.7 years ago Michael Love 43k

0

Entering edit mode

regarding your email to me: running pairs and finding more DEG in the case that you have one or more groups with high within-group variance is consistent with my recommendation in the Frequently Asked Questions section of the DESeq2 vignette:

http://bioconductor.org/packages/release/bioc/vignettes/DESeq2/inst/doc/DESeq2.html#if-i-have-multiple-groups-should-i-run-all-together-or-split-into-pairs-of-groups

in this case I recommend comparing pairs of groups.

ADD REPLY • link 7.7 years ago Michael Love 43k

0

Entering edit mode

Thank you very much your quick response of my question.

We did filter out genes with summation of counts: rowSum=0. Among three treatments, we find two treatments have high variations in gene expression among their 3 replicates and one treatment has small variations among its 3 replicates.

We find that using local fitType, we obtained more genes with significant differential expression, 20 of which have PCR data validations. If we use parameter fitType, we obtained a few genes less. The PCR data matches to local fitType better.

More strikingly, when we run three treatments together versus split into pairs of groups in DESeq2. Together gave us much less genes, but pairs gave us more genes with significant differential expression. PCR data matches pairs results much better than together results.

Any comment will be highly appreciated.

ADD REPLY • link 7.7 years ago liangc.mu ▴ 10

0

Entering edit mode

I think our messages crossed. Yes I'd recommend paired comparisons in this case.

ADD REPLY • link 7.7 years ago Michael Love 43k