Question

DESeq2 Interspecies comparisons

0

Entering edit mode

adi.saxena • 0

@adisaxena-20520

Last seen 4.5 years ago

We have performed differential expression analyses using DESeq2 to directly compare the same tissue in two species that diverged from a common ancestor about 50 million years ago. We added a line of code that takes into account gene length differences between species. Since DESeq2 was not initially intended for such analyses, we wonder about some general considerations -

Is there an assumption in DESeq2 that most gene expression is not different in the two samples? Can this be overcome if the mean and variance of rlog or vst transformed gene counts are similar?
Since these gene counts are continuous values, how 'similar' is similar?

Thanks in advance!!

deseq2 • 1.6k views

ADD COMMENT • link updated 4.5 years ago by Michael Love 43k • written 4.5 years ago by adi.saxena • 0

score 3 · Accepted Answer · 2020-08-03

3

Entering edit mode

Michael Love 43k

@mikelove

Last seen 8 hours ago

United States

Regarding the size factor estimation, it's not that some percent of genes must be equal in expression, but more that the median of log ratios provides the center of the "small" changes. Can you provide an MA plot (of the MLE LFCs) which gives a sense of the distribution of LFCs?

The issue about size factor scaling is mostly that we don't know a priori what represents similar expression level, without any type of inserted controls.

ADD COMMENT • link 4.5 years ago Michael Love 43k

0

Entering edit mode

Thanks for that reply @Michael Love! Here is the MA plot for my interspecies comparisons. This plot is for all the log2 fold changes in my comparisons.

ADD REPLY • link 4.5 years ago adi.saxena • 0

0

Entering edit mode

Oh, can you change ylim to see the extent of the LFC better? Also use MLE=TRUE

ADD REPLY • link 4.5 years ago Michael Love 43k

0

Entering edit mode

Hi @Michael Love,

Thanks for the suggestion. I did-

dds <- DESeq(ddsHTSeq, betaPrior = TRUE)

res <- results(dds, addMLE = TRUE)

plotMA(res,ylim = c(-21,21),main = 'DESeq2', MLE = TRUE)*

This is the resulting MA plot with a wider LFC scale and MLE=True.

(For comparison here is the MLE=FALSE MA plot)

Thanks in advance!

ADD REPLY • link 4.5 years ago adi.saxena • 0

0

Entering edit mode

To me this looks like the median ratio is effective.

Note that betaPrior=TRUE is provided for backwards compatibility but is no longer recommended (since v1.16), e.g. see vignette code examples for recommended LFC shrinkage.

ADD REPLY • link 4.5 years ago Michael Love 43k

0

Entering edit mode

Thank you so much, Michael! I want to understand this a bit better and have few naive followups:

The median ratio you refer to is LFC (M) : Mean of Normalised Counts (A) (and not the median of LFC) ?

You think the median ratio in my interspecies analysis is effective for identifying differentially expressed genes because most genes cluster around the central M=0 line? What would MA plots look like if the median ratios were not effective?

Thanks in advance!

ADD REPLY • link 4.5 years ago adi.saxena • 0

0

Entering edit mode

The median ratio is the method used for normalization, consult the original DESeq publication (2010).

One case when the median ratio is clearly not appropriate is when there are different modes of the LFC distribution.

But so long as you understand that the median ratio method is correcting for the predominant mode of the LFC distribution, then it is effective for that purpose.

ADD REPLY • link 4.5 years ago Michael Love 43k