3.7 years ago by

United States

I realize my comment was confusing and revised it as you answered -- all I meant is that the QL estimates inform which logFC are significant.

( All I mean with the log was that for NB counts c, var(c) ~ mu + d*mu^2, where mu = average count over biological replicates, d dispersion. Then we can write var(log(cpm)) as approximatelyT/mu + B, where T is technical and B is biological variance. So, yes, the log overdamps the cpm to B for large mu. I think we're talking past each other. )

Anyway, I'm just stuck interpreting the plotBCV and plotQLDisp and the huge common and tagwise trends they display....i.e. my plotQLDisp with robust = T is a linear function ranging from 0.6 to 1.0 in deviance over 0 to 10 average log2cpm. So just to be concrete, here's what I'm doing:

d_QL <- estimateDisp(dge, design)

#robust = T

fit_QL <- glmQLFit(d_QL, design, robust = T)

#replace likelihood tests with quasy likelihood F-tests for coefficients in the linear model (so just like glmLRT)

res_QL <- glmQLFTest(fit_QL)

plotBCV(d_QL)

plotQLDisp(fit_QL)

•

link
modified 3.7 years ago
•
written
3.7 years ago by
osa.zuta • **40**
Have any of you looked at the prior weights when limma is applied to single cell data? In other words, how different are the results if you just run a Lognormal gene-specific model w/o shrinkage? Single cell has a larger sample size than bulk RNA, so I wonder if shrinkage becomes less necessary or maybe even redundant for single cell.

70I suggest you post a new question.

23k