Question

about paired-samples in edgeR

1

Entering edit mode

Bogdan ▴ 670

@bogdan-2367

Last seen 13 months ago

Palo Alto, CA, USA

Dear all,

would appreciate a suggestion please : considering a replicated experiment comparing Normal vs Tumor :

Normal1 Normal2 Tumor1 Tumor2

when setting up the list of samples for differential analysis in edgeR,

group <- factor(c("normal","normal","tumor", "tumor"))
y <- DGEList(counts=x,group=group)
y <- calcNormFactors(y)
y <- estimateCommonDisp(y)

how could I indicate that Normal1 is paired with Tumor1, and Normal2 is paired with Tumor2 ? thank you !

-- bogdan

edger • 3.7k views

ADD COMMENT • link 8.5 years ago Bogdan ▴ 670

score 2 · Answer 1 · 2016-05-24

2

Entering edit mode

b.nota ▴ 370

@bnota-7379

Last seen 4.2 years ago

Netherlands

You'll have to add a factor with subjects in the design.

group <- factor(c("normal","normal","tumor", "tumor"))

subject <- factor(1, 2, 1, 2)

design <- model.matrix(~0+group+subject)

etc.

ADD COMMENT • link 8.5 years ago b.nota ▴ 370

3

Entering edit mode

Note that it should be:

subject <- factor(c(1,2,1,2))

ADD REPLY • link 8.5 years ago Aaron Lun ★ 28k

2

Entering edit mode

Also, there is no advantage gained by adding "0+" to the model formula. The analysis would be equivalent and slightly simpler without it.

ADD REPLY • link 8.5 years ago Gordon Smyth 51k

score 0 · Answer 2 · 2016-05-24

0

Entering edit mode

Bogdan ▴ 670

@bogdan-2367

Last seen 13 months ago

Palo Alto, CA, USA

Dear all, thank you all for your comments : shall I just write then :

design <- model.matrix(~group+subject)

Also, please could you let me know if there are major advantages/disadvantages of limma versus edgeR for DE analysis ?

ADD COMMENT • link 8.5 years ago Bogdan ▴ 670

1

Entering edit mode

Both packages will do the job when it comes to detecting DE genes in a routine analysis. limma (+ voom) is faster and has access to more methodology (e.g., duplicateCorrelation) compared to edgeR, courtesy of lots of things being easier when you assume normality. However, voom does rely on the presence of a well-fitted mean-variance trend to estimate the precision weights. For some applications (though not usually DE analyses), the spread of abundances is too low to stably fit the trend, and in such cases edgeR will give better performance. edgeR also handles low counts better, which is worth considering if you want to focus on lowly-expressed genes or are dealing with very low-coverage data (e.g., single-cell stuff).