Hello,

I used edgeR.

After calcNormFactors, I got the norm.factors, then I tried to use the counts number and norm.factors to calculate the logFC.

1. (478/1.0296636+619/1.0372521+628/1.0362662+744/1.0378383)/4=596

(483/0.9537095+716/0.9525624+240/0.9583181)/3=503

log2(503/596)=-0.244754

2. (478*1.0296636+619*1.0372521+628*1.0362662+744*1.0378383)/4=639

(483*0.9537095+716*0.9525624+240*0.9583181)/3=458

log2(458/639)=-0.480468

None of them seems to =-0.4315571.

My question is maybe calculate the logFC is more complicate?

Is that possible I can calculate it by myself?

Thanks in advance!

> y
An object of class "DGEList"
$counts
                Con1 Con2 Con3 Con4 DHT1 DHT2 DHT3
ENSG00000124208  478  619  628  744  483  716  240
ENSG00000182463   27   20   27   26   48   55   24
ENSG00000124201  180  218  293  275  373  301   88
ENSG00000124207   76   80   85   97   80   81   37
ENSG00000125835  132  200  200  228  280  204   52
16489 more rows ...

$samples
       group lib.size norm.factors
Con1 Control   976847    1.0296636
Con2 Control  1154746    1.0372521
Con3 Control  1439393    1.0362662
Con4 Control  1482652    1.0378383
DHT1     DHT  1820628    0.9537095
DHT2     DHT  1831553    0.9525624
DHT3     DHT   680798    0.9583181

> y <- estimateCommonDisp(y, verbose=TRUE)
Disp = 0.02002 , BCV = 0.1415
> y <- estimateTagwiseDisp(y);
> plotBCV(y);
> et <- exactTest(y);

> a<-et$table
> head(a)
                     logFC   logCPM      PValue
ENSG00000124208 -0.4315571 8.725537 0.002741666
ENSG00000182463  0.6909960 4.691815 0.007143860
ENSG00000124201 -0.0540919 7.511315 0.667158340
ENSG00000124207 -0.4226992 5.900093 0.026677865
ENSG00000125835 -0.2470948 7.098438 0.176266270
ENSG00000125834  0.3098867 5.803637 0.152476022

You should be using the quasi-likelihood framework (estimateDisp, glmQLFit and glmQLFTest), which offers a number of advantages over the classic and LRT methods. But long story short, yes, the calculation of the log-fold change is more complicated than taking group-wise averages and comparing them.

Check out ?predFC for a brief summary and the associated reference for precise mathematical details. The calculation in exactTest is slightly different due to the differences between the classic and GLM-based methods, but then again, you should be using the GLM-based methods anyway.

edgeR computes logFC values using negative binomial generalized linear models. There are some differences between the classic (exactTest) and GLM (glmFit) pipelines, but they all use generalized linear models.

You could theoreticaly reproduce edgeR's calculation if you are familiar with generalized linear models, but the computation is indeed much more complicated (and much better) than what you have done so far.