[BioC] strange log2FoldChange estimates

[Shi, Tao]

hi Michael,

I have a simple RNAseq experiment: a cell line treated with a compound at various doses and time.  Here are the htseq counts for one of gene below along with the treatment conditions and sizeFactor estimated by DESeq2.

> tmp
  count sizeFactor time dose 
1  11396  1.0069598    8    0 
2  12306  1.0913848    8    0 
3  11761  1.0611031    8  0.1 
4  12177  1.0764237    8  0.1 
5  12017  1.1675793    8    1 
6  10667  1.0375095    8    1 
7  11423  1.0693406    8   10 
8  11529  1.1097800    8   10 
9   7845  0.9187615   24    0 
10  8600  0.9620073   24    0 
11  3200  0.9502428   24  0.1 
12  2964  0.8491865   24  0.1 
13  1079  0.8893242   24    1 
14  1391  1.0884646   24    1 
15  1261  0.9130893   24   10 
16  1451  1.0182247   24   10 

The design matrix I used is : ~time + dose + time:dose

Here are the coefficient estimates for this particular gene:

baseMean                       7.266687e+03 
baseVar                        1.854839e+07 
allZero                        0.000000e+00 
dispGeneEst                    1.000000e-08 
dispFit                        6.694323e-04 
dispersion                     4.994657e-04 
dispIter                       1.000000e+01 
dispOutlier                    0.000000e+00 
dispMAP                        4.994657e-04 
Intercept                      1.238851e+01 
time8                          1.007756e+00 
time24                        -9.953678e-01 
dose0                          8.800724e-01 
dose0.1                        2.014315e-01 
dose1                         -5.854584e-01 
dose10                        -4.836570e-01 
time8.dose0                   -8.127168e-01 
time24.dose0                   8.201522e-01 
time8.dose0.1                 -1.467029e-01 
time24.dose0.1                 1.484047e-01 
time8.dose1                    5.176693e-01 
time24.dose1                  -5.226157e-01 
time8.dose10                   4.502645e-01 
time24.dose10                 -4.543508e-01 

And some of the contrasts I tested:

for time24.dose10 vs time24.dose0
baseMean log2FoldChange      lfcSE      stat    pvalue      padj 
7266.687      -1.274503 0.02810044 -45.35527         0         0 

for time8.dose10 vs time8.dose0 
baseMean log2FoldChange      lfcSE      stat    pvalue      padj 
7266.687       1.262981 0.02804676  45.03128         0         0 


The thing that puzzles me a lot is that, for the log2FC of the 2nd contrast, I can't see how it can be ~1.26 given the sizefactor corrected mean counts for the gene for these two conditions are 11296.41 and 10535.41, respectively. To me, the log2FC should be negative or close to 0.

Attach is the tmp matrix you saw above.  I know you can't replicate all the estimates using the tmp matrix alone without other genes in the dataset, but that's all I can provide for confidentiality.  Hopefully, you can figure out where the problem is.

I'm using DESeq2 v1.4.5, R3.1.0, under Win 7.

Thank you very much!

Tao