# NaN values after ComBat analysis on TCGA COAD RNAseq

I have FPKM-UQ data from COAD-TCGA.

I generated an expression set of this data using:

> edata = log2(data + 1)
> edata[1:2,1:2]
X01240896.3f3f.4bf9.9799.55c87bfacf36
1                                  8.540967
10                                13.528968
100                               16.296422
1000                              13.658199
10000                             15.143788
1                                  9.886537
10                                16.719682
100                               17.212312
1000                              10.317842
10000                             13.166767


Row name is entrezid, and column names are case_ids.

The batch data provided by TCGA which indicates a total batch num = 42

> omics[1:5,]
snf    boo order batch
X01240896.3f3f.4bf9.9799.55c87bfacf36   3 192807     1    22
X01f493d4.229d.47a6.baa8.32a342c65d01   1 A08907     3    29
X022f39e9.57ee.4b2b.8b3a.8929e3d69a37   2 A16S13     4    32
X02f9668c.71e6.485f.88b1.b37dc8bdd2ab   3 102113     5     7
> batch = omics\$batch
> batch
[1] 22  2 29 32  7  1  8  8  7 41 26 14 26 26  1  5 22  5 15 18 28 12 12  8 19
[26] 13 38  5 27  7 41  1 22 26 28 22 33 14 13  8 10 13 11 17  5 38  5  7  7 31
[51] 22 42 11  8  7 17 15 17 12 17  2 19  2 12 18 14 22 11 21 15 17 17 33 15 19
[76] 15  6  1 10 11  2 28 36 17  9 15 12 17 15 10  6  8 13 25 11 15  1 13 24  8
[101] 13 15 17  8  2 22 17 11 13 37 19 38 13 19  5 16  5  5 22 35 35 19 13 17 22
[126]  7  8  8 41 39 14  6 19 18 38 22 14 30 19 24 11 14 13 16  7 13  8 22 12  8
[151] 22 17  2 40 13 35 15  5 24 19 41 22 22 19 17  8 13 33 10 17 29 12  9 11 24
[176] 27  5 14  7 25 13 35  9 18 16 13 31 25 18 10 35 18 14 33  5 19 17  8 11 11
[201] 42  7 10  8 39 19  8  7 11 19 31 35 28 36 13 37 10 38  3 18 13 34 39 13  3
[226] 40 31 17 19 26  7 19 12 22 19 17 12 17 13 12 28 33 37 17 35 18 40 13 20 35
[251] 17  6 15 14 12 27  5 41 31 41 18 21 19 17 40  8 41  2  5 15  5 15 15  5 39
[276]  1 22 33 13 17  7 11 15  3  7 38 15  7 14 14 22 31 14 18  3 19  7  9  6  5
[301] 22  8 11  1 13 10 19  8 14 15  6 32  1  2 19  4  5 13 18 42 11 17  8 36 19
[326] 15 22 28 15  8 11  8 15 38 19 39  7 11 42 23  5  1 22 17 35 13 40 25  7 41
[351] 38  5  8 19 31  6 38  8 36  4 15 15  7 12 22 38 29 35 15  2 10  2 19 15 25
[376]  5 39 30 19 18 15 10 10 33  1 19 32 19 19 15 20 15  7 22 35 22 30 14 12 16
[401] 21 13 15 18  7  5  6 31 12  7 15 32 33 18 14 15 15 26 31  1 14 19 17 18  1
[426]  5 10 32


Unfortunately, when I run the ComBat analysis it generates all NANs, and no values:

> modcombat = model.matrix(~1, data=omics)
> combat_edata = ComBat(dat=edata, batch=batch, mod=modcombat, par.prior=TRUE, prior.plots=FALSE)
> combat_edata[1:5,1:2]
X01240896.3f3f.4bf9.9799.55c87bfacf36
1                                       NaN
10                                      NaN
100                                     NaN
1000                                    NaN
10000                                   NaN
1                                       NaN
10                                      NaN
100                                     NaN
1000                                    NaN
10000                                   NaN

• What happens if you make batch a factor? Jun 23 '17 at 21:44
• So when I normalize the data first by mean, then perform the log2 conversion, the batch correction works. Jun 23 '17 at 21:44
• Stipulating batch as factor generates the same NAN result. I'm thinking that maybe it wasn't close enough to a normal distribution without the mean normalization, so it wasn't able to calculate the correction? Jun 23 '17 at 21:47
• Hi Martin, thanks for your question and welcome to Bioinformatics Stack Exchange. I have changed your title to be more specific and match the question that you have asked. Having specific questions makes it more likely that people will be able to answer your problem (or be interested in answering it), and some people won't click through if they see a question that has a vague title.
– gringer
Jun 23 '17 at 21:51
• It's just interesting because the data is supposedly already normalized between samples using FPKM-UQ. So I'm not clear why the additional mean normalization is necessary. In particular, I need to do something like: data <- standardNormalization(data) to scale the mean at 0 w/ stdev @ 1, then log2(data + 1). Jun 23 '17 at 21:59

I think I might know what the problem is.

I plotted the log2 corrected values, and the distribution may not be normal because of the 0 values present in RNASeq data. If I understand correctly, this data is actually missing, whereas with my current analysis it is viewed by R as an actual value of 0. So either the missing data needs to be imputed, or the data needs to be scaled and normalized to mean of 0, then subsequently log2 transformed.

> data <- standardNormalization(data)
> data <- log2(data + 1)
> combat_data = ComBat(dat=data, batch=batch, mod=modcombat, par.prior=TRUE, prior.plots=FALSE)
> combat_data[1:5,1:2]
X01240896.3f3f.4bf9.9799.55c87bfacf36
100507661                           -0.13857428
57103                               -0.06700506
22838                                0.16295096
55567                               -0.13579991
6147                                 1.64990772