PCA on FPKM data

Question

[ipalmisa]

Hi I am running a PCA analysis using prcomp function on my FPKM data from Rnaseq analysis. I am scaling the FPKM data by using zFPKM from here (https://www.bioconductor.org/packages/release/bioc/html/zFPKM.html) The zero values in FPKM data are converted in -Infinite, so when I run prcomp funtion, I get Error in svd(x, nu = 0, nv = k) : infinite or missing values in 'x' Can you help me, please? Should I just remove the -infinite?

Answer 1

[piyushjo]

why do you want to scale? are you following some tutorial where they scale the FPKM values?

[ipalmisa]

I read here Question about PCA plot using RPKM/FPKM. the suggestion to scale

[Kevin Blighe]

prcomp() will, by default, only center your data. You can set it to additionally scale by using:

prcomp(x, center = TRUE, scale = TRUE)

If you are having trouble with zFPKM, then just try the above.

zFPKM converts 0 values to -Inf (negative infinities):

df <- data.frame(a=c(1,4,5,6,7,4,3,0),b=c(6,5,4,23,5,6,7,78))
 df
  a  b
1 1  6
2 4  5
3 5  4
4 6 23
5 7  5
6 4  6
7 3  7
8 0 78

zFPKM::zFPKM(df)
           a           b
1 -4.0512861  0.08036657
2 -0.2796893 -0.08062816
3  0.3274022 -0.27766979
4  0.8234321  1.26691969
5  1.2428194 -0.08062816
6 -0.2796893  0.08036657
7 -1.0623663  0.21648567
8       -Inf  2.34528431

[ipalmisa]

Yes, i have noticed that the zeros are scaled to minus infinite But when I try to use the prcomp on the FPKM data using pca=prcomp(my_data, center=TRUE, scale = TRUE) I get Error in prcomp.default(my_data, center = TRUE, scale = TRUE) : cannot rescale a constant/zero column to unit variance It works only with scale=FALSE I think here the zeros are the issue...

[Kevin Blighe]

I see. You have variables that are just all 0. You need to remove these like this:

df
  a  b
1 1  6
2 4  5
3 5  4
4 0  0
5 7  5
6 4  6
7 3  7
8 0 78

keep <- !apply(df, 1, function(x) all(x == 0))

df[keep,]
  a  b
1 1  6
2 4  5
3 5  4
5 7  5
6 4  6
7 3  7
8 0 78

[ipalmisa]

Thank you! It worked!

[ipalmisa]

ah ok! I will remove the ones that are Zero in ALL my samples, and keep the ones that are Zero only in some samples, right? Have I understood correctly? Thanks

[Kevin Blighe]

Yes, that is correct. I cannot see your data but you may also have other non-zero variables that have constant variance, but, first, try to remove just those variables that are all zero.

I also updated my original answer: A: Question about PCA plot using RPKM/FPKM.

Sorry to piyushjo for overtaking the thread.

[piyushjo]

It's alright. You are the pro! I was just trying to get more information but you had better answer.

[Kevin Blighe]

You are pro too. Everybody has much to offer in different areas.