lots of "NA" values after preprocessing by "sesame" compared to "minfi/ENmix"

[yazdan59]

Hello,
I did a preprocessing step on my methylation data using two different methods (one using "sesame" and the other using "minfi" and "ENmix" packages) to compare the results.
However, in terms of final Beta values, I see a big difference in "NA" values. Could anyone explain why "sesame" has such many "NA" values and what is the best strategy to remove them after preprocessing (because if I remove their corresponding rows, it remains only 365,218 probes out of 866,553)?

Here are my used commands on the input data:

# Method 1) "minfi"/"ENmix" Packages
load("RGset_1752.RData"))
# Performing "Illumina+RCP" method
Mset_illumina <- preprocessIllumina(RGset)
betaIllumina_rcp <- rcp(Mset_illumina)
# Create the MethylSet
Mset <- preprocessRaw(RGset)
# Create the RatioSet
Rset <- ratioConvert(Mset)
# Retrieve the Beta-values
bVals= Rset@assays@data@listData[["Beta"]]
dim(bVals)
# [1] 716145   1752
sum(is.na(bVals))
# [1] 0
min(bVals)
# [1] 0
max(bVals)
# [1] 0.999969

# Method 2) "sesame" Packages
load("SigDF_1752.RData"))
sdfs_norm <- openSesame(sdfs, prep="QCDPB", func=NULL)
sdfs_norm_betas <- lapply(sdfs_norm, getBetas)
sdfs_norm_betas_matrix <- do.call(cbind, sdfs_norm_betas)
dim(sdfs_norm_betas_matrix)
# [1] 866553     1752
sum(is.na(sdfs_norm_betas_matrix))
# [1] 278742766
sum(is.na(sdfs_norm_betas_matrix))/length(sdfs_norm_betas_matrix)
# [1] 0.1836007
min(sdfs_norm_betas_matrix)
# [1] NA
max(sdfs_norm_betas_matrix)
# [1] NA

Thanks,

[zwdzwd]

My guess is that it is related to the low detection success of your data. Sesame has a more stringent detection masking. If you can run the risk of relaxing this threshold, you can tweak the pOOBAH p-value threshold or use the new ELBAR method for detection. Or you can turn off masking fully, by using the mask=F in openSesame.

[yazdan59]

Thanks for your suggestion.
I try two scenarios: 1) using the ELBAR method and 2) running the preprocess (QCDPB) with no mask option in openSesame. And I will let you know about the results.
Just for using the ELBAR method, where is the best place to put its step ("I") into the "QCDPB" code? What is your recommendation?

[zwdzwd]

Sure.
I should replace P
so QCDIB.
I experimented with different orders but didn't notice a big difference in practice.

[yazdan59]

Hello again,
As we discussed, I tried two scenarios: 1) using the ELBAR method ("QCDIB") and 2) running the "QCDPB" and then obtaining beta with no mask option. Here are the results and also the result of "minfi" package:

# "sesame" Packages (first approach)
sdfs_norm_v2 <- openSesame(sdfs, prep="QCDIB", func=NULL)
sdfs_norm_v2_betas <- lapply(sdfs_norm_v2, getBetas)
sdfs_norm_v2_betas_matrix <- do.call(cbind, sdfs_norm_v2_betas)
dim(sdfs_norm_v2_betas_matrix)
# [1] 866553     1752
sum(is.na(sdfs_norm_v2_betas_matrix))
# [1] 185072778
min(sdfs_norm_v2_betas_matrix)
# [1] NA
max(sdfs_norm_v2_betas_matrix)
# [1] NA

# "sesame" Packages (second approach)
sdfs_norm <- openSesame(sdfs, prep="QCDPB", func=NULL)
sdfs_norm_v3_betas <- lapply(sdfs_norm, function(x) getBetas(x, mask=FALSE))
sdfs_norm_v3_betas_matrix <- do.call(cbind, sdfs_norm_v3_betas)
dim(sdfs_norm_v3_betas_matrix)
# [1] 866553     1752
sum(is.na(sdfs_norm_v3_betas_matrix))
# [1] 0
min(sdfs_norm_v3_betas_matrix)
# [1] 0.00370823
max(sdfs_norm_v3_betas_matrix)
# [1] 0.9977153

# "minfi"/"ENmix" Result
Mset_illumina <- preprocessIllumina(RGset)
betaIllumina_rcp <- rcp(Mset_illumina)
Mset <- preprocessRaw(RGset)
Rset <- ratioConvert(Mset)
bVals= Rset@assays@data@listData[["Beta"]]
dim(bVals)
# [1] 866553     1752
sum(is.na(bVals))
# [1] 0
min(bVals)
# [1] 0
max(bVals)
# [1] 0.999969

So, as you can see, running the ELBAR method caused less "NA" values (185,072,778) compared to "QCDPB" strategy (278,742,766). Also, using "no mask" option removed all "NA" values. So, it seems I should use "no mask" option to go to the next level analyses (calculating M-values, DMR, etc.). However, I do not know if in this situation, still we have used the strength of the QCDPB sesame preprocessing as mentioned in the papers which recommended its usage compared to other preprocessed methods (such as Illumina, ENmix, noob, rcp, etc.).

[zwdzwd]

Hi @yazdan59, Great. The background subtraction and dye bias correction should've been completed even though the problematic probes remain not masked. Just take caution interpreting the methylation signals from ones that are supposed to be masked.

[yazdan59]

Thanks for all your suggestions and consideration.