DEXSeq estimateDispersions error

[dadada4ever]

Hi everyone,
I run into a problem when using DEXSeq estimateDispersions command.
I know the error is because I have only 1 Tumor & Normal RNA-Seq data, which means there's no replicate, but if I still want to deal with the data with DEXSeq how can I solve the problem?

Samples:

                      condition replicate type
T_DEXSeq.txt     Tumor         1     paired-end
N_DEXSeq.txt     Normal        1     paired-end

When I set the nCores=2:

Estimating Cox-Reid exon dispersion estimates using 2 cores. (Progress report: one dot per 100 genes)
Error in function (classes, fdef, mtable)  :
   unable to find an inherited method for function "fData", for signature "try-error"
In addition: Warning message:
In mcLapply(allecs, FUN = funtoapply, mc.cores = mc.cores) :
   all scheduled cores encountered errors in user code

There's a similar question in http://permalink.gmane.org/gmane.sci...onductor/40321 which occured the same error in another procedure. And Simon said this rarely happened due to high dispersions.

When I only use one core the error said:

Dispersion estimation. (Progress report: one dot per 100 genes)
Error in FUN(c("ENSG00000000003", "ENSG00000000419", "ENSG00000000457",  : 
  Underdetermined model; cannot estimate dispersions. Maybe replicates have not been properly specified.
In addition: Warning messages:
1: In .local(object, ...) :
  Exons with less than 11 counts will be discarded. For more details read the documentation, parameter minCount
2: In .local(object, ...) :
  Genes with more than 70 testable exons will be kicked out of the analysis. For more details read the documentation, parameter maxExon

I'm wondering if I could do some manipulation to avoid the error? Such as change minCount or artificially remove the genes the error mentioned? Thanks!

[Simon Anders]

Seems that in the multicore setting, the correct error message gets overwritten by something less helpful. We need to check this.

In the single-core call, DEXSeq reports correctly reports: "Underdetermined model; cannot estimate dispersions. Maybe replicates have not been properly specified."

I guess I don't need to tell you that you cannot extract any biological results from comparing a single normal with a single tumour sample, and that your analysis is hence pointless, anyway.

[dadada4ever]

Seems that in the multicore setting, the correct error message gets overwritten by something less helpful. We need to check this.

In the single-core call, DEXSeq reports correctly reports: "Underdetermined model; cannot estimate dispersions. Maybe replicates have not been properly specified."

I guess I don't need to tell you that you cannot extract any biological results from comparing a single normal with a single tumour sample, and that your analysis is hence pointless, anyway.

Thank you for replying Simon.
I know and you have mentioned the meaningless of RNA-Seq data with no replicates for many times. To my standpoint, I knew DESeq has a section of "Working partially without replicates", that's the reason why I guess if there's some way to avoid this error in DEXSeq (maybe that's not very logical..), so your answer is NO, am I right?
Thank you.

[Simon Anders]

Exactly. In principle you can also use DEXSeq in the "blind" mode that we offered for DESeq, but I cannot see any circumstances where this would make sense.

[dadada4ever]

Exactly. In principle you can also use DEXSeq in the "blind" mode that we offered for DESeq, but I cannot see any circumstances where this would make sense.

I see. Thank you Simon.

[EGrassi]

Exactly. In principle you can also use DEXSeq in the "blind" mode that we offered for DESeq, but I cannot see any circumstances where this would make sense.

In principle but non in practice right now, am I wrong?
They are asking me to do so, I suppose that I need to compare DESeq and DEXSeq codes and try to add the blind support there.

ps.

> counts <- estimateDispersions(counts)
Dispersion estimation. (Progress report: one dot per 100 genes)
Error in FUN(c("ENSG00000000003", "ENSG00000000005", "ENSG00000000419",  :
  Underdetermined model; cannot estimate dispersions. Maybe replicates have not been properly specified.
In addition: Warning messages:
1: In .local(object, ...) :
  Exons with less than 11 counts will be discarded. For more details read the documentation, parameter minCount
2: In .local(object, ...) :
  Genes with more than 70 testable exons will be kicked out of the analysis. For more details read the documentation, parameter maxExon
> counts <- estimateDispersions(counts, nCores=nCores, method="blind")
Error in .local(object, ...) : unused argument(s) (method = "blind")
> counts <- estimateDispersions(counts, method="blind")
Error in .local(object, ...) : unused argument(s) (method = "blind")

[areyes]

In principle but non in practice right now, am I wrong? They are asking me to do so, I suppose that I need to compare DESeq and DEXSeq codes and try to add the blind support there.

Not necessarily, just remove the lines

if( nrow(mm) <= ncol(mm) )
   stop( "Underdetermined model; cannot estimate dispersions. Maybe replicates have not been properly specified." )

But of course, those lines have a VERY GOOD reason to be there

Alejandro

[EGrassi]

I see. Right now I have data without replicates and they are asking me to do a preliminary differential isoforms usage test, so I guess that I have to find a way Maybe I can try with one of the other ton of packages (MISO, etc), which is always an entertaining job...

I commented out that lines and recompiled DEXSeq but now the next step is having troubles, I guess that reading all the code is the right option now, which is more interesting than trying other packages randomly

counts <- estimateDispersions(counts, nCores=nCores)
Estimating Cox-Reid exon dispersion estimates using 3 cores. (Progress report: one dot per 100 genes)
.........................................................................................................................................................> counts <- fitDispersionFunction(counts)
Error in glmgam.fit(mm, disps[good], start = coefs) :
More columns than rows in X
In addition: Warning message:
In is.na(rows) : is.na() applied to non-(list or vector) of type 'NULL'
Error in fitDispersionFunction(counts) :
Failed to fit the dispersion function
>

[Simon Anders]

Of course, you cannot estimate dispersion without replicates. So you have to come up with some fantasy value and plug it in, with "fData(ecs)$dispersion <- 0.1".

The "0.1" here is just a place holder. I will not venture a guess what an appropriate value might be.

[EGrassi]

Of course, you cannot estimate dispersion without replicates. So you have to come up with some fantasy value and plug it in, with "fData(ecs)$dispersion <- 0.1".

The "0.1" here is just a place holder. I will not venture a guess what an appropriate value might be.

Yup, thanks. My idea was to check which value is used with the blind method in DESeq (which gave results superimposable with other tools) and give it a try. I'm well aware of your reasonable need for biological replicates, but I'm not the biggest statistical head here and when they ask me to do something which seems reasonable to them I have at least to try

[camelbbs]

Not necessarily, just remove the lines

if( nrow(mm) <= ncol(mm) )
   stop( "Underdetermined model; cannot estimate dispersions. Maybe replicates have not been properly specified." )

But of course, those lines have a VERY GOOD reason to be there

Alejandro

HI,

I want to ask how to remove this line. Since where I can find the raw scripts in R installation folders.

Thanks..

[Wolfgang Huber]

HI,

I want to ask how to remove this line. Since where I can find the raw scripts in R installation folders.

Thanks..

See http://www.bioconductor.org/developers
"Source code is stored in svn (user: readonly, pass: readonly)."
Go to the "R" directory of the DESeq package, and grep for the keyword you are looking for.

Best wishes
Wolfgang

[camelbbs]

See http://www.bioconductor.org/developers
"Source code is stored in svn (user: readonly, pass: readonly)."
Go to the "R" directory of the DESeq package, and grep for the keyword you are looking for.

Best wishes
Wolfgang

Hi Wolfgang,
Thanks very much. But while I downloaded the core_function.R and modified it, I get an error to load it:

> source("C:\\Users\\cjhe\\Desktop\\dexseq\\core_functions.R")
Error in setMethod("estimateSizeFactors", signature(object = "ExonCountSet"),  : 
  no existing definition for function ‘estimateSizeFactors’

[camelbbs]

See http://www.bioconductor.org/developers
"Source code is stored in svn (user: readonly, pass: readonly)."
Go to the "R" directory of the DESeq package, and grep for the keyword you are looking for.

Best wishes
Wolfgang

Hi Wolfgang,

While I comment the line:

if( nrow(mm) <= ncol(mm) )
   stop( "Underdetermined model; cannot estimate dispersions. Maybe replicates have not been properly specified." )

and run estimateDispersions, i got error:

Error in match.arg(start.method, c("log(y)", "mean")) : 
  'arg' must be NULL or a character vector

can you help for that? Thanks a lot!

[areyes]

Dear camelbbs,

have a look on this post:

http://seqanswers.com/forums/showthread.php?t=23589

bests,
Alejandro