[Michael Love]

Fitting a GLM is an iterative process, and so you need to define a stopping criteria. We look for relative changes to the likelihood to define convergence. Sometimes the GLM doesn't converge for some rows when the counts are mostly zero and with very sparse model matrices (although in the recent releases, typically all rows converge). You can likely ignore this message, or remove these rows from the results if you like: res = res[mcols(dds)$fullBetaConv,]

What package version are you using: packageVersion("DESeq2")?

[lmolokin]

Gotcha. Session info below. Really appreciate your help, Michael!

Aleksey

R version 3.1.1 (2014-07-10)
Platform: x86_64-w64-mingw32/x64 (64-bit)

locale:
[1] LC_COLLATE=English_United States.1252  LC_CTYPE=English_United States.1252   
[3] LC_MONETARY=English_United States.1252 LC_NUMERIC=C                          
[5] LC_TIME=English_United States.1252    

attached base packages:
character(0)

other attached packages:
[1] DESeq2_1.6.1

loaded via a namespace (and not attached):
 [1] acepack_1.3-3.3           annotate_1.44.0           AnnotationDbi_1.28.1     
 [4] base_3.1.1                base64enc_0.1-2           BatchJobs_1.4            
 [7] BBmisc_1.7                Biobase_2.26.0            BiocGenerics_0.12.0      
[10] BiocInstaller_1.16.0      BiocParallel_1.0.0        brew_1.0-6               
[13] checkmate_1.5.0           cluster_1.15.2            codetools_0.2-8          
[16] colorspace_1.2-4          datasets_3.1.1            DBI_0.3.1                
[19] digest_0.6.4              edgeR_3.8.2               fail_1.2                 
[22] foreach_1.4.2             foreign_0.8-61            Formula_1.1-2            
[25] genefilter_1.48.1         geneplotter_1.44.0        GenomeInfoDb_1.2.2       
[28] GenomicRanges_1.18.1      ggplot2_1.0.0             graphics_3.1.1           
[31] grDevices_3.1.1           grid_3.1.1                gtable_0.1.2             
[34] Hmisc_3.14-5              IRanges_2.0.0             iterators_1.0.7          
[37] labeling_0.3              lattice_0.20-29           latticeExtra_0.6-26      
[40] limma_3.22.1              locfit_1.5-9.1            MASS_7.3-33              
[43] methods_3.1.1             munsell_0.4.2             nnet_7.3-8               
[46] parallel_3.1.1            plyr_1.8.1                proto_0.3-10             
[49] RColorBrewer_1.0-5        Rcpp_0.11.3               RcppArmadillo_0.4.450.1.0
[52] reshape2_1.4              rpart_4.1-8               RSQLite_1.0.0            
[55] S4Vectors_0.4.0           scales_0.2.4              sendmailR_1.2-1          
[58] splines_3.1.1             stats_3.1.1               stats4_3.1.1             
[61] stringr_0.6.2             survival_2.37-7           tools_3.1.1              
[64] utils_3.1.1               XML_3.98-1.1              xtable_1.7-4             
[67] XVector_0.6.0

[Michael Love]

Ok just curious. The convergence of these rows is not something to worry about. You can just ignore these rows or remove them using the above code.

[lmolokin]

Ah, now I see the problem. You can't use nbinomWaldTest(dd) in the last step. You need to use all the lines of code from the thread I pointed you to, which was

dds <- estimateSizeFactors(dds)
dds <- estimateDispersionsGeneEst(dds, modelMatrix=m)
dds <- estimateDispersionsFit(dds)
dds <- estimateDispersionsMAP(dds, modelMatrix=m)
dds <- nbinomLRT(dds, full=m, reduced=m[,-idx])

where idx is a numeric vector of columns to remove from the model matrix for the LRT. There is no support with nbinomWaldTest in the current version to supply model matrices, but I've promised to work on this in devel.

Hi Michael,
I wanted to clarify the purpose of the reduced model in my code. Since my goal was not to compare deferentially expressed genes between a full and a reduced model, I assume that the purpose of including it above was just a work-around (a place-holder) to allow the workflow to run with a user supplied model matrix?

So in other words, when I run:

c2 <- results(dds, [B]test="Wald"[/B], contrast=list("txc.time2"))

...the results I get are as if I had run a nbinomWaldTest?

Aleksey

[Michael Love]

Yes, for version 1.6, this call to results will give Wald test statistics and p-values. Suppling list or numeric arguments to 'contrast' will work, as will using 'name'.

[Michael Love]

A note: DESeq2 versions >= 1.7 allow user-supplied model matrices to top level functions DESeq() and estimateDispersions(), so the above extra lines of code are no longer necessary.