DEXSeq questions and errors

[Udmurtia]

Hi,

I've been trying to use DEXseq for my project. (As background, I'm attempting to do pairwise comparison of wild-type versus samples where I've perturbed a gene that regulates alternative splicing. My samples are in triplicate.) I have three questions, and two errors that I've encountered. I've worked through the vignette with the pasilla data with no problems.

1. Do you have any recommendations for programs to create the input gtf file (to be flattened by dexseq_prepare_annotation.py)? I've tried cuffmerge, but it seems to be too liberal in its definition of gene models, joining adjacent genes. This seems to result in DEXseq missing differentially expressed exons that are "clearly" significant by eye from the mapped reads.
   
2. Could you please explain the meaning of the pasillaExons@dispFitCoefs statistic? Also, what are we looking for here, and what is an acceptable value?
   
3. All of my data are paired-end reads. Does that mean that the section of the vignette talking about giving this extra information to DEXSeq is irrelevant for me? i.e. is this method only for mixed data? (I'm referring to section 5: Additional technical or experimental variables.)

Following are warnings/errors that I've encountered.

1. When estimating log2 fold changes, I got 26 identical warnings. I did not get this error on the pasilla data subset. Are these important?
   

   > ecs <- estimatelog2FoldChanges(ecs)
    There were 26 warnings (use warnings() to see them)
   > warnings()
   Warning messages:
   1: In chol.default(XVX + lambda * I, pivot = TRUE) : matrix not positive definite
   2: In chol.default(XVX + lambda * I, pivot = TRUE) : matrix not positive definite
   ...etc.

2. The answer to this may depend on the response to (3) above. When attempting to take into account the paired-end nature of my data, I got the following error. I did not get this error on the pasiilla data subset. Can I fix this?
   
   

   formuladispersion <- count ~ sample + ( condition + type ) * exon
   ecs <- estimateDispersions( ecs, formula = formuladispersion , nCores=8 )
   
   Estimating Cox-Reid exon dispersion estimates using 8 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 ‘"character"’
   In addition: Warning message:
   In parallel::mclapply(allecs, FUN = funtoapply, mc.cores = mc.cores) :
     all scheduled cores encountered errors in user code

Thanks in advance.


<EDIT>========================================<EDIT>

1. Do you have any recommendations for programs to create the input gtf file (to be flattened by dexseq_prepare_annotation.py)? I've tried cuffmerge, but it seems to be too liberal in its definition of gene models, joining adjacent genes. This seems to result in DEXseq missing differentially expressed exons that are "clearly" significant by eye from the mapped reads.

I've done a bit more troubleshooting regarding this issue. If I manually split off the gene of interest, creating the "correct" gene model, DEXSeq does find significance, as expected. It seems that I really do need good gene models to input into DEXSeq. I've made a screenshot of the relevent gene, showing coverage of my 12 samples. I've marked with (1) where the alternative splicing is not detected using cuffmerge's gtf file. The upstream and downstream genes are connected to this gene according to cuffmerge. One example of this is marked by (2), where the low-level reads from adjacent genes overlap with each other. At (3), it's not so much a low level of reads, but longer mRNAs in some samples that overlap with other samples. Looking by eye, it seems obvious (to me) that these three genes are separate. Is there a way to make cuffmerge agree, and/or are there other tools that would do better?