Cuffdiff results for genes with low count for one treatment

[pinin4fjords]

Hi all,

I've a question related to the situation where one group of samples for Cuffdiff has near-zero counts, and another has very high counts.

I have a rat RNA-seq dataset, Illumina 78 bp paired-end, 12 samples over two groups (7/5), approx. 14M reads per sample. I've already done count-based differential expression analysis using (htseq-count -> DESEeq), and am running Cufflinks for comparison, and for the additional insights it can provide wrt differential splicing etc.

I have a problem with the gene-based results (in gene_exp.diff) in that I'm seeing some NOTEST statuses with associated values of 1 for 'q_value', in genes that I know are highly differentially expressed. I think this is due to low counts in one of my groups leading to this gene being ignored due to Cufflinks' '--min-alignment-count' setting, despite expression being very high in the other group. Assuming this is the case I believe I could partly correct for this by using the '--min-alignment-count' parameter to with a very low value.

So, my question is: what about cases where the alignment count is zero (or close to it) in one of the groups, but high in the other? That usually represents significant differential expression, but it doesn't seem like cufflinks would report that because it can't see sufficient counts for one of the groups.

I suspect I'm missing something, so I'd appreciate any pointers.

Many thanks,

Jon

[torrey]

I've used the --min-alignment-count flag to get around this in the past. Although, as it states in the manual you increase the test burden during multiple hypothesis correction by allowing more loci to be tested.

It would be nice if this flag was the sum of the alignment counts from both conditions. As it currently appears to be the alignment count of the least expressed sample.

[pinin4fjords]

Good to know- thanks, and confirms what I was thinking.