Large discrepancy between estimated and detected DTU

[John Smout]

Hi all,

I'm working on a small dataset (8 samples, 3 treatments) and trying to look at differential transcript usage/isoform switching alongside some other analyses. My data is from ONT sequencing, quantified using salmon.

When I run <importRdata()> I get "GUESTIMATED" gened with dtu like this:

            comparison estimated_genes_with_dtu
1  post_rep vs pre_rep                261 - 436
2 post_rep vs pregnant                357 - 595

3  pre_rep vs pregnant                191 - 319

However, when I go ahead and run the analysis with <isoformSwitchAnalysisPart1()>, I get way fewer genes with switches than estimated:

            Comparison nrIsoforms nrSwitches nrGenes
1  post_rep vs pre_rep          0          0       0
2 post_rep vs pregnant          6          5       4
3  pre_rep vs pregnant          0          0       0
4             Combined          6          5       4

I recognise that of course the results of the analysis will differ from the estimates ... but this is a pretty big difference, and does not seem that credible to me in biological terms. I would expect at least the first two comparisons to show at least some significant dtu. Is it normal for the final test results to be this different from the estimates? Is it possible that this is caused by unwanted effects (sva failed to run when importing my data) or by small sample size (smallest is n = 2)?

I previously analysed the same dataset using a different approach, also based on DEXSeq, detailed here: https://www.bioconductor.org/packages/release/workflows/vignettes/rnaseqDTU/inst/doc/rnaseqDTU.html This gave me dramatically different results, with over 100 genes showing significant DTU in comparison 2, and dozens for the other comparisons, including biologically relevant genes. However, that pipeline doesn't include the nice visualisations and integration of other tools that IsoformSwitchAnalyseR has.

I feel like something is amiss here, and I'm wondering if anyone can advise - is it worth continuing with using isoformSwitchAnalyseR, trying to fix the batch effects, adjust the alpha and so on, to try and see if I can capture more genes? And alternatively, is it possible to import results from the DEXSeq/stageR analysis directly into a switchAnalyzeRlist object to then use the downstream visualisation tools and plots?

Many thanks!

J.