Hello everyone. I was wondering if there is a direct way of statistical testing PSI (Percent Spliced In) values without the need of external software. Let's say I have the following PSI values obtained by De Novo alternative events prediction. For instance, I would like to compare if PSI values of events 1 [1,] and 2 [2,] of the four samples (2 controls [,1][,2] and two perturbated samples [,3][,4]) are statistically different.
[,1] [,2] [,3] [,4]
[1,] 0.12658228 0.06343907 0.1038961 0.11666667
[2,] 0.87341772 0.93656093 0.8961039 0.88333333
[3,] 0.9117984 0.92075472 0.94163424 0.91400491
[4,] 0.0882016 0.07924528 0.05836576 0.08599509
[5,] 0.01875378 0.02342787 0.02713178 0.02777778
[6,] 0.98124622 0.97657213 0.97286822 0.97222222
R packages like DEXSeq allow for differential splicing testing, but I am currently limited to only using the information I have (PSI or counts for each event) and not able to use any third programs.
Is there a way in which PSI can be statistically tested directly? Which statistical test would you recommend for the case? Given the ammount of events I believe multiple testing could be the best choice. But I would probably only test for one or two events in each experiment, which makes me believe a two group testing could be more than enough.
You're missing all the depth and exon length information to replicate what one of those callers would need, so you are pretty much going to have to do a crude series of t-tests or a negative binomial.
That is very true, Jeremy, thank you for answering. In fact, I also have the event start and end coordinates and the read counts. And I had them already normalized to FPKMs.
So taking into account PSI to filter interesting events, technically... using their related FPKMs would suffice in a two group t-test?