Hello! I've recently ran the rnasplice pipeline and am now wondering what is the best way to analyze all the different outputs? My goal is to evaluate the accuracy of each alternative splicing analysis tools' outputs against long-read RNA-seq data and to evaluate the differences in splicing predictions between each tool. But in general, I am just curious as to how to analyze all of these outputs since certain tools like SUPPA do not have documentation on how to analyze the outputs (at least to my knowledge). Any advice or help with this would be much appreciated, thank you!

Hi Mia! I am facing the same. It would be very helpful if anyone from the community is willing to share their experience with the downstream analysis, especially with SUPPA.

Hi, just bringing back this post. If anybody has any github repositories with their analyses code for the outputs of any of the tools used in this pipeline, that would be great as a reference. I can't seem to find any on my own, so any help is greatly appreciated!

The SUPPA GitHub README explains the output file formats. You'll get e.g. PSI values for each event/sample, as well as p-values for the contrasts you set up. Conceptually this should similiar to what you get from other tools, e.g. rMATS. What you do with it really depends on your questions. You could do summary statistics by counting events in comparisons that meet certain criteria (PSI and/or p-value cutoffs), or you can see if your genes of interest have any AS events. You might do the former if you are interested if/how a certain treatment affects splicing generally, and the latter if you want to explain a certain phenotype you see, which might be caused by one or a few specific AS events. If you see any events of interest, one of the first steps would be to see what it does it the protein sequence (nothing, truncation, frame shift, in-frame insertion or deletion, etc.).

Hi @Thomas Danhorn and all! Following up on your discussion, I wonder how to proceed with the PSI values cutoffs. While for comparisons (dPSI) we have p-values, for PSI values, even having most of the PSI per isoform/local event consistent among bioreps, some might vary quite a lot. How do you handle it? Do you have any experience? Here is an example:

Id    R1    R2    R3
AT1G01020;A5:Chr1:8464-8571:8464-8594:- 0.739432749728953  0.729771473690935  0.77637299452515
Id    R1    R2    R3
AT1G01140;A5:Chr1:66897-67017:66897-67333:-   1     0.583021776494144  0.223310239122538

In a comparison, the absolute PSI value don't have a lot of meaning, it is the change from one isoform to another that is interesting, i.e. the difference between two samples or preferably sample groups (e.g. genotypes, cell types, or pre-/post-treatment, etc.), so you have to look at the difference of two PSI values (averages when comparing groups), often called ∆PSI (in rMATS the column name is "IncLevelDifference"). I should mention that I have no experience with SUPPA (yet), but it seems its concepts are similar to e.g. rMATS (which I have used) and other packages, so I hope this information is directly applicable. So the ∆PSI is how many percentage points an alternative transcript changes in prevalence as the result of a gentype/cell type/treatment/etc. If this is a very small change, e.g. it changes from 31% (0.31) to 32% (0.32), it is questionable if this is biologically significant, even if if the p-value indicates statistical significance, so you can apply some cutoff to this difference. You have to be careful, though, because some alternate isoforms can have very powerful effect (e.g. an exon-skipping even remove the protein domain responsible for the enzymatic action [e.g. phosphorylation], but the binding domain is still intact, so the alternate protein becomes an inhibitor of the original protein form; MYD88 is an example). If that happens, even small amounts can have a noticeable biological effect. So you could take the absolute levels into account by looking at the relative difference (instead of the absolute one). So a change from 0.01 to 0.02 might be interesting, because there is a doubling involved (in case of an inhibitor, doubling that could make a large impact), whereas a shift from 0.31 to 0.32 is only a 3% difference. The cutoffs you apply ideally should depend on what you are looking for and what you can expect in your experiment

I agree with you. Still, prior to comparisons, I aim to show an overview of the number of splicing events in my 65 conditions (different splicing mutants/treatments). And this overview is crucial for my big biological question. I can only think of using the PSI values for it. But should I only exclude the N/A and zeros? I wonder if I should apply a cutoff when I see such a considerable variation for some genomic locations within bioreps. Any comments or suggestions? Unfortunately, all the publications that I reviewed also show an overview of the number of events and do not describe anything on the M&M.

I think by "event" people typically mean a change. Biochemical reactions (including the various steps of splicing) are rarely 100% but rather in some form of equilibrium, and the conditions determine where that equilibrium lies. So every cell where a gene is expressed likely has multiple isoforms of the transcript, and their ratios are determined by the conditions. So looking at PSI levels by themselves without comparison is not necessarily very meaningful. I assume you have some kind of controls (untreated/wild-type), so you can compare each of your genotype/treatment groups to a control and make a statements like "applying treatment X results in n events" for each treatment/genotype (and you can plot those numbers in a bar graph, or whatever; typically the events are split by type, like SE, RI, etc., so you can see if a treatment favors one event type over others). For that you need to define "event", and that is at the minimum a cutoff (like 0.05) applied to the p-value (adjusted for multiple testing, FDR), ideally combined with some kind of biological significance threshold involving the ∆PSI (either the absolute value or relative to e.g. the level in the control). Since this is an overview, you can't really fine-tune this cutoff to the specifics of the biology, so just pick something that weeds out most of the spurious stuff without throwing the baby out with the bath water. If you really want to use the PSIs for each sample directly, you could try to correlate the samples based on PSI values for each gene. (Not sure how best to handle genes that are not expressed in some samples -- maybe throw them out.) Then you can create correlation heatmaps or dendrograms showing how closely the samples are related. You could also try principal components analysis (PCA; plot two or three dimensions) and see if certain groups cluster together. I have never tried this, and given how noisy the PSI values seem to be (possibly a result of expression, see below), I am not sure if you'll get much out of it, but it's easy to try. (If you do it, let me know.) You were worried about the high variation among samples of the same group. If you use the comparison approach, where you get (adjusted) p-values, that should take care of the variation -- if it is too high, the difference is not going to be significant and it's a "non-event". Note that PSI by itself also does not mean much -- if the gene expression is very low, a difference of 1 count can change the PSI a lot. Hope this helps you think about it.

Indeed. Thanks a lot. Great help. 🙌