Significant alternative splicing detected by rMATS, but no corresponding transcript-level expression — expected behavior?

[Jia-nianhua]

Hello,

I encountered a situation where a statistically significant alternative splicing event was detected by rMATS, but this does not seem to correspond to any detectable transcript-level expression. I would like to understand whether this is expected.

For one example gene (maize geneA):
rMATS reports a significant splicing event:
FDR ≈ 0.047
IncDifference ≈ -0.115
Group1 shows higher inclusion (long exon usage)
However, transcript-level quantification (featurecounts -t exon -g transcript_id --countReadPairs ) shows:
Only one annotated transcript (T001) has measurable expression
Other annotated isoforms (e.g. T002, T003) have ~0 expression

Thank you~！

[EricKutschera]

rMATS checks the reads supporting the two isoforms of the splicing event. There is a diagram showing how reads are counted for the different splicing events: https://github.com/Xinglab/rmats-turbo/tree/v4.3.0#output

Since you mentioned long exon usage, it seems like an A3SS or A5SS event. For the event to be significant, there must be some reads supporting the junction from the short exon to the flanking exon and there must be some reads supporting the long exon

I'm not sure exactly what featurecounts does, but it has a parameter -O to allow reads that overlap multiple features. It might be that featurecounts only assigns reads to the transcript with the longer exon because any reads covering the shorter portion of the exon would overlap both isoforms

You could try looking at the region of this splicing event in IGV with your alignment files and annotation loaded to manually check how reads match up with the transcripts

[Jia-nianhua]

rMATS checks the reads supporting the two isoforms of the splicing event. There is a diagram showing how reads are counted for the different splicing events: https://github.com/Xinglab/rmats-turbo/tree/v4.3.0#output

Since you mentioned long exon usage, it seems like an A3SS or A5SS event. For the event to be significant, there must be some reads supporting the junction from the short exon to the flanking exon and there must be some reads supporting the long exon

I'm not sure exactly what featurecounts does, but it has a parameter -O to allow reads that overlap multiple features. It might be that featurecounts only assigns reads to the transcript with the longer exon because any reads covering the shorter portion of the exon would overlap both isoforms

You could try looking at the region of this splicing event in IGV with your alignment files and annotation loaded to manually check how reads match up with the transcripts

Hi, thank you for your help.

I encountered a question regarding a retained intron (RI) event.
In my rMATS results, the IncLevelDifference is negative, which I understand means that group2 has a higher inclusion level and therefore shows more intron retention.

However, when I visualized this event using a sashimi plot, the result seems inconsistent with this interpretation. In the plot, group1 appears to have higher RPKM coverage over the intronic region and also shows more splice junction reads around the event, which visually suggests stronger intron retention in group1.

I am a bit confused about how to reconcile these observations.
Could you help clarify whether I might be misinterpreting the sashimi plot, or if there are factors that could explain this discrepancy?

Thank you very much!

[EricKutschera]

The inclusion level is roughly inc_count/(inc_count + skip_count) where inc_count is the number of reads supporting the inclusion isoform (intron retention) and skip_count is the number of reads supporting the skipping isoform (intron spliced out): #286

The sashimiplot shows two splice junctions and I can't tell which junction counts from the picture correspond to the RI event in the rMATS output, but it seems like group 1 has higher junction counts than group 2 for both junctions. The intron coverage also seems higher in group 1. With both higher inc_count and higher skip_count it's not clear how the inclusion levels will compare between groups

The rMATS output file gives the counts in the columns like IJC_SAMPLE_1. If you ran with --individual-counts then the output file will also have columns corresponding to the different lines in the diagram I mentioned before. If the rMATS output counts don't match with what you see in the sashimiplot, it could be due to the differences in counting described in this post, but I wouldn't expect a large difference: Xinglab/rmats2sashimiplot#33