Quantification of EST/cDNA based retained intron transcripts

Question

Note: this question has also been asked on Biostars

We are seeing a retained intron transcript event for some RNAseq samples, and we want to assess at the sequence level which intron retention events are actually occurring so that we can do protein domain predictions. For example, we have transcript level expression support for DROSHA-203, which according to Ensembl is a retained intron transcript that potential gives rise to multiple splicing variants via the cDNA sequence information:

In theory, the orange sites indicate where a mutation might occur which could obliterate a splicing recognition site, thereby causing an retained intron event. So in total there are 6 potential mutational events which might give rise to 4 possible retention events. Given that 4 intronic regions might be retained, the number of unique retention events is a combinatorics question equal to 2^4, e.g., intron1 retained only, vs intron1 + 2 retained, all possible unique combination, etc, correct?

Is there a better way to more granularly quantify these possible transcripts? Or do we have to PCR validate all of them to figure out which one's are actually occurring? Is there a way that I can look at the reads mapping to the corresponding intronic regions to make even a probabilistic assessment of which transcripts are present and to what extent they are expressed? The libraries are prepped for total RNA with a ribozero depletion, so I should still be able to quantify any enrichment to these intronic regions, correct?

Answer 1

You can do full-length whole-transcriptome sequencing using either the direct RNA kit or a cDNA kit (with or without PCR) on a MinION. This will allow you to quantify isoforms directly, rather than estimating by looking at read pairs.

However, it'd be a bit expensive to do it that way if you're looking at a single transcript (and aren't barcoding to multiplex samples on the same run).

If you've already done Illumina whole-transcriptome sequencing (e.g. strand-specific paired-end sequencing), tools like Kallisto, RSEM, or Salmon can be used to estimate isoform expression based on the mapped reads. These tools use the information about transcript placement and/or similarity to alter the observed read counts in an attempt to better represent reality.

Unfortunately, if you want these tools to quantify intronic sequence (i.e. the $2^4$ possibilities you have suggested), then those sequences need to be included in the sequences provided. They won't estimate frequencies for non-existent sequence.

As an alternative to including every possible gene model, something like a SuperTranscript model might work. I'm not familiar with Alicia Oshlack's method, but it at least gives the appearance that it might reduce the complexity of searching for multiple unannotated isoforms within a gene.