Reads Assignment Strategy for Highly Similar Isoforms and Novel Isoform Validation

[yjliuhub]

Hi Andrey,
Thank you for developing and maintaining this excellent tool.
I have a specific question regarding the reads assignment strategy when two isoforms originating from the same gene are highly similar.
Example: I am investigating the PARVB gene, which has the following two highly similar isoforms:
Reference/Known Isoform: PARVB-201
Novel Isoform identified by IsoQuant: 86796 (a novel isoform ID generated during the run)
These two isoforms appear to differ only slightly, specifically in their first and last exons.I extracted the reads assigned to both 86796 and PARVB-201 from the Embryo.transcript_model_reads.tsv.gz file and visualized them in IGV.
I found that a significant portion of the reads assigned to the novel isoform 86796 do not contain the unique first exon specific to that isoform.
In many cases, these reads seem structurally more compatible with the known PARVB-201 isoform (which might only differ in the terminal exon or lack the unique start/end features of the novel isoform 86796).

Given this observation, I am confused about the underlying mechanism of reads distribution:
Could you please clarify the strategy IsoQuant employs for assigning full-length reads when multiple competing isoforms from the same gene are nearly identical or when the reads only cover the shared internal regions?
Specifically:

1. Does IsoQuant prioritize known (reference) isoforms over novel isoforms in cases of ambiguity?
2. How does the model account for reads that support the bulk of a novel isoform's structure but fail to include a single defining feature (like a unique starting exon)?
3. What is the recommended threshold or metric within IsoQuant's output to assess the confidence of reads assigned to a novel isoform, especially when the unique distinguishing features are missing in the assigned reads?
   Understanding this assignment strategy is crucial for the biological validation of these novel isoforms.
   Thank you very much for your time and assistance!
   Best regards,
   yijia

[yjliuhub]

I'm using IsoQuant version 3.7.1.

[andrewprzh]

Dear @yjliuhub

Thanks for the feedback. I will answer your questions first and then give an overall comment on the situation below.

Does IsoQuant prioritize known (reference) isoforms over novel isoforms in cases of ambiguity?

No, it does not.

How does the model account for reads that support the bulk of a novel isoform's structure but fail to include a single defining feature (like a unique starting exon)?

It does not, if there are no unique features, the read will be marked as ambiguous. However, features also inlcude polyA positions if a polyA tail is detected (see below).

What is the recommended threshold or metric within IsoQuant's output to assess the confidence of reads assigned to a novel isoform, especially when the unique distinguishing features are missing in the assigned reads?

Unfortunately, there are none. Thanks for the idea, which might be implemented at some point.

Overall, it's tough to figure out simply from this screenshot, but I assume that polyA sites may be involved. As you see, the upper isoform and most reads terminate earlier, while the second one seems to have a long terminal exon. If those reads have a polyA tail that overlaps with this long last exon, this will be sufficient for IsoQuant to assign them to the first isoform.

It may as well happen that there is another isoform which has an intron chain matching the second isoform but a distinct polyA site. IsoQuant currently does not report such novel isoforms due to the danger of false positives. Overall, detecting alternative polyA and TSS sites is a weak part of IsoQuant, and we hope to improve it soon.

Also, using polyA sites to assign isoforms is a debatable strategy, which we might also revise in the near future.

P.S. I also recommend updating to 3.10, it has some important fixes (not related to the issue though).

Best
Andrey