Question: Using IsoQuant after Iso-Seq preprocessing pipeline for bulk long-read RNA-seq

[ljh-9809]

Hi IsoQuant team,

I have bulk PacBio long-read RNA-seq data that has been preprocessed using the PacBio Iso-Seq pipeline (CCS → Lima → Isoseq3 refine → Isoseq3 cluster), and then mapped to a reference genome using pbmm2, resulting in a sorted and indexed BAM file.

My question is:
Is it appropriate to use this mapped BAM file (output from the Iso-Seq preprocessing pipeline including clustering) directly as input to IsoQuant via --bam for bulk isoform quantification?

Or is it recommended to use an earlier-stage reads (e.g., before clustering or before refine) instead, and let IsoQuant handle the downstream processing internally?

Any guidance on the recommended workflow when starting from Iso-Seq pipeline output would be greatly appreciated.

Thank you!

[andrewprzh]

Dear @ljh-9809

I suggest starting from CCS reads (with non trimmed poly-A tails if possible). I think discovery and especially quantification will produce more reliable results. The main reason is that initial read support information is lost during clustering.

Out of curiosity, you can also try running on clustered data, probably using -d transcripts instead of -d pacbio and see if the results are different. However, IsoQuant was never tested on such data and cannot say how reliable the results will be.

Best
Andrey

[ljh-9809]

Hi Andrey,

Thank you so much for the detailed and helpful response!

I have a follow-up question regarding filtering of the IsoQuant output before downstream analysis. Specifically, I'm planning to run SQANTI3 on the IsoQuant-generated GTF, and I'm wondering about recommended filtering strategies based on SQANTI3's artifact categories, such as removing transcripts flagged as intra-priming artifacts or RT switching artifacts.

Do you have any recommended criteria for filtering transcripts based on these SQANTI3 artifact flags? Or alternatively, are there any filtering options within IsoQuant itself that you would suggest applying beforehand to reduce such artifacts in the output?

Thank you again for your help!

[andrewprzh]

Dear @ljh-9809

Generally, IsoQuant outputs isoforms with high precision and additional filtration is not mandatory.
I have not used SQANTI myself, so I cannot say which filters can be useful.
If you find any suspicious transcripts, it would be interesting to see which categories are those.

Best
Andrey

[ljh-9809]

Hi Andrey,

Thank you again for your guidance. I have a follow-up question regarding the recommended CCS reads input.

You suggested starting from CCS reads with non-trimmed polyA tails. In my case, the data is multiplexed (multiple samples with PacBio Iso-Seq barcodes in a single SMRT Cell), so I need to demultiplex before running IsoQuant.

I tested three different input strategies:

Lima output (demuxed CCS, unmapped BAM) with -d pacbio_ccs --fl_data — closest to your recommendation, but SQANTI3 classification showed a very high proportion of genic_intron (38%), intergenic, and antisense categories
Clustered + aligned BAM with -d pacbio_ccs — cleaner results but uses post-clustering data
Clustered + aligned BAM with -d transcripts — cleanest results (50% FSM), as you suggested for curiosity
My questions:

When you say "CCS reads," do you mean reads before or after Lima demultiplexing? Since IsoQuant does not seem to handle Iso-Seq sample barcode demultiplexing internally, Lima appears to be a necessary preprocessing step.
For the Lima output approach, would you recommend using the refined (FLNC) reads from isoseq3 refine (which removes concatemers and artifacts but trims polyA) instead of raw Lima output? Or is the polyA preservation more important?
The noisy results from the Lima output (strategy 1) — is this expected behavior, or could there be a parameter issue?
Thank you for your help!

[andrewprzh]

Hi @ljh-9809

Unfortunately, I am not that familiar with IsoSeq pipeline, but I'll try to answer based on your quite comprehensive information.

Yes, Lima demultiplexing is a must, IsoQuant cannot handle that. However, as IsoQuant can perform barcode detection in single-cell data, I might consider including in the pipeline at some point.

If you do novel transcript discovery, polyAs are quite important and various artifacts should also be filtered out during that process. However, you can use FLNC reads with trimmed polyAs and also set --polya_trimmed option (you can set either all or stranded value), which would indicate that reads were polyA trimmed and IsoQuant must treat them as if the polyA is detected.

You can try both ways and see which outcome sounds more reasonable, I would be also curious to see how it works.

Best
Andrey

[ljh-9809]

Hi Andrey,

Thank you for the suggestion! I tested FLNC reads with both --polya_trimmed stranded and --polya_trimmed all (along with --fl_data). Here are the SQANTI3 classification results:

Category	stranded	all
Total transcripts	76,805	224,411
FSM	23,446 (30.5%)	28,439 (12.7%)
ISM	13,333 (17.4%)	12,071 (5.4%)
NIC	11,079 (14.4%)	23,461 (10.5%)
NNC	12,714 (16.6%)	11,197 (5.0%)
genic_intron	6,543 (8.5%)	92,373 (41.2%)
antisense	6,167 (8.0%)	12,648 (5.6%)
genic	1,621 (2.1%)	27,199 (12.1%)
intergenic	884 (1.2%)	16,092 (7.2%)
fusion	1,018 (1.3%)	931 (0.4%)
Mono-exon	1,747	160,315
Artifact %	21.1%	66.5%

stranded is clearly better , all caused a mono-exon explosion (1,747 → 160,315), mostly classified as genic_intron/genic/intergenic artifacts.

With stranded, the results are much improved compared to the Lima input I reported previously (genic_intron dropped from 38% to 8.5%). However, the overall artifact proportion (genic_intron 8.5% + antisense 8.0% ≈ 21%) is still considerably higher than the LRGASP benchmark results (~0.9% for PacBio cDNA, Long Only).

Is this level of genic_intron/antisense expected when running IsoQuant on real tissue samples, or are there additional IsoQuant parameters that might help reduce these categories?

Thank you!

[andrewprzh]

Dear @ljh-9809

Thanks for sharing insights! Yes, monoexonic (and in fact mono-intronic) alignments are not very reliably. For example, when --data_type nanopore is set, IsoQuant simply avoids reporting novel monoexonic transcript at all (you can control this by --report_novel_unspliced false).

Do you know what kind transcripts fall into these categories (genic_intron + antisense), e.g. monoexonic, monointronic etc?
You can also check if they originate from reads with low mapping quality (this filtering can be controlled using IsoQuant options).
Also, what happens if you --polya_trimmed none?

Best
Andrey