oarfish reports only shortest isoform despite long-read support for longer isoforms

[FenFotop]

Hi! I am running oarfish on single-cell long-read data generated with PacBio Kinnex. The dataset contains hematopoietic cells, which abundantly express the PTPRC (CD45) gene. PTPRC has well-known alternative isoforms, so we are treating it as a positive control.
In the alignment, we see some reads spanning the canonical isoform (ENST00000442510.8, that is expected to be expressed - see photo attached and table with reads from the minimap2 alignment). However, the oarfish output reports expression only for the shortest annotated isoform (ENST00000367364.5), with no counts assigned to the longer ones.
This is unexpected given both the known biology and the presence of long reads supporting longer isoforms.
Is this behavior expected? If so, are there any recommended parameter tweaks, annotation tricks, or general workarounds to help oarfish recover these longer isoforms? I like the tool and would prefer to adapt my setup rather than switch methods.
I provide the details of my minimap2 and oarfish runs. Thank you!!

Minimap2

minimap2 -t 16 -ax splice:hq --secondary=no /path/dataset_assembly.mmi  /path/Lib_dedup.fasta \
        | samtools sort -@16 -o Lib_alignment.bam

Oarfish
oarfish --single-cell --alignments /path/Lib_alignment_namesorted.bam --output Lib_out --threads 8 --filter-group nanocount-filters --model-coverage --bin-width 100

[rob-p]

Hi @FenFotop,

Thank you for the report. Certainly this is not the expected behavior, as the abundances should be estimated in light of all of the evidence. The first thing I would try would be to see the effect of different specific parameter choices in this case. For example, the nanocount-filters discard alignments that are too short with respect to the underlying transcript, or that are too far from the 5' or 3' end of the annotated transcript. Perhaps reads compatible with the longer transcript are being removed by some of those filters?

The other thing you might look at is the effect of the coverage model in this scenario (though, given the coverage patterns you are showing above, I'd expect the coverage model to try harder to assign some reads to the longer isoform, rather than the shorter).

If we have a few more data points on the effect of these parameters, then we can help diagnose any potential issues more easily.

Best,
Rob