Major discrepancies between ESPRESSO Q & other long-read quantification methods

[sbresnahan]

I'm working with ONT cDNA reads, quantifying transcripts we assembled from these reads with ESPRESSO. I'm seeing some extreme discrepancies in the isoform quantifications computed between ESPRESSO's quant module and other long-read quant methods (Oarfish, lr-kallisto, Isoquant, even just minimap2 + featurecounts allowing all multi-feature mapping). Attached are two figures:

CSH1_exprplot is boxplots of the read counts from these long-read methods for each isoform of CSH1 in n=66 ONT cDNA libraries. Red points indicate the total counts across libraries from ESPRESSO. To highlight an example, the purple arrow & star indicate the isoform with the highest # of reads assigned by ESPRESSO, but all long-read methods gave 0 counts across samples.

CSH1_txplot shows the models for isoforms we assembled for CSH1 from the same reads. Purple arrow & star indicate the isoform in question.

Why other long-read quant methods assign 0 counts to some isoforms but not others (even those sharing splice junctions and exon structures) is unclear to me. Any thoughts about what we are seeing here?

CSH1_exprplot.pdf
CSH1_txplot.pdf

[EricKutschera]

From the plots I can't pinpoint a specific issue. For the example you pointed out where ESPRESSO assigned a large read count and the other tools gave zero counts, one possible explanation is that the reads only match that isoform after ESPRESSO realigns the reads

Based on the plots it looks like there are around 100 isoforms and millions of reads just for this one gene, CSH1. That can make things difficult to debug. Do you get a similar read count difference for a smaller dataset and does this also happen for a gene with fewer isoforms?

One thing you can try is running the Q step with --tsv_compt. Then if there is an isoform that only ESPRESSO assigns reads to, you can check that .tsv output file to see which reads ESPRESSO assigned. Viewing those reads along with the GTF in IGV might show something interesting. ESPRESSO also produces intermediate files like chr17_read_final.txt which show details about how each read was handled. Among other things those files would show if the reads assigned to that particular isoform had any splice junctions realigned (corrected)