Why does flair quantify give results inconsistent with IGV?

[evans-ZH]

Hi, my data are direct RNA sequencing (DRS) reads, I found that the results from flair quantify don’t match what I see in IGV.
I ran flair quantify twice, and the -i INPUT were both generated by flair collapse and then processed with SQANTI3:

First run:
flair quantify
-i $INPUT
-r reads_manifest.tsv
--output $OUTPUT_DIR/flair.quantify
--threads $THREADS
--generate_map

Second run (following #556):
flair quantify
-i $INPUT
-r reads_manifest.tsv
--output $OUTPUT_DIR/flair.quantify_stringent
--threads $THREADS
--stringent --check_splice --quality -1
--isoform_bed $sqanti3_rescue.bed
The --isoform_bed file was generated by flair collapse and then processed with SQANTI3.

In the first run, the quantification seems closer to the expression levels observed in IGV, but sometimes tends to overestimate.
In the second run, the results are generally lower.

For example:

For one transcript, the counts were 108028 (first) vs 41522 (second), while IGV shows around 95k, lower than the first result.

For another, 71387 (first) vs 60196 (second), where IGV again seems closer to the first run.

Could you help me why is this happening, and which parameter settings should I use for DRS data to get quantification results consistent with the truth?

Thanks!