Question about SQANTI3 filtering false positive rate on RefSeq reference transcripts

[Xiaoguang-Shi]

Is there an existing issue for this?

• I have searched the existing issues

Have you loaded the SQANTI3.env conda environment?

• I have loaded the SQANTI3.env conda environment

Are you running SQANTI3 on its latest release?

• I have SQANTI3 updated to the latest release

Problem description

We would like to thank you for developing SQANTI3 for transcriptome quality control. We recently performed transcript assembly using long-read sequencing. When applying SQANTI3, a large proportion of our transcripts were classified as artifacts. To evaluate the reliability of the filtering, we ran SQANTI3 v6.0 QC + default filters on RefSeq reference transcripts (181,361 entries) using GENCODE v49 as the reference annotation. Results: 28,403 transcripts (15.7%) had RTS_stage=True, 21,617 (12.0%) were flagged as Artifact by the default rules filter, and 80,086 (44.2%) were flagged as Artifact by the ML filter. Is it normal that 15.7% of reference transcripts are labeled RTS_stage=True? If a substantial fraction of gold-standard reference transcripts carry the RTS=True label, does the RTS-based filtering risk removing genuine but novel transcripts in long-read discovery projects? Could you suggest parameter adjustments to reduce such false negatives?

SQANTI3 commands used:

# QC
sqanti3_qc.py \
    --isoforms refseq.gtf \
    --refGTF gencode.v49.annotation.gtf \
    --refFasta GRCh38.p14.genome.fa \
    --CAGE_peak refTSS_v4.1_hg38_nohdr.bed \
    --polyA_motif_list mouse_and_human.polyA_motif.txt \
    -o refseq_qc -d ./ -t 40 --report skip

# Rules filter
sqanti3_filter.py rules \
    --sqanti_class classification.txt \
    --output refseq_rules \
    --dir filter_rules \
    --cpus 40

# ML filter
sqanti3_filter.py ml \
    --sqanti_class classification.txt \
    --output refseq_ml \
    --dir filter_ml \
    --cpus 40

Code sample

No response

Error

No response

Anything else?

No response