Skip to content

TUSCO‑novel (Novel Isoform Stress Test)

Overview

TUSCO‑novel benchmarks a pipeline’s ability to detect truly novel isoforms by intentionally “lying” to the pipeline: the reference GTF is altered only for curated single‑isoform TUSCO genes so that the real expressed isoform appears novel relative to the supplied annotation.

Note: The curated TUSCO gene list is derived from the TSV panel (src/utilities/report_qc/tusco_<species>.tsv). TUSCO itself uses TSV panels only; no TUSCO GTF is required.

Concept

  • Alters single‑isoform, multi‑exon TUSCO genes to hide one real internal junction and replace it with a plausible synthetic junction (canonical GT–AG; reasonable intron length; still multi‑exon).
  • Tools that depend on the given annotation tend to degrade; reference‑agnostic discovery plus robust filtering (e.g., Iso‑Seq + SQANTI3 ML) better controls false positives.

Design constraints (for controlled difficulty and fairness): - Modify only internal splice junctions; preserve TSS/TTS and multi‑exon structure. - Use canonical motifs (GT–AG) and intron lengths within empirical bounds for the species/tissue. - Leave all non‑TUSCO genes unchanged to localize the perturbation.

How It Works

  1. Inputs: native reference GTF, genome FASTA, curated TUSCO single‑isoform gene list (human/mouse).
  2. For each multi‑exon TUSCO gene:
  3. Remove one true internal splice junction in the transcript.
  4. Insert a plausible synthetic junction (canonical; reasonable distance; preserves multi‑exon structure).
  5. Use this altered GTF as the only “reference” for reconstruction and for downstream classification/evaluation.

Simulator code: https://github.com/TianYuan-Liu/tusco-paper/blob/main/src/tusco_novel_simulator/tusco_novel_sim.py

When To Use

  • Quantify dependence on annotation vs sequencing data.
  • Assess novel discovery under under‑annotated or misleading references (species/tissues).

End‑to‑End Workflow

1) Prepare inputs

  • Genome FASTA: hg38.fa or mm10.fa.
  • Native reference GTF: e.g., GENCODE.
  • TUSCO gene list: derive from src/utilities/report_qc/tusco_<species>.tsv (human or mouse).
  • Extract the TUSCO gene identifiers to a text file, one per line (e.g., tusco_genes.txt).
  • Your reads/alignments as required by each pipeline (BAMs or CCS/FASTQ for Iso‑Seq).

2) Build the TUSCO‑novel GTF

Use the simulator to modify only TUSCO genes:

python tusco_novel_sim.py \
  --refGTF native.gtf \
  --genome hg38.fa \
  --tusco-list tusco_genes.txt \
  --out tusco_novel.gtf \
  --seed 42

Sanity checks: - Modify only TUSCO single‑isoform genes and only internal junctions. - Ensure synthetic junctions are canonical and yield valid multi‑exon transcripts. - Log which junction was modified per gene for reproducibility.

Provenance and determinism: - Record the commit of the simulator and configuration used. - Fix the RNG seed (e.g., --seed 42) and preserve the simulator log.

3) Run reconstruction with the TUSCO‑novel reference

  • StringTie2: bash stringtie aligned.bam -G tusco_novel.gtf -o stringtie.gtf -L
  • FLAIR (guide with altered annotation): bash flair correct -q reads.fastq -g hg38.fa -f tusco_novel.gtf -o flair_correct flair collapse -g hg38.fa -r reads.fastq -q flair_correct.bed -f tusco_novel.gtf -o flair_collapse
  • Bambu (R): provide tusco_novel.gtf as the annotation object.
  • Iso‑Seq + SQANTI3 ML: run Iso‑Seq reference‑free; use tusco_novel.gtf only for SQANTI3 classification.

4) Evaluate with SQANTI3 (TUSCO metrics)

  • TUSCO‑novel classification: bash python sqanti3_qc.py \ --isoforms <tool_output.gtf> \ --refGTF tusco_novel.gtf \ --refFasta hg38.fa \ --tusco human|mouse \ --report html -o <prefix_novel> -d <outdir>
  • Baseline (native reference): bash python sqanti3_qc.py \ --isoforms <tool_output.gtf> \ --refGTF native.gtf \ --refFasta hg38.fa \ --tusco human|mouse \ --report html -o <prefix_native> -d <outdir> Outputs are identical to the standard TUSCO report. See TUSCO Quick Start for report paths, metric definitions, and interpretation guidance.

Interpreting Results

  • Expect larger drops under TUSCO‑novel for reference‑guided tools (e.g., StringTie2, Bambu) if they rely on annotation for novel splice discovery.
  • Reference‑free discovery with stringent filtering (Iso‑Seq + SQANTI3 ML) typically:
  • Minimizes false positives (FDR), especially when junction support is enforced.
  • May show lower sensitivity/precision for exact TSS/TTS due to read length and end adjustments.
  • Compare native vs TUSCO‑novel: the gap indicates reliance on annotation vs data.

Report context: - TUSCO‑novel uses the same TSV panel and metrics as TUSCO. Interpreting Sn, nrPre, rPre, 1−FDR, PDR, and 1/red follows the same guidance as in the Quick Start.

Practical Tips

  • Restrict modification to multi‑exon TUSCO genes with well‑supported junctions; leave the rest of the annotation unchanged.
  • Use a fixed RNG seed and write a per‑gene change log.
  • Verify synthetic junctions against the genome (canonical motifs, reasonable intron sizes).
  • Keep pipeline parameters identical between native and TUSCO‑novel runs for fair comparison.

Limitations: - The stress test focuses on splice‑junction novelty; it does not simulate alternative TSS/TTS or structural variants. - Results can depend on the choice of TUSCO panel (species/tissue) and the simulator constraints; report both explicitly.

See Also

Reproducibility Checklist

  • SQANTI3 version/commit and command lines for both native and TUSCO‑novel runs.
  • Provenance of the TUSCO panel TSV (filename, species, checksum).
  • Simulator commit, configuration, RNG seed, and logs.
  • All generated logs (including tusco_report.log) and HTML reports archived.