Skip to content

DeepCatch v2.2.0 — Panel-Based Ultra-Sensitive MRD Detection Benchmark + Paper

Latest

Choose a tag to compare

@rollroyces rollroyces released this 11 Aug 07:31
· 241 commits to main since this release

📄 The paper

An Open-Source Benchmark for Panel-Based Ultra-Sensitive Detection of Circulating Tumor DNA from Real Tumor Mutations

  • paper/PAPER.md (markdown) · paper/paper.tex (LaTeX, bioRxiv-ready)
  • 20 real TCGA-LUAD patients, 5,738 mutations, GDC open access
  • Context-aware sequencing error model (CpG 10×, homopolymer 5×, clean baseline)
  • Three scoring methods at fixed specificity, no threshold optimization on test data

🎯 Key results (verified in results/real_tcga_validation.json)

Panel detection (20 patients × 5 seeds)

ctDNA fraction LLR AUC Fisher AUC Strand AUC Sens@95% Paired win
10% 1.000 1.000 1.000 1.000 1.000
5% 1.000 1.000 1.000 1.000 1.000
1% 1.000 1.000 1.000 1.000 1.000
0.5% 0.9995 0.9965 0.9955 1.000 1.000
0.1% 0.9210 0.8340 0.8310 0.770 1.000

Ultra-early assay sweep (0.1% ctDNA) → production specifications

Error rate Depth AUC Sens@95%
2e-3 5,000× 0.921 0.770
2e-3 50,000× 0.998 0.990
1e-4 (duplex-UMI) 5,000× 0.998 1.000
1e-4 (duplex-UMI) 50,000× 1.000 1.000

Conclusion: duplex-UMI error suppression (≤1e-4) or ~50,000× depth each achieve Sens@95% = 1.000 at 0.1% ctDNA.

🧬 Fragmentomics cross-validation

  • FLARE (GSE317007): 12 real cfDNA samples, 6 HNSCC patients, ONT sequencing
  • Reproduced published pan-cancer signature: CG depletion (ratio 0.034), AT enrichment (0.106), AAAA top motif

🔬 Engineering

  • 228/228 tests green (core + DL modules + real-data guard + smoke)
  • Full benchmark reproducible in one command:
    python real_tcga_validation.py --n-patients 20 --seeds 5 --cancer-types LUAD
  • Fixed foundation-data reproducibility bug (per-process hash() → hashlib.md5)
  • CI runs real test suite (core + DL jobs) + fail-loud guard that refuses synthetic fallback

📦 All data open-access

  • TCGA-LUAD MAFs (GDC) — auto-downloaded by pipeline
  • FLARE fragmentomics (GEO GSE317007)
  • GSE185307 cfDNA methylation (GEO)

🤝 Seeking expert review

See REVIEWERS.md. Feedback welcome via GitHub Issues or PR comments.

Honest framing: simulation benchmark with real tumor mutations as ground truth; plasma reads are simulated. Clinical validation on real plasma cfDNA is the essential next step.