Open-source. Reproducible. Public data only.
Five lines to a panel-based MRD score on real TCGA tumor mutations.
git clone https://github.com/rollroyces/deepcatch.git
cd deepcatch && python -m venv .venv && source .venv/bin/activate
pip install -e .
pytest -q # 256 tests, all pass
python scripts/real_tcga_validation.py --reportValidated end-to-end on real TCGA-LUAD mutations (GDC, open access) and FLARE/GSE317007 real cfDNA.
DeepCatch ships two detection channels that are individually useful and even better together:
Per-locus Poisson LLR aggregation across tumor-suppressor panel. Best at MRD-style ultra-low VAF (≤0.1%). Trained on real TCGA-LUAD mutations with a context-aware sequencing error model.
AUC 0.921 at 0.1% ctDNA on 20 TCGA-LUAD patients.
5-mer end motifs + fragment size distribution + DELFI/WPS coverage profiles. No tumor mutations required. Best at pan-cancer screening.
AUC 0.978 ± 0.002 on 627 cross-study samples from FinaleDB.
Two channels averaged. AUC 0.9886 — paired t-test vs each alone p < 0.0001, DeLong p < 0.0015 on every seed. The two channels are complementary.
Standalone cfdna-score CLI scores any sample against the trained model in <1 s with cache hit. SCORE_USAGE.md →
6-class TOO from fragmentomics alone: macro AUC 0.9533 (full), 0.9338 within-Cristiano (batch-effect removed). Top-1 79.4%, top-2 90.7%.
CpG methylation channel being developed in deepcatch-methylation. Phase 0 + Phase 1 complete; FinaleMe pretrained HMM decoder working (per-CpG β-values on BH01 chr22).
Every result on this page is 5-seed mean ± std with bootstrap 95% CI. No single-seed AUC is reported anywhere in the repo.
| Cohort | Samples | Task | AUC (mean ± std) |
|---|---|---|---|
| FinaleDB pan-cancer (Jiang 2015 + Cristiano 2019) | 627 | Cancer vs healthy | 0.978 ± 0.002 |
| Mutation × tumor-naive fusion (same 627) | 627 | Channel fusion | 0.989 ± 0.001 |
| Tumor-naive fragmentomics + synthetic mutation (5-channel) | 545 | Cross-channel | 0.9921 pooled |
| TCGA-LUAD panel LLR @ 0.1% ctDNA | 20 patients | Ultra-low VAF | 0.921 (Sens@95% = 0.770) |
| Multi-cancer 5-class OvR (LR+PCA(200)) | 236 cancer | Cancer subtyping | 0.970 ± 0.001 macro |
| Tissue-of-origin (within-Cristiano, 5-class) | 336 cancer | Tissue origin | 0.934 ± 0.002 macro |
| FLARE/GSE317007 (ONT real cfDNA) | 12 HNSCC | Cross-platform | Reproduces CG-depletion signature |
Headline comparisons vs published MCED tests at sens@99% specificity:
| Test | Cohort | Cancers | Sens @ 99% spec |
|---|---|---|---|
| Galleri (CCGA-3, methylation-based) | 4,023 | 50+ | 51.5% @ 99.5% |
| CancerSEEK (8-cancer) | 1,005 | 8 | ~70% @ ~99% |
| This — fragmentomics alone | 627 | 8 | 75.5% (CI 63.3–87.6) |
| This — fragmentomics + fusion | 627 | 8 | 84.3% (CI 80.0–92.9) |
See RESULTS.md for full per-cohort breakdown, ablations, and known limitations.
Markdown source: paper/PAPER.md
LaTeX source: paper/paper.tex
Pre-compiled PDF: paper/biorxiv_submission_v2.2.0.pdf
Author: Yu Ching Lam · ORCID 0009-0008-9113-769X · Independent Researcher, Hong Kong.
Standalone fragmentomics-only paper draft at deepcatch-methylation/BIORXIV_PAPER_FRAGMENTOMICS.md
Covers the 4 enhancements: multi-cancer, sens@spec, per-sample CLI, tissue-of-origin.
REVIEWERS.md — what we built, what we want feedback on, contact info.
AUDIT_REPORT.md — independent review notes.
MODEL.md — model card (intended use, limitations, ethics).
I'm Yu Ching Lam, an independent researcher in Hong Kong. No institution. No lab. No academic email.
What I have: open data, ORCID, GitHub, and time.
DeepCatch is built on public data only — TCGA GDC, FinaleDB, GEO, Zenodo. Every result in the repo is reproducible from open data with the provided scripts. No paywalled supplements, no institutional access required.
If this work is useful, a GitHub star helps reach more reviewers. Sponsorship funds compute and storage.