bio-ctdna-mutation-detection

What this skill does

## Version Compatibility

Reference examples tested with: pysam 0.22+, pandas 2.2+, VarDictJava 1.8+, GATK 4.5+, Ensembl VEP 111+

Before using code patterns, verify installed versions match. If versions differ:
- Python: `pip show <package>` then `help(module.function)` to check signatures
- CLI: `<tool> --version` then `<tool> --help` to confirm flags

If code throws ImportError, AttributeError, or TypeError, introspect the installed
package and adapt the example to match the actual API rather than retrying.

Notes specific to this skill: VarDict's `-c -S -E -g` are 1-based BED COLUMN INDICES, not genomic coordinates; var2vcf_valid.pl's `-E` suppresses the END tag (opposite meaning to VarDict's `-E`). VEP gnomAD flags are `--af_gnomade` (exomes)/`--af_gnomadg` (genomes); the bare `--af_gnomad` is a legacy alias that returns only exome AF, so prefer the explicit forms.

# ctDNA Mutation Detection

**"Detect mutations in my cfDNA sample"** -> Either scan a panel for unknown low-VAF somatic variants (de novo calling) or quantify a pre-specified mutation set across samples (tumor-informed tracking) — two different statistical problems.
- CLI: `vardict-java | teststrandbias.R | var2vcf_valid.pl` for de novo low-VAF calling on a consensus BAM
- CLI: `gatk Mutect2` with the read-orientation model for de novo calling with artifact filtering
- Python: `pysam` pileup of ref/alt counts at fixed loci for known-variant tracking (MRD)

## The Single Most Important Modern Insight -- detection is a signal-vs-noise problem, and tracking is not the same problem as calling

Low-VAF ctDNA detection is set by two limits that no caller can overcome: the per-base error floor (raw Illumina ~1e-3 caps naive VAF detection near 0.5-1%) and the number of tumor molecules physically present in the tube (1 ng cfDNA ~= 303 haploid genome-equivalents; at 0.01% VAF in 10 ng the expected mutant count is ~0.3 copies — there is nothing to detect at any depth). The achieved limit of detection is the worse of the two. Error suppression (UMI consensus -> ~1e-5, duplex -> <1e-7) and input mass move the floor; swapping VarDict for Mutect2 does not.

Critically, de novo CALLING and known-variant DETECTION are different statistical problems. De novo calling scans every covered position for an unknown alt and pays a multiple-testing tax across 1e5-1e6 loci, so per-locus thresholds must be stringent (practical LoD ~0.1-0.5% on UMI consensus). Tumor-informed detection tests ONE hypothesis — "is tumor present?" — by integrating signal across a pre-specified set of N patient-specific loci; the multiple-testing penalty collapses and per-locus signal that is individually indistinguishable from noise sums into a confident panel-level call. This is why per-locus LoD is poor while panel-integrated LoD reaches single-ppm. Conflating the two is the most common conceptual error in the field.

## Methods Landscape

| Method | Class | Citation | Role | When |
|--------|-------|----------|------|------|
| VarDict / vardict-java | de novo caller | AstraZeneca-NGS | sensitive low-VAF amplicon/capture calling with explicit strand-bias test | de novo panel calling on a UMI-consensus BAM |
| Mutect2 (tumor-only) | de novo caller | GATK | local-assembly somatic caller + learned orientation-bias artifact model | de novo calling needing FFPE/OxoG artifact filtering, PoN, germline resource |
| umi-varcal | UMI-aware caller | Sater 2020 *Bioinformatics* 36(9):2718 | own UMI-aware pileup + per-position Poisson test against local background | UMI-tagged BAM where a consensus-aware caller is wanted (floor ~0.3%) |
| CAPP-Seq / iDES | tumor-informed integration | Newman 2014 *Nat Med* 20:548; 2016 *Nat Biotechnol* 34:547 | hybrid-capture deep panel + molecular barcoding + in-silico background polishing | de novo ctDNA to ~0.02%; iDES stacks ~15x error suppression |
| INVAR | tumor-informed integration | Wan 2020 *Sci Transl Med* 12:eaaz8084 | integrate variant reads across 100s-1000s patient loci, background-weighted | MRD/monitoring with tumor WES; quantifies to ~1e-5, best ~2.5 ppm |
| MRDetect | tumor-informed integration | Zviran 2020 *Nat Med* 26:1114 | shallow WGS vs patient SNV compendium, read-level SVM noise model | MRD trading depth for breadth (~35x WGS, thousands of SNVs); ~1e-5 |
| PhasED-Seq | tumor-informed integration | Kurtz 2021 *Nat Biotechnol* 39:1537 | enrich phased (co-occurring) variants to suppress single-molecule error | sub-ppm MRD where phased variants are available |

Per-locus LoD for any of these is error- and sampling-limited (~0.1-0.5%); the tumor-informed methods reach ppm only by integrating across a known, large, patient-specific variant set. Methodology evolves — verify current best practice against each tool's live docs before committing to one.

## Decision Tree by Scenario

| Scenario | Recommended | Why |
|----------|-------------|-----|
| Tumor tissue available, MRD/monitoring of a known cancer | Tumor-informed tracking (INVAR/MRDetect/Signatera-class), or the pysam tracker below for a fixed list | Integrating across N pre-specified loci is the only route to ppm; CHIP excluded by construction (CHIP variants are not on the tumor list) |
| No tumor tissue, screening/discovery | de novo panel calling (VarDict or Mutect2) + matched WBC | Must scan for unknown variants; CHIP subtraction is mandatory or most calls are not tumor |
| VAF regime > 1% | Any standard caller on a deduplicated BAM | Above the raw error floor; consensus not strictly required |
| VAF regime 0.1-1% | UMI single-strand consensus + VarDict/umi-varcal/Mutect2 | Below the raw 1e-3 floor; consensus needed to recover real signal from error |
| VAF regime < 0.1% | Duplex consensus + tumor-informed integration | Single-strand consensus cannot remove one-strand deamination/oxidation; only duplex + panel integration reaches this regime |
| No matched WBC available | Do NOT report de novo calls as somatic-tumor | Without WBC subtraction, CHIP (the majority of non-germline cfDNA variants) is indistinguishable from tumor |

## CHIP -- the dominant false positive, not background

Clonal hematopoiesis of indeterminate potential (CHIP) is the single largest source of false-positive somatic calls in plasma, and it is the null hypothesis for any low-VAF cfDNA variant. Razavi 2019 sequenced cfDNA with matched white-blood-cell DNA (508 genes, >60,000x) and found that **53.2% of non-germline cfDNA variants in cancer patients and 81.6% in non-cancer controls** had features consistent with clonal hematopoiesis; only ~24.4% of cfDNA somatic variants in patients were also in the matched tumor (the remainder split between white-cell CHIP and variants of uncertain origin). These are bona fide somatic mutations — in cancer genes — that come from lysed leukocytes, not tumor. No error-suppression tier removes them because they are not errors.

The biology: CHIP arises in hematopoietic stem cells and rises steeply with age (Jaiswal 2014: ~10% prevalence over age 70), enriched for PPM1D/TP53/CHEK2 clones after prior chemo/radiation — exactly the monitored population. The canonical genes are DNMT3A, TET2, ASXL1 (the big three), then PPM1D, TP53, JAK2, SF3B1, SRSF2, GNB1, GNAS, CBL, ATM, CHEK2. TP53 and ATM are both CHIP genes and bona fide tumor suppressors, so a low-VAF TP53 cfDNA call is the ambiguous case par excellence.

The only reliable filter is matched buffy-coat/WBC subtraction: sequence the WBC fraction of the same draw at comparable depth and remove any cfDNA variant also present in WBC. gnomAD filtering removes germline only — CHIP variants are somatic and absent from germline databases, so they sail straight through. A canonical-CHIP-gene list (the example's `CHIP_GENES`) is a heuristic flag for extra scrutiny, NOT a substitute for WBC subtraction. See analytical-validation for the LoB/LoD statistics that quantify how confidently a subtracted call clears background.

## De Novo Calling with VarDict

**Goal:** Scan a target panel