Skip to content

Prostate Cancer Detection doc — MethylPipeline fitness analysis

Status: research / design note (not operator runbook).
Source analyzed: Prostate Cancer Detection.md
Related: BuffyCoat_vs_cfDNA_for_Cancer_Detection.md, SaMD SOP docs/usage/18-samd-study-lifecycle.md, evidence index docs/regulatory/validation-evidence-index.md
Interactive canvas: pca-detection-fitness (sync with bash scripts/sync_cursor_canvases.sh to open beside chat)

Date: 2026-07-10
Verdict: adopt-with-caveats

Executive verdict

The clinical gatekeeper framing (high NPV for Gleason ≥3+4 / GG≥2, separate indolent GG1) and the critique of buffy-only / shallow WGBS for early localized csPCa are scientifically directionally correct and align with MethylPipeline’s own analyte research. The wet-lab SOW (EM-seq + hybrid capture) and pan-cancer “hierarchical NN” are not “apply with few/no pipeline changes”—they are mostly upstream assay design and future ML architecture. MethylPipeline can support the software/evidence half of a gatekeeper program today via the SaMD ladder + staged cohorts + clinical metrics, but only after analyte/endpoint redesign and a few small, disease-agnostic extensions.

Assertion-by-assertion evaluation

Claim Support Risk
Gatekeeper must prioritize NPV ≥95% for csPCa (GG≥2), with sens/spec + prevalence Strong clinical consensus for rule-out tests; doc correctly notes NPV prevalence dependence 95% is a clinical target, not evidence MethylPipeline (or any assay) currently meets; must not become a Python default
Separate GG1 (3+3) from GG2 (3+4) Clinically sound transition point Biologically hard in liquid biopsy; current Buffy healthy-vs-PCa work does not test this
Estimate pattern-4 % burden Plausible clinical utility; doc admits need for pathology validation Speculative for cfDNA/urine methylation; continuous endpoint not in pipeline
Spatial localization for targeted biopsy High clinical value for imaging Out of scope for blood/urine methylation software; weak molecular claim
Localized PCa → very low ctDNA fraction; WGBS inefficient Well-supported literature consensus Implies assay redesign (depth/targeting), not just software knobs
Buffy alone unlikely to grade-discriminate Aligns with BuffyCoat_vs_cfDNA_for_Cancer_Detection.md Current Buffy SaMD research path is host-response, not gatekeeper-grade biology
Buffy as matched hematopoietic control Sound (esp. mutation assays; useful for methylation confounding) Needs paired study design; not automatic today
Prefer targeted deep methylation (urine/cfDNA) over WGBS for gatekeeper Sound engineering for low TF Wet-lab + CRO; pipeline consumes FASTQ→BAM→H5 regardless
Candidate genes (GSTP1, APC, RASSF1A, PITX2, …) Reasonable literature candidates; doc labels them candidates Must stay study/profile priors, never Python hardcoding
EM-seq + hybrid capture @ 2–5k× Operationally plausible SOW Outside MethylPipeline; chemistry is upstream
GRAIL = wide/shallow MCED vs narrow/deep gatekeeper Directionally correct tradeoff Some numbers (0.05% TF, “stellar” Stage I) are marketing-grade speculation
Modular pan-cancer via bigger probe set + TOO NN Wet-lab modularity OK “Hierarchical neural network” ≠ current MethylPipeline (ECDF/FeatureCuts/MC stability)

Recommendation feasibility matrix

Recommendation Feasibility in MethylPipeline
Report NPV/PPV/sens/spec with CIs for intended-use population Already possible — packages/methylvalidation/methyl_validation/clinical_performance.py, Wilson CIs; holdout Workflow 3
Gleason-staged labels (3+3 … 4+5) as disease stages Config-only — study manifest stages (usage ch.16); SaMD staged program
SaMD evidence ladder before clinical claims Already possible — samd_research → samd_holdout_enrichment → samd_pivotal; claims blocked pre-pivotal
Buffy research with dual_fc Already done — Buffy → samd_research + dual_fc
Switch primary analyte to cfDNA + fragmentomics Config-only / ops — regulatory.primary_analyte: cfdna, SamplePrep fragmentomics
Empirical panel via MC stability → freeze fixed_dmp_panel Already possible — discovery path; compatible with capture BED if regions are sequenced
Define positive class as csPCa (GG≥2) for rule-out NPV (GG1 as non-positive) Small code — today screening_binary is control_vs_pooled_disease only
Acceptance gate on NPV LCB (+ prevalence metadata) Small code — only min_sensitivity_lcb / min_specificity_lcb exist today
High-NPV operating-point selection Small code — metrics exist; rule-out threshold policy not first-class
Soft gene/region priors (BED/gene list) for enrichment/caps Small–medium code — prefer profile/site overlays; do not hardcode GSTP1 et al.
Paired plasma + buffy background Config + study design — analyte docs endorse; paired subtraction not a dedicated action
EM-seq / Twist-Agilent SOW Out of pipeline — wet-lab/CRO
Pattern-4 % continuous model Large redesign — new endpoint + labels + validation
Spatial lesion localization Reject for MethylPipeline — imaging/other modality
Pan-cancer TOO hierarchical NN Large redesign — not current architecture
New SQL NodeTypes for “prostate gatekeeper” Do not — architecture forbids

MethylPipeline fitness gaps & proposed extensions

What fits today (software/evidence)

  • Disease-agnostic MC stability, FeatureCuts (dual_fc), enricher/PPI, freeze, locked model, PCCP scaffolds
  • Multi-stage Gleason cohorts + OVR-style backends
  • Clinical performance reports with NPV; partitions; SaMD claim gating
  • cfDNA path with fragmentomics + analyte-match guards
  • Evidence index under docs/regulatory/ (Buffy package still stability-oriented / incomplete freeze→model for gatekeeper claims)

Gaps vs the document’s ideal gatekeeper

  1. Wrong primary biology for current Buffy work if the claim is pre-biopsy csPCa rule-out (doc + internal analyte research agree).
  2. Wrong screening definition for gatekeeper NPV: pooled “any disease” ≠ GG≥2 positive.
  3. No NPV LCB / prevalence-conditioned acceptance in config schema.
  4. No pattern-4 burden or spatial outputs.
  5. Discovery assumes WGBS-style genome-wide HDF5; targeted capture is upstream—pipeline can still run if extractor outputs compatible H5, but region-aware depth/QC assumptions may need profile tweaks.
  6. Evidence packages lack holdouts / pivotal partitions for live prostate studies.

Preferred extension style

Profile/site/program overlays + study manifests; workers keep consuming resolvedConfig; no disease-specific Python defaults; no new NodeTypes.

What NOT to do

  • Hardcode prostate genes, Gleason cutoffs, or “NPV≥0.95” into package DEFAULT_* / Pydantic defaults
  • Create prostate-specific DomainPrograms or SQL NodeTypes
  • Treat Buffy healthy-vs-PCa BA as gatekeeper clinical evidence
  • Enable allow_clinical_performance_claims before samd_pivotal + populated pivotal_validation
  • Bake GRAIL-comparison marketing numbers into acceptance criteria
  • Implement spatial localization or pan-cancer TOO NN “because the doc said so”

Top 5 actionable extensions (impact vs effort)

  1. Configurable clinical screening roles (e.g. positive = GG≥2; GG1 as control-side or separate class) + report NPV for that definition — high impact / small code
  2. min_npv_lcb (+ optional prevalence / intended-use metadata) in validation schema + clinical_performance gates — high impact / small code
  3. Stand up plasma (or urine) Gleason-staged study on samd_research + dual_fc with real locked_test partitions — high impact / config+ops (largest real blocker is data, not code)
  4. Rule-out operating-point selection (threshold policy optimizing NPV/sensitivity under prevalence assumptions) — medium-high impact / small–medium code
  5. Soft region/gene prior overlays (BED or gene list via profile/site actionConfig) to bias discovery/enrichment toward candidate DMRs without hardcoding — medium impact / small–medium code

Honorable mention: paired cfdna+buffy_coat manifests for background control (config/study design).

Open questions (product / study design)

  1. Is the near-term intended use still buffy host-response, or a commitment to a pre-biopsy gatekeeper claim (requires plasma/urine + pathology GG labels)?
  2. Are there (or can we acquire) pre-biopsy cohorts with GG1 vs GG≥2 labels and patient-disjoint holdouts?
  3. Will wet-lab move to EM-seq + hybrid capture, or stay WGBS for discovery then shrink the panel later?
  4. Are pattern-4 % and spatial localization in-scope SaMD claims, or aspirational only?
  5. For screening NPV: should GG1 count as “negative” (defer biopsy) or as a third category (not rule-out)?

Architecture constraints (locked for any follow-on work)

  • Disease-agnostic, config-not-code, four-layer config (site / profile / study / program)
  • SaMD ladder: samd_research → samd_holdout_enrichment → samd_pivotal + researchMode overlays
  • DomainPrograms are algorithm-generic fixtures (mc_stability, etc.); study facts in /work/projects/...
  • Workers consume resolvedConfig; no new SQL NodeTypes for science