Prostate Cancer Application Deep-Dive¶
Status: research / application synthesis (not an operator runbook).
Audience: scientists, physicians, and insurance / medical-director readers who need to understand why MethylPipeline’s analyte, assay, and validation choices matter for prostate cancer.
Follow-on operator surface: Prostate cancer options catalog (ch.25) — what to set (assay, alignment, extraction, tissue progression, SaMD ladder). Pattern: Methylation application packs.
Related notes
| Document | Role |
|---|---|
| BuffyCoat_vs_cfDNA_for_Cancer_Detection.md | Analyte biology and pairing |
| Prostate Cancer Detection.md | Gatekeeper clinical framing; EM-Seq + GRAIL contrast; wet-lab SOW |
| Prostate_Cancer_Detection_MethylPipeline_Fitness.md | What the platform can / cannot claim today |
| ANALYTE_PROFILES.md | Operator analyte config |
| validation-evidence-index.md | EV-PCA-* feasibility packages |
| SaMD study lifecycle | Research → holdout → pivotal ladder |
| Prostate cancer options catalog (ch.25) | Operator/researcher: what to set (procedures, tissue progression, commands) |
| florida-community-access-partnership-brief.md | Florida community-access brief; Wong/Wang MHL+OS as a published method class on GoliathOmics |
Claim boundary (read first)¶
MethylPipeline enforces process and configuration controls and helps assemble evidence scaffolds. Architecture, SaMD profiles, and filled feasibility packages are not FDA clearance, CLIA validation, or coverage-ready clinical performance.
Registered prostate evidence packages (EV-PCA-PLASMA-2026-06, EV-PCA-HGOOD-2026-06, EV-PCA-BUFFY-2026-07) are feasibility / engineering evidence. They currently lack patient-disjoint locked_test / pivotal_validation partitions and must not be cited for gatekeeper NPV, biopsy deferral, or payer medical-necessity claims. See samd-submission-scaffold.md.
This note synthesizes existing research and platform capabilities. It does not invent a winner for the unfinished linear vs pangenome_wgbs SamplePrep compare, and it does not equate an EM-Seq gatekeeper design with GRAIL Galleri claims or clearance.
End-to-end map¶
flowchart LR
analyte[AnalyteOrAssayChoice] --> path{WGBS_or_EMSeq}
path -->|WGBS| align[Linear_vs_PangenomeWGBS]
path -->|EMSeq_targeted| panel[LockedPanelBED]
align --> qc[QC_informME_deconv]
panel --> qcEm[QC_onTarget_depth]
qc --> stab[Stability_MC]
qcEm --> stab
stab --> bio[BiomarkerConfirm]
bio --> model[Model_MC_ECDF]
model --> holdout[LockedTest_Pivotal]
holdout --> claim[ClaimGate]1. Chosen analyte and assay modality¶
MethylPipeline is analyte-agnostic: the same DomainProgram / profile machinery runs for leukocyte DNA and plasma cfDNA. Analyte and chemistry change biology, QC defaults, and clinical interpretation—not which orchestration engine you use.
Present four options as a decision table, not a premature single winner.
Option decision table¶
| Option | Material / chemistry | Typical role | Physician workflow fit | Depth / design note | Procedure pack |
|---|---|---|---|---|---|
| 1. Buffy ~30× WGBS | Leukocyte DNA, genome-wide | Host-response / systemic epigenome; aggressiveness research | Elevated-PSA research into host signatures; not preferred alone for tumor-shed rule-out | ~30× often adequate for buffy | buffy_wgbs_pangenome_gene_fc (or linear alternate) |
| 2. Plasma ~30× WGBS | cfDNA, genome-wide + fragmentomics | Tumor-shed discovery; MCED-adjacent research | Early-detection research when TF allows; production only after SaMD ladder | TF-aware design; not “more buffy depth” | cfdna_wgbs_plasma |
| 3. Paired buffy + plasma | Same subject, both tubes | Plasma = tumor-oriented signal; buffy = host / hematopoietic background | Best when logistics allow (CHIP-analogue / leukocyte confounding control) | Separate manifests or matched IDs | Both procedures, matched cohorts |
| 4. Plasma EM-Seq + hybrid capture | cfDNA, targeted panel | Pre-biopsy gatekeeper geometry (inch-wide / mile-deep) | Elevated PSA / secondary markers → defer vs proceed to biopsy after pivotal evidence | Locked target_panel_bed; extreme on-target depth (~2–5k×), not genome-wide depth |
cfdna_emseq_targeted |
| 5. Plasma EM-Seq MHL + survival | Same chemistry as 4; haplotype blocks | mCRPC OS prognosis (Wong/Moffitt method class) | Oncology follow-up / nomogram research — not a GG≥2 gatekeeper | {chrom}-CG.mhap.h5 + survival_path; native MHB/MHL + Cox |
cfdna_emseq_mhl_survival |
Biological roles (options 1–3)¶
| Analyte | What methylation mainly reflects | Detection role |
|---|---|---|
| cfDNA (plasma) | Shed DNA including tumor-derived fragments when tumor fraction is high enough; fragmentomic structure | Direct tumor / MCED / monitoring-oriented signal |
| Buffy coat | Host immune / systemic epigenome (inflammation, aging, cell-type mix); matched hematopoietic background | Indirect risk / host-response; complementary layer |
Buffy-only shallow WGBS is a poor fit for a pre-biopsy csPCa (GG≥2) rule-out claim: localized tumors shed little ctDNA into leukocyte DNA, and host-response signatures do not substitute for tissue anchors (e.g. GSTP1-style tumor panels). Buffy remains valuable for host biology, cell-composition covariates, and paired background control. Details: BuffyCoat_vs_cfDNA_for_Cancer_Detection.md.
Option 4 — EM-Seq gatekeeper vs GRAIL (shared biology, different geometry)¶
The wet-lab and clinical framing in Prostate Cancer Detection.md recommends Enzymatic Methyl-Seq (EM-Seq) + custom hybridization capture for a prostate gatekeeper. The shipped procedure is cfdna_emseq_targeted: operator-supplied panel BED, elevated min_cov, SamplePrep via sample_prep_emseq, no genome-wide DMP hunt, no cell deconvolution by default.
Option 5 is the same wet-lab geometry with a different science path: native MHB/MHL + Cox overall survival (cfdna_emseq_mhl_survival). That is the Moffitt / Wong 2026 method class, not FeatureCuts. See florida-community-access-partnership-brief.md.
| Dimension | GRAIL Galleri (MCED) | Prostate EM-Seq gatekeeper (this option) |
|---|---|---|
| Biology | cfDNA methylation | Same |
| Chemistry philosophy | Targeted methylation enrichment | EM-Seq (gentler than bisulfite for low-yield cfDNA) + hybrid capture |
| Geometry | Miles-wide / moderate depth (many cancer types) | Inch-wide / mile-deep (≈50–100 prostate / GG≥2 regions, very high on-target depth) |
| Clinical objective | Asymptomatic multi-cancer screen; ultra-high specificity | Single-cancer pre-biopsy rule-out for clinically significant PCa |
| Early localized PCa | Structurally quiet for wide/shallow panels | Design goal: recover signal at very low tumor fraction |
Cite GRAIL only as shared methylation biology + contrasting width/depth/objective. Do not imply Galleri-equivalent performance, clearance, or MCED positioning.
When each option helps (and when it does not)¶
| Clinical question | Prefer | Avoid claiming from |
|---|---|---|
| Host / immune correlates of PCa aggression | Buffy WGBS + deconv | Buffy as tumor-shed NPV gatekeeper |
| Genome-wide discovery of plasma methylation features | Plasma WGBS + fragmentomics | Equating discovery BA with biopsy-deferral policy |
| Background control for plasma features | Paired designs | Unpaired buffy trained → plasma deployed without retrain |
| Pre-biopsy rule-out of GG≥2 | EM-Seq targeted (after pivotal evidence) | Feasibility WGBS healthy-vs-pooled-PCa BA alone |
2. Alignment options (WGBS) and EM-Seq SamplePrep¶
WGBS arms (options 1–2)¶
Three SamplePrep modes are not interchangeable:
| Mode | Mechanism | Role |
|---|---|---|
linear |
Parabricks fq2bam_meth + MethylExtract |
Biological baseline |
pangenome_wgbs |
methylGrapher C2T/G2A bisulfite-aware pangenome | Candidate production default for buffy research packs |
Stock pangenome / Giraffe |
Parabricks Giraffe | Engineering comparator only — no WGBS methylation parity |
How option “best of linear vs pangenome_wgbs” is decided: run the dual-align harness (sample_prep_test_bed.md; scripts/compare_sample_prep_linear_vs_wgbs.sh; plan linear-vs-wgbs-sampleprep-compare.plan.md). Metrics include alignment QC pass, CpG site yield, coverage, extraction QC—not classifier BA alone.
Status honesty: live compare runs have been incomplete (e.g. BAMs without full H5 / extraction QC). Until a report shows both arms through extraction with operator-set thresholds, do not lock a biological winner. After a passing compare, bake the winner into procedure packs (buffy_wgbs_*_gene_fc, plasma WGBS procedures).
EM-Seq arm (option 4)¶
EM-Seq uses a separate SamplePrep path (sample_prep_emseq.program.json, useEmseqTargeted). Alignment and extraction are panel-constrained. The linear vs pangenome_wgbs harness does not choose EM-Seq vs WGBS: that is a clinical assay-design choice (gatekeeper depth vs genome-wide discovery), not an aligner bake-off.
3. Improving read and signal quality¶
SamplePrep QC gates eligibility; informME and cell deconvolution improve interpretability and confounding control in modeling.
Guardrails (eligibility)¶
- Alignment QC and extraction QC (
overall_pass, disposition) keep failed samples out of discovery and modeling. - cfDNA adds fragmentomics QC/features when
primary_analyte: cfdna. - EM-Seq emphasizes on-target depth / coverage against the locked panel BED.
informME (pipeline.info_measures)¶
Read-level patterns (*.patterns.h5) feed information-theoretic measures (NME, MML, ESI, MSI, and related confirmation metrics). Outputs land under info_measures/ and can stack as covariates. See methylpipeline_informme_integration.md and architecture covariates in end-to-end-workflow.md.
Physician-facing meaning: methylation heterogeneity and information content, not only mean β, so noisy or low-complexity samples are visible.
Cell-type deconvolution (Houseman / HiTIMED)¶
| Method | Typical use |
|---|---|
| Houseman | Buffy six-cell Ω fractions → ALR covariates |
| HiTIMED | Hierarchical / analyte-driven leaves (e.g. cfDNA tumor_fraction leaves when configured) |
Action: pipeline.cell_deconvolution → cell_fractions/cell_fractions.csv. Buffy research procedures enable Houseman by default; plasma WGBS / EM-Seq gatekeeper procedures often use the no-deconv lifecycle (fragmentomics / panel depth carry more of the noise story). Theory: ch.07a methyldeconv.
Stacking into the model¶
Default ECDF second-stage can stack:
- Cell fractions (ALR on Houseman columns when present)
derived_measures(genome / chromosome entropy, PMD/PDR-style proxies)info_measures/readlevel_measures.csv
First-stage ECDF uses methylation features; second-stage logistic stacks class probabilities + covariates. Ablation artifacts (feature_family_ablation.json) document whether covariates moved the needle.
Physician-facing meaning: adjust for immune-cell mix and methylation complexity so a “cancer” call is less likely to be composition noise.
4. Benefit to physician and insurance¶
Clinical scenario¶
Elevated PSA (and/or secondary serum markers) → decision to biopsy. A useful molecular gatekeeper:
- Prioritizes high NPV for clinically significant PCa (GG≥2 / Gleason ≥3+4)
- Separates indolent GG1 from actionable disease when biology allows
- Complements—not replaces—mpMRI, PHI, 4Kscore
Buffy WGBS research informs host biology; EM-Seq targeted plasma is the chemistry geometry aimed at the gatekeeper window; plasma WGBS is the discovery / breadth path. Requirements source: Prostate Cancer Detection.md. Fitness gaps (GG≥2 positive class, NPV LCB gates): Prostate_Cancer_Detection_MethylPipeline_Fitness.md.
Payer / medical-director lens¶
| Question | Platform artifact | Status today |
|---|---|---|
| Analytical validity (reproducible pipeline, locked config) | Release manifests, CAAS / .action_results, locked model spec |
Available as process evidence |
| Clinical validity (holds on patient-disjoint holdout) | locked_test, WF3, clinical performance report |
Required; not present on current EV-PCA packages |
| Clinical utility (biopsies avoided vs csPCa missed) | Intended-use + prevalence-conditioned NPV | Policy + pivotal study design; not a software brochure claim |
| Change control / post-market | SaMD ladder, PCCP / monitoring scaffolds | Scaffolded; study-specific |
Map evidence to analytical → clinical validity → utility. Do not claim coverage readiness from feasibility BA.
5. Detecting cancer versus noise¶
Translate the platform’s statistical honesty into plain language.
Stability Monte Carlo (WF1)¶
Repeated random train/validation splits discover DMPs/genes; recurrence across iterations filters features that appear only because of a lucky split. BA-gated FeatureCuts (dmp_fc, gene_fc, dual_fc) further require that panels meet held-out balanced accuracy inside discovery. Operator detail: usage ch.05; theory: ch.12.
A locus that survives a high fraction of random splits is more likely biology than noise—still not a clinical claim until holdout and pivotal stages pass.
Holdouts and partitions¶
| Role | Meaning |
|---|---|
development_train |
Feature selection and model training |
locked_test |
Internal validation; never returns to training |
pivotal_validation |
Claim-oriented external cohort |
WF2 evaluates a frozen model under Monte Carlo on the development world; WF3 is the true held-out batch. Current prostate EV packages: no partitions — treat metrics as engineering only.
Claim gates¶
Code blocks clinical-performance claims until regulatory.stage reaches pivotal (or later) and allow_clinical_performance_claims is permitted. SaMD ladder: samd_research → samd_holdout_enrichment → samd_pivotal (usage ch.18).
6. Detecting and confirming biomarkers¶
Pipeline story (disease-agnostic):
- Discover — centroid + detector → statistical / biological DMPs
- Stabilize — MC recurrence + optional DMP/gene FeatureCuts
- Freeze — fixed panel for production
- Map — DMPs → genes / features
- Confirm biology — enricher (
cancer-core/cancer-extended) and optional STRING PPI hubs - Readiness — freeze-readiness / biological completeness gates before modeling
Genes improve interpretability for physicians; enricher/PPI supply independent pathway support, not circular proof from the same BA. Literature candidates (GSTP1, APC, RASSF1A, PITX2, …) stay study/profile priors—never hardcoded disease logic in Python (fitness analysis).
EM-Seq gatekeeper path: the “biomarker set” is largely the locked capture panel refined by MC/stability within sequenced regions—not a genome-wide DMP hunt. Buffy/plasma WGBS paths use genome-wide discovery then freeze.
Honest limitation: some buffy packages show empty gene panels at threshold despite DMP stability—gene-axis readiness must be checked before biological claims.
7. Finding a good model and validating it¶
Model backends¶
| Backend | Role |
|---|---|
| ECDF (+ optional covariate second stage) | Transparent Bayesian-style methylation classifier; preferred regulated path today |
| tabular_sklearn | Feature-matrix alternatives; useful in model-MC comparisons |
| Generative / hybrid | Research backends when enabled |
Model-MC ranks backends on held-out metrics (e.g. median BA); select_best_model records the winner. Post-model validation consumes frozen artifacts without retraining (usage ch.07–08).
What “good” means for a gatekeeper vs current runs¶
| Metric world | Gatekeeper product | Current EV-PCA feasibility |
|---|---|---|
| Positive class | Clinically significant PCa (GG≥2) | Often healthy vs pooled PCa |
| Primary operating characteristic | High NPV (prevalence-aware) + sensitivity/specificity CIs | Balanced accuracy / stability panel size |
| Holdout | Locked + pivotal partitions | Missing |
| Acceptance | NPV LCB / operating-point policy (product gaps) | BA gates inside FeatureCuts |
Gleason-staged manifests are already supported (usage ch.16); defining GG≥2 as the screening positive class and NPV LCB acceptance gates are called out as small, disease-agnostic extensions in the fitness note—not as present production claims.
SaMD ladder as the validation program¶
| Stage | Profile | Allowed narrative |
|---|---|---|
| Research / feasibility | samd_research |
Engineering, panel stability, method development |
| Holdout enrichment | samd_holdout_enrichment |
Internal locked-test enrichment |
| Pivotal | samd_pivotal |
Clinical performance claims (when evidence filled) |
Operator SOP: usage ch.18. Evidence registry: validation-evidence-index.md.
8. Current prostate evidence inventory (honesty)¶
| Package ID | Study tree | Analyte | Status |
|---|---|---|---|
EV-PCA-PLASMA-2026-06 |
Plasma_healthy_vs_PCa |
cfDNA | Draft feasibility |
EV-PCA-HGOOD-2026-06 |
H_PCa_good |
Buffy | Draft feasibility; gene-stable-at-threshold gaps reported |
EV-PCA-BUFFY-2026-07 |
Buffy_healthy_vs_PCa |
Buffy | Draft; incomplete model chain for gatekeeper claims |
None of these packages alone supports physician biopsy-deferral policy or payer coverage decisions.
Appendix A — Procedure and config quick map¶
| Intent | Set |
|---|---|
| Buffy research WGBS | regulatory.primary_analyte: buffy_coat + procedure buffy_wgbs_pangenome_gene_fc (or buffy_wgbs_linear_gene_fc until compare locks) |
| Plasma WGBS discovery | primary_analyte: cfdna + cfdna_wgbs_plasma |
| EM-Seq gatekeeper research | primary_analyte: cfdna + cfdna_emseq_targeted + operator target_panel_bed |
| Claims | Only after partitions + pivotal stage + filled evidence package |
Canonical analyte encoding: ANALYTE_PROFILES.md. Product roadmap context: Regulatory-Ready Platform for Multiomics Diagnostics.md.
Appendix B — Further reading¶
- Stability / freeze / model / post-model: usage ch.05–08
- Two workflows (WF1–3): theory ch.12
- Model creation: theory ch.15
- SamplePrep linear vs WGBS compare: sample_prep_test_bed.md
- Interactive canvases: analyte-comparison, pca-detection-fitness