Low-pass WGS for MCED · Evidence, prototype, and roadmap

Executive decision

Proceed—but build a disciplined fragmentomics platform, not an “LLM cancer test.”

Low-pass WGS is attractive because one library exposes several partially complementary signals. The defensible product path is a modular, high-specificity binary detector with an explicit no-call state. Tissue of origin and foundation models come later.

Classical fragmentomic experts + late fusion

Coverage/CNA, length, motifs, chromatin and repeats get separate normalization, models and ablations before out-of-fold stacking. Here “fusion” means combining model scores, not detecting gene fusions.

Gate

Clinical claim waits for intended-use data

Retrospective AUCs and computational downsampling do not establish screening sensitivity, PPV, mortality benefit or diagnostic burden.

Boundary-aware MIL, not text LLMs

First compare hashed k-mers, a small CNN/Transformer, frozen DNABERT-2/Caduceus and LEAF-1 under identical fragment bags and compute.

Data

Open empirical work now; controlled validation next

Start with FORESEE’s published CRC partition, CRAG matched coordinates and donor-grouped FinaleMe ULP-WGS; then run a raw-read pilot and independent multisite challenges.

Empirical public-data reanalysisSynthetic software testPublished paper resultCalculation / scenario

What I actually downloaded and ran

Three real-data analyses, including a confounding failure.

This section is new. It reports this project’s own computations on public patient-derived data. It does not silently substitute paper numbers, and it does not call processed-matrix reanalysis a raw-WGS reproduction.

Empirical · processed matrix

FORESEE / PRJNA755688

Downloaded and checksum-verified the authors’ released 512-value regional matrix and metadata. Modeling used 701 published-training samples; 384 published-holdout samples were not used for fitting or thresholds. The code detects the validation file’s hidden comma delimiter and rejects sample overlap.

Empirical · fragment coordinates

CRAG / Zenodo 6914806

Downloaded a 3.308 GB liver archive, verified its MD5, reconciled all 16 members to the official supplement, validated every gzip stream to EOF, and parsed 434,814,486 coordinates into 389,462,309 MAPQ/length-qualified fragments.

Empirical · confounding audit failed

FinaleMe / Zenodo 7779198

Downloaded and MD5-verified 1.895 GB of approximately 0.1× ULP-WGS coordinates. All 250,308,232 rows were parsed; 221,792,015 coordinate records passed canonical-autosome, MAPQ and length filters. Duplicate state is unknown, so these are not claimed as unique molecules. Repeated samples were aggregated into 42 donor groups.

FORESEE CRC result

Same-study published holdout: class separation with confounding risk.

The input is the authors’ released within-sample-normalized regional vector. It is linked to ordinary input-control cfDNA WGS, but the release does not identify which documented short/long/ratio/total channel the single vector represents. This is our new linear-model reanalysis, not reproduction of the authors’ ensemble.

Empirical SVM AUROC0.844Conditional stratified sample bootstrap 0.804–0.881; 220 CRC + 164 noncancers

95% development target53.6% sens.Observed specificity 94.5%; 9 FP

99% development target30.9% sens.Observed specificity 98.8%; 2 FP

Restricted label-shuffle diagnostic0.4980 / 100 fixed-C shuffles reached observed AUROC; model selection not repeated

How the held-out score was produced: each linear branch used 10 outer development folds, 5-fold inner tuning over a fixed C grid, fold-local imputation/scaling and Platt calibration learned only from inner out-of-fold margins. The published holdout score is the mean probability from the 10 calibrated outer models; 95%/99% target thresholds come only from nested development out-of-fold probabilities. Bootstrap intervals resample this holdout cohort and do not include training, site or model-selection uncertainty.

Empirical model / control	Held-out AUROC	Interpretation
Linear SVM fragment vector	0.844	Primary linear comparator for this post-publication reanalysis
Ridge logistic fragment vector	0.846	Separate linear-model comparator; the tiny difference does not select a winner
Age + sex only	0.594	Above chance; residual demographic structure remains
Filename-prefix number only	0.624	Technical proxy, not authoritative batch metadata; measurable structure remains

Batch warning, not a footnote: the 20 training filename-prefix groups and 11 validation groups have zero overlap. Within-prefix case/control concordance remains 0.832, so class separation persists under this coarse prefix-stratified comparison, but a one-number filename-derived proxy AUROC of 0.624 means technical confounding is not excluded. This reanalysis was post hoc, not blinded or preregistered.

CRC subgroup	Cases	AUROC vs all noncancers	Sensitivity at 95% target	Sensitivity at 99% target
Stage I	45	0.807	48.9%	31.1%
Stage II	87	0.834	50.6%	20.7%
Stage III	61	0.851	55.7%	36.1%
Stage IV	27	0.924	66.7%	51.9%

CRAG coordinate result

Eight matched early liver-cancer pairs: an exploratory engineering result, not a clinical estimate.

The fixed three-scalar leave-one-pair-out model reached AUROC 0.812 (pair-bootstrap 0.656–0.984). The code-defined exploratory 5 Mb coverage-dispersion score alone reached 0.969; all eight cases exceeded their matched control, exact two-sided sign-test p=0.0078.

Coordinate-level coverage score0.969Only 8 case/control pairs; 1 AUC step = 0.0156

Three-feature LOPO model0.812Exact 2⁸ label-swap p=0.0391

Declared batch-only control0.500Pairs share library/sequencing batch

Usable-fragment-count control0.141Coverage-dispersion vs depth ρ=-0.009

Important analytical limit: this coordinate benchmark uses fixed autosomal 5 Mb bins and neutral placeholder GC/mappability values. It is not GC/mappability-corrected, blacklisted or segmented CNA analysis. The exploratory eight-pair coverage separation is therefore a direction to reproduce with locked reference assets—not a production biomarker estimate.

Real-fragment computational downsampling

For each CRAG sample, real author-derived fragments were independently Bernoulli-thinned after separating short, long and other per-bin counts. This preserves a paired fragment as one unit. Approximate depth assumes paired 2×100 bp reads and a 3.1 Gb genome. It tests read-count noise on real libraries, but it is still not independent library preparation.

Fragments retained	Median fragments	Approx. aligned depth	Median coverage-score AUROC	95% replicate range	Median within-sample score SD
5%	1.24 M	0.08×	0.938	0.906–0.969	0.00151
10%	2.48 M	0.16×	0.969	0.938–0.969	0.00124
25%	6.21 M	0.40×	0.969	0.953–0.969	0.00091
50%	12.42 M	0.80×	0.969	0.969–0.969	0.00070
100%	24.84 M	1.60×	0.969	0.969–0.969	0.00000

At roughly 0.08×, median AUROC was 0.938 with replicate range 0.906–0.969; measurement variation fell as fragments increased. That supports the narrow claim that more counts reduce Monte Carlo noise. It does not prove clinical accuracy, optimal depth or library-level repeatability.

FinaleMe confounding result

The fragment model looked excellent—and batch alone looked almost identical.

This is the most important negative result in the package. FinaleMe contains 30 advanced cancer donors (castration-resistant prostate or metastatic breast) and 12 healthy donors, no early-stage cases or benign-disease controls. Longitudinal repeats were median-aggregated by donor before the primary fixed-model leave-one-donor-out calculation.

Post-hoc three-feature donor LODO AUROC0.969Conditional fixed-OOF-score bootstrap 0.917–1.000

Batch-incidence-only donor LODO AUROC0.964Nearly equals the fragment model; outcome is strongly encoded by batch

Usable-record-count-only donor LODO AUROC0.750Another technical proxy, not biology

Mixed-class batches2 / 2321 batches contain only one outcome class

Inference fails here: the full donor-grouped AUROC of 0.969 is not an interpretable cancer-performance estimate because batch-only AUROC is 0.964. Restricting to batches 8 and 10 leaves only 7 donors (3 cancer, 4 healthy); its AUROC of 1.000 and tiny cross-batch transfers are too underpowered to rescue the inference. LODO pools margins from separately fitted models without formal cross-fold calibration. This cohort is useful for exposing leakage/confounding, not for claiming early detection or generalization.

Exactly what was not empirically tested

Motif/chromatin/repeat AUROCs: synthetic only in this project. CRAG BED coordinates contain no bases, and the FORESEE vector’s exact released channel is not explicit.

LLM/foundation-model AUROC: not tested on patient data. The “genome-token” result below is an ordinal encoder of simulated coverage/length values—not DNA sequence, a Transformer, an LLM or a foundation model.

OOF fusion: means out-of-fold model-score stacking. It has nothing to do with gene fusions, and no cancer is assumed to carry a fusion.

Synthetic software stress test

A transparent simulation for exercising code—not patient performance.

All numbers in this section come from 720 feature-level simulated samples—not simulated FASTQ, BAM or WGS reads. The generator uses Poisson regional coverage with constructed site/GC/CNA-like templates, binomial regional short counts, multinomial 16-state motif counts, and Gaussian latent chromatin/repeat programs. The benchmark exercises donor/site quarantine, nested stacking, thresholds, QC/no-calls and negative controls; its biological signals were deliberately constructed and therefore cannot validate a cancer assay.

Synthetic — not patient data

Synthetic held-site AUROC0.940One predefined pair of simulated sites; individual bootstrap 0.906–0.971

Synthetic sensitivity at 99% target threshold57.6%Observed held-site specificity was 97.3%, not 99%; exact sensitivity CI 47.2%–67.5%

Synthetic observed specificity97.3%71 TN / 73 controls; 95% CI 90.5%–99.7%

Synthetic no-call rate4.4%Unconditional sensitivity 55.3%

Synthetic-only horizontal bar charts comparing held-site AUROC and sensitivity for nine simulated model families. — **SYNTHETIC — NOT PATIENT DATA.** Late fusion now uses fully nested expert and meta-model cross-fitting. AUROC and strict-threshold sensitivity tell different stories: early concatenation reached AUROC 0.935 but only 3.0% sensitivity at its development threshold.

Synthetic model	AUROC	Standardized partial AUROC, max FPR 1% (chance 0.5)	Sensitivity	Observed specificity	Brier
Simulated binned coverage (CNA-like latent signal; no segmentation)	0.660	0.548	0.0%	100.0%	0.492
Simulated regional fragment length	0.903	0.787	44.4%	100.0%	0.131
Simulated end motifs	0.908	0.665	10.1%	100.0%	0.221
Simulated latent chromatin view	0.699	0.528	8.1%	98.6%	0.221
Simulated latent repeat view	0.648	0.523	4.0%	100.0%	0.243
Early concatenation + logistic	0.935	0.756	3.0%	100.0%	0.230
Early concatenation + boosted trees	0.850	0.711	8.1%	100.0%	0.184
QC-only negative control	0.473	0.497	0.0%	100.0%	0.256
Site-only negative control	0.500	0.500	0.0%	100.0%	0.252
Quantized coverage/length ordinal-feature proxy	0.822	0.558	11.1%	100.0%	0.178
Fully nested out-of-fold late fusion	0.940	0.761	57.6%	97.3%	0.102

Synthetic negative controls behaved as expected in this single seeded experiment. QC-only AUROC was 0.473; a single permuted-label fit was 0.488. Site-only AUROC of 0.50 on unseen sites is nearly structural because those one-hot columns are inactive during training; it does not prove site confounding was removed. In a random patient split, site identity alone reached 0.504.

Synthetic count/noise stress

A constructed depth sweep, not a clinical depth experiment.

Five nominal depths were simulated with three seeded site-held-out replicates per depth using only coverage, length and motif experts. This is a reduced three-view model—not the five-expert main model, resequenced specimens or independently prepared libraries.

Synthetic-only line chart of AUROC, sensitivity and observed specificity over nominal simulated depth. — **SYNTHETIC — NOT PATIENT DATA.** Sensitivity uses a development threshold targeting 99% specificity; the plotted observed held-site specificity is allowed to differ. Do not interpret this as sensitivity fixed at 99% specificity or as an optimal depth.

Synthetic-only stage construction check. — **SYNTHETIC — NOT PATIENT DATA.** Stage signal was built in using median tumor fractions 0.15%, 0.35%, 0.9% and 2.8%, with signal strength scaling with tumor fraction. The rising curve is an expected construction check, not evidence of stage sensitivity. Issued n/sensitivity: I 37/21.6%, II 35/65.7%, III 21/95.2%, IV 6/100.0%.

Reference architecture

Keep every layer separate enough to diagnose.

The central design principle is separation: specimen physics, unique molecules, fragment measurements, classifiers, operating threshold and clinical truth must never collapse into one “depth-to-sensitivity” formula.

Literature synthesis / roadmap This reference architecture is a recommended analytical design, not a locally validated clinical pipeline.

Specimen

Plasma volume, input cfDNA mass, tube, processing delay, freeze-thaw, hemolysis and WBC contamination.

Library + reads

Recovery, conversion, duplicate saturation, pair completeness, MAPQ, insert distortion and actual usable unique depth.

Feature experts

Coverage/CNA, regional length, motifs, chromatin, repeats and optional sequence embeddings.

OOF calibration

Every transformation and expert fits inside donor/site-aware folds; only held-fold scores train fusion.

Decision

Locked high-specificity threshold, calibrated probability, technical no-call and biological abstention.

Clinical endpoints

Binary detection first; tissue origin conditional on true positives; diagnostic resolution and utility separately.

Method evaluation

Combine complementary failure modes—not duplicate views of the same fragments.

Most modalities are correlated because they arise from the same endpoints, lengths and regional counts. Calibrated late fusion and explicit ablation are safer than multiplying “independent” likelihoods.

Literature synthesis / roadmap The cards below evaluate published method families and propose gates. Motif, chromatin, repeat, inferred-methylation and CNA segmentation branches were not benchmarked on patient data in this project.

Regional length + coverage

Core now

The most reproducible lpWGS baseline and the center of DELFI-like methods.

Strength: Genome-wide aggregation works at roughly 0.5–3×.
Failure: GC, mappability, library chemistry, inflammation and cohort batch.
Decision: Reproduce first; lock binning/build/GC policy.

CNA / ichorCNA

Core now

A direct tumor signal and useful tumor-fraction/QC expert.

Strength: Orthogonal to nuclease/chromatin effects.
Failure: CNA-quiet, copy-neutral, focal and very-low-TF cancers.
Decision: Use as an expert, never universal ground truth.

Global + regional size

Core now

Strong, compact and easy to cross-check independently.

Strength: Robust biological direction and inexpensive extraction.
Failure: Nonspecific host biology and pre-analytics.
Decision: Include global histograms plus regional short/long ratios.

End / breakpoint motifs

Controlled

Compact nuclease and cleavage-context signatures.

Strength: Often strong internal discrimination.
Failure: End repair, chemistry, reference and sequencer sensitivity.
Decision: Add only after cross-kit concordance and FrEIA cross-check.

WPS / TFBS / chromatin

Research

Aggregates host and tissue regulatory programs.

Strength: Potential tissue biology and cross-cancer signal.
Failure: Atlas mismatch, immune/inflammatory signal and site batch.
Decision: Griffin/LIONHEART expert with leave-dataset-out testing.

Repeats + inferred methylation

Exploratory

Potentially orthogonal landscape and tissue-origin information.

Strength: Adds genomic contexts not captured by broad bins.
Failure: Ancestry/build dependence and sparse learned priors.
Decision: Ablate separately; defer product dependence.

Large language models and foundation models

The useful idea is fragment representation—not natural language.

A low-pass WGS sample is an unordered bag of sparse fragments. The architecture must respect fragment boundaries, reverse complements and permutation invariance. General DNA models learn reference grammar; they do not automatically learn plasma cancer biology.

Literature synthesis / proposed experiments No sequence encoder, Transformer or foundation model was run on patient data here.

Recommended AI path

Boundary-aware encoder → MIL → late fusion.

Represent each fragment with left flank, cut token, insert/read sequence, right cut, right flank, length, GC, strand, MAPQ and mate state. Aggregate 10k/50k/100k fragment bags with DeepSets or gated-attention MIL.

Natural-language LLM: literature, QC narrative and structured metadata only.
Generic DNA FM: frozen benchmark after a scratch baseline.
cfDNA FM: pretrain in-house only after external incremental value is proven.

Synthetic only The implemented “genome-token” score is an ordinal encoding of simulated coverage/length features. It is neither raw DNA sequence nor a Transformer, LLM or foundation model. Its AUROC was 0.822, versus 0.903 for simulated regional length and 0.940 for fully nested simulated late fusion. No patient-data AI performance was measured.

Hashed k-mer logisticSequence-content floor with exact reproducibility.

Must establish first

Small CNN / Transformer from scratchSame fragment bags, donor splits and compute accounting.

Beat k-mers externally

Frozen DNABERT-2 + CaduceusCompare generic sequence pretraining and reverse-complement symmetry.

Beat scratch fairly

LEAF-1 research benchmarkCleavage-aware cfDNA/ATAC embeddings with MIL.

License + overlap audit

In-house cfDNA pretrainingMasked bases, cut recovery, reverse-complement consistency and contrastive objectives.

Only after external gain

Out-of-fold multimodal fusionAdd the sequence score to classical experts; never train meta-model on in-sample scores.

Gain at ≤1% FPR

Do not start with: LLaMA over serialized motif numbers, one artificial sentence made from arbitrary read order, billion-parameter Evo fine-tuning, methylation models on ordinary WGS, cloud submission of patient reads, or joint detection/tissue-origin training.

Data and software

Public data can support real reanalysis—within strict intended-use limits.

This package now includes a same-study held-out reanalysis of 1,085 FORESEE CRC/control samples, exact coordinate processing of 16 matched CRAG liver samples, and a verified 77-file FinaleMe archive with donor IDs retained. Only those three resources were downloaded here; other cards are evaluated accessions and future-use boundaries. None is a prospective, asymptomatic, multisite MCED cohort. The next raw-read pilot is specified in a downloadable 16-run FORESEE manifest; the roughly 59.4 GiB SRA-Lite figure is a planning estimate, not a completed per-run transfer audit.

Open processed data, CC BY 4.0; raw SRA publicPRJNA755688 / SRP332946 / Zenodo 5170265

FORESEE CRC input-WGS fragmentomics

Released DELFI-style matrices: training 405 CRC plus 296 noncancers; validation 220 CRC plus 164 noncancers, including small rheumatoid arthritis, peptic-ulcer and COPD subsets

Assay: Ordinary paired-end plasma cfDNA input-control WGS; this reanalysis uses the released 512 regional within-sample normalized values
Burden: Processed matrices about 10.7 MB; 1,392 raw input-WGS SRA runs are roughly 5.76 TB
Best intended use: Immediate post-publication training/published-holdout CRC benchmark with stage and benign-control audits
Do not use for: Independent-site, prospective, raw-pipeline or multi-cancer validation

Open processed coordinates, CC BY 4.0; raw data controlledZenodo 6914806; raw study phs003062.v1.p1

CRAG open fragment-coordinate cohort

8 stage I-II HCC plus 8 matched controls; 25 stage I-III breast cancers plus 25 matched controls

Assay: hg19 fragment-coordinate files derived from plasma WGS
Burden: 66 subjects; approximately 13.4 GiB in five archives
Best intended use: Immediate length, coverage, endpoint, hotspot and WPS software benchmark
Do not use for: Clinical specificity or pan-cancer validation

Open processed coordinates, CC BY 4.0Zenodo 7779198

FinaleMe ULP-WGS prostate/breast cohort

77 records but only 42 donor groups: 43 CRPC prostate, 22 metastatic breast and 12 healthy records

Assay: Approximately 0.1x b37 fragment-coordinate files
Burden: Verified 1.895 GB archive; 77 gzip members
Best intended use: Donor-grouped ultra-low-pass feature and domain-shift stress test with explicit batch baselines
Do not use for: Early detection; repeated donor records and healthy-batch confounding prohibit ordinary sample splits

Open under source-specific termsIGSR / 1000 Genomes Project

1000 Genomes germline reference data

Population germline genomes from cellular DNA, not matched plasma cfDNA cancer/control specimens

Assay: Whole-genome germline sequencing across populations and technologies
Burden: Large, but reference panels and selected files can be staged selectively
Best intended use: Population allele frequencies, ancestry checks, mappability/CNA technical controls and germline masking
Do not use for: Primary cancer-versus-normal cfDNA training or biological sensitivity estimates

Public INSDC records; cite original study and retain source termsGSE71378 / SRP061633

Snyder plasma nucleosome-footprint cohort

60 paired-end runs: 48 cancers, 8 inflammatory noncancers, and four healthy-labelled records; one healthy record is pooled plasma

Assay: Public paired-end plasma WGS with very heterogeneous depth
Burden: Full study about 601 GB compressed; recommended six-run plumbing subset about 8.5 GB
Best intended use: Raw alignment, duplicate, insert-size, WPS and downsampling invariance tests
Do not use for: Training an MCED classifier or estimating specificity

Controlled EGA DAC/DUA; no redistributionEGAS00001003611 / EGAD00001005339

DELFI discovery cohort

Classifier cohort 208 untreated cancers across seven types and 215 healthy controls

Assay: 1-2x, 100-bp paired-end cfDNA WGS
Burden: About 4.4 TB; EGA metadata show a 537/538-sample discrepancy that must be reconciled
Best intended use: Reproduce the published regional fragmentation baseline and develop models inside grouped folds
Do not use for: Final independent or prospective validation

Controlled EGA DAC/DUA; no redistributionEGAS00001005340 / EGAD00001007796

LUCAS lung and benign-control cohort

129 lung cancers and 158 noncancer development samples; external 46 cancers and 385 controls

Assay: 1-2x, 100-bp paired-end cfDNA WGS
Burden: 872 BAMs, approximately 4.5 TB
Best intended use: Locked lung-cancer and benign-condition external challenge
Do not use for: Pan-cancer generalization by itself

Controlled; restrictive ovarian data-use policyEGAD00001004939 / EGAD00001006132

Mouliere fragment-size cohorts

344 plasma samples from 200 patients across 18 cancers plus 65 healthy individuals

Assay: Low-pass WGS around 0.4x
Burden: Approximately 933 GB across the two controlled datasets
Best intended use: Fragment-size and CNA method reproduction
Do not use for: Early-stage MCED sensitivity because disease is mainly advanced

Research availability does not provide one uniform commercial license; source terms controlFinaleDB

FinaleDB processed fragmentome collection

Uniformly processed coordinates and tracks for 2,579 datasets, 2,505 samples and 23 conditions

Assay: Derived fragment coordinates and feature tracks from many source studies
Burden: Feature-level downloads can be small; source studies vary
Best intended use: Rapid feature prototyping and hard-negative exploration
Do not use for: Redistribution or product training without source-by-source permission

Controlled, HDAC003271HRA005926 / PRJCA020703

Prospective early gastric cancer WGS

Development 110 stage I-II gastric cancers and 139 noncancers; validation 73 cancers and 94 noncancers

Assay: cfDNA WGS with size, CNA, nucleosome and SNV features
Burden: Not exposed before authorization
Best intended use: Locked early-stage single-cancer external challenge
Do not use for: Pan-cancer specificity

Public SRAPRJNA929650

SPOT-MAS public bisulfite cohort

239 breast cancers and 278 healthy controls

Assay: Paired-end bisulfite sequencing
Burden: 517 SRA experiments; about 1.26 TB
Best intended use: Adjacent methylation-plus-fragment modeling experiments
Do not use for: Direct validation of ordinary low-pass WGS because bisulfite conversion changes sequence and fragment distributions

Controlled dbGaPphs001417.v1.p1

ichorCNA metastatic ULP-WGS cohort

Longitudinal metastatic breast/prostate cancers and controls

Assay: Ultra-low-pass WGS around 0.1x
Burden: Controlled dbGaP
Best intended use: CNA and tumor-fraction algorithm validation
Do not use for: Interpreting the reported approximately 3% favorable-CNA threshold as early-cancer clinical LoD

Evaluated software stack

Tool	Role	License / IP	Decision
samtools + bwa-mem2 + mosdepth	Pair-preserving alignment, fixmate/sort/markdup/index, QC and actual usable-depth measurement	Mixed open-source licenses; review each distribution	Production plumbing baseline
FinaleToolkit 1.0.0	Length, coverage, WPS, DELFI, end/breakpoint motifs, MDS and cleavage profiles from BAM/CRAM/fragments	MIT	Primary independent feature extractor to benchmark this prototype against
ichorCNA	ULP-WGS copy-number segmentation and tumor-fraction estimate with panel of normals	GPL-3.0	Orthogonal CNA expert and QC feature; never use its estimate as early-cancer ground truth
Griffin	Length-specific GC correction and aggregated TFBS/chromatin profiles	Clear BSD source text; method paper discloses patent application	Primary chromatin expert after separate IP review
FrEIA	Independent end-motif, length, CNA and tumor-fraction features	MIT	Cross-implementation unit test for endpoint features
cfDNAPro	Independent Bioconductor insert-size/fragment distribution QC	GPL-3	Orthogonal size-distribution check
cfdnakit	Length-aware binning, panel of normals, CES and CNV	Bioconductor package terms	Sensitivity-analysis comparator to ichorCNA
delfi3 / reproduce_lucas_wflow	Reference reproduction of published regional fragmentation pipelines	Repository-specific; verify before reuse	Use to reproduce original baselines before adding corrections
LIONHEART	Cross-dataset open-chromatin correlation expert	Code open; shared models/resources explicitly research/noncommercial	Strong external comparator, not an unrestricted product dependency
DNABERT-2	Frozen generic DNA fragment embedding benchmark	Apache-2.0	First foundation-model comparator only after classical baseline lock
Caduceus	Reverse-complement-aware frozen DNA fragment embedding benchmark	Apache-2.0	Second generic foundation-model comparator
LEAF-1	Cleavage-boundary-aware fragment foundation model plus MIL	PolyForm Noncommercial / preprint resource terms	Best research benchmark after license acceptance; not MVP dependency

Reference build, blacklist, mappability, annotation, duplicate policy, read group, library kit, tube, site, lane and software version must be recorded in every sample manifest. hg19→hg38 remapping can change fragment endpoints, motifs, repeats and regulatory annotations.

Complete go-forward roadmap

Advance through evidence gates, not calendar promises.

Each phase ends with a falsifiable decision. Failure to transfer across sites or improve high-specificity partial AUC is a reason to simplify, redesign or stop—not to add a larger model.

Completed locally · reproduce

Phase 0 — Freeze the analytical contract

Keep the checked manifests, exact source hashes, canonical-autosome policy, MAPQ/length filters, pair-aware extraction, grouped folds, QC/no-call rules and output schema under version control. Cross-check future base-aware motif extraction with FinaleToolkit/FrEIA before claiming agreement.

GateCurrent coordinate extraction and seeded analyses are deterministic and tested. Reference GC/mappability correction and base-aware endpoints remain unfinished.

Now · open-data baseline

Phase 1 — Finish the public-data ladder

Use FORESEE’s published 701/384 processed-matrix partition as the empirical CRC baseline and FinaleMe’s failed batch-confounding audit as a mandatory-control template. Next process the specified 16-run FORESEE raw pilot spanning stages I–IV and selected noncancer controls. Preserve read pairs and deterministically thin pairs for depth studies.

GateRaw-to-feature concordance, no donor leakage, batch-only baselines, stable direction under repeated thinning, and no claim beyond each cohort’s cancer type and stage distribution.

Parallel · controlled access

Phase 2 — Multisite development, grouped from day one

Obtain independent cohorts plus authoritative site, kit, plate, lane and donor metadata. Reproduce a locked regional baseline first; add CNA, motifs, Griffin/LIONHEART, repeats and sequence branches one at a time. Use nested donor/site-aware folds and out-of-fold score fusion.

GateIncremental leave-site-out standardized partial AUROC and sensitivity at prespecified 99–99.5% specificity, with calibration, no-call and subgroup stability.

After lock · external

Phase 3 — Frozen LUCAS and gastric challenges

Freeze code, normalizers, feature selection, model, calibration, threshold and no-call before opening outcomes. No site adaptation. Report cancer-type external challenges, not pan-cancer validation.

GatePredeclared high-specificity performance, benign-disease specificity and confidence intervals; investigate every domain-shift failure.

6–18 months · analytical

Phase 4 — Real library depth and pre-analytic validation

Independently prepare 0.1/0.25/0.5/1/2/3× libraries across plasma volume, cfDNA input, recovery, duplicate saturation, tubes, delay, freeze-thaw, hemolysis, gDNA, lots, operators and instruments. Use ≥20 replicates near each decision boundary.

GateLocked LoB/LoD/LoQ-like analytical characteristics, repeatability, reproducibility, failure rate and selected depth/cost Pareto point.

18–36+ months · clinical

Phase 5 — Prospective intended-use MCED study

Blinded multisite asymptomatic cohort with age/sex/smoking matching, benign lesions, infection/inflammation, autoimmune disease, COPD, cirrhosis/HBV, renal disease, prior cancer and 12–24 month follow-up. Binary detection first; tissue origin evaluated among true cancer positives.

GateClinical sensitivity by type/stage, specificity, calibration, PPV, diagnostic resolution, harms, interval cancers and utility—not AUROC alone.

Validation and screening arithmetic

Specificity is a sample-size problem before it is a model problem.

With zero observed false positives, 598 controls are needed merely to put the one-sided 95% upper FPR near 0.5%, and 2,995 to put it near 0.1%. Real studies need more because false positives will occur and subgroups/sites must be estimated.

Controls, 0 FP → upper FPR 0.5%598Exact binomial inversion

Controls, 0 FP → upper FPR 0.1%2,995Exact binomial inversion

Synthetic plug-in PPV at 0.5%9.6%Uses unstable observed specificity

Synthetic conservative PPV corner2.4%Separate marginal bounds; sensitivity analysis, not a joint 95% CI

Report sensitivity at 98%, 99% and 99.5% specificity; partial AUC in that region; exact confidence intervals; calibration; benign-condition and demographic subgroups; PPV at intended prevalence; failure/abstention; downstream diagnostic yield; and unconditional results counting no-calls. Overall AUROC is secondary.

Planning economics

Cost per reportable result, with every assumption exposed.

This calculator is an editable scenario, not a vendor quote. Defaults are placeholders. Obtain local quotes and record access dates before a purchasing decision. One-time development/validation and capital are not included in the per-result number.

Nominal depth

Usable aligned fraction

Plasma + extraction / sample

USD

Library / sample

USD

Sequencing / raw Gb

USD

Compute + storage / sample

USD

Labor + routine QC / sample

USD

Controls / batch

USD

Patient fill / batch

samples

Operating overhead

Technical issuance

Evidence atlas

38 primary studies and official records, with the caveat beside the claim.

The table prioritizes fragmentomics and WGS while including the sequence-model and clinical-evidence context needed to make a defensible roadmap. Search by author, method, cohort or limitation.

Year	Published source	Method + cohort	Published finding + critical interpretation
2026	PublishedA pan-cancer compendium of 1,294 plasma cell-free DNA methylomes and fragmentomes enabling multicancer detectionPan-cancer compendium	cfMeDIP-seq methylome plus fragmentomics1,294 plasma methylomes/fragmentomes	Demonstrated cancer-type heterogeneity across length, motif, nucleosome and methylation features.Useful multimodal evidence but immunoprecipitation enriches a different molecular view than ordinary lpWGS.
2026	PublishedEnhanced multicancer screening assay through whole-genome methylation sequencing-based multimodal cell-free DNA analysisMultimodal retrospective study	Methylation, CNV, fragment-size ratio and distribution ensembleWhole-genome methylation sequencing	Reported sensitivity 93.2% at 95% specificity with strong early-stage results.Assay chemistry differs from ordinary lpWGS and 95% specificity is insufficient for many low-prevalence screening settings.
2026	PublishedFragmentation-induced coverage biases in cell-free DNA sequencing affect the clinical sensitivity of liquid biopsiesMechanistic methods paper	Coverage-bias analysiscfDNA WGS	Shows that epigenetic-context-dependent fragmentation changes locus coverage and effective sensitivity.Coverage cannot be treated as uniform Poisson sampling; fragment context and assay preparation alter callability.
2026	PublishedFragmentomic liquid biopsy enables early breast cancer detection, molecular subtyping and lymph node assessmentMulticenter breast-specific case-control study	TuFEst repeats, motifs, size and TF coverage503 breast cancers, 289 benign controls; 77.1% stage 0-IIa; 2x from about 1 mL plasma	In the labeled external cohort (95 malignant, 61 benign), AUC was 0.968 (95% CI 0.941-0.995), with 91/95 sensitivity and 53/61 specificity at the reported confusion-matrix point.Breast-specific, not MCED; libraries were 2x and the 1x/0.5x/0.1x experiments were computational downsamplings. The abstract's 95%/78.3% is not labeled as that external operating point.
2026	PublishedNCI Cancer Screening Overview: Multi-Cancer DetectionUS National Cancer Institute	Regulatory and clinical-evidence contextNot applicable	As of access, no MCD assay is FDA-authorized and no published randomized trial has shown mortality reduction.A research classifier must not be marketed or interpreted as a screening test; clinical utility and downstream harms remain separate endpoints.
2026	PublishedSemantic fragment representations for coordinate-free analysis of genomics dataPreprint	LEAF-1 boundary-aware 90.4M fragment foundation model plus MILPretrained on about 58 billion ATAC/cfDNA fragments; pan-cancer and held-cancer experiments	Reported pan-cancer AUC around 0.95 and frozen clear-cell renal cancer AUC 0.83.Best research benchmark, but July 2026 preprint, possible cohort overlap, small held-cancer test, noncommercial terms and approximately 2 GB embeddings per sample by default.
2026	PublishedToward generalizable prediction of cancer signal using a cell-free DNA language modelPeer-reviewed multi-cohort research	Fragmentia-AI fragment encoder and sample-level learningPanels and ultra-low-pass WGS at approximately 0.1-1% of conventional sequencing input	Reported cancer-associated fragment-level sequence patterns across assay contexts.Directly relevant architecture but proprietary/restricted resources and lack of independent replication limit reproducibility.
2025	PublishedA deep-learning model for quantifying circulating tumour DNA from the density distribution of DNA-fragment lengthsCross-cohort method validation	Fragle fragment-length density modelMultiple low-pass WGS cancer and control cohorts	Reported improved tumor-fraction quantification and longitudinal/MRD utility.Best treated as an auxiliary tumor-fraction/QC expert rather than a standalone MCED classifier.
2025	PublishedBenchmarking DNA foundation models for genomic sequence predictionCross-model benchmark	Independent DNA foundation-model comparisonGeneral genomic tasks	Found no uniformly superior DNA foundation model across tasks.Requires task-specific scratch baselines and fair compute/data budgets before adopting pretraining.
2025	PublishedCross-dataset pan-cancer detection by correlating cell-free DNA fragment coverage with open chromatin sites across cell typesLeave-one-dataset-out multi-cancer benchmark	LIONHEART cross-dataset chromatin correlationsNine datasets, 14 cancers, 1,449 cancers and 1,106 controls; 63.3% at 1-3x	Mean cross-dataset AUC 0.826; mean sensitivity 49.5% at 95% specificity.More credible generalization blueprint and materially lower than typical within-study performance.
2025	PublishedEarly detection of multiple cancer types using multidimensional cell-free DNA fragmentomicsLarge case-control plus preliminary prospective asymptomatic study	Mercury integrated WGS genetic/fragmentomic modelInternal 3,021 cancers/3,370 controls; independent 677/687; prospective asymptomatic 3,724	Independent sensitivity 87.4%, specificity 97.8%, tissue-origin accuracy 82.4%; prospective sensitivity 53.5%, specificity 98.1%.The large case-control-to-screening sensitivity drop is the clearest warning against quoting discovery performance as clinical performance.
2025	PublishedMethylBERT learns read-level DNA methylation patterns for cancer detectionAssay-specific foundation model	Sequence-plus-methylation TransformerBisulfite methylation reads	Demonstrated read-level representation and tumor-fraction deconvolution for methylation assays.Not applicable to ordinary WGS without methylation-preserving chemistry.
2024	PublishedA blood-based test for lung cancer screening using cell-free DNA fragmentomesProspective multicenter screening-eligible validation	DELFI-L101 lung-screen triage958 participants from 47 centers; held-out 248 cancers and 134 controls; about 3x	Validation sensitivity 84%, specificity 53%, and stage-I sensitivity 71%.Useful LDCT-triage evidence, not an MCED-grade high-specificity screen.
2024	PublishedCaduceus: Bi-directional equivariant long-range DNA sequence modelingGeneral genome representation model	Reverse-complement-aware DNA state-space modelPretrained genomic sequence	Biologically attractive reverse-complement symmetry and Apache-2.0 code make it a second frozen-encoder comparator.It still lacks cfDNA cleavage-boundary pretraining and sample-level validation.
2024	PublishedDNABERT-2: Efficient foundation model and benchmark for multi-species genomeGeneral genome representation model	117M-parameter DNA foundation modelPretrained reference sequences, not cfDNA samples	Provides an Apache-2.0 frozen encoder suitable for a controlled fragment-embedding benchmark.Reference-sequence grammar is not plasma-fragmentation biology; it must beat k-mer/CNN baselines under identical bags and budgets.
2024	PublishedEnd-motif transformer for cancer detection from cell-free DNAAssay-mixed retrospective research	Compact motif Transformer4,606 assay-mixed samples; WGS HCC subset about 129	Reported motif AUC improvement from approximately 0.826 to 0.895.Assay mixing and pretraining/evaluation overlap create transductive leakage concerns.
2024	PublishedFinaleMe: inferring regional methylation from cell-free DNA fragmentationExploratory method	HMM-derived methylation/tissue probabilities without bisulfite sequencing80 paired WGS/WGBS samples plus ULP breast/prostate data; evaluated to 0.1x	Inferred regional methylation probabilities from fragment length, coverage and CpG context.Sparse and prior-dependent at low depth; use only as an exploratory tissue-origin expert.
2024	PublishedGenome-wide repeat landscapes in cancer and cell-free DNALarge multi-assay method study	ARTEMIS repeat-element fragmentation2,837 tissue/plasma samples from 1,975 people; cfDNA generally 1-2x	LUCAS ARTEMIS AUC 0.82 and ARTEMIS plus DELFI 0.91.Repeat features can be orthogonal but are sensitive to ancestry, reference build and reused cohorts.
2024	PublishedInterpretable large language model for cancer detection from cell-free DNA fragmentomicsAssay-mixed retrospective research	LLaMA-7B over serialized end-motif ranks2,451 samples overall; WGS subset 129; external WGS 44 cancers and 30 controls	Reported external AUC 0.912 with specificity 0.667 at its high-sensitivity point.Overcapacity and weak simple-model comparisons provide no persuasive reason to use a natural-language LLM for a compact numeric vector.
2024	PublishedLanguage model enables end-to-end accurate detection of cancer from cell-free DNAAssay-mixed retrospective research	ACID causal sequence model over concatenated reads2,094 samples across WGS and bisulfite assays; small WGS evaluation	Reported WGS AUROC 0.924 and high accuracy using 10,000 reads.Alphabetical read concatenation is arbitrary and not permutation-invariant; reproduce only as a research comparator.
2024	PublishedNucleotide Transformer: building and evaluating robust foundation models for human genomicsGeneral genome foundation model	Large DNA TransformerReference genomes, not plasma WGS	Provides strong general sequence embeddings across genomics tasks.Large compute and noncommercial/share-alike weight terms make it a research comparator, not the MVP.
2024	PublishedSystematic comparison of cell-free DNA fragmentomic features for cancer detectionCross-dataset benchmarking study	Ten-pattern, cross-source comparisonFour data sources and eight cancers	End motifs were strong internally while length and coverage transferred more consistently.Dataset origin substantially changed performance, arguing for site-held-out evaluation and late fusion.
2024	PublishedTransformer-based representation learning and multiple-instance learning for cancer diagnosis from bisulfite-treated plasma cfDNALarge targeted-bisulfite research cohort	Read-level Transformer plus MIL5,389 samples across colorectal, liver, lung and noncancer classes	CRC cross-validation AUROC 0.980 and external HCC AUROC 0.910 were reported.Architecture informs a future methylation branch but ordinary lpWGS does not contain base methylation states.
2023	PublishedDetecting liver cancer using cell-free DNA fragmentomesMulticohort retrospective validation	DELFI liver classifier724 total participants in US/EU and Hong Kong cohorts, about 2.6x	The study reported 88% sensitivity at 98% specificity versus average-risk controls and 85% at 80% specificity among high-risk controls; an independent Hong Kong cohort was also included.Those two operating points summarize study/development contexts rather than both being independent external estimates; the change against high-risk controls shows why disease-context controls are essential.
2023	PublishedMultidimensional analysis of cell-free DNA fragmentomics for lung cancer detectionMulticohort retrospective study	Multimodal score fusion of CNA, size, end and breakpoint-motif features (not gene-fusion detection)432 lung cancers and 350 noncancers; original mean 11.08x, development around 5x, computationally downsampled to 0.5x	Reported cohort AUCs 0.974-0.987 and stage-I sensitivity 83.2%.The 0.5x claim relies on in-silico downsampling; independently prepared depth, LoD and repeatability were not established.
2022	PublishedA framework for clinical cancer subtyping from nucleosome profiling of cell-free DNACross-cohort reanalysis method	Griffin length-specific GC correction and TFBS profilesDELFI/LUCAS and other WGS cohorts; tested down to 0.1x	DELFI AUC 0.94 and fixed external LUCAS AUC 0.86 in reported analyses.Explicit DELFI-LUCAS batch effects prevented direct transfer and performance worsened at 0.1x.
2022	PublishedCRAG: de novo characterization of cell-free DNA fragmentation hotspots in plasma whole-genome sequencingSmall early-stage multi-cancer proof of concept	Endpoint fragmentation hotspotsOpen processed coordinates for 66 matched HCC/breast cancer and control subjects	Identified open-chromatin/transposon-associated hotspots with cancer and tissue-origin signal.Small matched cohorts can validate software directionality, not population specificity.
2022	PublishedGenome-wide mutational signatures in low-coverage whole genome sequencing of cell-free DNASmall internal repeated nested-CV classifier plus larger signature analyses	Singleton mutational spectra (Pointy) with optional ichorCNAClassifier cohort only 16 stage-IV CRC and 19 controls; larger DELFI reanalysis was used mainly for signature detection	Repeated nested CV in the 35-subject classifier cohort gave median AUC 0.96 across models, random forest 0.98 and random forest plus ichorCNA 1.00.These are not external results; the tiny advanced-disease cohort and sequencing-error, germline, age, smoking and platform sensitivity make this a high-control exploratory branch.
2022	PublishedHydroxymethylation profile of cell free DNA is a biomarker for early colorectal cancerSame-study held-out CRC validation with open processed matrices	Regional fragmentation from ordinary input-control cfDNA WGSReleased fragment matrices: 405 CRC plus 296 noncancers for training; 220 CRC plus 164 noncancers for validation	The paper reported input-WGS fragmentation-only AUC about 0.83 and 62% sensitivity at 95% specificity in its held-out internal validation.CRC only, post hoc same-study validation rather than independent-site MCED; technical/vendor structure and the released vector's exact channel identity require further audit.
2021	PublishedDetection and characterization of lung cancer using cell-free DNA fragmentomesDevelopment plus external retrospective cohort	DELFI lung classifier with clinical features129 lung cancers and 158 controls; external 46 cancers and 385 controls; 1-2x	Cross-validation AUC 0.90 and stage-I AUC 0.76; combined model reached 94% sensitivity at 80% specificity.Symptomatic/imaging-positive context and source differences between cases and controls do not represent population MCED.
2021	PublishedThe LIQUORICE algorithm identifies tissue-specific cell-free DNA signatures in cancer patientsRetrospective method study	Regulatory-region coverage and fragment length263 samples, including sarcomas and 22 healthy participants; median 12x with computational 1x/0.1x downsampling	Cancer-versus-healthy AUC reached 0.97 in selected analyses.Deep discovery, pediatric/high-shedding biology and in-silico downsampling limit ordinary lpWGS MCED interpretation.
2020	PublishedPlasma DNA end-motif profiling as a fragmentomic marker in cancer, pregnancy, and transplantationMechanistic and liver-cancer case-control study	Four-base endpoint motif diversityPlasma WGS	Showed disease-associated changes in endpoint motifs and motif diversity.End repair, library chemistry, sequencer and reference context can dominate; assay lock is mandatory.
2019	PublishedGenome-wide cell-free DNA fragmentation in patients with cancerFoundational retrospective case-control MCED study	DELFI regional short/long fragmentation and coverageClassifier: 208 cancers, 215 healthy controls, seven cancer types, 1-2x WGS and 504 5-Mb windows	AUC 0.94; 73% sensitivity at 98% specificity; reported stage I/II/III sensitivities 68%/72%/79%.Internal case-control cross-validation with healthy controls and batch/site risk; not prospective screening performance.
2019	PublishedOrientation-aware plasma cell-free DNA fragmentation analysis in open chromatin regions informs tissue of originMechanistic method	Orientation-aware fragmentationPlasma WGS	Strand-specific endpoints around open chromatin carry tissue information.Requires pinned genome build, strand policy and cell/tissue atlas; tissue-origin must be evaluated conditionally after detection.
2018	PublishedEnhanced detection of circulating tumor DNA by fragment size analysisRetrospective multi-cancer analytical study	Global fragment size, size selection and CNA344 plasma samples from 200 cancer patients; low-pass analysis around 0.4x	Short-fragment selection often enriched ctDNA more than twofold and size complemented CNA.Mostly advanced disease; physical size selection discards molecules and can introduce bias.
2017	PublishedScalable whole-exome sequencing of cell-free DNA reveals high concordance with metastatic tumorsLarge metastatic analytical cohort	ichorCNA ultra-low-pass copy number and tumor fraction1,439 samples from 520 metastatic breast/prostate patients; approximately 0.1x ULP-WGS branch	Favorable arm-level CNA profiles were detectable around a modeled 3% tumor fraction.The 3% figure is not a clinical early-cancer detection limit and CNA-neutral disease will be missed.
2016	PublishedCell-free DNA comprises an in vivo nucleosome footprint that informs its tissues-of-originMechanistic plasma-WGS study	Windowed protection score and nucleosome footprintsPublic plasma WGS with heterogeneous approximately 1x to deep coverage	Established that cfDNA coverage around nucleosomes reflects chromatin and cell-of-origin biology.Foundational feature biology, not an MCED validation study; very few independent healthy donors.
2016	PublishedInferring expressed genes by whole-genome sequencing of plasma DNAMechanistic method	TSS nucleosome occupancyPlasma WGS	Linked promoter coverage patterns to gene expression and tissue contributions.Supports a chromatin expert; sparse low-pass data require aggregation over predefined programs.

Sources were checked on 2026-08-01. Performance values retain each paper’s own population and operating point and are not suitable for direct rank-order comparison.

Reproducibility package

Take the code, evidence and exact benchmark outputs.

The download page provides a source bundle with feature extraction, simulation, model and validation code, tests, structured evidence, dataset accessions, checked manifests and the report builder, plus separate empirical FORESEE, CRAG and FinaleMe output files. It excludes the downloaded patient-derived raw/coordinate archives and the obsolete local web starter.

Low-pass WGS for multi-cancer early detection

Proceed—but build a disciplined fragmentomics platform, not an “LLM cancer test.”

Classical fragmentomic experts + late fusion

Clinical claim waits for intended-use data

Boundary-aware MIL, not text LLMs

Open empirical work now; controlled validation next

Three real-data analyses, including a confounding failure.

FORESEE / PRJNA755688

CRAG / Zenodo 6914806

FinaleMe / Zenodo 7779198

Same-study published holdout: class separation with confounding risk.

Eight matched early liver-cancer pairs: an exploratory engineering result, not a clinical estimate.

Real-fragment computational downsampling

The fragment model looked excellent—and batch alone looked almost identical.

Exactly what was not empirically tested

A transparent simulation for exercising code—not patient performance.

A constructed depth sweep, not a clinical depth experiment.

Keep every layer separate enough to diagnose.

Specimen

Library + reads

Feature experts

OOF calibration

Decision

Clinical endpoints

Combine complementary failure modes—not duplicate views of the same fragments.

The useful idea is fragment representation—not natural language.

Public data can support real reanalysis—within strict intended-use limits.

FORESEE CRC input-WGS fragmentomics

CRAG open fragment-coordinate cohort

FinaleMe ULP-WGS prostate/breast cohort

1000 Genomes germline reference data

Snyder plasma nucleosome-footprint cohort

DELFI discovery cohort

LUCAS lung and benign-control cohort

Mouliere fragment-size cohorts

FinaleDB processed fragmentome collection

Prospective early gastric cancer WGS

SPOT-MAS public bisulfite cohort

ichorCNA metastatic ULP-WGS cohort

Evaluated software stack

Advance through evidence gates, not calendar promises.

Phase 0 — Freeze the analytical contract

Phase 1 — Finish the public-data ladder

Phase 2 — Multisite development, grouped from day one

Phase 3 — Frozen LUCAS and gastric challenges

Phase 4 — Real library depth and pre-analytic validation

Phase 5 — Prospective intended-use MCED study

Specificity is a sample-size problem before it is a model problem.

Cost per reportable result, with every assumption exposed.

38 primary studies and official records, with the caveat beside the claim.

Take the code, evidence and exact benchmark outputs.