rewirebio.iobenchmarks

Use caseDNA and genomesResearch and clinical research

Select a plasma ctDNA methylation detection workflow

Which measured methylation workflow detects tumour-derived plasma DNA and, where separately supported, identifies tissue of origin?

What the evidence shows

Six bounded sources compare cfDNA methylation methods.

None is a prospective screening study and none measures clinical benefit. Jamshidi et al. 2022 (Cancer Cell), by the GRAIL developers, compared prototype cfDNA classifiers on the same CCGA substudy 1 participants at a 98% specificity threshold set post hoc within each set. On the validation set of 464 cancers and 362 non-cancers, analysed double-blinded, the whole-genome bisulfite methylation classifier detected 34% of cancers (95% CI 30% to 39%; 158 of 464), against 33% for variants with white-blood-cell correction, 30% for copy number with that correction, 27% for copy number, 16% for uncorrected variants and 2.6% for clinical risk factors alone; only the gap to the corrected variant classifier was not marked as significant.

Among 127 validation cancers detected by all three representative classifiers, a methylation origin classifier developed after unblinding assigned the correct tissue for 75% (95 of 127), against 41% for copy number and 35% for corrected variants, with 'other' counted as correct for six cancer types. These are case-control results from the assay developer, not screening performance. Sun et al. 2024 (Genome Biology), an independent benchmark whose authors did not develop the tools, deconvolved real plasma with five methods and reported the ROC-AUC of a random forest trained on the estimated cell-type fractions; the paper does not state how the random forest was validated, so the AUCs may be in-sample.

For 24 liver cancer patients and 32 healthy individuals (WGBS), AUCs were 0.96 (CelFEER), 0.93 (CelFiE), 0.91 (MethAtlas), 0.77 (cfNOMe) and 0.72 (UXM). On cfMethyl-Seq plasma (30 to 67 patients per cancer type, each against the same 193 healthy individuals), AUCs ranged from 0.72 (UXM, lung squamous cell carcinoma) to 0.93 (CelFiE, gastric cancer), and no method was highest for every cancer type. Giuili et al. 2025, a bioRxiv preprint not yet peer reviewed, ran ten reference-based tools inside the authors' DecoNFlow pipeline on in silico mixtures of tumour tissue or cell-line reads in healthy plasma reads.

CelFiE had the lowest overall median tumour-fraction limit of detection in each dataset: 0.0070 (WGBS-TT), 0.0100 (RRBS-TT) and 0.0010 (RRBS-CL). CIBERSORT and EpiDISH had overall medians of 0.1000 to 0.2500 on WGBS-TT and RRBS-TT. For five tools the RRBS-CL values disagree between the supplementary table and Figure 4A and are not shown.

Nguyen et al. 2023 (eLife), by the developers of the SPOT-MAS assay at Gene Solutions, compared three tissue-of-origin classifiers on combined methylation, fragment-length, copy-number and end-motif features from cancer patients only. In the validation cohort of 239 patients, overall accuracy was 0.53 (random forest), 0.69 (deep neural network) and 0.69 (graph convolutional network); the graph network was trained with the validation samples in its graph. Zeng et al. 2026 (Nature Cancer), by the cfMeDIP-seq developers, compared a cfMeDIP-seq methylation classifier with fragmentomic features of the same reads and their combinations, reporting means over 10 fold models.

In cross-validation on 356 cancers and 64 healthy donors, methylation alone reached AUC 0.986 and detected 0.949 of cancers at 99% specificity; on an independent validation set its sensitivity at 99% specificity was 0.633, against 0.766 for methylation combined with fragment ratios. These sensitivities are read on few healthy controls (six or seven per cross-validation fold and 17 in the validation set), so 99% specificity there allows no false positive among them, and Li-Fraumeni syndrome carriers without current cancer count as cancers in the validation sensitivities. The Source Data give the methylation-only validation AUC as 0.694 and the text as 0.674, so that value is not shown.

Xu and Feng 2026, a medRxiv preprint not yet peer reviewed, by authors who did not develop CpGPT, fine-tuned the CpGPT methylation foundation model on public targeted bisulfite cfDNA data from 254 gastrointestinal cancers and 46 non-cancer controls and built four classical classifiers themselves. On three train/validation/test splits with test partitions of about 30 samples, mean test AUROC was 0.9904 for CpGPT against 0.9573 (SVM), 0.9420 (elastic net), 0.8460 (gradient boosting) and 0.8408 (random forest); the candidate CpGs were selected using all 300 samples, including the test partitions. Neither study reports performance in a screening population.

Tissue-of-origin evidence here comes from two developer studies, one methylation-only classifier scored on already detected cancers and one multimodal workflow, and neither classifies people without cancer.

You bring
  • Plasma cell-free DNA for targeted/genome-wide methylation sequencing
  • The target specificity constraint for the intended use
You want
A sourced all-stage cancer-detection sensitivity at a declared specificity where one is reported, otherwise the reported proxy endpoint (detection ROC-AUC on patient plasma, tumour-fraction limit of detection in in silico mixtures, or tissue-of-origin accuracy among cancer patients), each with the method, cohort and validation design it comes from.
Evidence in this release
8 comparisons shown · 22 tools · 40 configurations
Kept up to date
Evidence last checked 11 October 2026 · re-checked monthly, next due 11 November 2026
Executive Takeaway

For "Select a plasma ctDNA methylation detection workflow", on Tumour-fraction limit of detection in in silico mixtures (DecoNFlow benchmark, Giuili et al. 2025), WGBS-TT (Limit of detection, a proxy protocol), EpiDISH_CP_ineq scored 0.0085, reported by A benchmark of DNA methylation deconvolution methods for tumoral fraction estimation using DecoNFlow.

Best reported valueEpiDISH_CP_ineqA benchmark of DNA methylation deconvolution methods for…
Limit of detection0.0085Tumour-fraction limit of detection in in silico mixtures (DecoNFlow benchmark, Giuili et al. 2025), WGBS-TT
Baseline on this protocolNone scoredNo margin is computed
Ranked tools1010 from independent studies

10 tools have a value on this protocol; 10 of them come from studies whose authors did not develop the tool. No baseline method is scored on this protocol. Reviewers judged this protocol a proxy for the use case.

Ranking

Ranked on Tumour-fraction limit of detection in in silico mixtures (DecoNFlow benchmark, Giuili et al. 2025), WGBS-TT (Limit of detection, lower is better). Proxy protocol Reviewers judged this protocol a proxy for the use case, not a direct test of it. Only results on this protocol are ranked; other tools are listed below the table.

9 tools in 10 rows, from 2 sources. 10 of 10 results (100%) come from an independent study.

Rank▲ModelLimit of detection (lower is better)EvidenceCoverageReported byLicense
#10.0085Source-checked Independent study1/8A benchmark of DNA methylation deconvolution…Not recorded
#2
CelFiE[-] Baseline
0.0100Source-checked Independent study2/8A benchmark of DNA methylation deconvolution…Not recorded
#20.0100Source-checked Independent study1/8A benchmark of DNA methylation deconvolution…Not recorded
#2
meth_atlas[-] Baseline
0.0100Source-checked Independent study2/8A benchmark of DNA methylation deconvolution…Not recorded
#50.0250Source-checked Independent study1/8A benchmark of DNA methylation deconvolution…Not recorded
#50.0250Source-checked Independent study2/8A benchmark of DNA methylation deconvolution…Not recorded
#70.0500Source-checked Independent study1/8A benchmark of DNA methylation deconvolution…Not recorded
#80.1750Source-checked Independent study1/8A benchmark of DNA methylation deconvolution…Not recorded
#80.1750Source-checked Independent study1/8A benchmark of DNA methylation deconvolution…Not recorded
#80.1750Source-checked Independent study1/8A benchmark of DNA methylation deconvolution…Not recorded
13 tools are evaluated elsewhere in this use case but not on this protocol, so not ranked

Baseline

MethAtlas and CelFiE deconvolution. MethAtlas (non-negative least squares) and CelFiE (expectation-maximisation) estimate tissue contributions to plasma DNA from a reference methylation atlas, without a trained classifier. Sun et al. 2024 benchmark them with three later deconvolution methods and found CelFEER and UXM performed best.

Machine-learning methods are compared with the baseline on the same data in 2 tables: Tumour-fraction limit of detection in in silico mixtures (DecoNFlow benchmark, Giuili et al. 2025), Deconvolution-based cancer detection on real plasma (Sun et al. 2024). Baseline rows are marked in each table.

Tools compared

Comparisons

Unified model comparison for this use case. Compare tools against protocol baselines in a single view.

Tumour-fraction limit of detection in in silico mixtures (DecoNFlow benchmark, Giuili et al. 2025)

Proxy evidenceIndependent study: Reported in a study whose authors did not develop the evaluated tool.

Limit of detection (lower is better), with the other metrics below each value. The best value in each column is highlighted.

ToolWGBS-TTRRBS-TTRRBS-CL
EpiDISH_CP_ineq2 sweepsconventional
0.0085 (best in column)median at 183M aligned reads0.0100median at 10M aligned reads0.0030median at 10M aligned reads
CelFiEBaselineconventional
0.0100median at 183M aligned reads0.0070* (best in column)median at 10M aligned reads0.0010 (best in column)median at 10M aligned reads
meth_atlasBaselineconventional
0.0100median at 183M aligned reads0.0100median at 10M aligned readsNot reported
MetDecodespecialist model
0.0250median at 183M aligned reads0.0250median at 10M aligned readsNot reported
UXMconventional
0.0250median at 183M aligned reads0.1000median at 10M aligned readsNot reported
PRMethspecialist model
0.0500median at 183M aligned reads0.0250median at 10M aligned readsNot reported
CIBERSORTspecialist model
0.1750median at 183M aligned reads0.0500median at 10M aligned readsNot reported
EpiDISH_RPCconventional
0.1750median at 183M aligned reads0.0500median at 10M aligned reads0.1000median at 10M aligned reads
EpiSCOREspecialist model
0.1750median at 183M aligned reads0.0500median at 10M aligned reads0.0500median at 10M aligned reads
Caveats and method

What each column measures:

  • WGBS-TT: Median tumour-fraction limit of detection of ten reference-based deconvolution tools on WGBS-TT mixtures, by sequencing depth and overall
  • RRBS-TT: Median tumour-fraction limit of detection of ten reference-based deconvolution tools on RRBS-TT mixtures, by sequencing depth and overall
  • RRBS-CL: Median tumour-fraction limit of detection of five reference-based deconvolution tools (of ten benchmarked) on RRBS-CL mixtures, by sequencing depth and overall
  • In silico mixtures of tumour tissue or cell-line reads into healthy plasma reads, not patient plasma
  • Medians on a discrete grid of tested fractions
  • Preprint, not peer reviewed
  • This dataset's values in Supplementary Table 3 agree with Figure 4A; the RRBS-CL block of the same table has a conflict that does not affect them.
  • Each median pools tumour types and the three DMR selection tools; the table does not list the pooled combinations. The two reference-free tools of the benchmark are not in Supplementary Table 3.
  • UXM, MetDecode, meth_atlas, PRMeth and CIBERSORT are excluded: their RRBS-CL values conflict between Supplementary Table 3 sheet A, sheet B and Figure 4A. The five tools shown agree across all three.
  • One neuroblastoma cell line (CLB-GA) mixed into healthy plasma reads; the reference holds nine other neuroblastoma cell lines.

Protocols: DecoNFlow WGBS-TT tumour-fraction limit of detection, DecoNFlow RRBS-TT tumour-fraction limit of detection, DecoNFlow RRBS-CL tumour-fraction limit of detection

Primary benchmark evaluation shown (Tumour-fraction limit of detection in in silico mixtures (DecoNFlow benchmark, Giuili et al. 2025)). 7 additional study protocols (CCGA substudy 1: cancer signal origin accuracy, CCGA substudy 1: cfDNA classifiers on the same participants at a post hoc 98% specificity, CpGPT vs classical classifiers on EpiPanGI Dx cfDNA methylation (Xu and Feng 2026) and others) are cited in individual tool drawer rows and under Question, sources and review.

Not covered yet

  • No per-run scored denominator beyond the printed test n=102 is inferred.
  • A correction/source-reuse note on this source is tracked separately from the assay sensitivity/specificity values and must not be conflated with them.
  • No result is comparable across sources: cohorts, inputs, metrics and validation designs differ.
  • The only same-participant comparison at a declared specificity (Jamshidi et al. 2022) is developer-run on prototype assays in a case-control set, with the 98% specificity threshold set post hoc within each set; four of its classifiers, including the pan-feature classifier, were developed after validation unblinding and are not shown here.
  • Sun et al. 2024 does not state the random forest validation split for its AUCs.
  • Giuili et al. 2025 is a preprint on in silico mixtures; in its Supplementary Table 3 the RRBS-CL values of five tools (UXM, MetDecode, meth_atlas, PRMeth and CIBERSORT) conflict with Figure 4A, and those values are not shown.
  • Nguyen et al. 2023 is author-reported by the assay developer; its Table S9 breast stage strata disagree with the per-patient table and are not shown, and its graph network was trained with the validation samples in its graph.
  • No independent methylation-only tissue-of-origin comparison on patient plasma with printed per-method values was found.
  • No assay-level comparison (cfMeDIP-seq, WGBS, EM-seq or targeted bisulfite) with a detection endpoint on the same plasma samples was found.
  • Jamshidi et al. 2022 tissue-of-origin accuracy is scored on 127 cancers detected by all three classifiers, comes from origin classifiers developed after unblinding, and counts 'other' as correct for six cancer types; the Results and Methods name different classifier triples for that set.
  • Zeng et al. 2026 compare methylation with fragmentomic features from one assay (cfMeDIP-seq) and one group; their sensitivities at 99% specificity are read on few healthy controls (six or seven per cross-validation fold, 17 in the validation set), and their methylation-only validation AUC is printed as 0.694 in the Source Data and 0.674 in the text.
  • Xu and Feng 2026 is a preprint whose candidate CpGs were selected using all 300 samples, including the test partitions of about 30 samples each.
  • No machine-learning method has been compared with MethAtlas or CelFiE on the same data in the linked evidence. The learned cancer-detection classifiers are evaluated on other cohorts.

Contribute evidence or propose a correction

How to cite

Select a plasma ctDNA methylation detection workflow. Rewire biological model benchmark database, release 2026-11-21-c8bea25e5abb. https://benchmarks.rewirebio.io/use-cases/plasma-ctdna-methylation/

Release 2026-11-21-c8bea25e5abb. Rankings and values change between releases, so keep the release id in the citation. Cite the original publication as well for any value you reuse: each result links its source.

The compilation and annotations made by Rewire in this release are offered under CC BY 4.0; values and text taken from cited sources stay under the terms of those sources and are not covered by this licence. Data licence

Question, sources and review

Who this is for, setting and exclusions

Inspect the cfMethyl-Seq repeated-split sensitivity/specificity evidence, the same-participant comparison of whole-genome methylation with other cfDNA classifiers at a post hoc 98% specificity (CCGA substudy 1), and the proxy comparisons of deconvolution methods and tissue-of-origin classifiers, before selecting a workflow and validation design for the intended cancer-detection population.

Who this is for
  • Clinical researchers scoping the available diagnostic-genomics evidence
  • Computational researchers comparing exact evaluated configurations
Biological setting
cfMethyl-Seq achieves 80.7% all-stage cancer-detection sensitivity (95% CI 68.6-90.7) at 97.9% specificity, under a repeated random 25% test split (cohort 217 cancers/191 non-cancers, test n=102). The other evidence is a developer comparison of a whole-genome bisulfite methylation classifier with mutation, copy-number, fragment and clinical classifiers on the same CCGA substudy 1 participants, an independent benchmark of five deconvolution methods on two plasma cohorts, a preprint benchmark of ten deconvolution tools on in silico mixtures, and one author-reported tissue-of-origin comparison on cancer patients recruited in Vietnam.
Clinical research scope
Clinical applicability is not established as prospective screening performance. The cfMethyl-Seq figure comes from a repeated random test split on one assembled cohort, and the CCGA substudy 1 figures from a developer case-control validation set with a post hoc 98% specificity threshold; the other comparisons are proxies (a random forest AUC with no stated validation split, in silico mixtures, and tissue of origin among cancer patients only), and none is an independent prospective validation.

Outside this use case

  • Prospective screening validation
  • Tissue-of-origin identification as a screening result: the ingested tissue-of-origin results classify cancer patients only, and Jamshidi et al. 2022 only cancers already detected by three classifiers, so origin calls in people without cancer are not measured
Evidence collection plan
Evidence collection planCollecting evidence

Already on record: 8 comparisons shown · 22 tools, in the evaluated evidence above. The status above describes only the specific comparison in this plan, which remains open; it does not mean no evidence has been collected. The plan defines a comparison to investigate; it does not establish model performance or suitability.

Comparison question

Which measured methylation workflow detects tumour-derived plasma DNA and, where separately supported, identifies tissue of origin?

Baselines, outcomes and validation requirements

Baselines to include

  • The conventional/author-introduced workflow measured in the linked primary source(s).

Outcomes to measure

  • The declared endpoint in this use case's active mapping(s); see evidence_gaps for what remains open.

Validation requirements

  • Independent held-out population matched to the intended clinical setting.
  • Qualified human scientific review before any clinical-validation claim.

Next collection task

A dedicated rewire-benchmarks protocol/run task with an independent prospective cohort and an explicit tissue-of-origin evaluation if pursued.

Sources, review and downloads

Automated source review · 2026-10-07 · Claude Sonnet AMP-integration worker, bounded transcription of Codex-checked primary values; independently reviewed by Codex (workbench/amp-supervision/primary-review.md, integration-review-corrections.md)

Bounded primary-source transcription, independently Codex-checked. No new model execution, independent experimental reproduction, qualified human scientific review or clinical validation.

Release 2026-11-21-c8bea25e5abb · question use-case-plasma-ctdna-methylation

Use-case input digest bfbb7ab9303becad1e3cb395568c26f41e72ddb67cd1a89a06d9eaa0f76bd95f

Download questions and review metadata (JSON) (gzip) · Verify release checksums (gzip)