ECOG-ACRIN CH Assay Comparison

Standardized CH detection August 2026 Update

Ashwin Kishtagari, MBBS, MSCI

Division of Hematology/Oncology, Vanderbilt University Medical Center

2026-08-13

Why this matters

Detection is not standardized

• Every major center uses a different CH assay, with different sensitivity thresholds and gene coverage. No systematic head-to-head comparison exists.

Our own data show how large the gap is. Genome or exome sequencing, compared with deep targeted sequencing:

Metric	Value
Overall sensitivity	28%
Positive predictive value	44%
Sensitivity at VAF 2–5%	9%
Sensitivity at VAF 5–10%	32%
Sensitivity at VAF 10–20%	67%

These ascertainment errors underestimate true effect sizes by >80%.

The consequence for prevention trials

A patient may qualify at one institution and not another. Without standardization, multi-site enrollment is not defensible.

The risk scores inherit the bias. CHRS and MNPredict were derived from shallow whole-exome and whole-genome biobank data — anchored to a systematic undercount of CH, in exactly the 1–10% VAF window that drives trial eligibility.

Event rates are too low to power a trial unselected. Annual progression is roughly 0.05% in CH with normal counts and 0.5% in CH with cytopenia. No prevention trial is feasible without enrichment — and enrichment starts with knowing what you detected.

Objectives

Primary and secondary objectives

Primary objective

Characterize and compare the technical performance of the CH assays in routine use at seven academic centers, to establish optimal detection methods for ECOG clinical trials.

Secondary objectives

1. Define sensitivity and specificity thresholds for each assay
2. Identify assay-specific strengths and limitations
3. Develop selection criteria for fit-for-purpose applications
4. Create guidelines for CH detection in clinical trials

This is fundamentally a benchmarking study. We are not trying to declare a winner — we want to understand strengths, limitations, and appropriate use cases for each approach, so an investigator can say “if my trial asks X, I should use Y.”

The consortium

Seven participating centers

Center	Investigator	Platform
Vanderbilt	Alex Bick	Twist · ~36 genes, error-corrected
Dana-Farber	Coleman Lindsley	Custom · 0.1% VAF LOD
Washington University in St. Louis	Kelly Bolton	Institutional · 9 genes, therapy-related
Memorial Sloan Kettering	Ross Levine	MSK-ACCESS · comprehensive
Mayo Clinic	Mrinal Patnaik	Mayo · 37 genes, CLIA
University of Pennsylvania	Daehyun Lee	VariantPlex · 74 genes (Archer/IDT)
Yale University	Jennifer Kwan	pending

A participation request and short assay survey went to all members of the ECOG-ACRIN CH working group. These centers returned assay details and agreed to receive blinded samples. Together they span the platforms currently used to call CH in the US, from clinical myeloid NGS panels to research-grade duplex sequencing.

ACTION before presenting: fill in the Yale investigator and platform. Lourdes Mendez is the Yale member of the CH subcommittee and is the obvious person to ask.

Study design

Phased design

A preliminary 40-sample pilot generates comparative data before scaling to the full 400-sample validation.

Phase 1 — Feasibility (N=40) · IN PROGRESS

Source — Vanderbilt CH repository

VAF range — 0.57% to 50.8%

Design — blinded, identical aliquots to all sites

Goals — inter-assay concordance; surface logistical issues before scale-up

Phase 2 — Full validation (N=400) · PENDING FUNDING

Sample size — 400 reference samples

Power — >95% to detect a 5-percentage-point difference in sensitivity (α = 0.05)

Extended — VAF stratification, longitudinal tracking, gene-specific performance

Deliverables — comparative manuscript; ECOG-ACRIN selection framework

Pilot cohort: 40-sample characterization

40 samples · 15 distinct genes · 0.57–50.8% VAF range · 4.4% median VAF

VAF distribution

VAF range	N	Detection category
<1%	5	Ultra-low detection
1–2%	7	Standard threshold
2–5%	10	Moderate confidence
5–15%	7	High confidence
15–51%	11	Very high confidence

Gene coverage (15 genes)

N	Genes
8	DNMT3A
5	TET2
3	JAK2, TP53, ASXL1, SRSF2, SF3B1
2	PPM1D, IDH2, GNB1, BRCC3
1	GNAS, SETBP1, BCORL1, KRAS

The two commonest CHIP drivers dominate, as expected — but the panel reaches splicing factors and TP53.

Distribution is weighted toward ≥2% VAF for robust cross-platform concordance in this preliminary phase.

Phase 2 reference panel

400 samples, prespecified so that sensitivity, specificity, and limit of detection can be estimated within each VAF tier and against unequivocal negatives.

Engineered · n = 200

8 drivers at 0.5, 1, 2, 5, 10, 20, 30% VAF

Cell-line DNA blended into wild-type gDNA; target VAF confirmed by ddPCR

40 carry 2–3 co-occurring drivers

Clinical · n = 40

BioVU patient DNA, VAF 0.5–30%

Serial draws where available

Negative controls · n = 160

No variant at 0.1% VAF by duplex sequencing

Donors <40 y and cord blood

Drivers: DNMT3A, TET2, ASXL1, JAK2, TP53, SF3B1, SRSF2, U2AF1

Engineered samples give sensitivity, limit of detection, and caller behavior with multiple clones. Clinical samples test performance in a real sample matrix. Negatives give specificity and false-positive rate. VAF tiers below 1% fix each assay’s detection floor — they are not clinically actionable thresholds. Prepared with Dr. Brian Shirts, VUMC Molecular Pathology.

The consensus truth standard

No single participating assay can serve as ground truth. Truth is prespecified in three tiers.

1. Engineered samples — variants at known VAF are true positives by construction

2. Clinical samples — true when called by at least 4 of 7 assays

3. Indeterminate stratum — samples with only 2 or 3 concordant calls are adjudicated by droplet digital PCR and the methylation readout, and reported separately

Technical validation framework

1. Serial dilutions Test sensitivity limits 0.1% to 10% VAF

4. Truth variant definition Consensus standard across the 400-sample panel

2. DNA input titration Assess input requirements 10 ng to 100 ng

5. Longitudinal tracking VAF trajectory analysis Clonal dynamics mapping

3. Computational downsampling Evaluate depth requirements 100× to 10,000×

6. Evolution metrics Growth rate calculation Emergence detection <0.5%

Novel: first systematic comparison of assays for **longitudinal CH.

The challenge is that we are not comparing apples to apples. Each assay was designed with different goals — some prioritize sensitivity, others specificity, others turnaround or breadth of coverage. The framework has to be fair to all approaches while still returning actionable information. Genes like ASXL1 and BCOR with homopolymer regions are where artifacts concentrate.

Statistical plan

Primary analyses

1. Sensitivity and specificity

• Binomial exact 95% confidence intervals
• McNemar test for paired comparisons
• ROC curves at varying VAF thresholds
• AUC with DeLong test

2. Inter-assay concordance

• Cohen’s κ for pairwise agreement
• Bland-Altman for VAF correlation
• Lin’s concordance correlation coefficient
• Fleiss’ κ across all assays

3. Reproducibility

• Coefficient of variation for replicates
• Intraclass correlation coefficient
• Repeatability and reproducibility study

Modeling

• Generalized linear mixed models (lme4), center as a random effect
• Bonferroni for primary endpoints; FDR control across genes

We did not just throw samples at seven labs and see what came back. When an investigator asks “can this assay reliably detect variants at 1% VAF?”, the answer needs a confidence interval, not anecdote.

Power and stratification

With 400 samples and an expected 60% positive fraction, we have >95% power to detect inter-assay differences in sensitivity of 5 percentage points at α = 0.05.

VAF stratification for analysis

VAF range	Analysis tier
0.5–1%	Detection floor — fixes each assay’s LOD
1–2%	Standard reporting threshold
2–5%	Where sequencing PPV is poorest
>5%	High confidence; expected cross-platform agreement

Expect substantial inter-center heterogeneity below 5% VAF and in low-complexity regions, with no single assay dominating — which is itself the argument for fit-for-purpose guidelines.

Status and timeline

MTA status

As of the August 2026 update

Institution	Investigator	Status
Dana-Farber	Coleman Lindsley	Completed
University of Pennsylvania	Daehyun Lee	Completed
Memorial Sloan Kettering	Ross Levine	Completed
Washington University	Kelly Bolton	Completed
Mayo Clinic	Mrinal Patnaik	Completed
Yale	Jennifer Kwan	Pending

ACTION before presenting: update this table with current status. This is the single highest-value slide for a committee that can actually unblock signatures.

Deliverables and timeline

Phase	Activity	Status
Months 1–3	MTAs, sample acquisition and distribution	In progress
Months 4–6	Assay testing and data generation	Planned
Months 7–9	Data analysis and comparison	Planned
Months 10–12	Guidelines development and dissemination	Planned

Phase 2 is when each site runs its assay on the same sample set. Phase 3 looks at concordance, sensitivity, specificity, and practical considerations like turnaround and cost. Phase 4 produces the publications and decision tools.

Asks and discussion

Where we stand

• Seven academic centers have agreed to run their CH assay on a common, blinded reference set
• Phase 1 pilot (N=40) is underway — Vanderbilt CH repository, VAF 0.57–50.8%, 15 genes
• Phase 2 (N=400) reference panel is designed, with a pre-specified consensus truth standard — the piece the field has been missing
• MTAs: 6 of 7 executed as of the April update; the rest pending signature
• The comparison feeds directly into a screening-pathway framework for ECOG-ACRIN MDS prevention trials

Bottom line up front, because committee time is short. The three things worth their attention: the truth standard, the MTA bottleneck, and the fact that this now has a defined downstream use in prevention trial design.

Acknowledgments

ECOG-ACRIN CH Subcommittee

• Alexander Bick — Vanderbilt
• Yael Kusne — Mayo Clinic
• Jennifer Eads — Penn
• Kristina Kirschner — Mayo Clinic
• Daniel Nathan — Mount Sinai
• Lachelle Weeks — DFCI
• Kalyan Nadiminti — Wisconsin
• Tamanna Haque — MSK
• Cathy Handy Marshall — Johns Hopkins
• Kelly Bolton — Washington University
• Janghee Woo — Emory

• Roni Tamari — MSK
• Sanam Loghavi — MD Anderson
• Uma Borate — Ohio State
• Bonnie Ky — Penn
• Jenica N. Upshaw — BIDMC
• Daehyun Lee — Penn
• Antoine N. Saliba — Mayo Clinic
• Urvi Shah — MSK
• Lourdes Mendez — Yale
• Virginia Corbett — Mount Sinai
• Yasmin Abaza — Northwestern

Also: Brian Shirts (VUMC Molecular Pathology) for reference panel construction, and the Bick laboratory for sequencing capacity.