ABNL-MARRO 001 — Arm A

ASH 2026 Analysis Plan · ASTX727 + Itacitinib in MDS/MPN Overlap Syndromes

Ashwin Kishtagari, MD, MSCI — VUMC Hematology & Oncology

Division of Hematology/Oncology, Vanderbilt University Medical Center

2 Jul 2026

Purpose

A working analysis plan for the 28-patient Arm A dataset — a cohort whose phase 2 was terminated early for toxicity:

For the statistician

• Endpoints, populations, and estimands
• Estimation-over-testing framing for an early-terminated arm
• Method-by-quantity mapping (§ Methods)

For Dr. Savona

• Strategic recommendation (two linked abstracts)
• The imaging novelty hook for ASH

All efficacy framing is descriptive. Phase 2 was closed early for toxicity after Stage 1 accrual (21 treatment-naïve of a planned 20 → 58), so the optimal Simon two-stage design was not completed and no formal test against the 35% null was performed. Safety/tolerability is co-primary to the story.

Orient the room: this document does double duty. The stats reader cares about the back half (populations, methods, estimand); Savona cares about the front half (recommendation, novelty, timeline). The key change since the last draft: phase 2 stopped early for toxicity — so the analysis is a final, descriptive readout of an early-terminated arm, and safety is now part of the headline, not just a section.

Executive summary & recommendation

We hold a frozen, early-terminated Arm A cohort (n = 28) of ASTX727 (oral decitabine/cedazuridine) + itacitinib (JAK1i) in CMML, aCML, MDS/MPN-RS-T, and MDS/MPN-U — a rare group where standardized response assessment is still immature. The phase 2 portion was closed early for toxicity after 21 treatment-naïve patients (Stage 1 of a planned optimal Simon two-stage design, target 20 → 58) and did not proceed to Stage 2. With the 7-patient phase 1b safety run-in, this totals 28 patients treated at the RP2D doublet.

Because early closure was toxicity-driven, the safety/tolerability profile — reported alongside a descriptive response signal — is itself the important, publishable ASH result in a disease group with few prospective data. Separately, the protocol carries a genuinely novel imaging objective: AI-based automated spleen (and liver) volumetric segmentation from non-contrast CT (Landman group) as an objective measure of clinical benefit.

Recommendation — submit TWO linked abstracts

Abstract 1 (Clinical): safety and tolerability leading to early termination, plus descriptive ORR by modified IWG criteria and molecular correlates.

Abstract 2 (Imaging/Methods): AI-quantified spleen-volume dynamics as an objective response biomarker, benchmarked against palpation and correlated with IWG response and symptom burden.

Dataset, source data & populations

Source data

• EDC freeze: Arm A freeze executed 12-Jun-2026; subject casebooks (PDF) delivered via Theradex sFTP.
• Cohort: 28 patients — ASTX727 35/100 mg FDC Days 1–5; itacitinib 300 mg daily (RP2D from phase 1b run-in).
• Composition: 7 phase 1b safety run-in (naïve + R/R) + 21 phase 2 Stage 1 (treatment-naïve only). Phase 2 closed early for toxicity; Stage 2 never opened. Tabulate the naïve vs R/R split explicitly — it conditions every efficacy statement.

Analysis populations

• Safety (lead population): all 28 receiving ≥1 dose of either drug.
• Efficacy (headline): pooled 28 with ≥1 post-baseline response (or EOT if off before C7); report phase + line composition as a footnote. Sensitivity: phase 2 Stage 1 only (n = 21, naïve) — the protocol-defined efficacy set.
• Imaging-evaluable: analyzable baseline + ≥1 on-treatment CT through the pipeline — report n explicitly.
• Molecular-evaluable: baseline + ≥1 follow-up NGS.

Exposure is truncated — interpret efficacy with care

Because phase 2 closed early for toxicity, on-treatment exposure and follow-up are truncated; some subjects discontinued before the Cycle 7 response window. Use best response through C7 plus EOT response, and read ORR in light of limited exposure. Capture reason for discontinuation (toxicity / progression / death / consent) per subject — it drives both the safety narrative and the swimmer plot.

Data-handling note

Casebooks are PDFs, not an analysis-ready dataset

Budget time to abstract the casebooks into a tidy, patient-level table (one row/subject; columns for phase, line, subtype, cycles received, best response, per-timepoint CBC/marrow/spleen/TSS, mutations, AEs and reason for discontinuation). Confirm with Theradex/Marco whether a SAS/CSV export of the frozen EDC can be released alongside the PDFs — that would save substantial manual curation before Aug 1.

Endpoints to be analyzed (per protocol)

Domain	Endpoint / measure
Primary efficacy (phase 2)	ORR = CR, PR, optimal/partial marrow response (MR), or clinical benefit (CB) as best response through Cycle 7 (modified MDS/MPN IWG).
Safety (co-lead here)	AE/DLT type, grade (CTCAE), frequency; AEs of special interest; dose modifications; discontinuations by reason; toxicities triggering early phase 2 closure; 30/60-day mortality.
Secondary efficacy	CR/MR rate; OS and PFS; morphologic marrow response.
Response components	Δ from baseline in Hgb, WBC, ANC, platelets, monocytes; transfusion independence; MPN-SAF TSS; spleen volume — at C3D1, C7D1, EOT.
Imaging (exploratory)	Automated spleen + liver volume from non-contrast CT (Landman) at Screening, C3D1, C7D1, EOT.
Molecular (exploratory)	Baseline mutation frequency and ΔVAF across a ≥37-gene myeloid NGS panel; molecular response / clonal dynamics.

For the statistician: the response window is anchored at Cycle 7 (assessments after 2 and 6 cycles, i.e., C3D1 / C7D1). Paired component timepoints are fixed at C3D1 / C7D1 / EOT — those drive the Wilcoxon paired comparisons later. Given toxicity-driven early stop, many patients contribute EOT (not C7) as their last assessment.

Abstract 1 — overview

Working title: “ASTX727 in Combination with Itacitinib in MDS/MPN Overlap Syndromes: Safety and Response from Arm A of the ABNL-MARRO 001 Study.”

Core message

First reporting of an all-oral DNMTi + JAK1-inhibitor doublet in MDS/MPN — a combination whose phase 2 was terminated early for toxicity — with response assessed by the modified IWG criteria the consortium is helping to validate. An honest safety-and-signal readout in a rare disease.

Centerpiece figure (one allowed): swimmer plot — time on treatment, best response, and reason for discontinuation — so one figure carries both the response signal and the toxicity-driven early discontinuations. (Waterfall as the alternative.)

Denominator discipline: every efficacy statement is conditioned on the phase + naïve/R-R composition and on the efficacy-evaluable population — state both up front.

Abstract 1 — planned analyses

Safety (lead — toxicity drove termination)

• TEAEs by CTCAE grade & relatedness; AEs of special interest (itacitinib: cytopenias, infection, transaminitis; ASTX727: myelosuppression, GI); DLTs from run-in; dose holds/reductions; discontinuations by reason; 30/60-day mortality.
• Characterize the specific toxicities that triggered early phase 2 closure — event type, timing, attribution, reversibility, and the formal basis for closure (safety review / DSMB). This is the co-headline.

Baseline / disposition

• CONSORT-style accounting: accrued → treated → evaluable, annotated with the early-termination point. Table 1: demographics, WHO subtype (CMML / aCML / MDS/MPN-RS-T / MDS/MPN-U), IPSS-M or CPSS-Mol risk where available, prior therapy, baseline cytogenetics & mutations.
• Exposure: cycles started/completed, dose intensity, time on treatment.

Efficacy (descriptive)

• ORR (pooled 28) with exact (Clopper–Pearson) 95% CI; breakdown by category (CR / PR / optimal-MR / partial-MR / CB); footnote phase + line composition. Sensitivity: phase 2 Stage 1 (n = 21, naïve).
• Best response by subtype — descriptive given n; pre-specified exploratory, not powered.
• Component responses — erythroid/transfusion-independence, platelet, spleen, symptom (TSS) — tabulated separately; feed the IWG-validation narrative.
• Time-to-event — OS and PFS by Kaplan–Meier with 95% CIs; median follow-up by reverse-KM. Descriptive at n = 28 with truncated follow-up.
• Change-from-baseline — Hgb, WBC, ANC, platelets, monocytes, TSS at C3D1/C7D1/EOT via Wilcoxon signed-rank (protocol-specified).

Molecular correlative (if NGS available in time)

• Baseline mutation landscape (oncoprint); association of mutations (ASXL1, SRSF2, TET2, RAS-pathway, JAK2) with response — descriptive.
• ΔVAF of dominant clones (baseline → best response); define molecular response and relate to morphologic/clinical response.

Abstract 1 — statistical framing

Design context — and why the design was not completed

Phase 1b run-in (3+3, n = 7) feeding phase 2 Stage 1 of an optimal Simon two-stage design: H₀ ORR = 35% vs. one-sided alternative 50%; α = 0.10, power 80%. Stage 1 = 20 treatment-naïve (futility if ≤7 responses before C7); otherwise +38 to N = 58 (reject H₀ if ≥25/58).

Phase 2 was terminated early for toxicity after 21 Stage-1 patients — the two-stage design was not completed and no test against the 35% null was performed. Per protocol, phase 1 responses are excluded from the phase 2 efficacy determination.

Estimation over testing

Report point estimates with confidence intervals. The pooled-28 ORR is descriptive, not a test against the Simon boundary. Reserve hypothesis tests for the protocol-specified paired comparisons (Wilcoxon signed-rank) of response components.

This is the slide the statistician will push on. Be explicit: (1) the arm stopped for toxicity, not at a Simon boundary; (2) pooled-28 ORR is a Clopper–Pearson estimate carried as the headline per team decision, with the phase-2 Stage 1 population (protocol-defined, phase-1 excluded) shown as a sensitivity; (3) efficacy is confounded by toxicity-truncated exposure. Confirm at kickoff exactly how disposition/discontinuation reasons were coded.

Abstract 2 — the ASH hook (why novel)

Working title: “Deep-Learning–Based Automated Spleen Volumetry on Non-Contrast CT as an Objective Measure of Clinical Benefit in MDS/MPN: Results from ABNL-MARRO 001 Arm A.”

Why this is novel

• Quantitative vs crude: IWG response in MDS/MPN still leans on palpable splenomegaly (cm below costal margin). Volumetric CT is far more sensitive and reproducible; an automated pipeline removes inter-reader variability entirely.
• Automated vs manual: manual spleen segmentation is slow and rarely done at scale; the Landman deep-learning approach makes per-timepoint volumetrics feasible across a multi-site trial.
• Disease-specific gap: spleen volumetrics are established in myelofibrosis (MRI/CT) but under-explored as a standardized, AI-derived response metric in MDS/MPN overlap syndromes specifically.

The imaging story is independent of the toxicity termination — it stands on methodology and travels well on its own. Early closure of the clinical arm does not weaken it.

Abstract 2 — planned analyses

• Pipeline description — segmentation model, training/validation provenance, QC, and agreement vs expert manual contours on a sample (Dice, volumetric % difference). Establishes credibility.
• Baseline burden — distribution of baseline spleen (and liver) volumes; correlation with palpation-based splenomegaly and disease subtype.
• On-treatment dynamics — absolute and % change at C3D1, C7D1, EOT; spaghetti + waterfall plots. Define a candidate volumetric spleen response threshold (≥35% / ≥50% reduction, mirroring MF conventions) and report the response rate.
• Concordance — agreement between volumetric spleen response and (a) palpation, (b) overall IWG response, (c) TSS change — Cohen’s κ, Bland–Altman, Spearman ρ.
• Added value — do volumetric changes reclassify any clinical-benefit calls palpation missed? Even a few discordant cases make a compelling figure.
• Exploratory prognostic — association of early (C3) volume change with subsequent IWG response / PFS — hypothesis-generating.

Threshold must be pre-specified a priori (≥35% vs ≥50%), not chosen after seeing the data.

Abstract 2 — feasibility gate (check FIRST)

One number drives the standalone decision

Before committing to a standalone abstract, confirm the imaging-evaluable n. Non-contrast CT was protocol-scheduled at Screening / C3 / C7 / EOT, but real-world CT capture across sites is the limiting factor — and early discontinuations from toxicity further thin the on-treatment scans.

Pull the actual count of subjects with a baseline + ≥1 on-treatment CT that has cleared the segmentation pipeline:

• Robust n (a healthy share of 28 with paired scans) → go standalone.
• Thin n → integrate as a sub-section of the clinical abstract.

Ask this week; it gates everything downstream.

Make the room agree that we get this count from Theradex + the Landman lab before any analysis time is spent. The toxicity-driven early stop makes the on-treatment scan count even more uncertain — check it early.

Analytical methods & tooling

Quantity → method

Quantity	Method
Response rates (ORR, CR/MR, spleen)	Exact Clopper–Pearson 95% CI
Paired Δ (counts, TSS, volume)	Wilcoxon signed-rank
OS / PFS	Kaplan–Meier, 95% CI; reverse-KM
Time on treatment / discontinuation	Swimmer; reasons tabulated
Method agreement	Cohen’s κ, Spearman ρ, Bland–Altman
VAF dynamics	Paired ΔVAF; clonal trajectories
Subgroup / mutation	Descriptive; Fisher exact where used

Implementation (R / Python)

• R: tidyverse · survival/survminer (KM/PFS) · gtsummary (Table 1 + response/AE tables) · binom (exact CIs) · swimplot/ggplot2 (swimmer/waterfall/spaghetti).
• Python: pandas (casebook abstraction/QC) · lifelines (survival cross-check) · matplotlib/seaborn; coordinate with Landman group on segmentation outputs (volumes per timepoint as a tidy table for merge).
• Reproducibility: single locked analytic dataset, scripted end-to-end, version-controlled; one figure each as the ASH centerpiece.

Decision rule — one abstract or two?

Resolve as soon as the imaging-evaluable count is known:

1. Robust imaging-evaluable n (paired CTs in a healthy share of 28) → TWO abstracts. Clinical carries safety/tolerability + early-termination and descriptive ORR/molecular; imaging carries volumetrics + concordance. Cross-reference.
2. Marginal imaging n, or pipeline/QC not finalized by mid-July → ONE abstract. Lead with safety and the toxicity-driven termination, then efficacy; include a dedicated objective spleen volumetrics paragraph + the spleen waterfall as the single figure.
3. Molecular (NGS) not back in time → drop molecular to a single descriptive line; it is not the headline and should not gate submission.

Novelty verdict

The spleen-segmentation work clears the bar for separate-abstract novelty on methodology. The clinical cohort is independently reportable — a toxicity-terminated doublet with a descriptive response signal is a legitimate, useful ASH result. Default to two; consolidate only if forced by n or timeline.

Timeline to August 1

Window	Milestone	Owner
This week	Kickoff with statistician; lock endpoints, populations, table/figure shells; request EDC export + imaging-evaluable count; confirm the toxicities/basis for early termination	AK / statistician / Theradex
Wk of Jul 6	Casebook abstraction → locked analytic dataset (incl. discontinuation reasons); confirm one-vs-two via gate; Landman volumes as tidy table	AK / data team / Landman lab
Wk of Jul 13	Primary analyses (safety/AE tables, ORR, KM, paired Δ, volumetrics, concordance); draft figures	Statistician / AK
Wk of Jul 20	Draft abstract(s); internal review (Savona, co-authors); finalize single figure each	AK / Savona / co-authors
Jul 27–31	Revisions, author approvals, ASH formatting/character check; submit ahead of Aug 1	AK

Critical-path risk

The binding constraints: (1) getting an analysis-ready export rather than only PDFs, (2) the imaging-evaluable count + segmentation QC, and (3) a precise, documented account of the toxicities that triggered early closure. Nail all three down this week.

Open questions to resolve at kickoff

1. Can Theradex release a SAS/CSV export of the frozen Arm A EDC alongside the casebook PDFs?
2. What were the specific toxicities and the formal basis (safety review / DSMB) for the early phase 2 termination, and the per-subject reason for discontinuation? — needed to state the reason precisely and build the swimmer plot.
3. How many subjects are imaging-evaluable (baseline + ≥1 on-treatment CT through the pipeline)? — drives one-vs-two.
4. Is the NGS (≥37-gene panel) data returned and matched to timepoints for the 28, or pending?
5. Confirm the naïve vs R/R split (R/R confined to the 7 phase 1b) and lock the pooled-28 ORR + phase-2 Stage 1 sensitivity presentation.
6. Confirm authorship/order and any consortium (IWG) or sponsor (Astex/Incyte) publication-review windows affecting the submission date.
7. Agree the volumetric spleen-response threshold a priori (≥35% vs ≥50%) so it is pre-specified.

Close by assigning owners. Items 1, 2, and 3 are the critical path — the termination detail (item 2) is new since last draft and must be pinned down before the safety narrative is written. Everything else can move in parallel.