Phenotype-Specific Associations Between Graft-Versus-Host Disease and Post-Transplant Outcomes

Abstract

Background:

Disease relapse remains the leading cause of failure following allogeneic hematopoietic cell transplantation (HCT). As novel prophylactic strategies increasingly aim to universally eliminate all forms of graft-versus-host disease (GVHD), how these approaches may inadvertently sacrifice graft-versus-leukemia (GVL) activity in the post-transplant cyclophosphamide (PTCy) era is unclear.

Objective:

To evaluate the associations between time-updated GVHD phenotypes and clinical outcomes, including relapse, non-relapse mortality (NRM), and overall survival (OS), in patients undergoing PTCy-based haploidentical or mismatched unrelated donor (MMUD) HCT.

Study Design:

This was a retrospective cohort study using the Center for International Blood and Marrow Transplant Research registry. Participants included 7,055 patients with hematologic malignancies who received a first haploidentical or MMUD HCT with PTCy from 2013–2021. Associations with relapse, NRM and OS were evaluated using multi-state time-dependent Cox proportional hazards models and a multi-state random survival forest (MS-RSF). Time-updated acute and chronic GVHD phenotypes included isolated grade II acute GVHD (aGVHD), grade III-IV aGVHD, mild chronic GVHD (cGVHD), and immunosuppressive therapy-requiring (IST-requiring) cGVHD.

Results:

IST-requiring cGVHD without antecedent aGVHD was associated with a lower modeled relapse hazard compared with remaining GVHD-free (Hazard Ratio [HR], 0.74; 95% confidence interval [CI], 0.62–0.89; False Discovery Rate (FDR)-adjusted p (q)=0.006) and lower overall mortality (HR 0.62; 95% CI, 0.53–0.73; q < 0.001). In contrast, grade III-IV aGVHD was associated with significantly higher modeled NRM (HR, 3.15; 95% CI, 2.57–3.84; q < 0.001). Mild cGVHD and isolated grade II aGVHD showed intermediate patterns without consistent associations. These associations were directionally consistent across multiple analytic approaches, including standard time-dependent Cox regression, dynamic and fixed landmark analyses, and MS-RSF.

Conclusions:

In this large PTCy-treated mismatched donor cohort, IST-requiring cGVHD was the GVHD phenotype most consistently associated with lower relapse incidence, whereas severe aGVHD was associated with higher NRM. These findings highlight heterogeneity in GVHD phenotypes and suggest that strategies distinguishing toxic acute GVHD from chronic alloreactivity patterns may better balance morbidity and long-term disease control. Given that relapse remains the predominant cause of post-transplant mortality, approaches aiming to universally eliminate all GVHD warrant careful reconsideration.

Introduction

The therapeutic potential of allogeneic hematopoietic cell transplantation (HCT) reflects a central tension: donor-derived immunity can eliminate residual malignant cells through graft-versus-leukemia (GVL) activity while also driving graft-versus-host disease (GVHD). Historical registry analyses conducted in the era of conventional prophylaxis linked both acute graft-versus-host disease (aGVHD) and chronic graft-versus-host disease (cGVHD) with lower relapse incidence. These observations shaped long-standing assumptions about the relationship between alloreactivity and antileukemic control.^1,2

However, the landscape has shifted with the widespread adoption of post-transplant cyclophosphamide (PTCy), which has markedly reduced lower non-relapse mortality (NRM).^3–5 Despite these gains, relapse remains the predominant cause of mortality after HCT.⁶ At the same time, emerging prophylactic strategies increasingly aim to abolish all forms of GVHD, raising concern that universal suppression of alloreactivity may inadvertently compromise GVL activity.

Whether complete suppression of GVHD can coexist with preserved antileukemic immunity in the PTCy setting remains unresolved. Recent studies in PTCy-based haploidentical transplantation have not identified a clear association between GVHD and relapse.^7,8 These analyses relied primarily on fixed or dynamic landmarking and categorized chronic GVHD using the Seattle limited/extensive criteria,⁹ which do not distinguish between mild, self-limited manifestations and clinically significant disease requiring systemic immunosuppression. Because immunosuppressive therapy-requiring (IST-requiring) cGVHD represents the subset most reflective of sustained, high-grade alloreactivity, its relationship to relapse risk may differ from milder phenotypes. Yet prior PTCy-era cohorts have been too small to evaluate these more granular distinctions or to assess whether specific GVHD trajectories, rather than static subtypes alone, are associated with clinical outcomes.

The threshold at which immune tolerance corresponds to preserved antileukemic activity in this therapeutic environment therefore remains incompletely defined. To address this gap, we examined the association between time-updated GVHD severity phenotypes and clinical outcomes using multi-state time-dependent survival modeling and machine-learning frameworks in a cohort of 7,055 patients who underwent haploidentical or mismatched unrelated donor (MMUD) HCT with PTCy.

Methods

This retrospective registry study included patients with hematologic malignancies who underwent first allogeneic haploidentical or mismatched unrelated donor (MMUD) HCT with PTCy from 2013 through 2021. Data were obtained from a publicly available Center for International Blood and Marrow Transplant Research (CIBMTR) dataset from the referenced publication.¹⁰ The CIBMTR did not participate in study design, analysis, or manuscript preparation.

Outcomes and Definitions

The primary outcome was relapse. Secondary outcomes included NRM and overall survival (OS). Acute GVHD was categorized as isolated Grade II or Grade III-IV. Chronic GVHD was categorized as mild (not requiring systemic IST) or IST-requiring.

Statistical Analysis

We used a dual-framework analytic strategy combining a multi-state continuous time-dependent Cox model with a multi-state random survival forest (MS-RSF). The Cox model provided transition-specific hazard estimates, and the MS-RSF characterized competing transitions and individual-level heterogeneity.

Multi-State Time-Dependent Cox Model

A four-state structure (no GVHD, GVHD onset, relapse, NRM) was used to represent the clinical course and to avoid immortal-time bias. Five transitions were modeled. For transitions occurring after GVHD onset, time was measured from GVHD onset (gap-time formulation), conditioning on entry into the GVHD state. A single stratified Cox model was fitted across transitions, with transplant center as the clustering unit for robust variance estimation.

Multi-State Random Survival Forests

Each transition was modeled using a separate RSF with 1,000 trees and log-rank splitting. Baseline covariates were included for all transitions, and GVHD subtype was included only for post-GVHD transitions. Discrimination was assessed using out-of-bag concordance indices. Variable importance was quantified using permutation-based variable importance (VIMP), and individual-level contributions were examined using Shapley Additive Explanation (SHAP) values.

Adjusted Cumulative Incidence and State-Occupation Probabilities

For each GVHD subtype, population-averaged cumulative incidence functions (CIFs) for relapse and non-relapse mortality were estimated using Aalen-Johansen estimators derived from MS-RSF-predicted survival curves. State-occupation probabilities were computed to depict the distribution of relapse, non-relapse mortality, and survival over time.

Moderation Analysis of GVL Associations

We next sought to determine whether the magnitude and direction of the GVHD-relapse association varied across patients with GVHD. Because the primary multi-state and time-dependent models identified chronic GVHD, but not acute GVHD, as the phenotype associated with relapse, moderation analyses focused specifically on cGVHD. To evaluate heterogeneity in potential GVL-related effects, we generated individual model-based simulated cumulative incidence functions (CIFs) for relapse under two cGVHD scenarios - mild versus IST-requiring - while holding all covariates constant. For each patient who developed cGVHD, the difference in 24-month relapse CIF defined the simulated relapse-difference score. Moderation was assessed across baseline covariates, and k-means longitudinal clustering was applied to identify distinct response patterns. These analyses were exploratory and intended to characterize variation in relapse risk within cGVHD phenotypes rather than to infer causality.

Sensitivity Analyses

Robustness of the primary Cox model was assessed using a standard (non-multi-state) time-dependent Cox model in Andersen-Gill format,¹¹ a dynamic landmark supermodel stacking monthly datasets from months 1 through 12, and fixed 6- and 12-month landmark analyses. Multiple testing was controlled using the Benjamini-Hochberg false discovery rate (FDR) procedure applied to all GVHD-related contrasts within each modeling framework with significance defined as q < 0.05. Details of statistical methods are provided in the supplementary appendix. Analyses were performed in R version 4.5.2. The MS-RSF code is publicly available on Zenodo at: https://zenodo.org/records/19243243

Results

Study Cohort

Among 7,055 patients, mean (standard deviation [SD]) recipient and donor ages were 49.4 (18.7) and 35.7 (11.5) years, respectively. Most transplants used haploidentical donors (85%) and peripheral blood grafts (PB, 76%). Diagnoses included acute myeloid leukemia (AML, 50%), myelodysplastic syndrome/myeloproliferative neoplasms (24%), and acute lymphoblastic leukemia (ALL, 22%). Fifty-five percent received reduced-intensity or non-myeloablative conditioning, and 45% had an HCT-comorbidity index score of ≥3. Median time to HCT was 7 months. Median follow-up via the reverse Kaplan-Meier method was 45.1 months (95% confidence interval [CI], 40.4–46.5) [Baseline characteristics seen in Table 1].

Table 1.

Baseline Patient, Disease, and Transplant Characteristics

Baseline Characteristic	Total Cohort (N = 7055)
Recipient Age (years), Median (Q1-Q3)	55.0 (36.0–64.0)
Donor Age (years), Median (Q1-Q3)	34.0 (27.0–43.0)
Recipient Sex, n (%)
Male	4,107 (58%)
Female	2,948 (42%)
Disease Type, n (%)
AML	3,534 (50%)
MDS/MPN	1,707 (24%)
ALL	1,519 (22%)
CML	295 (4%)
Disease Risk Index, n (%)
Low/Intermediate	4,614 (65%)
High/Very High	1,177 (17%)
Missing	1,264 (18%)
Donor Type, n (%)
Haplo	6,007 (85%)
MMUD	1,048 (15%)
Graft Type, n (%)
Peripheral blood	5,390 (76%)
Bone marrow	1,665 (24%)
Conditioning Intensity, n (%)
RIC/NMA	3,862 (55%)
MAC	3,140 (45%)
Missing	53
HCT Comorbidity Index, n (%)
0–2	3,824 (55%)
>=3	3,174 (45%)
Missing	57
Karnofsky Performance Status, n (%)
< 90%	2,890 (42%)
90–100%	4,029 (58%)
Missing	136
Race, n (%)
Other	1,572 (25%)
White	4,700 (75%)
Missing	783
Ethnicity, n (%)
Hispanic or Latino	1,153 (17%)
Non-resident of the U.S.	803 (12%)
Non-Hispanic or non-Latino	4,920 (72%)
Missing	179
Donor-Recipient Sex Match, n (%)
M-M	2,658 (38%)
M-F	1,623 (23%)
F-M	1,449 (21%)
F-F	1,325 (19%)
Donor-Recipient CMV, n (%)
−/−	1,424 (20%)
−/+	1,807 (26%)
+/−	697 (9.9%)
+/+	3,082 (44%)
Missing	45
Time from Diagnosis to HCT (months), Median (Q1-Q3); [Min-Max]	7.0 (4.6–14.6); [0.4–427.7]
Missing	5
Year of HCT, n (%)
2013	114 (1.6%)
2014	216 (3.1%)
2015	397 (5.6%)
2016	530 (7.5%)
2017	764 (11%)
2018	1,004 (14%)
2019	1,169 (17%)
2020	1,358 (19%)
2021	1,503 (21%)

Abbreviations: aGVHD, acute graft-versus-host disease; ALL, acute lymphoblastic leukemia; AML, acute myeloid leukemia; CI, confidence interval; cGVHD, chronic graft-versus-host disease; HCT, hematopoietic cell transplantation; IST, immunosuppressive therapy; KM, Kaplan-Meier; MDS, myelodysplastic syndromes; MPN, myeloproliferative neoplasms; N, number; NRM, non-relapse mortality; OS, overall survival.

GVHD Incidence and Clinical Trajectories

Acute GVHD occurred early, with median onset of 1.4 months (interquartile range [IQR], 0.9–2.0) for isolated Grade II and 1.2 months (IQR, 0.8–2.2) for Grade III-IV. The 3-month cumulative incidence was 15.5% (95% confidence interval [CI], 14.7–16.4) for isolated Grade II and 6.2% (95% CI, 5.6–6.7) for Grade III-IV [Table S1]. IST-requiring cGVHD was the predominant chronic phenotype, with a 12-month incidence of 21.6% (95% CI, 20.6–22.6) and median onset at 6.2 months, compared with 6.4% (95% CI, 5.8–7.0) for mild cGVHD. At 36 months, cumulative incidence of NRM was 20.1% (95% CI, 19.1–21.0) and relapse was 28.3% (95% CI, 27.2–29.4).

More than half of the cohort (52.5%) remained free of acute or chronic GVHD. The most common discrete trajectory was IST-requiring cGVHD without preceding aGVHD (16.6%). Isolated Grade II aGVHD progressed to IST-requiring cGVHD in 7.3% of patients, while 9.0% experienced Grade II aGVHD without subsequent cGVHD. Unadjusted time-dependent overall survival and cumulative incidence curves for relapse and NRM are shown in Figure S1.

Time-Dependent Multivariable Outcomes

Using the multi-state model, estimated state-occupation probabilities for an average patient at 12 months post-transplant were 0.35 for remaining GVHD-free, 0.18 for IST-requiring cGVHD, 0.20 for relapse, and 0.14 for NRM [Figure S2].

To evaluate the Markov assumption, we examined whether prior aGVHD was associated with outcomes among patients who subsequently developed cGVHD. A history of aGVHD was associated with a higher hazard of adverse events (hazard ratio [HR], 1.41; 95% CI, 1.09–1.81; P=0.008), indicating that prior GVHD history remained prognostically relevant. Accordingly, subsequent analyses emphasized complete GVHD trajectories rather than isolated GVHD states.

In adjusted Cox models, IST-requiring cGVHD without antecedent aGVHD was associated with lower modeled relapse hazard (HR 0.74; 95% CI, 0.62–0.89; P = 0.001; q = 0.006) and overall mortality (HR 0.62; 95% CI, 0.53–0.73; P < 0.001; q < 0.001) compared with remaining GVHD-free [Table 2, Tables S2–4]. When IST-requiring cGVHD followed isolated grade II aGVHD, the inverse association with relapse was weaker and no longer significant after FDR correction (HR 0.73; 95% CI, 0.55–0.97; P = 0.031; q = 0.065). In this trajectory, the hazard of NRM increased (HR 1.57; 95% CI, 1.17–2.10; P = 0.003; q = 0.009), although overall mortality remained modestly lower (HR 0.78; 95% CI, 0.62–0.96; P = 0.022; q = 0.051).

Table 2.

Multi-state Time Dependent Multivariable Cox Models for Relapse, Non-relapse Mortality, Overall Survival

Trajectory	Person-months	Events	HR (95% CI)	p-value	FDR q
Relapse
No aGVHD → No cGVHD	100,429	1280	Ref.
No aGVHD → Mild cGVHD	11,171	83	1.00 (0.77–1.31)	0.977	0.993
No aGVHD → IST-cGVHD	32,616	162	0.74 (0.62–0.89)	0.001	0.006
Grade II aGVHD → No cGVHD	15,380	221	0.95 (0.81–1.12)	0.552	0.682
Any aGVHD → Mild cGVHD	2,304	14	0.90 (0.50–1.64)	0.737	0.815
Grade II aGVHD → IST-cGVHD	11,744	58	0.73 (0.55–0.97)	0.031	0.065
Grade III-IV aGVHD → No cGVHD	5,309	85	1.11 (0.87–1.42)	0.382	0.573
Grade III-IV aGVHD → IST-cGVHD	2,528	21	1.16 (0.72–1.85)	0.543	0.682
Non-relapse Mortality
No aGVHD → No cGVHD	100,429	913	Ref.
No aGVHD → Mild cGVHD	11,171	48	1.50 (1.08–2.09)	0.016	0.043
No aGVHD → IST-cGVHD	32,616	114	1.11 (0.88–1.40)	0.382	0.573
Grade II aGVHD → No cGVHD	15,380	119	0.96 (0.78–1.19)	0.736	0.815
Any aGVHD → Mild cGVHD	2,304	10	1.54 (0.79–3.01)	0.203	0.388
Grade II aGVHD → IST-cGVHD	11,744	61	1.57 (1.17–2.10)	0.003	0.009
Grade III-IV aGVHD → No cGVHD	5,309	153	3.15 (2.57–3.84)	<0.001	<0.001
Grade III-IV aGVHD → IST-cGVHD	2,528	30	3.50 (2.36–5.18)	<0.001	<0.001
Overall Survival
No aGVHD → No cGVHD	116,783	1959	Ref.
No aGVHD → Mild cGVHD	12,165	113	0.89 (0.72–1.11)	0.306	0.536
No aGVHD → IST-cGVHD	35,105	227	0.62 (0.53–0.73)	<0.001	<0.001
Grade II aGVHD → No cGVHD	17,835	306	1.00 (0.88–1.14)	0.993	0.993
Any aGVHD → Mild cGVHD	2,340	21	0.86 (0.53–1.39)	0.533	0.682
Grade II aGVHD → IST-cGVHD	12,343	102	0.78 (0.62–0.96)	0.022	0.051
Grade III-IV aGVHD → No cGVHD	5,968	229	2.09 (1.79–2.44)	<0.001	<0.001
Grade III-IV aGVHD → IST-cGVHD	2,717	47	1.54 (1.13–2.11)	0.006	0.019

^*Models stratified by DRI. Cause-specific Hazard ratios reported for relapse and NRM. Multi-state model (counting process). Reference: “No aGVHD to No cGVHD”. Grade II & III-IV aGVHD to Mild cGVHD collapsed into “Any aGVHD to Mild cGVHD” due to limited sample.

Abbreviations: aGVHD, acute graft-versus-host disease; CI, confidence interval; cGVHD, chronic graft-versus-host disease; DRI, Disease Risk Index; HR, hazard ratio; IST, immunosuppressive therapy; FDR q: False Discovery Rate-adjusted p-value.

In contrast, trajectories involving grade III-IV aGVHD were consistently associated with markedly higher NRM hazards, regardless of subsequent cGVHD course. Compared with remaining GVHD-free, grade III-IV aGVHD without cGVHD had an HR of 3.15 (95% CI, 2.57–3.84; P < 0.001; q < 0.001), and progression to IST-requiring cGVHD did not mitigate this risk (HR 3.50; 95% CI, 2.36–5.18; P < 0.001; q < 0.001). These trajectories were also associated with higher overall mortality, with HRs of 2.09 (95% CI, 1.79–2.44; P < 0.001; q < 0.001) and 1.54 (95% CI, 1.13–2.11; P = 0.006; q = 0.019), respectively.

Formal interaction testing between IST-requiring cGVHD and all baseline covariates showed no significant interactions for relapse (all q > 0.20), indicating that the association was consistent across the full set of pre-specified characteristics, including disease type, disease risk index (DRI), donor type, graft source, HCT-CI, and conditioning intensity. We also evaluated whether this association differed across transplant eras (<2017 vs ≥2017), given changes in GVHD management after 2017 including the advent of post-transplant cyclophosphamide, and the emerging use of FDA-approved therapies including ruxolitinib, belumosudil and ibrutinib.^1–4 The relapse-protective association was present in both eras (HR 0.66 pre-2017 vs 0.77 post-2017), and the interaction was not statistically significant (p = 0.136).

Secondary analyses using standard time-dependent Cox regression, the dynamic landmark supermodel, and fixed 6- and 12-month landmarks yielded similar patterns [Tables S5–9]. For example, in the standard time-dependent Cox model, IST-requiring cGVHD remained associated with a lower relapse hazard (HR, 0.76; 95% CI, 0.65–0.89; P<0.001; q=0.002) and lower hazard of death (HR for OS, 0.81; 95% CI, 0.73–0.91; P<0.001; q=0.002), while Grade III-IV aGVHD consistently showed elevated NRM associations across all frameworks.

Multi-State Random Survival Forest Validation

The MS-RSF modeled the post-transplant course using four clinical states and five predefined transitions. Permutation-based variable importance (VIMP) demonstrated clear shifts in the predictors of relapse and NRM across disease states [Figure S3]. Before GVHD onset, relapse was most strongly associated with baseline disease biology (DRI, VIMP = 0.035), whereas early NRM was associated with recipient age (VIMP = 0.031). After GVHD developed, GVHD type and grade became the dominant predictors of subsequent outcomes, exceeding the contribution of all baseline covariates. GVHD subtype was the strongest predictor of post-GVHD NRM (VIMP = 0.099) and post-GVHD relapse (VIMP = 0.026), while donor type contributed minimally. These state-dependent shifts - where baseline disease and recipient age initially influence outcomes but are later superseded by the immunologic GVHD phenotype - were further supported by Shapley Additive Explanation (SHAP) value distributions [Figure 1].

Figure 1: SHapley Additive exPlanations (SHAP) Summary Plots of Feature Importance and Directionality Across Clinical Transitions.

This figure displays the association of baseline covariates and GVHD phenotype with the model-predicted 24-month cumulative hazard for four multi-state transitions. Variables on the y-axis are ordered by their overall contribution to model predictions. Each point represents an individual patient-level prediction. The x-axis shows the SHAP value, reflecting the direction and magnitude of a feature’s contribution to the predicted hazard: points to the right of zero indicate higher predicted hazard, and points to the left indicate lower predicted hazard. Point color corresponds to the underlying clinical value (e.g., older age, higher Disease Risk Index [DRI], or more severe GVHD shown in yellow/orange; lower values in dark purple). Prior to GVHD onset (top row), DRI contributes most strongly to predicted relapse hazard, while recipient age contributes most strongly to predicted NRM. Following GVHD onset (bottom row), GVHD type/grade becomes the dominant model-based contributor to predicted relapse and NRM relative to baseline covariates.

Post-GVHD Competing-Risks and Absolute Benefit

MS-RSF–derived state-probability curves demonstrated divergent post-GVHD trajectories across GVHD phenotypes [Figure 2]. Grade III-IV aGVHD showed rapid early NRM. IST-requiring cGVHD showed the lowest model-predicted relapse incidence and the most sustained OS profile. Mild cGVHD and isolated Grade II aGVHD showed intermediate patterns.

Figure 2. Stacked Cumulative Incidence of Competing Post-GVHD Events.

Population-averaged stacked area plots derived from the multi-state random survival forest illustrate the combined modeled probabilities of relapse (red) and non-relapse mortality (blue) following GVHD onset. The total height of the shaded region at each time point represents the model-predicted probability of experiencing either event.

Partial-dependence cumulative incidence curves were directionally consistent with these observations [Figure 3]. IST-requiring cGVHD had the lowest model-predicted relapse incidence (9.9% at 12 months, 15.0% at 36 months) and comparatively modest modeled NRM (8.2% at 12 months, 12.4% at 36 months). Grade III–IV aGVHD showed a steep early rise in modeled NRM, reaching 27.5% at 12 months and 31.2% at 36 months. These patterns were similar across haploidentical and MMUD donor subtypes [Figure S4].

Figure 3: Population-Averaged Marginal Cumulative Incidence of Relapse and Non-Relapse Mortality Following GVHD Onset.

Cumulative incidence functions were estimated using a discrete Aalen–Johansen approach applied to individual survival predictions from the multi-state random survival forest (MS-RSF). To avoid immortal-time bias and sparse-data artifacts from staggered entry, time zero is defined as the exact date of GVHD onset (gap-time). Curves represent population-averaged marginal probabilities, standardizing baseline covariate distributions across GVHD subtypes.

To generate model-based absolute risk estimates, we simulated hypothetical trajectories for patients alive and GVHD-free at 6 months post-transplant. Under simulated trajectories with IST-requiring cGVHD onset at this 6-month landmark, the model-predicted 36-month relapse incidence was lower than for trajectories remaining GVHD-free (15.0% vs. 26.0%), accompanied by a modest modeled difference in NRM (12.4% vs. 7.2%) [Figure 4].

Figure 4: Model-predicted cumulative incidence of relapse and non-relapse mortality following a 6-month landmark, comparing patients who developed IST-requiring chronic GVHD versus those who remained GVHD-free.

Model-predicted cumulative incidence functions (CIFs) for relapse (left panel) and non-relapse mortality (NRM; right panel) were generated under two hypothetical post-landmark trajectories: remaining GVHD-free versus developing IST-requiring chronic GVHD at the 6-month post-transplant landmark. For the “No GVHD” group (blue), conditional CIFs were derived from transitions 1 to 3 (relapse) and 1 to 4 (NRM), conditioning on event-free survival beyond 6 months. For the “IST-requiring cGVHD” group (red), CIFs were derived from transitions 2 to 3 (relapse after GVHD) and 2 to 4 (NRM after GVHD), assuming GVHD onset at the landmark. Solid lines represent population-averaged model predictions; shaded ribbons denote 95% confidence intervals reflecting across-patient variability.

Heterogeneity of GVL Benefit and Response Phenotypes

Simulation-based moderation analysis was performed to assess whether the magnitude of the modeled association between IST-requiring cGVHD and relapse varied across patients [Figure S5]. Across the cohort, 92% of individuals showed a lower model-predicted 24-month relapse incidence under the IST-requiring cGVHD scenario compared with the mild cGVHD scenario (mean absolute modeled difference, 4.5%; 95% CI, 4.4–4.6). The magnitude of this modeled difference varied by disease-risk features: patients with high or very-high DRI had the largest divergence in predicted relapse incidence (mean 7.7%; 95% CI, 7.3–8.0), whereas those with low or intermediate DRI had smaller divergences (mean 3.5%). Disease-specific patterns were similar, with ALL showing the greatest modeled difference (mean 5.7%) and AML showing a mean difference of 4.4% [Table S10].

To characterize patterns of heterogeneity, longitudinal k-means clustering was applied to the individual model-based hypothetical relapse-difference trajectories. This analysis identified groups representing distinct magnitudes of divergence in predicted relapse probabilities: large-difference (n = 561, 18%), moderate-difference (n = 1,318, 43%), and small-difference (n = 1,220, 39%) groups [Figure S6]. At 24 months post-GVHD onset, the large-difference group had a mean absolute modeled relapse-probability difference of 9.5% with minimal accompanying difference in modeled NRM (0.27%), whereas the small-difference group had a mean modeled relapse-probability difference of 1.5% and a net difference of 2.2%. Patients in the large-difference group were more likely to be younger, receive myeloablative conditioning, and have high or very-high DRI disease (47% vs. 4.5% in the small-difference group; P < 0.001) [Table S11]. Importantly, no subgroup showed a modeled increase in relapse probability under the IST-requiring cGVHD scenario, indicating that the direction of the association was consistent across the population.

Discussion

In this registry analysis of 7,055 patients undergoing PTCy-based haploidentical or MMUD HCT, IST-requiring cGVHD was independently associated with lower relapse and improved OS. This association was phenotype-specific: neither aGVHD of any grade nor mild cGVHD demonstrated a consistent association with relapse across analytic frameworks.

Using MS-RSF counterfactual simulations, we translated these relative associations into absolute risk differences. Patients who developed IST-requiring cGVHD at 6 months post-transplant had an estimated 11% lower 36-month relapse incidence compared with remaining GVHD-free, accompanied by a 5% higher NRM. Longitudinal clustering of individual counterfactual relapse-difference trajectories identified groups with differing magnitudes of predicted relapse reduction – with one group, typically younger patients with high-risk disease, exhibiting the largest predicted relapse differences with minimal difference in predicted NRM.

Our findings complement prior PTCy-based haploidentical transplantation studies that did not observe a clear association between GVHD and relapse.^7,8 Differences between studies likely reflect variation in cohort composition, GVHD phenotyping, graft source, and analytic structure rather than differences in underlying biology. Previous analyses included focused disease subsets - such as AML in remission (n=805)⁸ or refractory disease (n=670)⁷ - and used Seattle limited/extensive criteria for cGVHD classification, which do not distinguish IST-requiring from milder cGVHD phenotypes. In contrast, our larger cohort and time-updated phenotyping allowed evaluation of more granular GVHD states, including IST-requiring cGVHD as a distinct category.

The multi-state, continuous-time framework modeled GVHD as an evolving sequence, distinguishing de novo IST-requiring cGVHD from cGVHD following antecedent aGVHD. Markov assumption testing indicated that prior aGVHD carried prognostic information among patients who later developed cGVHD, supporting trajectory-based modeling rather than isolated state comparisons.

Across analytic frameworks, the direction of associations for IST-requiring cGVHD was consistent, although the magnitude varied. This variability is expected because fixed-landmark analyses condition on survival and GVHD status at the landmark, thereby excluding patients who relapse, die, or develop GVHD beforehand. In our cohort, 1,949 patients (29%) were excluded before the 6-month landmark and 2,682 (40%) before the 12-month landmark, removing a substantial proportion of early relapse and NRM events as well as most early IST-requiring cGVHD. As a result, the fixed-landmark models had smaller risk sets (N = 4,795 at 6 months; N = 4,062 at 12 months) and fewer relapse events (974 and 535, respectively), which naturally attenuated the estimated associations. In contrast, the time-dependent Cox models, dynamic landmark supermodel, and multi-state trajectory analysis incorporated evolving GVHD states without conditioning on survival to a fixed time point, preserving the full cohort (N = 6,744) and the complete set of relapse (n = 1,924) and NRM (n = 1,448) events. These frameworks therefore produced more stable estimates. Taken together, differences in effect size reflect the expected consequences of landmark conditioning rather than discordant biological signals.

Interpretation of mild cGVHD requires caution due to its limited sample size (361 patients) and short duration (median 5.9 months in state), which contributed to model-dependent variability. Mild cGVHD did not show a reproducible association with outcomes across frameworks, reinforcing that IST-requiring cGVHD is the phenotype most consistently associated with clinically meaningful GVL activity.

This study has limitations. The retrospective design precludes causal inference. Because the registry captures only the peak GVHD severity, we could not model severity progression within aGVHD or within cGVHD (e.g., Grade II to Grade III-IV, or mild to IST-requiring), although transitions between acute and chronic GVHD states were modeled. As a result, some patients classified as milder grades of GVHD may have died before developing more severe phenotypes. Additionally, granular details regarding organ involvement, dose and tapering of immunosuppression, and specific GVHD therapies was not available. The IST-requiring cGVHD category likely encompasses heterogeneous clinical courses, and the absence of longitudinal data on GVHD resolution limits evaluation of post-resolution trajectories. Future studies incorporating detailed organ involvement, disease severity, and immunosuppressive treatment data will be important to further refine the relationship between chronic GVHD phenotypes and transplant outcomes. Finally, because the cohort consisted exclusively of mismatched donor transplants, generalizability to fully matched donor setting remains uncertain.

Conclusion

In this large PTCy-treated mismatched donor cohort, IST-requiring cGVHD was the only GVHD phenotype reproducibly associated with lower relapse incidence, whereas severe aGVHD was uniformly toxic. These findings are consistent with heterogeneity in alloreactivity patterns and suggest that strategies aimed at globally eliminating all GVHD forms may risk suppressing a phenotype linked to sustained disease control. Approaches that continue to reduce toxic acute GVHD while preserving beneficial chronic alloreactivity may offer a more effective balance between morbidity and durable GVL.

Data Sharing and Availability:

The deidentified individual participant data and a data dictionary for the dataset analyzed in this study are publicly available for immediate download from the Center for International Blood and Marrow Transplant Research (CIBMTR) at https://cibmtr.org/CIBMTR/Resources/Publicly-Available-Datasets. Access is governed by the CIBMTR’s “Terms of Use and Conditions”; downloading the dataset implies agreement to these terms.

Code availability:

The R script used for the Multi-State Random Survival Forest (MS-RSF) modeling and moderation analysis is publicly available on Zenodo at: https://doi.org/10.5281/zenodo.19243242.