A molecular-based risk score for predicting leukemia-free survival in adult AML patients undergoing Allo-HSCT

Summary

Allogeneic hematopoietic stem cell transplantation (allo-HSCT) remains the cornerstone of curative therapy for acute myeloid leukemia (AML), yet precise molecular prognostic tools are currently insufficient. This study developed a prognostic model, AML-PRSS, integrating genomic and clinical factors from 389 adult AML patients undergoing their first allo-HSCT between 2013 and 2021. Seven genetic mutations significantly associated with leukemia-free survival (LFS) were categorized as favorable (DNMT3A, CEBPA bZIP domain, NPM1 without FLT3-ITD or with FLT3-ITD plus tyrosine kinase inhibitors), unfavorable (NRAS and GATA2), and high-risk (TP53 and U2AF1). Multivariate analysis identified molecular risk, cytogenetic risk, pre-transplant disease status, age, and hematopoietic cell transplant-comorbidity index (HCT-CI) score as independent predictors of LFS. AML-PRSS stratified patients into four risk groups with stepwise increasing hazard of LFS failure. Validation in an independent multi-center cohort of 266 patients confirmed robust predictive accuracy, highlighting AML-PRSS as an effective tool for personalized prognostication and clinical decision-making.

Graphical abstract

Highlights

• •AML-PRSS was developed to predict prognosis for AML patients undergoing allo-HSCT
• •AML-PRSS integrates genomic, clinical, and transplant factors into a unified system
• •AML-PRSS stratifies patients into four risk groups with distinct survival outcomes
• •AML-PRSS enables precise risk assessment and individualized treatment decisions

Hematology; Bioinformatics

Introduction

Acute myeloid leukemia (AML) is a heterogeneous group of clonal disorders arising from chromosomal and/or genomic alterations in hematopoietic stem/progenitor cells.^1,2,3 Although several new therapeutic modalities—including targeted therapy, epigenetic therapy, immunotherapy, and combination approaches—have been incorporated into the treatment of AML and have improved the survival of AML patients,^4,5,6 allogeneic hematopoietic stem cell transplantation (allo-HSCT) remains to be the fundamental curative treatment modality,⁷ particularly in patients classified as intermediate- or high-risk categories.

Prognostic evaluation for allo-HSCT recipients plays an essential role in precise risk stratification and informed decision-making throughout the transplantation process, ultimately leading to improved outcomes. Factors contributing to allo-HSCT prognostication primarily fall into three categories: disease-related clinical variables, transplant-specific factors, and individual parameters.⁸ To date, several prognostic models for AML patients undergoing allo-HSCT have been reported. The hematopoietic cell transplant-comorbidity index (HCT-CI) was one of the earliest models and remains the most widely used tool to predict the risk of non-relapse mortality (NRM) following allo-HSCT.⁹ Another useful tool, the disease risk index (DRI) and its refined version (R-DRI),^10,11 primarily focus on disease-related characteristics, particularly disease type and disease status at transplantation, which limits their predictive power due to the evolving complexity of AML genomics. With the increasing availability of next-generation sequencing (NGS) in clinical and research settings, significant progress has been achieved in understanding genomic profiles.^12,13 Molecular studies have led to the development of the risk stratification model proposed by the European LeukemiaNet (ELN) group,¹⁴ which incorporates updated genetic abnormalities.^15,16 However, this model is not specifically designed for HSCT recipients and has shown limited prognostic utility in the context of allo-HSCT.^17,18,19 Based on these risk scores, efforts have been made to integrate, validate, or modify them to establish new prognostic models.^{17,18,19,20,21,22,23,24,25} However, these combined models have not demonstrated significantly improved performance compared to single models, with a maximum predictive power of approximately 0.72 based on C-statistic estimate.^21,22,24

In this study, we first established a transplant-specific molecular panel linked to prognosis and then developed a comprehensive prognostic risk scoring system (AML-PRSS) that integrates clinical, genomic, and transplant-related factors for AML patients undergoing allo-HSCT, providing an effective tool for evaluating the prognosis of AML patients before transplantation.

Results

Clinical and molecular characteristics in the training cohort

The baseline characteristics of the 389 patients in the training cohort are presented in Table 1. The median time from diagnosis to transplant was 174 days (range, 47–1,892 days), the median time to relapse after transplant was 232 days (range, 28–2,065 days), and the median follow-up time for survivors was 1,238 days (range, 590–4,055 days).

Table 1.

Cohort characteristics

Characteristic	Patients
	n	%
Full cohort	389	100
Age, median (range)	42 (18–71)
18–39	177	45.5
40–60	173	44.5
>60	39	10.0
Sex
Female	175	45.0
Male	214	55.0
Type of AML (clinically defined)
De novo	326	83.8
Secondary	63	16.2
Cytogeneticsa
Favorable	62	15.9
Intermediate	264	67.9
Adverse	63	16.2
2022 ELN risk group
Favorable	98	25.2
Intermediate	184	47.3
Adverse	107	27.5
First induction therapy
Intensive chemotherapy	342	87.9
Non-intensive chemotherapy	47	12.1
Induction result
CR1	230	59.1
>CR1	40	10.3
Non-CR	119	30.6
Pre-transplant disease status
CR, MRD- (flow/molecular)	138	35.5
CR, MRD+ (flow/molecular)	132	33.9
NCR	60	15.4
Relapse	59	15.2
R-DRI groupb
Low	48	12.3
Intermediate	193	49.6
High	132	33.9
Very High	16	4.1
HCT-CI score
0–1	274	70.4
>1	109	28.1
Missing	6	1.5
Donor type
Matched related	69	17.7
Matched unrelated	65	16.7
Haploidentical	255	65.6
Donor to recipient gender
Female to Male	84	21.6
Others	305	78.4
ABO compatibility
ABO-matched	205	52.7
ABO-mismatched	184	47.3
Conditioning regimen
Myeloablative	308	79.2
Reduced intensity	81	20.8
GVHD prophylaxis
LD-ATGc+MTX+CNI+MPA (Matched related/unrelated)	101	25.9
PTCyd+CNI+MPA (Haploidentical)	35	9.0
HD-ATGe+MTX+CNI+MPA (Haploidentical)	71	18.3
LD-ATG/PTCyf+CNI+MPA (Haploidentical+Matched unrelated)	182	46.8
Graft
PBSCs	332	85.3
PBSCs combined with UCB	57	14.7
MNC cell dose, x108/kg, median (IQR)	15.2 (10.5–21.0)
CD34+ cell dose, x106/kg, median (IQR)	9.7 (6.5–13.3)
CD3+ cell dose, x108/kg, median (IQR)	3.2 (2.3–4.2)
UCB CD34+cell dose, x104/kg, median (IQR)	7.4 (4.4–10.1)
UCB nuclear cell dose, x107/kg,median (IQR)	2.3 (1.9–3.1)

ELN, European LeukemiaNet; CR, complete remission; NCR, no complete remission; MRD, minimal residual disease; R-DRI, refined disease risk index; HCT-CI, hematopoietic cell transplant comorbidity index score; GVHD, graft-versus-host disease; LD-ATG, low dose anti-thymocyte globulin; HD-ATG, high dose anti-thymocyte globulin; PTCy, post-transplantation cyclophosphamide; MTX, methotrexate; CNI, calcineurin inhibitor; MPA, mycophenolic acid; PBSCs, peripheral blood stem cells; UCB, umbilical cord blood; MNC, mononuclear cells.

^aCytogenetic risk groups are categorized according to 2022 ELN risk classification.

^bR-DRI was reported according to the criteria of Armand et al. in 2014.

^cATG dose: 1.5–2 mg/kg/d for 3 days.

^dPTCy dose: 50 mg/kg/d for 2 days.

^eATG dose: 2.5 mg/kg/d for 4 days.

^fATG dose: 2.5 mg/kg/d for 2 days; PTCy dose: 50 mg/kg/d for 1 day.

As identified by targeted NGS using the AML screening panel (Table S1) of diagnostic (pre-induction) bone marrow samples, a total of 343 patients (88.2%) harbored at least one gene mutation. Additionally, 30 mutated genes were detected in at least two patients (>0.5%), with the frequency of each mutation illustrated in Figure 1. Genes are organized into broad functional classes to enhance readability and navigation.^26,27 Mutations in genes associated with signaling pathways and transcription factors were observed in 162 (41.6%) and 108 (27.8%) patients, respectively. This was followed by mutations in genes involved in epigenetic regulation (133 patients, 34.2%), tumor suppression (74 patients, 19.0%), chromatin modification (46 patients, 11.8%), and RNA splicing (34 patients, 8.7%). TP53 mutations and FLT3-ITDs without concurrent NPM1 mutations, both indicative of adverse molecular risk,^28,29,30 were identified in 26 (6.7%) and 35 (9.0%) patients, respectively. Tyrosine kinase inhibitors (TKIs) were administered post-transplantation to 28 patients with FLT3-ITDs.

Figure 1.

Genetic mutation profile at initial diagnosis

Each patient is represented in an individual column, while genes are listed in rows. Mutations are organized by gene function. Colors indicate the presence of mutations, with distinct hues representing functional categories. The proportion of patients in the cohort with each alteration is reported on the right.

Transplant outcomes

Four patients experienced primary graft failure. Among successfully engrafted patients, the median times for neutrophil and platelet engraftment were 13 and 14 days post-transplant, respectively. Within 180 days post-transplant, the cumulative incidences (CIs) of acute GVHD (aGVHD) and grade 2–4 aGVHD were 27.5% and 12.3%, respectively. Within two years, the CIs of overall chronic GVHD (cGVHD) and moderate to severe cGVHD were 33.2% and 14.5%, respectively.

For the entire cohort, the CIs of NRM were 9.8% [7.1%–13.2%] at 1 year and 12.1% [9.2%–16.0%] at 2 years, and the CIs of relapse were 14.9% [11.6%–18.9%] at 1 year and 18.0% [14.4%–22.3%] at 2 years. At 2 years post-transplant, overall survival (OS), leukemia-free survival (LFS), and graft-versus-host disease-free survival (GRFS) were 73.4% [68.7%–77.5%], 69.9% [65.2%–74.3%], and 50.3% [46.9%–54.1%], respectively (Table S2; Figure S2).

Genetic determinants of post-transplant prognosis

We first developed a molecular risk model based on diagnostic baseline genetic characteristics. 18 genes mutated in at least 12 patients (>3%) were enrolled in the prognostic analysis. Univariable survival analysis (Table S3) revealed that genetic mutations associated with prolonged LFS and OS included DNMT3A, NPM1 without FLT3-ITD or with FLT3-ITD plus TKI, and bZIP domain of CEBPA. In contrast, mutations correlated with inferior LFS and OS included ASXL1, NRAS, TP53, GATA2, U2AF1, and WT1. Multivariate analysis (Table 2) indicated that most significant mutations were primarily related to relapse. Notably, NRAS mutation was specifically associated with NRM (sHR = 1.82, p = 0.041), while U2AF1 mutation was linked to both relapse (sHR = 2.84, p = 0.024) and NRM (sHR = 3.41, p = 0.013). Based on the total scores assigned by hazard ratios (Tables S4 and S5), we established a molecular risk panel for LFS, which includes: favorable mutations in DNMT3A, the bZIP domain of CEBPA, NPM1 without FLT3-ITD or with FLT3-ITD plus TKI; unfavorable mutations in NRAS and GATA2; and high-risk mutations in TP53 and U2AF1. Based on the molecular risk panel, patients were stratified into three groups with significantly distinct outcomes, primarily driven by the risk of relapse, as shown in Figure 2.

Table 2.

Multivariate analysis of diagnostic genetic mutations

Variables	LFS			Relapse			NRM
	HR	95% CI	P	sHR	95% CI	P	sHR	95% CI	P
DNMT3A	0.62	0.34–1.02	0.053	0.59	0.26–1.07	0.065	0.70	0.30–1.61	0.297
NPM1+, FLT3-ITD-/FLT3-ITD+ with TKI	0.55	0.26–0.97	0.042	0.41	0.13–0.92	0.039	0.92	0.37–2.31	0.762
CEBPA, bZIP	0.52	0.25–0.94	0.032	0.21	0.02–0.74	0.024	0.95	0.33–2.67	0.785
NRAS	1.52	1.02–2.50	0.048	1.30	0.67–2.52	0.340	1.82	1.06–3.27	0.041
TP53	3.74	2.26–6.17	0.000	4.08	2.12–7.86	0.000	2.07	0.87–4.96	0.101
GATA2	1.95	1.12–3.37	0.018	2.56	1.28–5.90	0.026	1.22	0.53–2.84	0.548
U2AF1	3.15	1.89–6.04	0.001	2.84	1.15–7.02	0.024	3.41	1.29–9.04	0.013

sHR: subdistribution hazard ratio from competing-risk regression model.

Figure 2.

Post-transplant outcomes by molecular risk group in the training cohort

(A) Leukemia-free survival (LFS).

(B) Overall survival (OS).

(C) Cumulative incidence of relapse (CIR).

(D) Non-relapse mortality (NRM).

Clinical characteristics associated with post-transplant prognosis

Non-genetic factors, which have been shown or have the potential to impact the prognosis of AML patients following allo-HSCT, were also evaluated in this study.⁸ Univariable analysis identified several pre-transplant factors significantly associated with LFS and OS, including age, AML type, cytogenetic risk group (ELN2022 classification), induction effect, pre-transplant disease status (including MRD status), and HCT-CI score (Table S6). Multivariate Cox analysis identified independent pre-transplant non-genetic factors associated with inferior LFS, including age>60 (HR = 1.84, p = 0.027), adverse cytogenetics (HR = 3.45, p < 0.001), pre-transplant disease status of non-CR (HR = 4.13, p < 0.001), and HCT-CI>1 (HR = 1.54, p = 0.047) (Table S7). Competing-risk analysis for relapse and NRM indicated that age and HCT-CI were specifically linked to NRM, while karyotype and pre-transplant disease status were associated with both outcomes (Table S8).

Development and performance of the AML-PRSS model

The significant clinical and genetic variables previously identified were incorporated into a multivariable Cox frailty model. The final model included molecular risk, cytogenetics, pre-transplant disease status, age, and the HCT-CI score. Each variable was assigned a weighted score based on HR, and the prognostic scoring system (AML-PRSS) was developed according to the sum scores of these five risk factors (Tables 3 and 4).

Table 3.

Five significant pre-transplant variables for LFS included in the AML-PRSS model

Variables	Subgroups	HR	95% CI	p value	Score assigned
Molecular risk	Low	Ref.			0
	Intermediate	2.15	1.28–3.63	0.004	1.5
	High	2.86	1.63–5.02	0.000	2
Cytogenetics	Favorable	Ref.			0
	Intermediate	1.86	1.01–3.46	0.048	1
	Adverse	3.40	1.71–6.76	0.000	2
Pre-transplant disease status	CR, MRD-	Ref.			0
	CR, MRD+	1.28	0.87–1.99	0.319	0.5
	Non-CR	3.31	2.06–5.33	0.000	2
Age	≤60	Ref.			0
	>60	1.78	1.06–3.00	0.030	1
HCT-CI	0–1	Ref.			0
	>1	1.46	1.02–2.28	0.043	0.5

Table 4.

Prognostic risk groups and the corresponding total scores in the AML-PRSS model

Risk group (risk score)	Sum score	N	HR	95% CI	Total number
Low (0–1.5)	0–1	77	Ref.		112
	1.5	35	0.93	0.32–3.14
Intermediate (2–3.5)	2	21	2.22	0.74–7.05	170
	2.5	77	2.95	1.16–7.47
	3	33	3.31	1.49–7.38
	3.5	39	3.38	1.39–8.27
High (4–4.5)	4	21	6.35	2.51–16.10	53
	4.5	32	6.80	2.93–15.77
Very High (≥5)	5	27	9.93	4.34–22.71	54
	5.5	10	10.94	3.96–30.20
	≥6	17	16.76	7.00–40.10

Patients were categorized into four risk groups: low (score 0–1.5, n = 112), intermediate (score 2–3.5, n = 170), high (score 4–4.5, n = 53), and very high (score ≥5, n = 54). Throughout follow-up, the risk of LFS failure increased progressively across risk groups. Compared with the low-risk group, hazard ratios were 3.28 (95% confidence interval [CI]: 1.71–6.30) for the intermediate-risk group, 6.98 (3.46–14.09) for the high-risk group, and 12.42 (6.36–24.26) for the very high-risk group patients (all p < 0.001; Table S9; Figure 3). Consistent with these findings, 3-year LFS rates declined markedly from 90% (95% CI: 83–94%) in the low-risk group to 71% (64–78%), 49% (35–62%), and 26% (15–39%) in the intermediate-, high-, and very high-risk groups, respectively (p < 0.001; Table S10). The overall risk model reflected the patients’ composite risk of relapse (Figure 3C) and NRM (Figure 3D).

Figure 3.

Prognostic stratification of AML patients in the training cohort based on the AML-PRSS model

(A) Leukemia-free-survival (LFS).

(B) Overall survival (OS).

(C) Cumulative incidence of relapse (CIR).

(D) non-relapse mortality (NRM).

Validation of the AML-PRSS model

The baseline characteristics and transplant outcomes of 266 patients in the validation cohort are presented in Tables S11 and S12, respectively. The validation set had similar patient characteristics to the training cohort, with no variables showing statistical significances.

We first validated the molecular risk panel, which demonstrated significant prognostic ability for LFS and OS, as well as a notable predictive impact on relapse and NRM, as shown in Figure S3. Subsequently, according to our AML-PRSS model, patients were successfully stratified for LFS, OS, relapse and NRM, as shown in Figure 4 and Table S13. Compared to the low-risk group (n = 49), patients in the intermediate- (n = 146, HR = 2.36, p = 0.092), high- (n = 41, HR = 6.08, p = 0.001), and very high-risk (n = 30, HR = 15.66, p < 0.001) groups had significantly lower LFS. The corresponding 2-year LFS for patients in the low-, intermediate-, high-, and very high-risk groups was 89%, 83%, 59%, and 30% (p < 0.001), and 2-year OS was 96%, 88%, 70%, and 33% (p < 0.001), respectively (Table S14). The overall risk model also represented patients’ composite risk of relapse (Figure 4C) and NRM (Figure 4D), further confirming the model’s accuracy and stability.

Figure 4.

Prognostic stratification of AML patients in the validation cohort based on the AML-PRSS model

(A) Leukemia-free-survival (LFS).

(B) Overall survival (OS).

(C) Cumulative incidence of relapse (CIR).

(D) Non-relapse mortality (NRM).

The C-indexes for the AML-PRSS, R-DRI, DRI, and ELN2022 genetic risk models were 0.75 (95% CI: 0.69–0.80), 0.67 (95% CI: 0.60–0.74), 0.66 (95% CI: 0.60–0.73), and 0.64 (95% CI: 0.57–0.71), respectively. The 1-year AUCs for these models were 0.79 (95% CI: 0.73–0.85), 0.71 (95% CI: 0.63–0.79), 0.70 (95% CI: 0.62–0.77), and 0.68 (95% CI: 0.59–0.76). The 2-year AUCs were 0.78 (95% CI: 0.72–0.83), 0.69 (95% CI: 0.61–0.76), 0.68 (95% CI: 0.61–0.75), and 0.65 (95% CI: 0.57–0.73) (Table 5). Compared to the R-DRI and ELN2022 genetic risk stratification models, the AML-PRSS model showed significantly superior discriminative ability in predicting prognosis after allo-HSCT.

Table 5.

Comparisons of the performance of the AML-PRSS with other risk models in the validation cohort

Models	1-year AUC (95% CI)	2-year AUC (95% CI)	C-index (95% CI)	p value
PRSS	0.789 (0.726–0.852)	0.775 (0.718–0.834)	0.748(0.688–0.804)	0.000
R-DRI	0.711 (0.633–0.789)	0.687 (0.614–0.760)	0.671(0.601–0.738)	0.113
DRI	0.697 (0.621–0.773)	0.675 (0.605–0.746)	0.661(0.601–0.726)	0.201
ELN 2022	0.676 (0.592–0.757)	0.646 (0.567–0.725)	0.638(0.568–0.710)	Ref.

AUCs were calculated from the area under time-dependent ROC curves associated with mortality. The Harrell C-statistics were obtained using the STATA stcox postestimation command. The p value reflects the difference in C-index.

Discussion

This study developed and validated a comprehensive transplant-specific prognostic scoring system, AML-PRSS, for AML patients undergoing allo-HSCT. It demonstrates significant prognostic power for LFS, and a notable predictive impact on OS, relapse, and NRM. This system is potentially applicable to all allo-HSCT patients, regardless of age, AML subtype, conditioning regimen, donor type, or graft source. The AML-PRSS incorporates five pre-transplant key variables: clinical factors (age, pre-transplant disease status), a transplant comorbidity factor (HCT-CI score), and genomic factors (molecular risk and cytogenetics). This integration is particularly valuable given the heterogeneity and complexity of AML.

Cytogenetics, molecular risk, and pre-transplant disease status were closely associated with both relapse and NRM in the AML-PRSS model, underscoring the pivotal role of genomics and disease status prior to transplantation in determining post-transplant outcomes. The cytogenetic risk category has long been recognized as one of the most powerful prognostic factors, establishing cytogenetics as a fundamental component of any risk stratification.^31,32,33 The cytogenetic risk classification in this study followed the ELN2022 criteria, which is generally consistent with other major stratification systems.³³

Pathological mutations have been identified closely linked to outcomes in AML patients undergoing allo-HSCT;^34,35,36 however, a comprehensive molecular panel specifically for transplantation is still lacking. The molecular risk panel in the AML-PRSS model demonstrated significant prognostic value for LFS and OS, and predictive value for relapse and NRM. In contrast to ELN2022, this study additionally identified DNMT3A as a favorable mutation, and NRAS and GATA2 as unfavorable prognostic mutation markers. Prognosis of DNMT3A mutations has long been controversial, and the benefits observed in our study potentially due to the increasing use and positive therapeutic effects of hypomethylating agents.^37,38 GATA2 mutation posed a considerable risk for relapse. As a critical regulator of normal and leukemic stem cells, GATA2 mutations are likely to disrupt normal differentiation and survival pathways in hematopoietic cells.^39,40 NRAS mutation was identified as being associated with NRM in our study. NRAS mutation typically leads to continuous activation of the RAS pathway, potentially causing immune and metabolic dysfunction, which may contribute to poor treatment responses, disruption of GVL effect, and development of GVHD.^41,42,43 Furthermore, while we identified NPM1 mutations without FLT3-ITD as favorable, we also observed that the impact of FLT3-ITD could be mitigated by TKIs. Additionally, most secondary-type mutations included in ELN, except for U2AF1, were absent in our panel. This aligns with previous studies, suggesting that most aging- or MDS-related mutations could be overcome through transplantation.^{34,44,45,46,47} The molecular risk panel facilitates individualized risk assessment, guides MRD monitoring, and provides insights into precise, molecular-targeted treatment approaches.

Pretransplant disease status is another crucial factor in the AML-PRSS model, with patients not in CR facing significantly higher risks of both relapse and NRM.^15,48,49 The prognostic impact of MRD status was also studied in this study. MRD-positive status prior to transplantation posed a higher risk of inferior LFS compared to MRD-negative status; however, this difference was not statistically significant. Increasing evidences suggest that MRD positivity is primarily a reflection of baseline genetic risk, which is established at the time of diagnosis.^{50,51,52,53,54,55} High-risk genetic mutations tend to cause MRD persistence due to their aggressive nature, inherently affecting post-transplant outcomes. Therefore, AML patients with persistent MRD positivity before transplantation may benefit from strategies targeting molecular clearance and reduction of leukemic clones, such as epigenetic regulators, molecular target inhibitors,^{4,5,56,57,58,59} and novel cellular immunotherapies,^6,60,61 pending results from ongoing clinical trials.

Notably, external validation in a multi-center cohort demonstrated that AML-PRSS maintains robust discriminative capacity, underscoring its generalizability beyond single-center settings. However, in our external validation cohort, the low- and intermediate-risk groups did not demonstrate statistically significant separation across LFS, OS, relapse, or NRM. This discrepancy likely reflects several factors: the small size and low event rate of the low-risk subgroup, which limited statistical power; minor differences in patient mix which may have attenuated the survival gap; and a shorter median follow-up, which reduced sensitivity to late relapses in lower-risk patients. Despite the lack of statistical significance, the low-risk group maintained numerically superior outcomes—higher LFS and OS rates and lower relapse and NRM incidence. We therefore preserved the four-tier schema to retain clinical granularity and enable risk-adapted approaches (e.g., standard surveillance versus consideration of maintenance therapy). Moreover, this framework supports future recalibration: larger prospective cohorts may refine cut-off values or consider merging strata if justified by robust data.

The AML-PRSS model offers several advantages over existing prognostic models. Traditional models predominantly focus on individual risk factors, resulting in limited clinical utility. In contrast, AML-PRSS integrates clinical, genomic, and transplant-related factors into a unified scoring system. This multifactorial approach directly addresses the biological heterogeneity of AML in the transplantation setting, providing more precise risk stratification compared to single-domain models. Furthermore, the AML-PRSS demonstrated robust predictive power in both training and independent validation cohorts, confirming its reliability across diverse patient populations and clinical settings. In direct comparisons, AML-PRSS outperforms the ELN2022, DRI, and R-DRI models with superior discriminative performance. This enhanced discrimination enables clinicians to confidently identify lower-risk patients for standard post-transplant monitoring, and to recognize higher-risk patients who may benefit from intensified surveillance, maintenance therapies, or enrollment in clinical trials. Finally, although the model was primarily designed to predict LFS, its strong performance also extended to OS, relapse and NRM. By integrating genomic and clinical domains into a cohesive tool—and validating it in an independent, multi-center cohort—AML-PRSS fills the critical gap left by prior models and provides a robust framework for personalized risk stratification in AML patients undergoing allo-HSCT.

Limitations of the study

Our study has several limitations. First, the retrospective design may introduce potential biases, such as selection bias and information bias, which could affect the generalizability of the findings. Notably, 66% of transplants in our cohort used haploidentical donors—reflecting center-specific practice—and this donor distribution may not mirror settings where matched related or unrelated donors are more common. Future prospective validations in multicenter cohorts with a more diverse mix of donor types would be essential to confirm the broad applicability of AML-PRSS. Second, although the AML-PRSS was validated in an independent multi-center cohort, the sample size and median follow-up were relatively modest, potentially limiting power to detect late events. Furthermore, all participants were treated within Chinese centers; prospective validation in larger, ethnically and geographically diverse cohorts with extended follow up is warranted to confirm the universal utility of AML-PRSS. Third, while molecular information was well-collected at diagnosis and prior to transplantation, it was limited during post-transplant follow-up. Although baseline genetic risk is fundamental and crucial for prognosis, dynamic molecular monitoring could provide additional insights. Additionally, our model relies solely on high-depth targeted DNA sequencing; integrating transcriptome data (e.g., RNA-seq) could uncover additional prognostic biomarkers and further refine risk stratification.

Future research should prioritize prospective validation of AML-PRSS, assess its applicability across diverse clinical settings, explore its role in guiding personalized therapeutic interventions, and consider incorporating multi-omics data—including transcriptomic, proteomic, epigenomic, and immune-profiling approaches—alongside emerging biomarkers to further enhance prognostic accuracy.

Resource availability

Lead contact

Further information requests for resources should be directed to and will be fulfilled by the lead contact, Xianmin Song (shongxm@sjtu.edu.cn).

Materials availability

This study did not generate new unique reagents.

Data and code availability

• •Data: All data associated with this study are present in the paper or the supplemental information.
• •Code: This manuscript did not generate any new code. All original R/STATA code used for data analysis will be made available upon reasonable request to the lead contact.
• •All other items: Any additional information required to reanalyze the data reported in this paper is available from the lead contact upon reasonable request.

Acknowledgments

We acknowledge all authors and patients participating in this study. This work was supported by National Key R&D Program of China (2019YFA0111000 for X.S.), a three-year development project from Shanghai Shen Kang Hospital Development Center (SHDC2020CR1012B for X.S.), the National Natural Science Foundation of China (81570148 for X.S.).

Author contributions

S.L. and H.W. collected the data, performed the data analysis, and wrote the manuscript; X.S. and L.W. designed and directed the study and revised the manuscript; X.H., F.Z., X.N., Y.D., J.W., Y.C., J.Y., Y.T., H.Q., C.H., and K.Z. contributed for the clinical data and patients’ treatment; X.W. curated the sequencing results and bioinformatic analysis. All authors reviewed and approved the final manuscript.

Declaration of interests

The authors declare no competing interests.

STAR★Methods

Key resources table

REAGENT or RESOURCE	SOURCE	IDENTIFIER
Biological samples
Training cohort (AML patients)	Shanghai General Hospital, Shanghai Jiao Tong University	Ethics Approval No.: 【2024】396; Period: Jan 2013-Dec 2021
Validation cohort (AML patients)	Multi-center collaboration consortium: •Shanghai General Hospital, Shanghai Jiao Tong University •Ruijin Hospital Affiliated to Shanghai Jiao Tong University School of Medicine •Tongji Hospital Affiliated to Tongji University School of Medicine •Xinhua Hospital Affiliated to Shanghai Jiao Tong University School of Medicine •Changhai Hospital, Naval Medical University •960th Hospital of the Chinese People’'s Liberation Army Joint Logistics Support Force	Ethics Approval No.: 【2024】396; Period: Sep 2018-Aug 2022
Critical commercial assays
NovaSeq 6000 System	Illumina, San Diego, CA	Model: NovaSeq 6000; S4 Flow Cell
Software and algorithms
STATA	StataCorp LLC, College Station, TX	Version 14.1; Website: https://www.stata.com
R Statistical Software	R Foundation for Statistical Computing, Vienna, Austria	Version 4.0.2; Website: https://www.r-project.org
GraphPad Prism	GraphPad Software, San Diego, CA	Version 10.0; Website: https://www.graphpad.com/
IBM SPSS Statistics	IBM Corporation, Armonk, NY	Version 26.0; Website: https://www.ibm.com/products
COSMIC Database	Wellcome Sanger Institute;	https://cancer.sanger.ac.uk/cosmic
ClinVar Database	National Center for Biotechnology Information (NCBI)	https://www.ncbi.nlm.nih.gov/clinvar

Experimental model and study participant details

Patients

A total of 389 eligible adult patients diagnosed with AML (excluding with PML-RARA fusion gene) at Shanghai General Hospital were included in the training cohort and underwent their first allo-HSCT between January 2013 and December 2021. To validate the prognostic system, we used an independent, multi-center external cohort of 266 adult AML patients who underwent their first allo-HSCT at one of the six participating centers (Shanghai General Hospital, Ruijin Hospital, Tongji Hospital, Xinhua Hospital, Changhai Hospital, and the 960th Hospital of the Chinese People's Liberation Army Joint Logistics Support Force) between September 2018 and August 2022. The baseline demographic and disease characteristics of the 389 patients in the training cohort and 266 patients in the validation cohort are summarized in Tables 1 and S11, respectively. All patients were followed up through outpatient clinics, hospital medical records, or telephone calls, with the follow-up endpoints set at December 31, 2023 for the training cohort and July 31, 2024 for the validation cohort. The study protocol was approved by the institutional review boards, and written informed consent was obtained from each participant or their legal guardians in accordance with the Declaration of Helsinki.

Method details

Transplant procedures

Allo-HSCT was primarily offered to AML patients in first complete remission (CR1), as well as to those with primary refractory or relapsed AML, irrespective of their remission status after salvage chemotherapy. Patients in remission typically received 1-2 cycles of consolidation therapy before proceeding to allo-HSCT. Donors were selected based on high-resolution typing for HLA-A, -B, -C, -DRB1, and -DQB1 loci, including matched related donors (≥8/10 loci), haploidentical donors (≥3/10 loci mismatched), and unrelated donors (≥8/10 loci) from the China Marrow Donor Program (CMDP). Myeloablative conditioning (MAC) was administered to patients under 55 years old with an HCT-CI score of 2 or less, whereas patients aged 55 and older or with an HCT-CI score above 2 received reduced-intensity conditioning (RIC). The details of the conditioning regimens for all patients are shown in Figure S1. All patients received peripheral blood stem cells mobilized with granulocyte colony-stimulating factor (G-CSF) as grafts, while a minority of patients concurrently received a single unrelated umbilical cord blood (UCB). The graft-versus-host disease (GVHD) prophylaxis regimens consisted of Anti-Human Thymocyte Globulin (ATG, Thymoglobuline®) and/or post-transplant cyclophosphamide (PTCy), combined with cyclosporine A (CsA) and mycophenolate mofetil (MMF), with or without methotrexate (MTX), as specified in Table 1.

Genomic data generation and processing

Genetic mutation data were obtained from diagnostic bone marrow samples (mononuclear cells) using targeted next-generation sequencing (NGS). No RNA sequencing or gene expression data were generated. We employed a standardized hematologic malignancy panel covering 78 genes recurrently mutated in AML (Table S1). Sequencing was performed using the Illumina NovaSeq 6000 platform, with average sequencing depth >1000×. Raw sequencing data underwent rigorous quality control, including alignment (BWA algorithm), variant calling (GATK pipeline, Mutect2), and manual curation to exclude sequencing artifacts and polymorphisms. Mutations with variant allele frequency (VAF) ≥2% were considered positive. Non-pathogenic SNPs (based on COSMIC, ClinVar databases) and synonymous variants were excluded from the analysis. All genetic analyses were performed at centralized clinical genetics laboratory. Molecular analysis was blinded to clinical characteristics and locked before merging with clinical data.

Definitions

AML diagnosis was based on the 2016 World Health Organization classification.⁶² Patients with isolated extramedullary disease were excluded. Pre- and post-HSCT information, as well as clinical data, were collected for all patients from our transplantation database. Cytogenetic risks were classified as favorable, intermediate, or adverse according to the ELN2022 classification.¹⁶ Measurable residual disease (MRD) was evaluated using two complementary methods: multicolor flow cytometry to detect residual leukemic blasts with a sensitivity threshold of 0.01% (10⁻⁴); and molecular quantification of disease-specific markers and fusion transcripts using real-time quantitative PCR and/or targeted NGS, with each assay validated to a lower limit of detection between 10⁻⁴ and 10⁻⁵.

The primary endpoint was leukemia-free survival (LFS), defined as the time from transplantation to relapse or death from any cause. Secondary endpoints included overall survival (OS), defined as the time from transplantation to death from any cause or censoring at the last follow-up; non-relapse mortality (NRM), defined as death without prior relapse or progression, with relapse as a competing event; and relapse, defined as any occurrence of leukemia recurrence, with NRM as a competing event. Patients were censored at the time of subsequent HSCT or at the last follow-up if alive.

Neutrophil engraftment was defined as the first of 3 consecutive days with an absolute neutrophil count greater than 0.5×10⁹/L, and platelet engraftment as the first of 7 consecutive days with an untransfused platelet count greater than 20×10⁹/L. Graft failure was defined as described in the literature.⁶³ Acute graft-versus-host disease (aGVHD) was diagnosed and graded according to the modified Glucksberg grading system,⁶⁴ and chronic graft-versus-host disease (cGVHD) was diagnosed and graded according to the 2014 National Institutes of Health (NIH) consensus criteria.⁶⁵

Quantification and statistical analysis

Development and validation of AML-PRSS

Data from the training set were used to develop the prognostic risk scoring system (AML-PRSS). Univariate cox regression analysis was used to assess the prognostic ability of all potential baseline AML-related risk factors. All significant variables (P < 0.05) identified in the univariable analysis were subsequently entered into a multivariable Cox proportional hazards model to identify independent predictive variables. Weighted points were assigned to the significant factors proportionally to their hazard ratios (Table S4). A risk score was then calculated by summing the weighted points of all significant risk factors. Based on the total scores, patients were stratified into distinct risk groups, thereby forming the AML-PRSS.

The performance of the AML-PRSS model was validated using data from an independent validation cohort. The model's discriminative ability was assessed using Harrell’s C-statistics, with confirmatory analysis performed by estimating the area under time-dependent receiver operating characteristic curves (AUCs) associated with the risk of death.⁶⁶ Finally, Kaplan-Meier survival curves were generated for the risk groups defined by the model.

Statistical analysis

OS and LFS were estimated using the Kaplan-Meier method, and differences were determined by log-rank tests. Cumulative incidences of NRM and relapse were estimated within competing risk frameworks and compared using Gray’s test. Multivariable analysis was performed using Cox proportional hazard models for OS and LFS, and Fine and Gray models for NRM and relapse. Risk groupings were derived from the results of univariable and multivariable models. All P-values were two-sided, with a significance level set at 0.05. All analyses were performed using STATA 14.1 (StataCorp, College Station, TX), R version 4.0.2, and IBM SPSS Statistics 26.0.

Published: July 7, 2025