Prognostic impact of variant allele frequency in intensively treated patients with NPM1-mutated AML: a PETHEMA study

Key Points

• •NPM1^low defines a distinct subset of NPM1mut AML enriched in secondary AML features and linked to inferior survival despite NPM1 mutation.
• •Mutation acquisition ordering and clonal architecture modulate prognosis in NPM1mut AML and may inform future risk stratification.

Visual Abstract

Abstract

NPM1-mutated acute myeloid leukemia (AML) is genetically well defined, but clinical outcomes remain heterogeneous, suggesting that quantitative clonal features may refine current risk stratification. We analyzed 688 intensively treated NPM1-mutated (NPM1mut) AML integrating variant allele frequency (VAF), mutation order, and clonal architecture inferred by PyClone and ClonEvol. Comutations were present in 97% of patients (median, 3 per case), dominated by DNMT3A (49%), FLT3-ITD (46%), TET2 (22%), and IDH2 (20%). Prognostic modeling of NPM1 VAF identified an optimal cutoff of 31.44%, defining NPM1^high and NPM1^low groups. NPM1^low correlated with splicing-related alterations and independently predicted inferior overall survival (hazard ratio [HR], 1.46; P = .037) and relapse-free survival (HR, 1.40; P = .036). Gene-specific VAF analyses revealed divergent effects across partners: high DNMT3A, FLT3-OTHER (non-ITD/TKD), KRAS, and PTPN11 burdens were adverse, whereas high IDH2 VAF was protective. Combined models showed that patients with NPM1^high and favorable comutation VAFs had the best outcomes, whereas dual unfavorable burdens conferred the poorest survival. Mutation ordering inferred from VAFs positioned NPM1 after epigenetic and splicing lesions but before signaling and transcription-factor mutations. Noncanonical orders, such as early FLT3-OTHER/TKD or WT1 before NPM1, significantly stratified outcomes. Clonal reconstruction revealed predominantly linear evolutionary trajectories (84.3%), with increased mutational burden and clonal diversity associating with inferior survival. Notably, intraclonal colocalization of NPM1 with IDH1 or TET2 was associated with improved outcomes, whereas colocalization with WT1 predicted dismal prognosis. These results demonstrate that quantitative and structural dimensions of clonality refine the biological and prognostic landscape of NPM1-mutated AML beyond mutational status alone.

NPM1-mutated acute myeloid leukemia (AML) generally has a favorable prognosis, but outcomes remain heterogeneous, suggesting that clonal features may refine risk assessment. Gil et al show that a lower NPM1-mutant disease burden identifies a distinct subset that is enriched for secondary AML features and associated with inferior survival. Mutation order and clonal composition further influence prognosis, highlighting the biological heterogeneity of NPM1-mutated AML and supporting more individualized risk stratification.

Introduction

Acute myeloid leukemia (AML) is a clinically and genetically heterogeneous malignancy driven by the stepwise accumulation of somatic mutations in hematopoietic progenitors. Although next-generation sequencing has delineated recurrent genetic subgroups with distinct phenotypic and prognostic features.^{1, 2, 3} Among these subgroups, NPM1-mutated AML represents a genetically well-defined yet biologically diverse entity. NPM1 mutations are present in approximately one-third of adult patients, define a distinct diagnostic category and typically confer favorable prognosis in the absence of FLT3-ITD or adverse cytogenetics, according to the European Leukemia Net 2022 (ELN 2022) risk stratification.^{4, 5, 6} However, substantial heterogeneity in clinical outcomes persists within NPM1-mutated AML, suggesting that factors beyond genotype contribute to disease heterogeneity.

Previous work has shown that specific comutation patterns modulate prognosis, such as the adverse impact of the NPM1:FLT3-ITD:DNMT3A triad or the potentially mitigating influence of IDH2 or PTPN11.^7,8 However, mutational status alone does not reflect the quantitative contribution of individual lesions to the leukemic clone. In contrast, variant allele frequency (VAF), as a surrogate measure of clonal dominance, captures the relative contribution of individual mutations to the leukemic burden.^9,10 Across multiple AML genotypes, higher VAFs of mutations affecting TP53, TET2, DNMT3A, or NRAS have been associated with adverse outcomes, whereas elevated GATA2 VAF has been linked to improved survival.¹¹ Beyond clonal dominance, the temporal sequence of mutation acquisition represents an additional layer of biological complexity. Comparative analysis of corrected VAFs adjusted for underlying copy number alterations, such as loss of heterozygosity or deletions, enables inference of mutation acquisition order.¹² Mutations with higher VAFs are generally interpreted as early, founding events present in most leukemic cells, whereas lower VAF lesions are assigned to later arising subclones. In NPM1-mutated AML, this approach has consistently suggested that NPM1 mutations act as clonal founder events, preceding the acquisition of cooperating alterations affecting signaling pathways or transcriptional programs.^2,13

This temporal sequence of mutation acquisition further shapes leukemia phenotype and prognosis. Patients in whom splicing-related mutations precede RTK/RAS-pathway alterations or in whom transcription factor lesions arise before splicing mutations experience inferior outcomes, mirroring evolutionary principles previously established in myelodysplastic and myeloproliferative neoplasms.^14,15 Moreover, specific clinical features such as leukocytosis or blast burden correlate with the subclonal abundance of certain mutations.

Within this evolutionary context, leukemic progression can be conceptualized through distinct evolutionary trajectories. Clonal modeling frameworks describe either linear evolution, characterized by the sequential accumulation of mutations within a dominant clone, or branched architectures, in which parallel subclones emerge and coexist.¹⁶

Together, these findings suggest that clonal composition represents an additional layer of prognostic information beyond mutational status or comutation profiles. NPM1-mutated AML, as a genetically well-defined entity with limited cytogenetic complexity, provides an optimal framework to dissect how quantitative and structural features of mutations modulate leukemic behavior and clinical outcome.

In this study, we assess the prognostic relevance of VAF, mutation acquisition ordering, and evolutionary trajectories in a large cohort of 688 intensively treated patients with NPM1-mutated AML, aiming to determine how clonal architecture refines current risk stratification and informs the biological heterogeneity of this disease entity.

Methods

Study design

We conducted a retrospective, multicenter, observational study including patients diagnosed with NPM1-mutated AML across 7 reference laboratories of the Programa Español de Tratamientos en Hematología (PETHEMA) cooperative group. Diagnostic evaluations including morphological, immunophenotypic, and cytogenetic analyses were performed locally according to standardized protocols. Molecular and clinical data were subsequently harmonized and centrally curated within the PETHEMA AML registry (ClinicalTrials.gov identifier: NCT02607059).

Patients and inclusion criteria

Between January 2016 and April 2025, data from 688 adult patients with NPM1-mutated AML treated with intensive chemotherapy regimens were prospectively entered into the PETHEMA AML registry. Pediatric patients (aged <18 years) and patients with acute promyelocytic leukemia were excluded. Consecutive enrollment ensured a representative, unselected real-world cohort.

All participating central laboratories obtained approval from their respective ethics committees, and written informed consent was obtained in accordance with the Declaration of Helsinki.

Frontline treatment

All patients received intensive induction chemotherapy, most commonly anthracycline-cytarabine–based regimens (62.5%), followed by consolidation with chemotherapy alone or hematopoietic stem cell transplantation according to institutional practice. Detailed information on treatment regimens is provided in Table 1 and in the supplemental Methods (available on the Blood website).

Table 1.

Main characteristics and treatment of the cohort

Feature	Total (N = 688) n (%)
Age, y
<45	93 (13.5)
45-60	301 (43.8)
61-75	279 (40.6)
>75	15 (2.2)
Sex
Male	317 (46.1)
Female	371 (53.9)
AML type
De novo	601 (87.4)
Secondary	76 (11.0)
NA	11 (1.6)
ECOG
0	332 (48.3)
1	233 (33.9)
2	46 (6.7)
3	14 (2.0)
4	5 (0.7)
NA	58 (8.4)
Cytogenetics
Not evaluable	70 (10.2)
Normal	479 (69.6)
Altered	78 (11.3)
NA	61 (8.9)
NPM1 type
A	520 (75.6)
B	39 (5.7)
D	52 (7.5)
Non-ABD	77 (11.2)
ELN 2022 risk group
Favorable	362 (52.6)
Intermediate	315 (45.8)
Adverse	11 (1.6)
Frontline treatment
3+7	430 (62.5)
IDAC	2 (0.3)
3+7+GO	10 (1.4)
FLAG-based	6 (0.9)
3+7+FLT3i	188 (27.3)
Other intensive chemotherapy	36 (5.2)
CPX-351	16 (2.3)
Transplantation in first CR/CRi
Autologous	94 (13.7)
Allogeneic	157 (22.8)
No	363 (52.8)
NA	74 (10.8)
WBC, median (range)	29.9 (0.5-401.24)
Hemoglobin, median (range)	9 (1.89-15.9)
Platelet, median (range)	68 (0.90-939)
Bone marrow blasts, median (range)	73 (1-100)
LDH, median (range)	499 (1-13 190)

ECOG, Eastern Cooperative Oncology Group performance status; ELN, European Leukemia Net; FLAG, fludarabine + cytarabine + granulocyte-colony stimulating factor; GO, gemtuzumab ozogamicin; IDAC, cytarabine; NA, not available; y, years.

Molecular studies

Targeted next-generation sequencing was performed in 7 PETHEMA designated central laboratories following harmonized technical and reporting standards.¹⁷ Variants were annotated and classified according to consensus guidelines. Pathogenic or likely pathogenic variants with a VAF ≥5% were reported, except for clinically relevant hot spot mutations, which were reported down to 1%. Details on sequencing performance, gene panels, and variant annotation are provided in the supplemental Methods.

End points

The primary end point was overall survival (OS), defined as the time from AML diagnosis to death from any cause. Relapse-free survival (RFS) was calculated from diagnosis to morphological relapse or death, and event-free survival (EFS) was calculated from diagnosis to induction failure after 1 or 2 chemotherapy cycles, relapse, or death. Response criteria followed ELN guidelines: complete remission (CR) or CR with incomplete hematological recovery (CRi) required no extramedullary disease, <5% bone marrow blasts, and no circulating peripheral blood blasts. When available, minimal residual disease (MRD) was centrally assessed after induction in peripheral blood by standardized real-time quantitative polymerase chain reaction for mutant NPM1 transcripts. MRD positivity was defined as follows: ([NPM1 copies/ABL1 copies] × 10⁵) ≥1.

Statistical analysis

Associations between genomic features and clinical variables were explored using appropriate statistical tests for categorical and continuous data. Survival outcomes were analyzed using Cox proportional hazards models, and cumulative incidence of relapse was estimated using Fine and Gray models with death as a competing event. Optimal VAF cutoffs were identified using maximally selected rank statistics. Mutation acquisition order and clonal architecture were inferred based on corrected VAFs using established computational approaches. Additional methodological details are provided in the supplemental Methods. All tests were 2-sided, and P values <.05 were considered statistically significant, unless otherwise stated.

Results

Patient characteristics

We analyzed a total of 688 adults with NPM1-mutated AML enrolled in the PETHEMA registry. The median age of the cohort was 58.5 years (range, 18-91). Women accounted for 53.9% of the cohort. Most patients corresponded to de novo AML (87.4%), whereas 11.0% were secondary, and 1.6% lacked information on disease origin. Baseline laboratory parameters showed a median white blood cell (WBC) count of 29.9 × 10⁹/L (range, 0.5 × 10⁹/L to 401.2 × 10⁹/L), hemoglobin of 9 g/dL (range, 1.89-15.9), platelet count of 68 × 10⁹/L (range, 0.9 × 10⁹/L to 939 × 10⁹/L), lactate dehydrogenase (LDH) of 499 U/L (range, 1-13 190), and bone marrow blasts 73% (range, 1%-100%). Among the cohort, 157 patients (22.8%) underwent allogeneic transplantation, 94 (13.7%) received autologous transplantation, and 363 (52.8%) did not undergo transplantation in first CR/CRi (Table 1).

Mutational landscape of NPM1-mutated AML

Regarding NPM1 mutation, type A was the most prevalent (75.6%), followed by type D (7.5%), type B (5.7%), and non-ABD variants (11.2%). Comutations were present in 669 of 688 patients (97.2%), with a median of 3 mutations (including NPM1) per case (range, 0-9). The most frequently comutated genes were DNMT3A (49.4%) and FLT3-ITD (46.4%), followed by TET2 (21.9%), IDH2 (20.2%), FLT3-TKD (14.8%), NRAS (14.4%), IDH1 (14.1%), and PTPN11 (12.9%). Mutations in SRSF2 (9.4%), FLT3-OTHER (non-ITD/TKD; 8.9%), KRAS (7.4%), and WT1 (7.4%) were also recurrent, whereas alterations in transcription factors such as CEBPA (4.5%) or myelodysplasia-related genes such as ASXL1 (2.6%) and RUNX1 (2.0%) were less common (supplemental Table 1).

VAF affects clinical presentation features and survival

To explore the relationship between NPM1 VAF and clinical outcome, we first analyzed VAF as a continuous variable. Modeling its association with survival identified a nonlinear risk pattern, with an inflection point around 30% VAF, suggesting that the prognostic effect of NPM1 burden differs above and below this threshold (supplemental Figure 1). The maximally selected rank statistic identified an optimal cutoff of 31.44%, which was subsequently used to define NPM1^low (<31.44%) and NPM1^high groups (≥31.44%; Figure 1A). VAF measurements were highly concordant across the participating reference laboratories (adjusted R² = 0.11).

Figure 1.

Clinical and molecular features of NPM1 VAF. (A) Density plot showing the distribution of NPM1 VAF across the cohort, with the optimal cut-off (31.44%) determined by the maximally selected rank statistic. (B) Baseline clinical features according to NPM1^low (<31.44%) and NPM1^high (≥31.44%) groups. (C) NPM1^low and NPM1^high distribution according to age groups. (D) Distribution of NPM1 types across VAF groups. (E) Frequency of comutations grouped by functional class by NPM1 VAF group. (F) Frequency of comutated genes by NPM1 VAF group. BM, bone marrow; ns, not significant.

Clinical and molecular features of NPM1 allelic burden

The NPM1 VAF showed a robust, though not strictly, unimodal distribution, with a median of 36.9% (range, 4.1%-54.7%). NPM1^high cases accounted for 72.7% of patients and were overrepresented at both extremes of age including both the youngest (aged <45 years, 76.4%) and the oldest individuals (aged >75 years, 84.6%; P < .001). In contrast, secondary AML was more frequent among NPM1^low than NPM1^high cases (16.67% vs 8.93%; P = .008). At baseline laboratory parameters, NPM1^high was associated with higher WBC count (median 46.3 × 10⁹/L vs 16.8 × 10⁹/L; P < .001), increased LDH levels (P = .012), a greater proportion of bone marrow blasts (P < .001), and decreased platelet count (P = .015), whereas hemoglobin levels did not differ significantly between groups (P = .052; Figure 1B-C).

Regarding NPM1 isoforms, NPM1^high was markedly overrepresented in type B (88.9%), whereas type D showed the lowest proportion of high VAF (56.0%; Figure 1D). Moreover, NPM1^high cases were underrepresented among patients harboring comutations in splicing (eg, SF3B1 and SRSF2; 58.3%) or chromatin/cohesin genes (eg, EZH2; 65.9%; Figure 1E-F).

Prognostic impact of NPM1 VAF

NPM1^high patients achieved higher CR (CR/CRi) rates (84.0% vs 73.6%; P = .003) after the first induction cycle but did not show significant difference in relapse incidence (24.9% vs 26.6%; P = .449; Figure 2A-B). Consistently, cumulative incidence of relapse (CIR) was not significantly different between groups (hazard ratio [HR], 1.40; P = .121). Additionally, CR/CRi after 1 or 2 induction cycles was higher for NPM1^high than NPM1^low cases (87.0% vs 80.9%), although this did not reach statistical significance (P = .061). Median OS was not reached in the NPM1^high group compared with 50 months in NPM1^low cases (HR, 0.69; 95% confidence interval [CI], 0.52-0.91; P = .0079; Figure 2C). Similarly, EFS (57.1 vs 21.4 months) and RFS (57.1 vs 21.8 months) were both superior among NPM1^high patients (EFS [HR, 0.76; 95% CI, 0.59-0.97; P = .027]; RFS [HR, 0.73; 95% CI, 0.56-0.94; P = .013]; Figures 2D-E). These findings were reproduced in the internal validation cohort, in which NPM1^high cases showed significantly improved OS (HR, 0.42; 95% CI, 0.22-0.78; P = .0046), RFS (HR, 0.54; 95% CI, 0.32-0.89; P = .015), and EFS (HR, 0.60; 95% CI, 0.37-0.97; P = .034). Consistent results were also observed in independent publicly available cohorts with comparable age distribution, confirming the favorable prognostic impact of higher NPM1 VAF (OS HR, 0.54; 95% CI, 0.31-0.95; P = .029; supplemental Figure 2). We did not observe a treatment-related temporal effect associated to OS, RFS, or EFS in our cohort.

Figure 2.

Clinical response and survival outcomes according to NPM1 VAF. (A) Histogram depicting CR (CR/CRi) rates after first induction cycle by NPM1 VAF group. (B) Relapse rates according to NPM1 VAF. (C) Kaplan-Meier curves for OS according to NPM1 VAF group. (D) Kaplan-Meier curves for RFS. (E) Kaplan-Meier curves for EFS.

To explore the transcriptional programs associated with NPM1 VAF, we analyzed publicly available RNA sequencing data. NPM1-mutated cases were stratified into NPM1^high and NPM1^low groups according to the VAF threshold used in our study. Genes enriched in the NPM1^low group were predominantly associated with myeloid differentiation programs, inflammatory signaling, and granulocytic/monocytic components. In contrast, expression of homeobox-related genes did not significantly differ between groups, suggesting that the canonical homeobox transcriptional program is preserved and that differences may arise from alternative biological axes.

Regarding the FLT3-ITD subgroup, outcomes differed according to FLT3 inhibitor exposure, as expected, with improved OS, RFS, and EFS among treated patients; however, FLT3 inhibitor use did not modify the prognostic impact of NPM1 VAF (supplemental Figure 3).

In a subset of patients with available postinduction MRD data, MRD positivity was more frequent in NPM1^high than in NPM1^low cases (70% vs 56%; P = .039). Exploratory analyses integrating NPM1 VAF and MRD suggested a trend toward inferior OS among NPM1^high MRD⁺ compared with NPM1^high MRD⁻ cases (HR, 1.7; 95% CI, 0.78-3.70; P = .182), whereas patients with NPM1^low MRD⁺ showed the poorest outcomes. Differences were more evident for CIR, in which NPM1^high MRD⁺ showed an increased risk of relapsed compared to NPM1^high MRD⁻ (P = .134). Importantly, the trend toward a slightly higher relapse risk in the NPM1^high group should be interpreted cautiously, because the lower rate of allogeneic transplantation in this subset (23% vs 27.1% in NPM1^low) may at least partly account for the observed difference (supplemental Figure 4).

Outcomes were further stratified according to transplantation status, and the prognostic impact of the VAF appeared to be treatment dependent. Differences in OS between NPM1^high and NPM1^low patients were most pronounced among those who did not undergo transplantation (P = .039). In contrast, this separation was attenuated and no longer statistically significant in patients undergoing autologous or allogeneic transplantation (supplemental Figure 5).

Univariable Cox analyses were first performed for recurrent gene mutations present in at least 10 patients to identify molecular predictors of OS, RFS, and EFS (supplemental Table 2). These associations remained largely consistent after adjustment for age, sex, AML type, targeted therapy, cytogenetic risk, transplant, NPM1 variant type, and prognostic comutations identified in univariate analyses. In the multivariate Cox models for OS, RFS, and EFS, NPM1^low status retained an independent worse effect for OS (HR, 1.46; P = .037) and RFS (HR, 1.40; P = .036), with a nonsignificant but concordant trend for EFS. Notably, non-ABD NPM1 type consistently predicted inferior outcomes across all end points (OS [HR, 2.11; P < .001]; RFS [HR, 2.04; P < .001]; EFS [HR, 1.90; P = .002]). Among co-occurring mutations, SRSF2 emerged as a strong and independent adverse prognostic factor, retaining significance for OS (HR, 2.11; P = .004), RFS (HR, 2.62; P < .001), and EFS (HR, 2.52; P < .001). FLT3-ITD also maintained an independent negative impact across end points, with consistent associations with inferior OS (HR, 1.86; P < .001), RFS (HR, 1.51; P = .007), and EFS (HR, 1.55; P = .004). In contrast, DNMT3A showed a selective adverse effect on relapse-related outcomes, being significantly associated with shorter RFS (HR, 1.66; P < .001) and EFS (HR, 1.60; P = .002) but not with OS. Overall, transplantation remained the strongest predictor of favorable outcome (Table 2).

Table 2.

Multivariable Cox regression model including clinical covariates and selected genetic variables in the intensive therapy cohort

Feature	OS				RFS				EFS
	HR	95% CI low	95% CI high	P value	HR	95% CI low	95% CI high	P value	HR	95% CI low	95% CI high	P value
Age (continuous)	1.02	0.98	1.03	.035	1.01	1.00	1.03	.136	1.01	0.99	1.02	.301
Sex: female vs male	0.93	0.67	1.29	.662	0.97	0.73	1.30	.853	0.98	0.74	1.29	.863
AML type: secondary vs de novo	1.14	0.70	1.87	.602	1.08	0.69	1.70	.722	1.17	0.76	1.80	.476
Cytogenetics: altered vs normal	1.08	0.72	1.64	.706	1.11	0.77	1.61	.572	1.07	0.74	1.55	.705
NPM1 type: B vs A	1.05	0.50	2.21	.899	0.84	0.40	1.74	.641	0.93	0.47	1.87	.846
NPM1 type: D vs A	1.19	0.68	2.10	.542	0.89	0.52	1.52	.666	0.84	0.49	1.44	.529
NPM1 type: non-ABD vs A	2.11	1.37	3.26	<.001	2.04	1.37	3.05	<.001	1.90	1.27	2.83	.002
NPM1 VAF: low vs high	1.46	1.02	2.07	.037	1.40	1.02	1.91	.036	1.32	0.98	1.80	.072
HCT: autologous vs no	0.27	0.15	0.50	<.001	0.33	0.20	0.54	<.001	0.30	0.19	0.50	<.001
HCT: allogeneic vs no	0.30	0.19	0.46	<.001	0.29	0.20	0.43	<.001	0.26	0.18	0.39	<.001
ETV6 (mut)	2.44	0.89	6.70	.083	2.48	1.00	6.15	.051	2.16	0.87	5.36	.096
FLT3-ITD (mut)	1.86	1.32	2.61	<.001	1.51	1.12	2.04	.007	1.55	1.15	2.08	.004
FLT3-OTHER (mut)	0.54	0.28	1.04	.066	0.59	0.35	1.01	.056	0.62	0.37	1.05	.075
SRSF2 (mut)	2.11	1.26	3.51	.004	2.62	1.67	4.10	<.001	2.52	1.62	3.91	<.001
KRAS (mut)	1.81	1.09	3.02	.022	1.56	0.99	2.48	.058	1.59	1.01	2.50	.044
NRAS (mut)	0.74	0.44	1.24	.252	0.73	0.47	1.13	.156	0.65	0.42	1.01	.058
DNMT3A (mut)	1.27	0.92	1.76	.151	1.66	1.23	2.24	<.001	1.60	1.19	2.14	.002
PTPN11 (mut)	1.13	0.66	1.92	.658	0.89	0.55	1.44	.644	1.00	0.62	1.60	.986

The category listed after “vs” was used as the reference, except for gene mutation where the reference category was wild type. Boldface indicates significant values.

HCT, hematopoietic cell transplantation; mut, mutated; OTHER, Non-ITD/TKD.

Gene-specific VAF thresholds refine prognostic across comutations

To determine whether the prognostic relevance of clonal burden extended beyond NPM1, we applied the same analytical framework to all recurrently mutated genes within the cohort (n ≥ 20 cases per gene). Optimal cutoffs were established individually for each gene, generating a gene-specific VAF threshold map (supplemental Table 3).

The prognostic impact of allelic burden was strongly gene dependent. High VAF in DNMT3A, FLT3-OTHER, KRAS, and PTPN11 was associated with inferior OS, whereas in NPM1 and IDH2 correlated with improved outcomes. The effects of other signaling or epigenetic regulators such as FLT3-TKD, NRAS, or TET2 were largely neutral (supplemental Table 3).

Patients with DNMT3A^high status exhibited significantly shorter survival than those with DNMT3A^low (HR, 1.73; 95% CI, 1.11-2.7; P = .014), whereas IDH2^high conferred a protective effect (HR, 0.46; P = .011). Conversely, higher VAFs in RTK/RAS pathway genes such as FLT3-OTHER, KRAS, and PTPN11 cases showed significantly poorer outcomes (HR range, 2.3-3.3; P < .05). A similar, albeit less robust, trend was observed for SRSF2 and WT1 (HR > 2.3; P < .1; Figure 3).

Figure 3.

Gene-specific optimized VAF thresholds and their association with OS. Kaplan-Meier survival curves showing OS according to high vs low VAF for selected genes, using data-driven optimal thresholds.

When combining NPM1 and comutational VAFs in Cox models, the interaction emerged as a major prognostic determinant. Patients with NPM1^high and a favorable VAF profile of the comutated gene exhibited the best outcomes, whereas those in whom the comutation or both NPM1 and the comutation displayed unfavorable VAF profiles experienced the poorest survival. The magnitude of this effect varied across genes, being most pronounced for PTPN11 (HR, 17.0; 95% CI, 3.3-88.7; P < .001), followed by SRSF2 (HR, 3.76; P = .043) and IDH2 (both HR, 2.33; P < .05).

These results suggest that the relative allelic burden between NPM1 and its comutations captures an additional layer of biological heterogeneity, prompting further exploration of these relationships.

Ordering of mutation acquisition

To determine whether the quantitative hierarchy observed between NPM1 and its comutations reflected the temporal sequence of mutation acquisition, we inferred pairwise mutation ordering using corrected VAFs and a Bradley-Terry likelihood model (Figure 4A).

Figure 4.

Mutational hierarchy and prognostic impact of inferred acquisition order. (A) Pairwise mutational precedence matrix derived from corrected VAF. Each square represents the proportion of cases where the mutation in Gene 1 was inferred to occur before that in Gene 2. The color scale denotes the fraction of unambiguous cases, and the circle size reflects the number of co-occurrences. (B) Summary of inferred mutational hierarchy through Bradley-Terry model across functional gene categories. Density plots show the relative VAF distribution per gene, and the forest plot summarizes their average temporal order estimates (mean ± 95% CI). (C) Kaplan-Meier survival curves according to inferred pairwise mutation order for FLT3-OTHER, FLT3-TKD, and WT1. (D) Kaplan-Meier survival analyses grouped by functional mutation category.

NPM1 was acquired after early preleukemic lesions involving epigenetic regulators (DNMT3A and TET2) or splicing factors (SRSF2 and SF3B1) and before mutations in signaling genes (FLT3-ITD, FLT3-TKD, NRAS, KRAS, and PTPN11) and transcription factors (CEBPA and GATA2; Figure 4B). This ordering pattern positions NPM1 as a transforming event that bridges clonal hematopoiesis and overt leukemic expansion.

Survival analyses according to inferred order of mutation acquisition revealed that specific acquisition sequences carried distinct prognostic implications. In particular, the relative timing of NPM1 with FLT3-OTHER, FLT3-TKD, and WT1 significantly stratified patient outcomes. The noncanonical order acquisition patterns in which FLT3-OTHER (HR, 0.26; P = .048) or FLT3-TKD (HR, 0.34; P = .057) preceded NPM1 showed superior survival compared with cases in which NPM1 occurred earlier. In contrast, early acquisition of WT1 (HR, 7.58; P = .028) conferred a markedly adverse outcome (Figure 4C). These findings mirror their pairwise VAF relationships and align with our previous observations linking high WT1 allelic burden to poor prognosis.

When mutations were grouped into functional categories, 2 noncanonical ordering patterns significantly stratified survival. The most pronounced effect was observed in patients in whom NPM1 preceded tumor suppressor mutations (WT1 and TP53), who showed markedly improved outcomes (HR = 0.40, P = .013). In contrast, when NPM1 occurred after a transcription factor mutation (CEBPA and GATA2) survival was significantly worse (HR, 1.9; P = .025; Figure 4D).

Clonal architecture and evolutionary trajectories

To characterize the structural organization and evolutionary dynamics of leukemic clones, we reconstructed clonal architecture across the cohort. The median number of inferred clones was 2, and most patients exhibited linear evolutionary trajectories (84.3%). Higher mutational burden was associated with inferior OS, and this association was mainly driven by patients with linear trajectories (Figure 5A-B). In contrast, branched evolution patients with higher mutation counts initially appeared to fare better, although survival curves converged over time with those carrying fewer mutations (Figure 5C). No significant difference in OS was observed when comparing linear and branched configurations (Figure 5D).

Figure 5.

Clonal architecture and evolutionary trajectories in NPM1-mutated AML. (A) Kaplan-Meier curves showing OS according to mutational burden. (B) OS according to mutational burden restricted to patients with linear trajectories. (C) OS among cases with branched evolution. (D) Comparison of OS between linear and branched evolutionary trajectories. (E) Kaplan-Meier curves for OS according to the number of clones. (F) Kaplan-Meier curves for OS according to the mutational burden per clone. (G) Frequency of clonal colocalization between NPM1 and recurrent comutations, stratified by “same clone” (blue) vs “different clone” (red). (H) OS according to clonal relationship between NPM1 and TET2. (I) RFS according to clonal relationship between NPM1 and IDH1. (J) RFS according to clonal relationship between NPM1 and WT1.

Patients harboring a greater number of clones also experienced reduced OS (Figure 5E). Consistently, the number of mutations positively correlated with the number of inferred clones (Kruskal-Wallis P < .001), although survival did not differ significantly between patients with higher mutational density per clone (Figure 5F).

To explore whether the clonal colocalization of specific lesions influenced prognosis, we assessed the clonal relationship between NPM1 and recurrent comutations (Figure 5G). TET2 and IDH1 mutations occurring within the same clone as NPM1 were associated with markedly improved outcomes, reflected in superior OS (TET2; HR, 0.43; 95% CI, 0.21-0.88; P = .022), RFS (IDH1; HR, 0.21; 95% CI, 0.06-0.67; P = .009), and EFS (IDH1; HR, 0.28; 95% CI, 0.10-0.79; P = .016; Figure 5H-I). In contrast, patients in whom WT1 co-occurred with NPM1 in the same clone exhibited significantly shorter EFS (HR, 3.74; 95% CI, 1.64-8.51; P = .002) and RFS (HR, 2.89; 95% CI, 1.25-6.70; P = .013), consistent with the adverse effect of high WT1 allelic burden observed in preceding analyses (Figure 5J). Among these, IDH1 stood out for showing a marked prognostic divergence depending on its clonal relationship with NPM1, as segregation into distinct clones was associated with significantly poor outcomes.

Discussion

This study demonstrates that quantitative dimensions of clonal architecture provide independent biological and prognostic insight in NPM1-mutated AML. Using a large, uniformly genotyped, real-world cohort of 688 intensively treated patients, we show that NPM1 allelic burden discriminates disease phenotype, comutation patterns, and outcome. NPM1 VAF ≥31.44% identified patients with a more proliferative disease (higher WBC, LDH, and blasts) who experienced superior CR/CRi rates, OS, EFS, and RFS. Although different studies have investigated the prognostic relevance of NPM1 VAF, interpretation has been hampered by using arbitrary cutoffs, most commonly medians or quartiles, leading to heterogeneous and sometimes contradictory results across cohorts.^18,19 Our approach based on restricted cubic splines addresses this limitation by defining a biologically and clinically informed threshold.

Biologically NPM1^high defined a distinct group compared with NPM1^low. It was overrepresented in type B cases, which in our series had an OS comparable to type A, and underrepresented in type D, which tended to inferior outcome. The overall comutational landscape observed in our cohort was consistent with patterns recently reported by Hernández-Sánchez et al.²⁰ Conversely, NPM1^low clustered with a distinct genetic background, characterized by fewer canonical NPM1 partners (DNMT3A and FLT3-ITD) and a relative enrichment in splicing and chromatin/cohesin lesions, including myelodysplasia-related genes such as SRSF2, which are linked to secondary AML and are associated with poor prognosis. This shift toward a secondary-like genomic context provides a plausible explanation for the inferior outcomes despite the presence of the same sentinel lesion.^{21, 22, 23} Given that secondary-type genomic features are more frequent in older patients, cohorts enriched for younger individuals may underrepresent the NPM1^low subgroup, which may partly contribute to heterogeneity across published studies evaluating VAF-based stratification.

NPM1^high remained associated with higher CR/CRi rates and improved outcomes across all end points, even after multivariable adjustment. These findings contrast with earlier, smaller reports that described no effect or even unfavorable outcomes in patients with higher NPM1 VAF.^18,19,24 The most likely explanation for this discrepancy includes limited sample size and heterogeneous VAF stratification (median- or quantile-based cutoffs). In line with recent data, we also confirmed that non-ABD NPM1 variants retain adverse prognostic impact in multivariable models and that autologous and allogeneic transplant remain dominant favorable covariates.^{25, 26, 27} Despite the strong prognostic effect of transplantation, multivariable analyses indicate that NPM1 VAF captures baseline risk beyond established covariates, which is most evident in patients managed without transplant and becomes less discriminative in the context of intensive postremission strategies. These results are consistent with the strong risk-modifying effect of transplantation rather than a loss of biological relevance of NPM1 VAF and suggest that quantitative clonal features may be modulated by subsequent therapeutic interventions.

Although MRD positivity was more frequent among NPM1^high cases, these patients consistently showed superior survival, arguing against high VAF as a marker of chemoresistance and instead supporting its interpretation as a surrogate of clonal dominance within a biologically more homogeneous leukemia. By contrast, in NPM1^low, lower detection of NPM1 transcripts may be less reassuring, because the NPM1-mutated clone may represent only part of a more complex disease architecture. This interpretation is consistent with our results, in which NPM1^low MRD⁺ patients had the poorest outcomes, whereas MRD further refined relapse risk within NPM1^high cases. Overall, these findings suggest that baseline NPM1 VAF may provide prognostic information for OS and CIR beyond that captured by MRD alone. In particular, it may be especially useful in identifying MRD^- patients who still retain a relatively high risk of relapse and who might therefore benefit from more intensive postremission strategies, including consideration of allogeneic transplantation.

Extending this quantitative framework to recurrent comutations revealed that allelic burden is gene specific. High VAF in DNMT3A, FLT3-OTHER, KRAS, and PTPN11 identified patients with significantly inferior survival, and this was particularly pronounced for RTK/RAS lesions. These alterations are frequently subclonal, and our data suggest that their adverse effects only emerge when they become clonally dominant.^28,29 By contrast, high VAF in IDH2 as observed in NPM1 was associated with improved outcome. One plausible explanation is that high burden IDH2 reflects early, founder clone participation with NPM1 and reflects the canonical NPM1 leukemogenic trajectory, whereas low burden indicates parallel evolution and increased clonal complexity.^30,31

Consistent with this model, patients with NPM1^high together with a favorable VAF status of the partner mutation experienced the best outcomes, whereas those in whom the partner or both lesions fell within an adverse VAF stratum exhibited the worst behavior. This underscores that prognosis in NPM1-mutated AML is determined less by the mere presence of specific mutations than by which clone dominates the leukemic population at presentation.

As expected from previous bulk and single-cell studies, our temporal modeling showed that NPM1 was typically acquired after epigenetic/splicing lesions and before transcription-factor or signaling mutations, consistent with its established role as driver in leukemogenesis.^10,14,32 However, we identified noncanonical acquisition sequences such as early FLT3-OTHER/TKD or WT1, which clearly stratified outcome. These patterns suggest that deviation from canonical mutation order reflects alternative trajectories with distinct clinical consequences.^15,33 Additionally, clonal reconstruction showed that increased clonal complexity and higher mutation burden were associated with inferior survival, particularly in cases exhibiting linear evolutionary patterns. The importance of clonal context was illustrated by IDH1, when it was in the same clone as NPM1, outcomes were favorable; but when it resided in a separate clone, prognosis deteriorated. This supports a model in which the same mutation can be prognostically opposite depending on its clonal location.

This study has some limitations. In particular, the lack of significant differences between linear and branched evolutionary trajectories may be attributable to the use of targeted bulk sequencing at diagnosis. In fact, some studies based on whole-exome sequencing or single-cell DNA sequencing have reported higher rates of branched trajectories and better outcomes.^14,34 Thus, these results should be interpreted with caution. Moreover, although clonal architecture and mutation ordering were inferred from bulk VAFs using validated computational tools, these remain approximations and should ideally be confirmed using single-cell or high-depth longitudinal data sets.^35,36 Importantly, MRD, a well-established and powerful modulator of prognosis in NPM1-mutated AML, was not available in all patients and therefore could not be incorporated into the multivariable survival models. Nevertheless, we performed an exploratory analysis in a substantial subset of cases with available postinduction MRD data, which supported the complementary prognostic value of integrating baseline NPM1 allelic burden with MRD response dynamics. Finally, despite the large cohort size, some comutations or functional categories were infrequent, warranting external validation before clinical translation.

In summary, our findings indicate that risk in NPM1-mutated AML is multilayered by mutation identity, allelic burden, order of acquisition, and clonal location. The observation that prognosis also depended on whether key lesions reside within the NPM1 containing founder clone or in a separate branch highlights the importance of integrating both quantitative and structural dimensions of clonality. Accordingly, future risk stratification models should move beyond binary mutation status to incorporate quantitative genomics features, integrate MRD to capture posttherapy clonal reshaping, and evaluate whether scores can better select patients for transplant or targeted intensification. The emergence of menin inhibitors, which show high activity in NPM1-mutated AML, may further modify the prognostic impact of baseline molecular features and should be considered in future validation of VAF-based risk models. Similarly, the broader incorporation of FLT3 inhibitors, including emerging strategies exploring their use beyond FLT3-mutated AML, may further modulate the prognostic relevance of baseline genomic features in future therapeutic contexts.

Conflict-of-interest disclosure: The authors declare no competing financial interests.