Prognostic Model Combining Mutational and Cytogenetic Profiles in Acute Myeloid Leukemia Treated with Venetoclax and Hypomethylating Agents

Abstract

Venetoclax (VEN) combined with hypomethylating agents (HMA) improves outcomes for patients with newly diagnosed acute myeloid leukemia (AML) who are ineligible for intensive chemotherapy, yet overall survival (OS) remains variable. We analyzed 506 AML patients treated with front-line HMA/VEN at Moffitt Cancer Center to develop a genetics-based prognostic model. In multivariate analysis, mutations in TP53, KRAS, JAK2, U2AF1, CBL and cytogenetic lesions del(7q)/-7, del(17p)/-17/i(17q), del(20q), and MECOM rearrangements predicted inferior OS, whereas IDH1/2 mutations were favorable. A point-based system stratified patients into low-, intermediate-, and high-risk groups with median OS of 54.2, 22.3, and 7.5 months, respectively, (p<0.0001; C-index 0.648). External validation (n=126) retained prognostic separation (median OS 24.7, 17.4, and 4.3 months, p=0.0005; C-index 0.626). Compared to existing HMA/VEN-specific models, our model demonstrated superior low- vs. intermediate-risk discrimination (31.9-month separation, p=0.002; HR=0.45, p=0.003), with comparable C-index. Our model supports personalized risk stratification for HMA/VEN-treated AML, pending broader validation.

INTRODUCTION

The combination of the BCL2 inhibitor venetoclax (VEN) with a hypomethylating agent (HMA)—either azacitidine or decitabine—has marked a significant advancement in the treatment of acute myeloid leukemia (AML) (1–3). Compared to HMA monotherapy, HMA/VEN demonstrates significantly higher response rates and improved overall survival (OS), as shown in the pivotal VIALE-A trial (3, 4). Accordingly, the HMA/VEN regimen has become the standard front-line therapy for AML patients ineligible for intensive chemotherapy due to advanced age, frailty, or significant comorbidities (5).

Despite its efficacy, treatment response and survival outcomes vary widely among patients (1–3). To address this heterogeneity, extensive efforts have focused on identifying key genetic alterations linked to outcomes under HMA/VEN therapy and developing prognostic models that better predict clinical outcomes and refine risk stratification (6–10). Prognostic models typically integrate clinical variables with genetic data obtained through next-generation sequencing (NGS) and cytogenetic analyses (11–17). ELN 2022, a widely used risk stratification system incorporating both mutational and cytogenetic features, was derived primarily from intensively treated AML cohorts (18). Several studies have since demonstrated its suboptimal performance in HMA/VEN-treated patients (19, 20).

More recent mutation-centric models tailored to the HMA/VEN setting have emerged but with notable inconsistencies. ELN 2024 and refined ELN 2024 classify NPM1, IDH1, IDH2, and DDX41 mutations as favorable, while designating FLT3-ITD, KRAS, NRAS, PTPN11, and TP53 mutations as intermediate or adverse risk factors (15, 16); however, these models do not incorporate cytogenetic abnormalities. The BEAT-AML ELN 2024 (BEAT-AML) system similarly emphasizes mutations (e.g., IDH2, TP53, KRAS, and MLL2) while excluding cytogenetic profiles and was developed in patients aged over 60 years (13). In contrast, the Mayo genetic risk model integrates both mutational and cytogenetic profiles, classifying IDH2 mutation as favorable and TP53, KRAS, KMT2A-rearrangements, and ELN-adverse cytogenetics as adverse (17). However, it does not evaluate the prognostic impact of individual cytogenetic lesions. Notably, markers identified as favorable in ELN 2024 or refined ELN 2024 (i.e., NPM1, IDH1, IDH2, DDX41) were not uniformly retained as favorable in the BEAT-AML or Mayo model (13, 17). This cross-study variability limits generalizability and clinical utility.

To address these gaps, we developed an enhanced genetic risk model that integrates both mutation profiles and cytogenetics status to predict OS in newly diagnosed AML patients treated with HMA/VEN. This model was derived in an institutional cohort from Moffitt Cancer Center and independently validated in a combined AML cohort from Nice University Hospital and the University of Maryland.

RESULTS

Patient Characteristics

A total of 506 consecutive patients diagnosed with AML were included in the study (Figure S1). Of these, 36.8% (n=186) were female and 63.2% (n=320) were male (Table 1). The median age at AML diagnosis was 74.3 (18.8–92.6) years. A total of 42.7% (n=216) had sAML and 23.7% (n=120) patients had tAML. Among sAML cases, 4.5% (n=23) represented progression from prior therapy-related MDS (tMDS). Notably, 16.4% (n=83) of patients with sAML had received HMA prior to their AML diagnosis (Table S1). ELN 2022 risk classification was available for 99.2% (n=502) of patients, with 7.4% (n=37) classified as favorable, 14.3% (n=72) as intermediate, and 78.3% (n=393) as adverse risk (Table 1). ELN 2024 risk was assessed in all patients, with 54.7% (n=277) classified as favorable, 19.0% (n=96) as intermediate, and 26.3% (n=133) as adverse risk (Table 1). Detailed clinical and demographic characteristics are summarized in Table 1 and Table S1.

Table 1. Baseline characteristics of the study cohort.

Demographic, laboratory, and molecular profiles stratified by best response during HMA/VEN treatment.

Variables	CR/CRi (n=273)	Non-CR/CRi (n=175)	Response not assessed (n=58)	All patients (n=506)	P values*
Age at diagnosis, years (range)	74.7 (18.8–90.9)	74.0 (37.1–89.2)	72.7 (21.5–92.6)	74.3 (18.8–92.6)	0.388
Gender, n (%)
Female	98 (35.9)	64 (36.6)	24 (41.4)	186 (36.8)	0.971
Male	175 (64.1)	111 (63.4)	34 (58.6)	320 (63.2)
AML subtype, n (%) **
De Novo	139 (50.9)	47 (26.9)	20 (34.5)	206 (40.7)	<0.001
sAML	86 (31.5)	104 (59.4)	26 (44.8)	216 (42.7)
tAML	60 (22.0)	41 (23.4)	19 (32.8)	120 (23.7)
ELN 2022 risk, n (%)	N=270	N=175	N=57	N=502
Favorable	27 (10.0)	5 (2.9)	5 (8.8)	37 (7.4)	<0.001
Intermediate	48 (17.8)	17 (9.7)	7 (12.3)	72 (14.3)
Adverse	195 (72.2)	153 (87.4)	45 (78.9)	393 (78.3)
ELN 2024 risk (%)	N=273	N=175	N=58	N=506
Low-risk	168 (61.5)	79 (45.1)	30 (51.7)	277 (54.7)	0.001
Intermediate-risk	51 (18.7)	37 (21.1)	8 (13.8)	96 (19.0)
High-risk	54 (19.8)	59 (33.7)	20 (34.5)	133 (26.3)
NGS results, n (%)	N=270	N=175	N=58	N=506
TP53	54 (20.0)	59 (33.7)	20 (34.5)	133 (26.3)	0.001
ASXL1	64 (23.7)	42 (24.0)	8 (13.8)	114 (22.5)	0.943
RUNX1	60 (22.2)	38 (21.7)	6 (10.3)	104 (20.6)	0.899
TET2	54 (20.0)	40 (22.9)	8 (13.8)	102 (20.2)	0.471
DNMT3A	57 (21.1)	28 (16.0)	10 (17.2)	95 (18.8)	0.180
SRSF2	60 (22.2)	26 (14.9)	7 (12.1)	93 (18.4)	0.055
NPM1	37 (13.7)	7 (4.0)	6 (10.3)	50 (9.9)	<0.001
FLT3-ITD	22 (8.1)	16 (9.1)	3 (5.2)	41 (8.1)	0.714
FLT3-TKD	15 (5.6)	8 (4.6)	1 (1.7)	24 (4.7)	0.647
Cytogenetics, n (%)	N=270	N=175	N=57	N=502
Complex Karyotype	70 (25.9)	64 (36.6)	20 (35.1)	154 (30.7)	0.017
Del(5q)/-5	47 (17.4)	49 (28.0)	15 (26.3)	111 (22.1)	0.008
Del(7q)/-7	38 (14.1)	52 (29.7)	17 (29.8)	107 (21.3)	<0.001
Trisomy 8	47 (17.4)	32 (18.3)	11 (19.3)	90 (17.9)	0.813
Del(17)/-17/i(17q)***	25 (9.3)	27 (15.4)	7 (12.3)	59 (11.8)	0.048
Del(20q)	22 (8.1)	18 (10.3)	7 (12.3)	47 (9.4)	0.441
Inv(3)/t(3;3)/MECOM	9 (3.3)	11 (6.3)	2 (3.5)	22 (4.4)	0.142
KMT2A-rearranged	6 (2.2)	7 (4.0)	2 (3.5)	15 (3.0)	0.277
HMA used, n (%)
Azacitidine	207 (75.8)	99 (56.6)	38 (65.5)	344 (68.0)	<0.001
Decitabine	66 (24.2)	76 (43.4)	20 (34.5)	162 (32.0)
Response post C1, n (%)	N=203	N=142	NA	N=345
CR/CRi	96 (47.3)	0 (0)		96 (27.8)
MLFS	57 (28.1)	14 (9.9)		71 (20.6)
PR	38 (18.7)	27 (19.0)		65 (18.8)
No response	12 (5.9)	101 (71.1)		113 (32.8)
Best response, n (%)	N=273	N=175	NA	N=448
CR/CRi	273 (100)	0 (0)		273 (60.9)
MLFS	0 (0)	22 (12.6)		22 (4.9)
PR	0 (0)	36 (20.6)		36 (8.0)
No response	0 (0)	117 (66.9)		117 (26.1)
Allo-SCT, n (%)	64 (23.4)	8 (4.6)	1 (1.7)	73 (14.4)	<0.001

^*P-values were calculated using the Chi-square test or Fisher’s exact test for categorical variables and Student’s t-test for continuous variables. Comparisons were made between CR/CRi and non-CR/CRi groups.

^**Detailed information on sAML and tAML is provided in Table S1.

^***There was a single i(17q) case in the entire cohort, which belonged to the non-CR/CRi group.

Abbreviation: sAML (secondary AML), tAML (therapy-related AML), Allo-SCT (allogeneic stem cell transplant), CR (complete response), CRi (CR with incomplete count recovery), MLFS (morphologically leukemia free status), PR (partial response), OS (overall survival).

Mutational and Cytogenetic profiles

NGS using a 98-gene myeloid mutation panel was performed in 97% (n=491) of patients; the remainder (n=15) were profiled with alternative panels. The most frequent mutations at diagnosis were TP53 (26.3%, n=133) followed by ASXL1 (22.5%, n=114), RUNX1 (20.6%, n=104), TET2 (20.2%, n=102), and DNMT3A (18.8%, n=95) (Table 1 and Figure S2). These findings are consistent with those reported in other study cohorts (10–13, 17). Notably, CEBPA bZIP-domain mutations were rare (0.8%, n=4); accordingly, all CEBPA variants (4.2%, n=21) were analyzed as a single category. Cytogenetics were assessed in 99.2% (n=502) of patients. Among these, lesion-level abnormalities were documented in 87.4% (n=442) by both karyotyping and FISH; 4.3% (n=22) by karyotyping only; 7.5% (n=38) by FISH only. Complex karyotype (defined as ≥3 independent chromosomal abnormalities) was observed in 30.7% (n=154), while del(5q)/-5, del(7q)/-7, and trisomy 8 were identified in 22.1% (n=111), 21.3% (n=107), and 17.9% (n=90) of patients, respectively (Table 1 and Figure S2). Importantly, a significant proportion of mutations and cytogenetic abnormalities co-occurred within individual patients, supporting the inclusion of both mutational and cytogenetic profiles in outcomes analyses (Figure S3).

Treatment Response to HMA and VEN

In our AML cohort, 68% (n=344) of patients received VEN plus azacitidine, and 32% (n=162) received VEN plus decitabine as first-line therapy (Table 1). The median duration of treatment was 3 cycles (range 1–71) (Table S1 and Figure S4A–C). Treatment response after cycle 1 was assessed in 68.2% (n=345) of patients, among whom 27.8% (n=96) achieved complete remission (CR) or CR with incomplete hematologic recovery (CRi), while 72.2% (n=249) had morphologic leukemia-free state (MLFS), partial response (PR), or no response (NR) (Table 1). Over the course of treatment, including subsequent cycles, 88.5% (n=448) of patients were evaluated for response. Among these, 60.9% (n=273) achieved CR/CRi, 4.9% (n=22) achieved MLFS, 8% (n=36) had PR, and 26.1% (n=117) had NR. The CR/CRi rate was significantly lower in patients with prior HMA exposure compared to HMA-naïve patients (27.8%, n=20/72 vs. 67.3%, n=253/376, p<0.001). In patients who achieved CR/CRi, measurable residual disease (MRD) assessment by multiparameter flow cytometry (MFC) was performed in 39.6% (n=108/273). Of these, 42.6% (n=46/108) were MRD-negative, while 57.4% (n=62/108) were MRD-positive. Allo-SCT was performed in 14.4% (n=73) of patients. The distribution of mutations and cytogenetic abnormalities by response status is detailed in Table 1 and Figure S5A–C.

In univariate analysis, mutations in JAK2 and TP53, as well as cytogenetic abnormalities including del(7q)/-7, del(5q)/-5, del(17p)/-17/i(17q), and complex karyotype were associated with lower rates of CR/CRi (Table S2 and Figure 1A). In contrast, mutations in IDH2, NPM1, and DDX41 were associated with higher CR/CRi rates (Table S2 and Figure 1A). In multivariate analysis, including genetic alterations significantly associated with response in univariate analysis, IDH2 (OR=0.22, p=0.001) and DDX41 (OR=0.24, p=0.013) mutations remained independent predictors of achieving CR/CRi. Conversely, JAK2 mutation (OR=3.48, p=0.013) and del(7q)/-7 (OR=2.21, p=0.008) were independently associated with failure to achieve CR/CRi (Table S2 and Figure 1B). Notably, although TP53 mutations are associated with inferior response and have been implicated in VEN resistance (10, 21), they were not independently associated with lower response rates in our cohort, consistent with findings from recent studies (11, 17).

Figure 1. Association of Mutational and Cytogenetic Aberrations with Treatment Response and OS.

Univariate (A) and multivariate (B) analyses of odds ratios (OR) for failure to achieve CR/CRi. Univariate (C) and multivariate (D) analyses of hazard ratios (HR) for OS, censored at the time of Allo-SCT. The x-axis indicates the OR or HR, and the y-axis lists individual genetic alterations.

Risk Factors Associated with Overall Survival

We first evaluated OS based on clinical factors, censoring patients at the time of Allo-SCT to account for its potential impact on survival outcomes. Consistent with prior studies (17), patients who achieved CR/CRi had significantly longer median OS compared to those who did not (22.3 vs. 5.4 months, p<0.0001) (Figure S6A). Among responders, those with MRD-negative status had a longer median OS compared to MRD-positive patients (not reached (NR) vs. 22.7 months, p=0.074), though the difference did not reach statistical significance, likely due to the limited number of patients assessed for MRD (39.6%, n=108/273). Patients with de novo AML had significantly longer median OS compared to those with sAML or tAML (22.7 vs. 8.6 months, p<0.0001) (Figure S6B). Additionally, prior exposure to HMA was associated with significantly shorter median OS compared to HMA-naïve patients (5.6 vs. 15.0 months, p<0.0001) (Figure S6C).

Given that the primary goal of this study was establishing a genetic risk model, we focused on mutations and cytogenetic abnormalities in the downstream analyses. In a univariate analysis, mutations in IDH1 (HR=0.55, p=0.055), IDH2 (HR=0.52, p=0.002), NPM1 (HR=0.37, p<0.001), and CEBPA (HR=0.39, p=0.023) were associated with longer OS, whereas mutations in TP53 (HR=2.44, p<0.001), KRAS (HR=1.71, p=0.021), JAK2 (HR=2.47, p<0.001), U2AF1 (HR=1.72, p=0.008), CBL (HR=2.84, p<0.001), as well as del(5q)/-5 (HR=2.05, p<0.001), del(7q)/-7 (HR=2.48, p<0.001), del(17p)/-17/i(17q) (HR=2.73, p<0.001), del(20q) (HR=2.34, p<0.001), trisomy 8 (HR=1.45, p=0.012), complex karyotype (HR=2.66, p<0.001), and MECOM rearrangements (HR=2.62, p<0.001) were associated with significantly worse OS (Table S3 and Figure 1C).

Genetic Risk Model

Genetic variables with significant prognostic relevance in univariate screening (p<0.05) were entered into a multivariate Cox model. TP53 was modeled as a binary covariate (mutant vs. wild type) because neither allelic status (mono-allelic vs. multi-hit) nor VAF strata (<20% vs. ≥20%) showed differential OS under HMA/VEN (Figures S7A–C). Although IDH1 had a univariate p=0.055, it was included per our prespecified criterion (Figure S1). In the multivariate model, IDH1 (HR=0.49, p=0.039) and IDH2 (HR=0.53, p=0.009) were associated with improved OS, whereas TP53 (HR=1.67, p=0.038), KRAS (HR=1.81, p=0.018), JAK2 (HR=2.44, p=0.001), U2AF1 (HR=2.30, p<0.001), CBL (HR=2.29, p=0.012), del(17p)/-17/i(17q) (HR=1.90, p=0.004), del(7q)/-7 (HR=1.40, p=0.045), del(20q) (HR=1.58, p=0.034), and MECOM rearrangements (HR=2.38, p=0.002) were independently associated with inferior OS (Table S3 and Figure 1D).

Despite strong co-occurrence between TP53 mutation and complex karyotype or del(17p)/-17/i(17q), collinearity analyses confirmed that these features could be appropriately modeled as separate covariates without compromising model discrimination or calibration (Figures S8A–D and S9A–D). In the full multivariable model, complex karyotype lost independent prognostic significance, whereas both TP53 and del(17p)/-17/i(17q) remained independently prognostic and were retained in the final risk model.

To develop a genetic risk model for predicting OS, we next created a scoring system based on these variables. A score of −1 point was assigned for the presence of favorable mutations: IDH1 or IDH2, while a score of +1 was assigned for each of the following adverse features: TP53, KRAS, JAK2, U2AF1, CBL mutation, del(7q)/-7, del(17p)/-17/i(17q), del(20q), or MECOM rearrangements (Figure S1). Nested bootstrap analysis (B=1,000) confirmed the robustness of retained predictors, with high selection frequency and consistent risk directionality across resampled datasets; sensitivity analyses further supported the inclusion of all 11 variables, including IDH1, in the final model (Figure S10A–C).

Patients were stratified into three risk groups based on their cumulative score: low-risk (<0), intermediate-risk (0), and high-risk (>0). A total of 10.5% (n=53), 43.5% (n=220), and 46.0% (n=233) were classified into low-, intermediate-, and high-risk, respectively. KM analysis demonstrated that this model effectively stratified patients by OS, with median OS of 54.2, 22.3, and 7.5 months in the low-, intermediate-, and high-risk groups, respectively (p<0.0001), and a concordance index (C-index) of 0.648 (Figure 2A).

Figure 2. Comparison of the Moffitt prognostic model with pre-existing models.

OS outcomes stratified by Moffitt genetic risk categories (A), and risk classifications defined by existing prognostic models (B-F). P-values were calculated using the log-rank test.

Comparison of Moffitt Genetic Risk Model with Other Models

We next compared the performance of our genetic risk model with previously published models, including ELN 2022, ELN 2024, refined ELN 2024, the Mayo, and the BEAT-AML risk models (13, 15–18), using OS censored at Allo-SCT. In the ELN 2022 model, the median OS for favorable, intermediate, and adverse risk groups were 55.7, 25.2, and 9.8 months, respectively, with a C-index of 0.57 (p<0.0001). However, the difference between the favorable and intermediate groups was not statistically significant (p=0.127) (Figure 2B). Similarly, in the ELN 2024 model, the median OS for low-, intermediate-, and high-risk groups was 21.4, 16.8, and 6.7 months, respectively, with a C-index of 0.612 (p<0.0001), but the low- vs. intermediate-risk comparison reached only marginal significance (p=0.0514) (Figure 2C).

The refined ELN 2024 model demonstrated clearer stratification among risk groups, with median OS of 33.3, 15.4, and 6.6 months for favorable, intermediate, and adverse risk groups, respectively, and a C-index of 0.640 (p<0.0001) (Figure 2D). The Mayo model showed stepwise decline in median OS, with 34.3, 23.2, and 7.1 months for low-, intermediate- and high-risk groups, respectively, and a C-index of 0.647 (p<0.0001) (Figure 2E). The BEAT-AML model yielded median OS of 25.7, 15.4, and 6.6 months for favorable, intermediate, and adverse categories, respectively, with a C-index of 0.633 (p<0.0001) (Figure 2F). While all three models achieved statistically significant risk stratification, the median OS differences between low- and intermediate-risk groups were 17.9 months (refined ELN 2024), 11.1 months (Mayo), and 10.3 months (BEAT-AML), substantially lower than the 31.9-month difference observed with our model (HR=0.45, p=0.003) (Figures S11A–B).

To understand the basis for improved OS discrimination, we analyzed reclassification patterns between the Moffitt model and each comparator model. A substantial fraction of patients in each risk group was reclassified by our model (Figures S12A–E and S13A–C). Among patients upclassified to higher risk categories, adverse cytogenetics (del(7q)/-7, del(17p)/-17/i(17q), del(20q), or MECOM rearrangement) were present in 0.0–33.3% of low/favorable-risk groups and 25.0–61.5% of intermediate groups (Figures S13D–F). Reclassification patterns were statistically significant for all models except Mayo (Table S4), and upclassified patients had significantly higher rates of adverse cytogenetics compared to non-upclassified patients, particularly in intermediate-risk groups across all models (p<0.001) and in favorable risk groups for ELN 2024 (p=0.028) and Mayo (p=0.007) (Table S5). These findings suggest that integrating cytogenetic data identifies biologically high-risk disease missed by mutation-centric approaches, thereby improving the discriminatory power of the risk model. Because ELN 2024 and refined ELN 2024 were developed using uncensored OS, we repeated our analyses without censoring at Allo-SCT; our model maintained the strongest OS stratification (Figures S14A–F).

Independent Prognostic Value and External Validation of the Moffitt Genetic Prognosis Model

We next assessed whether the Moffitt genetic risk model retained prognostic significance after accounting for clinical covariates. Risk separation persisted irrespective of prior HMA exposure (Figure 3A–B). In a multivariate Cox model adjusting for AML subtype, prior HMA exposure, and Allo-SCT modeled as a time-dependent covariate, the Moffitt risk classification remained independently prognostic for OS (high- vs. low-risk, HR=4.90, p<0.001; intermediate- vs. low-risk, HR=1.98, p=0.008) (Figure S15). To validate our findings, we analyzed a combined external cohort of newly diagnosed AML patients (n=126) treated with HMA/VEN at Nice University Hospital (n=95) and the University of Maryland (n=31). In the Nice cohort, the median age was 78 (44–89) years and 44.2% were female; in the Maryland cohort, the median age was 73 (30–89) and 45.2% were female (Table S6). Pooled across cohorts, the most frequent mutation was ASXL1 (28.6%, n=36), followed by TET2 (27.0%, n=34), TP53 (26.2%, n=33), RUNX1 (20.6%, n=26), and SRSF2 (18.3%, n=23) (Table S6). The most common cytogenetic abnormality was complex karyotype (27.0%, n=34), followed by del(5q)/-5 (20.6%, n=26), trisomy 8 (18.3%, n=23), del(7q)/-7 (14.3%, n=18), and del(17p)/-17/i(17q) (11.9%, n=15). All patients received HMA/VEN as first-line therapy; the median treatment duration was 2 cycles (range, 1–36) and 61.1% (n=77) achieved CR/CRi (Table S6). Using our risk model, 6.3% (n=8), 44.4% (n=56), and 49.2% (n=62) of patients were classified as low-, intermediate-, and high-risk, respectively (Figure 3C). OS differed significantly by risk group, with median OS of 24.7, 17.4, and 4.3 months (p=0.0005, C-index 0.626) for low-, intermediate-, and high-risk groups, respectively.

Figure 3. Stratified Kaplan-Meier Analyses by Prior HMA Exposure and External Validation of the Moffitt Prognostic Model.

(A-B) OS by Moffitt genetic risk category in HMA-naïve patients (A; n=423) and in patients with prior HMA exposure (B; n=83). (C) External validation cohort. OS in newly diagnosed AML patients treated with HMA/VEN, stratified by Moffitt risk categories (total, n=126; Nice University Hospital, n=95; University of Maryland, n=31). P-values were calculated using the log-rank test.

Effect of Allogeneic Stem Cell Transplant and Subsequent Therapy

Finally, we assessed the impact of Allo-SCT on OS. Because a higher proportion of patients achieving CR/CRi underwent Allo-SCT (23.4% vs. 4.6%), raising concern for immortal-time and selection bias, we fit a multivariable Cox model treating Allo-SCT as a time-dependent covariate. In this model, Allo-SCT remained independently associated with improved OS (HR=0.20, p<0.001) (Figure S15). To evaluate effect modification by genetic risk, we additionally fit a time-dependent Cox model including an Allo-SCT x risk-group interaction. Stratified estimates from this interaction model showed benefit in the intermediate-risk (HR=0.27, p=0.001) and high-risk (HR=0.18, p<0.001) groups, whereas the formal interaction test was not statistically significant (p=0.32), indicating no strong evidence that the magnitude of benefit differed across risk categories (Figure S16A). In a sensitivity analysis restricted to IDH1/2-mutant patients (n=81), subsequent IDH1/2 inhibitor exposure following relapse of disease (n=12, 14.8%) did not significantly influence OS (HR=1.73, p=0.305) (Figure S16B).

DISCUSSION

HMA/VEN has significantly improved treatment outcomes in AML, particularly in patients who are unfit or ineligible for intensive chemotherapy (3, 10). However, treatment response and OS remain highly variable, largely due to the underlying heterogeneity of AML subtypes, genetic alterations, and clinical factors (3, 10). While the ELN 2022 classification provides comprehensive risk stratification for patients receiving intensive chemotherapy by systematically integrating both mutations and individual cytogenetic abnormalities, its performance is limited in the HMA/VEN setting (18). In response, several HMA/VEN-specific prognostic models have recently been developed (12–17). However, these models focus predominantly on mutations, with ELN 2024, refined ELN 2024, and BEAT-AML excluding cytogenetic data entirely, while the Mayo model categorized cytogenetics into broad ELN-adverse risk group without evaluating individual abnormalities as independent predictors. This mutation-centric approach may underestimate the prognostic impact of specific cytogenetic lesions, leading to potential misclassification of risks. Further, substantial cross-study variability persists; markers designated as favorable in ELN 2024 or refined ELN 2024 (i.e., NPM1, IDH1, IDH2, DDX41) are not uniformly retained as favorable in other models. In addition, the prognostic significance of genetic markers enriched in sAML/tAML populations remains insufficiently characterized due to their underrepresentation in prior cohorts. Therefore, a comprehensive prognostic model that systematically integrates both mutations and individual cytogenetic abnormalities across diverse AML subtypes—analogous to the ELN 2022 system—remains an unmet need for the growing population of patients treated with HMA/VEN.

In our studies, we identified specific mutational and cytogenetic alterations associated with both favorable and adverse treatment response. Consistent with previous studies (12, 13, 17), adverse-risk genetic factors, such as TP53 mutation and del(5q)/-5, del(7q)/-7, del(17p)/-17/i(17q), and complex karyotype were more frequently observed in patients who failed to achieve CR/CRi. In contrast, NPM1, IDH2, and DDX41 mutations were more common among patients who achieved CR/CRi. Multivariate analysis identified IDH2 and DDX41 mutations as favorable predictive markers for treatment response, whereas JAK2 mutation and del(7q)/-7 were associated with inferior response. Notably, although TP53 mutation and copy number loss in chromosomes 5 and 17 were predictive of poor response in univariate analysis, their significance was lost in multivariate analysis. These findings suggest that a substantial proportion of patients harboring these traditionally adverse genetic alterations may still respond to HMA/VEN therapy, although the duration of response appears to be shorter (22).

Consistent with prior HMA/VEN-specific models (12–17), we found IDH1 and IDH2 mutations to be associated with favorable OS, whereas TP53 and KRAS mutations were linked to adverse OS. Notably, NRAS mutations were not associated with adverse OS in our cohort, consistent with previous reports (16, 17), suggesting that the molecular consequences of KRAS and NRAS mutations may differ. Understanding these divergent signaling and biological effects will be important for future refinement of risk stratification.

Along with these previously established adverse-risk mutations, we also identified JAK2, U2AF1, and CBL mutations as being significantly associated with inferior OS. Of note, 42.7% of patients in our cohort had sAML, in which these mutations are more commonly observed (23, 24). The higher proportion of patients with these mutations in our cohort may account for the discrepancies between our findings and those reported in previous studies. Similarly, NPM1 and DDX41 are classified as favorable in ELN 2024 and refined ELN 2024 but are not designated as favorable in the Mayo or BEAT-AML model (13, 16, 17). In our study, both were associated with improved response but did not confer a statistically significant OS benefit, suggesting that their prognostic impact may depend on co-occurring genetic alterations and cohort composition (13, 16, 17).

Notably, 66.4% of patients in our cohort harbored cytogenetic abnormalities, likely reflecting the higher proportions of patients with sAML and tAML. Multivariate analysis identified del(7q)/-7, del(17p)/-17/i(17q), del(20q), and MECOM rearrangements as poor prognostic markers for OS. However, other ELN 2022 defined adverse cytogenetics, such as del(5q)/-5, complex karyotype, and KMT2A-rearrangements, were not significantly associated with OS in our analysis. While del(17p)/-17 and MECOM rearrangements have been previously associated with poor prognosis in AML overall (10, 18, 25), our findings highlight their potential prognostic significance in the context of HMA/VEN therapy and underscore the value of evaluating individual cytogenetic lesions rather than relying solely on broader ELN category.

Our integrated scoring model, validated in an external cohort (C-index 0.648 and 0.626, respectively), outperformed the ELN 2022 and ELN 2024 systems and showed comparable performance to the BEAT-AML, refined ELN 2024 and Mayo models, as reflected by similar C-indices. Notably, our model demonstrated superior separation between low- and intermediate-risk groups (median OS difference 31.9 months, p=0.002; HR=0.45, p=0.003), exceeding all comparator models. A key distinction of our approach is the incorporation of specific cytogenetic abnormalities as individual predictive markers, which allowed for better discrimination of OS between low- vs. other risk groups (Figure S17A–C). Each adverse cytogenetic alteration was independently evaluated and scored, allowing for more refined risk stratification. This approach reclassified a subset of patients initially categorized as favorable or intermediate by the comparator models into higher risk groups based on the presence of additional adverse cytogenetic features, thereby contributing to the observed enhanced separation between risk groups. Nonetheless, model performance should be interpreted in the context of our cohort’s genetic composition, which differed from those used to derive comparator models. These findings underscore the need for collaborative efforts to develop a unified and definitive prognostic model, validated across larger and more diverse AML datasets.

Data regarding which subgroups derived the most benefit from Allo-SCT following HMA/VEN remain limited. Our analysis demonstrated an OS benefit from Allo-SCT in both intermediate- and high-risk groups, though the small number of low-risk patients who underwent transplantation precludes definitive conclusions in that subgroup.

We acknowledge several limitations. First, the external validation cohort was relatively small, reducing precision within strata. Independent validation in larger cohorts is warranted. Second, del(7q) and −7 were pooled in our analysis, as FISH-based assessment cannot reliably distinguish interstitial deletions from whole-chromosome loss; whether these entities carry distinct prognostic implications under HMA/VEN warrants further investigation. Third, our model primarily focused on baseline genetics. Clinical characteristics (e.g., AML subtype, prior HMA therapy) and treatment-related factors (e.g., response) can modify risk over time. Further, MRD was assessed in a subset of responders, precluding definitive evaluation of its prognostic contribution. Prospective studies integrating baseline clinical and genetic factors with standardized response and MRD assessments may further improve risk prediction.

In conclusion, our study identified key genetic alterations linked to clinical outcomes in AML patients treated with HMA/VEN and developed a prognostic model that improves existing risk stratification systems based on intensive chemotherapy. These findings underscore the importance of further validation in larger and diverse cohorts to refine predictive tools for this patient population.

METHODS

Patients

We retrospectively identified study subjects at the Moffitt Cancer Center (Moffitt) from 1/2018 to 4/2025. Eligible patients were age 18 years and older with pathologic diagnosis of AML as defined by the 5^th World Health Organization (WHO) classification (26), who received HMA/VEN for their front-line treatment and had available mutation data (Figure S1). Patients with secondary AML (sAML) and therapy-related AML (tAML) were also included in our studies, but patients with acute promyelocytic leukemia (APL) with t(15;17)/PML::RARA were excluded. For all cases included, bone marrow (BM) biopsies, as well as molecular profiles, were reviewed by hematopathologists (L.Z. and D.Q.) to confirm the diagnosis. Responses to therapy were evaluated according to ELN 2022 AML response criteria (18). Clinical variables including age, sex, complete blood count (CBC) with differential, BM and peripheral blood (PB) blast counts, cytogenetics, mutation profiles, ELN 2022 and ELN 2024 risk categories were captured at the time of diagnosis. This study was conducted in accordance with the Declaration of Helsinki and was approved by the MCC Scientific Review Committee and Institutional Review Board (MCC #21675). Written informed consent was obtained from all patients under the Total Cancer Care (TCC) protocol.

Assessment of cytogenetics

We performed cytogenetic analyses using standard Trypsin-Giemsa banding technique by Laboratory Corporation of America (Burlington, NC, USA) in accordance with the International System for Human Cytogenomic Nomenclature, 2016 (ISCN 2016, Karger). Further, fluorescence in situ hybridization (FISH) using probe sets designed for del(5q), del(7q), +8, del(17p), del(20q), t(8;21)/RUNX1::RUNX1T1, t(15;17)/PML::RARA, t(9;22)/BCR::ABL1, inv(16)/CBFB::MYH11, MECOM or EVI1/3q26, and MLL/11q23 were used as per the manufacturer’s instructions (Vysis, Downers Grove, IL, USA). For validation cohorts, cytogenetic assessment was likewise performed by local laboratories using conventional karyotyping and FISH with comparable probe sets.

Targeted Exome Sequencing

A hybridization capture-based NGS assay targeting 98 genes associated with myeloid diseases was performed as previously described (27). Genomic DNA was extracted from BM mononuclear cells (BM-MNCs) or PB mononuclear cells (PB-MNCs), then subjected to targeted genome sequencing using an Illumina NextSeq 500 or NextSeq 550 (Illumina Inc., San Diego, CA) instrument. The lower limit of detection was a variant allele frequency (VAF) of 3% or greater for point mutations and 5% or greater for indels. Variants were annotated and assessed for pathogenicity using the following public databases: OncoKB, PubMed, ClinVar, the Catalogue of Somatic Mutations in Cancer (COSMIC), dbSNP (www.ncbi.nlm.nih.gov/SNP), and the Exome Variant Server (Table S7). For the validation cohorts, targeted NGS used the Oncomine Myeloid Assay GX v2 (45 genes) in the Nice University Hospital cohort (28) and the Oncomine Myeloid Assay (40 genes) in the University of Maryland cohort (Table S7).

Statistical Analysis

Baseline clinical and disease variables (e.g., age, sex, cytogenetics, and mutations) were described at AML diagnosis. The primary endpoint was OS based on different risk stratifications. OS was defined from the first HMA/VEN dose to death or last follow-up. For descriptive Kaplan-Meier (KM) analyses, patients were censored at allogeneic stem cell transplant (Allo-SCT). When evaluating transplant effects, Allo-SCT was modeled as a time-dependent covariate. Survival curves were estimated by KM and compared with the log-rank test. Associations between mutations or cytogenetic abnormalities and treatment response or OS were assessed using univariate and multivariate logistic regression or Cox proportional hazard models, reported as odds ratios (ORs) or hazard ratios (HRs) with 95% confidence intervals (CIs). Variables present in <2% of patients were excluded from these analyses. Each mutation or cytogenetic abnormality was analyzed as an independent, non-mutually exclusive covariate. For TP53 mutations, prognostic impact by allelic status was examined using three categories: wild type, mono-allelic (a single pathogenic TP53 mutation without del(17p)/-17/i(17q)), and multi-hit mutations (defined as the presence of TP53 mutation in conjunction with del(17p)/-17/i(17q) or ≥2 TP53 mutations).

For risk model development, candidate genomic features were prespecified (Figure S1). In the univariate screen, we evaluated only mutations and cytogenetic abnormalities with frequency ≥2% to avoid unstable estimates from rare events. Variables advanced to the multivariate Cox model if the univariate p<0.05. For borderline associations (0.05≤p<0.1), variables advanced only if classified as ELN 2024 favorable (NPM1, IDH1, IDH2, DDX41) to reduce exclusion due to cohort-specific sparsity or confounding. For ELN 2024 favorable variable advanced via this borderline rule, sensitivity analyses compared models with vs. without the variable to evaluate incremental fit (likelihood-ratio test), discrimination (C-index), and risk re-classification. Pairwise dependence was assessed by cross-tabulation with φ (phi) and χ² tests (or Pearson correlation where appropriate), and multivariable collinearity was evaluated using variance inflation factors (VIF) in the Cox model (with nested likelihood-ratio comparisons of models omitting each term as a secondary check). Internal robustness was examined using a nested bootstrap (B=1,000) that reapplied the full selection pipeline in each resample to quantify re-selection frequency, effect-direction stability, and HR distributions. Categorical variables were compared using Chi-square or Fisher’s exact test and continuous variables with Student’s t-test. Two-sided p<0.05 was considered statistically significant. All analyses were performed using R (v4.5.0). The development and validation of the prognostic model were designed, conducted, and reported in accordance with the TRIPOD guidelines (29).

SIGNIFICANCE.

This study identifies key mutational and cytogenetic markers associated with treatment response and overall survival in AML patients receiving HMA/VEN therapy. By integrating these genetic markers, our new prognostic model offers improved risk stratification, guiding personalized treatment strategies for patients ineligible for intensive chemotherapy.

Data Availability

Baseline clinical and genetic data for the study cohort are provided in Table S8. Additional requests for access to de-identified data can be made via email correspondence to the corresponding author.