Redefining prognostication of de novo cytogenetically normal acute myeloid leukemia in young adults

About 50% of acute myeloid leukemia (AML) showed normal cytogenetics (CN) with leukemogenesis driven putatively by recurrent mutations. These mutations occur singly or in combination, as dominant clones or subclones 1 – 4 . Induction with daunorubicin and cytarabine is the standard for young and ﬁ t patients, achieving ﬁ rst complete remission (CR1) in 60 – 80% cases. Post-remission strategies include consolidation with high-dose cytarabine and allogeneic hematopoietic stem cell transplantation (allo-HSCT). The latter may reduce the risk of relapse but is associated with mortality and long-term morbidities. The European LeukemiaNet (ELN) guidelines, based on cytogenetic and genetic risk strati


Dear Editor,
About 50% of acute myeloid leukemia (AML) showed normal cytogenetics (CN) with leukemogenesis driven putatively by recurrent mutations. These mutations occur singly or in combination, as dominant clones or subclones [1][2][3][4] . Induction with daunorubicin and cytarabine is the standard for young and fit patients, achieving first complete remission (CR1) in 60-80% cases. Postremission strategies include consolidation with highdose cytarabine and allogeneic hematopoietic stem cell transplantation (allo-HSCT). The latter may reduce the risk of relapse but is associated with mortality and longterm morbidities. The European LeukemiaNet (ELN) guidelines, based on cytogenetic and genetic risk stratification, provide general recommendations on prognostication and allo-HSCT for AML 5 . We performed next-generation sequencing (NGS) for young patients with de novo CN-AML, diagnosed between 2003 and 2019, who were treated with a relatively uniform algorithm to examine the prognostic impact of mutation combinations. Machine learning was used to generate prediction model and its performance was compared with that based on ELN guidelines. Clinical treatment and methodology are described in Supplemental Materials (see also Supplemental Fig. S1).
NPM1 and CEBPA DM mutations were associated with superior CR/CRi rates and RUNX1 and ASXL1 mutations with inferior CR/CRi rates after first induction (Supplemental Table S4). To examine the factors affecting survivals, age, gender, white blood cell count (WCC), daunorubicin dose (60 versus 90 mg/m 2 ), achievement of CR/CRi after induction or salvage chemotherapy, allo-HSCT at CR1 as well as individual gene mutations were analyzed by univariate analysis. Age and WCC varied with LFS, event-free survival (EFS), and OS as continuous functions and were defined as numerical data (Supplemental Fig. S3). High-dose daunorubicin and HSCT at CR1 were associated with superior LFS, EFS, and OS and achievement of CR/CRi was associated with superior EFS and OS, whereas high WCC and FLT3-ITD and DNMT3A mutations were associated with inferior LFS, EFS, and OS (Supplemental Table S5A). High-dose daunorubicin appeared to negate the adverse prognosis of DNMT3A mutations, consistent with previous reports 6 (Supplemental Fig. S4). Subsequently, these factors were evaluated in multivariate analysis. The prognostic impacts of FLT3-ITD and DNMT3A mutations, achievement of CR/CRi, and HSCT at CR1 have remained unchanged but those of high-dose daunorubicin have become largely insignificant (Supplemental Table S5B). NPM1 mutation was associated with superior LFS and EFS but not OS.
NPM1, DNMT3A, and FLT3-ITD were further evaluated for their relative impacts on LFS (Supplemental Fig.  S5) and OS (Supplemental Fig. S6). DNMT3A mutation negated the prognostic impact of NPM1 mutation and FLT3-ITD, attesting to its overriding impact on prognosis amidst co-existing mutations. FLT3-ITD also negated the prognostic impact of NPM1 but not DNMT3A mutation. NPM1 mutation had no significant impact on the adverse prognostic effects of DNMT3A mutation and FLT3-ITD. Their combinations showed variable LFS and OS (Supplemental Fig. S7) and were further categorized into five groups (Supplemental Table S6). Sole NPM1 mutation (Category 1) showed superior LFS and OS while sole FLT3-ITD (Category 4) and presence of DNMT3A mutation (Category 5) showed inferior LFS and OS. Patients of wild type for all 3 genes (Category 2) and with co-existing NPM1 mutation and FLT3-ITD (Category 3) showed intermediate LFS. However, their OS were indistinguishable from that of Category 1 (Fig. 1C, D). When outcomes were censored at HSCT, Category 1 remained superior, Categories 2 and 3 were intermediate, and Categories 4 and 5 remained inferior (Fig. 1E, F). Subgroup analyses were performed to evaluate the prognostic impact of other recurrent mutations on the five categories. IDH1R132H was associated with inferior LFS and OS in Category 2 exclusively (Supplemental Fig. S8). Other mutations had no significant impact on these categories or their occurrences were too low for comparison (Supplemental Table S7).
To examine whether prognostication by ELN 2017 guidelines might apply to young patients with CN-AML, the present cohort was classified according to the stipulated risk groups, based exclusively on gene mutations. High FLT3-ITD was defined by VAF ≥0.33, corresponding to an allelic ratio of ≥0.5 (Supplemental Table S8). There was a trend toward a difference in LFS and OS among the three risk groups. However, it was statistically insignificant ( Fig. 2A, B). We examined the impact of DNMT3A mutation on each ELN-defined risk groups in our patients. DNMT3A mutation negatively impacted on LFS and OS in the favorable (Supplemental Fig. S9A, B) and intermediate risk groups (Supplemental Fig. S9C, D) but not in the unfavorable risk group (Supplemental Fig.  S9E, F). Incorporating DNMT3A mutation into the ELN risk categorization as an unfavorable risk factor separated the three risk groups and significantly improved the risk stratification (Fig. 2C, D).
The genetic makeup of leukemic clones was extremely diverse (Fig. 2E and Supplemental Fig. S10). Of the 401 patients on whom NGS was performed, 383 patients showed detectable mutation of genes in the AML panel with at least 217 distinct clonal subtypes. The most common subtypes comprised co-dominant NPM1 and DNMT3A mutations, usually in conjunction with other co-dominant or subclone mutations. Sole NPM1 (1.00%) or DNMT3A mutations (0.50%) or their codominance without subclones (0.50%) were relatively uncommon. NPM1 mutation was infrequently found in subclones, and in those rare circumstances, the dominant clones were mostly DNMT3A or IDH2R140Q mutations. FLT3-ITD occurred most frequently as subclones. However, in 2.74% patients, FLT3-ITD occurred as the sole mutation, suggesting its role as leukemic driver early in the leukemic hierarchy 7 . CEB-PA DM occurred predominantly as a sole mutation in 5.24% patients. Forty-six patients (11.47%) were negative for all common or ELN risk-defining mutations, viz. NPM1, DNMT3A, FLT3, IDH1/2, CEBPA, ASXL1, RUNX1, and TP53. They showed rare mutations, some of which, including those of spliceosome genes 8 , were dominant and sole mutations, suggesting pathogenetic role in leukemogenesis (Supplemental Materials, Supplemental Fig. S11, and Supplemental Table S9).
The database built up in this study formed a foundation for the development of prediction model (https:// redefiningprognosis.shinyapps.io/denovo_cnaml/) that might inform clinical decision. Its application was highlighted by two hypothetical patients (Fig. 2F). The information provided quantitative measurement of survival benefits of individual patients based on their demographics and genotypes. Its performance was compared with that of the ELN risk stratification-based model based on concordance index. Using the present cohort of 401 patients as a training set, our prediction model showed a 4.6% higher concordance over the ELNbased model (Fig. 2G). To validate these models, a subset of The Cancer Genome Atlas patients comprising 83 de novo CN-AML patients aged ≤60 years was used as a validation cohort. Patients who received HSCT at refractory stage were not included as they were not represented in the training set. Again, our model showed 6.45% higher concordance over ELN-based model. The difference in concordance in both cohorts was statistically significant. We proposed that this multistage model might provide more personalized guidance to inform post-remission therapy with particular reference to allo-HSCT 9 . Our findings corroborated with recent reports demonstrating room to refine risk stratification based on sequencing and transcriptomic results of patients enrolled into clinical trials 10 .
In conclusion, in young patients with de novo CN-AML who received conventional induction chemotherapy, consolidation, and allo-HSCT, incorporation of DNMT3A mutation into risk stratification significantly improved their prognostication. Predication model based on machine learning of our database generated a more and overall survival (D) after incorporating DNMT3A mutation as unfavorable risk group. Leukemia-free survival was the duration from CR to the last follow-up, leukemia relapse, or death. Overall survival was the duration from diagnosis to last follow-up or death. E A bubble diagram showing clonal heterogeneity in CN-AML. Each mutation was represented by a distinct color except the checker that represented any of the rare mutations as shown. The size of each outer bubble (dominant or co-dominant) indicated the prevalence of patients with that genotype. Inner bubble indicated subclone and its size represented the clone size relative to that of the dominant or co-dominant clones. Horizontal bisection of inner bubbles indicated occurrence of either one of the mutations, whereas vertical bisection indicated occurrence of both mutations. F Sediment plots of two hypothetical patients who received allogeneic HSCT at first complete remission (CR1) or not, based on prediction model using machine learning of the present cohort of patients. The shaded areas indicated the time courses of different outcomes. Upper panels: A 25-year-old female patient presenting with white cell counts of 10 × 10 9 /L and genotype category 1 (NPM1 mutation only) who achieved CR1 after first induction and is considering allo-HSCT at CR1. Her chances of leukemia-free survival would be 87% at 2 years and 79% at 5 years post HSCT. If she declines HSCT, the chances would be reduced to 64% at 2 years and 52% at 5 years. Lower panels: A 25-year-old female patient presenting with white cell counts of 100 × 10 9 /L and genotype category 5 (DNMT3A mutation) is considering allo-HSCT at CR1. Her chances of leukemia-free survival would be 68% at 2 years and 55% at 5 years. Should this patient decline HSCT, her chances of surviving the leukemia would become 14% at 2 years and only 5% at 5 years with a 77% likelihood of death in relapse. G Histogram showing concordance index in the present cohort (training set) and a cohort of young patients (≤60 years old) with CN-AML in the TCGA cohort (validation set). The green bars indicated results from the reported prediction model and the yellow bars represented results if we categorize patients based on ELN 2017 risk stratification. The error bars indicated a standard error of mean. ***P < 0.001. personalized tool to guide post-remission therapy. The diverse clonal heterogeneity and the pathogenetic significance of mutation combinations provided important leads for future mechanistic study.