Clinical impact of the genomic landscape and leukemogenic trajectories in non-intensively treated elderly acute myeloid leukemia patients

To characterize the genomic landscape and leukemogenic pathways of older, newly diagnosed, non-intensively treated patients with AML and to study the clinical implications, comprehensive genetics analyses were performed including targeted DNA sequencing of 263 genes in 604 patients treated in a prospective Phase III clinical trial. Leukemic trajectories were delineated using oncogenetic tree modeling and hierarchical clustering, and prognostic groups were derived from multivariable Cox regression models. Clonal hematopoiesis-related genes (ASXL1, TET2, SRSF2, DNMT3A) were most frequently mutated. The oncogenetic modeling algorithm produced a tree with five branches with ASXL1, DDX41, DNMT3A, TET2, and TP53 emanating from the root suggesting leukemia-initiating events which gave rise to further subbranches with distinct subclones. Unsupervised clustering mirrored the genetic groups identified by the tree model. Multivariable analysis identified FLT3 internal tandem duplications (ITD), SRSF2, and TP53 mutations as poor prognostic factors, while DDX41 mutations exerted an exceptionally favorable effect. Subsequent backwards elimination based on the Akaike information criterion delineated three genetic risk groups: DDX41 mutations (favorable-risk), DDX41wildtype/FLT3-ITDneg/TP53wildtype (intermediate-risk), and FLT3-ITD or TP53 mutations (high-risk). Our data identified distinct trajectories of leukemia development in older AML patients and provide a basis for a clinically meaningful genetic outcome stratification for patients receiving less intensive therapies.


INTRODUCTION
Acute myeloid leukemia (AML) is a disease primarily affecting older individuals with a median age of 68 years at diagnosis [1].With the advent of the hypomethylating agents (HMA), such as azacitidine and decitabine [2][3][4], and the HMA-based combination therapies with venetoclax [5,6] or ivosidenib in IDH1-mutated AML [7], there have been significant advances in the therapy of older, unfit patients with AML.Nevertheless, outcome of older individuals remains unsatisfactory due to frequent comorbid conditions and in particular the underlying disease genetics [8].
The genomic landscape of AML has mostly been studied in younger patients who received intensive chemotherapy [9][10][11][12].Data of older patients with newly diagnosed AML receiving less intensive therapies are scarce.However, the available data indicate that the genomic landscape is different from that of younger patients [13][14][15].Also, the widely used European LeukemiaNet (ELN) genetic risk classifications [16,17] have been developed exclusively based on patients who received intensive chemotherapy and may warrant modifications in older patients receiving HMA-based therapies.
The international randomized multi-center phase 3 ASTRAL-1 trial evaluated safety and efficacy of the second-generation HMA guadecitabine (SGI-110) in treatment-naïve AML patients not eligible for intensive chemotherapy in comparison to a treatment choice of either decitabine, azacitidine, or low-dose cytarabine [18].The trial did not meet its co-primary endpoints of improvement of complete remission rate or of overall survival by guadecitabine in the overall population while ad hoc exploratory analysis favored guadecitabine in patients receiving at least 4 cycles.The ASTRAL-1 trial is the largest study performed in older patients receiving less intensive therapy and offers a unique opportunity to gain insights into the genomic landscape and its clinical impact of AML in older patients.

METHODS Patients
In the ASTRAL-1 trial (NCT02348489), 815 patients with previously untreated AML and unfit for intensive chemotherapy were randomly assigned to guadecitabine or treatment choice consisting of azacitidine, decitabine, and low-dose cytarabine [18].604 patients gave informed consent for molecular studies and were included in this study; of these patients, n = 278 bone marrow and n = 326 peripheral blood samples were available.Baseline patient and disease characteristics are given in Table 1.Similar to the overall trial population, there was no difference in outcome by treatment arm in the 604 patients (Supplementary Fig. S1).
Data from conventional cytogenetic analysis and/or fluorescence in-situ hybridization was available for 558 (92%) patients.Data on copy-number variations based on conventional cytogenetics were complemented by data from Illumina HumanMethylation EPIC BeadChip arrays analysis which was performed in 477 patients.
The study was approved by the institutional review board at each participating center.Written consent was given by all patients according to the Declaration of Helsinki.

Gene mutation analyses
Targeted sequencing (mean read depth: 905x [369-1379]) was performed on the entire coding region of 263 genes involved in hematologic disorders (Supplemental Table S1) using SureSelectXT HS from Agilent Technologies (Santa Clara, CA, USA) for library preparation and a pairedend sequencing (read length: 2 ×100 base pairs) on a HiSeq 2000 platform (Illumina, San Diego, CA, USA).All sequencing data were analyzed using an in-house computational pipeline [19].The analysis of internal duplications in the FLT3 gene was performed using GeneScan-based fragment length analysis [20].All gene mutation analyses were performed centrally at Ulm University Hospital.

Statistical analyses
The Kaplan-Meier method was used for estimation of survival curves.Differences between survival curves were tested by logrank tests.Mutual exclusivity and co-occurrence of mutations and cytogenetic aberrations with a frequency of ≥ 4% were tested via Fisher's exact test using a modified Benjamini-Hochberg procedure to control the false discovery rate (FDR) for discrete test statistics [21].Pairwise associations were measured by odds ratios.Bradley-Terry models were used to assess temporal order of acquisition of mutations (again reduced to those with a frequency of ≥ 4%) based on variant allele frequencies [22].Oncogenetic trees as proposed by Desper et al. [23] were used to model the dependencies in the sequence of mutation acquisition (for mutations with a frequency of ≥4%) in AML.Trees were developed by an algorithm using non-parametric bootstrap resampling (1000 samples) [24].To validate the results, a second approach based on maximum likelihood estimation was applied [25].For clustering observations based on Dirichlet processes, mutations with a frequency of ≥1% were selected.Bayesian Dirichlet processes, which use a mixture model with an infinite prior distribution to model the number and proportion of clusters, were used to classify patients into subgroups in an unsupervised manner [26].Prognostic factors for overall survival (OS) were identified using a multivariable Cox regression with subsequent backward elimination based on the Akaike information criterion (AIC).The full model included age, sex, ECOG performance status ( ≥ 2 vs.<2), white blood cell counts (log10-transformed), treatment arm, and mutations with a frequency of ≥4%.The results of the reduced model were visualized using predicted survival curves according to Kalbfleisch and Prentice [27] for all combinations of the resulting mutations while clinical variables were fixed at the median or mode.The prediction performance of the full model and the reduced model was compared to a random survival forest (500 trees) based on the same variables and a basic model including the five clinical variables only.The comparisons were based on (integrated) Brier scores using internal validation via 0.632+ bootstrap (R = 1000) as implemented in the package pec.All analyses were conducted in R version 4.

Oncogenetic tree model using NGS mutational data
To delineate leukemia-initiating trajectories in order to understand whether the leukemogenic in elderly patients differs from that of younger, we constructed an oncogenetic tree by inferring the sequence of mutation acquisition and illustrating the relationships among genetic alterations (Fig. 2) [24].
The algorithm yielded a stable and reproducible oncogenetic tree with five branches with ASXL1, DDX41, DNMT3A, TET2, and TP53 emanating from the root.The data suggest that these genes represent the initiating events which predispose to additional events with further distinct branches.
The tree originating from the ASXL1 node gave rise to several individual clones with EZH2, NRAS, RUNX1, SRSF2, and U2AF1; BCOR, JAK2, KRAS, PHF6, and SF3B1 originated from RUNX1; and IDH2 R140 and STAG2 from the SRSF2 node.Noteworthy, the branches originating from the ASXL1 node include 8 of the 9 genes defining the ICC category "AML with myelodysplasia-related gene mutations" [28].
In case of the DNMT3A node, the oncogenetic tree indicated that the descendent clone will most likely acquire an NPM1 mutation prior to a FLT3-ITD or a PTPN11 alteration.In general, nodes with mutations in signaling genes (NRAS, KRAS, JAK2, FLT3, PTPN11) were located at the very end of each branch, representing last events in driving leukemia, which was also confirmed by a separate analysis using Bradley-Terry models (Supplementary Fig. 1 Mutational and cytogenetic landscape of older patients with acute myeloid leukemia.Mutational (A) and cytogenetic (B) profile, as well as distribution of AML according to the International Consensus Classification (C) in 604 older patients with newly diagnosed AML.A Genes with mutations present in ≥4% of AML.B Cytogenetic abnormalities present in ≥4% of AML; abnormalities were determined by conventional chromosome analysis, fluorescence in-situ hybridization, and EPIC-array analysis.Frequencies given in percent.
Interestingly, the branches with DDX41 mutations, which were often germline, and TP53 terminated at the node without further branching suggesting that mutations in these genes do not depend on or constitute preconditions to further alterations.

Clustering using mutational and cytogenetic data
To further understand the genetic subgroups in elderly AML patients, now using cytogenetic in addition to molecular data, we performed clustering by hierarchical Dirichlet processes (HDP) which resulted in 5 distinct groups (class 1-5) and an additional group (class 0) whose mutational and cytogenetic data did not lead to a precise classification (Fig. 3).
Class 4 (n = 9 cases only) was characterized by various structural and numerical chromosomal alterations.
Upon further analysis of class 0 (6%, n = 32), three distinct, nonoverlapping subgroups were identified.The first subgroup was defined by DDX41 alterations, harboring the largest proportion of all DDX41 mutations (n = 18, 58% of all DDX41 mut in the cluster analysis) and was predominantly associated with a normal karyotype (89%, 16 of 18).The second subgroup harbored 6 of the 8 core-binding-factor leukemia cases.The third group (n = 5) was defined by MECOM(EVI1) rearrangements.

Current classifications and risk stratifications applied to older AML patients
Based on our observation that leukemia development, the mutational and cytogenetic landscape in elderly AML patients differ from those of younger individuals we aimed to evaluate the informative value of the current classifications and risk stratification in our cohort of older, not-intensively treated population.
Regarding prognostic value, both risk classifications did not provide clinically meaningful separation of the survival curves (Fig. 4A  and B).In both, the adverse-risk group did worst, however, the intermediate-and favorable-risk curves were largely overlapping.
Overall survival according to the ICC categories of AML.The generated survival curves did not show a clear separation especially between AML with mutated NPM1, AML with myelodysplasia-related gene mutations, and AML with myelodysplasia-related cytogenetic abnormalities (Fig. 4C).With regards to overall survival, patients with "AML with mutated TP53" had a dismal prognosis with a 2-year OS of 6%, followed by "AML with myelodysplasia-related gene mutations", "AML with mutated NPM1", "AML with myelodysplasia-related cytogenetic abnormalities" with 2-year OS rates of 13%, 18%, 21%, respectively (Supplementary Table S6)."AML not otherwise specified (NOS)" were associated with a comparatively favorable prognosis with a 2-year OS rate of 32%.The most commonly mutated gene in "AML NOS" was DDX41 in 27% (16/60) of the cases and accounted for 49% (16/33) of all DDX41 mutated cases (Supplementary Table S4).The second largest portion with n = 10 of all n = 33 DDX41 mutated cases were found in the category "AML with myelodysplasia-related gene mutations".To evaluate the prognostic impact of DDX41 mutations within those ICC subgroups we built Kaplan-Meier estimates (Supplementary Fig. S5).Mutated DDX41 improved the survival in "AML NOS" and was also associated with improved outcome within the category of "AML with myelodysplasia-related gene mutations".

Association of genetic landscape with outcome
In order to address the limitations of the current risk stratifications and classifications in predicting prognosis we assessed whether the genetic markers detected in our study exert an impact on Fig. 2 Oncogenetic tree model using a modeling algorithm by Szabo.In an oncogenetic tree model, the root represents a state of disease before occurrence of mutations.Each node represents a gene mutation and each branch represents a distinct biologic clone thus illustrating the different clones and temporal sequence of acquisition of mutations.Fig. 3 Unsupervised hierarchical clustering using mutational and cytogenetic data.Hierarchical Dirichlet Processes were employed to build the cluster plot.Mutations and cytogenetic data that was present ≥1% of AML were used.The distribution of ELN risk strata and ICC entities across the newly identified classes is indicated by different colors.
survival.Furthermore, we sought out to identify a new genetic risk classification that is better suited for older, non-intensively treated AML patients.

Impact of mutational and cytogenetic features on OS.
In a multivariable Cox model including known clinical prognostic factors, treatment arm, and all gene mutations ("full model"), all clinical variables except treatment arm (Fig. 5) showed a significant detrimental effect on OS (increasing age, male sex, higher ECOG status, increasing WBC).Mutations with an adverse effect were TP53, FLT3-ITD, SRSF2, and in trend U2AF1.In contrast, DDX41 showed a highly beneficial effect on OS (HR 0.41; Fig. 4 Prognostic value of current AML classifications and of proposed new genetic risk categories for older AML patients.Overall survival by European LeukemiaNet (ELN) 2017 (A) and ELN 2022 (B) genetic risk classification, by ICC categories (C) and proposed risk categories for older AML patients who are not eligible for intensive chemotherapy derived from multivariable Cox models of the current study (D).A, B Both risk classifications did not provide clinically meaningful separation of the survival curves.In cases previously stratified according to the 2017 ELN stratification, the 2022 ELN stratification entailed a change of strata in 14% (n = 82) of the patients, with re-classification to a more adverse-risk category in 13% (n = 75) and to a more favorable in 1% (n = 6) of the cases.D After applying a backwards elimination algorithm on the multivariable Cox model, a reduced prognostic model yielded genetic factors with a significant impact on OS: DDX41, FLT3-ITD, and TP53.The results of the reduced model were visualized using predicted survival probabilities for all combinations of the resulting mutations while clinical variables are fixed at the median or mode.This led to a stratification into three risk categories: DDX41 mut as favorable, DDX41 wt / TP53 wt /FLT3-ITD neg as intermediate, and TP53 mut or FLT3-ITD pos as adverse.0.24-0.69);and for ASXL1, a trend towards longer OS was observed.
A second model (Supplementary Fig. S6) with addition of cytogenetic abnormalities, yielded very similar results with the difference that instead of TP53 mutations "complex karyotype" was found as a highly significant adverse factor, which may be explained by the strong interaction between the two genetic variables (Supplementary Fig. S7).
OS curves, survival times for selected gene mutations and chromosome abnormalities are given in the Supplement (Supplementary Figs.S8 and S9; Supplemental Tables S7-S10).
Identification of biologically relevant mutations for survival outcome.After applying backward elimination on this model, WBC Fig. 5 Impact of clinical, mutational, and cytogenetic features on overall survival.Forest plot displaying hazard ratios based on results from Cox regression analysis using clinical variables [age, sex, ECOG performance status ( ≥ 2 vs. <2), white blood cell counts (log 10 -transformed)], treatment, and all mutations with a frequency of ≥4%; HR hazard ratio, CI confidence interval.
and ECOG status as well as DDX41, FLT3-ITD and TP53 remained as prognostic variables in the Cox model for OS.To demonstrate the estimated effects of these variables, the survival probabilities evolving from this model for an exemplary patient with median WBC (3.6 × 10 9 /L) and an ECOG status <2 are depicted in Fig. 4d conditional on the mutational status of DDX41, FLT3-ITD and TP53. Figure 4D is not to be interpreted as a Kaplan-Meier curve as it does not depict estimated survival probabilities for the mutation groups, but rather as predicted survival curves resulting from the multivariable Cox model.Notably, DDX41, TP53 mutations, and FLT3-ITD virtually never co-occurred in a patient (DDX41/FLT3-ITD, n = 0; TP53/FLT3-ITD, n = 2; TP53/DDX41, n = 2).The curves show that AML with mutated DDX41 have a markedly favorable OS, while AML with FLT3-ITD or with mutated TP53 have a similar adverse impact on OS.AML without DDX41 and TP53 mutations and without the presence of FLT3-ITD are associated with an intermediate prognosis.
Although the predicted survival curves show a clear separation, the full model as well as the reduced model yield no improvement in terms of the prediction error compared to a model including the clinical variables only (Supplementary Fig. S10).A model that was able to show better prediction performance compared to the simple clinical model was a random survival forest based on the same variables as the full model.However, predictions from a random forest are gained by averaging the predictions from the single trees, which can be complex models themselves.Hence, it is not possible to find a simple representation of the random forest, which can be used to define new clinically relevant risk groups without losing prediction accuracy.
Nevertheless, we argue that the mutations indicating a prognostic effect are biologically meaningful in the context of AML in elderly patients and even though the model shows no improvement in prediction performance, it does serve as hypothesis generating in terms of future investigations.

DISCUSSION
This large study enabled us to identify the mutational landscape, to derive distinct disease trajectories and to describe the clinical implications in older, unfit patients with newly diagnosed AML.One important strength of the study is that the data were derived from a single large prospective clinical trial of patients largely treated with an HMA.
In line with previous small studies in older AML patients [13][14][15], the genetic profile largely differed from that of younger patients, with a predominance of mutations in genes involved in CHIP (ASXL1, TET2, SRSF2, and DNMT3A) (Fig. 1A-C), and of adverse-risk molecular and cytogenetic features accounting for 73% by the 2022 ELN risk classification (Fig. 1A-C, Table 1) [17].
CHIP mutations are considered a marker of aging as they accumulate throughout life in individuals without a hematologic phenotype [30][31][32][33].The high prevalence of CHIP mutations in this cohort implies that leukemia of the elderly frequently arises from clonal hematopoiesis, which is supported by our Bradley-Terry model reconstructing the order of temporal acquisition of mutations (Supplementary Fig. S4) [22].
To further analyze leukemia-initiating trajectories, we applied an oncogenetic tree modeling algorithm inferring the sequence of mutation acquisition [24] which produced 5 branches pinpointing to distinct founding clones (Fig. 2).Of note, the ASXL1 node contained 8 of the 9 genes defining "AML with myelodysplasiarelated gene mutations" and 7 of the 8 genes previously reported to define AML with antecedent MDS or myeloproliferative neoplasm [28,34], providing further evidence for this category as a distinct AML subset.The DNMT3A node gave rise to subbranches defined by NPM1, IDH1, IDH2 R172 mutations.This temporal sequence is in line with previous observations that preleukemic DNMT3A mutations precede the acquisition of leukemia-driving mutations and may persist in remission after intensive chemotherapy [34][35][36][37].Further founding clones encompassed TP53 and DDX41 mutations that did not give rise to further subclones, on the one hand providing evidence that these mutations define distinct entities, on the other hand indicating that these leukemia-initiating events are dependent on less transformative events.In fact, both AML with TP53 and DDX41 mutations had significantly less co-mutations compared to AML with wildtype TP53 and DDX41, a finding consistent with previous observations [38][39][40][41].HDP clustering of genetic changes aiming at identifying distinct biologic subgroups largely recapitulated our oncogenetic tree model and provided further support for the recently published ICC categories.
Mutations of DDX41 have recently been recognized as one of the most common predisposition genes for hereditary AML/MDS syndromes and occurring predominantly in older male patients [39][40][41][42].In our cohort, 5.5% of patients exhibited a DDX41 mutation, with about two-thirds of cases predicted to be of germline origin, which is in line with previous reports in younger patients [39][40][41][42].Overall, the available data, i.e., the findings of DDX41 mutations being of frequent germline origin, representing leukemia-initiating events, and the paucity of co-occurring molecular and cytogenetic lesions, provide strong evidence that AML with mutated DDX41 may represent a clinico-pathologic entity of AML.
With regard to the clinical impact of the genomic landscape, both 2017 and 2022 ELN risk stratifications failed to identify clinically meaningful risk groups in this cohort of older patients who received less intensive therapy (Fig. 4A, B).A recent retrospective analysis within the VIALE-A [6] and the preceding phase 1 trial [5] also indicated that ELN risk groups are not prognostic in patients treated with azacitidine plus venetoclax [43].
To identify clinically relevant prognostic factors, we performed multivariable Cox regression analysis for OS including clinical variables and gene mutations.The clinical variables age, sex, and in particular ECOG performance status and WBC retained a strong prognostic impact.Among gene mutations, DDX41 mutations were identified as a highly prognostic favorable marker, whereas FLT3-ITD, SRSF2, and in particular TP53 mutations were unfavorable prognostic factors (Fig. 5).In the reduced model, ECOG performance status, WBC, as well as DDX41, FLT3-ITD, and TP53 mutations remained as the most relevant prognostic factors.Figure 4c shows a representation of the model in which three genetically defined risk groups are illustrated by predicted survival curves that are derived from the multivariable model, a favorablerisk group defined by DDX41 mutation, an adverse-risk groups defined by FLT3-ITD or TP53 mutations, while the remaining patients determine the intermediate-risk group.
An association of DDX41 mutations with a favorable outcome in AML has recently been described, although in younger intensively treated patients [39][40][41][42].Similarly, in MDS patients mutated DDX41 has been shown to be associated with a favorable OS after hypomethylating agents, despite being a predictor of AML transformation [44].In contrast to the data in younger patients, NPM1 mutation was not associated with a favorable prognosis (Fig. 4C, Supplementary Fig. S8), which in part may be related to a different co-mutation pattern found in older AML patients, e.g., almost half of NPM1-mutated AML (45%) had co-occurring myelodysplasia-related gene mutations, most frequently in SRSF2 and ASXL1 (Supplementary Fig. S11).The same is true for NRAS and KRAS mutations that have been suggested to confer resistance to less intense treatment regimens, however did not provide independent prognostic information in our multivariable analysis [14,43,45,46].
Although HMA monotherapy is no longer standard of care for older, unfit patients, similar prognostic factors appear to be relevant for patients treated with azacitidine plus venetoclax, thus, our model may retain its validity also for these patients.In retrospective post-hoc analyses of the VIALE-A trial, both AML with FLT3-ITD and mutated TP53 have been shown to be associated with unfavorable outcome even when treated with the combination [47,48].TP53 multihit and FLT3 mutations have also been identified as top genetic predictors of adverse outcomes in the Molecular International Prognostic Scoring System for Myelodysplastic Syndromes [44].
Outcome data for azacitidine/venetoclax-treated patients exhibiting a DDX41 mutation are not yet available.Data from a recent retrospective analysis suggest DDX41 mutation might be associated with effectiveness of monotherapy with HMA [49].Thus, based on available data, one can assume that these patients will belong to a group of high benefit and long survival times also when treated with the azacitidine and venetoclax.
In conclusion, the data from our study provide unprecedented insights into the genomic landscape of AML in older patients.We identified distinct trajectories of leukemia development, providing support for the new ICC AML disease categories as well as for DDX41 mutations defining a new clinico-pathologic entity of AML.Our proposed genetic risk model will need to be validated in independent data sets to evaluate whether it may be applicable more broadly also to doublet and triplet HMA-based combination therapies, including targeted agents which are currently in clinical development for patients ineligible for intensive chemotherapy [50].

Table 1 .
and Supplementary TableS2 and S3.Categorization of cases according to the International Consensus Baseline characteristics of the 604 AML patients.