Acute myeloid leukemia in the next-generation sequencing era

Acute myeloid leukemia (AML) is the most frequent form of acute leukemia in adults. While the prognosis is still dismal for many AML patients, the recent progress in diagnostics and risk stratification combined with the emergence and approval of novel therapeutic substances has significantly improved treatment outcomes for selected AML subgroups [1]. Most knowledge about AML diagnostic and treatment recommendations is based on controlled clinical intervention trials (CCIT). Despite all the advantages of CCIT, a downside of this approach is that specific patient groups are often excluded from these studies (i.e., frail, older and/or comorbid patients). Consequently, establishing real-world retrospective and prospective registries has evolved as an important quality control, yielding complementary value to the data from CCIT [2‐5]. The relevance and transferability of CCIT to patient cohorts not represented in CCIT can be derived from these registries. Even for patient collectives included in CCIT, these registries enable an essential validation step within a real-world setting. Finally, they serve as essential quality assurance tools, enabling the comparison of treatment and outcome data between individual centers.

Next-generation sequencing (NGS) enables the parallel analysis of many genes in a high-throughput fashion [6, 7]. It revealed important insights into the pathogenesis of AML, which spurred the development of novel and targeted therapeutic approaches. Some of those, including the FLT3 inhibitors midostaurin [8] gilteritinib [9] and quizartinib [10], the IDH inhibitors enasidenib [11], olutasidenib [12] and ivosidenib [13] or the BCL2 inhibitor venetoclax [14], have already achieved approval by the U.S. Food and Drug Administration (FDA) and/or European Medicines Agency (EMA) as mono-treatment or combination treatment. Moreover, it refined AML diagnosis and risk stratification and is nowadays well implemented in the routine clinical work-up of AML. Even more, comprehensive NGS analysis is now a prerequisite to fulfilling the 2022 classification and risk stratification algorithms [1]. Most NGS data describing the molecular landscape of AML are derived from patient cohorts treated within CCIT. Given the recent establishment of NGS, real-world data about its feasibility in the clinical routine and data about the molecular landscape of patients not included in CCIT are scarce.

The Medical University of Graz, Austria (Med Uni Graz) is a tertiary cancer care center serving a population of approximately 1.2 million people. It was among the first hematology centers in Europe to include NGS profiling in the routine clinical diagnostic work-up of AML in 2013 [15]. Therefore, we aimed to perform a retrospective, unselected data analysis of AML in patients treated within the NGS era and to analyze epidemiological data and treatment outcomes as part of a real-world observational study. In addition, we analyzed the feasibility, usefulness and results of NGS, particularly in patient groups not represented in CCIT. Moreover, we used this registry to determine potential progress in AML treatment compared to data before the NGS era and/or before the introduction of recently licensed novel treatment approaches.

The study cohort consisted of unselected consecutive adult patients treated for AML between January 2013 and April 2023 at the Division of Hematology of the Med Uni Graz. Diagnostics and risk stratification of myeloid neoplasms were performed according to standard criteria [1]. To evaluate AML treatment progress over time, we also collected data from unselected consecutive adult AML patients treated at the same institution between 2002 and 2008. For NGS, genomic DNA extracted from bone marrow (BM) aspirates or biopsies was analyzed for mutations in up to 49 myeloid-associated genes using an Ion Torrent next-generation sequencing (NGS) platform (Thermo Fisher Scientific, Waltham, MA, USA), as described previously [7, 16‐18]. In more detail, we analyzed the whole coding region of BCOR, BCORL1, CEBPA, DDX41, DNMT3A, ELANE, ETNK1, ETV6, GATA2, GNB1, HAX1, NF1, PHF6, PIGA, PPM1D, PRPF8, SF3B2, SFRP1, SRP72, STAG2, TP53, and ZRSR2, as well as mutational hotspot regions of NPM1, ASXL1, BRAF, CALR, CBL, CSF3R, CXCR4, ETNK1, EZH2, FLT3, IDH1, IDH2, JAK 2, KIT, KRAS, MPL, NRAS, PTPN11, RUNX1, SETBP1, SF3B1, SRSF2, STAT3, STAT5B, TET2, U2AF1 and WT1 genes. All diagnostic, treatment and outcome data and the results from NGS were retrieved from openMEDOCS, a regional hospital-based documentation system and continuously recorded in a dedicated internal database (Research, Documentation and Analysis database, RDA; https://​imi.​medunigraz.​at/​services).

Importantly, we also aimed for continuous biobanking of high-quality diagnostic and follow-up leukemia samples. Therefore, leukemic blasts were isolated from peripheral blood and/or BM by Ficoll (GE HealthCare, Chicago, IL, USA) density gradient centrifugation and subsequently stored in DMSO in the vapor phase of liquid nitrogen (Supplementary Fig. 1) [19‐21].

The study was approved by the institutional review board of the Med Uni Graz (EK 30-464 ex 17/18 and EK 35-079 ex 22/23). Every patient signed an informed consent for data recording and analysis as well as for the biobanking of the respective specimens.

For statistical analysis, group differences in continuous variables were compared by Mann-Whitney U test. In contrast, Fisher’s exact test was employed to compare all dichotomous variables in patient specimens. Kaplan-Meier survival curves were used to generate figures for overall survival; differences between groups were assessed using the log-rank test. Univariable Cox proportional hazards models were used to investigate the association of group (study cohort, NGS cohort) and other parameters with the incidence of all-cause mortality. A multivariable analysis was conducted on parameters with p < 0.20. Hazard ratio (HR) with 95% confidence interval (CI) are reported. A P-value < 0.05 was considered to be statistically significant. The analyses were performed with GraphPad Prism vs. 10.1.2 (Boston, MA, USA) while the statistical software R (version 4.3.2; https://​www.​R-project.​org/​) with the survival package (version 3.5‑7; https://​CRAN.​R-project.​org/​package=​survival) was used for the survival analysis (Cox models).

In this study, we performed a retrospective and unselected data analysis of AML in patients treated within the NGS era at a tertiary care center. We aimed to record and analyze epidemiological data and treatment outcomes as part of a real-world observational study. In addition, we aimed to analyze the feasibility and results of NGS, particularly in patient groups not represented in CCIT. Moreover, we use this registry to analyze potential progress in AML treatment compared to data before the NGS era and/or before the introduction of recently licensed novel treatment approaches.

The use of NGS has revolutionized the diagnostics and risk stratification of AML and helped to identify novel targets for molecularly tailored treatment approaches. Consequently, it is a necessary tool for fulfilling the ELN 2022 risk stratification, where molecular information has gained importance compared to previous versions. Unfortunately, most NGS data describing the molecular landscape of AML are derived from patient cohorts treated within CCIT, and real-world data about its feasibility in the clinical routine and data about the molecular landscape of patients not included in CCIT are scarce. We performed a retrospective, unselected data analysis of 284 AML patients treated between 2013–2023 at a tertiary care cancer center. NGS was routinely performed at diagnosis during this time. We initially focused on the feasibility and benefits of NGS within a clinical routine setting. We primarily used BM aspirates for NGS analyses; however, in cases where aspirates were unavailable NGS was performed from formalin-fixed paraffin-embedded (FFPE) BM trephine biopsies. This strategy resulted in evaluable NGS results in all cases allocated to NGS analyses. Importantly, we also present data showing that NGS results from BM aspirates and FFPE trephine biopsies are comparable, indicating that our approach is feasible for a routine clinical setting. In this respect, it is also worth mentioning that NGS was successful significantly more often than conventional cytogenetics and that 10 cases with missing cytogenetics could still be risk-stratified according to the molecular analyses. Despite these results, it is important to emphasize that the joint assessment of NGS and cytogenetics is essential for accurate risk stratification in many patients, particularly as our data suggest that the period between BM sampling and availability of the results is comparable between these assays. Another aspect of NGS applicability in the clinical routine is the processing time, defined by the number of days passing between BM sampling and the availability of a finished report. We show that the median NGS processing time in a clinical routine setting is only 16 days and substantially decreased with the advances of NGS technology to 10 days in 2022. Considering recent observations that a delayed treatment start of more than 15 days in stable patients does not negatively affect outcomes [34], our data indicate that it is possible to wait for molecular profiling in a routine clinical setting. This statement is further supported by our observation that NGS identified a molecularly targetable lesion in 107/284 (38%) of patients. Taken together, our data clearly demonstrate that comprehensive NGS profiling is feasible within a routine clinical setting. It is a valuable tool for ELN risk stratification and targeted treatment decisions and should be an integral part of diagnosis at every center treating AML. In this respect, our data will provide argumentation help for implementation in those centers where organizational and/or financial issues precluded such an approach in the past.

Considering the molecular landscape of AML, we observed similar mutation patterns as described in CCIT and previous real-world cohorts [22‐25]. This information is also relevant, as it strengthens the relevance and applicability of data from CCIT within this setting. Additionally, our molecular data confirm previous reports that the mutational landscape differs between younger and older patients. We have noted an increase of mutations in TET2, IDH2, SRSRF2 and TP53 in older patients, which matches well with previous reports [17, 25, 35‐37]. The same is true when mutations were clustered to functional classes. In these analyses, we observed enrichment of mutations affecting DNA methylation and the spliceosome in old people, whereas aberrations activating cellular signalling were more frequent in younger patients, which matches well with the reported increased frequency of receptor tyrosine kinase and cellular signalling mutations in these age groups [25, 38]. Finally, we also observed higher frequencies of MDS-related mutations in old people, which is in accordance with data from clonal evolution data of AML [22, 26, 39, 40].

Also, regarding outcome data, our cohort matches well with internationally published data. Firstly, our analyses of OS again show that any treatment is better than no treatment and that survival can be extended with all treatment types currently available [41]. In this respect, we could also compare outcome data with a historical cohort of 163 AML patients treated at the same institution in the pre-NGS era. As low-intensity treatment with HMA or HMA/VEN was unavailable at this time, we focused on ICT regimens in these analyses. In agreement with other registries, we show that survival has significantly improved in recent times [5, 42, 43]. While there will certainly be an influence of improved supportive medicine, including ameliorated intensive care treatment plans for AML patients, our subanalyses demonstrated that the increased use of allografting and better response rates to first line therapy are major determinants of this success. Considering allo-HSCT, it might be postulated that improved risk stratification through NGS helped to better identify patients for allografting. In addition, we show that the increased frequency of allo-HSCTs performed in the more recent cohort is also due to a higher age limit for this complex therapeutic approach. Indeed, median age at transplantation was significantly higher in the more recently treated cohort. Considering improved response rates in more recent times, we have also seen the advent of novel molecularly targeted treatment approaches in the NGS era. We have performed exploratory analyses and tested two of these substances, the FLT3 inhibitor midostaurin and the CD33 antibody-drug conjugate GO; these drugs are added to conventional ICT regimens [8, 27, 28]. Midostaurin improved CR/CRi rates and survival, whereas GO did not affect CR/CRi rates but extended survival in favorable/intermediate risk patients without allo-HSCT consolidation. Hence, our data suggest that improved survival might not only mediated by the increased frequency of allografting but also by the success of novel and molecularly targeted treatment approaches. Of course, the smaller sample sizes of the tested cohorts precluding multivariable calculations limit these analyses and necessitate further validations in independent cohorts. One issue to discuss at this point is the observation that midostaurin improved CR/CRi rates in our study. In contrast, no effect on CR rates was seen in the phase 3 RATIFY licensing trial [8]. Interestingly, other real-world observations agree with our study and have seen the same contradictory results with higher CR/CRi rates in the patients treated with midostaurin [44]. As stated in this paper, a possible explanation for this disagreement with the RATIFY trial is that our trial used CR and CRi to define CR rates. In contrast, the RATIFY trial used CR only. Another potential reason might be due to the definition of CR. The RATIFY trial followed a stringent CR definition by only counting CRs occurring on or before 60 days of starting therapy (including two induction cycles); however, an expanded CR definition (CRs during protocol treatment and those in the 30 days following treatment discontinuation) was used in additional subanalyses. It showed significantly higher CR rates in patients randomized to midostaurin compared to placebo [8]. In our study, several CR/CRis occurred during this expanded CR definition, which might also contribute to this discrepancy. Finally, the limitation of our small sample size might have contributed to this discrepancy compared to the RATIFY trial results.

Our database also allowed us to analyze NICT regimens in older and unfit patients. As seen in the recently published VIALE‑A trial [14], adding the BCL‑2 inhibitor venetoclax to HMAs significantly improved the outcome compared to HMAs alone. The success of HMA/VEN is also relevant as it is increasingly used in fit patients replacing ICT. The recent NCCN guidelines for AML enable the use of HMA/VEN in patients with adverse risk stratification [45] as the outcome with ICT is unsatisfactory in these patient cohorts. A series of retrospective real-world data support this approach [46‐49]. Our registry also enabled us to address this issue by comparing the differences between HMA/VEN and ICT in older patients ≥ 60 years and ELN2022 adverse risk. When analyzing OS, we observed that patients treated with ICT performed significantly better; however, when data were censored for allo-HSCT, the advantage of ICT was lost, suggesting that the extended OS in patients treated with ICT within this cohort is mainly mediated through allografting. In agreement with these data, CR/CRi rates were comparable between these groups. Although it has to be noted that these data are limited by the small number of patients treated with HMA/VEN, they further support the NCCN guidelines and the abovementioned retrospective observations. They might lead to the hypothesis that a paradigm shift for older patients with adverse risk classification from ICT-based regimens to the NICT HMA/VEN could happen soon. This could also hold for patients still eligible for allo-HSCT, with HMA/VEN used as an induction regimen instead of ICT. Whether HMA/VEN should be used as maintenance after allografting in such a scenario cannot be answered with this registry and is currently being addressed within the VIALE‑T trial. Of note, HMA/VEN was not used in ICT-eligible patients with ELN2022 intermediate and favorable risk; therefore, conclusions about a comparison between ICT and HMA/VEN outside the adverse risk group cannot be drawn from this study.

Ultimately, we could connect this registry to a state of the art leukemia biobank, which enables the corroboration of therapy-relevant molecular findings, familial analyses, and the establishment of novel methods. In addition, it is a precious resource for leukemia research and is open for internal and external collaboration projects.

Taken together, we present a comprehensive real-world registry of AML patients treated at a tertiary care cancer center in the NGS era that is additionally linked to a leukemia biobank containing high-quality biospecimens of these patients. The registry is a crucial tool for quality control and assurance and also helps to validate data from CCIT in a real-world setting. These data might even help to approach clinical questions that have not been studied in CCIT yet and design relevant CCITs for these questions. One example is the treatment of older patients with adverse risk stratification with ICT or HMA/VEN, where our data support current NCCN guidelines [45] and other real-world registries [46‐49] and suggest HMA/VEN as a valuable option in these patients. Finally, our data report important confirmatory data considering the molecular landscape in AML patients and underline the importance of routinely performed NGS profiling at diagnosis for correct diagnosis, risk stratification, and treatment planning. Despite all these benefits, mentioning this study’s limitations is important. Firstly, the retrospective nature of this study introduces a potential bias regarding patient selection, unequal treatment regimens, maintenance treatment, and others. Additionally, the limited number of patients in this study, particularly in some subgroups, which also precluded multivariable models assessing the impact of confounding factors, is another limiting factor. This is particularly true for analyses of novel agents, including midostaurin, GO, and HMA/VEN (the latter particularly in the adverse risk group). Furthermore, the follow-up of these cohorts is shorter than in the ICT group, introducing another potential bias.