Newly diagnosed adult AML patients, defined as 18 years or older, were included for this study. Patients with acute promyelocytic leukemia were excluded. The patient selection and bone marrow sample acquisition was done between October 2016 and July 2018. At the end, a total of 52 patients were included for analyses. Their medical records were reviewed and analyzed for demographics, baseline disease characteristics, factors related to treatment, response to treatment, adverse events including treatment related mortality, and survival outcomes. Data available up to March 2019 were used. Their bone marrow samples collected at AML diagnosis were subjected to flow cytometry and NGS.

The diagnosis of AML was made according to the WHO classification of hematopoietic neoplasms [11], which requires identification of 20% or more leukemic blasts in the bone marrow. Secondary AML was defined as AML following myelodysplastic syndrome (MDS) or myeloproliferative neoplasm (MPN) confirmed prior to the diagnosis of AML. Cytogenetic studies were performed onsite, whose satisfactory performance was monitored by a national external quality assurance scheme. Bone marrow cells were cultured for 24 hours then karyotype was analyzed using the standard G-banding technique. The karyotypes were constructed and chromosomal abnormalities were reported in accordance with the International System for Human Cytogenetic Nomenclature [12]. Prognostic grouping of cytogenetics was performed according to Southwest Oncology Group (SWOG) criteria [13]. Risk stratification was mainly based on cytogenetics, and NGS results for those with available data, according to 2019 National Comprehensive Cancer Network (NCCN) guidelines [14].

The study was conducted in compliance with all national and international ethical standards for research with humans and for research using radiopharmaceuticals. This study was conducted according to the Declaration of Helsinki and was approved by the Institutional Review Board of Seoul National University Hospital (IRB No. H-1902-140-101). Written informed consent was obtained from all patients. All authors had access to the study data and reviewed and approved this study.

Mononuclear cells were isolated from the patient’s bone marrow aspirates by ficoll gradient centrifugation and cryopreserved in serum-free medium. Stored cells were thawed to Iscove’s Modified Dulbecco’s Medium and washed with fluorescence-activated cell sorting (FACS) buffer (2% fetal bovine serum, Dulbecco’s phosphate-buffered saline). Cells were stained with following anti-human monoclonal antibodies: CD45-APC/cy7 (557833), CD34-APC (555824), CD38-BV421 (562445), CD90-PE (555596), CD123-PE/Cy7 (560826), and CD45RA-PerCP/Cy5.5 (563429) (BD Bioscience). Analyses were performed on a FACSCanto II (High Throughput Sampler, BD Bioscience, Franklin Lakes, NJ, USA) and FlowJo version 10.0 program (BD Bioscience). First, the expression profiles of CD34 and CD38 markers were analyzed (Fig. 1A). Then, the phenotypes of LSCs were further analyzed per methods proposed by Goardon et al. [15]; multipotent progenitor (MPP)-like LSC (CD34⁺CD38^–CD90^–CD45RA^–), lymphoid primed multipotent progenitor (LMPP)-like LSC (CD34⁺CD38⁺CD90^–CD45RA⁺) and granulocyte-macrophage progenitors (GMP)-like LSC (CD34⁺CD38⁺CD123⁺CD45RA⁺). The CD45 and side scatter properties were used for gating bulk leukemic blast cells and sub-gated from the cells [16], as shown in Fig. 1B. End results showed that there was a dominantly expressed LSC phenotype in each patient (Fig. 1C). In some cases, two or more LSC phenotypes co-existed. The percentage of cell fractions satisfying each LSC phenotype markers in CD34⁺ cells was calculated.

The DNA capture probes for 76 target genes were designed using the Agilent SureDesign web-based application (https://earray.chem.agilent.com/suredesign/, Agilent Technologies, Santa Clara, CA, USA). The target regions included protein coding exons with 10 bp intron flanking regions and hot spot regions of the 20 genes involved in recurrent translocations. DNA was extracted on a Chemagic 360 instrument (Perkin Elmer, Baesweiler, Germany). The genomic DNA was sheared using Covaris S220 focused-ultrasonicator (Covaris, Woburn, MA, USA). We used 50 ng of total input genomic DNA. A library preparation was performed according to Agilent’s SureSelectQXT Target Enrichment protocol (Agilent Technologies). Paired-end 150-bp sequencing was using NextSeq 550 Dx platform (Illumina, San Diego, CA, USA). Targeted sequencing raw data was obtained in FASTQ format.

The produced sequencing data was analyzed using the Seoul National University Hospital First Hemic Panel Analysis Pipeline (FHPAP). First, we performed the quality control of the FASTQ file and analyzed only the data that passed the criteria. Pair-end alignment to HG19 reference genome was performed using BWA-men and the GATK Best Practice [17,18]. After finishing the alignment step, an “analysis-ready BAM” was produced and a second quality control is performed to determine if further variant calling is appropriate. In the pipeline, SNV, InDel, CNV, translocation, FLT3 ITD, and KMT2A PTD were analyzed using at least more than two analysis tools including in-house and open-source software. Along with in-house software, GATK UnifiedGenotyper, SNVer, LoFreq were used for SNV/InDel detection [18-20], Delly and Manta for Translocation discovery [21,22], and THetA2 and CNVKit for purity estimation and CNV calling [23,24], respectively. FLT3 ITD and KMT2A PTD were analyzed by only the in-house software. Detected variants was annotated by SnpEff with various database such as RefSeq, COSMIC, dbSNP, ClinVar, and gnomAD. Then germline variant was filtered using the population frequency of these databases (> 1% population frequency) [25-30]. Finally, the variants were confirmed throughout a comprehensive review of a multidisciplinary molecular tumor board.

The relapse free survival (RFS) and overall survival (OS) curves were estimated using the Kaplan–Meier method. The RFS was derived from the date AML diagnosis to that of relapse or death from any cause, while the OS was defined as the time from AML diagnosis to death from any cause. If patients survived, RFS and OS was censored on the last date of follow-up when no death or relapse was confirmed. Complete remission rate, based on the revised recommendations of the Center for International Blood and Marrow Transplant Research (CIBMTR) group, with induction was compared. Induction TRM was also analyzed, and defined as mortality during induction chemotherapy. The differences between groups were assessed using a Mann–Whitney test or one-way analysis of variance for continuous variables, and Pearson chi-square test for categorical variables, as indicated. All data were analyzed using the SAS Enterprise Guide version 6.1 (SAS Institute, Cary, NC, USA) and the statistical software R (www.r-project.org). A p value of < 0.05 was considered statistically significant.

The baseline characteristics of the enrolled patients are presented in Table 1. There were 40 de novo AML and 12 secondary AML patients. Among secondary AML patients, there were four post-MPN AML patients and eight post-MDS AML patients. The median age of the whole group was 62 years.

When different LSC phenotypes were considered, secondary AML group showed higher fraction of CD34⁺CD38^– cells (p = 0.0245) and MPP-like LSCs (p = 0.0037) compared to de novo AML group (Fig. 2A). When secondary AML was further classified into post-MPN versus post-MDS, there was a statistically significant increase in percentage of MPP-like LSCs in post-MPN AML compared to post-MDS AML (p = 0.0485) (Fig. 2B). Collectively, secondary AML, particularly post-MPN AML, was associated with higher fraction of the MPP-like LSC phenotype compared to its de novo counterpart.

To determine the association between the age and LSC phenotype compositions, patients were divided into two group according to age at AML diagnosis at cutoff of 65 years (Supplementary Fig. 1). The CD34⁺CD38^– cell fraction was higher for younger patients (i.e., ≤ 65 years), but there were no differences in LSC phenotypes between the two age groups.

While data on conventional cytogenetics was available in all patients, the NGS data was not available in 14 patients due to lack of quality-assured samples. The majority of the patients were classified as intermediate risk by both cytogenetics risk stratification and combined risk stratification based on NGS data (Table 1). No correlation was seen between cytogenetics risk and LSC phenotypes (Supplementary Table 1). On the other hand, LMPP-like LSC was predominant in the poor risk group by combined risk stratification (p = 0.0052) (Fig. 3). This was true for both de novo AML (n = 28, p = 0.0401) and secondary AML (n = 10, p = 0.0871) patients (Supplementary Table 2).

For survival analyses, the patients were grouped according to their dominant LSC phenotypes (Fig. 4). There were ten patients in MPP-like LSC group, 11 in LMPP-like LSC group, and 31 in GMP-like LSC group. For the entire cohort, the mean RFS was 10.8 months and mean OS 12.3 months (Supplementary Table 3). Supplementary Fig. 2 represents the survival curves of the patients undergoing treatment per dominant LSC phenotype. In light of small number of patients with heterogeneous treatment schema, only those who underwent treatment regardless of intensity were included for survival analyses. Furthermore, to compensate the innate differences in survival among different AMLs, de novo AML and secondary AML were analyzed separately. There was no correlation between RFS or OS and LSC phenotype in either group.

The mutational landscape of enrolled patients is shown in Fig. 4. We looked for specific mutations that were enriched in each LSC phenotype group. Mutations in CEPBA (6 out of 7), DNMT3A (4 out of 7), and IDH1 (3 out of 5) were predominantly clustered in GMP-like LSC group. On the other hand, mutations in NRAS (4 out of 6) and KRAS (3 out of 4) were dominantly found in MPP-like LSC group, while mutations in TP53 (4 out of 5) and ASXL1 (5 out of 6) were often found in LMPP-like LSC group.

This observation was verified by comparing the LSC phenotype composition of specific mutation carriers versus non-carriers (i.e., wild-type). The differences in LSC expression per CEBPA, DNMT3A, and IDH1 mutations were not statistically significant. Interestingly, KRAS mutants showed higher fraction of MPP-like LSC compared to KRAS wild-types (Fig. 5). Also, the fraction of GMP-like LSC was significantly lower in KRAS mutants. NRAS mutants also showed similar tendencies, but the difference did not reach statistical significance. For TP53 and ASXL1 mutants, LMPP-like LSC was present in higher fraction. For TP53 mutants, the expression of LMPP-like LSC was significantly higher by 3.4-folds (p = 0.0276) compared to wild-type. ASXL1 mutants also showed trends towards higher expression of LMPP-like LSC (p = 0.0654).