http://orcid.org/0000-0001-5537-6767
Figures
Abstract
Acute myeloid leukaemia (AML) is characterised by phenotypic heterogeneity, which we hypothesise is a consequence of deregulated differentiation with transcriptional reminiscence of the normal compartment or cell-of-origin. Here, we propose a classification system based on normal myeloid progenitor cell subset-associated gene signatures (MAGS) for individual assignments of AML subtypes. We generated a MAGS classifier including the progenitor compartments CD34+/CD38- for haematopoietic stem cells (HSCs), CD34+/CD38+/CD45RA- for megakaryocyte-erythroid progenitors (MEPs), and CD34+/CD38+/CD45RA+ for granulocytic-monocytic progenitors (GMPs) using regularised multinomial regression with three discrete outcomes and an elastic net penalty. The regularisation parameters were chosen by cross-validation, and MAGS assignment accuracy was validated in an independent data set (N = 38; accuracy = 0.79) of sorted normal myeloid subpopulations. The prognostic value of MAGS assignment was studied in two clinical cohorts (TCGA: N = 171; GSE6891: N = 520) and had a significant prognostic impact. Furthermore, multivariate Cox regression analysis using the MAGS subtype, FAB subtype, cytogenetics, molecular genetics, and age as explanatory variables showed independent prognostic value. Molecular characterisation of subtypes by differential gene expression analysis, gene set enrichment analysis, and mutation patterns indicated reduced proliferation and overrepresentation of RUNX1 and IDH2 mutations in the HSC subtype; increased proliferation and overrepresentation of CEBPA mutations in the MEP subtype; and innate immune activation and overrepresentation of WT1 mutations in the GMP subtype. We present a differentiation-dependent classification system for AML subtypes with distinct pathogenetic and prognostic importance that can help identify candidates poorly responding to combination chemotherapy and potentially guide alternative treatments.
Citation: Schönherz AA, Bødker JS, Schmitz A, Brøndum RF, Jakobsen LH, Roug AS, et al. (2020) Normal myeloid progenitor cell subset-associated gene signatures for acute myeloid leukaemia subtyping with prognostic impact. PLoS ONE 15(4): e0229593. https://doi.org/10.1371/journal.pone.0229593
Editor: Francesco Bertolini, European Institute of Oncology, ITALY
Received: November 15, 2019; Accepted: February 10, 2020; Published: April 23, 2020
Copyright: © 2020 Schönherz et al. This is an open access article distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited.
Data Availability: All relevant data are within the manuscript and its Supporting Information files.
Funding: HEJ, 30-06-2015, Karen Elise Jensen Fonden, http://www.kejfond.dk/. JSB, 27251, Det Obelske Familefond, http://obel.com/. MB, 153-2014-10, Innovationsfonden, https://innovationsfonden.dk The funders had no role in study design, data collection and analysis, decision to publish, or preparation of the manuscript.
Competing interests: The authors have declared that no competing interests exist.
Introduction
Compelling evidence demonstrates that acute myeloid leukaemia (AML) is of clonal origin and represents the progeny of a single cell that enters leukaemic transformation due to multiple genetic events that impair cell differentiation and apoptosis and invoke uncontrolled cell proliferation. However, the evolution from the first somatic mutation to full-blown AML is not well mapped. The simplest models predict that each newly acquired somatic mutation during oncogenesis confers a selective advantage that drives successive waves of clonal expansion and deregulated differentiation, with the fittest clone becoming dominant at diagnosis and during relapse. [1,2] The understanding of linage-specific progenitor commitment during AML transformation and subsequent clonal evolution is fundamental to the pathogenesis and treatment of AML. We speculate that the genetic abnormality in clinically relevant clones can be traced back to the normal compartment or cell of origin (COO)–as we have previously shown for lymphoid B-cell malignancies [3–6]–and that knowledge of the COO and its deregulation could provide novel molecular and oncogenic insight into AML subtypes.
Myeloid and lymphoid malignancies are particularly well suited for evaluating the cellular origin on malignant transformation due to our understanding of the normal haematopoietic hierarchy and the availability of analytical tools for the examination of phenotypically defined subpopulations at the single-cell level. Recently, we described a procedure to identify and study the gene expression of immunophenotype-based flow-sorted minor subsets. [7] These subsets can be profiled by global gene expression, and using statistical modelling, we were able to define normal subset-specific B-cell-associated gene signatures (BAGS) for assignment and prognostic evaluation in B-cell malignancies. [4,8,9] We found that the diagnostic heterogeneity in multiple myeloma, chronic lymphocytic leukaemia, and diffuse large B-cell lymphoma reflects a link between differentiation and oncogenesis–a contribution often overlooked.
The phenotypic heterogeneity of AML was initially defined by cytology decades ago, when the French-American-British (FAB) Cooperative Group developed a classification system based on morphologic and cytochemical phenotypes associated with normal end stage myelopoiesis. However, as cytogenetics was introduced, the prognostic impact of the FAB subtypes (M0-M7) became unclear. [10–16] Immunophenotyping demonstrated that poorly differentiated AML blasts are deregulated and differentially disrupted early stem or progenitor cells; however, no uniform leukaemic phenotype has been defined by CD34 and CD38 membrane markers. [17] Here, we hypothesise that the phenotypic heterogeneity of AML is a consequence of deregulated differentiation with transcriptional reminiscence of the normal stem or progenitor COO phenotype. We generated myeloid progenitor cell subset-associated gene signatures (MAGS) and assigned subtypes in clinical cohorts to study their prognostic and potential pathogenetic impacts.
Material and methods
Data sets
This study is based on data sets already published and publicly available. Sorted normal myeloid data from healthy donors were retrieved from the Gene Expression Omnibus Database (GEO) and are referred to as the GSE63270, [18] GSE42519, [19] GSE19599, [20] GSE17054, [21] and GSE19429 [22] cohorts. For GSE63270, [18] and GSE17054, [21] human bone marrow mononuclear cells (BMMCs) from healthy donors were purchased from ALLCELLS (Emeryville, CA) where collection protocols and donor informed consent are approved by an institutional review board (IRB) in compliance to State and Federal regulations. For GSE42519, [19] GSE19599, [20] and GSE19429 [22] BMMCs from healthy donors were collected and analysed after IRB approval and written informed consent was obtained. The samples were previously phenotyped by multiparametric flow cytometry (MFC) and fluorescence-activated cell sorting (FACS) into myeloid subpopulations, and gene expression profiles (GEPs) were obtained using an Affymetrix Human Genome U13z½3 Plus 2.0 Array GeneChip (Affymetrix, Santa Clara, CA). Sorted normal myeloid data were restricted to those five cohorts due to the limited availability of publicly available data sets with comparable FACS and GEP protocols. The subpopulation compositions and cell sorting details are summarised in S1 and S2 Tables. The myeloid subpopulations of interest for the present study were CD34+/CD38- haematopoietic stem cells (HSCs), CD34+/CD38+/CD45RA+ granulocytic-monocytic progenitors (GMPs), and CD34+/CD38+/CD45RA- megakaryocyte-erythroid progenitors (MEPs).
Clinical AML data were retrieved from the GEO database and The Cancer Genome Atlas (TCGA) database and are referred to as the GSE6891 [23,24] and TCGA [25] cohorts, respectively. For GSE6891, [23,24] and TCGA [25] bone marrow aspirates from AML patients were obtained and analysed after patients provided written informed consent in accordance with the Declaration of Helsinki, and the study was approved by all participating institutional review boards. The data sets were chosen because the GEPs were generated in a comparable manner as those of the sorted normal myeloid data sets, and they contained information on age, gender, FAB subtype, cytogenetic risk score, WBC, disease outcome, and genetic aberrations. The metadata of both cohorts are summarised in the S1 Data (S1 and S3 Tables).
Statistical analysis
All statistical analyses were performed with R version 3.3.3. [26] The statistical analysis is summarised here; for comprehensive documentation, see the (S1 Data). This study followed the guidelines of omics-directed medicine [27–29] whenever possible. Prior to the statistical analysis, all gene expression data were background corrected and normalised using the Robust Multichip Average (RMA) [30] algorithm and summarised at the gene level using a Brainarray custom CDF for the Affymetrix Human Genome U133 Plus 2.0 GeneChip (version 20.0). The clinical cohorts were RMA normalised per cohort, whereas the normal myeloid cohorts were combined into one meta-cohort prior to RMA normalisation. Six normal myeloid HSC, GMP, and MEP samples were randomly selected and used as a training cohort. The remaining 38 normal myeloid samples were used as a validation cohort.
The MAGS classifier for the HSC, MEP, and GMP subsets was obtained from regularised multinomial regression with the cell type as the discrete outcome and the median-centred GEP of the training cohort as the explanatory variable. The model was fitted with an elastic net penalty. [31] The regularisation parameters were determined by cross-validation, and the parameters with the lowest multinomial deviance were chosen. Genes with non-zero coefficients were defined as predictive. Based on those genes, assignment probabilities were estimated to identify the most likely MAGS subtype of a sample. The prediction accuracy of the MAGS classification was validated in the validation cohort. To compensate for cohort-based technical batch effects, the validation cohort was median centred and scaled to have the same variance as the training cohort. The same procedure was applied to the clinical cohorts followed by cohort-based MAGS subtype assignment according to the subtype with the highest predicted probability score and assigning 15% of the samples with the lowest assignment probability as unclassified (UC). To improve statistical power, clinical cohorts were combined into a clinical meta-cohort following MAGS assignment. For biological characterisation, the GEPs of the clinical meta-cohort were batch corrected using the empirical Bayes approach, ComBat, [32] implemented in the Bioconductor package “sva” (version 3.18.0). [33]
Survival analyses were performed using the Kaplan-Meier method, log-rank test statistics, and Cox proportional hazards regression analysis for the individual cohorts and the combined clinical meta-cohort. In addition, the prognostic importance of 13 other explanatory variables (FAB subtype, cytogenetic risk score, CEBPA aberrations, FLT3-itd aberrations, FLT3-tkd aberrations, IDH1 aberrations, IDH2 aberrations, KRAS aberrations, NPM1 aberrations, NRAS aberrations, white blood cell count (WBC), age, and cohort) was investigated by univariate Cox proportional hazards regression analysis in the TCGA (N = 122), GSE6891 (N = 439), and meta-cohort (N = 561), limited to samples with complete records for all investigated variables (for detailed information, see S1 Data). MAGS was evaluated as an independent explanatory variable by multivariate Cox proportional hazards regression analysis for overall survival in all three cohorts, including explanatory variables tested with prognostic effects in the univariate regression analysis in the respective cohorts. Cytogenetic risk score could not be investigated as a confounding variable in the meta-cohort due to differences in cytogenetic risk group stratification between the two cohorts. In the GSE6891 cohort, cytogenetic risk score stratification was based on cytogenetic abnormalities only, [23] whereas in the TCGA cohort, cytogenetic risk score stratification was based on cytogenetics and molecular genetics. [25,34,35] Instead, MAGS was tested for prognostic impact against the cytogenetic risk scores independently in the two clinical cohorts.
MAGS validation through biological characterisation
The biological characterisation and identification of biological processes associated with MAGS subtypes were investigated through differential gene expression (DGE) analysis, gene set enrichment analysis, and the identification of subtype-specific mutation patterns for well-documented driver mutations. To increase detection power, analyses were conducted for the meta-cohort, but they were restricted to clinical samples with MAGS assignment probabilities ≥ 0.75, if not stated otherwise.
DGE analysis.
Was performed in R using the limma Bioconductor package (version: 3.26.9). [36,37] To identify MAGS subtype-specific differences, clinical samples assigned to one subtype were compared with the remaining samples (Rest), resulting in the following three comparisons: i) HSC vs. Rest, ii) GMP vs. Rest, and iii) MEP vs. Rest. For summary statistics, p-values were adjusted using the Benjamini and Hochberg algorithm, [38] and genes with p ≤ 0.001 were defined as differentially expressed.
Enrichment analysis.
Was conducted using two different approaches: a classical Gene Ontology (GO) annotation using Fisher’s exact test to identify over-represented GO terms in differentially expressed genes and a computational gene set enrichment analysis (GSEA) approach developed by the Broad Institute that uses a pre-ranked gene list of all profiled genes (S1 Data). The GSEA was performed using the GSEA desktop application (version 3.0) [39] using 2000 permutations of gene set randomisation and the default settings. Gene sets included in the analysis were selected from the Molecular Signature Database (MSigDB, v6.0) [40] using the Hallmark, [39,41] C2-CP, [39] and C3-TFT MSigDB collections. [39,42] Gene sets including fewer than 15 or more than 500 genes were excluded from the analysis. Gene sets with an adjusted p ≤ 0.05 and a false discovery rate (FDR) q-value for normalised enrichment scores (ES) ≤ 0.25 were considered significantly enriched. [39]
Identification of subtype-specific mutation patterns.
Was performed for well-documented driver mutations across seven AML oncogenes (CEBPA, IDH1, IDH2, FLT3 [including both the FLT3-itd and the FLT3-tkd aberrations], NPM1, NRAS, and KRAS) in the meta-cohort, irrespective of the MAGS assignment probabilities but limited to samples with recorded mutation information (N = 587: NGSE6891 = 457, NTCGA = 130). In addition, a second analysis was conducted for 112 genes associated with AML that had been previously characterised and classified. [43] Mutation records for those genes were extracted from exome-wide somatic mutation data available for a subset of the TCGA cohort (N = 130). [25] Potential associations with MAGS subtypes were investigated for each mutation using Fisher’s exact tests with a significance cut-off level of 0.05.
Results
Generation and validation of MAGS
The transcriptomic identity of normal myeloid subsets was validated by principal component analysis (Fig 1). Batch effects were partially removed using RMA normalisation (Fig 1A and 1B), and the subset identity could be confirmed for the three progenitor compartments through subset-specific segregation into discrete clusters (Fig 1C and 1D), allowing subsequent identification of MAGS. The MAGS classifier with the smallest deviance determined by cross-validation consisted of 92 genes (S4 Table, S1 Fig). The HSC subtype signature included 44 predictive genes, 30 of which were subtype-specific (68.2%); the GMP subtype signature included 37 predictive genes, 20 of which were subtype-specific (62.2%); and the MEP subtype signature included 33 predictive genes, 19 of which were subtype-specific (57.6%; Fig 2). The highest overlap of predictive genes was between the HSC and GMP subsets (N = 8), followed by GMP and MEP (N = 6) and HSC and MEP (N = 6). The prediction accuracy of the MAGS classifier was validated using sorted normal myeloid samples (N = 38: NHSC = 26, NGMP = 7, NMEP = 5), showing a prediction accuracy of 78.95% when all samples were assigned to one of the three MAGS subtypes and 90.63% when defining 15% of the samples with the lowest MAGS assignment probability as UC. For both assignment strategies, the prediction accuracy of the GMP and MEP subtypes was 100%. The MAGS assignment inconsistencies were restricted to the HSC subtype (Table 1A and 1B). Moreover, the majority of the samples wrongly assigned belonged to the GSE19429 cohort (six of the eight samples).
(A) The PCA conducted for the meta-cohort of normal myeloid samples that were RMA normalized before combination and includes all myeloid subsets available. (B) The PCA conducted for the meta-cohort of normal myeloid samples that were combined prior to RMA normalization and includes all myeloid subsets available. (C) The PCA conducted for the meta-cohort described in (B) but limited to samples identified as early (HSC) and late (GMP, MEP) hematopoietic progenitors by FACS. (D) The PCA conducted for the meta-cohort described in (B) but limited to samples that were included in the training-cohort and median centered. Each data point represents the expression profile of one sample. In A-B, samples are color-coded based on their cohort identity; for C-D, samples are color-coded according to their myeloid progenitor subset identity (HSC, GMP, MEP). Samples with similar expression profiles will cluster together. Axis labels indicate the principal component (PC) plotted and the proportion of the variance explained by that PC. PCs are derived by orthogonal data transformation to reduce dimensionality and represent the directions of the data that explain a maximal amount of variation.
MAGS assignment of clinical samples and prognostic impact
Clinical AML samples from two independent cohorts of adult patients diagnosed with de novo AML were classified into MAGS subtypes (S3 Table). We allowed 15% of the samples within each cohort to be assigned as UC, resulting in an assignment probability cut-off ≥ 0.71 (TCGA cohort = 0.71, GSE6891 cohort = 0.72). An unambiguous MAGS subtype assignment was achieved, and the subtype frequencies did not vary between the two clinical cohorts (Table 2). Subtype frequencies ranged from 28.1–31.2% in the GSE6981 cohort and 26.4–30.8% in the TCGA cohort when ignoring the UC-assigned samples. Furthermore, the GMP subtype was the most frequently assigned in both cohorts, followed by MEP and HSC.
Two-sided Fishers exact tests were used to determine significantly different distributions across data sets (p = 0.99).
The prognostic impact of the MAGS subtypes was analysed both individually and collectively in a meta-analysis combining the MAGS-assigned samples of the GSE6891 and TCGA cohorts. The MAGS assignment showed a significant prognostic association with overall survival (Fig 3; log-rank test p ≤ 0.001). The lineage-committed MAGS subtypes GMP and MEP had superior prognoses compared with the undifferentiated AMLs captured by the HSC subtype. This was supported by univariate Cox regression analysis conducted for the GSE6891 and the cohort-adjusted clinical meta-cohort, revealing significant differences between the GMP and HSC (GSE6891: HR = 0.63, p < 0.001; meta-cohort: HR = 0.64, p < 0.001), the MEP and HSC (GSE6891: HR = 0.53, p ≪ 0.001; meta-cohort: HR = 0.51, p ≪ 0.001), and the UC and HSC (GSE6891: HR = 0.70, p = 0.03; meta-cohort: HR = 0.60, p < 0.001; Table 3) subtypes. In the TCGA cohort, significant differences were only observed between the MEP and HSC (HR = 0.44, p = 1.9e-03) and the UC and HSC (HR = 0.37, p = 1.6e-03; Table 3) subtypes. Moreover, multivariate Cox proportional hazards analysis conducted for the three cohorts (TCGA, GSE6891, meta-cohort) demonstrated that the MAGS subtypes added significant prognostic information that was not already explained by FAB subtype, cytogenetics, molecular genetics (well-documented driver mutations in CEBPA, FLT3, IDH1, IDH2, KRAS, NPM1, or NRAS), or age (Table 4).
Kaplan Meier survival curves were generated for overall survival and P-values were estimated using a log-rank test. Only samples with complete survival information were included and the number at risk per MAGS subtype are provided for each cohort.
Analyses were performed for overall survival. Samples with missing survival information (TCGA cohort: N = 11) were excluded. Table columns indicate the total sample size (N), the number of patients that died (Events) per MAGS subtype, the associated hazards ration (HR), the 95% lower and upper confidence intervals (CI), and the estimated p-value (P).
Cohorts were limited to samples with complete records for all explanatory variables investigated. Results are shown for the (A) univariate Cox regression analysis per explanatory variable and (B) associated multivariate Cox regression analyses limited to confounding variables tested significant in univariate Cox regression analyses. Table columns are as described in Table 3. Analyses were performed for overall survival.
DGE and functional annotation of enriched gene sets
To assess biological differences between MAGS subtypes, we performed DGE analysis on 573 samples and compared each subtype with the combined other subtypes: HSC vs. Rest, GMP vs. Rest, and MEP vs. Rest. The largest number of differentially expressed genes (DEGs) was identified for the GMP subtype (NDEG = 6414), followed by the MEP (NDEG = 4279) and HSC (NDEG = 4071) subtypes. The most distinct DGE profile (number of subtype-specific DEGs) was discovered for the GMP subtype with 1657 DEGs, followed by MEP with 935 and the HSC subtype with 776 (S2 Fig). The top DEGs for the GMP and MEP subtypes overlapped (HBD, ALAS2, SPTA1, KLF1, EPB42, AHSP, and SELENBP1; S5A–S5C Table). They were upregulated in the MEP subtype and downregulated in the GMP subtype. Moreover, most of those genes were involved in erythrocyte differentiation (KLF1, ALAS2, and AHSP) or erythrocyte membrane or haemoglobin functions (HBD, SPTA1, and EPB42), which indicates transcriptional discrimination between erythrocytes and other cells. Hence, the results provide biological proof of concept that the MAGS classification of clinical AML samples enables separation into megakaryocyte-erythroid linage and granulocytic-monocytic linage COO subtypes. In contrast, the top DEGs associated with the HSC subtype were subtype-specific and did not reflect any lineage commitment.
To further investigate potential functional patterns associated with the MAGS subtypes, gene enrichment analysis was performed for only the DEGs and taking all genes into consideration. Enrichment analysis of DEGs annotated to GO terms associated with biological processes identified 1066 GO terms enriched for the MEP subtype, followed by 697 for the GMP subtype and 556 for the HSC subtype. Moreover, general patterns of enriched biological processes differed between MAGS subtypes, suggesting different pathogenic courses (S6 Table). The DEGs detected for the HSC subtype were mainly enriched for cell cycle and metabolic processes (S6A Table), whereas DEGs in the GMP and MEP subtypes were mainly enriched for immune system processes and cellular responses to external stimuli, suggesting late progenitor cell functions (S6B and S6C Table).
The GSEA revealed a general downregulation of genes associated with gene sets enriched in the HSC (Table 5A) and MEP subtypes (Table 5B), whereas genes associated with gene sets enriched in the GMP subtype were predominantly upregulated (Table 5C). Biological processes associated with subtype-specific enriched gene sets differed between subtypes. For the HSC subtype, enrichment patterns were characterised by a low cell-cycle activity signature with reduced metabolic rates, suggesting low cell proliferation or a prolonged quiescence phase (S7A Table, S3A Fig). The gene set enrichment patterns observed for the GMP subtype revealed a high metabolic activity signature with innate immune activation (S7B Table, S3B Fig), whereas the MEP subtype had a high cell-cycle activity signature with impaired innate immune activity (S7C Table, S3C Fig). Moreover, the MEP subtype was enriched for genes involved in the metabolism of heme- and erythroblast differentiation, which were downregulated in the GMP subtype, further supporting our hypothesis that malignant cells possess transcriptional reminiscence of the COO.
GSEA analysis was conducted for three comparisons: (A) HSC vs. Rest, (B) MEP vs. Rest, and (C) GMP vs. Rest. The total number of enriched gene sets (Total) with p-value ≤ 0.01 and FDR ≤ 0.25, and the numbers of enriched gene sets detected per group within each comparison are presented.
Annotation of genetic mutation patterns
Potential associations between the MAGS subtypes and well-documented mutations in seven AML-associated oncogenes recorded for both the TCGA and the GSE6891 cohort (CEBPA, FLT3 [including both the FLT3-itd and FLT3-tkd aberrations], IDH1, IDH2, KRAS, NPM1, and NRAS) were investigated in the meta-cohort. Two genes, CEBPA and IDH2, showed subtype-specific mutation patterns. Mutations occurring in the CEBPA gene were associated with the MEP subtype (p = 5.79e-08), which was observed especially for the CEBPA double mutation. Mutations detected in the IDH2 gene were more frequently observed in the HSC subtype (p = 0.015; S8A Table). Furthermore, a set of 112 genes previously shown to harbour AML driver mutations were investigated for MAGS subtype-specific mutation patterns in the reduced TCGA cohort (N = 130). Mutations were detected in 68 genes, revealing significant subtype-specific mutation patterns for RUNX1, RUNX1T1, TP53, and WT1. RUNX1 mutations were associated with the HSC subtype (p = 0.005), while mutations detected in the WT1 gene were associated with the GMP subtype (p = 0.031). Mutations detected in the TP53 gene were negatively correlated with the GMP subtype, and mutations in RUNX1T1 were associated with the UC samples (p = 0.02; S8B Table).
Discussion
In AML, GEPs have successfully identified molecular cancer subtypes for stratifying patients into responders vs. non-responders and predicting survival. [19,44–49] These molecular classification systems are generally based on the GEPs of leukaemic cells or on well-documented oncogenic driver mutations and cytogenetic aberrations associated with AML oncogenesis. [16,21,40–45] Here, we examined and validated a classification system using the GEPs of normal myeloid progenitor cell compartments to classify AML into subtypes based on transcriptional reminiscence of the COO. We showed that the MAGS subtypes of AML cases are associated with prognosis. This observation supports the idea that one or more MAGS subtypes have pathogenic impact. The conclusions may be important for future diagnostic phenotyping and the implementation of individual precision therapy, although there are conceptual, molecular, statistical, and clinical considerations that need to be discussed before clinical implementation and validation.
Our concept is that AML heterogeneity is a consequence of deregulated differentiation and that there is transcriptional reminiscence of the COO. Combining MFC, FACS, and GEP methodologies to phenotype the myeloid progenitor cells in normal bone marrow samples enabled the development of MAGS that differentiate between early (HSC) and late (GMP, MEP) progenitors by tracing transcriptional reminiscence expression patterns of the COO in end-stage AML samples. The MAGS classification assigned comparable subtype frequencies to AML samples within and across independent clinical cohorts. In a meta-analysis of 691 adult patients with de novo AML, we demonstrated a significant prognostic association with post-therapy outcome. Moreover, multivariate Cox proportional hazards analyses in the two clinical cohorts as well as in the meta-cohort supported that MAGS subtyping is independent of FAB subtype, cytogenetic risk score (not investigated in the meta-cohort due to different risk score stratifications across cohorts), and molecular genetics (well-documented driver mutations in CEBPA, FLT3 [including both FLT3-itd and the FLT3-tkd aberrations], IDH1, IDH2, KRAS, NPM1, and NRAS). These results suggest that we have identified distinct pathogenic mechanisms that require further investigation.
The prognostic impact of MAGS-assigned subtypes conferred a superior prognosis to the lineage-committed GMP and MEP subtypes compared with the undifferentiated HSC subtype. The adverse association between leukaemic stem cell phenotypes and survival is well documented in primary AML. [50,51] In agreement with our findings, those studies demonstrated that leukaemic stem cell signatures were independent prognostic predictors that were associated with adverse clinical outcome.
The molecular characterisation of MAGS subtype-specific mutation patterns revealed that RUNX1 and IDH2 mutations had non-random associations with the HSC subtype and that there was a tendency for the DNMT3A mutation to be overrepresented in this subtype. Thus, the HSC subtype is associated with well-documented driver mutations in key haematopoietic and epigenetic regulators involved in AML initiation. [25,52,53] In normal haematopoiesis, RUNX1 plays a critical role in HSC compartment maintenance, proliferation, and haematopoietic differentiation, while in malignancy, genetic aberrations in RUNX1 have been associated with highly aggressive AML and poor prognosis. [43,54–57] In contrast, IDH2 and DNMT3A are key epigenetic regulators, and pre-leukaemic driver mutations in these genes occur early in AML development and are associated with the pre-leukaemic HSC compartment. [2,17,52,58] The significant enrichment of stem cell-specific mutations in AML samples with the HSC subtype indicates a consensus between the transcriptional COO detected by MAGS and the genetic COO, suggesting that AML transformation is not accompanied by subsequent differentiation in the HSC subtype or that HSC-like transcriptional gene signatures were reactivated after AML transformation in more differentiated cells. Similarly, CEBPA mutations were overrepresented in the MEP subtype. CEBPA is a myeloid transcription factor involved in the balance between cell proliferation and terminal differentiation, especially granulocyte differentiation. Loss of CEBPA function in AML contributes to leukaemogenesis by blocking granulocytic differentiation, which is accompanied by the increased accumulation of earlier stem and myeloid progenitors as well as erythroid and megakaryocytic progenitors. [59–62] According to the WHO classification and the current ELN guidelines, patients with mutated CEBPA, particularly those with biallelic mutated CEBPA, represent a cohort with favourable prognosis, [54] which is in general accordance with current findings of superior prognosis in the MEP subtype. However, MAGS-based survival curves revealed smaller differences between subtypes than previously reported findings. [63] Although a positive association between MEP and CEBPA aberrations was observed, this is of associative and not definitive nature, as only 24 out of 147 MEP classified cases showed CEBPA mutations and recorded both contain mono and biallelic CEBPA mutations, possibly contributing to a lower overall survival range than previously reported. [63]
Functional annotation revealed that genes involved in cell-cycle activity and metabolic processes were downregulated in the HSC subtype. This suggests slower cell proliferation or even a prolonged quiescence phase compared with the GMP or MEP subtype. The poor prognosis of the HSC subtype, thus, might reflect the inefficacy of antiproliferative chemotherapeutics, such as cytarabine, in eradicating slow or non-proliferating leukaemic cells. [64,65] Hence, reactivation of cell-cycle activity might be a critical step to re-establish chemotherapy sensitivity, as previously demonstrated in xenograft AML models. [66,67] In addition, the GMP subtype was characterised by enhanced innate immune activity, especially through Toll-like receptor (TLR) signaling, which was impaired in the MEP subtype. Enhanced expression of TLRs has been associated with haematopoietic malignancies, [68–71] including AML, [68,72,73] but their role in pathogenesis remains unclear. Nevertheless, enhanced TLR signaling could activate inflammatory cytokine secretion and downstream effectors, which might explain the observed upregulation of IL-6 JAK-STAT3 signaling and coregulation of TNF-α signaling through NF-κB and IFN-γ signaling in the GMP subtype. Accordingly, inhibition of TLRs or downstream effectors may confer therapeutic benefit in the GMP subtype but not necessarily in the MEP or HSC subtype, as described previously. [68,74]
Statistical models used restricted multinomial regression to estimate the MAGS assignment probability for each sample. MAGS subtypes were defined a priori based on FACS and were independent of the GEP used to build the classifier and subsequent MAGS assignment in clinical samples. Furthermore, samples with low assignment probabilities were labelled UC. The frequency of UC samples in other gene expression-based COO classifications is approximately 15%. [4,75] The probability cut-offs observed for MAGS assignment in the clinical cohorts, when allowing for the assignment of 15% of the samples as UC, exceeded 0.70, which is well above the random assignment probability of one out of four. Furthermore, the prognostic robustness of MAGS was successfully validated for a wide range of assignment frequency cut-offs for the UC subtype (S4A–S4C Fig). The prediction accuracy of the MAGS classification was rather low at 78.95%, but defining 15% of the samples with a low assignment probability as UC improved the prediction accuracy to 90.63%. Incorrect subtype prediction was restricted to the HSC subtype, especially to the GSE19429 cohort, for which FACS information was limited. The findings, thus, may be associated with differences in the FACS procedures and poorly defined progenitor populations. This is further supported by recent findings indicating that FACS surface markers are limited in their capacity to fully capture the differentiation stage of haematopoietic progenitor cells. [76] As the prediction accuracy of the MAGS classification is highly dependent on the number and quality of normal myeloid reference populations, it may be improved by increasing the sample size of the training cohort, avoiding interlaboratory batch effects, and optimising the isolation and characterisation of normal haematopoietic cell compartments for a priori subtype assignments.
Clinical considerations: Overall, patient survival was associated with the MAGS-assigned AML progenitor subtypes, independent of age, FAB subtype, and cytogenetic risk scores. These findings support the idea that initial hits in oncogenesis occur in the stem and progenitor cell compartments. MAGS subtype-specific mutation patterns of well-documented driver mutations also support the potential clinical impact of MAGS subtyping. Combination chemotherapy still forms the backbone of AML treatments; however, patients with relapsed or refractory diseases have an unmet need for predictive tests and precise companion diagnostics. This need may be fulfilled using MAGS subtyping with predictive information to guide targeted therapy. In agreement with previous work of our group, [4,8] the current analyses indicate that such information is available at diagnosis and could be used for the identification of candidates needing more precise strategies. We believe our results support the future inclusion of gene expression profiling in randomised prospective clinical trials aimed at improving AML treatment.
In summary, we have developed and documented a novel classification system that associates normal myeloid progenitor subsets with AML subtypes and prognosis. The MAGS subtypes have different clinical courses, drug resistance mechanisms, and molecular pathogenesis. However, further studies are needed to examine subtype-specific therapeutic strategies. Interestingly, the results imply a consensus between the genetic and transcriptional COOs, suggesting a minor impact of cell plasticity in leukaemic end stage cells. Future prospective studies will be needed to prove this concept using clinical endpoints.
Supporting information
S1 Table. Summary of cohort properties for sorted normal myeloid data sets and clinical AML data sets.
https://doi.org/10.1371/journal.pone.0229593.s002
(DOCX)
S2 Table. FACS antibody panels used to sort early (HSC) and late (GMP, MEP) myeloid progenitor cell subsets.
https://doi.org/10.1371/journal.pone.0229593.s003
(DOCX)
S3 Table. Molecular metadata information available for the GSE6891 and TCGA cohorts.
https://doi.org/10.1371/journal.pone.0229593.s004
(DOCX)
S4 Table. List of predictive genes defining the myeloid progenitor cell subset-associated gene signatures.
https://doi.org/10.1371/journal.pone.0229593.s005
(DOCX)
S5 Table. Top differentially expressed genes.
https://doi.org/10.1371/journal.pone.0229593.s006
(DOCX)
S6 Table. Top twenty Gene Ontology (GO) terms enriched for differentially expressed genes.
https://doi.org/10.1371/journal.pone.0229593.s007
(DOCX)
S7 Table. Top-enriched gene sets identified through GSEA analysis.
https://doi.org/10.1371/journal.pone.0229593.s008
(DOCX)
S8 Table. MAGS subtype-specific mutation patterns.
https://doi.org/10.1371/journal.pone.0229593.s009
(DOCX)
S9 Table. Association between the MAGS and FAB subtype.
https://doi.org/10.1371/journal.pone.0229593.s010
(DOCX)
S1 Fig. Identification of regularisation parameters through cross-validation.
https://doi.org/10.1371/journal.pone.0229593.s011
(DOCX)
S2 Fig. Venn diagram of MAGS subtype-specific differentially expressed genes.
https://doi.org/10.1371/journal.pone.0229593.s012
(DOCX)
S4 Fig. Prognostic validation of the assigned MAGS subtypes for the clinical meta-cohort.
https://doi.org/10.1371/journal.pone.0229593.s014
(DOCX)
Acknowledgments
The authors thank Roel Verhaak from the Department of Bioinformatics and Computational Biology, the University of Texas MD Anderson Cancer Center Houston, Texas, USA and Peter Valk from the Department of Haematology, Erasmus University Medical Center, Rotterdam, Netherlands for kindly providing survival data for the GSE6891 cohort.
References
- 1. Greaves M, Maley CC. Clonal evolution in cancer. Nature. 2012. pmid:22258609
- 2. Shlush LI, Zandi S, Mitchell A, Chen WC, Brandwein JM, Gupta V, et al. Identification of pre-leukaemic haematopoietic stem cells in acute leukaemia. Nature. 2014;506: 328–333. pmid:24522528
- 3. Johnsen HE, Bergkvist KS, Schmitz A, Kjeldsen MK, Hansen SM, Gaihede M, et al. Cell of origin associated classification of B-cell malignancies by gene signatures of the normal B-cell hierarchy. Leukemia and Lymphoma. 2014. pmid:23998255
- 4. Dybkær K, Bøgsted M, Falgreen S, Bødker JS, Kjeldsen MK, Schmitz A, et al. Diffuse large B-cell lymphoma classification system that associates normal B-cell subset phenotypes with prognosis. J Clin Oncol. 2015;33: 1379–88. pmid:25800755
- 5. Rasmussen T, Honoré L, Johnsen HE. Identification and characterisation of malignant cells using RT-PCR on single flow-sorted cells. Med Oncol. 1998.
- 6. Rasmussen T, Jensen L, Honoré L, Andersen H, Johnsen HE. Circulating clonal cells in multiple myeloma do not express CD34 mRNA, as measured by single-cell and real-time RT-PCR assays. Br J Haematol. 1999;107: 818–24. Available: http://www.ncbi.nlm.nih.gov/pubmed/10606890
- 7. Bergkvist KS, Nyegaard M, Bøgsted M, Schmitz A, Bødker JS, Rasmussen SM, et al. Validation and implementation of a method for microarray gene expression profiling of minor B-cell subpopulations in man. BMC Immunol. 2014;15: 1–12. pmid:24423010
- 8. Bødker JS, Brøndum RF, Schmitz A, Schönherz AA, Jespersen DS, Sønderkær M, et al. A multiple myeloma classification system that associates normal B-cell subset phenotypes with prognosis. Blood Adv. 2018;2: 2400–2411. pmid:30254104
- 9. Nørgaard CH, Jakobsen LH, Gentles AJ, Dybkær K, El-Galaly TC, Bødker JS, et al. Subtype assignment of CLL based on B-cell subset associated gene signatures from normal bone marrow–A proof of concept study. Calin G, editor. PLoS One. 2018;13: e0193249. pmid:29513759
- 10. Bennett JM, Catovsky D, Daniel MT, Flandrin G, Galton DA, Gralnick HR, et al. Criteria for the diagnosis of acute leukemia of megakaryocyte lineage (M7). A report of the French-American-British Cooperative Group. Ann Intern Med. 1985;103: 460–462. pmid:2411180
- 11. Bennett JM, Catovsky D, Daniel MT, Flandrin G, Galton DA, Gralnick HR, et al. Proposed revised criteria for the classification of acute myeloid leukemia. A report of the French-American-British Cooperative Group. Ann Intern Med. 1985;103: 620–625. pmid:3862359
- 12. Bennett JM, Catovsky D, Daniel M ‐ T, Flandrin G, Galton DAG, Gralnick HR, et al. Proposal for the recognition of minimally differentiated acute myeloid leukaemia (AML‐MO). Br J Haematol. 1991;78: 325–329. pmid:1651754
- 13. Bennett JM, Catovsky D, Daniel M ‐ T, Flandrin G, Galton DAG, Gralnick HR, et al. Proposals for the Classification of the Acute Leukaemias French‐American‐British (FAB) Co‐operative Group. Br J Haematol. 1976;33: 451–458. pmid:188440
- 14.
Jaffe ES, Harris NL, Stein H, Vardiman JW. Pathology and Genetics of Tumours of Haematopoietic and Lymphoid Tissues. World Health Organization Calssification of Tumours. Lyon: IARC Press; 2001.
- 15.
Swerdlow SH, Campo E, Harris NL, Jaffe ES, Pileri SA, Stein H, et al. WHO Classification of Tumours of Haematopoietic and Lymphoid Tissues. World Health Organization Classification of Tumours. Lyon, France: IARC; 2008.
- 16. Vardiman JW, Thiele J, Arber DA, Brunning RD, Borowitz MJ, Porwit A, et al. The 2008 revision of the World Health Organization (WHO) classification of myeloid neoplasms and acute leukemia: Rationale and important changes. Blood. 2009. pp. 937–951. pmid:19357394
- 17. Corces-Zimmerman MR, Hong W-J, Weissman IL, Medeiros BC, Majeti R. Preleukemic mutations in human acute myeloid leukemia affect epigenetic regulators and persist in remission. Proc Natl Acad Sci U S A. 2014;111: 2548–53. pmid:24550281
- 18. Jung N, Dai B, Gentles AJ, Majeti R, Feinberg AP. An LSC epigenetic signature is largely mutation independent and implicates the HOXA cluster in AML pathogenesis. Nat Commun. 2015. pmid:26444494
- 19. Rapin N, Bagger FO, Jendholm J, Mora-Jensen H, Krogh A, Kohlmann A, et al. Comparing cancer vs normal gene expression profiles identifies new disease entities and common transcriptional programs in AML patients. Blood. 2014. pmid:24363398
- 20. Andersson A, Edén P, Olofsson T, Fioretos T. Gene expression signatures in childhood acute leukemias are largely unique and distinct from those of normal tissues and other malignancies. BMC Med Genomics. 2010;3: 1–8. pmid:20092628
- 21. Majeti R, Becker MW, Tian Q, Lee T-LM, Yan X, Liu R, et al. Dysregulated gene expression networks in human acute myelogenous leukemia stem cells. Proc Natl Acad Sci U S A. 2009. pmid:19218430
- 22. Pellagatti A, Cazzola M, Giagounidis A, Perry J, Malcovati L, Della Porta MG, et al. Deregulated gene expression pathways in myelodysplastic syndrome hematopoietic stem cells. Leukemia. 2010;24: 756–764. pmid:20220779
- 23. De Jonge HJM, Valk PJM, Veeger NJGM, Ter Elst A, Den Boer ML, Cloos J, et al. High VEGFC expression is associated with unique gene expression profiles and predicts adverse prognosis in pediatric and adult acute myeloid leukemia. Blood. 2010;116: 1747–1754. pmid:20522712
- 24. Verhaak RGW, Wouters BJ, Erpelinck CAJ, Abbas S, Beverloo HB, Lugthart S, et al. Prediction of molecular subtypes in acute myeloid leukemia based on gene expression profiling. Haematologica. 2009;94: 131–134. pmid:18838472
- 25. Ley TJ, Miller C, Ding L, Raphael BJ, Mungall AJ, Robertson G, et al. Genomic and epigenomic landscapes of adult de novo acute myeloid leukemia. N Engl J Med. 2013;368: 2059–2074. pmid:23634996
- 26.
The R Development Core Team. R: A Language and Environment for Statistical Computing. Vienna, Austria, Austria: R Foundation for Statistical Computing, Vienna, Austria; 2016.
- 27. Brazma A. Minimum information about a microarray experiment (MIAME)—Successes, failures, challenges. TheScientificWorldJournal. 2009. pp. 420–423. pmid:19484163
- 28. McShane LM, Altman DG, Sauerbrei W, Taube SE, Gion M, Clark GM. Reporting recommendations for tumor marker prognostic studies. J Clin Oncol. 2005;23: 9067–72. pmid:16172462
- 29. McShane LM, Cavenagh MM, Lively TG, Eberhard DA, Bigbee WL, Williams PM, et al. Criteria for the use of omics-based predictors in clinical trials. Nature. 2013;502: 317–20. pmid:24132288
- 30. Irizarry RA, Hobbs B, Collin F, Beazer-Barclay YD, Antonellis KJ, Scherf U, et al. Exploration, normalization, and summaries of high density oligonucleotide array probe level data. Biostatistics. 2003;4: 249–64. pmid:12925520
- 31. Friedman J, Hastie T, Tibshirani R. Regularization Paths for Generalized Linear Models via Coordinate Descent. J Stat Softw. 2010;33. pmid:20808728
- 32. Johnson WE, Li C, Rabinovic A. Adjusting batch effects in microarray expression data using empirical Bayes methods. Biostatistics. 2007. pmid:16632515
- 33. Leek JT, Johnson WE, Parker HS, Jaffe AE, Storey JD. The SVA package for removing batch effects and other unwanted variation in high-throughput experiments. Bioinformatics. 2012;28: 882–883. pmid:22257669
- 34. Mroźek K, Marcucci G, Nicolet D, Maharry KS, Becker H, Whitman SP, et al. Prognostic significance of the European LeukemiaNet standardized system for reporting cytogenetic and molecular alterations in adults with acute myeloid leukemia. J Clin Oncol. 2012;30: 4515–4523. pmid:22987078
- 35. Patel JP, Gönen M, Figueroa ME, Fernandez H, Sun Z, Racevskis J, et al. Prognostic relevance of integrated genetic profiling in acute myeloid leukemia. N Engl J Med. 2012;366: 1079–1089. pmid:22417203
- 36. Phipson B, Lee S, Majewski IJ, Alexander WS, Smyth GK. Robust hyperparameter estimation protects against hypervariable genes and improves power to detect differential expression. Ann Appl Stat. 2016. pmid:28367255
- 37. Ritchie ME, Phipson B, Wu D, Hu Y, Law CW, Shi W, et al. Limma powers differential expression analyses for RNA-sequencing and microarray studies. Nucleic Acids Res. 2015. pmid:25605792
- 38. Benjamini Y, Hochberg Y. Controlling the False Discovery Rate: A Practical and Powerful Approach to Multiple Testing. J R Stat Soc Ser B. 1995;57: 289–300.
- 39. Subramanian A, Tamayo P, Mootha VK, Mukherjee S, Ebert BL, Gillette MA, et al. Gene set enrichment analysis: A knowledge-based approach for interpreting genome-wide expression profiles. Proc Natl Acad Sci U S A. 2005;102: 15545–15550. pmid:16199517
- 40. Liberzon A, Subramanian A, Pinchback R, Thorvaldsdóttir H, Tamayo P, Mesirov JP. Molecular signatures database (MSigDB) 3.0. Bioinformatics. 2011;27: 1739–1740. pmid:21546393
- 41. Liberzon A, Birger C, Thorvaldsdóttir H, Ghandi M, Mesirov JP, Tamayo P. The Molecular Signatures Database Hallmark Gene Set Collection. Cell Syst. 2015. pmid:26771021
- 42. Xie X, Lu J, Kulbokas EJ, Golub TR, Mootha V, Lindblad-Toh K, et al. Systematic discovery of regulatory motifs in human promoters and 3′ UTRs by comparison of several mammals. Nature. 2005;434: 338–345. pmid:15735639
- 43. Papaemmanuil E, Gerstung M, Bullinger L, Gaidzik VI, Paschka P, Roberts ND, et al. Genomic Classification and Prognosis in Acute Myeloid Leukemia. N Engl J Med. 2016;374: 2209–2221. pmid:27276561
- 44. Bullinger L, Döhner K, Bair E, Fröhling S, Schlenk RF, Tibshirani R, et al. Use of Gene-Expression Profiling to Identify Prognostic Subclasses in Adult Acute Myeloid Leukemia. N Engl J Med. 2004. pmid:15084693
- 45. Ebert BL, Galili N, Tamayo P, Bosco J, Mak R, Pretz J, et al. An erythroid differentiation signature predicts response to lenalidomide in myelodysplastic syndrome. PLoS Med. 2008. pmid:18271621
- 46. Raponi M, Lancet JE, Fan H, Dossey L, Lee G, Gojo I, et al. A 2-gene classifier for predicting response to the farnesyltransferase inhibitor tipifarnib in acute myeloid leukemia. Blood. 2008;111: 2589–2596. pmid:18160667
- 47. Theilgaard-Mönch K, Boultwood J, Ferrari S, Giannopoulos K, Hernandez-Rivas JM, Kohlmann A, et al. Gene expression profiling in MDS and AML: Potential and future avenues. Leukemia. 2011. pmid:21445077
- 48. Valk PJM, Verhaak RGW, Beijen MA, Erpelinck CAJ, van Doorn-Khosrovani SB van W, Boer JM, et al. Prognostically Useful Gene-Expression Profiles in Acute Myeloid Leukemia. N Engl J Med. 2004;350: 1617–1628. pmid:15084694
- 49. Kohlmann A, Bullinger L, Thiede C, Schaich M, Schnittger S, Döhner K, et al. Gene expression profiling in AML with normal karyotype can predict mutations for molecular markers and allows novel insights into perturbed biological pathways. Leukemia. 2010;24: 1216–1220. pmid:20428199
- 50. Eppert K, Takenaka K, Lechman ER, Waldron L, Nilsson B, Van Galen P, et al. Stem cell gene expression programs influence clinical outcome in human leukemia. Nat Med. 2011. pmid:21873988
- 51. Gentles AJ, Plevritis SK, Majeti R, Alizadeh AA. Association of a leukemic stem cell gene expression signature with clinical outcomes in acute myeloid leukemia. JAMA—J Am Med Assoc. 2010;304: 2706–2715. pmid:21177505
- 52. Shih AH, Abdel-Wahab O, Patel JP, Levine RL. The role of mutations in epigenetic regulators in myeloid malignancies. Nat Rev Cancer 2012 129. 2012;12: 599–612. pmid:22898539
- 53. Bacher U, Schnittger S, Haferlach T. Molecular genetics in acute myeloid leukemia. Curr Opin Oncol. 2010;22: 646–655. pmid:20805748
- 54. Döhner H, Estey E, Grimwade D, Amadori S, Appelbaum FR, Büchner T, et al. Diagnosis and management of AML in adults: 2017 ELN recommendations from an international expert panel. Blood. 2017;129: 424–447. pmid:27895058
- 55. Gaidzik VI, Teleanu V, Papaemmanuil E, Weber D, Paschka P, Hahn J, et al. RUNX1 mutations in acute myeloid leukemia are associated with distinct clinico-pathologic and genetic features. Leukemia. 2016;30: 2160–2168. pmid:27137476
- 56. Metzeler KH, Herold T, Rothenberg-Thurley M, Amler S, Sauerland MC, Görlich D, et al. Spectrum and prognostic relevance of driver gene mutations in acute myeloid leukemia. Blood. 2016. pp. 686–698. pmid:27288520
- 57. Sood R, Kamikubo Y, Liu P. Role of RUNX1 in hematological malignancies. Blood. 2017;129: 2070–2082. pmid:28179279
- 58. Shlush LI, Chapal-Ilani N, Adar R, Pery N, Maruvka Y, Spiro A, et al. Cell lineage analysis of acute leukemia relapse uncovers the role of replication-rate heterogeneity and microsatellite instability. Blood. 2012;120: 603–612. pmid:22645183
- 59. Hasemann MS, Lauridsen FKB, Waage J, Jakobsen JS, Frank AK, Schuster MB, et al. C/EBPα Is Required for Long-Term Self-Renewal and Lineage Priming of Hematopoietic Stem Cells and for the Maintenance of Epigenetic Configurations in Multipotent Progenitors. PLoS Genet. 2014;10. pmid:24415956
- 60. Mancini E, Sanjuan-Pla A, Luciani L, Moore S, Grover A, Zay A, et al. FOG-1 and GATA-1 act sequentially to specify definitive megakaryocytic and erythroid progenitors. EMBO J. 2012;31: 351–365. pmid:22068055
- 61. Pabst T, Mueller BU, Harakawa N, Schoch C, Haferlach T, Behre G, et al. AML1-ETO downregulates the granulocytic differentiation factor C/EBPα in t(8;21) myeloid leukemia. Nat Med. 2001;7: 444–451. pmid:11283671
- 62. Zhang P, Iwasaki-Arai J, Iwasaki H, Fenyus ML, Dayaram T, Owens BM, et al. Enhancement of hematopoietic stem cell repopulating capacity and self-renewal in the absence of the transcription factor C/EBPα. Immunity. 2004. pmid:15589173
- 63. Pastore F, Kling D, Hoster E, Dufour A, Konstandin NP, Schneider S, et al. Long-term follow-up of cytogenetically normal CEBPA-mutated AML. J Hematol Oncol. 2014;7. pmid:25214041
- 64. Dick JE. Stem cell concepts renew cancer research. Blood. 2008;112: 4793–4807. pmid:19064739
- 65. Drolle H, Wagner M, Vasold J, Kütt A, Deniffel C, Sotlar K, et al. Hypoxia regulates proliferation of acute myeloid leukemia and sensitivity against chemotherapy. Leuk Res. 2015;39: 779–785. pmid:25982178
- 66. Essers MAG, Trumpp A. Targeting leukemic stem cells by breaking their dormancy. Mol Oncol. 2010;4: 443–50. pmid:20599449
- 67. Saito Y, Uchida N, Tanaka S, Suzuki N, Tomizawa-Murasawa M, Sone A, et al. Induction of cell cycle entry eliminates human leukemia stem cells in a mouse model of AML. Nat Biotechnol. 2010;28: 275–280. pmid:20160717
- 68. Monlish DA, Bhatt ST, Schuettpelz LG. The Role of Toll-Like Receptors in Hematopoietic Malignancies. Front Immunol. 2016;7: 390. pmid:27733853
- 69. Muzio M, Scielzo C, Bertilaccio MTS, Frenquelli M, Ghia P, Caligaris-Cappio F. Expression and function of toll like receptors in chronic lymphocytic leukaemia cells. Br J Haematol. 2009;144: 507–516. pmid:19036098
- 70. Rolf N, Kariminia A, Ivison S, Reid GS, Schultz KR. Heterodimer-specific TLR2 stimulation results in divergent functional outcomes in B-cell precursor acute lymphoblastic leukemia. Eur J Immunol. 2015;45: 1980–1990. pmid:25867213
- 71. Wei Y, Dimicoli S, Bueso-Ramos C, Chen R, Yang H, Neuberg D, et al. Toll-like receptor alterations in myelodysplastic syndrome. Leukemia. 2013;27: 1832–1840. pmid:23765228
- 72. Okamoto M, Hirai H, Taniguchi K, Shimura K, Inaba T, Shimazaki C, et al. Toll-like Receptors (TLRs) are expressed by myeloid leukaemia cell lines, but fail to trigger differentiation in response to the respective TLR ligands: Correspondence. British Journal of Haematology. 2009. pp. 585–587. pmid:19673881
- 73. Rybka J, Butrym A, Wróbel T, Jaźwiec B, Stefanko E, Dobrzyńska O, et al. The expression of Toll-like receptors in patients with acute myeloid leukemia treated with induction chemotherapy. Leuk Res. 2015;39: 318–322. pmid:25624047
- 74. Ignatz-Hoover JJ, Wang H, Moreton SA, Chakrabarti A, Agarwal MK, Sun K, et al. The role of TLR8 signaling in acute myeloid leukemia differentiation. Leukemia. 2015;29: 918–926. pmid:25283842
- 75. Lenz G, Wright G, Dave SS, Xiao W, Powell J, Zhao H, et al. Stromal gene signatures in large-B-cell lymphomas. N Engl J Med. 2008;359: 2313–23. pmid:19038878
- 76. Paul F, Arkin Y, Giladi A, Jaitin DA, Kenigsberg E, Keren-Shaul H, et al. Transcriptional Heterogeneity and Lineage Commitment in Myeloid Progenitors. Cell. 2015;163: 1663–77. pmid:26627738