Abstract

   In 2019 it is estimated that more than 21,000 new acute myeloid
   leukemia (AML) patients will be diagnosed in the United States, and
   nearly 11,000 are expected to die from the disease. AML is primarily
   diagnosed among the elderly (median 68 years old at diagnosis).
   Prognoses have significantly improved for younger patients, but as much
   as 70% of patients over 60 years old will die within a year of
   diagnosis. In this study, we conducted a reanalysis of 2,213 acute
   myeloid leukemia patients compared to 548 healthy individuals, using
   curated publicly available microarray gene expression data. We carried
   out an analysis of normalized batch corrected data, using a linear
   model that included considerations for disease, age, sex, and tissue.
   We identified 974 differentially expressed probe sets and 4 significant
   pathways associated with AML. Additionally, we identified 375 age- and
   70 sex-related probe set expression signatures relevant to AML.
   Finally, we trained a k nearest neighbors model to classify AML and
   healthy subjects with 90.9% accuracy. Our findings provide a new
   reanalysis of public datasets, that enabled the identification of new
   gene sets relevant to AML that can potentially be used in future
   experiments and possible stratified disease diagnostics.

   Subject terms: Acute myeloid leukaemia, Gene expression,
   Transcriptomics

Introduction

   Acute myeloid leukemia (AML) is a heterogeneous malignant disease of
   the hematopoietic system myeloid cell lineage^[26]1–[27]5. AML is best
   characterized by terminal differentiation in normal blood cells and
   excessive production and release of cells at various stages of
   incomplete maturation (leukemia cells). As a result of this faster than
   normal, and uncontrolled growth of leukemia cells, healthy myeloid
   precursors involved in hematopoiesis are suppressed, and ultimately can
   soar to death within months from diagnosis if untreated^[28]1,[29]6.
   AML accounts for 70% of myeloid leukemia and nearly 80% of acute
   leukemia cases, making it the most common form of both myeloid and
   acute leukemia^[30]1,[31]7. The number of new AML cases is increasing
   each year – in 2019 alone, an estimated 21,450 new AML patients will be
   diagnosed, and nearly 10,920 are expected to die from the
   disease^[32]8.

   According to the 2016 World Health Organization (WHO) newly revised
   myeloid neoplasms and acute leukemia classification system^[33]9, AML
   prognosis criteria for classification are highly dependent on the
   presence of chromosomal abnormalities, including chromosomal deletions,
   duplications, translocations, inversions, and gene fusions. AML is
   diagnosed predominantly through microscopic, cytogenetic, and molecular
   genetic analyses of patients’ blood, and/or bone marrow samples.
   Microscopic examination may be used to detect distinctive features
   (e.g. Auer rods) in cell morphology, cytogenetic analysis to identify
   chromosomal structural aberrations (e.g., t(8;21), inv(16), t(16;16),
   or t(9;11)), and molecular genetic analysis to identify gene fusion
   (e.g., RUNX1-RUNX1T1 and CBFB-MYH11), and mutations in genes frequently
   mutated in AML (e.g., NPM1, CEBPA, RUNX1,
   FLT3)^[34]1,[35]3,[36]5,[37]10–[38]12. Such cytogenetic and molecular
   genetic analyses are used to identify prognosis markers for classifying
   AML patients into three risk categories: favorable, intermediate, and
   unfavorable, currently based primarily on the European LeukemiaNet
   (ELN) 2017 classification^[39]3,[40]10 (see Estey^[41]3 for a recent
   review, including ELN assessments). A large group of AML patients
   present normal karyotypes and lack chromosomal
   abnormalities^[42]3,[43]5,[44]10,[45]11,[46]13. These patients are
   classified as intermediate risk, and often have heterogeneous clinical
   outcome with standard therapy with risk of AML
   relapse^[47]3,[48]5,[49]14.

   Additionally, AML prognosis worsens with age, and older patients
   respond less to current treatments, with poorer clinical outcomes
   compared to younger patients^[50]15,[51]16. AML can occur in people of
   all ages but is primarily diagnosed among the elderly (>60 years old),
   with a median age of 68 years at diagnosis^[52]8. Recent advances in
   AML biology have expanded our understanding of its complex genetic
   landscape, and led to significant improvement in prognoses and
   therapeutic strategy for younger patients^[53]2,[54]16. For elderly
   patients, prognoses remain grim and the main therapeutic strategy,
   remission induction therapy followed by an intensive consolidation
   phase (post-remission), had remained nearly unchanged over the past
   three decades^[55]1,[56]2,[57]4,[58]5,[59]10,[60]16,[61]17. More
   recently, however, new therapeutic agents have been approved for older
   AML patients^[62]4,[63]5, and these include venetoclax (combined with
   decitabine or azacitidine)^[64]18, midostaurin (combined with standard
   chemotherapy)^[65]19, and gilteritinib^[66]20. It is expected that the
   new therapeutic agents will improve prognosis for older AML patients
   (where in the past up to 70% of AML patients aged 65 or older were
   reported to die within a year following diagnosis^[67]21). While it is
   apparent that the nature of AML changes with age, still little is known
   about the extent of these associations and how they vary with patient
   age^[68]2,[69]22,[70]23, and current indications from ELN and the
   National Comprehensive Cancer Network (NCCN) essentially consider age
   as a surrogate variable that is used only in conjunction with other
   treatment‐related mortality factors^[71]3,[72]5,[73]10. Taking into
   consideration age in the identification of changes in AML global gene
   expression may lead to improved early diagnosis and improvement in
   treatment approaches for elderly patients. To further complicate
   matters, AML has multiple driver mutations and competing clones that
   evolve over time, making it a very dynamic disease^[74]13,[75]24.

   Multiple gene expression analyses of AML have been carried out, 25 of
   which have been systematically compared by Miller and
   Stamatoyannopoulos^[76]25, who analyzed information on 4,918 genes, and
   identified 25 genes reported across multiple studies, with potential
   prognostic features. In this study, we performed a comprehensive gene
   expression analysis of 2,213 AML patients and 548 healthy subjects, by
   re-analyzing publicly available gene expression microarray data from 37
   curated studies (a reanalysis following strict inclusion criteria) and
   identified disease-, age- and sex-related gene expression changes
   associated with AML. The differentially expressed gene sets were
   associated to signaling pathways relevant in AML, and also used to
   train and test a predictive model of AML or healthy status. We believe
   that our results may lead to improved AML early detection, and
   diagnostic testing with target genes, which collectively can
   potentially serve as age- and sex-dependent biomarkers for AML
   prognosis, as well as new treatment targets with mechanisms of action
   different from those used in conventional chemotherapy.

Results

Data curation and gene expression pre-processing

   We searched the Gene Expression Omnibus (GEO) public repository, based
   on our systematic workflow and inclusion criteria, Fig. [77]1a,b.
   Overall, 2,132 datasets were screened, and 643 selected (577 were
   excluded as non-Affymetrix, various platform arrays). From the 66
   remaining corresponding studies, 34 were excluded due to: lack of
   metadata, using non-peripheral blood or non-bone marrow tissues, or
   being cell line or cell-type specific, or analyzing treated subjects.
   After this curation we obtained 34 age-annotated gene expression
   datasets from 32 different studies covering 2,213 AML patients and 548
   healthy individuals. These 34 datasets were reanalyzed, starting from
   raw microarray data, to perform a gene expression analysis of variance
   and functional pathway enrichment analysis (see online Methods).
   Table [78]1 provides a description of each dataset with a sub-table
   summary of all curated data used in this study. After pre-processing
   each individual dataset separately, Fig. [79]1b, we performed the
   statistical analysis on 44,754 probe sets which were common across all
   samples (Affymetrix expression microarray data).

Figure 1.

   [80]Figure 1
   [81]Open in a new tab

   General approach, data curation, and analysis workflow summary. The
   flowchart shows in (a) the five main steps that summarize our method of
   approach for our study, and in (b) the curation and screening criteria
   for raw gene expression and annotation data files curation, data
   pre-processing, supervised machine learning for missing metadata
   prediction, and batch effects correction. (c) The analysis included a
   linear model analysis of variance (ANOVA) coupled with Tukey’s Honestly
   Significant Difference (HSD) post-hoc tests, and KEGG pathway and GO
   enrichment. Finally, we performed a machine learning classification of
   AML based on our findings.

Table 1.

   Summary table gene expression datasets used in this study.
   Author, Year GEO accession Disease Status* Affymetrix platform id:
   Number of samples used & Sample source* Refs*
   (A) Curated datasets used in linear model analysis (34 datasets from 32
   studies)
   Zatkova et al., 2009 [82]GSE10258 AML [83]GPL570: 8 BM ^[84]68
   Tomasson et al., 2008 [85]GSE10358 AML [86]GPL570: 300 BM ^[87]69
   Metzeler et al., 2008 [88]GSE12417 AML

   [89]GPL570: 73 BM & 5 PB

   [90]GPL96/97: 160 BM & 2PB
   ^[91]55
   Wouters et al., 2009, Taskesen et al., 2011 [92]GSE14468 AML
   [93]GPL570: 482 BM & 43 PB ^[94]70, [95]71
   Figueroa et al., 2009 [96]GSE14479 AML [97]GPL570: 16 BM ^[98]72
   Klein et al., 2009 [99]GSE15434 AML [100]GPL570: 231 BM & 20 PB
   ^[101]73
   Lück et al., 2011 [102]GSE29883 AML [103]GPL570: 10 BM & 2 PB ^[104]74

   Li et al., 2013,

   Herold et al., 2014,

   Janke et al., 2014,

   Jiang et al., 2016
   [105]GSE37642 AML

   [106]GPL570: 140 BM

   [107]GPL96/97: 422 BM
   ^[108]56– [109]59
   Bullinger et al., 2014 [110]GSE39363 AML [111]GPL570: 11 BM & 2 PB NYP
   Opel et al., 2015 [112]GSE46819 AML [113]GPL570: 8 BM & 4 PB ^[114]75
   TCGA et al., 2015 [115]GSE68833 AML [116]GPL570: 183 BM NYP
   Cao et al., 2016 [117]GSE69565 AML [118]GPL570: 12 PB ^[119]76
   Bohl et al., 2016 [120]GSE84334 AML [121]GPL570: 25 BM & 20 PB NYP
   Li et al., 2011 [122]GSE23025 AML [123]GPL570: 21 BM & 13 PB ^[124]77
   Warren et al., 2009 [125]GSE11375 Healthy [126]GPL570: 26 PB ^[127]78
   Green et al., 2009 [128]GSE14845 Healthy [129]GPL570: 1 PB NYP
   Wu et al., 2012 [130]GSE15932 Healthy [131]GPL570: 8 PB NYP
   Karlovich et al., 2009 [132]GSE16028 Healthy [133]GPL570: 22 PB
   ^[134]79
   Krug et al., 2011 [135]GSE17114 Healthy [136]GPL570: 14 PB NYP
   Kong et al., 2012 [137]GSE18123 Healthy [138]GPL570: 17 PB ^[139]80
   Sharma et al., 2009 [140]GSE18781 Healthy [141]GPL570: 25 PB ^[142]81
   Rosell et al., 2011 [143]GSE25414 Healthy [144]GPL570: 12 PB ^[145]82
   Schmidt et al., 2006 [146]GSE2842 Healthy [147]GPL570: 2 PB ^[148]83
   Meng et al., 2015 [149]GSE71226 Healthy [150]GPL570: 3 PB NYP
   Tasaki et al., 2017 [151]GSE84844 Healthy [152]GPL570: 30 PB ^[153]84
   Leday et al., 2018 [154]GSE98793 Healthy [155]GPL570: 64 PB ^[156]85
   Shamir et al., 2017 [157]GSE99039 Healthy [158]GPL570: 121 PB ^[159]86
   Tasaki et al., 2018 [160]GSE93272 Healthy [161]GPL570: 35 PB ^[162]87
   Clelland et al., 2013 [163]GSE46449 Healthy [164]GPL570: 24 PB ^[165]88

   Lauwerys et al., 2013

   Ducreux et al., 2016
   [166]GSE39088 Healthy [167]GPL570: 46 PB ^[168]89, [169]90
   Xiao et al., 2011 [170]GSE36809 Healthy [171]GPL570: 35 PB ^[172]91
   Zhou et al., 2010 [173]GSE19743 Healthy [174]GPL570: 63 PB ^[175]92
   (B) Covariate datasets (used for batch correction and for testing
   predictive models)
   Jiang et al., 2018^# [176]GSE107968^* 2 AML; 1 Healthy [177]GPL570: 3
   BM NYP
   Greiner et al., 2015^# [178]GSE68172^* 20 AML; 5 Healthy [179]GPL570:
   25 PB ^[180]64
   Majeti et al., 2009^# [181]GSE17054^* 9 AML; 4 Healthy [182]GPL570: 13
   BM ^[183]65
   Bacher et al., 2012^# [184]GSE33223^* 20 AML; 10 Healthy [185]GPL570:
   30 PB ^[186]66
   Mills et al., 2009^# [187]GSE15061^* 404 AML; 138 Healthy [188]GPL570:
   542 BM ^[189]67
   (C) Analysis datasets summary statistics
   Disease state Sample source Affymetrix platform id Unique probe sets
   AML Healthy BM PB [190]GPL570 [191]GPL96/97 [192]GPL570 [193]GPL96/97
   2,213 548 2,090 671 2,177 584 54,675 44,760
   [194]Open in a new tab

   Summary of datasets used in our analysis and disease classification.
   *GEO, Gene Expression Omnibus; AML, acute myeloid leukemia; Refs.,
   references; NYP, not yet published; [195]GPL570, Affymetrix Human