Integrating genetic and epigenetic factors in chronic myeloid leukemia risk assessment: toward gene expression-based biomarkers.

Cancer treatment is constantly evolving from a one-size-fits-all towards bespoke approaches for each patient. In certain solid cancers, including breast and lung, tumor genome profiling has been incorporated into therapeutic decision-making. For chronic phase chronic myeloid leukemia (CML), while tyrosine kinase inhibitor therapy is the standard treatment, current clinical scoring systems cannot accurately predict the heterogeneous treatment outcomes observed in patients. Biomarkers capable of segregating patients according to outcome at diagnosis are needed to improve management, and facilitate enrollment in clinical trials seeking to prevent blast crisis transformation and improve the depth of molecular responses. To this end, gene expression (GE) profiling studies have evaluated whether GE signatures at diagnosis are clinically informative. Patient material from a variety of sources has been profiled using microarrays, RNA sequencing and, more recently, single-cell RNA sequencing. However, differences in the cell types profiled, the technologies used, and the inherent complexities associated with the interpretation of genomic data pose challenges in distilling GE datasets into biomarkers with clinical utility. The goal of this paper is to review previous studies evaluating GE profiling in CML, and explore their potential as risk assessment tools for individualized CML treatment. We also review the contribution that acquired mutations, including those seen in clonal hematopoiesis, make to GE profiles, and how a model integrating contributions of genetic and epigenetic factors in resistance to tyrosine kinase inhibitors and blast crisis transformation can define a route to GE-based biomarkers. Finally, we outline a four-stage approach for the development of GE-based biomarkers in CML.

Introduction

Chronic myeloid leukemia (CML) is a clonal disorder of the hematopoietic stem cell compartment defined and driven by the BCR-ABL1 gene rearrangement and the tyrosine kinase it encodes.[1] Clinically, it is accompanied by an expansion of mostly myelo-erythroid progenitors that maintain the ability to differentiate terminally into neutrophils. Prior to the introduction of ABL1 tyrosine kinase inhibitors (TKI), most patients would progress to a terminal blast crisis (BC) stage marked by the acquisition of additional genetic abnormalities within an average of 5-7 years.[2] In this stage, the clinico-pathological features were the inexorable accumulation of either myeloid or lymphoid progenitors that had acquired aberrant self-renewal properties, broad resistance to cytotoxic therapies, and eventual patient demise from bone marrow failure.[2] The arrival of TKI at the turn of the century resulted in remarkable responses, such that most individuals treated in chronic phase (CP) CML can expect to achieve near-normal life expectancies.[3] Nevertheless, CML-related deaths are still reported, mainly due to resistance and progression to BC, especially in the first few years of treatment.[4]

Current treatment aims and features of an ideal biomarker

Current therapeutic aims are directed at achieving sufficiently deep molecular responses that the risks of BC transformation are effectively negligible and, in the longer-term, increasing the rates of treatment-free remission. [5-7] Clinical guidelines toward achieving deep molecular responses have been reviewed elsewhere,[8] and at their core, prescribe the measurement of BCR-ABL1 transcript levels using the International Scale (IS) every 3 months as a readout of the depth of the response to TKI. In turn, the depth of TKI response serves as a critical biomarker guiding patient management and prognostication (Figure 1).

Figure 1.

Features of an ideal chronic myeloid leukemia biomarker. Curves indicate changes in BCR-ABL1 transcript levels, meausred using the International Scale (BCR-ABL1IS), following initation of tyrosine kinase inhibitor (TKI) therapy in patients with chronic phase (CP) chronic myeloid leukemia (CML). The corresponding molecular response (MR) value is provided next to the BCR-ABLIS value. Green, orange, and red curves are representative of patients in European LeukemiaNet 2020 ‘optimal’, ‘warning’, and ‘failure’ cateogories respectively. Major (MMR) and deep (DMR) molecular remissions are defined as 0.1% (MR 3) and 0.01% (MR4) BCRABLIS, respectively. Green, orange, and yellow bullet points indicate guidelines for each category of response. Gray boxes describe predictive capabilities of an ideal biomarker. HCT: hematopoietic stem cell transplant.

Given current treatment goals, an ideal biomarker would accurately predict patients who will achieve a deep molecular response with first-line TKI, or require a switch to alternative therapy, and, among those who achieve a deep molecular response, those who will be able to stop TKI successfully (Figure 1). The biomarker would be informative from the time of diagnosis and prior to TKI initiation, since this would enable early stratification of patients for therapy with a first-generation versus a second/third-generation TKI, allosteric BCRABL1 inhibitor, a clinical trial, or preparation for allogeneic transplantation. Additionally, among patients who meet the criteria for stopping TKI therapy, the ideal biomarker would identify additional therapies that would enhance treatment-free remissions. Finally, gene expression (GE)-based biomarkers should be clinically robust, and widely available among centers and regions in both low and high Human Development Index countries.

Why gene expression-based biomarkers?

Contributions from genetic and epigenetic mediators to TKI resistance and BC transformation are well documented, [10-14] and it is axiomatic that genetic or epigenetic factors mediating these outcomes will contribute to a cell’s GE signature. Accordingly, GE signatures offer a molecular profile that integrates risk factors encoded by both mutations and epigenetic states. However, faithfully extracting and interpreting GE-based information in clinical settings is challenging. Barriers to adoption include technical limitations, logistical factors, as well as differences in study design and data analysis, and are described below.

Despite the barriers, recent advances in technological and computational platforms are enabling the interrogation of patient samples on an unprecedented scale, and are being translated into robust technical assays on patient material that are reproducible in clinical laboratories. [15] Such advances may eventually result in the identification of pretreatment biomarkers that not only predict TKI resistance but suggest alternative non-BCR-ABL1-targeting therapies to pre-empt the emergence of clinical resistance. Accordingly, it is timely to review the results of GE studies using primary patient material annotated for clinical outcomes, and assess how genetic and epigenetic factors associated with treatment outcome contribute to GE signatures. In doing so, it is also important to develop models incorporating the interplay between genetic and epigenetic factors, and determine how best to use the resulting GE outputs to understand and predict CML drug resistance and transformation. Finally, it is incumbent on the CML community to outline the practical steps needed for the clinical development of GE-based biomarkers in CML.

Gene expression signatures associated with resistance to tyrosine kinase inhibitors

Since the beginning of the TKI era, a variety of diagnostic material from CP patients has been used to discover TKI-resistance GE signatures (Table 1). Here, we review the key conclusions from these studies.

Table 1.

Gene expression profiling studies comparing responders and non-responders to tyrosine kinase inhibitors.

Gene expression using peripheral blood

In the earliest research by Kaneta et al.[16] and McLean et al.[17] microarray studies were conducted on blood from imatinib responders and non-responders. Apart from CBLB, which was downregulated in responders, there was no overlap between the two datasets. De Lavallade et al. conducted microarray studies on peripheral blood mononuclear cells to identify a 105-gene set that was enriched in imatinib non-responders, comprising mainly genes in cell cycle and DNA repair pathways.[18] However, the GE signature could be validated only in an imatinibtreated cohort but not in a cohort treated with interferona. As a targeted approach, the expression of 21 genes associated with TKI responses and disease progression was studied by Zhang et al.[19] Increased PTGS1 expression was the only gene that differentiated primary imatinib-resistant patients from responders, while 15 genes distinguished CP from BC. Twelve genes distinguished imatinib- responsive from secondary imatinib-resistant CML without BCR-ABL1 mutations, of which LYN, JAK2, PTPN22 and CEBPA downregulation was shared with BC samples. The study concluded that at least some features of secondary imatinib resistance overlap with BC transformation.

More recently, Kok et al. conducted microarray-based analysis on diagnostic blood from 96 CP patients from the TIDEL-II trial to predict failure of early molecular response,[20] which correlates with inferior long-term outcomes. [21,22] Three hundred sixty-five differentially expressed genes were identified which were enriched for ‘cell cycle’ and ‘stemness’ (MYC, HOXA9, b-catenin) but depleted for ‘immune-response’ categories in the group with early molecular response failure. A binary classification model was built to predict early molecular response failure based on 17 genes and the signature was validated in an independent cohort. Of these, eight genes IGFBP2, SRSF11, BAX, CDKN1B, BNIP3L, FZD7, PRSS57, and RPS28 intersected with findings of previous CML TKIresistance and progression studies. This study demonstrated that GE information from diagnostic samples could predict events long in the future, including major molecular response (MMR) at 24 months, MR4.5 at 5 years, and BC transformation.

Gene expression using bone marrow

Independently, a series of studies used unselected bone marrow for comparisons of GE between groups of patients with different treatment responses. Frank et al. identified a 128 GE signature associated with imatinib resistance, specifically in an interferon-a pre-treated cohort. Differentially expressed genes were involved in apoptosis (CASP9, TRAP1), DNA repair (MSH3, DDB2), oxidative stress protection (GSS, PON2, VNN1) and centrosomes (ID1).[23] Villuendas et al.[24] identified 46 differentially expressed genes of which a six-gene prediction score (BIRC4, FZD7, IKBKB, IL-7R, TNC, VWF) that correlated with imatinib resistance after interferon-a failure developed. Differentially expressed genes were involved in cell adhesion (TNC and SCAM-1), drug metabolism (COX1 or PTGS1), protein tyrosine kinases (MKNK1), and phosphatases (BTK and PTPN22). Notably, the MKNK1/2 kinases have been shown by two independent groups to be involved in BC transformation.[25,26] In contrast to the prior studies, Crossman et al. found no differentially expressed genes between the imatinib responder categories. The use of mixed peripheral blood and bone marrow samples, unselected white blood cells and a heterogeneous cohort of patients in late CP and heavily pre-treated, were suggested as potential reasons for the negative results.[27] The important conclusion was that GE comparisons should be made on purified CD34+ cells. Indeed, in a meta-analysis comparing six published GE studies in CML, DDX11, MSH5, and RAB11FIP3 were the only genes coincident between any two of the studies.[28] The small differences in differential GE between responder groups, different GE platforms, different statistical methods and different sources of cells profiled were suggested reasons for the poor intersection. The disappointing results from unselected peripheral blood and bone marrow provided the impetus to isolate and study CD34+ fractions.

Gene expression using CD34+ cells

McWeeney et al. were the first group to use CD34+ cells from diagnostic bone marrow.[11] Cell adhesion genes were upregulated in imatinib-resistant patients suggesting that CD34+ cells may establish more adhesive interactions with the bone marrow milieu. The enrichment for bcatenin binding targets suggested activated Wnt/b-catenin signaling in imatinib-resistant patients, a feature shared with CD34+ progenitors from BC.[26,29] The authors concluded that primary resistance to imatinib might reflect more advanced disease progression. A 75-probe minimal gene classifier predicted 88% of responders and 83% of nonresponders in a validation cohort. Importantly, the authors of this paper compared their GE signatures to those predicting early BC transformation, as discussed below, and provided an important resource for validation and comparison of other CD34+-based GE datasets.

Single-cell-based gene expression analysis

Recent advances in single-cell analysis have enabled novel GE-based insights on the roles of tumor cell heterogeneity and clonal evolution under the selective pressure of therapeutics, with obvious applications in biomarker development.[30] Leukemia stem cell (LSC) heterogeneity was characterized by Warfvinge et al. by combining highthroughput immunophenotyping with single-cell GE profiling with a defined panel of genes.[31] LSC sub-fractions with more primitive and quiescent signatures had a higher persistence after TKI therapy with the most TKI-insensitive population identified as Lin–CD34+CD38low/– CD45RA–cKIT–CD26+ stem cells. Giustacchini et al. used the Smart-seq2 platform to combine single-cell RNAsequencing analysis with BCR-ABL1 transcript detection using purified stem cells. A sub-population of BCR-ABL1+ quiescent stem cells enriched for hematopoietic stem cell (HSC) signatures was found to persist during TKI therapy.[32] Intriguingly, the BCR-ABL1- cells in CML patients were enriched for inflammatory, tumor growth factor-b and tumor necrosis factor-a hallmarks and discriminated between the TKI-responder groups.

In addition to LSC-derived signatures, the GE signature of immune cells can be equally instructive. For example, plasmacytoid dendritic cells, the major producers of interferon- a in vivo, promoted resistance to nilotinib in CML patients.[33] These studies imply that the cytokines released by immune cells in the bone marrow microenvironment, and the transcriptomic changes that they bring about on the LSC, may activate cytokine-dependent TKI resistance programs.[34] Together, these single-cell studies demonstrate that GE signatures within malignant and nonmalignant compartments in CML are prognostically informative. We anticipate the discovery of additional biomarkers among discrete cell types which have the potential to be assayed by platforms available in standard pathology laboratories, e.g., by flow cytometry or immunohistochemistry.

Gene expression signatures associated with blast crisis progression

Transcriptomic comparisons between the CP and BC stages have uncovered progression-related signatures that can herald BC transformation (Table 2). In the pre-TKI era, the time to BC transformation from CP varied between patients, and to understand this difference, Yong et al. compared CD34+ cells from leukapheresis samples provided by patients who progressed to BC within 3 years (aggressive leukemia) or after more than 7 years (indolent leukemia) following diagnosis.[13] The study identified that lower CD7 with higher PR3 and ELA2 expression at diagnosis was associated with longer survival. Intriguingly, when the GE signatures identified by Yong et al. and McWeeney et al. were compared, a significant overlap was found.[11] This important study demonstrated that biological processes associated with TKI resistance and early BC transformation overlapped, and that CD34+ cells from different sources (bone marrow vs. peripheral blood) contained this information.

Table 2.

Gene expression profiling studies comparing chronic phase and acute phase.

In the post-TKI era, a landmark study by Radich et al. identified distinct transcriptional programs during BC progression. [25] About 3,000 genes were associated with the BC stage with a dysregulated WNT/b-catenin pathway, decreased Jun B and FOS, and higher PRAME expression. The Radich dataset was subsequently used to compute a six-gene signature comprising NOB1, DDX47, IGSF2, LTB4R, SCARB1, and SLC25A3 to predict progression.[35] Independently, Zheng et al. isolated CD34+ cells and identified 34 differentially expressed genes as cells transited from CP to BC. Among the misregulated genes, SOCS2 and CD52 were downregulated while HLA-related genes were overexpressed in BC.[36]

To understand the biological mechanisms underlying GE changes in TKI resistance and BC, a recent study tested the hypothesis that prognostically important genes were enriched for targets of the polycomb repressive complex (PRC; see below).[10] Importantly, target genes of PRC-associated silencing in BC progression were enriched for downregulated genes identified in the datasets of both McWeeney et al. and Yong et al.[10] The cross-validation of these three independent datasets suggests important lessons for the development of GE-based risk assessment: (i) the discovery of reproducible GE-based biomarkers is possible when homogeneous CD34+ populations are used; (ii) the processes of TKI resistance and BC transformation are biologically convergent despite genetic heterogeneity:[10] and (iii). PRC-regulated processes contribute to silencing of prognostically informative genes.

Contribution of somatic mutations to gene expression signatures

Recent reviews have described the range and frequency of specific genetic mutations in patients who developed TKI resistance and/or BC.[37] For many of these genes, there is strong preclinical information indicating that their associated mutations contribute to or are even sufficient to produce TKI resistance or transformation phenotypes (summarized in Table 3).[38-48] These studies imply that genetic mutations alter GE profiles, and here we review their contributions to GE changes in CML since these changes may represent useful GE-based biomarkers.

Table 3.

Functional effects of frequently mutated genes in blast crisis.

For RUNX1 mutations, the Mustjoki group identified an accompanying GE signature in BC samples.[49] They found that RUNX1 mutations were associated with the upregulation of stemness, B-cell markers, interferon and immune signaling and transcription factors regulating plasmacytoid dendritic cell development.

In analogous work, the overexpression of an IKZF1 dominant-negative mutant in CD34+ cells from CP patients increased STAT5 expression, a pathway associated with imatinib resistance,[50] and enhanced transformation.[51] RAG expression status was recently assessed in diagnostic samples, given the role of RAG recombination as a mediator of IKZF1 deletions.[52] Notably, RAG1/2 and DNTT upregulation at diagnosis suggested imminent lymphoid BC transformation within 12 months (8/8 patients), demonstrating that GE signatures can reliably predict transformation.

Despite limited functional interrogation of ASXL1 using CML patient material, insertion sites within ASXL1 promoted BC progression in a CP mouse model subjected to transposition-based mutagenesis.[53] Transgenic expression of truncated protein ASXL1aa1-587 in mice increased HSC self-renewal, and Brd4 occupancy and chromatin accessibility around genes required for stemness, and predisposed mice to myeloid malignancies.[54] However, the clinical relevance of diagnostic ASXL1 mutations is still unclear because some patients with ASXL1 variants at diagnosis can achieve a MMR after TKI therapy.[14] Furthermore, ASXL1 mutations frequently disappeared when monitored in the long-term during TKI therapy (personal observation by Dr. Dennis Kim). Meanwhile, direct evidence for contributions of other mutations to CML GE signatures is currently lacking, and we have to infer them from studies in other malignancies (Table 3).

Lessons from clonal hematopoiesis

Clonal hematopoiesis is the clinical phenomenon by which populations of hematopoietic cells expand and carry a somatic mutation that is at least 2% of the variant allele fraction.[55] The common genes comprise DNMT3A, TET2, and ASXL1, and others also found in CML individuals, including RUNX1, BCORL1, and TP53.[56] Individuals with clonal hematopoiesis are at increased risk of developing hematologic malignancies, and it is therefore pertinent to ask whether clonal hematopoiesis-related mutations also confer increased risk of TKI resistance or progression. A study by Kim et al. has highlighted important features of clonal hematopoiesis-related mutations in CML.[57] Firstly, they may occur in a non-Philadelphia chromosome-positive clone and predate the development of CML, and are unrelated to the CML clone. Secondly, even when a specific mutation occurs in the Philadelphia chromosome-positive clone, it only confers a relative risk of TKI resistance or progression. Indeed, patients with RUNX1 mutations have been documented to achieve MMR (personal observation, Dr. Dennis Kim). Nevertheless, Kim et al. concluded that mutations in genes regulating epigenetic function (TET2, ASXL1 among them) were associated with a higher risk of inferior TKI responses.

There are also strong preclinical data indicating that clonal hematopoiesis-related mutations result in subtle but important changes in GE in HSC. For example, Dnmt3a-deficient HSC show a loss of DNA methylation in regions enriched for self-renewal genes such as Meis1, Evi1 and HoxA9.[58] In Tet2-deficient mice, the loss of DNA demethylation is accompanied by an expansion of the stem and progenitor cell compartments, and eventual myeloproliferation.[45] In ASXL1-deficient mice, an increase in self-renewal capacity of stem cells is observed, through the loss of PRC1-mediated gene repression.[59] Another interesting aspect of hematopoietic stem and progenitor cells harboring inactivating mutations of DNMT3A and TET2 is that they both led to increased cytokine production in peripheral myeloid cells, including interleukin-6 and interleukin-1b.[60,61] Furthermore, mutations associated with clonal hematopoiesis are frequently found in monocytes, granulocytes, and natural killer cells compared to B or T cells, suggesting that their effects may also be manifest in multiple differentiated cell types within the hematopoietic compartment.[62] Together, these observations are relevant to the search for prognostic GE signatures in CML for the following reasons: (i) increased inflammation and cytokine production is associated with LSC persistence,[32] and disease progression;[10,63] (ii) prognostic GE changes may be found in both CD34+ and CD34– fractions of peripheral blood or bone marrow mononuclear cells; and (iii) changes in natural killer cell function and number may predict treatment-free remissions, and presumably contain informative natural killer cell GE signatures.[64-66]

Epigenetic contributions to gene expression signatures

Polycomb repressive complex-associated gene expression changes

Among the most well studied epigenetic complexes in CML are the polycomb group (PcG) proteins.[67] The polycomb group proteins assemble into two complexes, PRC2 and PRC1, which modify histones through repressive H3K27 trimethylation (H3K27me3) and H2AK119 monoubiquitination (H2AK119ub1),respectively, and in general repress gene expression.[68]

From a mechanistic standpoint, the most commonly occurring mutations in CML appear to converge in their ability to interact with and function in conjunction with the polycomb group proteins. ASXL1 functions in transcriptional repression through its interaction with PRC2 and BAP1.[69] BCORL1 is a transcriptional co-repressor that interacts with PCGF1, a core complex of the PRC1.1 complex. [70] The RUNX1-CBF-b heterodimer mediates transcription by binding to RUNX sites, but also represses transcription by interacting and recruiting BMI1 of the PRC1 complex to target sites.[71] IKZF1 regulates transcription by interacting with repressive epigenetic complexes such as HDAC1, HDAC2, CHD3, CHD4, and SWI/SNF complex, and also recruits PRC2 to target gene loci in T cells.[72] Thus, while the commonly mutated genes in CML have their own exclusive roles in transcriptional regulation, they also share a striking commonality as modulators of the PRC. Whether mutated variants of RUNX1, ASXL1, IKZF1, and BCORL1 drive aberrant PRC recruitment and GE in CML remains to be determined.

In this respect, a recent study determined that lymphoid and myeloid BC transcriptomes are highly congruent, and that both undergo PRC-driven epigenetic reprogramming towards a convergent transcriptomic state.[10] PRC-dependent epigenetic reprogramming was attributed to gain- and loss-of-function mutations in members of the PRC1 and PRC2 complexes, respectively. Of these, ongoing BMI1/PRC1 activity contributes to maintaining the BC transcriptome, while EZH2/PRC2-binding was instructional for DNA hypermethylation-dependent gene repression. Importantly, the integrative model proposed by Ko et al. suggests that enrichment for PRC-dependent GE signatures at diagnosis can predict disease transformation and TKI resistance, as highlighted above.[10]

We also note that dysregulated regulation of PRC has been identified as a key feature of TKI-resistant LSC in CP. EZH2 expression was higher in CML LSC than in normal HSC, and CML LSC have a stronger dependence on the PRC2-EZH2 axis for survival and TKI resistance.[73,74] Likewise, higher BMI1 levels at diagnosis correlated with disease progression from CP to BC.[12] while BMI1 overexpression in CP CD34+ cells increased proliferation and self-renewal,[75] and transformed B-lymphoid progenitors in vivo.[76]

DNA methylation-associated gene expression changes

Many studies have examined the role of DNA methylation as a regulator of aberrant GE in CML pathogenesis. In candidate-based approaches, genes involved in cell cycle regulation (P16, P53, PLCD1, PER3, HIC1), differentiation (HOXA4, DLX4, DDIT3, SPI1) proliferation (CDH13, DAPK1), apoptosis (BIM), Wnt regulation (sFRP1, CBY1), LSC maintenance (MTSS1), and cell signaling (Jun B, SOCS2) were identified as targets of DNA methylation.[67,7]

Recent unbiased genome-wide methylome analyses have solidified the concept of aberrant DNA methylation as a driver of resistance and transformation. The number of differentially methylated regions in CP increased from ~600 to ~6,500 CpG sites in BC.[78] BC was associated with heightened DNA hypermethylation, and to a lesser extent hypomethylation, around promoters of genes involved in stem cell fate, differentiation and leukemia-related functions. [10] Mechanistically, differential DNA methylation patterns in CML have been attributed to underlying DNMT3A/TET2 mutations, PRC2-dependent epigenetic re-programming, and cytosolic sequestration of Tet2 by BCR-ABL1.[79] Notably, the physiological targeting of DNA hypermethylation using 5-aza-2’-deoxycytidine ameliorated disease phenotypes in a mouse model of CP disease, [80] while low-dose decitabine displayed clinical activity in patients refractory to imatinib,[81] suggesting DNA methylation does indeed contribute to TKI resistance.

Based on the biological insights gleaned so far, it is possible that progression-related DNA methylation signatures may already be evident at diagnosis, particularly in patients presenting with advanced CP.[10] The DNA methylation status of specific target genes might therefore be useful in the timely identification of such patients for more aggressive therapies. Furthermore, given that DNA methylation is a relatively stable epigenetic and biochemical mark, there are practical advantages to developing DNA methylation-based biomarkers rather than transcript- based readouts, especially for the development of robust clinical-grade tests (Figure 3).

Figure 3.

Stages of development of gene expression-based biomarkers. In chronic myeloid leukemia (CML), the development of gene expression-based biomarkers can be divided into three stages following an initial discovery phase. These stages will each determine the analytical validity, clinical validity, and clinical utility of the tests in question. Examples of CML-specific issues or questions that are pertinent to each stage are outlned in boxes under each stage. GE: gene expression; IHC: immunohistochemistry; FC: flow cytometry; RT-PCR: reverse transcriptase polymerase chain reaction; ISH: in-situ hybridization; LCM: laser capture microdissection; scRNA-seq: single-cell RNA sequencing; ATAC-seq: assay for transposase-accessible chromatin sequencing; BM: bone marrow; PBMC: peripheral blood mononuclear cell; MNC: mononuclear cells; FFPE: formalin-fixed paraffin-embedded tissues; PB: peripheral blood; NK: natural killer cells; MDSC: myeloid-derived suppressor cells; TKI: tyrosine kinase inhibitor; NCCN: National Comprehensive Cancer Network; ELN: European LeukemiaNet; BC: blast crisis; EFS: event-free survival; DFS: diseasefree survival; PFS: progression-free survival; OS: overall survival, TFR: treatment-free remission; 95% CI: 95% confidence interval; DMR: deep molecular response..

Gene expression profiles and mutations: ‘seed and soil’ revisited

As described above, it will be important to develop CML models that integrate the interaction between genetic and epigenetic factors in driving drug resistance and disease transformation. In this respect, the effects of specific mutations may be cell-context dependent, with differential effects on GE and function depending on the cell type being examined. This is particularly the case for mutations affecting transcription factors, for which cell states, and their attendant chromatin accessibility profile, determine whether the mutated transcription factor has access to its target genes.

To integrate contributions from both the above features, we propose a model in which the cell of origin, with its attendant epigenetic and transcriptional program, determines the ability of specific mutations to contribute to biological and clinical outcomes (Figure 2). This model is a derivative of the ‘seed and soil’ hypothesis of cancer initiation.[82] The model will be useful for hypothesis testing, and likely explains an important feature of BCRABL1 itself. It has been shown in murine models that only when expressed in HSC, but not more committed progenitors, can BCR-ABL1 induce a myeloproliferative disorder. This is likely because BCR-ABL1 is incapable of conferring self-renewal capacity upon committed progenitors, indicating that CML cells rely on BCR-ABL1-independent mechanisms for stemness programs. These findings are in contrast to those for other leukemia fusion genes (e.g., MLL-ENL, MLL-AF9, MOZ-TIF2) which are capable of conferring self-renewal and transform progenitor cells.[83] Relatedly, the model may also explain a naturally occurring phenomenon whereby normal individuals found to carry the BCR-ABL1 fusion in their peripheral blood mononuclear cells apparently never develop CML.[84] Here, the model would posit that the BCR-ABL1 fusion is occurring in a long-lived progenitor without selfrenewal function.

Figure 2.

Diagrammatic representation of the ‘seed and soil’ model of chronic myeloid leukemia. The model proposes that both acquired mutations and the cell state of the mutation-acquiring cell contribute to the process of full transformation to blast crisis (BC). A ‘strong’ mutation is defined as being able to confer selfrenewal function on a progenitor cell that does not possess inherent self-renewal capacity. A ‘weak’ mutation is unable to confer self-renewal function and can only transform a cell with native self-renewal ability, i.e., a stem cell. For both strong and weak mutations, it is likely that additional genetic and epigenetic events are necessary to confer the full suite of features required for BC transformation. The model is also based on the recent finding that BC progenitors which harbor different somatic mutations share a common or core transcriptome enriched for stemness, quiescence, and inflammatory gene expression signatures.[20] HSC: hematopoietic stem cell; LSC: leukemia stem cell; MPP: multipotential progenitor; LMPP: lymphoid-primed multipotent progenitor; CMP: common myeloid progenitor; GMP: granulocyte- macrophage progenitor; TKI-S/R: tyrosine kinase inhibitor-sensitive/resistant cells. Rx: treatment; CP: chronic phase; BC: blast crisis; CML: chronic myeloid leukemia; GE: gene expression.

Analogous to the situation regarding cancer initiation by leukemia fusion genes, mutations devoid of selfrenewal function may only confer an increased risk of BC transformation when they occur in a target cell that already possesses physiological self-renewal function. According to this model, mutations in RUNX1 that are sufficient to induce BC-like disease in mice (Table 3) may be deemed a ‘strong’ biological seed that can transform many cell types within the hematopoietic hierarchy. Such mutations would be expected to induce disease progression in the majority of patients who harbor such mutations, which is indeed the case.[37] However, a minority of CP patients with RUNX1 mutations continue to enjoy sustained deep maolecular responses,[57] suggesting the existence of other important factors that modulate RUNX1 function. Along the same lines, ASXL1 was recently identified as the most frequently mutated gene at diagnosis in nine patients, the majority (n=6) of whom eventually developed BC, while a minority (n=2) achieved a MMR.[14]

In contrast to the above examples, the prognostic impact of ‘weak’ seeds is much less clear. In a study by Kim et al., at least four different patterns were observed for TET2 mutations.[57] One pattern is seen in patients with TKI resistance when both TET2 and ABL1 variant allele frequencies increased following TKI therapy, while another is seen when the TET2 variant allele frequency reduces after TKI treatment in patients with disease progression. In other cases, TET2 mutations were also detected within Philadelphia chromosome-negative cells, and here, patients showed complex outcomes following TKI therapy, with some achieving MMR and others showing TKI resistance. These observations suggest that the effect of TET2 mutations are highly contextual.

Challenges ahead but room for optimism

As described above, the discovery of a limited and tractable set of genes that is prognostic across a majority of CML patients has been challenging for clinical, biological, and technical reasons. Nevertheless, there is room for optimism. In the setting of breast cancer, GE panels comprising 21 genes that encompass various aspects of breast cancer biology have been found to be predictive of therapeutic response, and minimized the use of additional therapy without compromising survival.[85] Among liquid tumors, a recent study in acute myeloid leukemia demonstrated that a parsimonious 17-gene GE score, derived from a larger set of stemness-conferring genes, predicts resistance to initial therapy.[86] Interestingly, this score was independent of cytogenetic and mutational risk factors, and suggests that biological factors (e.g., stemness) transcend traditional genetics-based groupings.[87]

Encouragingly in CML, two recent reports suggest that it is possible, using peripheral blood samples taken at diagnosis or 3 months after diagnosis, to predict deep molecular responses and also sustained treatment-free remissions. In the first study, the Adelaide group showed that the rate of decline of BCR-ABL1 transcripts during first-line TKI therapy (calculated from baseline and 3- month BCR-ABL1 transcript levels) predicts success of treatment-free remission.[88] The time taken for BCR-ABL1 transcripts to halve was the strongest independent predictor of sustained treatment-free remission: 80% in patients with a halving time of <9.35 days versus 4% if the halving time was >21.85 days (P<0.001). In a separate study, Radich et al. reported that GE signatures from peripheral blood taken prior to TKI initiation can distinguish individuals who will achieve a deep molecular response (MR4.5) at 5 years from those who will have suboptimal responses.[89] Thus, biological information encoded in GE data can predict very long-term clinical outcomes in CML, and it is therefore conceivable that GE-based data will be able to identify patients in whom TKI therapy can be safely discontinued. More importantly, these early reports suggest that despite the likely existence of diverse resistance mechanisms within the study populations, final common paths, readout either as dynamic measures of BCR-ABL1 transcript levels, or peripheral blood GE signatures are indeed discoverable.

Stages in developing gene expression-based risk assessment

The stages of developing GE-based tests has been outlined in recent reviews and consensus statements, and comprise at least three phases that assess: analytical validity (reliably measuring the genotype of interest), clinical validity (ability to segregate patients into biologically and clinically important subsets), and clinical utility (ability to improve clinical decision making).[90,91] In this section, we summarize the pertinent stages and highlight issues of particular relevance to GE-based biomarker development in CML (Figure 3).

Stage 0 is the discovery stage, which is where the field is currently. Here, we highlight three important components, which include the use of technical approaches for unbiased discovery, the simultaneous interrogation of leukemic and non-leukemic clones from the same sample (since both have been shown to be prognostic), and the use of robust statistical and computational pipelines to discover minimal prognostic genes sets. The advent of single-cell-based technologies and their application to well-annotated cohorts will facilitate this step.

In stage I, the minimal gene set has to be converted into a clinical test that accurately and reproducibly measures the GE phenotype. The test platform needs to be robust, as well as sensitive, specific and reliable. The assay should be developed for tissues that are collected as part of routine clinical care. Ideally, any additional processing of material beyond what is routine should be minimized, e.g., CD34+ selection, and should utilize standard procedures available in clinical laboratories, such as flow cytometry and bone marrow immunohistochemistry. An example would be detecting GE signatures of interest by a panel of antibodies for use in flow cytometry or immunohistochemistry applications. It is preferable that the samples used for analytical validation are from wellcharacterized patients representative of ‘real-world’ settings and, ideally, validated in at least one independent cohort. Sample size and power calculations should be determined prior to starting the study, and analytic sensitivity and specificity for the test should be available at the end of the study. At the end of stage I, a locked-down test should be evaluated in stage II, that of clinical validation.

In stage II, the locked-down test will be evaluated for its ability to differentiate between clinically meaningful outcomes in modern CML practice. The samples to be tested should be obtained from well-annotated cohorts representative of the broader population, and the test conducted on tissues in a blinded manner with respect to testing and result reporting. Ideal populations include patients who have been treated uniformly in clinical trials. At the end of this stage, the ability of the test to predict clinical outcome should be available as a test score, with clearly defined positive and negative predictive values.

The final stage, stage III, will be the determination of clinical utility. This stage would entail the use of the GEbased test to improve clinical decision-making, and would require the study to demonstrate that meaningful outcomes are improved when the test is used compared to when the test is not used. Besides clinical outcomes such as improved progression-free survival and overall survival, additional measures such as cost-effectiveness, avoidance of toxicities, quality of life and psychological parameters should also be assessed. Such studies may also incorporate the contribution of pharmacological factors (e.g., drug metabolism and side effects, patient compliance) to overall outcomes. Given the relative rarity of CML, it is envisaged that this will be a multicenter international study.

Conclusion

Genetic and epigenetic events contribute to the emergence of BCR-ABL1-independent clones that result in clinical TKI resistance and, if unopposed, BC transformation. Long-term TKI responses, including successful TKI stoppage, can be predicted by slower declines in BCRABL1 transcript levels during first-line TKI therapy,[88] suggesting that genetic and epigenetic factors contributing to TKI resistance are present at diagnosis. Recent studies describe a convergent GE signature common to the majority of BC progenitors.[10] Elements of this common or core transcriptome can be detected in CD34+ cells from CP patients at risk of TKI resistance or early transformation, [11,13] and specific mutations have been shown to contribute additional nuances to the core transcriptome. [49] These observations are consistent with a ‘seed and soil’ model that may be helpful for hypothesis generation (Figure 2). Emerging technologies, particularly multimodal single-cell-based approaches, will facilitate the discovery of genetic and epigenetic biomarkers at presentation. This initial discovery phase has to be followed by the translation of GE-based information into validated analytical tests, and subsequently, the determination of clinical validity and utility. This process will be a multi-year, multi-institution international effort akin to that for the development of a genetic-based risk assessment. [85,90,91] The integration of both gene mutation- and gene expression-based biomarkers into the care of CML patients will be an important step to achieving the ultimate goal of CML research: the cure of the majority of our patients.