Acute myeloid leukemia (AML) is a heterogeneous disease with a complex cytogenetic and molecular landscape.¹ Following conventional induction chemotherapy, patients are assigned to risk-adapted post-remission consolidative therapies.² Although the majority of AML patients respond to induction chemotherapy, refractory disease is common and relapse is a major challenge.² Currently, risk stratification is based on cytogenetic diagnostics for recurrent structural genomic abnormalities and targeted sequencing-based diagnostics for recurrent gene mutations.³ However, the origin of relapse has been traced back to therapy-resistant leukemia cells referred to as minimal residual disease (MRD), containing leukemic stem cells (LSC).⁴ The cancer stem cell concept attributes the origin of relapse to these therapy-resistant leukemia cells exhibiting specific gene expression signatures related to stemness properties.⁵ The early detection of these drug-tolerant persister cells enables allocation of patients to salvage therapies or enrollment in clinical trials prior to overt AML relapse (Figure 1A).

The combination of venetoclax and the hypomethylating agent azacitidine has become the standard of care for patients with newly diagnosed AML ≥75 years of age or those who have comorbidities that preclude the use of standard intensive chemotherapy, as a phase III clinical trial demonstrated durable remissions.⁶ Furthermore, there is a role for venetoclax + azacitidine in refractory/relapsed AML patients to induce remission prior to allogeneic stem cell transplantation and in newly diagnosed patients in whom intensive chemotherapy is not justifiable (given leukemic organ infiltration or a serious infectious complication in neutropenia).⁷

Since MRD assessment is routinely performed either by cytogenetics, targeted sequencing detecting specific molecular alterations present at diagnosis, such as NPM1 mutations, or by multiparameter flow cytometry, AML evolution trajectories and dynamic properties of therapy-resistant AML cells have not so far been captured. The era of single-cell multi-omics provides precious insights into clonal dynamics and enables tracing of distinct subclones and the detection of therapy-resistant leukemia cells during the course of AML therapy. Here, we discuss LSC biology and emphasize resistance mechanisms and therapeutic vulnerabilities thereby highlighting a role for single-cell multi-omics to detect MRD, to uncover clonal dynamics of AML and to identify new therapeutic targets aiming to prevent the re-emergence of AML.

Mature blood and immune cells exhibit tremendous diversity in cell morphology and function. The majority of these cells are derived from multipotent hematopoietic stem cells (HSC), at the top of the hierarchy within the hematopoietic organization.⁸ The hallmark of these HSC is their capacity to self-renew maintaining the resident HSC population and generating various progenitors that proliferate and differentiate into mature blood cells and immune cells. By contrast, committed progenitors have limited and steadily decreasing self-renewal capacity, are exposed to lineage fate and are exhausted within a certain time.⁹ At steady state, most HSC are inactive or in a long-term quiescent, but reversible G₀ phase of the cell cycle to maintain their long-term function, a state termed dormancy.^10,11 Classically, HSC have been considered as a discrete homogeneous population. However, more recent studies showed significant HSC heterogeneity including early lineage priming and the presence of lineage-biased HSC within the HSC compartment.^8,12,13 HSC reside in a highly specialized, hypoxic bone marrow microenvironment referred to as a niche. The niche concept was proposed in 1978 and is now viewed as a complex network that provides molecular mechanisms and physical interactions that are essential for HSC localization, maintenance and differentiation.^14,15

The sequential acquisition of somatic mutations contributing to subsequent clonal evolution constitutes a basic principle in cancer biology.¹⁶ This sequence was introduced in studies investigating mutations across different stages of colorectal cancer establishing that genetic alterations cause phenotypic manifestations.¹⁷ Since the acquisition of somatic mutations in HSC results in a mutated progeny that is endowed with a Darwinian fitness advantage, these cells are empowered to clonal expansion and will dominate the site in which they originate.¹⁶ Additional mutations have the potential to strengthen the growth advantage, resulting in different subclones contributing to independent phylogenetic lineage trees within a tumor reminiscent of a branching evolution.¹⁶ Thus, many different subclones are conceivable alongside the dominant clone at diagnosis and these do not contribute significantly to the tumor bulk population. This view shows clearly that tumors are not a collection of homogeneous cells with equal capacity for proliferation but rather a heterogeneous assembly consisting of differently functioning cells working together to maintain tumor growth as a pathophysiological organ.¹⁶ The LSC phenotype and its plasticity are shaped by distinct mechanisms including gene mutations, epigenetic modifications that result in specific gene expression programs and the metabolic states that shape a patient’s unique leukemic cell heterogeneity. Complexity is further enhanced through the crosstalk between leukemic cells and non-tumor elements, referred to as the tumor microenvironment.¹⁸

Stem cells can be functionally identified by testing self-renewal in clonal serial in vivo repopulation assays.¹⁶ AML is a prime example in which the capacity to self-renew is tested in xenotransplantation assays where LSC engraft in immune-deficient recipient mice giving rise to leukemia.¹⁹ This was first achieved 30 years ago, when it was possible to engraft normal human hematopoietic cells and leukemic cells in mice.^20-22 Engraftment and the potential to initiate leukemia was restricted to the flow-sorted CD34⁺CD38^– fraction, establishing that AML is organized as a hierarchy with CD34⁺CD38^– leukemia-initiating cells sitting at its top.^21,23 The ability of xenografts to capture even rare relapse-relevant LSC enables comprehensive investigations of therapy resistance and therapeutic approaches. The origin of leukemic cells in relapse samples can be traced back by using specific mutations as lineage tracking marks. Therefore, cells within the diagnostic and relapse samples sharing the same mutational profile will most likely originate from the same founder LSC. Individual variant allele frequencies can then be used to follow the evolution of LSC clones from diagnosis to relapse.²⁴ Leukemic cells capable of engrafting in NSG mice have been demonstrated to be transiently quiescent in the G₀ phase of the cell cycle. Following serial transplantation, a rare quiescent long-term leukemia-initiating cell population with extensive self-renewal capacity and an extremely low proliferation rate was identified.²⁵ These data suggest that only LSC subsets drive relapse. Studies with paired diagnostic/relapse samples provide evidence that relapse arises from re-emergence or clonal evolution of a pre-existing and chemotherapy-resistant clone generated before treatment.^26-28 Thus, a role for LSC in AML relapse has been demonstrated by combining sequencing of purified AML subpopulations and xenotransplantation assays from paired diagnostic/relapse samples identifying the presence of genetically diverse LSC at diagnosis and two distinct patterns of relapse based on the cell type from which relapse originates.²⁸ In the relapse origin-primitive (RO_P) group, a rare population of cells with an HSC-like phenotype already present at diagnosis generates bulk blasts that exhibit extensive myeloid differentiation. By contrast, in the relapse origin-committed group (RO_C), relapse originates from cells with an immunophenotype of a more committed progenitor. In both groups of patients relapse is linked to stem cell properties, manifested either as a primitive LSC population giving rise to relapse or as stemness transcriptional programs that are retained in the more differentiated bulk population.²⁸ These findings have considerable implications for cancer biology as well as for how AML should be monitored and treated. The identification of distinct relapse patterns emphasizes that improved methods (including single-cell multi-omics as discussed below) tracking the complex evolutionary history of AML within individual patients are inevitable in the design of further clinical trials as an attempt to overcome LSC-mediated therapy resistance and relapse. Furthermore, the shared functional and transcriptional stemness properties that underlie both cellular origins of relapse emphasize the importance of integrating new therapeutic approaches targeting stemness properties to prevent AML relapse (Table 1).

LSC can give rise to leukemic blasts that carry leukemia-related mutations and are characterized by a differentiation block. By contrast, pre-leukemic stem cells (pre-LSC) harbor recurrent pre-leukemic variants and maintain differentiation and maturation abilities capable of giving rise to mature functional progenitor cells bearing the same variants.²⁹

Clonal hematopoiesis (CH), also called CHIP (clonal hematopoiesis of indeterminate potential), is an age-related condition defined as the presence of myeloid malignancy-associated somatic driver mutations in the peripheral blood without diagnostic criteria for hematologic malignancies.^30,31 CH is associated with an increased risk of leukemia and increased mortality largely mediated by cardiovascular disease.^32,33 The latter is considered to be caused by a hyperinflammatory phenotype mediated by monocytes and macrophages bearing CH mutations that show increased production of pro-inflammatory cytokines, such as interleukin-1β and interleukin-6, in mice.^34,35 DNA methyltransferase 3A (DNMT3A) mutations are the most common driver of this state and most variants exhibit reduced protein stability correlating with strengthened clonal expansion and AML development.³⁶ The tet methylcytosine dioxygenase 2 (TET2) gene has a functionally opposite effect on DNA methylation and is also recurrently mutated in myeloid malignancies and CH.^{37-3 9} These CH mutations confer a selection advantage to the mutated cell resulting in clonal expansion.⁴⁰ Thus, DNMT3A mutation-bearing HSC in AML remission samples, without coincident NPM1 mutations present in AML blasts, have a competitive multilineage repopulation advantage over non-mutated HSC in xenografts, thereby establishing their identity as pre-leukemic HSC.²⁹ These early mutations in pre-LSC precede leukemic transformation and define a pre-leukemic state capable of generating the entire hematopoietic hierarchy (Figure 1B). AML can evolve from such a clonally expanded pre-LSC pool detected in remission samples, indicating that pre-LSC survive chemotherapy and might serve as a reservoir for clonal evolution leading to recurrent disease.²⁹ This was shown by performing deep targeted sequencing of commonly mutated leukemia genes, which revealed that DNMT3A mutations occur in an ancestral cell that gives rise to both T cells and the dominant AML clone present at diagnosis.²⁹ Xenograft repopulation assays then demonstrated that phenotypically defined DNMT3A-mutated HSC were functional pre-LSC endowed with a competitive repopulation advantage.²⁹ These data also showed that mutations in healthy HSC or at least multi-potent progenitors can serve as the cell-of-origin for myeloid leukemias in humans.

Thus, the accumulation of mutated mature blood cells arising from CH clones/pre-LSC can have an impact on atherosclerosis via monocytes/macrophages but also contribute to clonal expansion and in some rare cases give rise to frank leukemia.³⁷ Within individuals with CH, those with a high risk of developing AML can be identified in predictive models, thereby distinguishing between benign CH and the pre-leukemic state.⁴¹

In contrast, other leukemias are considered to be related to specific translocations that can be detected in these cases. AML with chromosomal rearrangements inv(16)(p13q22) or t(16;16)(p13;q22) – collectively referred to as inv(16) – and t(8;21)(q22;q22) are classified as core-binding factor (CBF) leukemias and result in the oncogenic fusion proteins CBFB-MYH11 and RUNX1-RUNX1T1 (AML1/ETO), respectively.⁴² Unique translocations are likely not sufficient to drive leukemogenesis alone and additional mutational events are needed for leukemic evolution and relapse.^42-44 However, in those AML cases, pre-LSC/LSC might not evolve from a pre-existent CH clone but structural variations may spontaneously occur in stem and progenitor cells or after exposure to genotoxic agents including chemotherapy.

The mechanisms of clonal fitness in CH constitute a field of highly competitive research and there is evidence that a pro-inflammatory phenotype contributes to cardiovascular disease. This suggests that clonal expansion is also strengthened through inflammatory pathways; although this needs to be explored further, recent data demonstrate that an enhanced inflammatory response in TET2-mutated mice correlates with progression of myeloid neoplasms.^45,46

Although LSC remain difficult to isolate because of their scarcity, their pronounced similarity to healthy HSC and their phenotypic plasticity, novel technologies now allow the identification of complex heterogeneous cell mixtures at single-cell resolution. Moreover, more sophisticated multi-omics single-cell approaches are now available to capture surface proteins next to the transcriptomes (CITEseq; cellular indexing of transcriptomes and epitopes by sequencing),^110,111 chromatin accessibility (ATAC-seq; assay for transposase-accessible chromatin with sequencing) and importantly can also integrate mutational profiling (single nucleotide variations and structural variants)¹¹² and/or tracking of clonal dynamics based on mitochondrial marker mutations (TARGET-seq,¹¹³ GoT [genotyping of transcriptomes]¹¹⁴ and MutaSeq¹¹⁵). Although it is becoming increasingly evident that dynamic changes in metabolism play critical roles in LSC function and treatment resistance, approaches based on mass spectrometry of bulk samples and metabolic flux analysis both require large numbers of cells. These are often not available from patients, in particular if smaller subpopulations such as LSC need to be analyzed and thus the development of better, high-resolution, single-cell technologies is much wanted in this area. Further technical advances, each with its inherent merits and limitations, will pave the way towards a more comprehensive understanding of clonal dynamics and the distinct (transient) single-cell states responding to AML therapy driving the continued AML evolution. The integration of single-cell genotyping adds an additional layer of information, thus allowing the capture of even rare clones and the comparison of the networks active in various (pre-)leukemic subclones and wildtype cells within the same patient. This technical progress also offers new opportunities to analyze rare CH clones in the pre-leukemic state and to capture and characterize residual, therapy-resilient relapse-initiating leukemia cells including LSC present in patients’ MRD.

AML is a highly heterogeneous disease characterized by a complex network of genetically distinct subclones arising in a branching evolution alongside the predominant clone. The cells within each genetically identical subclone show their own clone-specific molecular features and develop a specific non-genetically driven hierarchy of cellular differentiation. There is overwhelming evidence that cancer stem cells and stemness properties are clinically relevant, in particular for AML. In this review we have discussed that overcoming therapy resistance in AML requires not only eradication of bulk tumor cells but also the capture and efficient targeting of therapy-resistant leukemia cells, including LSC. The presence of genetically diverse LSC at diagnosis highlights a major limitation of therapies that target only the specific properties of the dominant clone. The identification of specific patterns of AML relapse demonstrated that these will require different therapies given their distinct stem cell biology.²⁸ Since LSC harbor inherent resistance mechanisms including phenotypic plasticity, dormancy and senescence, conventional chemotherapy is increasingly being added to or replaced by targeted therapeutic strategies to act specifically on LSC properties. Given the chemotherapy resistance of LSC, targeted strategies need to be integrated into first-line regimens to prevent LSC-mediated AML relapse. Venetoclax + azacitidine is a promising approach which is currently reserved for relapsed/refractory patients and newly diagnosed patients of older age or with comorbidities. This combination targets at least some LSC, but the molecular basis of treatment refractoriness and resistance still needs to be better explored and overcome in this setting as well. Nevertheless, BH3 mimetics are among the most promising strategies to treat AML, including LSC. Importantly, with the success of venetoclax + azacitidine and availability of this combination for first-line therapy, the selection of patients who would benefit from either standard chemotherapy or upfront venetoclax + azacitidine treatment is a challenge. Biomarkers need to be developed to stratify patients and clinical trials to monitor LSC-targeting efficacy in AML first-line regimens need to be implemented and used to study, understand and overcome LSC-mediated therapy resistance. Furthermore, new methods of disease monitoring have to be established to track LSC subclones and improve future clinical trials. Despite impressive rates of response to venetoclax + azacitidine, the combination is not curative since LSC exhibit molecular and metabolic plasticity becoming resistant to BCL-2 inhibition (i.e. high expression of MCL-1 or BCL-xL). Importantly, while venetoclax + azacitidine efficiently targets at least some LSC, it may not target all of them in both intra- and inter-patient settings, and the surviving LSC are the drivers of relapse. Thus, many different clinical trials investigating combinatorial therapeutic approaches to target LSC vulnerabilities and thereby attempt to eradicate relapse-initiating cells through different mechanisms are currently being explored in clinical settings (Table 1).

The era of single-cell multi-omics provides unprecedented opportunities to characterize relapse-initiating cell populations and allows tracking of individual clonal architectures and underlying biological networks. These technical innovations may offer new ways to trace the drivers of relapse, including LSC, within upcoming clinical trials and to identify and target therapy-resistant cells with resilience phenotypes that repopulate leukemia. Many of the cited methods have recently been applied to address basic and translational research questions and offered novel, critical insights into AML biology. However, it seems very difficult that these can be included in clinical routine diagnostics, at least at present, as they are not easily scalable. Nevertheless, these methods will be crucial to characterize the few relapse-initiating cells and to translate the understanding of LSC biology into novel therapeutic strategies for AML therapy. Results obtained from these newer, low-throughput and expensive technologies need to be translated into scalable tools that can, after clinical validation, be implemented in routine clinical practice. Such tools need to fulfill clinical standards regarding specificity, feasibility and cost-efficiency and have to be validated in larger cohorts of patients. Overall, there is increasing evidence for patient-specific approaches that address individual therapeutic vulnerabilities. An inevitable strategy to prevent AML recurrence and improve clinical outcome in the future is the integration of LSC-targeting agents into first-line treatments which may lead to a decrease in relapse frequency and an increase of cure rates of AML patients.