Acute myeloid leukemia (AML) is an aggressive blood cancer where immature myeloid cells accumulate in the bone marrow and blood, impairing normal blood cell production. Although approximately 60-70% of patients initially respond to chemotherapy, about 75% eventually relapse and die from the disease within 5 years - one of the most discouraging statistics in oncology.
A fundamental reason for this poor outcome is intratumoral heterogeneity: AML tumors are not composed of identical cancer cells. Instead, they contain a spectrum of cell types at different stages of differentiation, with different molecular properties, proliferative capacities, and sensitivities to treatment. This diversity makes AML difficult to eliminate completely with any single therapeutic strategy.
Of particular clinical importance are leukemia stem cells (LSCs) - a subset of primitive AML cells that can self-renew indefinitely, remain quiescent (dormant) during chemotherapy, and later re-initiate disease. LSCs are thought to be the primary drivers of relapse, yet identifying and characterizing them has been difficult because they share many molecular features with normal blood-forming stem cells.
Understanding the full spectrum of cell types within AML tumors, how they relate to normal blood cell development, and how they vary between patients is essential for developing precision therapies that target the specific cell types driving disease in individual patients. This is the challenge that single-cell technologies are uniquely positioned to address.
Researchers at Massachusetts General Hospital and the Broad Institute adapted a high-throughput single-cell RNA sequencing (scRNA-seq) technology called Seq-Well to profile individual cells from bone marrow aspirates of 16 AML patients and 5 healthy donors. In total, 38,410 individual cells were profiled - an unprecedented dataset for AML research that captures the full cellular ecosystem of each tumor.
A key technical challenge was distinguishing cancer cells from normal cells in the same sample. The team developed a method to simultaneously perform scRNA-seq and single-cell genotyping by amplifying gene segments near known AML mutations from the same cells being transcriptomically profiled. This allowed them to determine not only which genes each cell expressed, but also whether it carried cancer-specific mutations - definitively marking it as malignant or normal.
For mutations in genes distant from the 3' end of transcripts, the team also applied Oxford Nanopore long-read sequencing - a newer technology that produces much longer DNA sequences than conventional methods. Long reads enabled detection of the FLT3-ITD internal tandem duplication (a 60 base pair insertion), phase mutations across the same allele, and identify gene fusion breakpoints in individual cells.
The combined transcriptional and genetic data were integrated using a random forest machine learning classifier that was first trained on genotyped cells and then used to classify all remaining cells as malignant or normal based on their gene expression patterns alone. This classifier achieved greater than 95% sensitivity and greater than 99% specificity in cross-validation, enabling reliable annotation of the full dataset.
The machine learning classifier identified six distinct types of malignant AML cells, each resembling a specific stage of normal blood cell development along the spectrum from primitive stem cells to mature myeloid cells: HSC-like (hematopoietic stem cell-like), progenitor-like, GMP-like (granulocyte-monocyte progenitor-like), promonocyte-like, monocyte-like, and cDC-like (conventional dendritic cell-like).
Critically, the relative proportions of these six cell types varied dramatically between different patients' tumors. Some AMLs consisted predominantly of one or two primitive cell types, while others contained a broad spectrum across all six categories. This patient-to-patient variability in cellular hierarchy had not been fully appreciated before and has important implications for why different patients respond differently to the same treatments.
The cell type compositions inferred by scRNA-seq correlated closely with standard clinical measurements of AML differentiation - supporting the accuracy of the method - but the single-cell approach provided far more detailed and nuanced information. For example, the classification revealed the presence of cDC-like cells in some tumors that clinical flow cytometry tests would not have detected.
When cell type-specific gene signatures were used to score 179 AML samples from The Cancer Genome Atlas (TCGA), the tumors clustered into seven groups with distinct cellular hierarchies. Each of these groups was strongly enriched for specific genetic mutations - revealing a striking correspondence between the genetics of an AML and the types of cells it contains.
The link between AML genetics and cellular hierarchy was remarkably precise. RUNX1-RUNX1T1 fusion tumors (t(8;21) translocations) were almost exclusively composed of GMP-like cells. CBFB-MYH11 fusion tumors (inv(16) translocations) showed a spectrum enriched for monocyte-like and cDC-like cells. Acute promyelocytic leukemia (APL) with PML-RARA fusions had its own distinct GMP-like signature. These near-perfect associations suggest that specific genetic lesions directly determine the differentiation block at which AML cells accumulate.
The study uncovered a particularly important distinction between two types of FLT3 mutations. Tumors with the FLT3-ITD (internal tandem duplication) mutation were enriched for primitive, undifferentiated cells with high HSC/progenitor signatures. In contrast, tumors with FLT3-TKD (tyrosine kinase domain point mutation) were enriched for more differentiated, monocyte-like cells - despite both mutations activating the same kinase. FLT3-ITD is clinically associated with worse outcomes, and these data suggest a mechanistic explanation: FLT3-ITD creates a stronger differentiation block that maintains more primitive, chemotherapy-resistant cells.
This FLT3 finding was confirmed at multiple levels: in bulk TCGA data, in single-cell analysis of individual patients, within genetic subclones of the same tumor, and in functional experiments expressing different FLT3 variants in a leukemia cell line. FLT3-ITD expression specifically increased the fraction of primitive CD34+ cells, confirming that the mutation actively suppresses differentiation rather than merely marking an aggressive subtype.
Primitive AML cells (HSC-like and progenitor-like) displayed distinctly dysregulated transcriptional programs compared to their normal counterparts: they abnormally co-expressed stemness genes (like HOXA9, BMI1, MEIS1) alongside myeloid differentiation genes (like MPO, ELANE) that are normally mutually exclusive. This aberrant co-expression pattern - termed lineage priming - may be what allows primitive AML cells to both self-renew and partially differentiate, sustaining disease while generating diverse cell types.
To test whether the identified cell types have prognostic significance, the researchers used the malignant cell type signatures to partition the 179 TCGA AMLs into two groups based on the balance of primitive versus more differentiated cell types. Patients whose tumors had higher expression of HSC/progenitor-like gene signatures had dramatically worse outcomes compared to those with higher GMP-like signatures (P less than 0.0001 by log-rank test).
This survival difference was more pronounced than either signature individually, suggesting that the relative balance between primitive and partially differentiated cells - rather than absolute levels of either alone - is what matters prognostically. A tumor dominated by primitive cells appears particularly dangerous, potentially because primitive cells are more therapy-resistant and more capable of re-initiating disease after treatment.
The prognosis-relevant gene signatures identified in this study, derived from direct single-cell data rather than bulk tissue averages, are more precise and specific than signatures used in prior studies. They are specifically calibrated to malignant cells and exclude confounding signals from normal bone marrow cells present in the same samples.
These findings support a model for clinical translation: in the future, measuring the cellular hierarchy of an AML at diagnosis - through gene expression testing or flow cytometry using markers identified in this study - could refine risk stratification and guide decisions about treatment intensity, stem cell transplantation timing, and the use of novel targeted agents.
While most AML research focuses on primitive cells, this study revealed that differentiated monocyte-like AML cells also play an important role - not in driving tumor growth directly, but in helping the tumor evade destruction by the immune system. AML tumors contained fewer T-cells and cytotoxic T-lymphocytes (CTLs) than normal bone marrow, and had relatively more T-regulatory cells (T-regs), consistent with an immunosuppressive tumor environment.
In laboratory experiments, sorted CD14+ monocyte-like cells from AML cell lines and primary patient tumors inhibited T-cell activation up to 10-fold in co-culture assays - a dramatic immunosuppressive effect. In contrast, CD34+ primitive AML cells had little effect on T-cell activation, demonstrating that the immunosuppressive function is specifically a property of differentiated monocyte-like cells.
The monocyte-like AML cells express diverse immunomodulatory genes including TNF pathway genes (TRAIL/TNFSF10, TNFAIP2), IL-10 pathway genes (STAT1, HMOX1), and surface markers associated with immunosuppressive myeloid cells (CD206/MRC1 and CD163). Expression of these markers varied between patients, and high CD206 and CD163 expression correlated with worse clinical outcomes in the TCGA cohort.
This finding has significant implications for immunotherapy: AML has been less responsive to immune checkpoint inhibitors than other cancers, and the immunosuppressive activity of malignant monocyte-like cells may be a key reason. Strategies to eliminate or reprogram these cells - or to combine immunotherapy with approaches targeting the monocyte-like population - may be necessary to achieve durable immune-mediated tumor control.
This landmark study provides the most comprehensive characterization to date of the cellular ecosystem of AML tumors, generating a single-cell atlas of AML cell states, regulators, and markers that will serve as a valuable resource for the field. The combination of transcriptomic and genetic profiling in individual cells resolves questions about AML heterogeneity that bulk sequencing approaches cannot address.
Key contributions include: the identification of six malignant cell types that map onto normal hematopoietic development; the discovery that AML genetics strongly predicts cellular hierarchy; the mechanistic insight that FLT3-ITD actively blocks differentiation; the documentation of dysregulated stemness programs in primitive AML cells; and the unexpected finding that differentiated AML cells contribute to immunosuppression.
The technologies developed here - combined scRNA-seq and genotyping, nanopore-based mutation detection in single cells, and machine learning classifiers for distinguishing malignant from normal cells in complex mixtures - are broadly applicable to other cancers and will enable similar cellular ecosystem analyses across oncology.
Clinically, these findings point toward novel therapeutic strategies: targeting primitive cell programs (like HOXA9 or BMI1) to overcome self-renewal, modulating FLT3 signaling to force differentiation, and combining conventional therapies with immunotherapy approaches that counteract the immunosuppressive monocyte-like AML cell population. This single-cell atlas provides the mechanistic foundation for all of these strategies.