We obtained 29 m⁶A regulators, which including 11”writers” (METTL3, METTL14, METTL16, WTAP, METTL5, ZC3H13, CBLL1, RBM15, RBM15B, PCIF1, ZCCHC4),3 “erasers”(ALKBH5, FTO, ALKBH3) and 15 “readers”(IGF2BP3, IGF2BP2, YTHDC1, YTHDC2, YTHDF1, YTHDF2, YTHDF3, HNRNPA2B1, HNRNPC, HNRNPH1, LRPPRC, ELAVL1, FMR1, RBM33, RBMX) (15–21). m⁶A modification activity was scored using the AUCell R package, which evaluates the enrichment of predefined gene sets at the single-sample level based on the area under the recovery curve of gene expression rankings. To assess the robustness of the scoring strategy, we additionally applied the singscoreR package (22), an independent rank-based method that calculates normalized scores by comparing the relative expression ranks of signature genes within each sample, thereby minimizing the influence of between-sample normalization and technical variability.

To explore macrophage differentiation dynamics, pseudotime trajectories were constructed using Monocle3 (v1.2.9). Pseudotime trajectory analysis was performed using Monocle3. The Seurat object was first converted into a cell_data_set. Data were preprocessed using preprocess_cds with principal component analysis (num_dim = 50) to capture major transcriptional variation. Cells were clustered using the Leiden algorithm implemented in cluster_cells (k = 20, resolution = 1e−3). Trajectory inference was conducted using learn_graph with default settings (use_partition = TRUE, close_loop = FALSE) to learn a principal graph representing transcriptional state transitions. Pseudotime ordering was performed using order_cells, with the root defined as the Mono.M cluster, which represents the biologically earliest cell state based on lineage marker expression and relative pseudotime distribution.

Publicly available single-cell RNA sequencing (scRNA-seq) datasets from patients with acute myeloid leukemia (AML) were retrieved from the Gene Expression Omnibus (GEO), ArrayExpress, and published supplementary sources. Inclusion criteria were (1): availability of raw or preprocessed gene expression matrices (2), annotation or identifiable markers for immune cells, and (3) inclusion of bone marrow or peripheral blood mononuclear cells from AML patients. A total of 10 datasets encompassing n = 129 AML patients were included (Supplementary Table S1).Raw count matrices were processed using the Seurat package (v4.3.0) in R. Cells with fewer than 500 detected genes, >20% mitochondrial gene content, or doublet-like profiles were excluded. Genes expressed in fewer than 5 cells were filtered out. Datasets were normalized using Harmony to correct for batch effects.

Cell clustering was performed using principal component analysis (PCA), followed by uniform manifold approximation and projection (UMAP) for dimensionality reduction. Clusters were identified using a shared nearest neighbor (SNN) graph–based clustering approach with a resolution of 0.5, followed by marker-based cell type assignment (23, 24). Immune cell subsets were annotated based on canonical markers (25–27). Macrophage-lineage cells were subsetted for downstream analysis.

Differential gene expression (DGE) analysis was performed using FindMarkers in Seurat with Wilcoxon rank-sum test. FindAllMarkers performs pairwise comparisons of each cluster against all other cells to identify genes that are significantly up- or down-regulated. Adjusted p-values < 0.05 and |log2FC| > 0.25 were considered significant. To identify the functions of DEGs in AML subsets. Gene Ontology (GO) was conducted using the Metascape (https://metascape.org/) The results of enrichment analysis are shown using corrected p values and normalized enrichment scores.

Intercellular communication in AML cells and immune cells was analyzed using the CellChat package. Expression matrices were extracted from the Seurat object, and ligand-receptor interactions were identified using the CellChatDB.human database. Communication probabilities (computeCommunProb) and pathway-level communication strengths (computeCommunProbPathway) were calculated, and intercellular communication networks were aggregated (aggregateNet). Differences between tumor and normal tissues were compared by signaling pathway activity (netAnalysis_computeCentrality) and visualized via network diagrams (netVisual_diffInteraction) and heatmaps (netVisual_heatmap).

The upregulated DEGs and m⁶A regulators were used to explore their interactions using the Search Tool for the Retrieval of Interacting Genes/Proteins (STRING) online database (https://string-db.org/), with a median confidence score cutoff of 0.4. The protein-protein interaction (PPI) network was visualized and analyzed using Cytoscape (version 3.9.0).

Overall survival (OS) was defined as the time from diagnosis to death or last follow-up. Differences in OS among subtypes within each cohort were assessed using Mantel-Cox log-rank tests implemented in the R package survival. Kaplan–Meier curves for each cluster were visualized with the survminer package. To evaluate the predictive performance of the risk model, 1-, 3-, and 5-year receiver operating characteristic (ROC) curves were generated using the survivalROC package (v1.0.3.1). Associations analyzed using Cox proportional hazards regression implemented in the survival R package. Hazard ratios with 95% confidence intervals were used to summarize the associations, and the results were visualized using ggplot2.

Bone marrow samples from AML patients were collected at Qilu Hospital of Shandong University, with informed consent obtained and ethical approval granted. All the patients have signed the informed consents.

Total RNA was extracted using TRIzol reagent (Thermo Fisher Scientific, #15596026) following the manufacturer’s instructions. Complementary DNA (cDNA) was synthesized from 500 ng RNA using the PrimeScript™ RT Master Mix (Takara, Japan, #RR036A) in a 20 µL reaction under the following conditions: 37°C for 15 min, followed by 85°C for 5 s to inactivate the reverse transcriptase. Quantitative real-time PCR (qRT-PCR) was performed with TB Green Premix Ex Taq™ II (Takara, #RR820A) on a LightCycler 480 II system (Roche) using the following cycling program: 95°C for 30 s, followed by 40 cycles of 95°C for 5 s and 60°C for 30 s. Melt curve analysis was performed to confirm specificity of amplification. GAPDH was used as an endogenous control. Primer sequences for all genes analyzed are listed in Supplementary Table 3.

Mononuclear cells were isolated from bone marrow aspirates of AML patients by Ficoll-Paque density gradient centrifugation. After washing with PBS containing 2% fetal bovine serum (FBS), cells were incubated with human Fc receptor blocking reagent (Miltenyi Biotec) to minimize nonspecific binding, followed by staining with fluorochrome-conjugated antibodies against CD11b and CD68. Macrophages were defined as CD11b⁺CD68⁺ cells and sorted using a BD FACSAria™ III cell sorter. The purity of the sorted macrophage population was routinely greater than 95%, as confirmed by post-sort analysis.

All experiments were performed with a sample size of at least three biological replicates (n ≥ 3). Data are presented as the mean ± standard deviation (SD). Differences between two groups were evaluated using an unpaired t-test or correlation analysis, as appropriate. Two-way ANOVA was applied to assess the main and interactive effects of two factors between groups. A P value < 0.05 was considered statistically significant. Statistical analyses were performed using GraphPad Prism 9.0 or the R statistical environment.

To investigate the role of different kinds of m⁶A regulators in shaping macrophage states within the AML microenvironment, we reanalyzed publicly available scRNA-seq datasets (Supplementary Table 1), comprising a total of 856,320 cells derived from the bone marrow and peripheral blood of AML patients (Figure 1A; Supplementary Figure 1A). Based on distinct lineage-specific marker gene sets, the overall cellular landscape of AML patients was delineated, confirming the presence of diverse hematopoietic and immune lineages (Figures 1B, C). To further validate our cell type annotations, we extracted the original annotations from the GSE116256 dataset and systematically compared them with our own classifications. In addition, we examined the marker genes associated with the original clusters and compared them with those defining our clusters, demonstrating a high degree of consistency between our annotations and the previously published classifications (Supplementary Figures 1B, C). Through this validation, a total of 140,166 macrophages were identified and confirmed to represent normal immune cells rather than malignant cells (25–27), which were subsequently included in downstream stratification analyses (Figure 1D). This revealed five transcriptionally distinct macrophage phenotypes: monocyte-like macrophages, inflammatory (M1-like) macrophages, proliferating macrophages, immunosuppressive M2-like macrophages, and a unique population of SPP1⁺ LAMs characterized by high expression of exhaustion and remodeling markers.

To systematically assess m⁶A modification states, we curated a panel of 29 well-characterized regulators (Supplementary Table 2), comprising three classes (writers, erasers and readers), and applied AUCell to score their expression within individual cells. The resulting area under the ROC curve (AUC) reflects the discriminative ability of each score, with higher values indicating better predictive performance. m⁶A scoring uncovered distinct modification patterns across macrophage subsets (Figure 1E). To further validate the robustness of the m⁶A scoring strategy, we employed singscore as an independent scoring method (22). Singscore-based evaluation of m⁶A writers, erasers, and readers yielded highly consistent results with our original m⁶A score (Supplementary Figure 1E). Monocyte-derived macrophages were prominently marked by elevated expression of writer genes, while M1-like macrophages were enriched for eraser activity. In contrast, both M2-like and proliferative macrophages exhibited high reader scores. Notably, the SPP1⁺ LAMs exhibited minimal scores of m⁶A system across all three categories, suggesting uniformly low scores of m⁶A regulators, potentially associated with functional exhaustion. The cell proportion analysis further substantiated these relationships, underscoring the specific associations between the three m⁶A regulator states and distinct macrophage phenotypes (Figure 1F), suggesting that m⁶A modification may potentially regulate macrophage plasticity in AML.

We focused our analysis on macrophages with specific m⁶A scores (writer-high, eraser-high, reader-high and m⁶A-deficient). We compared the molecular and functional profiles of these four m⁶A-regulated macrophage subsets. Differentially expressed genes (DEGs) for each subset were identified using FindAllMarkers function, comparing cells in the subset (target group) against all other macrophages (comparison group).While some Differentially Expressed Genes (DEGs) were shared among the subsets (Figure 2A), each exhibited a distinct DEGs and displayed unique expression patterns characteristic of their respective states (Figures 2A, B).

We next examined the functional programs enriched in each subset. m⁶A-deficient macrophages exhibited stress-related transcriptional signatures, including hypoxia, nutrient deprivation, and pyroptotic inflammation. These features were consistent with an exhausted or dysfunctional phenotype, characterized by impaired metabolic activity and diminished immune function. Writer-high macrophages were enriched in proliferation-related and translation-related pathways, indicating a highly active functional state that aligned with their association with monocyte-derived, actively cycling macrophages. Eraser-high cells exhibited strong enrichment for pro-inflammatory programs and chemokine signaling, reflecting an M1-like immune-activating profile. Reader-high macrophages showed dominant signatures related to chromatin remodeling, gene silencing, and post-transcriptional control, suggestive of an immunoregulatory or suppressive phenotype (Figure 2C; Supplementary Figures 2A, B). Together, these findings indicated that distinct m⁶A modification states may influence macrophage biological functions by regulating epigenetic modifications, thereby shaping the fate of macrophages within the AML microenvironment.

We used Monocle3 to infer pseudotime trajectories and assess the potential role of m⁶A modification in shaping macrophage fate decisions during differentiation. Monocyte-derived macrophages were located at the root of the trajectory, which bifurcated into distinct branches corresponding to inflammatory M1-like, immunosuppressive M2-like, and terminally exhausted macrophage states (Figure 3A). Systematic changes in the expression of canonical macrophage markers (IL1B, SPP1, and MRC1) along the pseudotime trajectory further validated the inferred differentiation path and its biological relevance (Figure 3B).

Next, we examined the distribution of macrophage subsets in and their corresponding m⁶A modification states along pseudotime. We observed that the early trajectory was dominated by Mono.M cells with high writer activity, followed by a transient expansion of eraser-high macrophages corresponding to M1-like states. Along the differentiation trajectory, reader-high cells, encompassing M2-like and proliferative macrophages, progressively dominated the population. Terminally exhausted macrophages showed a widespread decline in m⁶A-related signals, consistent with an m⁶A-deficient phenotype (Figure 3C).

Finally, we evaluated the dynamic expression of individual m⁶A regulators across pseudotime. Most writer (METTL3, METTL14), eraser (FTO, ALKBH5), and reader (YTHDF1, IGF2BP2) genes showed coordinated expression changes along the differentiation trajectory (Figure 3D). These findings support the idea that different classes of m⁶A regulators drive macrophage differentiation at distinct stages, writer-high macrophages dominate early development, followed by eraser-high cells driving M1 activation, then reader-high macrophages promoting an M2 phenotype, and finally m⁶A silencing leading to macrophage exhaustion, highlighting distinct roles of m⁶A regulators in macrophage fate.

To reveal how different classes of m⁶A regulators orchestrate macrophage state transitions and fate determination, we constructed correlation-based regulatory networks centered on subset-enriched m⁶A regulators and DEGs. In writer-high cells, we observed consistent upregulation of writer genes, including METTL3, METTL14, WTAP, and RBM15 (Figure 4A, left). Downstream genes of these categories exhibit dense network connectivity and reflect a highly active, metabolically dynamic state. This state is characterized by elevated transcriptional activity (RPL17, UPF3A, RSRP1), enhanced metabolic processes (SOAT1, SLC25A36, FTH1), and increased proliferative capacity (ACTG1, CTNNA1) (Figure 4A, right). These findings indicate that m⁶A writers potentially promote macrophage activation, metabolic activity, and proliferation, contributing to unique phenotypic states in the AML microenvironment.

In eraser-high macrophages (Figure 4B), which were primarily associated with inflammatory states, ALKBH5 and FTO exhibited strong positive correlations with genes involved in cytokine production and innate immunity, including IRAK3, ALOX5, and NLRP3. This network architecture corresponded to the inflammatory phenotype characteristic of M1-like macrophages and suggest that m⁶A erasers may be involved in macrophage reprogramming, promoting their transition toward an inflammatory phenotype. Reader-high cells displayed a distinct regulatory landscape (Figure 4C), where high expression of YTHDF1, IGF2BP2, and YTHDC1 correlated with genes involved in RNA processing, chromatin remodeling, and immune evasion pathways (PRPF3, RPS6, TRIM28). m⁶A readers may orchestrate a program favoring immune evasion and cellular plasticity through enhanced RNA processing and chromatin remodeling.

m⁶A-deficient macrophages showed uniformly low expression of m⁶A regulators (Figure 4D). The gene network extensively connects immune-related genes (TYROBP, FCER1G, S100A8 and S100A9), potentially influencing immune and inflammatory responses that contribute to the development of exhaustion. In addition, metabolic-related genes (FTL and TSPO) are also involved, possibly regulating metabolism and cellular energy status, thereby impacting the exhaustion state. m⁶A-deficient represented a distinct regulatory state characterized by low expression of m⁶A regulators and a chaotic and intricate regulatory landscape, leading to disorganized gene expression and impaired cellular function. In summary, distinct m⁶A regulators drive divergent macrophage states through specific downstream gene networks, writers promote activation and proliferation, erasers support inflammatory programs, readers facilitate immune evasion by modulating RNA processing, and m⁶A-deficient cells exhibit dysregulated networks leading to exhaustion (Figure 4E).

We examined the cell–cell communication patterns of macrophages to assess how distinct m⁶A states influence their interactions within the AML microenvironment. Focusing on ligand–receptor networks, our analysis revealed an interaction landscape characterized by marked differences among the four subsets (Figures 5A, B). Notably, eraser-high macrophages exhibited reduced outgoing signaling. In contrast, writer-high cells primarily function as signal senders but have limited incoming interactions. Reader-high macrophages showed few detectable connections with other macrophage subsets and mainly interact with m⁶A-deficient cells, suggesting they may exist in an isolated state. Conversely, m⁶A-deficient macrophages maintained extensive outgoing communication, potentially exerting broader influence on other cells and the tumor microenvironment.

Dissecting the ligand-receptor pairs revealed subset-specific signaling axes. Writer-high cells predominantly transmitted MIF-(CD74+CD44) signaling to other macrophage subsets (Figure 5C), a known axis for modulating macrophage activation. Interestingly, eraser-high cells also engaged this pathway, underscoring its central role in immune activation and its capacity to reprogram surrounding cells toward a pro-inflammatory and activated state (Figure 5D). Reader-high macrophages exhibited limited intercellular interactions, with CD99-CD99 homotypic signaling being the most prominent, particularly directed toward exhausted macrophages (Figure 5E). CD99-CD99 interactions are known to promote cell proliferation, which may lead to an expansion of exhausted macrophages and thereby contribute to the formation of an immunosuppressive niche in the AML microenvironment. Similarly, m⁶A-deficient cells exhibited enhanced CD99-CD99 signaling toward reader-high cells, promoting the formation and aggravation of the immunosuppressive microenvironment (Figure 5F). Interestingly, m⁶A-deficient macrophages uniquely received HLA-CD related signals from other subsets, a signal typically associated with antigen presentation and adaptive immune activation, potentially indicative of a process aimed at reinvigorating these cells, reversing their exhausted phenotype and reinstating immune functionality.

In addition, we performed cell–cell communication analyses between macrophage subsets and T cells. The results showed that the eraser-high subset was significantly enriched in pro-inflammatory signaling pathways, suggesting a role in inflammatory activation and immune cell recruitment. Cell–cell communication analysis between macrophages and AML cells revealed that eraser-high macrophages predominantly interacted with AML cells through immune-related pathways, indicating an immune-interactive, antigen presentation–associated microenvironment. In contrast, Reader-high and m⁶A-deficient subsets exhibited signaling patterns associated with enhanced leukemic cell proliferation and immune evasion (Supplementary Figures 3A–D). These findings suggest that the eraser-high subset may exert its effects by promoting effector T-cell activation, thereby influencing AML progression.

Four m⁶A-regulated macrophage subsets modulate macrophage cell states, with eraser-high macrophages exhibiting pro-inflammatory characteristics, and reader-high as well as m⁶A-deficient subsets particularly involved in driving immunosuppressive programs. We therefore hypothesized that these macrophage states are closely linked to patient prognosis. To investigate the clinical relevance of macrophage m⁶A states, we analyzed three independent RNA-seq datasets from AML patients. We first applied AUCell to score the activity of m⁶A regulators in each patient (Figure 6A). The m⁶A abundance inferred from RNA-seq partially represents the m⁶A abundance of macrophages. In all cohorts, the m⁶A-deficient and reader-high populations were consistently dominant, suggesting a TME landscape skewed toward exhaustion and immunosuppression. To evaluate whether the m⁶A scores were associated with clinical and molecular characteristics, we examined the relationships between the three m⁶A scores and key clinical variables, including age, sex, FAB subtypes, and genetic alterations (28). We found that none of the three m⁶A scores showed a strong or consistent association with patient sex or age groups. Similarly, pairwise comparisons within groups revealed no significant differences in m⁶A scores across FAB subtypes or genetic alterations, with no clear enrichment of any specific category (Supplementary Figure 4A). Overall, these results indicate that the m⁶A-related scores are largely independent of conventional clinical features and known genetic alterations, supporting their potential role as independent prognostic indicators in AML.

We next examined whether these macrophage states correlated with patient prognosis. Kaplan-Meier survival analysis revealed that in both the GSE106291 and GSE146173 cohorts, elevated eraser activity, which may indicate a higher abundance of eraser-high macrophages, was associated with the most favorable survival outcomes (Figures 6B, C). In contrast, both reader-high and m⁶A-deficient scores, indicating a higher abundance of the corresponding m⁶A-regulated macrophages, were associated with poorer survival outcomes, with the latter predicting the worst prognosis. The writer-high score group was excluded from subsequent analyses owing to limited sample size and lack of statistical power. We constructed time-dependent ROC curves to evaluate the predictive capacity of m⁶A-regulated macrophage surrogate signature. The model demonstrated moderate predictive accuracy, as confirmed by the AUC curve analysis (Figure 6D). In GSE106291, the model showed relatively high specificity and PPV, particularly at longer follow-up times, indicating robust identification of high-risk patients. In contrast, in GSE146173, the model exhibited higher sensitivity and NPV, especially at 1 year, suggesting improved ability to exclude patients with favorable outcomes (Supplementary Table 5).

We performed a comparative analysis of the performance of two established models (Model1 (29): Risk score = 0.5890×UCP2 + 0.2590×DOCK1 -0.2193×SLC14A1 + 0.2553×SLC25A1, Model2 (30): ALKBH5 (coefficient= 0.0348), HNRNPA2B1 (coefficient=-0.0054), YTHDF3(coefficient=-0.0219), and METTL14 (coefficient=-0.0024). By calculating the concordance index (C-index) for our eraser-high prognostic model with favorable prognosis and the two previously published models, we found that our model demonstrated good discriminative ability, performing comparably to or even better than the existing models (Supplementary Figure 4B).

To validate the reliability of our findings, we collected clinical samples from AML patients and conducted long-term follow-up to record their prognostic outcomes. First, macrophages were isolated from patient samples using flow cytometry (Figure 6E). Subsequently, we examined the mRNA expression levels of three erasers (FTO, ALKBH5 and ALKBH3) in AML patient-derived macrophages. Based on a five-year survival threshold, patients were classified into good-prognosis and poor-prognosis groups. The results revealed that patients who survived for more than five years exhibited significantly higher expression levels of all three molecules compared to those with poor prognosis (Figure 6F). These observations from clinical samples were consistent with our bioinformatic analyses, supporting the notion that eraser-high macrophages are associated with a favorable prognosis in AML. Collectively, these findings suggest that patient-level m⁶A activity, reflecting the corresponding m⁶A-regulated macrophage states, may serve as a prognostic marker in AML.