Identification and validation of lymphangiogenesis-related genes for predicting acute myeloid leukemia prognosis: insights from bulk RNA sequencing and single-cell RNA sequencing analyses

Abstract

Acute myeloid leukemia (AML) is a hematologic malignancy, and lymphangiogenesis can affect the proliferation, invasion, and other biological behaviors of leukemia cells. This study explored lymphangiogenesis-associated mechanisms in AML. AML datasets were downloaded from public databases. Differential expression analysis, univariate Cox regression, and machine learning were used to identify prognostic lymphangiogenesis-related genes (LYMRGs) and build a risk model. Prognostic analyses included enrichment pathway, genetic mutation, immune microenvironment, and drug sensitivity analyses. Dataset GSE116256 explored LYMRG expression in key cells; GSE142698 and RT-qPCR verified prognostic LYMRG expression. A 6-LYMRG (ANGPT1, HGF, MAPK8, PCNA, TBL1XR1, TLR4) risk model was the optimal prognostic signature. Moreover, pathways like cytokine-cytokine receptor interaction and immune cells such as macrophages were found to be associated with risk stratification in AML patients. Mutational patterns differed between different risk AML patients. High-risk AML patients showed greater sensitivity to UMI.77, vorinostat, BI.2536, tozasertib, daporinad, carmustine, MIM1, and WEHI.539. Furthermore, significant changes in prognostic LYMRG expression were observed during key cell (including progenitor cells, monocyte-derived dendritic cells, and erythroblasts) differentiation. Importantly, GSE142698 and RT-qPCR confirmed that HGF, MAPK8, PCNA, and TBL1XR1 were abundantly expressed, while TLR4 showed low expression in AML patients. This study identified ANGPT1, HGF, MAPK8, PCNA, TBL1XR1, and TLR4 as the key prognostic indicators for AML. The lymphangiogenesis-associated risk model provided an efficient tool for predicting patient survival and might facilitate the development of personalized treatment strategies for AML.

Graphical abstract

Supplementary Information

The online version contains supplementary material available at 10.1007/s10238-026-02067-w.

Introduction

Acute myeloid leukemia (AML) represents an aggressive hematologic malignancy characterized by marked cellular heterogeneity, arising from the aberrant differentiation of hematopoietic stem and progenitor cells (HSPCs). Among adults, it constitutes the predominant form of rapidly developing leukemia. Its characteristic is the abnormal proliferation of primitive and immature cells in the bone marrow (BM), which inhibits normal hematopoiesis. Clinical manifestations typically include recurrent infections due to neutropenia, anemia resulting from impaired erythropoiesis, and bleeding complications secondary to thrombocytopenia [1]. Although the clinical remission rate of AML has reached around 70%−80% at present, still 50%−70% of patients will relapse [2, 3]. This indicates that multiple different treatment regimens need to be used in combination to treat AML to increase long-term efficacy [4]. Therefore, further exploration of the pathogenesis of AML at the biological level and screening of new prognostic-related genes are of great importance. Such investigations could provide a scientific foundation for risk stratification, treatment optimization, and ultimately improved patient survival outcomes.

Lymphangiogenesis refers to the process of neolymphatic vessel formation within and surrounding tumor tissues, as well as the dilation of pre-existing lymphatic networks. This process has been demonstrated to play a critical role in lymphogenous metastasis of solid tumors and is closely associated with adverse patient prognosis [5]. However, the biological functions and underlying mechanisms of lymphangiogenesis in acute myeloid leukemia (AML) remain subjects of considerable debate and represent a significant research gap. On one hand, existing studies have predominantly focused on angiogenesis in AML, confirming that increased bone marrow microvessel density and aberrant expression of factors such as vascular endothelial growth factor (VEGF) are correlated with AML progression and prognosis [6, 7]. In contrast, direct investigations into lymphangiogenesis in AML remain relatively scarce. On the other hand, most research has concentrated on the potential direct regulatory effects of lymphangiogenic factors on AML cell proliferation, survival, or migration, rather than classical lymphangiogenesis [8]. To date, no systematic study has been reported on how lymphangiogenesis-related genes (LYMRGs) exert specific biological functions in AML or influence patient outcomes. Therefore, a comprehensive investigation into the role and mechanisms of LYMRGs in AML, along with clarifying the relationship between lymphangiogenesis and AML pathogenesis, holds significant importance for identifying novel therapeutic targets and prognostic biomarkers.

Single-cell RNA sequencing (scRNA-seq) is a revolutionary technology that enables high-throughput transcriptomic analysis at individual cell resolution [5]. This technology facilitates the exploration of cellular heterogeneity within complex tissues and enables inference of intercellular communication networks, providing unprecedented resolution in transcriptomic characterization [6]. ScRNA-seq enables refined AML subtype classification and guides personalized therapeutic strategies by elucidating disease heterogeneity, facilitating risk stratification, and identifying novel therapeutic targets throughout the disease continuum [7]. In the studies of AML and lymphangiogenesis, combining bulk RNA sequencing (RNA-seq) and scRNA-seq can achieve a comprehensive exploration from gene expression patterns to the cellular heterogeneity level, and better reveal the molecular mechanisms of AML occurrence and development.

This study mainly explores the prognostic value of LYMRGs in AML through bulk RNA-seq data, and screens out the prognostic LYMRGs to provide new therapeutic targets for the disease. Based on these prognostic LYMRGs, a risk model is constructed to analyze the biological pathways participated by patients in the high-risk and low-risk groups, and to explore the differences in the immune microenvironment, genomic variations, immunotherapy, and drug sensitivity. scRNA-seq research provides a unique opportunity to reveal key cell heterogeneity and molecular characteristics. Finally, the expression of prognostic LYMRGs is verified through reverse transcription-quantitative polymerase chain reaction (RT-qPCR) to determine whether it is consistent with the results of bioinformatics. This multi-dimensional research strategy provides new theoretical support and reference basis for understanding gene expression regulation, disease mechanisms, and the development of new drugs. It is expected to promote the targeted lymphangiogenesis therapy for AML to enter a new stage of development.

Materials and methods

Data acquisition

Gene expression profiles and clinical outcome information for AML cases were obtained from The Cancer Genome AtlasA (TCGA) repository (http://cancergenome.nih.gov/, last accessed January 13, 2025). Four AML-related transcriptomic datasets (GSE142698, GSE114868, GSE12417, and GSE116256) were sourced from the Gene Expression Omnibus (GEO) platform (http://www.ncbi.nlm.nih.gov/geo/). Comprehensive sample characteristics for these datasets are detailed in Table 1. Furthermore, we compiled 660 genes associated with lymphangiogenesis (LYMRGs) from a published study [8], as listed in Supplementary Table 1.

Table 1.

The detailed information of included datasets in the study

Datasets	Platform	AML cases	Controls	Sample type	Category of this study	Purpose
GSE142698	GPL19956	24	24	Blood samples	Assisted dataset	Identification of DEGs1
GSE114868	GPL17586	194	20	Bone marrow tissue samples	Assisted dataset	Identification of DEGs2
TCGA-LAML	-	132	-	Peripheral blood samples	Training set	Construction of risk model, etc.
GSE12417	GPL570	79	-	Peripheral blood and bone marrow mononuclear cell samples(1/78)	Validation set	Validation of risk model
GSE116256	GPL18573	16 (GSM3587923, GSM3587925, GSM3587927, GSM3587931, GSM3587940, GSM3587946, GSM3587950, GSM3587953, GSM3587959, GSM3587963, GSM3587969, GSM3587980, GSM3587984, GSM3587988, GSM3587990, GSM3587992)	4 (GSM3587996, GSM3587997, GSM3587998, GSM3588000	Bone marrow tissue samples	scRNA-seq dataset	Gene expression and distribution analysis

Differential expression analysis, functional enrichment analysis, and protein-protein interaction (PPI) network

Gene expression differences (DEGs1 and DEGs2) between AML cases and healthy controls were identified in datasets GSE142698 and GSE114868 using limma software (version 3.58.1) [9], applying criteria of |log2FC| > 0.5 and p-value < 0.05. Subsequently, visualization was performed through volcano plots and heatmaps generated with ggplot2 (v 3.5.1) [10] and ComplexHeatmap (v 2.18.0) [11] packages, respectively. To identify overlapping expression patterns, the VennDiagram package (v 1.7.3) [12] was utilized to find common upregulated and downregulated genes across both datasets (GSE142698 and GSE114868). Differentially expressed lymphangiogenesis-related genes (DE-LYMRGs) were obtained by identifying the overlap between consistently regulated DEGs and the lymphangiogenesis gene set.

To elucidate the molecular mechanisms and cellular pathways associated with DE-LYMRGs in AML pathogenesis, functional enrichment analysis was conducted using clusterProfiler (version 4.10.1) [13] for both Gene Ontology (GO) terms and Kyoto Encyclopedia of Genes and Genomes (KEGG) pathways with statistical significance set at p < 0.05. Additionally, protein-protein interaction networks were constructed via the STRING platform (http://www.string-db.org/) using a confidence threshold of 0.4, with network visualization performed through Cytoscape (version 3.9.1) [14] to investigate interconnections between DE-LYMRG-associated proteins.

Establishment and assessment of a risk model

To assess the prognostic significance of DE-LYMRGs in AML, 132 patients with complete survival data from TCGA-LAML were randomly divided into training (“TCGA-LAML1”, n = 92) and internal validation (“TCGA-LAML2”, n = 40) cohorts at a 7:3 proportion. This division enabled both model construction and internal validation for predicting AML patient overall survival (OS). Survival-relevant LYMRGs were first screened from DE-LYMRGs in the training set through univariate Cox regression (HR ≠ 1, p < 0.2) combined with proportional hazards assumption testing (p > 0.05) using the survival package (v 3.7-0.7) [15]. Subsequently, Least absolute shrinkage and selection operator (LASSO) regularization via glmnet package (v 4.1–8.1) [16] with 10-fold cross-validation was employed to refine prognostic gene selection and optimize model performance. The resulting prognostic signature was formulated using the following mathematical expression:

)

In this equation, coef represents the regression coefficient while expr indicates the gene expression level for each prognostic LYMRG. TCGA-LAML1 patients were stratified into high-risk (HRG) and low-risk (LRG) categories based on median risk score cutoff values. Visual representations depicted risk score distributions, patient survival status, and prognostic LYMRG expression patterns across the two risk categories. Survival analysis was performed using survminer (v 0.4.9) [17] for Kaplan-Meier survival plotting and survivalROC (v 1.0.3.1) [18] for receiver operating characteristic curve generation to assess model predictive performance. Risk models demonstrating area under the curve (AUC) values above 0.6 for 1-, 2-, and 3-year survival predictions were deemed to possess acceptable discriminative capacity. Finally, this study adopted the same methodology to validate the model in the TCGA-LAML2 internal validation cohort and the GSE12417 external dataset, so as to verify its reproducibility and prognostic reliability.

Screening of independent prognostic factors and construction of nomogram

Independent prognostic factor screening was performed based on all AML samples with complete survival information in the TCGA-LAML1 dataset. According to the literature [19], cytogenetic characteristics and molecular genetic risk factors were classified into four tiers: Normal, Favorable, Intermediate, and Adverse. The survival package [20]was used to conduct univariate Cox regression analysis on risk scores and different characteristics (age, gender, Cytogenetic, Molecular, CALGB) (Hazard Ratio (HR) ≠ 1, p < 0.05), followed by the PH assumption test (p > 0.05). For the factors that met the PH assumption test, the survival package was applied to conduct multivariate Cox regression analysis (HR ≠ 1, p < 0.05), as well as conduct the multivariate PH assumption test (p > 0.05). The influencing factors that passed the multivariate PH assumption test were identified as independent prognostic factors for AML. Subsequently, the rms package (v 8.0–0) [21]was used to construct a nomogram for predicting the overall survival of AML patients based on these independent prognostic factors. To evaluate the predictive performance of the nomogram, calibration curves for 1-, 2-, and 3-year survival rates of AML patients were generated using the rms package, and corresponding ROC curves (AUC > 0.7) were generated using the timeROC package (v 0.4) [22] to comprehensively assess the accuracy and reliability of the nomogram prediction model. The rmda package (v1.6) [23] was used to perform decision curve analysis (DCA) for 1-, 2-, and 3-year overall survival (OS) and visualize the results to further evaluate the clinical value of the nomogram.

Gene set enrichment analysis (GSEA)

GSEA was conducted to further determine the differential biological functions between the HRG and LRG in TCGA-LAML1. Notably, the DESeq2 package (v 1.40.2) [24] was employed to perform differential analysis between the two groups, yielding the log2FC values of DEGs3. These log2FC values of DEGs3 were then sorted in ascending order. Based on the sorting results, the clusterProfiler package (v 4.10.1) was employed to perform GSEA (|NES| >1, adj.p < 0.05). Notably, “c2.cp.kegg.v7.0.symbols.gmt” (as background gene set) was retrieved from the Molecular Signature Database (MSigDB) (https://github.com/broadinstitute/ssGSEA2.0/blob/master/db/msigdb/c2.cp.kegg.v7.0.symbols.gmt).

DNA damage repair pathway analysis

The DNA damage repair pathway is essential for maintaining the stability of cellular genetic material, and its proper function is critical for cell survival. In TCGA-LAML1, the DNA damage repair pathway (including homologous recombination (HR), nucleotide excision repair (NER), base excision repair (BER), mismatch repair (MMR), Fanconi anemia (FA), checkpoint factors (CPF), and nonhomologous end-joining (NHEJ)) was utilized as a reference gene set, GSVA package (v 1.50.5) [25] was employed to perform ssGSEA to evaluate pathway activity scores in the HRG and LRG. Pathway activity score differences were then compared between these groups via the limma package (v3.56.2) (p < 0.05, |t| > 2.0). Additionally, the Spearman correlation between risk scores and differentially activated DNA repair pathways was assessed using the cor function from the corrplot package (v 0.95) [26].

Immune microenvironment analysis

The progression and drug resistance of AML are closely linked to the high degree of heterogeneity of immune cells within the tumor microenvironment (TME) [27]. Using the GSVA package (v 1.53.28), the ssGSEA algorithm was employed to calculate the infiltration scores of 28 immune cells [28] between the HRG and LRG in TCGA-LAML1. The Wilcoxon test was employed to obtain differential immune cells (DICs) (p < 0.05). Spearman analysis between prognostic LYMRGs and DICs was carried out to probe into their relationships (|correlation coefficient (cor)| > 0.30 and p < 0.05). Immune checkpoints act as “brake pads” used by tumor cells to suppress the killing ability of immune cells. The Wilcoxon test (p < 0.05) was conducted to investigate the expression patterns of immune checkpoints [29] between risk groups. Immune Phenotype Score (IPS) data for AML patients were downloaded from the Cancer Immunome Atlas (TCIA, https://tcia.at/). Variations in IPS components (including MHC_IPS, EC_IPS, SC_IPS, CP_IPS, AZ_IPS) between risk groups were evaluated (p < 0.05).

Genetic mutation analysis

Genetic mutation analysis involves the number of non-synonymous mutations per megabase (Mb) in the tumor genome, reflecting the degree of genetic instability of AML [30]. In TCGA-LAML1, somatic mutation profiles (including single-nucleotide variants (SNVs) and insertions/deletions (INDELs)) in the HRG and LRG were systematically analyzed. The maftools package (v 2.16.0) [31] was employed to visualize and summarize mutated genes and major oncogenic pathways between the two groups.

Drug sensitivity analysis

Drug sensitivity analysis was carried out via the Genomics of Drug Sensitivity in Cancer (GDSC) database (https://www.cancerrxgene.org/) to provide management recommendations for AML. In TCGA-LAML1, the oncoPredict package (v 1.2) [32] was applied to compute the half-maximal inhibitory concentration (IC50) of 198 conventional chemotherapeutic agents to infer drug sensitivity. The correlation between prognostic LYMRGs and the IC50 of chemotherapeutic drugs was evaluated by Spearman analysis. Key drugs with a correlation coefficient > 0.6 with prognostic LYMRGs were selected for further analysis. The Wilcoxon test was applied to analyze the variations in medication sensitivity of key drugs for AML clinical treatment between the HRG and LRG (p < 0.05). Spearman analysis between key drugs and prognostic LYMRGs was performed to probe into their relationships via the cor function (|cor| > 0.30 and p < 0.05).

Identification of key cells using scRNA-seq data in AML

In all samples of GSE116256, the Seurat package (v 5.1.0) [33] was used to transform scRNA-seq data into Seurat objects. Firstly, scRNA-seq data were filtered via the subset function to select high-quality cells (nFeature_RNA between 200 and 4,000, mitochondrial proportion less than 10%, and genes expressed in at least three cells). The LogNormalize function was used to standardize the data, and the vst method from the FindVariableFeatures function was utilized to extract the top 2,000 highly variable genes (HVGs). The ScaleData function was applied to scale scRNA-seq samples. Top principal components (PCs) were determined by principal component analysis (PCA) using the RunPCA and JackStrawPlot functions (p < 0.05). Uniform manifold approximation and projection (UMAP) was applied for cluster determination via the FindNeighbors and FindClusters functions (resolution = 0.7). Then, cells were annotated as different types according to marker genes via the FindAllMarkers function, the singleR package (v 2.2.0) [34], and the literature [35]. The expression of marker genes across annotated cells was visualized.

The ReactomeGSA package (v 1.14.0) [36] was employed to examine the biological pathways in which the annotated cells were involved in the development of AML (p < 0.05). Moreover, a histogram was created to display the proportions of various annotated cells in all samples. The abundance differences among annotated cells between AML and normal groups were analyzed (p < 0.05). The annotated cells with significant differences were defined as differential cells. UMAP analysis was conducted to examine the distribution of prognostic LYMRGs at the single-cell level. The expression disparities of prognostic LYMRGs between AML and normal groups in the annotated cells were analyzed. Annotated cell types with significant differences (p < 0.05) and those that significantly affect the AML process according to relevant literature were screened out as key cells.

Heterogeneity and pseudo-time analysis of key cells

In AML and normal samples from GSE116256, key cells were subjected to a secondary round of dimensionality reduction and clustering using the same method. Subsequently, the key cells were annotated as different subtypes via the FindAllMarkers function and CellMarker database (http://xteam.xbio.top/CellMarker/) into different key cell subtypes. Variations in the expression levels of prognostic LYMRGs were assessed in the key cell subtypes. Moreover, the Monocle package (v 2.30.1) [37] was utilized for pseudo-time analysis to explore their differentiation status across different subgroups. Variations in the expression levels of prognostic LYMRGs were assessed during key cell differentiation.

Expression validation of prognostic LYMRGs by bioinformatics and reverse transcription quantitative PCR (RT-qPCR)

Statistical analysis of prognostic LYMRG expression variations between AML cases and healthy controls in the GSE142698 dataset was performed via Wilcoxon rank-sum testing with significance set at p < 0.05. For experimental validation, peripheral blood specimens were obtained from 5 AML cases and 5 healthy subjects at Chifeng Municipal Hospital under approved ethical guidelines (Ethics Committee approval: CK20250807), with written informed consent from all participants. RNA extraction was performed using Trizol reagent (Ambion, Texas, USA), followed by reverse transcription to generate cDNA with the SweScript First Strand cDNA synthesis kit (Servicebio, Wuhan, China) according to manufacturer protocols. Quantitative PCR reactions were conducted in 10 µL volumes under optimized conditions. Expression quantification of prognostic LYMRGs was achieved through the 2^−ΔΔCt comparative method. Custom RT-qPCR primers were synthesized by Sangon Biotech (Shanghai, China) with sequences detailed in Table 2. GAPDH served as the housekeeping gene for expression normalization.

Table 2.

The primer sequences

Primer	Sequence
TLR4-F	GATAGCGAGCCACGCATTCA
TLR4-R	TTAGGAACCACCTCCACGCA
ANGPT1-F	TGCTGAACGGTCACACAGAG
ANGPT1-R	GTACTGCCAGCACACTCCTT
HGF-F	TGATACCACACGAACACAGC
HGF-R	GCAAGAATTTGTGCCGGTGT
MAPK8-F	CGGTCTTGCAGCCTTACAGT
MAPK8-R	TTGCTTTTGTCAGGCACAGC
PCNA-F	GTAGCAGAGTGGTCGTTGTCT
PCNA-R	TCCTTGAGTGCCTCCAACAC
TBL1XR1-F	AGGAGGGGAATTTCCTTGTGC
TBL1XR1-R	ACAGGATATAACCCTGGAAACGG
GAPDH-F	ATGGGCAGCCGTTAGGAAAG
GAPDH-R	AGGAAAAGCATCACCCGGAG

Statistical analysis

All datasets were analyzed with R software (version 4.3.3) and GraphPad Prism (v 10). To evaluate differences in the relative mRNA expression levels of prognostic LYMRGs, we used a t-test. Additionally, the Wilcoxon test was applied for comparative analyses, where a p < 0.05 was considered statistically significant.

Results

Identification and functional assessment of 21 DE-LYMRGs

GSE142698 revealed 220 DEGs1 (Supplementary Table 2), including 134 up-regulated genes and 86 down-regulated genes in the AML group (Fig. 1a). GSE114868 revealed 6,826 DEGs2 (Supplementary Table 3), containing 3,789 up-regulated genes and 3,037 down-regulated genes in the AML group (Fig. 1b). The volcano plots depicted the top 10 (up/down) regulated DEGs based on the |log2FC| values from largest to smallest. These DEGs had diverse expression patterns between the AML and control samples in the heat maps. The intersection of up-regulated DEGs from GSE142698 and GSE114868 yielded 76 genes (Fig. 1c), while the intersection of down-regulated DEGs from the two datasets identified 45 genes (Fig. 1d). The union of these two gene sets was further intersected with 660 LYMRGs, resulting in 21 DE-LYMRGs (Fig. 1e, Supplementary Table 4).

Fig. 1.

Analysis of differentially expressed genes (DEGs) and identification of candidate genes in acute myeloid leukemia patients. (a-b) Volcano plot and heatmap of differential gene expression analysis in the dataset. In the left volcano plot, orange dots represent upregulated genes, green dots represent downregulated genes, and gray dots represent genes with no significant difference or minimal fold change. The right panel shows a heatmap; the upper part displays the expression density heatmap, and the lower part shows the expression heatmap. Pink indicates highly expressed genes, and blue indicates lowly expressed genes. (a) GSE142698 dataset. (b) GSE114868 dataset. (c) Intersection of upregulated genes from the two datasets. (d) Intersection of downregulated genes from the two datasets. (e) Intersection result between the merged genes from panels C and D and lymphangiogenesis-related genes

GO and KEGG analyses were conducted on the 21 DE-LYMRGs to elucidate the molecular biological processes. A total of 1,136 GO terms were enriched, such as “chemotaxis”, “cell chemotaxis”, “external side of plasma membrane”, “phosphatidylinositol 3-kinase complex”, “DNA-binding transcription factor binding”, and “receptor tyrosine kinase binding” (p < 0.05, Fig. 2a, Supplementary Table 5). Furthermore, KEGG analysis demonstrated that 21 DE-LYMRGs were enriched in 125 KEGG pathways, like “PI3K-Akt signaling pathway”, “HIF-1 signaling pathway”, “focal adhesion”, “Rap1 signaling pathway”, “FoxO signaling pathway”, and “JAK-STAT signaling pathway” (p < 0.05, Fig. 2b, Supplementary Table 6). The protein interactions among the 21 DE-LYMRGs were investigated, yielding a PPI network comprising 21 proteins and 62 interaction pairs. Within this network, MYC, IL6, PIK3R1, TLR4, PRCM1, and CREBBP exhibited extensive interactions with other proteins (Fig. 2c). These analyses provided insights into the molecular mechanisms underlying AML.

Fig. 2.

Enrichment analysis of candidate genes. (a) Bar plot of Gene Ontology (GO) enrichment analysis. The y-axis represents GO terms, and the x-axis represents the number of genes enriched in the pathway. (b) Results of Kyoto Encyclopedia of Genes and Genomes (KEGG) enrichment analysis. (c) Protein-protein interaction (PPI) network of candidate genes. Node colors are mapped based on the number of edges, with darker colors indicating more interactions with other nodes

Excellent performance of the prognostic LYMRG-related risk model in AML patients

The univariate Cox regression analysis and PH assumption test identified six survival-related LYMRGs (including ANGPT1, HGF, MAPK8, PCNA, TBL1XR1, and TLR4) notably linked to OS in TCGA-LAML1 (Fig. 3a, Supplementary Fig. 1). Next, to optimize gene selection and develop a more reliable risk model, 10-fold cross-validated LASSO analysis (optimal lambda = 0.0251123) was performed to screen out the six most predictive prognostic LYMRGs: ANGPT1, HGF, MAPK8, PCNA, TBL1XR1, and TLR4 (Fig. 3b). Subsequently, the specific formula for the AML patient risk model was: risk score = (0.0265) * ANGPT1 expression + (−0.0079) * HGF expression + (−0.0031) * MAPK8 expression + (−0.0915) * TBL1XR1 expression + (0.0066) * PCNA expression + (0.0084) * TLR4 expression. Based on these six prognostic LYMRGs, a prognostic LYMRG-related risk model was constructed, and the AML samples were classified into HRG (n = 46) and LRG (n = 46) using the median risk score (−0.8083402) in TCGA-LAML1. The risk score profile and survival condition plot illustrated that there was a lower survival probability and more deaths in the HRG. The expression patterns of prognostic LYMRGs varied significantly across different risk groups (Fig. 3c). Kaplan-Meier (K-M) curves displayed that high-risk AML patients possessed notably lower survival probability, suggesting that these patients experienced a poorer prognosis (p = 0.0014, Fig. 3d). The risk model was shown to accurately forecast OS in AML patients, with AUC values of 0.750, 0.747, and 0.851 for 1-, 2-, and 3-year prognoses, respectively (Fig. 3e). Furthermore, the same median risk score cutoff was applied to stratify AML samples in both the TCGA-LAML2 and GSE12417 datasets to evaluate the stability of the risk model, which yielded comparable results to TCGA-LAML1 (Fig. 3f-k). Overall, a prognostic LYMRG-related risk model was successfully established and revealed good performance in predicting the OS of AML patients, suggesting its potential as a clinical tool for prognostic stratification.

Fig. 3.

Identification of prognostic genes and construction and evaluation of the prognostic model. (a) Univariate Cox regression analysis screened for prognosis-related genes. From left to right: survival-related genes with prognostic value, corresponding P-values, hazard ratios, and their confidence intervals. Smaller confidence interval ranges indicate higher reliability. (b) Least absolute shrinkage and selection operator (LASSO) regression analysis. The left panel shows the coefficient penalty plot (x-axis: log Lambda, y-axis: coefficients). The right panel shows the cross-validation error curve (x-axis: log Lambda, y-axis: deviance explained). (c-e) Construction and evaluation of the prognostic model in the training set (n = 92) based on univariate Cox regression results. (c) KM survival curves: the upper part shows survival probability (y-axis) over time (x-axis); the lower part shows sample grouping (y-axis) over time (x-axis), with numbers indicating sample size. (d) Upper panel: risk curve (red: high-risk samples, blue: low-risk samples). Middle panel: survival status distribution (red: deceased, blue: alive). Lower panel: expression heatmap of prognostic genes in high- and low-risk groups. (e) ROC curves for 1, 2, and 3 years (x-axis: False positive rate, y-axis: True positive rate). (f-h) Construction and evaluation of the prognostic model in the internal validation set (n = 40). (i-k) Construction and evaluation of the prognostic model in the GSE12417 validation set (n = 79)

Nomogram had good predictive performance

Univariate Cox analysis showed that age, risk score, CALGB treatment regimen, and risk score were significant risk factors, while Intermediate in cytogenetic stratification was a protective factor (p < 0.05), and all these factors passed the PH assumption test (p > 0.05) (Fig. 4a)(Table 3). Subsequently, further multivariate Cox regression analysis combined with PH assumption test (p > 0.05) confirmed that risk score and age were independent factors affecting the prognosis of AML patients (p < 0.05) (Fig. 4b) (Table 4). Then, a nomogram model was constructed based on risk score and age to predict the 1-year, 2-year, and 3-year survival probabilities of AML patients (Fig. 4c). The calibration curve could intuitively reflect the consistency between the predicted probabilities of the model and the actual survival probabilities of patients. Among them, the predicted probabilities of 1-year, 2-year, and 3-year survival had a high degree of overlap with the reference line, indicating that the prediction accuracy of the nomogram was good (Fig. 4d). ROC analysis was then conducted to further evaluate the prognostic prediction accuracy of the model at different time points. The results showed that at 1-year, 2-year, and 3-year time points, the predicted AUC values of the model were 0.8, 0.81, and 0.88, respectively, suggesting that the prediction model had certain accuracy (Fig. 4e). The DCA results show that the average returns of the model over 1, 2, and 3 years are greater than those of ALL treatment strategies (Fig. 4f). In conclusion, the constructed nomogram model could provide accurate decision support for the prognosis of AML patients.

Fig. 4.

Screening of independent prognostic factors and construction of nomograms. (a) Univariate Cox analysis of risk score and different clinical features. (b) Multivariate Cox analysis of risk score and different clinical features. The x-axis represents the Hazard Ratio (HR); the red dot indicates the point estimate of HR, and the line segments on both sides denote the 95% Confidence Interval (95% CI) range. (c) Nomogram for predicting 1-, 2-, and 3-year survival rates of AML patients. Each variable’s corresponding line segment is labeled with interval scales (representing the range of the variable), and the length of the segment reflects the contribution of this factor to the outcome event. For scoring-related labels:“Points” in the figure refers to the individual score corresponding to each variable under different value levels; “Total Points” refers to the cumulative total score obtained by summing the individual scores of all variables. (d) Calibration Curve. The x-axis represents the predicted risk probability, and the y-axis represents the actual risk probability. (e) ROC curve. The x-axis represents the false positive rate and the y-axis represents the true positive rate; curves are differentiated by color according to different years. (f) DCA curve. The region where the DCA value is greater than the “ALL” and “None” curves indicates the net benefit of the model. The x-axis represents the high-risk threshold (which can be interpreted as the disease probability), and the y-axis represents the benefit

Table 3.

Proportional hazards (PH) test for univariate Cox analysis

Varibles	PH.p
Age	0.090
Riskscore	0.615
Cytogenetic	0.205
CALGB	0.393
Molecular	0.051
Gender	0.778

Table 4.

Proportional hazards (PH) test for multivariate Cox analysis

Varibles	PH.p
Age	0.142
riskScore	0.538

Biological mechanisms associated with risk scores in AML

GSEA proved highly beneficial in clarifying broad and integrated modifications in biological mechanisms [38]. In AML, GSEA identified 80 significantly differentially enriched pathways between the two risk groups, like “cytokine-cytokine receptor interaction”, “hematopoietic cell lineage”, and “graft versus host disease” (adj p < 0.05, Fig. 5a, Supplementary Table 7).

Fig. 5.

Enrichment analysis between high- and low-risk groups and DNA damage repair pathway analysis. (a) GSEA analysis between high- and low-risk groups. Part 1: Enrichment Score (x-axis: gene set rank based on correlation). Part 2: Hit markers indicating genes within the tested gene set. Part 3: Rank value distribution of all genes. (b) Significantly differential DNA damage repair pathways between high- and low-risk groups. (c) Correlation analysis between risk scores and differential DNA damage repair pathways. Red dot position represents the correlation coefficient between the specific pathway and the risk score. **, P < 0.01

Given that cancer risk is often increased by defects in intracellular DNA repair pathways, the association between risk scores and DNA damage repair pathways was investigated. Then, seven DNA damage repair signatures were analyzed: HR, MMR, BER, NER, NHEJ, CPF, and FA. Downregulation of all DNA repair pathways was observed in the HRG, though it was not statistically significant (p > 0.05, Fig. 5b). Notably, a significant positive correlation with risk scores was shown by CPF (cor = 0.32, p < 0.01) and BER (cor = 0.18, p < 0.05) (Fig. 5c). These findings suggested that disrupted DNA repair mechanisms and altered signaling pathways might serve as potential therapeutic strategies for risk-stratified AML management.

Immune landscape associated with risk scores in AML

Tumor-infiltrating immune cells had a remarkable impact on AML progression and were closely associated with the clinical outcomes of AML patients. Immune cell infiltration analysis results demonstrated notable differences in the infiltration scores of five DICs between the HRG and LRG. Specifically, the LRG exhibited a significant increase in central memory CD4 T cells, plasmacytoid dendritic cells, macrophages, effector memory CD8 T cells, and type 1 T helper cells (p < 0.05) (Fig. 6`a-b). A correlation heatmap analysis revealed significant negative correlations between TBL1XR1/MAPK8 and five types of DICs, whereas TLR4 and the risk score exhibited significant positive associations with these cells. Among the analyzed genes, TLR4 demonstrated the strongest positive association with macrophages (cor = 0.72, p < 0.001), and MAPK8 demonstrated the strongest negative correlation with type 1 T helper cells (cor = −0.60, p < 0.001) (Fig. 6c, Supplementary Table 8). The abnormal immune infiltration associated with lymphangiogenesis in AML carries critical clinical significance. Additionally, notable differences in the expression of 11 immune checkpoints were identified between risk groups. HLA-E, LGALS9, HLA-DRB5, CD276, PDCD1, KIR3DL1, KIR3DL2, KIR2DL4, KIR2DS4, KIR2DL1, and KIR2DL3 showed elevated expression in the low-risk AML group (p < 0.05, Fig. 6d). The high-risk AML patients exhibited a significantly higher CP_IPS immune phenotype score (p < 0.05, Fig. 6e). These findings suggested that high-risk AML patients might derive greater benefit from immunotherapy.

Fig. 6.

Immune microenvironment analysis. (a) Heatmap of the abundance of 28 immune infiltrating cells in high- and low-risk groups (y-axis: immune cell proportion; top color bar: sample type, green for PD samples, red for Control samples). (b) Differences in the infiltration of each immune cell type between high- and low-risk groups (x-axis: immune cells, y-axis: immune cell infiltration score; blue: high-risk group, red: low-risk group). *: p < 0.05, ns: not significant. (c) Correlation analysis among differential immune cells. Red indicates positive correlation, blue indicates negative correlation. *: p < 0.05, **: p < 0.01, ***: p < 0.001. (d) Differences in the expression of immune checkpoint genes between high- and low-risk patients (x-axis: groups, y-axis: gene expression; yellow: high-risk group, blue: low-risk group). (e) Differences in TIDE score and TMB score between high- and low-risk groups (x-axis: group information, y-axis: score). *: p < 0.05, ns: not significant

Comparisons of somatic mutations under different risk levels

It was reported that the gradual accumulation of gene mutations is closely linked to the occurrence and development of AML [39]. The mutation types and SNV types in TCGA-LAML, as well as different risk groups, were examined. Analysis of the mutation data revealed that missense mutation was the predominant classification of genetic variation in TCGA-LAML1 (Fig. 7a). Compared with the HRG, the LRG had a decreased proportion of SNPs and a greater proportion of C-T mutations (Fig. 7b-c). TP53 exhibited the greatest mutation frequency across genes (Fig. 7d). Compared with the LRG, the HRG had a higher overall mutation rate, suggesting greater genomic instability in the HRG. In the HRG, the top 10 mutant genes were IDH2, AHNAK2, ANKS1A, APC, ARSJ, ASXL1, ASXL2, BRWD1, DNMT3A, and DOCK5, with IDH2 boasting a dominant mutation rate of 17% (Fig. 7e). Conversely, the top 10 mutant genes comprised FLT3, NPM1, TP53, ANAPC4, BPIFC, DNMT3A, PTPN11, SENP6, XPO4, and ABCA6 in the LRG, where FLT3, NPM1, and TP53 displayed mutation rates of 11% (Fig. 7f). Subsequently, the mutation profiles of major oncogenic pathways across risk groups were examined. Results indicated that oncogenic mutation pathways containing RTK-RAS, WNT, NOTCH, and PI3K were detected in both groups. Notably, in the HRG, 10 genes were affected in the RTK-RAS pathway, with pathway aberrations exhibited by eight out of 24 samples (Fig. 7g). In the LRG, five genes were affected in the NOTCH pathway, with pathway abnormalities shown by four out of 37 samples (Fig. 7h). In conclusion, the integration of mutation status and risk score demonstrated improved predictive ability regarding the prognosis of AML patients.

Fig. 7.

Genomic variation analysis. (a) Mutation types and counts in high- and low-risk groups (blue: high-risk group, red: low-risk group). (b) Counts of INS, DEL, and SNP mutations (blue: high-risk group, red: low-risk group). (c) Proportion of mutations by stage in high- and low-risk groups (blue: high-risk group, red: low-risk group). (d) Mutation frequency of different genes. (e-f) Distribution of the top 15 most frequently mutated genes in high- and low-risk groups. The plot consists of four parts: the lower left color and annotations represent mutation types; the upper bar plot represents Tumor Mutational Burden (TMB); the right bar plot shows the proportion of somatic mutation types. (e) High-risk group. (f) Low-risk group. (g-h) Mutation patterns in major oncogenic pathways between high- and low-risk groups. Left: proportion of mutations in different oncogenic pathways. Right: proportion of samples affected in different oncogenic pathways (x-axis: mutation proportion or sample proportion affected, y-axis: oncogenic pathways). Red bar length indicates the proportion of mutations or affected samples in the training set for each pathway. (g) High-risk group. (h) Low-risk group

Analysis of drug sensitivity correlations between risk groups

The GDSC database was utilized to forecast the IC50 values of commonly used chemotherapeutic drugs for AML patients in various risk groups. Spearman analysis between chemotherapy drugs and prognostic LYMRGs was used to identify 13 key drugs (including vorinostat, KU.55933, BI.2536, tozasertib, PF.4708671, daporinad, IRAK4, selumetinib, carmustine, AZD6738, UMI.77, MIM1, WEHI.539) (cor > 0.6, Fig. 8a, Supplementary Table 9), and 10 key drugs revealed notable differences (p < 0.01) in IC50 values between risk groups. Notably, low-risk AML patients showed significantly lower IC50 for PF.4,708,671 and selumetinib (p < 0.05) (Fig. 8b-c). This indicated that low-risk AML patients might be more sensitive to PF.4,708,671 and selumetinib, whereas high-risk patients might exhibit greater susceptibility to drug resistance. In contrast, the IC50 values of UMI.77, vorinostat, BI.2536, tozasertib, daporinad, carmustine, MIM1, and WEHI.539 were notably lower in high-risk AML patients (p < 0.05), indicating that these drugs might provide potential therapeutic advantages for high-risk AML patients (Fig. 8d-k). Significant associations were observed between these key drugs and prognostic LYMRGs, with TLR4 demonstrating the strongest positive correlation with vorinostat (cor = 0.69, p < 0.0001), and MAPK8 demonstrating the strongest negative correlation with MIM1 (cor = −0.52, p < 0.0001). These findings indicated a potential association between prognostic LYMRGs and drug sensitivity, offering valuable insights for personalized therapeutic strategies in AML.

Fig. 8.

Drug sensitivity analysis. (a) Correlation analysis between prognostic genes and the IC50 values of targeted drugs. (b-k) Differences in IC50 values of key drugs between high- and low-risk groups, and scatter plots of correlation between prognostic genes and IC50 values of key drugs. Left scatter plot: correlation between prognostic gene expression (y-axis) and drug IC50 values (x-axis). Right box plot: difference in drug IC50 values between high- and low-risk groups (x-axis: group, y-axis: IC50 value). *: P < 0.05, **: P < 0.01, ***: P < 0.001

Identification of key cells in the comprehensive AML single-cell landscape

Following integration and filtering of the original GSE116256 dataset, data comprised 27,899 cells and 20,362 genes before quality control (QC), whereas 27,899 cells and 20,268 genes were retained after QC (Supplementary Fig. 2a). After sample normalization, the top 2,000 HVGs were selected (Supplementary Fig. 2b). PCA and scree plots were employed to determine that the top 20 PCs were retained for downstream analysis (p < 0.05, Supplementary Fig. 2c-d). Then, the UMAP clustering plot was used to visualize the 21 distinct clusters (Fig. 9a). These cell clusters were annotated into 15 cell types: macrophages, monocytes, granulocyte-monocyte progenitor cells (GMP), progenitor cells (Prog), hematopoietic stem cells (HSC), T cells, monocyte-derived dendritic cells (Mono-DC), erythroblasts, pre-dendritic cells (Pre-DC), natural killer T cells (NKT), natural killer cells (NK), CD34 + hematopoietic stem cells (HSC(CD34+)), plasmacytoid dendritic cells (Plasma-DC), B cells and plasma cells (B_Plasma), and pro-B cells (Fig. 9b). The plot was created to visualize the expressions of marker genes (Supplementary Fig. 2e, Supplementary Table 10). Subsequently, significant pathways, such as “cytokine signaling in immune system”, “hemostasis”, “adaptive immune system”, “post-translational protein modification”, and “mitotic cell cycle”, were enriched in the annotated cells (p < 0.05, Fig. 9c).

Fig. 9.

Single-cell analysis. (a) Dimensionality reduction and clustering analysis (x and y axes: UMAP dimensions; colors represent different clusters). (b) Cell annotation analysis. (c) Enrichment analysis of annotated cell types (x-axis: cell types, y-axis: pathway names; heatmap color represents NES, red: higher score/upregulation, blue: lower score/downregulation). (d) Abundance differences of cell types across different samples. *: P < 0.05, **: P < 0.01, ***: P < 0.001, ns: not significant. (e) Distribution of prognostic genes in cell types with differential abundance. (f) Expression differences of prognostic genes among samples in cell types with differential abundance. *: P < 0.05, **: P < 0.01, ***: P < 0.001, ns: not significant

The histogram showed the relative abundance of the 15 annotated cell types, with hematopoietic stem cells, T cells, and macrophages having a high abundance in AML and normal samples (Supplementary Fig. 2f-g). Progenitor cells, Mono-DC, and erythroblasts with notable differences (p < 0.05) in abundance were defined as differentially abundant cells between the AML and normal samples (Fig. 9d). Prognostic LYMRGs were expressed in all differentially abundant cell types. Specifically, ANGPT1 and TBL1XR1 exhibited significant expression variability (p < 0.05) among different samples within each of the three differential cell populations (Fig. 9e-f). Given that progenitor cells serve as the origin of leukemia stem cells and are key drivers of differentiation arrest [40, 41], Mono-DC act as critical antigen-presenting cells (APCs) in anti-tumor immunity, with functional defects promoting immune escape [42], and erythroblasts were markers of suppressed erythropoiesis and characteristic cells of rare AML subtypes [43], these three cell types were significantly associated with AML. Progenitor cells, Mono-DC, and erythroblasts were thus selected as key cells for subsequent analysis.

Pseudo-time landscapes of progenitor cells, Mono-DC, and erythroblasts

To decipher the differentiation dynamics of AML lineage cells, we reconstructed the developmental trajectories of progenitor cells, Mono-DC, and erythroblasts. Progenitor cells were classified into two sub-clusters and annotated as distinct subtypes: Prog(FTH1+) and Prog(LYZ+) (Fig. 10a-b). HGF was significantly upregulated in the Prog(LYZ+) subtype (Fig. 10c). Pseudotime analysis was performed to delineate progenitor cell differentiation trajectories, which resolved into seven distinct states. Notably, State 5 and Prog(FTH1+) cells exhibited earlier differentiation kinetics, whereas State 1 and Prog(LYZ+) cells demonstrated delayed differentiation (Fig. 10d). Furthermore, we mapped the expression patterns of prognostic LYMRGs along the pseudotime axis. ANGPT1 was predominantly enriched in the Prog(LYZ+) stage, MAPK8 and TBL1XR1 were in the Prog(FTH1+) stage, while HGF, TLR4, and PCNA showed concurrent enrichment in both Prog(LYZ+) and Prog(FTH1+) stages (Fig. 10e).

Fig. 10.

Progenitor cell communication and pseudotime analysis. (a) Dimensionality reduction, clustering, and annotation of key cells. (b) Annotation of key cell subpopulations. (c) Expression analysis of prognostic genes across subpopulations of key cells. (d) Pseudotime analysis of key cells. Left: trajectory of key cell developmental time; Middle: trajectory map of key cells; Right: trajectory map of key cell subpopulations. (e) Expression patterns of prognostic genes at different pseudotime points during key cell differentiation. Colored areas represent gene expression density distributions at different stages: yellow (early), purple (middle), red (late)

Mono-DC were classified into three distinct subtypes and annotated as Mono-DC(FCER1A+), Mono-DC(TOP2A+), and Mono-DC(SNHG25+) subtypes (Fig. 11a-b). TBL1XR1 and PCNA were significantly expressed across all three Mono-DC subtypes, while TLR4 showed specific enrichment in Mono-DC(SNHG25+) and Mono-DC(TOP2A+) (Fig. 11c). Pseudotime trajectory analysis revealed a differentiation hierarchy originating from Mono-DC(FCER1A+) and bifurcating into Mono-DC(TOP2A+) and Mono-DC(SNHG25+) lineages (Fig. 11d). Furthermore, we mapped the expression dynamics of prognostic LYMRGs along this pseudotemporal axis. ANGPT1, PCNA, TBL1XR1, and TLR4 were primarily enriched in the Mono-DC(FCER1A+) and Mono-DC(SNHG25+) stages, MAPK8 was preferentially expressed in Mono-DC(FCER1A+), and HGF was specifically upregulated in Mono-DC(SNHG25+) (Fig. 11e).

Fig. 11.

Communication and pseudotime analysis of CD141 + dendritic cells. (a) Dimensionality reduction, clustering, and annotation of key cells. (b) Annotation of key cell subpopulations. (c) Expression analysis of prognostic genes across subpopulations of key cells. (d) Pseudotime analysis of key cells. Left: trajectory of key cell developmental time; Middle: trajectory map of key cells; Right: trajectory map of key cell subpopulations. (e) Expression patterns of prognostic genes at different pseudotime points during key cell differentiation. Colored areas represent gene expression density distributions at different stages: yellow (early), purple (middle), red (late)

Erythroblasts were classified into three distinct subtypes: Erythroblast(GLRX+), Erythroblast(SERPINB1+), and Erythroblast(SLC4A1+) (Fig. 12a-b). PCNA was significantly expressed across all three erythroblast subtypes, while TBL1XR1 was specifically enriched in Erythroblast(SERPINB1+) and Erythroblast(GLRX+) (Fig. 12c). Pseudotime trajectory analysis revealed a differentiation pathway originating from Erythroblast(GLRX+) and bifurcating into Erythroblast(SERPINB1+) and Erythroblast(SLC4A1+) lineages (Fig. 12d). Mapping the expression dynamics of prognostic LYMRGs along this pseudotemporal axis showed that ANGPT1, HGF, MAPK8, PCNA, and TBL1XR1 were primarily enriched in the Erythroblast(SERPINB1+) stage, whereas TLR4 was specifically upregulated in Erythroblast(GLRX+) (Fig. 12e). This study revealed the differentiation trajectories of AML-related cells and their association with prognostic LYMRGs, providing potential targets for AML precision therapy and prognosis evaluation.

Fig. 12.

Communication and pseudotime analysis of Proerythroblasts. (a) Dimensionality reduction, clustering, and annotation of key cells. (b) Annotation of key cell subpopulations. (c) Expression analysis of prognostic genes across subpopulations of key cells. (d) Pseudotime analysis of key cells. Left: trajectory of key cell developmental time; Middle: trajectory map of key cells; Right: trajectory map of key cell subpopulations. (e) Expression patterns of prognostic genes at different pseudotime points during key cell differentiation. Colored areas represent gene expression density distributions at different stages: yellow (early), purple (middle), red (late)

Expression confirmation of ANGPT1, HGF, MAPK8, PCNA, TBL1XR1, and TLR4

Analysis of prognostic LYMRGs in GSE142698 revealed that ANGPT1, HGF, MAPK8, PCNA, TBL1XR1, and TLR4 were important contributors to AML. The GSE142698 results demonstrated that the expression levels of ANGPT1, HGF, MAPK8, PCNA, and TBL1XR1 were notably higher in AML patients (p < 0.05). Additionally, AML patients exhibited notably lower TLR4 expression (p < 0.05, Fig. 13a). The expression trends of prognostic LYMRGs in the RT-qPCR results were consistent with those observed in the bioinformatics analysis (GSE142698) (p < 0.01, Fig. 13b). These results revealed that prognostic LYMRGs might play pivotal roles in the progression of AML.

Fig. 13.

Expression validation of prognostic genes. (a) Expression differences of prognostic genes between AML samples and control samples in the transcriptomic pediatric blood dataset. *: P < 0.05, **: P < 0.01. (b) Expression differences of prognostic genes among different clinical samples. *: P < 0.05, **: P < 0.01, ns: not significant

Discussion

AML pathogenesis is intimately linked to dysregulated vascular and lymphatic networks within the BM microenvironment. The interactions among lymphangiogenic cytokines maintain the survival rate of leukemia precursor cells in the BM. During AML progression, dysregulated pro-angiogenic cytokines promote both angiogenesis and lymphangiogenesis, subsequently disrupting normal hematopoietic cell function and immune cell homeostasis within the BM [44]. This study, using extensive research based on large-scale RNA sequencing and scRNA-seq data, clarified the complexity of AML, revealed the detailed diversity of cell populations and their impact on tumor progression and outcomes, and successfully constructed an AML risk model based on 6 prognostic LYMRGs (ANGPT1, HGF, MAPK8, PCNA, TBL1XR1, and TLR4). This model can effectively stratify patients into high-risk and low-risk groups, and has demonstrated robust prognostic prediction performance in multiple datasets, providing a new tool for risk stratification of AML patients.

The findings of this study are in line with the recognized core pathogenic mechanisms in the AML field and provide a new integrated perspective. Dysfunction of DNA repair and abnormal activation of key signaling pathways such as PI3K/AKT and JAK/STAT are important bases for the occurrence and development of AML [45–47]. Our analysis results show that the risk classification based on LYMRGs is not independent of these classic mechanisms but has a profound biological association with them. Studies have shown that factors related to lymphangiogenesis are not only mediators of microenvironment remodeling but also can affect the proliferation, survival, DNA damage response, and immune escape of leukemia cells through paracrine or direct actions, which are the core malignant phenotypes [48]. Therefore, the LYMRGs in this model can be regarded as a key molecular node connecting the microenvironment promoting lymphangiogenesis with the intrinsic oncogenic program of tumor cells.

ANGPT1 produces angiopoietin-1, belonging to the angiogenic growth factor angiopoietin family. In healthy physiological states, ANGPT1 is essential for proper development of neural, vascular, and reproductive organ systems [49]. Conversely, in AML pathology, ANGPT1 shows dysregulated overexpression. Through molecular interactions with diverse cellular factors, it promotes malignant blast cell proliferation and stemness maintenance, ultimately driving tumor progression and poor clinical prognosis.

HGF stands for hepatocyte growth factor gene. Under normal circumstances, the HGF/c-MET (mesenchymal-epithelial transition factor) signaling pathway is crucial for tissue repair, inflammation control, and immune regulation [50]. Nevertheless, the increased expression of HGF has been identified as a primary compensatory pathway involved in the emergence of resistance to MET inhibitors among AML patients [51]. Research findings indicate that HGF expression in BM specimens collected during treatment-refractory phases is 29-fold greater compared to levels detected in remission states.

MAPK8 represents the mitogen-activated protein kinase 8, a gene responsive to diverse environmental stress stimuli. Upon stimulation, this kinase undergoes nuclear translocation where it modulates stress-responsive gene networks through transcription factor phosphorylation, facilitating cellular adaptation to adverse conditions. The MAPK signaling cascade plays an essential role in controlling cell growth, maturation, and vascular development, with pathway dysregulation being implicated in numerous malignancies [52, 53].

PCNA functions as a central protein in DNA replication and damage repair processes, showing increased levels in AML [54]. Within the malignant context, cytosolic PCNA enhances tumor metabolic activity and expansion, especially within leukemia stem cell populations that rely on mitochondrial energy production for viability [55]. Consequently, therapeutic strategies directed against cytoplasmic PCNA could potentially destabilize LSC metabolic balance and help minimize relapse risk.

TBL1XR1 is a protein encoded by the TBL1XR1 gene in humans. It is a key component of the nuclear receptor co-repressor (NCoR)/retinoic acid and thyroid hormone receptor silencing mediator (SMRT) complex. By recognizing and binding to ubiquitinated histones, it promotes chromatin remodeling and thereby inhibits the transcriptional expression of specific genes [56, 57]. This is the first report of this phenomenon in myeloid-dominated acute leukemia.

TLR4 is a protein encoded by the TLR4 gene in humans and belongs to the TLR family. TLR4 is a pattern recognition receptor. It recognizes intracellular and extracellular ligands, mediates inflammation inhibition, and maintains an environment favorable for AML cells [58], which is conducive to the development of AML. Therefore, TLR4 has broad research prospects in the field of AML.

Furthermore, we investigated the mRNA expression of ANGPT1, HGF, MAPK8, PCNA, TBL1XR1, and TLR4 using RT-qPCR. The expression patterns of these genes in AML versus normal tissues corroborated our computational predictions, supporting the clinical relevance of the identified prognostic LYMRGs as potential biomarkers and therapeutic targets for AML. The LYMRG-based prognostic risk model serves as a robust clinical tool that provides novel insights into AML molecular heterogeneity and facilitates the development of precision medicine strategies.GO and KEGG pathway enrichment analysis revealed that DE-LYMRGs were significantly enriched in the PI3K-Akt, HIF-1, FoxO, and JAK-STAT signaling pathways. Additionally, GSEA identified “hematopoietic cell lineage” as the most significantly different pathway between high-risk and low-risk patient groups. Among them, the PI3K/AKT pathway is involved in a variety of biological processes, and its activation is associated with the pathogenesis and progression of various tumors [59]. Dysregulated PI3K/AKT signaling promotes cell proliferation, enhances DNA repair capacity, increases survival signals, and contributes to chemotherapy resistance in AML. In preclinical models of AML, PI3K/AKT pathway inhibitors have been shown to stimulate apoptosis and limit leukemia growth [60]. The JAK/STAT signaling pathway plays a significant role in the development and progression of AML. Overactive JAK-STAT signaling is associated with abnormal proliferation of tumor cells, so selective targeting of the JAK-STAT pathway has been proven to be an effective treatment strategy for AML [61, 62]. The FoxO signaling pathway is associated with chemotherapy resistance in various tumors. FoxO signaling plays a role in the chemosensitivity of chemotherapy-resistant cell culture models, and FoxO signaling is significantly upregulated in relapsed/refractory AML patients, and is expected to become a new target for the treatment of relapsed/refractory AML [63].

We conducted an analysis of immune cell infiltration patterns in high-risk and low-risk patient groups. The results demonstrated notable differences in the abundance of five distinct types of immune cells between the high-risk and low-risk cohorts of AML patients. We investigated the potential role of risk scores and immune cells. For instance, macrophages exhibit remarkable plasticity within the BM microenvironment, where distinct subpopulations can modulate AML blast behavior through complex cellular interactions [64]. M2-polarized macrophages support leukemic cell survival and proliferation while simultaneously suppressing T cell activation and proliferation. In contrast, M1-like macrophages act as antigen-presenting cells or produce pro-inflammatory cytokines to activate cytotoxic T lymphocytes, leading to T cell proliferation and secretion of IFNγ [65]. pDCs are the main contributors to type I interferon responses. However, in the context of cancer, pDCs have been shown to have a stronger tolerogenic state, thereby inducing regulatory T cells and producing less type I interferon, which is necessary for tumor clearance. Additionally, in AML, pDCs express lower levels of HLA-DR; thus they are unable to present antigens as effectively as cells from healthy donors, resulting in tumor immune escape [66].

We also explored the relationship between the 6 prognostic LYMRGs and the differential immune cells, and the results showed significant correlations with multiple differential immune cells. This correlation reflects the adverse role of prognostic LYMRGs in the immune microenvironment, thereby influencing the progression of AML.

By utilizing the gene mutation data from TCGA, we analyzed the mutation gene profiles of the two risk groups. This analysis revealed that the risk population involved mutations in genes such as IDH2, AHNAK2, TP53, etc., which are key drivers of AML progression. Among them, isocitrate dehydrogenase (IDH), encoded by the IDH1 and IDH2 genes, which is a homodimeric enzyme that mainly functions in the citric acid cycle. It has been reported that IDH1 and IDH2 mutations are found in 5% to 15% of AML cases [67]. These mutations lead to excessive production of D-2-hydroxyglutarate, thereby interfering with cellular metabolic regulation [68]. The AHNAK2 gene belongs to the nuclear protein AHNAK family and encodes a giant scaffold protein that participates in cell adhesion, signal transduction, and cytoskeletal organization. It drives abnormal proliferation of leukemia cells by activating the Wnt/β-catenin signaling pathway, upregulating the expression of cyclins (such as Cyclin D1) and anti-apoptotic proteins, and interacting with the RAS pathway to further support cell survival and the formation of drug resistance. TP53 is one of the most critical tumor suppressor genes, encoding the p53 protein that plays essential roles in maintaining genomic stability, cell cycle regulation, and apoptosis.TP53 mutations are associated with AML development, therapeutic resistance, and adverse clinical outcomes [69].

Beyond mutational alterations, therapeutic resistance represents a major clinical challenge in AML management. Predicting drug sensitivity allows for personalized treatment options, taking into account individual differences in response to chemotherapy drugs. Therefore, we assessed IC50 differences for anti-tumor agents between risk groups to identify potential therapeutic opportunities. Eight compounds (UMI.77, vorinostat, BI.2536, tozasertib, daporinad, carmustine, MIM1, and WEHI.539) demonstrated differential sensitivity patterns between high-risk and low-risk patients. However, clinical validation of these predictions remains necessary. For example, vorinostat may have a slightly lower efficacy as a single treatment for AML, but when combined with other chemotherapy drugs, it can provide significant therapeutic benefits. More importantly, recent studies may fill the gap between necroptosis and small molecule inhibitors to develop new strategies for better targeted AML treatment [70]. Carmustine is a traditional chemotherapy drug that can cross the blood-brain barrier and is often used in consolidation chemotherapy for AML or to prevent central nervous system leukemia. However, due to side effects such as BM suppression and liver toxicity, it is currently mostly used in combination with other drugs (such as in high-dose chemotherapy regimens for AML).

scRNA-seq technology provides a powerful approach for dissecting tumor heterogeneity at single-cell resolution, enabling the identification of potential therapeutic targets. By identifying specific cell types critical to AML pathogenesis, scRNA-seq analysis elucidates cellular dynamics and pathological processes underlying disease progression. For example, the self-renewal and fate determination of HSPCs are tightly regulated by intrinsic molecular pathways and extrinsic signals derived from the surrounding microenvironment. FOXP1 is a novel essential regulator for the maintenance of human HSPCs and the growth and survival of human AML cells [71]. The AML microenvironment can inhibit the differentiation, maturation, and function of Mono-DC. Immature or dysfunctional Mo-DCs are unable to present leukemia-associated antigens to T cells, leading to the failure to initiate anti-leukemia immune responses. At the same time, they highly express inhibitory molecules (such as PD-L1, IDO, CTLA-4 ligands, etc.) and secrete inhibitory cytokines (IL-10, TGF-β), directly inhibiting the activation and proliferation of T cells and promoting immune escape. Moreover, dynamic expression patterns of prognostic LYMRGs during cellular differentiation support their role in AML progression through key cell regulation. The expression regulation of prognostic LYMRGs at specific stages of key cell differentiation may become potential targets for intervening in AML progression. For example, TLR4 is a pattern recognition receptor characterized by binding to bacterial cell wall components such as lipopolysaccharide [72]. Endogenous TLR ligands originate from damaged host tissues or activated cells and subsequently bind to immune cells to initiate downstream signal transduction [73]. ct-CD45 is an inhibitor of DCs or T cells (Mono-DC) activated by plate-bound antibodies. ct-CD45 acts as an immunomodulator on DCs through non-classical TLR4 activation, thereby altering the polarization of co-cultured T cells [74].

In conclusion, this study successfully established a robust six-gene lymphangiogenesis-related prognostic model (ANGPT1, HGF, MAPK8, PCNA, TBL1XR1, and TLR4) that accurately stratifies AML patients and provides novel insights into the molecular mechanisms underlying disease progression through comprehensive bulk and scRNA-seq analyses. The integration of immune microenvironment characterization, mutational profiling, and drug sensitivity predictions demonstrates the clinical utility of our approach for personalized AML management. However, several limitations should be acknowledged, During the stage of differential expression analysis, public datasets from different sources were included. Although the core genes were selected through the intersection strategy to ensure greater robustness, the heterogeneity of sample types may theoretically cause potential bias in the detection of certain genes that are mainly expressed in specific cells (such as stromal cells). The relatively small experimental validation cohort (n = 10), the computational nature of drug sensitivity predictions requiring clinical validation, and the focus on transcriptomic data without proteomic or functional validation of the identified targets. Future research should prioritize expanding the RT-qPCR validation to larger, multi-center patient cohorts to confirm the generalizability of our findings across diverse populations. Additionally, investigating the functional roles of individual LYMRGs through in vitro and in vivo experiments would provide mechanistic insights essential for therapeutic development. Finally, clinical trials testing LYMRG-guided treatment selection could directly assess the translational potential of this prognostic framework, ultimately advancing precision medicine approaches in AML treatment.

Conclusion

This study successfully established a robust six-gene lymphangiogenesis-related prognostic model (ANGPT1, HGF, MAPK8, PCNA, TBL1XR1, and TLR4) for AML through comprehensive bulk and scRNA-seq analyses, demonstrating excellent predictive performance (AUC > 0.75) across multiple independent cohorts and revealing significant associations with immune microenvironment alterations, mutational landscapes, and drug sensitivity profiles. Single-cell analysis provided unprecedented insights into the role of prognostic LYMRGs in key cellular populations and their differentiation trajectories, while RT-qPCR validation confirmed the clinical relevance of our computational predictions. This LYMRG-based prognostic framework represents a powerful tool for precision medicine in AML, enabling personalized risk stratification and treatment selection while advancing our understanding of lymphangiogenesis in AML pathogenesis.

Supplementary Information

Below is the link to the electronic supplementary material.

Abbreviations

AML

Acute myeloid leukemia

LYMRGs

Lymphangiogenesis-related genes

RT-qPCR

Reverse transcription quantitative real-time polymerase chain reaction

scRNA-seq

Single-cell RNA sequencing

RNA-seq

RNA sequencing

BM

Bone marrow

DEGs

Differentially expressed genes

PPI

Protein-protein interaction

DE-LYMRGs

Differentially expressed lymphangiogenesis-related genes

GSEA

Gene set enrichment analysis

OS

Overall survival

KEGG

Kyoto Encyclopedia of Genes and Genomes

GO

Gene Ontology

HRG

High-risk

LRG

Low-risk

cor

Correlation coefficient

TCGA

The Cancer Genome Atlas

GEO

Gene Expression Omnibus

BER

Base excision repair

LASSO

Least absolute shrinkage and selection operator

ROC

receiver operating characteristic

OS

overall survival

HR

Homologous recombination

NER

Nucleotide excision repair

CIs

confidence intervals

AUC

Area under the curve

MSigDB

Molecular Signatures Database

TIME

tumor immune microenvironment

MMR

Mismatch repair

FA

Fanconi anemia

CPF

Checkpoint factors

NHEJ

Nonhomologous end-joining

DICs

Differential immune cells

IPS

Immune Phenotype Score

TCIA

The cancer immunome atlas

Mb

Megabase

SNVs

Single-nucleotide variants

INDELs

Insertions/deletions

GDSC

Genomics of drug sensitivity in cancer

IC50

Inhibitory concentration

HVGs

Highly variable genes

PCs

Principal components

PCA

Principal component analysis

UMAP

Uniform manifold approximation and projection

K-M

Kaplan-Meier

Author contributions

MFZ and WTQ contributed to study design; WTQ, RW, MHZ retrieved and analyzed the data; WTQ wrote the main manuscript; TZ, XXS prepared the fgures; MFZ contributed to supervision, review, and revision of the manuscript. All authors read and approved the final manuscript.

Funding

This work is sponsored by the major special project on public health science and technology in Tianjin (24ZXGZSY00120); Tianjin Key Medical Discipline (Specialty) Construction Project (TJWJ2023XK010), and the Science and Technology Project of Tianjin Municipal Health Committee (TJWJ2024QN040, TJWJ2023XK010, TJWJ2023QN027, TJWJ2022XK018, TJWJ2022QN030). And the incubation Fund of Tianjin First Central Hospital(No. 2025FYQN04). This work is also funded by Tianjin Key Clinical Specialty Construction Project and Tianjin Key Medical Discipline Construction Project (Grant No.TJYXZDXK-3–001 A-004).

Data availability

The datasets [GSE142698, GSE114868, GSE12417, and GSE116256] for this study can be found in the [Gene Expression Omnibus (GEO)] [http://www.ncbi.nlm.nih.gov/geo/], and The datasets [Gene expression profiles and clinical outcome information for AML cases] for this study can be found in the [The Cancer Genome AtlasA (TCGA)] [http://cancergenome.nih.gov/].

Declarations

Conflict of interest

The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

Consent for publication

Not applicable.

Ethical approval and consent to participate

This study was approved by the Ethics Committee of Chifeng Municipal Hospital(CK20250807). All samples were collected after obtaining informed consent.