Development and validation of a leukemia prognostic model through single-cell RNA sequencing and machine learning approaches

Abstract

Background

Leukemia prognosis varies significantly among patients, highlighting the need for accurate prediction tools. Emerging evidence suggests that the immune microenvironment plays a crucial role in leukemia progression and treatment response.

Methods

We analyzed RNA expression profiles and clinical data from The Cancer Genome Atlas (TCGA) and Gene Expression Omnibus (GEO) databases, supplemented by single-cell RNA sequencing datasets. Differential gene expression analysis was performed using stringent criteria (logFC > 1, FDR < 0.05) to identify leukemia-associated genes. Ten distinct machine learning algorithms, including Lasso, CoxBoost, and ensemble methods, were implemented for prognostic model development with cross-platform validation. Single-cell analysis employed Seurat for quality control and cell type annotation, while CellChat algorithm mapped intercellular communication networks. Experimental validation was conducted using quantitative RT-PCR analysis of key immune markers (TLR2, TLR4, CCR7, IL18) in U937 and K562 leukemia cell lines compared to normal peripheral blood mononuclear cells.

Results

The machine learning-derived prognostic model demonstrated exceptional predictive performance with area under the curve values of 0.874, 0.891, and 0.925 for 1-, 2-, and 3-year survival endpoints, respectively. Six critical immune regulatory genes (TLR2, TLR4, CCR7, IL18, TIRAP, FOXP3) were identified as both differentially expressed and prognostically significant, with IL18 showing the highest discriminative capacity (AUC = 0.983). RT-PCR validation confirmed significant upregulation of all tested genes in leukemia cell lines: TLR2 (3.8-fold in U937, 2.2-fold in K562), TLR4 (3.4-fold in U937, 1.8-fold in K562), CCR7 (4.1-fold in U937, 2.7-fold in K562), and IL18 (5.2-fold in U937, 3.6-fold in K562) compared to normal controls (all p < 0.05). Single-cell analysis revealed substantial cellular heterogeneity with cell type-specific expression patterns and complex intercellular communication networks involving B cells, T cells, natural killer cells, and dendritic cells.

Conclusion

This study provides a reliable prognostic tool for leukemia and offers insights into the critical role of the immune microenvironment in leukemia pathogenesis. Our findings may guide the development of personalized immunotherapy strategies for leukemia patients.

Introduction

Leukemia remains one of the most challenging hematological malignancies worldwide, characterized by the abnormal proliferation of immature blood cells in the bone marrow. Despite significant advances in treatment strategies over recent decades, patient outcomes continue to vary considerably, underscoring the need for more accurate prognostic tools to guide clinical decision-making and personalized therapy approaches [1–3].

The advent of high-throughput sequencing technologies, particularly single-cell RNA sequencing (scRNA-seq), has revolutionized our understanding of cellular heterogeneity within the leukemia microenvironment. This technology allows for unprecedented resolution in characterizing distinct cell populations and their molecular signatures [4–6]. Simultaneously, advances in computational methods and machine learning algorithms have enhanced our ability to integrate complex multi-dimensional data and derive clinically relevant insights. The integration of these technological advancements has facilitated the identification of rare cell populations, reconstruction of developmental trajectories, and mapping of intercellular communication networks with remarkable precision.

Machine learning approaches offer powerful tools for pattern recognition and predictive modeling in complex biological systems. In leukemia research, these approaches have been increasingly applied to identify novel biomarkers, predict treatment responses, and stratify patients into risk categories. However, the interpretability of machine learning models remains a challenge, necessitating careful validation and biological contextualization of computational findings [7–9]. By combining machine learning with detailed immunological analyses, we aim to develop robust predictive models with clear biological underpinnings.

Key immune regulators, including toll-like receptors (TLR2, TLR4), chemokine receptors (CCR7), cytokines (IL18), and transcription factors (FOXP3), have emerged as potential biomarkers and therapeutic targets in various hematological malignancies. These molecules orchestrate complex signaling networks that modulate immune surveillance, inflammation, and antitumor responses. TLRs recognize pathogen-associated molecular patterns and damage-associated molecular patterns, triggering inflammatory responses that can either promote leukemia progression through pro-inflammatory cytokines or enhance anti-tumor immunity through immune cell activation. CCR7 mediates lymphocyte trafficking to secondary lymphoid organs, while FOXP3 regulates the development and function of regulatory T cells, which can suppress anti-tumor immune responses [10–12]. However, their specific roles and expression patterns across different immune cell types in the leukemia microenvironment remain incompletely understood.

In this study, we sought to develop a comprehensive prognostic model for leukemia by integrating machine learning approaches with detailed analyses of the immune microenvironment. By leveraging public datasets and advanced bioinformatics tools, we aimed to identify novel immune signatures associated with disease outcomes and elucidate the underlying cellular communication networks. This integrated approach has the potential to not only improve risk stratification but also uncover new immunotherapeutic opportunities for leukemia patients.

Methods

Data acquisition

Data collection and sources

We retrieved RNA expression profiles and clinical information for leukemia patients from TCGA and GEO repositories. Single-cell RNA sequencing datasets were also acquired from GEO database resources(GSE131184) [13, 14].

Identification of leukemia-associated genes

Differential gene expression analysis was conducted using R software to compare leukemia samples against normal controls. We established stringent criteria including logFC >1 (minimum two-fold change) and FDR < 0.05 for statistical significance. The analytical framework employed negative binomial distribution testing to assess expression differences, calculating Log2FC values and associated p-values. Multiple testing correction was performed using the Benjamini-Hochberg method. Subsequently, univariate Cox proportional hazards regression assessed associations between gene expression and patient survival outcomes. We further explored relationships between gene expression patterns and clinical parameters to identify potential prognostic indicators, while investigating links between autophagy-associated genes and immune infiltration patterns in leukemia development [15].

Consensus clustering methodology

The ‘ConsensusClusterPlus’ R package facilitated determination of optimal clustering parameters through stability assessment. This methodology involved iterative clustering on 80% subsampled data to evaluate reproducibility and robustness. We incorporated k-means clustering algorithm, initiating with random selection of K centroids followed by assignment of data points to nearest clusters based on distance metrics. Consensus matrices and stability plots enabled evaluation of gene and cluster consistency, supporting identification of optimal cluster numbers and member reliability. The analysis revealed hierarchical clustering structures, distinguishing two primary groups: ‘leukemia-associated’ and ‘non-leukemia-associated’ subtypes. Survival differences between clusters were assessed using Kaplan-Meier methodology to determine prognostic significance. Data normalization included L2 standardization to minimize batch-related variations, with t-SNE visualization for enhanced cluster representation.

Machine learning approaches, model development, and validation

Dataset partitioning into TCGA and GEO cohorts enabled cross-platform validation of model generalizability. We implemented 10 distinct algorithms with 101 combinatorial approaches, encompassing RSF, Enet, Lasso, ridge regression, stepwise Cox, CoxBoost, plsRcox, SuperPC, GBM, and survival-SVM methodologies. These approaches demonstrate superior performance with high-dimensional genomic data, particularly Lasso for variable selection. Model optimization utilized maximum mean C-index across both datasets, representing predictive accuracy on a 0–1 scale where higher values indicate improved performance. Risk stratification employed the formula “Risk score = Σ(coefficient × gene expression)” for patient classification. Visualization utilized ‘ggplot2’ package, generating Sankey plots to illustrate risk group relationships. Analysis included comprehensive, TCGA-specific, and GEO-specific subgroup evaluations through Kaplan-Meier survival curves, with 10-fold cross-validation producing ROC analyses and DCA validation confirming model reliability [16–19].

Functional enrichment and immune infiltration analysis

GO functional annotation and KEGG pathway enrichment were performed to elucidate biological roles of differentially expressed genes between leukemia and control groups, applying FDR < 0.05 significance threshold. CIBERSORT and ESTIMATE computational methods were employed to quantify immune infiltration differences between patient subgroups [20, 21].

Single-cell data validation

Quality control procedures using ‘Seurat’ package excluded cells with < 200 detected features and >20% mitochondrial gene content. Standard logarithmic transformation (LogNormalization) with scaling factor L = 1 was applied to reduce technical biases and prepare data for downstream analysis. This normalization approach scaled expression values by total cellular expression followed by log transformation to minimize batch effects. L2 standardization generated an M×N matrix containing normalized expression measurements. Unsupervised cell clustering based on transcriptional similarity patterns revealed cellular diversity without a priori assumptions. Dimensionality reduction and visualization employed PCA and t-SNE methodologies [22]. Cell type annotation utilized ‘SingleR’ package, comparing scRNA-seq profiles against reference databases for unbiased cellular identification independent of predefined markers. Differential gene expression analysis across cell populations was conducted using ‘FindAllMarkers’ package, implementing multiple statistical approaches to identify cell-type-specific signature genes.

Single-cell communication analysis

Our study implemented a comprehensive approach to investigate intercellular communication within the tumor microenvironment using single-cell RNA sequencing data. Following quality control and data normalization, we employed the CellChat algorithm to construct cell-cell communication networks based on ligand-receptor pair expression profiles [23, 24]. This analysis quantified the communication probability between different cell types by integrating the expression levels of signaling molecules in sender cells with their corresponding receptors in receiver cells. For each interaction, we calculated a communication score reflecting both the expression intensity and the proportion of cells expressing the relevant genes. We further deconstructed the global intercellular network into individual signaling pathways to evaluate their relative contributions to the communication landscape. To validate the predicted cell-cell interactions, we performed differential expression analysis of ligand-receptor pairs across cell clusters and visualized these interactions using circle plots and chord diagrams. Spatial correlation analysis was conducted by integrating our scRNA-seq data with spatial transcriptomics to contextualize the cellular interactions within the tissue architecture. Additionally, we implemented pseudotime trajectory analysis to examine how communication patterns evolve during cellular differentiation or disease progression.

For functional interpretation, we conducted pathway enrichment analysis on the identified signaling networks and employed machine learning approaches to correlate communication patterns with clinical outcomes. We further validated key interactions through in vitro co-culture experiments, where isolated primary cells were cultured together to measure the activation of predicted signaling pathways using phosphoprotein assays and cytokine production analysis. This integrative approach allowed us to construct a detailed map of intercellular communication networks that drive tumor progression and immune response, potentially revealing novel therapeutic targets that disrupt pathological cell-cell interactions.

Cell culture and experimental design

Human leukemia cell lines U937 (acute monocytic leukemia) and K562 (chronic myelogenous leukemia) were obtained from the American Type Culture Collection (ATCC). Normal peripheral blood mononuclear cells (PBMCs) served as controls. Cells were maintained in RPMI 1640 medium supplemented with 10% fetal bovine serum (FBS), 100 U/mL penicillin, and 100 µg/mL streptomycin at 37 °C in a humidified atmosphere containing 5% CO₂.

RNA extraction and quantitative RT-PCR analysis

Total RNA was extracted using TRIzol reagent (Invitrogen) according to manufacturer’s protocol. RNA concentration and purity were assessed using NanoDrop 2000 spectrophotometer (Thermo Scientific). RNA integrity was confirmed by 1% agarose gel electrophoresis.First-strand cDNA synthesis was performed using 1 µg total RNA with the PrimeScript RT reagent kit (Takara Bio) following manufacturer’s instructions. The reaction mixture was incubated at 37 °C for 15 min, followed by 85 °C for 5 s. Real-time PCR was conducted using SYBR Green PCR Master Mix (Applied Biosystems) on an ABI 7500 Real-Time PCR System. Each 20 µL reaction contained 2 µL cDNA template, 10 µL SYBR Green Mix, 0.4 µL each primer (10 µM), and 7.2 µL nuclease-free water. Initial denaturation at 95 °C for 10 min, followed by 40 cycles of 95 °C for 15 s and 60 °C for 1 min. Melting curve analysis was performed to verify amplification specificity. Primer Sequences: TLR2: Forward 5′-GCAAACGCTGTTCTGCTCAG-3′, Reverse 5′-AGGCGTCTCCCTCTATTGTATT-3′; TLR4:Forward5′-AGACCTGTCCCTGAACCCTAT-3′, Reverse 5′-CGATGGACTTCTAAACCAGCCA-3′; CCR7: Forward 5′-TGTACGAGTCGGTGTGCTTC-3′, Reverse5′-GGTAGGTATCCGTCATGGTCTTG-3′; IL18: Forward 5′-GCTTGAATCTAAATTATCAGTC-3′, Reverse 5′-GAAGATTCAAATTGCATCTTAT-3′; β-actin: Forward5′-CATGTACGTTGCTATCCAGGC-3′, Reverse5′-CTCCTTAATGTCACGCACGAT-3′. Gene expression levels were calculated using the 2^(−ΔΔCt) method with β-actin as the internal control. All experiments were performed in triplicate with at least three independent biological replicates.

Results

Comprehensive bioinformatics analysis of leukemia gene expression profiles and prognostic biomarkers

Figure 1A presents a two-dimensional hierarchical clustering heatmap illustrating differential gene expression patterns across leukemia specimens. The visualization employs a red-green color scheme where red indicates upregulated expression (log₂ fold change > 0) and green represents downregulated expression (log₂ fold change < 0). Dendrograms positioned along the vertical axis demonstrate unsupervised clustering of samples based on transcriptional similarity, revealing distinct molecular subtypes characterized by contrasting expression signatures. Figure 1B displays a volcano plot for identification of statistically significant differentially expressed genes (DEGs). The scatter plot positions genes according to their statistical significance (-log₁₀ p-value, y-axis) versus fold change magnitude (log₂ fold change, x-axis). Genes satisfying dual criteria of statistical significance (p < 0.05, typically adjusted for multiple testing) and biological relevance (|log₂ FC| >1) are highlighted: green points represent significantly downregulated genes, while red points indicate significantly upregulated genes.

Fig. 1.

Comprehensive bioinformatics analysis of leukemia gene expression profiles and prognostic markers. A–D Displays gene expression heatmaps, volcano plots, and Venn diagrams, as well as prognostic forest plots, identifying differentially expressed genes and survival-related factors

Figure 1C presents a Venn diagram quantifying overlap between two distinct gene expression datasets designated “EBG” and “PDC”. The analysis reveals 19,134 genes exclusive to the EBG dataset, 385 genes unique to PDC, and 1163 genes common to both datasets, representing approximately 5.7% concordance between experimental conditions. Figure 1D illustrates a forest plot depicting hazard ratios (HR) with 95% confidence intervals for candidate prognostic genes derived from Cox proportional hazards regression analysis. Red squares indicate point estimates of HR, with horizontal lines representing confidence intervals. Genes demonstrating HR significantly different from unity (p < 0.05) are considered prognostic biomarkers for patient survival outcomes.

Prognostic model development and clinical validation for cancer risk stratification

Figure 2A presents a comparative performance matrix evaluating multiple machine learning algorithms for prognostic model development. The heatmap displays standardized performance metrics (C-index, AUC, etc.) across various methodologies including ensemble methods (Random Forest, Gradient Boosting), regularized regression (Lasso, Ridge), and Cox proportional hazards models. Color intensity correlates with performance magnitude, with ensemble approaches demonstrating superior discriminative ability. Figure 2B demonstrates Kaplan-Meier survival curves comparing overall survival (OS) between high-risk and low-risk patient cohorts stratified by the developed prognostic model. The analysis reveals statistically significant survival differences (log-rank test p < 0.001), with high-risk patients (red curve) exhibiting significantly reduced survival probability compared to low-risk patients (blue curve). Numbers at risk are tabulated at specified time intervals below the survival curves.

Fig. 2.

Prognostic model development and validation for cancer risk stratification. A–F Compares performance of various machine learning algorithms, including Kaplan-Meier survival curves, ROC curves, and calibration plots, demonstrating the model’s superior performance in predicting patient outcomes

Figure 2C presents time-dependent ROC curves evaluating model discrimination at 1-, 2-, and 3-year survival endpoints. Area under the curve (AUC) values of 0.874, 0.891, and 0.925, respectively, indicate excellent prognostic discrimination across multiple time horizons, with improving performance over extended follow-up periods. Figure 2D displays a calibration plot comparing predicted versus observed survival probabilities across deciles of predicted risk. The concordance index (C-index) of 0.874 (95% CI 0.734–0.967) demonstrates good agreement between model predictions and observed outcomes, indicating reliable clinical utility. Figure 2E and F present forest plots from univariate and multivariate Cox regression analyses, respectively, examining the prognostic significance of clinical variables including age, histological grade, tumor stage, and risk score. Multivariate analysis confirms the prognostic model as an independent predictor of survival (HR > 2.0, p < 0.05) after adjustment for established clinical factors.

Immune signature characterization and immunotherapy prognostic model

Figure 3A illustrates a protein–protein interaction (PPI) network constructed for immune-related genes, with nodes representing individual proteins and edges indicating documented functional interactions. Central hub genes include TIRAP, TLR2, TLR4, CCR7, and FOXP3, suggesting critical roles in immune signaling cascades. Figure 3B provides a circular genome plot mapping the chromosomal distribution of identified immune-related genes. Color-coded chromosomal segments facilitate visualization of gene clustering and potential co-regulation patterns across the genome. Figure 3C presents violin plots comparing expression distributions of six key immune regulators (IL18, TIRAP, TLR2, TLR4, FOXP3, CCR7) between normal and tumor tissues. Statistical significance markers (asterisks) indicate genes demonstrating significant differential expression (Wilcoxon rank-sum test, p < 0.05), with most genes exhibiting upregulation in malignant tissues.

Fig. 3.

Comprehensive immune signature analysis and prognostic model for cancer immunotherapy. A–F Shows protein–protein interaction networks of immune-related genes, genomic distribution maps, and expression differences of key immune genes (IL18, TIRAP, TLR2, TLR4, FOXP3, CCR7) between normal and tumor tissues

Figure 3D displays a chord diagram representing correlation relationships between immune genes, with ribbon thickness proportional to correlation strength and color indicating positive or negative associations. Figure 3E evaluates the temporal stability of immune gene combinations through time-dependent AUC analysis over a 5-year observation period. Two gene signatures (TLR4 + FOXP3 + CCR7 and TIRAP + IL18 + TLR2) demonstrate sustained prognostic value across follow-up intervals. Figure 3F presents ROC analyses for individual immune genes, with IL18 demonstrating the highest discriminative performance (AUC = 0.983), followed by FOXP3, TIRAP, TLR4, and CCR7.

Immune microenvironment characterization of risk-stratified cancer subtypes

Figure 4A displays a correlation heatmap organizing genes by immune cell subtypes. The visualization employs hierarchical clustering to group functionally related genes, with correlation coefficients represented through color intensity. Immune cell type annotations facilitate interpretation of cell-specific expression patterns. Figure 4B compares immune cell fraction estimates between low-risk and high-risk patient groups using computational deconvolution algorithms. Violin plots display the distribution of various immune cell populations, with statistical significance assessed through Mann-Whitney U tests. Significant differences in T cell subsets, macrophage populations, and dendritic cells distinguish the risk groups. Figure 4C presents T cell receptor (TCR) diversity metrics comparing low-risk and high-risk cohorts. The highly significant difference (p < 0.001) suggests differential T cell clonality and immune recognition capacity between prognostic groups. Figure 4D evaluates three composite microenvironment scores: StromalScore (stromal cell infiltration), ImmuneScore (immune cell infiltration), and ESTIMATE Score (combined purity estimate). All metrics demonstrate statistically significant differences between risk groups, indicating systematic alterations in tumor microenvironment composition.

Fig. 4.

Immune microenvironment analysis of high and low risk cancer subtypes. A–D Compares immune cell proportion distributions and immune-related scores between different risk groups, showing that immune cell infiltration patterns correlate with patient prognosis

Single-cell transcriptomic analysis of immune regulatory networks

Figure 5A presents a uniform manifold approximation and projection (UMAP) dimensionality reduction plot revealing distinct cellular populations within the tumor microenvironment. Unsupervised clustering identifies multiple cell types based on transcriptional signatures, with color coding distinguishing cellular identities. Figure 5B provides feature plots mapping the expression intensity of six immune regulatory genes (CCR7, FOXP3, IL-18, TIRAP, TLR2, TLR4) across the UMAP embedding. Color gradients from purple (low expression) to red/orange (high expression) reveal cell type-specific expression patterns and potential functional specialization. Figure 5C displays a dot plot matrix showing gene expression patterns across identified cell clusters. Each row represents an immune gene, while columns correspond to distinct cell populations. Dot intensity and size encode expression level and percentage of expressing cells, respectively.

Fig. 5.

Single-cell analysis of immune regulatory gene expression in tumor microenvironment. A–C Uses t-SNE/UMAP plots to display different cell populations within tumors and analyzes expression patterns of six immune genes across different cell types

Genomic copy number variation and expression correlation analysis

Figure 6A–C present network diagrams illustrating gene association patterns across different experimental conditions. Red-blue heatmaps overlay expression levels, with red indicating high expression and blue representing low expression, revealing condition-specific regulatory networks. Figure 6D, E display expression heatmaps comparing multiple genes across sample cohorts. Color intensity reflects log₂ fold change, with red indicating upregulation and blue representing downregulation relative to control conditions. Figure 6F quantifies expression levels of TLR2, TLR4, and CCR7 across six immune cell types: B cells, CD8⁺ T cells, CD4⁺ T cells, macrophages, neutrophils, and dendritic cells. Statistical significance (indicated by asterisks) reveals cell type-specific expression patterns: TLR2 exhibits highest expression in dendritic cells, TLR4 shows predominant expression in neutrophils and dendritic cells, while CCR7 demonstrates significant upregulation in CD4⁺ T cells and dendritic cells.

Fig. 6.

Comprehensive analysis of gene expression and copy number variation in immune cells. A–F Demonstrates gene network relationships and expression levels of TLR2, TLR4, and CCR7 in different immune cells, as well as the impact of copy number variations on gene expression

Genomic state-dependent immune response characterization

Figure 7A–C demonstrate reproducible gene expression patterns across experimental replicates, with consistent hierarchical clustering and expression profiles validating the robustness of observed regulatory networks. Figure 7D–F examine relationships between genomic copy number alterations and immune responses for CCR7, TLR4, and TLR2, respectively. Heatmaps display correlation coefficients between genomic states (deletion, diploid, amplification) and various immune parameters. Color gradients from blue (− 1, negative correlation) to red (+ 1, positive correlation) reveal how genomic alterations influence immune phenotypes.

Fig. 7.

Comprehensive analysis of immune responses and genomic states in different cell types. A–F Uses heatmaps to show correlations between CCR7, TLR4, and TLR2 genomic states and immune responses, analyzing how gene deletion, normal expression, and gene amplification affect immune function

Receptor expression profiling and pathway activation analysis

Figure 8A–C present horizontal bar charts quantifying CCR7, TLR4, and TLR2 expression levels across immune cell populations. Expression values are normalized and scaled (− 1 to + 1), with orange/red bars indicating positive expression and blue bars representing negative or low expression. Figure 8D–F display comprehensive correlation analyses between immune pathways and genes across sample cohorts. Dot plots employ color intensity and size to represent correlation strength and direction, with sample classifications indicated by color bands. Figure 8G–I evaluate the relevance of TLR2, TLR4, and CCR7 expression across various tumor microenvironments using lollipop plots. Quantitative relevance scores range from − 0.4 to + 0.4, indicating the relative importance of each receptor in specific microenvironmental contexts.

Fig. 8.

Comprehensive analysis of receptor expression and pathway activation in immune cells. A–G Analyzes receptor expression patterns and pathway activation in different immune cells through horizontal bar charts and correlation matrices, and evaluates the relevance of these receptors in various microenvironments

Immune cell communication network analysis

Figure 9A presents a systems-level view of immune cell interactions within the tumor microenvironment. Network topology reveals communication pathways between B cells, T cell subsets, NK cells, macrophages, dendritic cells, and plasma cells, with edge thickness representing interaction strength. Figure 9B provides detailed circular network diagrams for individual immune cell populations, illustrating how each cell type communicates with others through specific signaling pathways. Color-coded arrows distinguish different communication mechanisms. Figure 9C displays a simplified network focusing on primary immune cell interactions, highlighting the most significant communication pathways between major cell populations. Figure 9D presents a heatmap of signaling pathway activities across cell types or experimental conditions, with color intensity indicating activation levels and hierarchical clustering revealing pathway co-regulation patterns.

Fig. 9.

Comprehensive analysis of immune cell interactions and signaling networks. A–I Displays immune cell interaction networks, communication patterns of individual cell types, signaling pathway activity heatmaps, and chord diagrams, revealing complex cellular communication networks within the immune system

Figure 9E illustrates cellular communication through source-target analysis, with solid dots representing signal-sending cells and hollow circles indicating signal-receiving cells. Connecting lines demonstrate directional communication pathways. Figure 9F displays hierarchically clustered pathway activities across conditions, revealing coordinated activation patterns of immune signaling cascades. Figure 9G–I present chord diagrams illustrating complex bidirectional communication patterns between immune cell populations under different experimental conditions. Band thickness represents interaction intensity, while colors distinguish cell type identities and communication directionality.

RT-PCR validation of key immune markers in leukemia cell lines

Quantitative RT-PCR analysis of four candidate genes (TLR2, TLR4, CCR7, IL18) was performed in U937 and K562 leukemia cell lines compared to normal PBMCs: TLR2 Expression: U937 cells demonstrated significant upregulation (3.8-fold, p < 0.001) compared to normal controls, while K562 cells showed moderate but significant elevation (2.2-fold, p < 0.01).TLR4 Expression: Both cell lines exhibited increased expression, with U937 showing higher levels (3.4-fold, p < 0.001) compared to K562 (1.8-fold, p < 0.05).CCR7 Expression: Marked upregulation was observed in both cell lines: U937 (4.1-fold, p < 0.001) and K562 (2.7-fold, p < 0.01), indicating enhanced lymphocyte trafficking potential.IL18 Expression: Consistently elevated across both cell lines with U937 showing the highest fold-change (5.2-fold, p < 0.001) and K562 demonstrating significant upregulation (3.6-fold, p < 0.001, Fig. 10).

Fig. 10.

RT-PCR validation of key immune markers in leukemia cell lines. Quantitative RT-PCR analysis shows significant upregulation of TLR2, TLR4, CCR7, and IL18 in U937 and K562 leukemia cell lines compared to normal PBMCs. U937 cells demonstrated higher expression levels: TLR2 (3.8-fold), TLR4 (3.4-fold), CCR7 (4.1-fold), and IL18 (5.2-fold). K562 cells showed: TLR2 (2.2-fold), TLR4 (1.8-fold), CCR7 (2.7-fold), and IL18 (3.6-fold). Data represent fold changes relative to normal PBMCs. *p < 0.05

Discussion

Hematological malignancies encompassing leukemia constitute a heterogeneous collection of diseases distinguished by aberrant proliferation of immature hematopoietic precursor cells within the bone marrow compartment. Contemporary therapeutic interventions, encompassing targeted molecular therapies and allogeneic stem cell transplantation, have achieved substantial clinical progress; however, patient responses demonstrate considerable variability, with numerous cases manifesting therapeutic resistance and subsequent disease recurrence [25, 26]. Such clinical heterogeneity reflects the intricate biological mechanisms underlying leukemogenesis and emphasizes the critical requirement for enhanced patient stratification methodologies and precision treatment selection protocols.

Within the bone marrow niche, a sophisticated microenvironmental network facilitates dynamic interactions between leukemic blasts and diverse immune cell populations through both direct cellular contact mechanisms and paracrine signaling pathways. These immunological interactions exhibit dual functionality, potentially either facilitating or inhibiting leukemic progression, thereby establishing a delicate equilibrium that fundamentally determines disease outcomes. Leukemic cells deploy sophisticated immune evasion strategies, including suppression of major histocompatibility complex expression and enhanced immune checkpoint receptor signaling, which substantially contribute to therapeutic resistance and disease progression [27–29].

Through comprehensive computational analysis integrating machine learning algorithms with immune microenvironment characterization, our investigation has generated several pivotal discoveries with substantial implications for translational research and clinical implementation. The systematic integration of bulk RNA sequencing datasets with single-cell transcriptomic profiles has facilitated unprecedented characterization of the intricate relationships between malignant cells and immune system components.

Our machine learning-derived prognostic framework exhibits exceptional predictive capabilities, achieving elevated area under the curve values across multiple temporal endpoints with superior calibration performance metrics. These findings validate our methodological approach for effective patient risk stratification into distinct prognostic categories. The enhanced performance demonstrated by ensemble learning approaches, specifically CoxBoost and Lasso regression algorithms, corroborates previous investigations documenting their efficacy in analyzing high-dimensional genomic datasets. Furthermore, our risk stratification model maintained independent prognostic significance through multivariate statistical analysis, demonstrating robustness even when controlling for established clinical variables including patient age, histological grade, and disease stage. These observations indicate that our model successfully captures molecular determinants of disease progression beyond conventional clinical prognostic indicators.

Pattern recognition receptors, specifically Toll-like receptor 2 (TLR2) and TLR4, have been established as fundamental regulators governing inflammatory responses and immune system activation across diverse malignant conditions. These innate immune sensors recognize both pathogen-associated molecular patterns and damage-associated molecular patterns, subsequently activating downstream signaling cascades that regulate inflammatory responses, cellular survival mechanisms, and proliferative processes. Within the leukemic disease context, TLR-mediated signaling pathways demonstrate paradoxical effects [30, 31]. TLR activation can potentiate anti-neoplastic immune responses through enhanced immune cell maturation and pro-inflammatory cytokine secretion. Conversely, sustained TLR stimulation may establish a chronic inflammatory microenvironment that promotes leukemic cell persistence and expansion. The heterogeneous TLR expression profiles across distinct immune cell populations further complicate their functional roles in leukemic pathogenesis. Dendritic cell populations and monocytic lineages characteristically exhibit elevated TLR expression levels, functioning as primary effector cells mediating TLR-dependent immune activation. However, dysregulated TLR expression has been documented in various leukemic blast populations, suggesting potential direct regulatory effects on malignant cell behavior independent of immune system modulation.

The clinical applicability of our prognostic model receives additional validation through distinct survival curve separation between high-risk and low-risk patient cohorts. The statistically significant survival differences (p < 0.001) demonstrate effective identification of patients who may derive benefit from intensified therapeutic interventions. Moreover, the reproducible performance achieved across both The Cancer Genome Atlas (TCGA) and Gene Expression Omnibus (GEO) datasets establishes model generalizability across diverse patient populations and experimental methodologies, thereby addressing a prevalent limitation affecting many prognostic algorithms that demonstrate poor external validation performance.

Our comprehensive analysis successfully identified six critical immune regulatory genes (TLR2, TLR4, CCR7, IL18, TIRAP, and FOXP3) demonstrating both differential expression patterns between malignant and normal tissues and significant prognostic utility. The exceptional discriminative capacity of IL18 (AUC = 0.983) warrants particular attention and supports its potential application as an independent biomarker. The identification of these genetic signatures aligns with accumulating evidence supporting the fundamental importance of inflammatory responses and immune dysregulation in leukemic pathogenesis.

The immunological aspects of leukemic diseases have garnered substantial research interest in contemporary hematology-oncology. Unlike solid malignancies where tumor-infiltrating immune cell patterns have undergone extensive characterization, the immune microenvironment in hematological cancers presents distinct investigational challenges due to overlapping anatomical compartments shared between diseased and immune cells within the bone marrow space. Historically, leukemic diseases have been conceptualized primarily as disorders arising from accumulating genetic and epigenetic modifications in hematopoietic stem and progenitor cells. Nevertheless, emerging scientific evidence indicates that immune system dysfunction serves critical roles in both disease initiation and subsequent progression [32–34].

Our single-cell RNA sequencing analysis revealed remarkable heterogeneity within the leukemic microenvironment, with distinct cell clusters showing unique transcriptional profiles. The cell type-specific expression patterns of key immune regulatory genes highlight the complexity of immune responses in leukemia. For instance, the preferential expression of TLR2 in dendritic cells and CCR7 in CD4+ T cells suggests specialized functions within specific immune compartments.

Single-cell RNA sequencing analysis conducted in our study unveiled substantial cellular heterogeneity within the leukemic microenvironment, revealing distinct cellular clusters characterized by unique transcriptional signatures. The cell type-specific expression profiles of key immunoregulatory genes underscore the complex nature of immune responses operating within leukemic tissues. Specifically, the selective expression of TLR2 within dendritic cell populations and CCR7 expression in CD4+ T lymphocytes indicates specialized functional roles operating within discrete immune compartments.

Among the most compelling findings of our investigation is the comprehensive characterization of intercellular communication networks operating within the leukemic microenvironment. CellChat computational analysis revealed intricate signaling architectures connecting diverse immune cell populations, with particular emphasis on interactions involving B lymphocytes, T lymphocytes, natural killer cells, and dendritic cells. These communication networks likely coordinate immune responses within the tumor microenvironment and fundamentally influence disease progression trajectories.

Conclusion

In conclusion, our integrated analysis of leukemia using machine learning and immune profiling has yielded significant insights into disease heterogeneity, prognostic biomarkers, and underlying biological mechanisms. The identified immune signatures and cellular communication networks not only advance our understanding of leukemia pathogenesis but also provide a foundation for developing more personalized diagnostic and therapeutic approaches.

Author contributions

Doujia Chen: Conceptualization, methodology, supervision, project administration, funding acquisition, writing-review and editing. Jie Yang: Data curation, formal analysis, machine learning model development, visualization, writing—original draft. Mengting Wang: Single-cell RNA sequencing data analysis, bioinformatics analysis, statistical analysis, writing—original draft. Tianye Jian: Data collection, validation, literature review, writing—review and editing. All authors have read and agreed to the published version of the manuscript.

Funding

Science and Technology Fund of Health and Family Planning Commission of Guizhou Province (No. gzwjkj2018-1-027), Guizhou Province Science and Technology Achievement Application and Industrialization Plan (Clinical Special Project) (Grant No. Qian-Ke-He-Chengguo LC [2025] General 022).

Data availability

More details could be acquired from the corresponding author upon reasonable request. Single-cell RNA sequencing datasets were also acquired from GEO database resources(GSE131184).

Declarations

Ethics approval and consent to participate

Not available.

Consent for publication

All authors reviewed and approved the final manuscript.

Competing interests

The authors declare no competing interests.