Fig 2: Marker genes, pathway enrichment, and key features of tumor subtypes, particularly the inflammatory subpopulation. (A) Bubble plot illustrating the expression patterns of subtype-specific marker genes across different tumor subpopulations. (B) Dot plot summarizing the results of Gene GO and KEGG pathway enrichment analyses for each tumor subcluster. (C) GSEA plot showing a significant enrichment of the “Chemokine signaling pathway” in the inflammatory tumor subpopulation. (D) Bar graphs depicting KEGG, Molecular Function (MF), Cellular Component (CC), and Biological Process (BP) enrichment analyses for the inflammatory tumor subpopulation, highlighting key immune-related pathways such as leukocyte transendothelial migration. (E) Volcano plot of differentially expressed genes between inflammatory and non-inflammatory tumor subpopulations, with CCL20 prominently labeled as a significantly upregulated gene.

Fig 3: Cell-cell communication landscape between tumor subclusters and T cell subsets via cell chat analysis. (A) Cell communication network plots illustrating the overall interaction strength (left) and the number of interactions (right) between tumor subclusters and T cell subsets. Thicker and darker lines represent stronger communication. (B) Heatmap of interaction strength between different tumor subclusters as signal senders (Source) and T cell subsets as receivers (Target). Deeper red indicates stronger interaction, highlighting the potent signaling from the inflammatory tumor subpopulation (Tumor_Inflammatory) to Tregs. (C) Bubble plot of significant ligand-receptor pairs. This plot shows the molecular pairs significantly enriched in signals sent from the “Tumor_Inflammatory” subpopulation to various T cell subsets. Dot size represents the significance of the P-value, and color indicates the communication probability. The CCL20-CCR6 interaction targeting Tregs is highlighted in red. (D) Circos plot (circle plot) specifically depicting the CCL20-CCR6 signaling pathway. The arrow direction indicates signal flow, and the line width represents signal strength. The results clearly show that CCL20 signals are predominantly emitted by inflammatory and EMT-type tumor cells and specifically target Tregs.

Fig 4: Trajectory inference and transcriptional regulatory network in tumor cell evolution. (A) Pseudotime trajectory of tumor cells constructed using the Monocle2 algorithm. The color gradient from dark blue to dark red represents the pseudotime progression, indicating the evolutionary direction from an early state toward a late, more malignant state. (B) The same pseudotime trajectory map annotated with tumor subpopulations. The results show that the Tumor_AT1/2_like subpopulation is predominantly located at the start of the trajectory, while the Tumor_Inflammatory and Tumor_EMT subpopulations are primarily distributed toward the trajectory end. (C) Expression dynamics of key genes (CCL20, TIMP1, C3, FN1) along the pseudotime trajectory. The fitted black curves illustrate the trend of gene expression over the evolutionary process, demonstrating a significant upregulation of CCL20 toward the late stage. (D) Violin plots showing the expression levels of key transcription factors (NFKB1, STAT1, RELA, IRF1) across tumor subpopulations, revealing their specific high expression in the Tumor_Inflammatory subpopulation. (E) Heatmap of regulon activity from SCENIC analysis. The columns represent individual cells, and rows represent transcription factors. The red color indicates high regulon activity, uncovering that NF-κB and STAT family transcription factors dominate the regulatory network in the inflammatory tumor subpopulation.

Fig 5: Development and external validation of a prognostic model Based on an ensemble of machine learning algorithms. (A) Heatmap of Algorithm Performance Screening. The heatmap displays the predictive performance (C-index) of dozens of algorithm combinations (Lasso, Random Forest, SVM, GBM) in the training set and multiple external validation cohorts (GSE75037, GSE10072, GSE72094, GSE118370). Darker colors (brown/yellow) indicate higher model prediction accuracy. (B) Feature Coefficients in the Optimal Model (Rf + Gbm). A bar chart shows the feature genes incorporated into the optimal model and their corresponding coefficients. Blue bars represent risk genes (coefficient > 0), and orange bars represent protective genes (coefficient < 0). Notably, CCL20 exhibits a high positive coefficient, identifying it as a primary risk-contributing factor. (C) Kaplan-Meier Survival Analysis Based on the Model-Derived Risk Score. The survival curve demonstrates that patients in the high-risk group have a significantly worse overall survival rate compared to those in the low-risk group (Hazard Ratio [HR] = 2.01, 95% Confidence Interval [CI]: 1.49-2.71, P < 0.001).