Endothelial-derived PANoptosis factor IL33 is a potential immunotherapy in breast cancer

Summary

Breast cancer (BC) heterogeneity necessitates prognostic and therapeutic biomarkers. PANoptosis remains incompletely understood with regard to its role in BC immunomodulation and clinical outcomes. Here, we integrated transcriptomic data from 7067 patients with BC across multiple cohorts. Patients were stratified via non-negative matrix factorization based on PANoptosis-related gene expression, revealing three distinct clusters. Cluster C3 exhibited the poorest overall survival, characterized by upregulated lipid metabolism and suppressed immunity. A prognostic signature was constructed using a random survival forest model and served as an independent prognostic factor. Single-cell RNA sequencing of 31 tumors identified IL-33-expressing endothelial subclusters (ACKR1⁺ and FBLN5⁺) as key regulators of the tumor immune microenvironment. Reduced IL-33 expression correlated with increased M2 macrophages and CD4⁺ T cell depletion, while sex-specific analysis revealed IL-33 as a predictor of immunotherapy response. Our findings underscore the role of PANoptosis and IL-33⁺ endothelial cells in shaping BC immunity and prognosis.

Graphical abstract

Highlights

• •NMF stratified patients with breast cancer into 3 PANoptosis-related clusters
• •A 15-gene signature was constructed as an independent predictor for breast cancer
• •Core PANoptosis signature IL-33 is mainly expressed in endothelial cells
• •IL-33 correlates with immunotherapy response with a sex-specific dimorphism in efficacy

Immunology; Bioinformatics; Cancer

Introduction

Breast cancer (BC) remains the most prevalent malignancy among women worldwide, contributing significantly to cancer-related mortality, particularly in Asian and African populations.^1,2 Its pronounced molecular heterogeneity—categorized into distinct subtypes (e.g., luminal A/B, HER2-positive, and triple-negative) based on hormone receptor status (ER/PR), HER2 expression, and proliferation markers—underpins variability in clinical behavior, treatment response, and prognosis.³ Despite advances in traditional pathological markers (e.g., tumor size and lymph node status⁴) and multigene assays (e.g., oncotype DX⁵), identifying reliable prognostic biomarkers remains challenging due to tumor heterogeneity, intrinsic resistance mechanisms, and limitations in cost and accessibility of existing tools. Consequently, biomarkers are urgently needed to refine risk stratification and guide personalized therapy.

PANoptosis, a recently defined lytic and inflammatory programmed cell death (PCD) pathway, integrates molecular components from pyroptosis, apoptosis, and necroptosis via multiprotein PANoptosome complexes.⁶ This pathway amplifies immune responses through cytokine release and crosstalk with interferon and JAK-STAT signaling, thereby bridging cell death and immune activation.⁷ While PANoptosis exhibits a “double-edged” role in infections and other cancers (e.g., promoting host defense or tissue damage), its function in BC is poorly characterized.⁸ Preliminary evidence suggests that BC microenvironment cells may exhibit PANoptosis-associated activity linked to immune response, but systematic studies exploring its role in BC immunology, progression, and treatment resistance are lacking.^9,10,11 Especially, a comprehensive map of PANoptosis-driven immune regulatory functions—and their impact on remodeling the tumor microenvironment (TME) through mechanisms such as T cell dysfunction or fibroblast activation—remains unestablished.

To systematically explore whether regulatory interactions exist between PANoptosis and immune networks in BC, and whether characteristic gene expression patterns derived from this axis correlate with clinical outcomes and therapeutic responses. We integrate transcriptome data (bulk/scRNA-seq) from clinical cohorts to identify PANoptosis-associated networks in BC subtypes, define distinct gene expression signatures linked to prognosis, and assess their clinical utility as biomarkers for immune evasion and treatment stratification. This work provides PANoptosis-immune regulatory axes with dual diagnostic and therapeutic value.

Results

PANoptosis gene-based NMF classification unveils distinct breast cancer subtypes with divergent survival

To investigate the contribution of PANoptosis-related genes to BC heterogeneity, we evaluated a custom set of 568 genes implicated in necroptosis, pyroptosis, and apoptosis (Table S2). An initial Cox regression analysis identified 67 PANoptosis-associated genes significantly correlated with overall survival (OS) (p < 0.05; Table S3). Using these genes, we performed non-negative matrix factorization (NMF) clustering on the transcriptomic data from 1026 patients with TCGA-BRCA, segregating the cohort into three distinct molecular clusters based on gene expression patterns (Figures 1, S1A, and Table S4).

Figure 1.

Molecular subtyping of breast cancer based on PANoptosis signatures

(A) Heatmap depicts the three distinct PANoptosis clusters identified by the non-negative matrix factorization (NMF) algorithm in the TCGA-BRCA cohort.

(B) Circos plot shows the distribution of conventional breast cancer molecular subtypes within each PANoptosis cluster.

(C) Sankey diagram illustrates the relationships between PANoptosis clusters, RFS status, OS status, and molecular subtypes.

(D) Kaplan-Meier curve analyzes OS differences among patients across the three PANoptosis clusters.

(E) Kaplan-Meier curve analyzes RFS differences among patients across the three PANoptosis clusters.

(F and G) Boxplots display the significant enrichment scores of selected HALLMARK gene sets in the PANoptosis clusters. Statistical significance was determined by the Wilcoxon test: ∗: p < 0.05, ∗∗: p < 0.01, ∗∗∗: p < 0.001, and ∗∗∗∗: p < 0.0001.

Clusters 1 (C1) and 3 (C3) were predominantly composed of luminal subtypes, whereas cluster 2 (C2) was enriched with estrogen receptor-independent subtypes, particularly TNBC and HER2-positive disease (Figure 1B and Table 1). Although C1 and C3 displayed heterogeneous molecular subtype distributions, C2 and C3 shared similar clinical attributes, including T stage and lymph node metastasis status. Critically, survival analysis revealed that patients in C3 had significantly poorer OS compared to those in C1 and C2, whereas C2 was associated with the most favorable outcomes (Figures 1C–1E). These results suggest that PANoptosis-related gene expression patterns may serve as an independent prognostic indicator beyond conventional molecular subtypes or clinical factors.

Table 1.

Clinical information of each PANoptosis cluster in TCGA-BRCA cohort

	Level	C1	C2	C3	P
n	–	163	260	603	–
subtype (%)	basal	4 (2.5)	94 (36.2)	84 (13.9)	<0.001
	Her2	0 (0.0)	36 (13.8)	38 (6.3)	–
	LumA	136 (83.4)	79 (30.4)	322 (53.4)	–
	LumB	1 (0.6)	40 (15.4)	153 (25.4)	–
	normal	22 (13.5)	11 (4.2)	6 (1.0)	–
T_stage (%)	T1	48 (29.4)	71 (27.3)	160 (26.5)	<0.001
	T2	77 (47.2)	165 (63.5)	371 (61.5)	–
	T3	38 (23.3)	24 (9.2)	72 (11.9)	–
N_stage (%)	LN_negative	68 (42.2)	134 (51.7)	292 (49.5)	0.151
	LN_positive	93 (57.8)	125 (48.3)	298 (50.5)	–
Age (%)	≤35	3 (1.8)	12 (4.6)	18 (3.0)	0.005
	≥60	63 (38.7)	105 (40.4)	306 (50.7)	–
	35∼60	97 (59.5)	143 (55.0)	279 (46.3)	–

GSVA using the HALLMARK gene sets further illuminated the distinct biological behaviors of each cluster (Figure S1B). C1 was markedly enriched in angiogenesis, Hedgehog signaling, and epithelial-mesenchymal transition pathways (Figure S1C). In contrast, C2 demonstrated elevated activity in estrogen response and immune-related pathways (Figure S1D). Cluster 3 was defined by the upregulation of lipid metabolism processes—especially adipogenesis and fatty acid metabolism (Figure 1F)—coupled with reduced apoptotic activity and suppressed immune responses (Figure 1G). Together, these findings highlight a strong connection between PANoptosis-associated gene signatures and key biological processes, influencing both clinical outcomes and potential therapeutic vulnerabilities in BC.

Identifying specific biological pathways associated with each PANoptosis cluster

To further characterize the biological features distinguishing the clusters, we performed DEG analysis, which identified distinct sets of upregulated and downregulated genes specific to C1, C2, and C3 (Figures 2A and S2A). Reactome pathway analysis indicated that C1 was marked by the activation of pathways related to fibril organization, whereas C2 exhibited enrichment in interleukin- and chemokine-mediated immune responses. Notably, C3 displayed a unique dual functional signature: It showed enhanced neuro-regulatory capacity, particularly in neuronal synaptic plasticity, coupled with suppressed immune activity, as reflected by inhibited chemokine signaling pathways (Figures 1B and 1C). Subsequent GSEA using Reactome and Gene Ontology Biological Process (GOBP) gene sets further delineated cluster-specific pathway activations and suppressions (Figures S2B–S2D). Intersection analysis of the significantly enriched pathways across clusters (Figure 2D) reinforced these findings, confirming C1’s association with extracellular matrix reorganization, C2’s link to immune response activation, and C3’s distinct profile—characterized by the upregulation of short-term neuronal synaptic plasticity pathways and concurrent downregulation of cellular responses to chemokines (Figure 2E).

Figure 2.

DEGs and pathway enrichment analysis across PANoptosis clusters in breast cancer

(A) Volcano plot displays DEGs among the PANoptosis clusters. (B) Boxplots showing the top five enriched Reactome pathways for cluster-specific upregulated DEGs.

(C) Boxplots illustrate the top five enriched Reactome pathways for cluster-specific downregulated DEGs.

(D) Dot plot presents GSEA results of enriched GOBP pathways for cluster-specific DEGs.

(E) GSEA plots highlight enriched GOBP pathways for DEGs unique to each PANoptosis cluster.

WGCNA analysis uncovers core DEGs associated with PANoptosis clusters

To identify core genes associated with PANoptosis-derived clusters in BC, we conducted a WGCNA. After removing outlier samples (Figure S3A), a soft thresholding power of 5 was chosen to achieve a scale-free topology fit index (R² = 0.85) (Figure S3B). We analyzed the top 5,000 most variable genes, which were grouped into 11 distinct co-expression modules (Figure 3A). Module-trait relationship analysis demonstrated that the black module was significantly positively correlated with C1, whereas the blue and brown modules showed strong associations with C2 and C3, respectively (Figures 3B and 3C). Accordingly, genes within the black module were upregulated in C1, with analogous expression patterns observed for the brown module in C3 and the blue module in C2 (Figures S3C and S3D).

Figure 3.

WGCNA results linked to PANoptosis clusters

(A) Heatmap of WGCNA results identifies gene modules with high covariance based on PANoptosis-related gene expression.

(B) Heatmap depicts module-trait relationships.

(C) Dot plot shows correlations between module membership and gene significance for the black, blue, and brown modules.

(D) Cytoscape network illustrates relative expression levels, PPI networks, and enriched reactome pathways of module genes.

(E) Venn diagrams represent overlaps between cluster-specific DEGs and WGCNA module genes.

We subsequently constructed PPI networks to examine functional interactions among genes in each module and performed Reactome pathway enrichment analysis. Consistent with prior differential gene expression findings (Figure 1): (1) the black module was enriched in extracellular matrix organization pathways, (2) the blue module was associated with fatty acid metabolism and regulation of the neuronal system, and (3) the brown module was overrepresented in immune response-related pathways (Figure 3D). Further screening of module-specific DEGs identified 143, 321, and 160 genes uniquely correlated with the C1, C2, and C3 clusters, respectively (Figure 3E). These cluster-specific gene signatures may represent critical biomarkers for predicting clinical outcomes in BC and provide insights for developing targeted therapeutic interventions.

Construction of a prognostic model based on PANoptosis cluster-specific DEGs in breast cancer

To assess the prognostic value of DEGs associated with PANoptosis clusters in BC, we analyzed a meta-cohort of 7067 patients (Figure S4A), divided into training and testing sets. An initial Cox regression across individual datasets (TCGA-BRCA, Metabric, SCAN-B, and GEO) identified 54 candidate prognostic genes (Figure S4B). Subsequently, we employed a machine learning-based integrated model, incorporating 10 distinct algorithms, to identify that the random survival forest (RSF) showed the highest C-index in the training cohort (Figure S4C). We then applied the RSF model to construct a prognostic signature, with key predictive factors illustrated in Figure S4D. To refine the model, a second RSF training round was performed using genes with variable importance scores >0.01, leading to a final 15-gene signature termed the PANop Index (Figure 4A and Table S5). Notably, 65.8% of patients in PANoptosis cluster C3 were classified as PANop Index^High (Figure 4B), and conversely, 79.4% of patients with PANop Index^High fell into C3 (Figure 4C and Table 2).

Figure 4.

Construction and validation of the prognostic model

(A) Variable importance plot of the 10 genes selected by the RSF algorithm.

(B) Circos plot shows the distribution of PANop Index^High and PANop Index^Low groups across PANoptosis clusters in the TCGA-BRCA cohort.

(C) Circos plot illustrates the distribution of PANoptosis clusters within PANop Index^High and PANop Index^Low groups.

(D and E) Kaplan-Meier curves for OS and RFS of PANop Index^High and PANop Index^Low groups in the training cohort (D) and testing cohort (E).

(F and G) Forest plot of univariate (F) and multivariate (G) Cox regression analyses for OS prediction using the PANop Index, N stage, T stage, age, and grade (training cohort).

(H) Calibration curves for the nomogram predicting 1-year OS (Training cohort).

(I) DCA compares the nomogram with individual predictors (PANop Index, N stage, age).

(J and K) Kaplan-Meier curves for OS and RFS of Nomogram^High and Nomogram^Low groups in the training cohort (J) and testing cohort.

(L and M) ROC curves evaluate the nomogram’s performance for OS prediction in the training cohort (L) and testing cohort (M).

Table 2.

Clinical information of each PANop Index group in the metadata of breast cancer

	Level	PANop IndexHigh	PANop IndexLow	P
n	–	3533	3534	–
Subtype (%)	basal	734 (20.8)	479 (13.6)	<0.001
	Her2	502 (14.2)	261 (7.4)	–
	LumA	1118 (31.6)	1958 (55.4)	–
	LumB	1044 (29.5)	493 (14.0)	–
	normal	135 (3.8)	343 (9.7)	–
T _Stage (%)	T1	1157 (33.9)	1857 (55.1)	<0.001
	T2	1986 (58.2)	1338 (39.7)	–
	T3	229 (6.7)	169 (5.0)	–
	T4	41 (1.2)	8 (0.2)	–
Lymphnode Metastasis (%)	negative	1810 (53.8)	1994 (59.0)	<0.001
	positive	1553 (46.2)	1388 (41.0)	–
Grade (%)	grade_1	189 (7.0)	503 (18.6)	<0.001
	grade_2	1083 (39.9)	1405 (51.9)	–
	grade_3	1442 (53.1)	800 (29.5)	–
Age (%)	≤35	91 (2.6)	100 (2.9)	<0.001
	≥60	2169 (62.2)	1643 (47.5)	–
	35∼60	1228 (35.2)	1719 (49.7)	–
Death (%)	no	2096 (59.3)	3214 (90.9)	<0.001
	yes	1437 (40.7)	320 (9.1)	–
Recurrence (%)	no	1226 (62.9)	1261 (83.3)	<0.001
	yes	722 (37.1)	252 (16.7)	–
Chemotherapy (%)	no	1751 (62.6)	1833 (64.0)	0.257
	yes	1048 (37.4)	1029 (36.0)	–
Hormonal_therapy (%)	no	489 (36.9)	282 (38.6)	0.475
	yes	837 (63.1)	449 (61.4)	–
Menopausal (%)	post_menopausal	976 (82.7)	487 (71.1)	<0.001
	pre_menopausal	204 (17.3)	198 (28.9)	–

In the training cohort, the PANop Index significantly correlated with both OS and relapse-free survival (RFS) (Figure 2D). When patients were stratified by the median PANop Index, those in the PANop Index^High group showed significantly worse OS and RFS in the testing cohort (Figures 4D, 4E, S4E, and S4F), a trend consistently observed across all separate datasets (Figures S5A and S5B). To rigorously assess generalizability, we applied the PANop Index to an independent external validation cohort. In the GSE86166 cohort, the PANop Index^High group showed significantly poorer OS (Figure S4F) and a trend toward worse RFS (Figure S4G).

Univariate Cox analysis within the training cohort identified the PANop Index, age, T stage, N stage, tumor grade, chemotherapy, and endocrinotherapy as significant predictors of OS (Figure 4F). Multivariate analysis further established the PANop Index, age, T stage, and N stage as independent prognostic factors (Figure 4G). Using these predictors, we built a nomogram to estimate OS probability, which achieved a C-index of 0.843 (95% CI: 0.832–0.855) in the training cohort and 0.737 (95% CI: 0.71–0.762) in the testing cohort. Calibration curves indicated strong agreement between predicted and observed 1-year survival (Figure 4H). Decision curve analysis (DCA) revealed that the PANop Index alone performed comparably to the full nomogram and outperformed other predictors (Figure 4I). Stratification by nomogram-derived risk scores also revealed significant OS differences between high- and low-score groups in both training (Figure 4J) and testing (Figure 4K) cohorts. The nomogram demonstrated high predictive accuracy for 1-, 3-, and 5-year OS in the training cohort (AUC = 0.902, 0.867, 0.866, Figure 4L) and testing cohort (AUC = 0.803, 0.757, 0.714, Figure 4M). Especially, we evaluated the predictive efficacy of the nomogram in an independent cohort GSE21653, and found that the risk is significantly correlated with the poorer RFS (Figure S5C). The nomogram also demonstrated optimal accuracy for 1-, 3-, and 5-year OS in the training cohort (AUC = 0.715, 0.721, 0.755, Figure S5D). In summary, we have developed and validated a cluster-specific prognostic model—the PANop Index—derived from PANoptosis-related DEGs, which robustly predicts survival in a large BC meta-cohort.

IL-33 is the only PANoptosis gene identified in the model and correlated with the activation of anti-tumor immunity

Based on the analysis of epithelial-mesenchymal transition (EMT) and stemness characteristics, cluster C1 demonstrated a higher EMT score but lower stemness levels, whereas C3 exhibited intermediate values for both parameters (Figures 5A and 5B). These observations imply that EMT and stemness may not be the principal determinants of the unfavorable survival outcomes associated with certain clusters. To elucidate the biological underpinnings of the PANop Index and its relationship with PANoptosis clusters, we examined the interplay between the TME, PANoptosis clusters, and the PANop Index. We applied several deconvolution algorithms—including Cibersort, EPIC, ESTIMATE, Immunophenoscore (IPS), MCP-counter, TIMER, and xCell—to profile immune infiltration in the TCGA cohort (Figure 5C). Subsequent Cibersort analysis indicated that cluster C3 was characterized by the reduced infiltration of antitumor immune cells—such as CD4⁺ T cells, CD8⁺ T cells, and M1 macrophages—alongside increased M2 macrophage presence relative to other PANoptosis clusters. In contrast, clusters C1 and C2 displayed an opposing immune cell infiltration profile (Figure 5D). Mantel tests further supported a significant negative correlation between T cells and M2 macrophages, with PANoptosis clusters exerting a substantial influence on the abundance of these immune subsets (Figure 5E).

Figure 5.

Tumor microenvironment characterization by PANoptosis clusters and PANop Index groups

(A and B) Violin and boxplots of EMT scores (A) and stemness scores (B) across PANoptosis clusters.

(C) Heatmap summarizes immune cell infiltration estimates from Cibersort, EPIC, ESTIMATE, IPS, MCP-counter, TIMER, and xCell algorithms (TCGA-BRCA cohort).

(D) Boxplots of Cibersort scores for TME components per PANoptosis cluster (Wilcoxon test).

(E) Heatmap of Mantel test results correlates PANoptosis clusters with Cibersort scores.

(F) Boxplots of Cibersort scores for TME components by PANop Index group (Wilcoxon test).

(G) Venn diagram shows the intersection of PANoptosis genes and RSF-derived genes.

(H) Boxplots of IL-33 expression across PANoptosis clusters (Wilcoxon test).

(I) Boxplots of IL-33 expression by PANop index group (Wilcoxon test).

(J) Boxplots of IL-33 expression by PANop index group within each PANoptosis cluster (Wilcoxon test). ∗: p < 0.05, ∗∗: p < 0.01, ∗∗∗: p < 0.001, and ∗∗∗∗: p < 0.0001.

We next evaluated the relationship between the PANop Index and TME components across all cohorts (TCGA, SCAN-B, and METABRIC). A consistent positive correlation emerged between the PANop Index and M2 macrophage infiltration, whereas a negative correlation was observed with CD4⁺ memory resting T cells (Figure S6A). Moreover, the PANop Index^High group recapitulated the TME features seen in specific PANoptosis clusters, particularly in terms of these immune cell populations (Figure 5F). Differential expression analysis of PANop Index genes between patients with PANop Index^High and PANop Index^Low revealed elevated expression of NXPH4 in the high-risk group, whereas the remaining PANop Index genes were more highly expressed in the low-risk group (Figure S6B). However, expression patterns of these genes varied considerably across individual PANoptosis clusters (Figure S6C).

To identify central PANoptosis-related factors, we screened for genes overlapping between the PANoptosis gene set and the prognostic model. This approach identified IL-33 as the only gene common to both (Figure 5G). IL-33 expression was significantly reduced in PANoptosis cluster C3 (Figure 5H) and in the PANop Index^High group across the entire cohort (Figure 5I). Within clusters C2 and C3 specifically, patients with a high PANop Index also displayed lower IL-33 expression (Figure 5J). Notably, IL-33 levels correlated negatively with M2 macrophages and positively with CD4⁺ memory resting T cells—a pattern opposite to that of the PANop Index (Figure S6D). These results position IL-33 as a key PANoptosis-related gene potentially involved in remodeling the TME and influencing clinical outcomes in BC.

IL-33-expressing endothelial cells are key regulators in TME

To delineate the expression patterns of core PANop Index genes at single-cell resolution, we analyzed the scRNA-seq data from the GSE161529 database. Following rigorous quality control and integration, 202,889 high-quality cells from 31 patients with BC were retained for subsequent analysis. Unsupervised clustering revealed 20 distinct cell clusters across ER-positive, TNBC, and HER2-positive samples (Figure S7A and Table S6). After excluding low-quality clusters, the remaining cells were annotated into ten major lineages—epithelial cells, T cells, NK cells, B cells, plasma cells, myeloid cells, mast cells, endothelial cells, fibroblasts, and perivascular-like (PVL) cells—using well-established marker genes from CellMarker 2.0 (Figure 6A). We visualized the expression of both cell-type-specific markers (Figure S7B) and PANopIndex genes (Figure 6B). Notably, IL-33 expression was predominantly localized to endothelial cells (Figure S7C), corroborating our bulk RNA-seq findings and underscoring the potential contribution of endothelial cells to PANoptosis-associated BC progression.

Figure 6.

scRNA-seq analysis of breast cancer samples

(A) UMAP plot annotates 10 major cell types from 31 patients, with bar plots showing proportions in ER⁺, HER2⁺, and TNBC subtypes.

(B) Dot plot displays the expression levels of PANop Index genes across cell clusters and types.

(C) UMAP visualization of re-clustered endothelial, T/NK, and myeloid cells.

(D–F) UMAP plots link endothelial (D), myeloid (E), and T/NK (F) cell phenotypes to PANoptosis clusters via Scissor analysis (TCGA-BRCA bulk RNA-seq).

(G) Density plot of IL-33 expression in endothelial, T/NK, and myeloid cells.

(H) Density plot of IL-33 expression across endothelial subpopulations.

(I) Boxplots of Scissor analysis results associating cell phenotypes with PANoptosis clusters.

(J) Boxplots of Scissor analysis results associating cell phenotypes with PANop Index groups.

(K) Heatmap of transcriptional clustering across cell types.

Given the strong association between PANoptosis-related genes (including IL-33) and immune infiltration patterns observed in bulk sequencing (Figure 5), we performed deeper subclustering on T/NK cells, myeloid cells, and endothelial cells (Figure 6C and Table S7). This refined analysis identified: (1) seven endothelial subpopulations (ACKR1⁺, CCL21⁺, FBLN5⁺, PGF⁺, SPARC⁺, TIMP4⁺, and TOP2A⁺ endothelial cells; Figure 6D); (2) five myeloid subsets (APOC1⁺ macrophages, APOE⁺ macrophages, C15orf48⁺ dendritic cells, MMP9⁺ macrophages, and TOP2A⁺ macrophages; Figure 6E); and (3) four T/NK subtypes (CXCL13⁺ CD4⁺ T cells, GZMB⁺ NK cells, GZMK⁺ CD8⁺ T cells, and IL7R⁺ CD8⁺ T cells; Figure 6F). Expression patterns of PANop Index genes across these subclusters are displayed in Figure S7D. Consistent with prior observations, IL-33 was markedly enriched in endothelial cells (Figure 6G), with the highest expression detected in ACKR1⁺ and FBLN5⁺ subpopulations (Figure 6H).

We next employed the Scissor algorithm to link single-cell transcriptional profiles with bulk transcriptomic phenotypes from the TCGA cohort, identifying cells whose expression signatures aligned with specific PANoptosis clusters or PANopIndex groups (Figures 6I and 6J). ACKR1⁺ and FBLN5⁺ endothelial cells displayed a transcriptional phenotype associated with non-C3 and PANopIndex^Low status, whereas CCL21⁺ and PGF⁺ endothelial cells were preferentially associated with the C3 cluster and PANopIndex^High phenotype. Transcriptomic similarity analysis further indicated a high degree of correlation between ACKR1⁺ and FBLN5⁺ endothelial cells (Figure 6K), suggesting possible functional synergy in modulating the TME.

IL-33-expressing endothelial cells are potential therapeutic targets for breast cancer

Analysis of the top 50 marker gene-enriched pathways across cell clusters revealed a pronounced enrichment of extracellular matrix (ECM)–related pathways in endothelial and macrophage subclusters, aligning with bulk transcriptome findings (Figure 7A). We further examined ligand-receptor interactions among these cell types and identified several key signaling axes (Figure S8A). Specifically, the MIF and SPP1 pathways emerged as dominant mediators of endothelial-myeloid crosstalk, while FBLN5⁺ endothelial cells were found to regulate T and NK cells primarily through CXCL-mediated signaling (Figures 7B and 7C). In particular, ACKR1⁺ endothelial cells modulated macrophage and T cell activity via the MIF pathway (Figure S8B), and FBLN5⁺ endothelial cells activated the CXCL12–CXCR4 axis in both macrophages and T cells (Figure S8C). Together, these results underscore a critical network of cellular crosstalk among endothelial, T, and myeloid cells, with IL-33–expressing endothelial subclusters—particularly ACKR1⁺ and FBLN5⁺ populations—playing a central role in the regulation of antitumor immunity.

Figure 7.

Role of IL-33 in immunotherapy response

(A) Boxplots of the top 50 marker gene-enriched pathways per cell type.

(B) Ligand-receptor interaction analysis highlights MIF, SPP1, and CXCL pathways among endothelial, T/NK, and myeloid cells.

(C) Heatmap of cell-cell interactions mediated by MIF, SPP1, and CXCL signaling networks.

(D–G) Kaplan-Meier curves for OS of patients with IL-33^High and IL-33^Low cancer (D), all melanoma (E), male melanoma (F), or female patients with melanoma (G) treated with immunotherapy (KM-Plotter database).

To evaluate clinical relevance, we assessed the relationship between IL-33 expression and response to immunotherapy using the KM-Plotter database.^12,13 Notably, elevated IL-33 expression was associated with improved immunotherapy outcomes across multiple cancer types (Figure 7D). A sex-stratified analysis in melanoma uncovered a striking dimorphism: High IL-33 expression in male patients correlated with poorer treatment efficacy, whereas in females, it predicted optimal immunotherapy response (Figures 7E–7G). This sexual dimorphism highlights the importance of considering sex-specific mechanisms in IL-33–mediated immune regulation and therapeutic development.

Discussion

The integration of transcriptome data in this study reveals a paradigm in BC immunotherapy, centered on the PANoptosis pathway and its regulatory role in the TME. Our findings demonstrate that endothelial cell-derived IL-33 serves as a critical orchestrator of anti-tumor immunity through PANoptosis modulation, with significant prognostic and therapeutic implications.

The identification of three distinct patient clusters based on PANoptosis-related gene expression patterns provides unprecedented stratification capabilities. Cluster C3, characterized by suppressed immunity and lipid metabolic reprogramming (Figures 2B and 2C), exhibited the poorest survival outcomes (Figures 1D and 1E). The distinct lipid metabolic reprogramming observed in the C3 cluster, coupled with its immunosuppressive TME (characterized by enriched M2 macrophages and depleted CD4⁺ T cells), presents a compelling avenue for mechanistic exploration. Accumulating evidence suggests that altered lipid metabolism in cancer cells is not merely a bioenergetic adaptation but also an active regulator of immune evasion. For instance, lipid accumulation in the TME can promote the differentiation and recruitment of immunosuppressive M2 macrophages and impair the anti-tumor function of T cells.^14,15,16 While our study establishes a strong association between this metabolic-immune phenotype and poor prognosis through the PANop Index. Importantly, emerging evidence linking metabolic dysregulation to immune evasion, though the direct involvement of PANoptosis in this process represents a rising finding.^17,18 The prognostic PANop Index, derived from random forest modeling, demonstrated robust predictive power, suggesting its potential utility in clinical risk stratification. Several works have generated predictive models based on PANoptosis-related genes in patients with as gastric,¹⁹ ovarian²⁰ or lung cancer.²¹ Especially, our model presented better predictive efficacy in a meta-data compared to other works in BC,^9,10,11 suggesting an optimal clinical application.

Our discovery of IL-33 as the sole prognostic PANoptosis gene co-occurring in the signature challenges existing paradigms. While previous studies reported context-dependent pro-tumorigenic roles in gastric,²² pancreatic,²³ and BCs,²⁴ focusing on the pre-metastatic role of IL-33, our data reveal protective effects in BC. Mechanistically, scRNA-seq analysis localized IL-33 expression to ACKR1/FBLN5 endothelial subpopulations, which appear to modulate immune cell trafficking via MIF-CXCR4 and CXCL12 signaling. The observed inverse correlation between IL-33 and M2 macrophage infiltration suggests a previously unrecognized role in TME polarization. However, Shani et al.²⁴ reported a pro-tumorigenic and pro-metastatic role of fibroblast-derived IL-33 in BC lung metastases. Importantly, a recent comprehensive review consolidates evidence that IL-33 within the TME can be produced by a variety of cell types, including stromal fibroblasts and endothelial cells, underscoring the plasticity and context-dependence of its cellular sources.²⁵ Furthermore, IL-33 derived from eosinophils enhances the efficacy of immunotherapy in metastatic BC, which is consistent with our hypothesis that endothelial-derived IL-33 may also play the pivotal role in the immunotherapy of BC.²⁶ Therefore, our findings do not contradict but rather complement prior work, collectively painting a more complex picture where IL-33’s cellular origin and function are dynamically shaped by disease stage and the local microenvironment. While the definitive role and spatial distribution of IL-33-expressing endothelial cells warrant further validation using spatial transcriptomics or multiplex immunohistochemistry in future studies.

Especially, the striking sex dimorphism in IL-33-mediated immunotherapy response warrants particular attention. Female patients with high IL-33 expression showed significantly improved response to immunotherapy, while males exhibited worse outcomes. Some studies have found the relationship between estrogen and IL-33 in BC. For instance, IL-33 induces endocrine resistance of BC by promoting cancer stem cell properties.²⁷ Higher expression of IL-33 can be detected in TNBC compared to estrogen-dependent BC.²⁸ Importantly, a meta-analysis uncovered that immune checkpoint inhibitors can improve OS for patients with advanced cancers, but the magnitude of benefit is sex-dependent.²⁹ These findings emphasize the necessity for estrogen- or sex-stratified analysis in immunotherapy trials and suggest potential hormonal modulation strategies based on IL-33.

However, our exploratory analysis on the association between IL-33 expression and response to immune checkpoint blockade relied on immunotherapy cohorts from the KM-Plotter database, which are largely comprised of patients with melanoma and non-small cell lung cancer. This represents an important limitation, as the tumor immune microenvironment and mechanisms of therapeutic response can differ substantially across cancer types. Therefore, the observed association and the intriguing sex-specific dimorphism, while providing a compelling biological hypothesis, require direct validation in dedicated BC immunotherapy cohorts with available transcriptomic and clinical outcome data. Nevertheless, this analysis serves as a valuable proof-of-concept, demonstrating the potential translational relevance of IL-33 in cancer immunotherapy and highlighting a direction for future breast-cancer-specific investigations.

Conclusively, our study reveals an association between PANoptosis-related molecular patterns and BC immunotherapy response, and highlights the potential regulatory role of IL-33-expressing endothelial cells within this context. The integration of bulk and single-cell transcriptome analysis with clinical outcomes provides a framework for identifying targets and patient stratification strategies in precision oncology.

Limitations of the study

Several limitations of our work merit consideration. The retrospective design introduces potential confounding, necessitating prospective validation. While transcriptomic analysis provides cellular resolution, future studies should incorporate spatial transcriptomics to map IL-33 endothelial-immune cell interactions within the TME architecture. Additionally, functional validation using endothelial-specific IL-33 knockout models would confirm causality.

Resource availability

Lead contact

Requests for further information and resources should be directed to and will be fulfilled by the lead contact, Yingpu Li (liyingpu@hrbmu.edu.cn).

Materials availability

This study did not generate new unique reagents.

Data and code availability

Data

Processed data have been uploaded as Supplementary Processing Data S1, and clinical information has been uploaded as Supplementary Processing Data S2 in this work. All raw transcriptomic data are available in public databases under accession numbers listed in the STAR Methods.

Code

This paper does not report original code.

Other items

Any additional information required to reanalyze the data reported in this paper is available from the lead contact upon request.

Acknowledgments

We thank the Cancer Genome Atlas (TCGA), the Molecular Taxonomy of Breast Cancer International Consortium (METABRIC), the Sweden Cancerome Analysis Network - Breast (SCAN-B) initiative, and the Gene Expression Omnibus (GEO) for providing the transcriptomic and clinical data. This work was supported by the following grants: Natural Science Foundation of Heilongjiang Province (PL2025H178) and Basic Scientific Research Business Expenses of Colleges and Universities in Heilongjiang Province (31031240062 and 31031250057).

Author contributions

Y.L., ZG.L., and S.Z. conceptualized the research. Y.Q. conducted the sample curation and data analysis. ZH.L. and L.J. supported dataset curation. Y.L., Y.Q., and S.Z. contributed to the analysis methods and interpretation of the data. ZG.L., L.J., and S.Y. helped with manuscript preparation and data discussion. writing – original draft, Y.Q. and Y.L. All co-authors participated in revising the manuscript, read and approved the final manuscript for publication. funding acquisition Y.L. and S.Z.

Declaration of interests

The authors declare no competing interests.

Declaration of generative AI and AI-assisted technologies in the writing process

During the preparation of this work, the authors used AI-assisted technologies (DeepSeek) in order to improve the readability and language of the manuscript. After using this tool, the authors reviewed and edited the content as needed and take full responsibility for the content of the published article.

STAR★Methods

Key resources table

REAGENT or RESOURCE	SOURCE	IDENTIFIER
Deposited data
TCGA-BRCA RNA-seq data	The Cancer Genome Atlas	https://portal.gdc.cancer.gov/
METABRIC microarray data	Molecular Taxonomy of Breast Cancer International Consortium	https://www.cbioportal.org/study/summary?id=brca_metabric
SCAN-B RNA-seq data	Sweden Cancerome Analysis Network - Breast	https://www.ncbi.nlm.nih.gov/geo/query/acc.cgi?acc=GSE96058
GEO microarray dataset (GSE20685, GSE58644, GSE58812, GSE88770)	Gene Expression Omnibus	https://www.ncbi.nlm.nih.gov/geo/
External validation dataset (GSE21653, GSE86166)	Gene Expression Omnibus	https://www.ncbi.nlm.nih.gov/geo/
Single-cell RNA-seq data	Gene Expression Omnibus	https://www.ncbi.nlm.nih.gov/geo/query/acc.cgi?acc=GSE161529
Software and algorithms
R(v4.3)	The R Foundation	https://www.r-project.org/
IOBR (v0.99.8)	Zeng et al.30	https://github.com/IOBR/IOBR
NMF (v0.28)	Gaujoux and Seoighe31	https://cran.r-project.org/package=NMF
DESeq2(v1.42.1)	Love et al.32	https://bioconductor.org/packages/DESeq2
clusterProfiler(v4.12.3)	Wu et al.33	https://bioconductor.org/packages/clusterProfiler
WGCNA (v1.73)	Langfelder and Horvath34	https://cran.r-project.org/package=WGCNA
Mime1(version 0.0.0.9000)	Liu et al.35	https://github.com/l-magnificence/Mime1
randomForestSRC (version 3.3.1)	Ishwaran and Kogalur	https://cran.r-project.org/package=randomForestSRC
survival(version 3.7–0)	Terry Therneau	https://cran.r-project.org/package=survival
survminer (version 0.4.9)	Kassambara et al.	https://cran.r-project.org/package=survminer
rms (version 6.8–1)	Frank Harrell	https://cran.r-project.org/package=rms
Seurat(v4.4)	Hao et al.36	https://satijalab.org/seurat/
Harmony (v1.2.1)	Korsunsky et al.37	https://github.com/immunogenomics/harmony
CellChat(v2.1.2)	Jin et al.38	https://github.com/sqjin/CellChat
Scissor(v2.0)	Sun et al.39	https://github.com/sunjunqiang/Scissor
Other
CellMarker 2.0 Database	Hu et al.40	https://bio-bigdata.hrbmu.edu.cn/CellMarker/
KM-Plotter Database	Kovacs et al.41	https://kmplot.com/analysis/
Cytoscape (v3.10)	Cytoscape Consortium	https://cytoscape.org/
STRING (v12.0)	STRING Consortium	https://string-db.org/

Experimental model and study participant details

This study is a retrospective bioinformatics analysis based on publicly available transcriptomic datasets. No experimental models (cell lines, animals, plants, or microbes), human participants, or primary cultures were generated or used in this work. All data were obtained from public repositories with approved ethical approval and informed consent from the original published studies. No cell authentication or mycoplasma testing was applicable. Sex-specific analysis was performed in the results section to evaluate the association between sex and immunotherapy response, with relevant results reported in the main text.

Method details

Bulk transcriptome data collection and integration

RNA-seq data were compiled from TCGA, METABRIC, and SCAN_B⁴² databases. Further microarray data from GEO databases include GSE20685,⁴³ GSE58644,⁴⁴ GSE58812,⁴⁵ and GSE88770⁴⁶ were used to integrate with RNA-seq data for generating a meta-cohort of 7067 breast cancer patients (as detailed in^47,48). The database inclusion required: (i) sample size >100 breast cancer patients, (ii) availability of full clinical annotations (e.g., survival status, subtype), and (iii) public accessibility with raw data, enough gene number recognized (e.g., GPL570 platform for microarray data). We excluded patients with incomplete clinical data or non-primary tumors. The summary of clinical information of the meta-cohort was listed in Table S1. External databases including GSE21653 (n = 252)⁴⁹ and GSE86166 (n = 366)⁵⁰ were used for independent validation. Bioinformatic analysis was performed using R. To address batch effects, we applied the ComBat function from the “sva” R package.⁵¹ Transcript per million (TPM) normalization and principal component analysis (PCA) were subsequently conducted using the “IOBR” package.³⁰

Non-negative matrix factorization (NMF) analysis

PANoptosis genes were curated from established sources^52,53 (detailed in Table S2). To explore the association of these PANoptosis genes with breast cancer patient outcomes, we performed dimensionality reduction using non-negative matrix factorization (NMF) with the “NMF” R package.³¹ The optimal number of clusters (k) was determined by jointly maximizing the cophenetic correlation coefficient and silhouette score.

Differential expression and pathway analysis in TCGA-BRCA

Differentially expressed genes (DEGs) within the TCGA-BRCA cohort were identified using the “DESeq2” R package.³² Genes were classified as DEGs if they met the thresholds of an adjusted p-value <0.05 and |log₂ fold change (FC)| > 1. To functionally characterize these DEGs, we performed pathway analysis using: (i) Molecular Signatures Database (MSigDB) v7.5.1: Functional enrichment across HALLMARK and other curated pathways,⁵⁴ (ii) ReactomePA R package (v1.46.0): Pathway analysis based on Reactome annotations.⁵⁵ Additionally, Gene Set Enrichment Analysis (GSEA) and Gene Set Variation Analysis (GSVA) were conducted using the “clusterProfiler” package³³ for comprehensive pathway and gene set exploration.

Weighted gene co-expression network analysis (WGCNA) of PANoptosis clusters in breast cancer patients

The “WGCNA” R package was employed to construct a gene co-expression network using the top 5,000 most variably expressed genes.³⁴ Hierarchical clustering analysis identified gene modules, with parameters set to a minimum module size of 30 and a merge threshold height of 0.25. We then assessed network interactions by calculating gene significance (GS) and module membership (MM), and evaluated module-trait correlations with clinical characteristics. Genes exhibiting high GS (|GS| > 0.8) and MM (|MM| > 0.8) within target modules were identified as key candidates for further analysis. Finally, the “Pheatmap” package (v1.0.12) visualized associations between modules and PANoptosis clusters.

Prognostic model validation and clinical utility assessment

A cohort of 7067 breast cancer patients with complete outcome data was randomly partitioned into Training (70%) and Testing (30%) cohorts. Genes with the p value < 0.05 for Cox regression survival analysis were selected for further prognostic model construction. To select the most accurate prognostic model in the Training cohort, we utilized the R package “Mime1”, which includes 10 individual machine-learning algorithms.³⁵ The random survival forest (RSF) algorithm was selected to develop the final prognostic model.

An initial RSF model was fitted using all features in the Training cohort. Variable importance (VIMP) measures were calculated for each feature (gene). Features were then ranked in descending order of importance to generate an ordered list. A nested cross-validation (CV) procedure was employed to obtain a robust performance estimate and optimize hyperparameters. After completing the five rounds of nested CV, the parameter set associated with the highest validation C-index was selected as the final optimal hyperparameters. The resultant risk score, termed the PANop Index, predicted individual overall survival (OS). All patients were stratified into PANopIndex^High and PANopIndex^Low subgroups based on cohort-spPANop Indexfic median risk scores. Kaplan-Meier curves were generated to visually compare survival outcomes between these groups, utilizing the “survival” and “survminer” packages.

Univariate and multivariate Cox regression analyses were performed using the “forestmodel” package to evaluate the association between the PANop Index and key clinicopathological variables (age, T stage, N stage, tumor grade), identifying independent predictors of overall survival (OS). Based on these predictors, a prognostic nomogram was constructed via the “rms” package to quantify individual OS probability, with calibration curves visually validating its predictive accuracy against observed outcomes. Decision curve analysis (DCA) implemented by “ggDCA” further demonstrated the nomogram’s clinical net benefit across risk thresholds. Model performance was rigorously assessed through time-dependent receiver operating characteristic (ROC) analysis, reporting area under the curve (AUC) values and diagnostic accuracy metrics. Finally, the Kaplan-Meier Plotter (kmplot.com) evaluated survival differences in immunotherapy cohorts to explore the PANop Index’s therapeutic relevance.⁴¹

Protein-protein interaction (PPI) and TME analysis

PPI analysis of PANop Index genes was performed using the STRING database (v12.0, available at https://string-db.org/). Interacting proteins within each cluster were subjected to Reactome pathway enrichment analysis, with the top five enriched pathways selected for visualization. The PPI network was constructed and visualized via Cytoscape, highlighting hub gene significance and pathway associations. For TME component quantification, the infiltration scores of stromal/immune cells were calculated using the “IOBR” R package.³⁰ This tool integrates seven established algorithms—CIBERSORT, EPIC, ESTIMATE, IPS, MCP-counter, TIMER, and xCell—to comprehensively profile TME heterogeneity.

Correlation and mantel testing analysis

Spearman correlation analysis between individual TME components and the PANoptosis Index (PANop Index) was performed via Spearman’s rank correlation analysis. Results were visualized using the corrplot and ggcorrplot R packages. Correlations with |ρ| > 0.3 and p < 0.05∗∗ were defined as statistically significant, with ρ denoting the Spearman correlation coefficient. Mantel test analysis was conducted using the linkET package to evaluate matrix-level associations (e.g., between TME heterogeneity and PANop Index spatial patterns). The Mantel statistic (r) and significance were reported to quantify dependency structures.

Single-cell RNA sequencing data integration and analysis

Single-cell RNA sequencing (scRNA-seq) data from 31 primary breast tumor samples (GSE161529) were retrieved from the GEO database. The Seurat R package was utilized for data integration and preprocessing.³⁶ Cells with a gene expression count below 200 or exceeding 5000 were excluded from the analysis. Furthermore, cells with a mitochondrial content greater than 20% were also filtered out. After quality control, 210,293 high-confidence cells were retained for downstream analysis. Batch effect correction was performed using the Harmony algorithm to mitigate technical variations across samples.³⁷ Uniform Manifold Approximation and Projection (UMAP) was applied for dimensionality reduction and visualization of cell clusters in two-dimensional space. Cell type annotation was executed by matching gene expression signatures against the CellMarker 2.0 database.⁴⁰ To investigate intercellular communication, ligand-receptor interactions between annotated clusters were quantified via the CellChat package.³⁸ Finally, the Scissor algorithm was employed to integrate bulk RNA-seq profiles with scRNA-seq data, enabling the identification of phenotype-associated cell subpopulations.³⁹

Quantification and statistical analysis

All statistical analyses were performed using R software (v4.3). Comparisons between two groups were conducted using the Wilcoxon rank-sum test. Survival curves were plotted using the Kaplan–Meier method and compared via the log rank test. Univariate and multivariate Cox regression analyses were used to identify independent prognostic factors for overall survival. Correlations between variables were evaluated using Spearman rank correlation analysis. The Mantel test was used to assess matrix-level associations between TME components and PANoptosis clusters.

Statistical significance was defined as ∗p < 0.05, ∗∗p < 0.01, ∗∗∗p < 0.001, ∗∗∗∗p < 0.0001. The n value represents the number of breast cancer patients or single cells, as specified in the corresponding figure legends. Center values represent median or mean, and dispersion values represent standard deviation (SD) where applicable. All statistical details, including sample sizes, statistical tests, and significance thresholds, are provided in the figure legends or main text.

Published: April 15, 2026