SAP18 drives vasculogenic mimicry in esophageal squamous cell carcinoma: a machine learning and multi-omics investigation

Abstract

Vasculogenic mimicry (VM), a novel endothelial-independent blood perfusion pathway, is linked to advanced stage and poor prognosis in esophageal squamous cell carcinoma (ESCC). In this study, by integrating single-cell RNA sequencing and transcriptomic data and employing a machine learning framework incorporating 117 algorithmic combinations, we constructed a robust 4-VM-related gene prognostic model for ESCC. Consensus clustering further stratified patients into two subtypes. The high-risk subtype (C2) was characterized by unfavorable prognosis, activated stroma, enrichment of M2 macrophages, and multidrug resistance. As the core regulatory hub of this model, SAP18 was markedly upregulated in ESCC tissues and showed positive correlations with aggressive clinicopathological features. Mechanistically, SAP18 binds to PIK3CB, activating the AKT/mTOR signaling cascade and upregulating HIF-1α, thereby conferring VM-forming capability to epithelial-derived tumor cells. Both in vitro and in vivo experiments confirmed that knockdown of SAP18 significantly suppressed malignant phenotypes in ESCC. Pharmacological intervention using the highly selective AKT inhibitor MK-2206 effectively abolished VM network formation and profoundly inhibited tumor growth. Our integrated multi-omics and functional analyses decipher the molecular architecture of VM in ESCC, nominating SAP18 as a precise prognostic biomarker and therapeutic target, and providing a foundation for individualized VM-targeted strategies in ESCC management.

Introduction

Esophageal cancer (EC) remains a major global health burden and is one of the leading causes of cancer death worldwide¹. Esophageal squamous cell carcinoma (ESCC) is the most common type, especially in East Asia^2,3. Although treatment has improved through chemoradiation, targeted agents, and immunotherapy, the survival rate for patients with advanced or recurrent disease is still very poor⁴. This limited progress reflects the deep biological diversity within ESCC, which makes it difficult to predict outcomes or guide therapy⁵. There is an urgent need to define precise molecular subgroups and identify reliable biomarkers that can guide personalized treatment decisions^6,7.

Formation of new blood vessels is a well-known feature of cancer and plays a key role in tumor growth and metastasis^8,9. However, drugs that block traditional endothelial blood vessels have shown only temporary and limited benefit in ESCC^10,11. This has led to growing interest in an alternative process called vasculogenic mimicry (VM). VM refers to the ability of tumor cells themselves, rather than endothelial cells, to form vessel-like structures that supply nutrients and promote tumor spread^12–15. The presence of VM has been linked to aggressive behavior and poor outcomes in many cancers. In ESCC, evidence is still emerging, with a recent study identifying the PLEKHA1-TACC2 fusion gene as a possible driver of VM¹⁶. Yet, the biological mechanisms behind VM and its clinical relevance in ESCC are still unclear. Understanding how VM develops and how it affects disease progression could open new directions for therapy.

Modern sequencing and data analysis have transformed how cancer biology can be studied. The combination of bulk RNA sequencing and single-cell RNA sequencing (scRNA-seq) allows researchers to capture both the broad landscape and the fine details of tumor biology^17,18. Machine learning methods can then be applied to integrate these data and identify genes that predict survival or treatment response. Such integrative approaches have already produced useful clinical models in several cancer types^19,20. Applying similar methods to ESCC may help uncover new molecular subtypes and targets for therapy.

In this study, we first analyzed ESCC specimens and found a close association between VM number and patient prognosis. Then, we analyzed ESCC using both bulk and single-cell transcriptomic data from multiple independent datasets. We started with a group of known VM-related genes and built a prognostic model using a comprehensive set of machine learning algorithms. The final model, composed of 4 genes, was able to divide patients into groups with clearly different outcomes. We also found two major molecular subtypes of ESCC with distinct immune and metabolic features. One subtype showed poorer survival, higher stromal content, more M2 macrophage infiltration, and lower predicted response to targeted drugs.

Among the genes identified, SAP18 (Sin3-associated protein 18 kDa) emerged as a key regulator of VM and a strong predictor of poor prognosis. SAP18 is part of the Sin3-HDAC1 transcriptional repression complex and has been linked to gene regulation and chromatin remodeling^21,22, but its role in cancer development is still not well understood. Through both computational and experimental approaches, including cell line studies, xenograft models, zebrafish models, and patient-derived organoids, we found that SAP18 promotes VM formation, tumor growth, and metastasis. Further analysis suggested that SAP18 may drive these effects through the PI3K/AKT/mTOR/HIF-1α signaling pathway. These results suggest that SAP18 could serve as a clinically relevant biomarker and potential therapeutic target in ESCC.

Together, our findings outline the molecular basis of VM in ESCC and highlight SAP18 as a link between metabolic activity, immune suppression, and VM. This integrated framework provides a step toward applying VM biology to precision oncology and offers a path for developing targeted treatment strategies for ESCC.

Results

Association of VM with prognosis and clinical characteristics in patients with ESCC

Figure 1 shows the main analysis process of this study. Since its initial description, the clinical significance and formation mechanisms of VM have been a major focus of scientific investigation. To investigate the pathological characteristics and potential significance of VM in ESCC clinical samples, this study collected 150 ESCC clinical tissue samples from the First Affiliated Hospital of Zhengzhou University between 2012 and 2019. Through the observation of the above specimens, we found that the amount of VM was significantly associated with the stage and prognosis of ESCC patients. Patients with a higher TNM stage often had a higher number of VM, and their survival prognosis was also poorer (Fig. 2a–c). We next conducted preliminary observations to explore whether ESCC organoids might exhibit phenomena related to VM. IHC analysis (CD34/PAS) revealed the presence of PAS-positive channels within poorly differentiated ESCC organoids. These channels appeared devoid of CD34-positive endothelial cells, suggesting the potential presence of putative VM-like structures formed by the tumor cells (Fig. 2d). The tube formation assay demonstrated that most of the tested ESCC cell lines (KYSE150, KYSE510, KYSE30, KYSE410) exhibited a significant capacity to form tubular structures. In contrast, KYSE450 showed a comparatively weaker tube-forming ability than the other cell lines, but nevertheless retained this capability. Representative results of tube formation under a standard light microscope and following Calcein-AM staining observed under a fluorescence microscope are presented in Fig. 2e and f, respectively.

Fig. 1.

Flowchart of this study.

Fig. 2. Analysis of clinical specimens of ESCC.

a, b The relationship between the number of VM and tumor stage. c The relationship between the number of VM and patient survival. d Observation of tube formation in patient-derived ESCC organoids, red arrows indicate suspected VM structures. e Results of tube formation assay in ESCC cell lines visualized by light microscopy. f Results of tube formation assay in ESCC cell lines visualized by fluorescence microscopy (Calcein-AM staining). Red arrows indicate VM channels. Data are presented as mean ± SD. ***P < 0.001.

Identifying cell types through scRNA-seq analysis

We performed single-cell analysis on 11 tumor samples and divided all cells into 18 clusters (0–17) (Fig. 3a). Meanwhile, we analyzed the specifically expressed genes of each cell cluster and annotated the cell types based on these genes. Based on canonical cell-type markers, these clusters were annotated as 11 major cell types: T/NK cells (n = 12412), Monocytes/Macrophages (n = 7007), Epithelial cells (n = 4405), B cells (n = 4037), Fibroblasts (n = 3322), Endothelial cells (n = 1802), Neutrophils (n = 1324), Proliferative (n = 812), Mast cells (n = 700), Plasma (n = 698), Pericytes (n = 663) (Fig. 3b). The stacked violin plot shows the most representative marker genes for each cell type (Fig. 3c). The dot plot shows the top 5 specifically expressed genes of each cell (Fig. S1a). The number and relative proportions of each cell type across samples are shown in Figs. 3d and S1b. In addition, the specific expression genes of each cell type were also demonstrated (Fig. 3e). Notably, as illustrated in Fig. 3e, S100A family genes (such as S100A8 and S100A9) were significantly upregulated specifically within the epithelial cell cluster. This pattern aligns with the classical biological lineage characteristics of ESCC, further validating the accuracy of our epithelial cell annotation^23,24. In this context, epithelial cells were identified as the malignant tumor population based on their characteristics of esophageal squamous origin, and they served as the core malignant component used for subsequent VM analysis. By intersecting epithelial cell markers with 191 VM-related genes (VMRGs), we obtained 7 single-cell VMRGs (scVMRGs) (HSPB1, CLDN4, IGFBP2, SDC1, KRT14, SLC2A1, SERPINB5) (Fig. 3f).

Fig. 3. Integration, clustering, and analysis of scRNA-Seq data.

a t-SNE plot showing unsupervised clustering of all cells (n = 11), resulting in 18 distinct clusters (0–17). b t-SNE plot displaying annotated cell types based on canonical marker genes: T/NK, Monocytes/Macrophages, Epithelial cells, B cells, Fibroblasts, Endothelial cells, Neutrophils, Proliferative, Mast cells, Plasma, Pericytes. c Expression of classical markers in different cell types. d The proportion of different cell types in each ESCC scRNA data. e The specifically expressed genes of each cell. f Venn diagram of Epithelial cell marker genes and 191 VMRGs. g Results of the subgroup analysis of epithelial cells: a total of 5 subgroups were identified. h t-SNE plot of expression level of scVMRGs in Epithelial cells by using the “AUCell” function. i Expression differences of scVMEGs among various epithelial cell subsets. j Based on the expression of scVMEGs, epithelial cells were divided into high-VM and low-VM groups. k The volcano plot shows the DEGs between normal tissues and tumor tissues. l The intersection of high-VM group-specific genes and DEGs.

Identification of VMRGs differentially expressed in ESCC

We stratified the epithelial cells into distinct subgroups, identifying 5 clusters (Fig. 3g). Given the established role of VM in tumor progression, we next applied the AUCell algorithm to quantify VM activity in individual epithelial cells based on the expression of these 7 scVMRGs (Fig. 3h). The violin plots depicted the expression levels of the scVMRGs across different epithelial cell subpopulations, revealing a relatively higher expression in the EP4 subpopulation (Fig. 3i). Based on the expression level, we classified the epithelial cells into a high-VM group and a low-VM group, and extracted the specifically expressed genes in the high-VM group (Fig. 3j). Finally, we identified 1577 genes specifically upregulated in the high-VM group for all subsequent functional analyses. Meanwhile, we conducted enrichment analysis on these specifically expressed genes. The results of the enrichment analysis indicated that they were related to the carcinogenic pathway of reactive oxygen species (Fig. S1c). As shown in the volcano plot (Fig. 3k), a total of 2302 significantly differentially expressed genes (DEGs) were identified through the comparison between ESCC and normal tissues. Intersection of DEGs with high-VM group feature genes identified 181 genes for downstream analysis (Fig. 3l).

Development and evaluation of the VMRGs prediction model

Using the mime package, we incorporated the aforementioned 181 genes and constructed 117 predictive models, with GSE53624 serving as the training set and GSE53622 as the validation set. For each model, the C-index was calculated for both the training data and the validation sample, and the results are shown in Fig. 4a. The optimal integrated model, combining StepCox(both) with Generalized Boosted Regression Models (GBM), demonstrated the highest average C-index of 0.695 during cross-validation. Among all the combinations of machine learning, the AUC value of StepCox(both) with GBM is the highest, suggesting that this model has a higher prediction accuracy (Fig. S2a, b). The StepCox(both) with GBM model incorporates four genes: CSRP1, GGH, SAP18, and TRIM29. This four-gene signature was referred to as the VMRG model. We applied the optimal model to all three cohorts and performed a univariate Cox regression meta-analysis to evaluate its prognostic value. The analysis yielded hazard ratios (HRs) of 4.67 and 2.41, respectively (all P < 0.05). Both fixed- and random-effects meta-analysis models indicated a significantly elevated risk (P < 0.001) (Fig. 4b). The VMRGs model, identified from our analysis, was further evaluated using time-dependent ROC curves (Fig. 4c). Similarly, the 1-year and 3-year AUC values of the VMRGs model consistently remained at a high level in almost all cohorts and often ranked first (Fig. S2a, b). These findings indicate that the VMRGs model has a high degree of predictive accuracy. Based on the median risk score calculated from the VMRGs model, patients were stratified into high- and low-risk groups. Subsequent log-rank tests revealed significantly poorer survival outcomes in the high-risk group across all cohorts (P < 0.05; Fig. 4d, e). The heat map shows the expression distribution of model genes in high-risk and low-risk patients across two datasets (Fig. 4f).

Fig. 4. Development and validation of the VMRGs prediction model.

a The heat map summarizes 117 predictive models on three large datasets and their corresponding C-index. b The results of the univariate regression analysis of the predictive model. c The ROC curve of the VMRGs model for predicting patient survival. d The OS between the high-risk group and the low-risk group defined by the best model was compared using the Kaplan–Meier curve. e Visualization of patient survival status and risk score distribution. f The heat map shows the expression distribution of the model genes.

Validation of the VMRGs model’s superiority compared to existing published predictive models

To evaluate the superiority of the VMRGs model, we compared it with five previously published EC-related predictive models. The risk scores of these models were calculated in two independent cohorts, and the corresponding HRs were derived (Fig. S3a). The results showed that the VMRGs model was consistently and significantly associated with poor prognosis in all cohorts (all HR > 1, all P < 0.05), demonstrating better predictive ability than the other published models. Moreover, in terms of discriminative performance, the VMRGs model achieved mid-to-high-level C-index and AUC values in both cohorts, ranking first and second, respectively (Fig. S3b–d).

Molecular subtypes of ESCC were constructed based on VMRGs

Using consensus clustering of the GSE53624 cohort, we identified two robust molecular subtypes (C1 and C2) for ESCC (Fig. 5a). PCA revealed the differences in the two-dimensional space distribution among the two ESCC subgroups (Fig. 5b). And, the survival analysis results showed that the C2 subgroup had the poorest prognosis (Fig. 5c). Furthermore, the sankey diagram and heat map provide visual evidence that the majority of the C2 subgroup belong to the high-risk group (Fig. 5d). The GSVA enrichment analysis results indicated that there were significant differences in pathways between group C1 and group C2. Cluster C2 was mainly enriched in Notch, EMT, TGF-β, and Wnt/β-catenin signaling pathways, and also showed potential associations with angiogenesis, involving the regulation of multiple pathways related to vascular formation (Fig. S4a–c).

Fig. 5. Identification of molecular subgroups in ESCC patients based on VMRGs.

a Identification of two molecular subgroups. b PCA illustrate the spatial distribution of the two molecular subgroups. c The survival differences between the two molecular subgroups. d Sankey diagram demonstrated the relationship between molecular clusters, risk groups, and survival status. e The differences in immune cells among different molecular clusters. f The relationship between immune checkpoint gene expression and different molecular clusters. g The differences in stromal score, immune score, ESTIMATE score, and tumor purity among different molecular clusters. h The differences in TIDE scores among different molecular clusters. i Differences in drug sensitivity between the two distinct molecular subtypes.

Analysis of differences in immune infiltration and drug sensitivity among different clusters

To comprehensively characterize the tumor immune microenvironment (TME) across molecular clusters, we applied the CIBERSORT algorithm to the GSE53624 cohort. This analysis revealed a significant increase in the infiltration of immunosuppressive M2 macrophages in the high-risk cluster C2 (Fig. 5e). At the level of immune checkpoint expression, cluster C2 showed distinct patterns, with up-regulation of TNFRSF4, but downregulation of TNFRSF18 and HHLA2 compared to C1 (Fig. 5f). Furthermore, assessment with the ESTIMATE algorithm indicated that cluster C2 was associated with a stroma-rich microenvironment, exhibiting significantly elevated Stromal and ESTIMATE scores alongside reduced tumor purity (Fig. 5g). Moreover, the C2 subgroup has relatively high CAF, Exclusion, and TIDE scores, suggesting that the C2 subgroup has immune evasion and is difficult to benefit from immunotherapy (Fig. 5h). Meanwhile, the results indicated that the C2 molecular subpopulation was insensitive to multiple anti-tumor drugs, with significantly higher IC50 values (Fig. 5i).

VM-related gene SAP18 is a poor prognostic factor for ESCC

Based on the TCGA database, we performed a pan-cancer analysis of VMRG model gene expression. The results revealed significantly elevated mRNA levels of GGH, SAP18, and TRIM29 in EC, while CSRP1 expression showed no significant difference (Fig. 6a–d). We further evaluated the prognostic value of these VMRG model genes using datasets from GSE53624, GSE53622, and TCGA, assessing overall survival (OS), disease-specific survival (DSS), and progression-free interval (PFI). Survival analysis demonstrated that high SAP18 expression was significantly associated with poorer OS, DSS, and PFI; TRIM29 expression was correlated with OS and PFI; and the other two genes were only associated with OS in the GSE53622 or GSE53624 dataset (Fig. 6e–h). Subsequently, patients were stratified into high- and low-expression groups based on median expression values, and univariate Cox regression was performed. The results indicated that only SAP18 remained significantly associated with patient prognosis in two independent datasets (Fig. 6i, j). By integrating these findings, we identified SAP18 as a key VM-driving gene that is consistently linked to patient prognosis (Fig. 6k).

Fig. 6. The relationship between SAP18 and survival.

a CSRP1 mRNA levels in TCGA tumor samples compared to paired normal tissues. b GGH mRNA levels in TCGA tumor samples compared to paired normal tissues. c SAP18 mRNA levels in TCGA tumor samples compared to paired normal tissues. d TRIM29 mRNA levels in TCGA tumor samples compared to paired normal tissues. e Explore the relationship between CSRP1 expression and OS, DSS, and PFI of patients in different datasets respectively. f Explore the relationship between GGH expression and OS, DSS, and PFI of patients in different datasets respectively. g Explore the relationship between SAP18 expression and OS, DSS, and PFI of patients in different datasets respectively. h Explore the relationship between TRIM29 expression and OS, DSS, and PFI of patients in different datasets respectively. i Univariate analysis of the VMRGs in the GSE53625 dataset. j Univariate analysis of the VMRGs in the TCGA dataset. k Core prognostic gene screening.

Association of SAP18 with the tumor immune microenvironment and immunotherapy response

Further research has found that the expression of SAP18 is positively correlated with the expression of multiple immunoinhibitors and the infiltration of immune cells, and similar results have been shown in multiple datasets. Notably, our analysis revealed that SAP18 expression was significantly inversely correlated with the infiltration of unpolarized M0 macrophages, while showing essentially no association with M1 macrophages. In contrast, SAP18 expression exhibited a significant positive correlation with M2 macrophage infiltration (Fig. 7a–c). Altogether, these findings suggest that elevated SAP18 is linked to a reduction in M0 macrophages and a relative enrichment of pro-tumor M2 macrophages, potentially driving an immunosuppressive M2-polarized tumor microenvironment that facilitates VM and tumor progression. This is further supported by the strong positive correlation observed between SAP18 and the immunosuppressive factor IL-10RB (Fig. 7b). At the same time, it has a significant positive correlation with the scores reflecting the status of the tumor microenvironment (stromal score, immune score, and ESTIMATE score) (Fig. 7c). Meanwhile, in the immunotherapy dataset, we found that the expression of SAP18 is associated with the efficacy of immunotherapy. Patients with high expression of SAP18 can obtain better therapeutic benefits from immunotherapy (Fig. 7d). We leveraged data from the TCGA cohort to investigate potential associations between SAP18 expression and various genomic alterations, including somatic mutations, and copy number variations (losses and gains). Our analysis indicated that the majority of common genetic alterations were distributed similarly between groups characterized by high and low SAP18 expression. A distinct finding emerged from the analysis of EC cases, where a significant positive correlation was observed between elevated SAP18 expression and the mutational frequency of USH2A (Fig. 7e).

Fig. 7. The correlation between SAP18 expression and immune cell infiltration.

a Heatmap show the relationship between SAP18 and multiple immune cell infiltrations and ESTIMATE results in several EC datasets. b Heatmap shows the relationship between SAP18 and multiple immunosuppressive molecules in multiple EC datasets. c The relationship between SAP18 expression and Stromalscore, Immunescore, and ESTIMATEscore in the ESCC dataset. The relationship between SAP18 expression and infiltration of macrophage subtypes in the ESCC dataset. d The relationship between SAP18 expression and the outcome of immunotherapy. e Genetic alterations linked to SAP18 expression were analyzed via a heatmap in a cohort of EC patients from TCGA.

SAP18 participates in the malignant progression of ESCC by promoting VM formation, cell proliferation, and metastasis

As a molecule involved in malignant processes such as proliferation and metastasis by regulating signaling pathways in various malignancies, the specific role and mechanism of SAP18 in ESCC remain unclear, particularly its impact on VM, which urgently needs clarification. The analysis of protein in 6 pairs of ESCC patient tissues showed that the expression of SAP18 protein was elevated in tumors (Fig. 8a, b). IHC staining analysis revealed that the protein expression level of SAP18 in ESCC tissues was significantly higher than that in adjacent normal tissues (Fig. 8c, d). Further correlation with clinicopathological features showed that SAP18 expression level was significantly positively correlated with adverse clinical indicators such as TNM stage (Fig. 8e, f).

Fig. 8. SAP18 promotes vasculogenic mimicry and malignant progression in ESCC.

a, b Western blotting analysis and corresponding quantification of SAP18 protein levels in 6 paired ESCC tumor (T) and normal (N) tissues. c, d Representative IHC images and staining scores of SAP18 in ESCC tumor tissues and adjacent normal tissues. e, f Representative IHC images and staining scores of SAP18 in ESCC tissues at different clinical stages (Stage I–II vs. Stage III–IV). g–j Verification of SAP18 knockdown efficiency. Western blot analysis of SAP18 protein levels in KYSE150 and KYSE410 cells. qRT-PCR analysis of SAP18 mRNA relative expression in KYSE150 and KYSE410 cells stably transfected with shNC, shSAP18#1, or shSAP18#2. k–n Matrigel tube formation assay evaluating the VM capacity of KYSE150 and KYSE410 cells following SAP18 knockdown. The number of VM structures was quantified for KYSE150 and KYSE410. o, p Western blot analysis of VE-cadherin and SAP18 protein expression in KYSE150 and KYSE410 cells upon SAP18 knockdown. q, r Representative images of subcutaneous xenograft tumors and quantification of tumor volume from nude mice injected with KYSE150-shNC or KYSE150-shSAP18 cells. s Representative IHC images of SAP18, VE-cadherin, and CD34/PAS dual staining in the xenograft tumor tissues. Red arrows indicate VM channels, and blue arrows indicate endothelium-dependent blood vessels. β-Actin was used as an internal loading control for Western blotting. Data are presented as mean ± SD. *P < 0.05; **P < 0.01; ***P < 0.001.

To investigate the biological function of SAP18 in ESCC, particularly its regulatory role in VM formation, we generated stable SAP18-knockdown ESCC cell lines (KYSE150 and KYSE410) using lentiviral vectors. The knockdown efficiency was rigorously confirmed via Western blot and qRT-PCR (Figs. 8g–j, S5a, b). Subsequent in vitro functional assays demonstrated that the depletion of SAP18 significantly impaired the capillary tube formation capacity of ESCC cells (Fig. 8k–n). Furthermore, Western blot analysis revealed a substantial reduction in the expression of VE-cadherin, a critical molecular marker for VM, following SAP18 knockdown, suggesting that SAP18 facilitates VM formation, potentially by upregulating VE-cadherin (Fig. 8o, p). Additionally, Transwell, colony formation, and CCK-8 assays corroborated that SAP18 knockdown significantly attenuated the invasive, metastatic, and proliferative capacities of ESCC cells in vitro (Fig. S6a–e).

We further validated these findings using an in vivo subcutaneous xenograft model in nude mice. Tumors derived from the SAP18-knockdown group (shSAP18) exhibited a significantly reduced volume compared to the control group (shNC), confirming that SAP18 downregulation suppresses ESCC tumor growth in vivo (Fig. 8q, r). Crucially, CD34/PAS dual staining of the tumor sections revealed a marked reduction in VM channels (identified as PAS-positive and CD34-negative (PAS+/CD34-) networks) in the SAP18-depleted tumors. Parallel IHC staining confirmed a concurrent decrease in VE-cadherin expression within the tumor tissues, reinforcing the premise that SAP18 promotes ESCC progression by driving VM formation in vivo (Fig. 8s).

To further evaluate the biological function of SAP18 in vivo, we performed a subcutaneous tumorigenesis assay in nude mice and zebrafish metastasis model experiment. The zebrafish metastasis model experiment showed that after knockdown of SAP18, the number of metastatic tumor masses formed by ESCC cells in zebrafish was significantly reduced, suggesting that SAP18 can promote the metastasis process of ESCC in vivo (Fig. 9a, b). Furthermore, we found that knockdown of SAP18 could significantly inhibit the lung metastasis ability of the ESCC cell line KYSE150 (Fig. 9c, d).

Fig. 9. SAP18 knockdown suppresses the metastasis of ESCC cells in vivo.

a, b The zebrafish model was used to analyze the dissemination and metastasis of ESCC cells. c, d Mouse lung metastasis model was used to evaluate the lung metastasis ability of ESCC cells. Data are presented as mean ± SD. **P < 0.01; ***P < 0.001.

Based on the comprehensive results of in vitro and in vivo experiments, SAP18 participates in the malignant progression of ESCC by promoting VM formation, cell proliferation, and metastasis, and its high expression may be a potential molecular marker for poor prognosis in ESCC patients.

SAP18 promotes vasculogenic mimicry via the PI3K/AKT/mTOR/HIF-1α signaling axis

To gain deeper insights into the signaling pathways involving SAP18, we performed enrichment analysis. GSEA for Hallmark gene sets revealed that SAP18 is predominantly implicated in oxidative phosphorylation (OXPHOS), glycolysis, the PI3K/AKT/mTOR signaling pathway, and various metabolic pathways (Fig. 10a). Furthermore, Over-Representation Analysis (ORA) for GO and KEGG indicated a significant association of SAP18 with ATP synthesis, OXPHOS, fatty acid metabolism, and carbohydrate metabolism pathways (Fig. S7a, b). Among them, the PI3K/AKT/mTOR signaling pathways have been widely confirmed to be closely related to VM^25,26. To further explore the functional mechanism of SAP18, we used LC-MS/MS to identify interacting partners of SAP18 and regulatory molecules associated with the above signaling pathways, and selected PIK3CB for further analysis. To further validate the mass spectrometry findings, Co‑IP assays were performed to examine the interaction between SAP18 and PIK3CB. Physical interactions between SAP18 and PIK3CB were detected in both KYSE150 and KYSE410 cells. In addition, immunofluorescence staining showed distinct colocalization of SAP18 and PIK3CB. Collectively, these results confirm that SAP18 physically interacts with PIK3CB.

Fig. 10. SAP18 promotes VM in ESCC through the PI3K/AKT/mTOR/HIF-1 signaling pathway.

a The results of GSEA-HALLMARK enrichment analysis. b Co-IP combined with mass spectrometry to identify binding partners of SAP18. c Co-IP demonstrated an interaction between SAP18 and PIK3CB in ESCC cells (KYSE150 and KYSE410). IgG served as the negative control, and input represented the whole-cell protein lysate used as the positive control. d Immunofluorescence colocalization in ESCC cells: red (SAP18), green (PIK3CB), and blue (nucleus). Fluorescence intensity distribution plots also confirmed the colocalization of SAP18 and PIK3CB. e, f Western blot reveals changes in protein expression of PIK3CB/AKT/mTOR/HIF-1α pathway in ESCC cells after knocking down or overexpressing SAP18. g Molecular mechanism of SAP18 in ESCC. h Tubule formation was conducted in three groups (OE-SAP18, OE-SAP18-DMSO, OE-SAP18-MK2206). i, j The animal models were divided into two groups: the DMSO control group and the MK2206 treatment group. k IHC staining for VM (CD34-/PAS+) tubes and statistical analysis, red arrows indicate VM, and blue arrows indicate blood vessels. β-actin serves as an internal reference. Data are presented as mean ± SD.**P < 0.01.

We further evaluated the regulatory effect of SAP18 on this signaling cascade. Western blot analysis revealed that knockdown of SAP18 led to a reduction in PI3KCB protein expression. And it also demonstrated that SAP18 knockdown significantly decreased the phosphorylation levels of p‑AKT, and p‑mTOR, as well as the protein expression of the downstream effector HIF‑1α, whereas the total protein levels of AKT, and mTOR remained largely unaltered (Fig. 10e). In contrast, overexpression of SAP18 notably upregulated the phosphorylation of these kinases and the expression of PI3KCB and HIF‑1α (Fig. 10f, g).

To determine whether the PI3K/AKT/mTOR axis is indispensable for SAP18‑mediated VM formation, we employed the highly selective AKT inhibitor MK2206. In vitro tube formation assays revealed that MK2206 treatment markedly abrogated the VM‑forming capacity of OE-SAP18 ESCC cells (Fig. 10h). In vivo xenograft models further validated these observations: mice administered MK2206 displayed significantly reduced tumor volumes relative to the DMSO‑treated control group (Fig. 10i, j). Additionally, dual immunohistochemical staining for CD34 and PAS in tumor tissues confirmed that MK2206 treatment effectively suppressed the formation of CD34-/PAS+ VM networks (Fig. 10k).

Based on the enrichment analysis and experimental results, we hypothesize that SAP18 promotes VM formation potentially via the PI3K/AKT/mTOR/HIF-1α pathway, and that MK2206 could serve as a targeted therapeutic agent to suppress VM in ESCC. These results further underscore the pivotal and multifaceted role of SAP18 in driving ESCC progression.

Collectively, these findings support a novel mechanism whereby SAP18 activates the PI3K/AKT/mTOR signaling cascade through binding to PIK3CB, thereby upregulating HIF-1α to promote VM formation. These results highlight the critical role of SAP18 in driving ESCC progression and indicate that MK2206 may serve as a promising potential targeted therapeutic agent for suppressing VM in ESCC.

Discussion

Treatment for advanced or recurrent ESCC remains challenging, and progress in improving survival has been modest⁴. Understanding the biological processes that drive tumor progression and treatment resistance is essential for developing new strategies. When conventional endothelial angiogenesis fails to adequately support tumor growth, VM acts as an alternative microvascular circulatory mechanism that provides essential support for tumor invasion and metastasis^12–15. Previous studies have established that VM is closely associated with therapeutic resistance and poor survival across multiple malignant tumors^27–29. However, the molecular drivers and clinical translational potential of VM in ESCC have not been fully characterized. In the present study, analysis of clinical ESCC specimens confirmed that VM is significantly correlated with advanced clinical stage and poor prognosis, and serves as an independent risk factor for unfavorable outcomes in ESCC patients. More importantly, by integrating single-cell and bulk RNA-seq data, our study is the first to comprehensively uncover the multifaceted roles of SAP18 as a core hub in regulating VM in ESCC.

Firstly, by integrating multi-omics data, we were able to identify a specific set of genes associated with VM and construct a 4-gene prognostic signature. This model not only predicts patient treatment outcomes but also demonstrates significantly superior predictive performance compared to previously reported prediction models. Additionally, based on the 4-gene prognostic signature, we identified two distinct molecular subtypes with divergent clinical outcomes and immune infiltration characteristics. The C2 molecular subtype shared features of poor prognosis, high stromal and immune scores, and enrichment of M2 macrophages, suggesting that VM is closely linked to an immunosuppressive tumor environment. These findings extend earlier work describing the interaction between VM and immune infiltration in cancer³⁰.

Furthermore, in this study, we demonstrated that the C2 subtype is associated with resistance to multiple therapeutic agents. Previous studies have indicated that VM can lead to reduced drug sensitivity, particularly to anti‑angiogenic therapies^31,32. This is because VM consists of tumor cell‑derived, endothelial‑independent tubular networks that serve as an alternative microcirculation, supplying nutrients and promoting tumor dissemination. Conventional anti‑angiogenic agents, such as VEGF pathway inhibitors, primarily act by inhibiting endothelial cell‑mediated neovascularization. In contrast, VM channels lack an endothelial lining, which may allow tumor cells to evade the effects of such drugs, thereby maintaining blood supply and ultimately leading to treatment failure³³. Our work, at the molecular subtyping level, confirms that the VM‑enriched C2 subtype represents a group with intrinsic multidrug‑resistant potential in ESCC. This provides a new perspective for understanding therapeutic resistance in ESCC and suggests that intervention strategies targeting the C2 subtype may represent an effective approach to overcome such resistance. Notably, the subtyping is based on the expression profiles of four key genes: CSRP1, GGH, SAP18, and TRIM29. Targeting these genes for precise therapy could therefore be a promising direction to overcome drug resistance and immunosuppression.

Based on a comprehensive analysis of four VM-associated genes, SAP18 was ultimately selected for further investigation. SAP18 is a component of the Sin3-HDAC1 complex that regulates chromatin remodeling and transcriptional repression^21,22. Survival analysis in a larger cohort revealed that SAP18 expression was closely associated with patient prognosis. Moreover, analysis of real-world patient specimens demonstrated a significant correlation between SAP18 levels and advanced tumor stage. Both in vitro and in vivo experiments confirmed that knockdown of SAP18 expression effectively suppressed multiple tumorigenic phenotypes, including VM formation, proliferation, and migration. Collectively, our findings identify SAP18 as a pivotal regulator that drives VM in ESCC.

While earlier reports have associated SAP18 expression with poor prognosis in ESCC³⁴, its biological function has not been clearly defined. Informed by our enrichment analysis, which implicated SAP18 in the PI3K/AKT/mTOR pathway, which has been established as critical for VM formation^12,35. To elucidate the underlying mechanism of SAP18 in the PI3K/AKT/mTOR signaling pathway, we utilized Co-IP coupled with LC-MS/MS to screen for SAP18-interacting proteins. By intersecting these candidates with known regulators of the PI3K/AKT/mTOR pathway, we ultimately identified PIK3CB for further investigation. As a crucial subunit of PI3K, PIK3CB has been previously demonstrated to activate the PI3K/AKT/mTOR pathway in ESCC by promoting downstream AKT phosphorylation, thereby driving tumor progression³⁶. Our study further reveals that SAP18 mediates the activation of this downstream cascade by regulating PIK3CB expression, ultimately facilitating the formation of VM in ESCC. Numerous studies have confirmed that the PI3K/AKT/mTOR signaling pathway can enhance the activity of transcription factors such as FoxK1, NF-κB, SP1, and c-MYC, thereby promoting the transcription of HIF-1α³⁷, which is a key molecule driving tumor progression and VM formation. As a master transcriptional regulator, HIF-1α subsequently induces VM-associated markers, notably VE-cadherin, conferring epithelial-derived tumor cells with the structural capacity to form vascular networks³⁸. Crucially, the functional dependency of SAP18-driven VM on this signaling axis was validated using the highly selective AKT inhibitor MK-2206. Pharmacological blockade with MK-2206 not only markedly abrogated the capillary tube-forming capacity of ESCC cells in vitro but also profoundly suppressed xenograft tumor growth and the formation of CD34-/PAS + VM networks in vivo. Collectively, these data firmly establish the PI3K/AKT/mTOR/HIF-1α signaling axis as the central engine of VM in ESCC, positioning MK-2206 as a compelling targeted therapeutic agent to uncouple SAP18-driven microvascular remodeling.

In addition, enrichment analysis also indicated that SAP18 participates in both glycolysis and OXPHOS pathways. These metabolic programs are known to support the energy demand of rapidly growing tumors and are often linked to immune suppression^9,39. Previous studies have suggested that enhanced glycolysis can drive VM formation⁴⁰ and that OXPHOS contributes to metastasis and therapeutic resistance^41–44. The association of SAP18 with these pathways suggests that it may coordinate metabolic reprogramming and vascular mimicry to sustain ESCC progression. These findings provide insights into the biological heterogeneity of ESCC and suggest potential applications for patient stratification and targeted therapy against VM.

Furthermore, we also found that SAP18 expression correlated with the infiltration of M2 macrophages and the expression of immunosuppressive checkpoint molecules. Tumor-associated macrophages with an M2 phenotype can promote angiogenesis, tissue remodeling, and immune evasion⁴⁵. The link between SAP18 and this immune profile implies that targeting SAP18 may have a dual benefit by disrupting both VM formation and the immunosuppressive microenvironment. This observation is consistent with recent studies highlighting the potential of targeting M2-like macrophages to overcome therapy resistance in solid tumors⁴⁵. Meanwhile, we found that patients with high expression of SAP18 could achieve better immunotherapy effects. This indicates that SAP18 might be a potential predictive marker for immunotherapy. Mutation analysis in the present study revealed a significantly higher frequency of USH2A gene mutations in populations with high SAP18 expression. Notably, accumulating evidence has clearly demonstrated that lung cancer patients harboring USH2A missense mutations exhibit substantially improved clinical responses to immune checkpoint inhibitors (ICIs), including prolonged OS and PFS. This finding suggests that high SAP18 expression may be associated with a favorable immune microenvironment driven by USH2A mutations⁴⁶. Therefore, we hypothesize the existence of a potential “SAP18-USH2A functional axis,” which may modulate the immunogenicity of tumor cells or the host anti-tumor immune response, thereby affecting the therapeutic efficacy of ICIs. Considering its immunosuppressive effect, it may play a role similar to that of immune checkpoint molecules. Meanwhile, it has been described that the process of VM and the immune system are interdependent, forming a mutually reinforcing vicious cycle that promotes tumor immune evasion and malignant progression³⁰. Notably, SAP18 may serve as a key regulatory hub mediating the crosstalk between these two processes.

Together, these findings support a model in which SAP18 enhances VM and tumor aggressiveness through the PI3K/AKT/mTOR/HIF‑1α signaling axis and immune modulation. The identification of SAP18 as a molecular driver of VM has several clinical implications. First, its expression may serve as a biomarker for patient stratification, helping to identify ESCC cases with poor prognosis or resistance to standard therapy. Second, SAP18 inhibition may represent a therapeutic opportunity. Agents that modulate chromatin remodeling or cellular metabolism could potentially block SAP18-driven pathways. Finally, integrating SAP18 status into molecular profiling frameworks could improve the prediction of treatment response in future clinical trials.

Organoid models have emerged as valuable tools for studying tumor biology, as they can largely preserve the structural and functional characteristics of patient-derived tumors^47–52. In our preliminary observations, putative VM-like structures were noted within patient-derived ESCC organoids, where tumor cells formed tubular patterns. However, it must be acknowledged that the characterization of VM formation in this organoid model remains incomplete in the present study. Specifically, the current organoid model lacks appropriate specific controls to definitively distinguish cancer cells from co-existing stromal and endothelial cell populations within the 3D architecture. Therefore, whether tumor cells can independently form VM in organoids is still a tentative observation, and these initial findings warrant further exploration and validation through more rigorous experimental designs in future research. Despite these limitations, this platform holds promise as a potential opportunity for drug screening and mechanistic studies in systems that approximate the clinical setting. Previous studies have utilized organoids to test anti-angiogenic and metabolic inhibitors^53,54; a similar approach may, in the future, be applied to evaluate potential compounds targeting SAP18-mediated VM.

There are limitations to acknowledge. The prognostic model was developed using retrospective datasets and will need validation in larger, prospective cohorts. Although our experiments confirm a functional role of SAP18 in VM formation, the precise molecular interactions and downstream targets of SAP18 require further study. Additional work using metabolomic profiling or CRISPR-based perturbation could help define its role in energy metabolism and immune regulation. Despite these limitations, our findings provide a foundation for future research on SAP18 and VM as therapeutic targets in ESCC. Furthermore, it is important to acknowledge the methodological considerations in our current identification of VM. In this study, we utilized CD34/PAS dual staining to delineate endothelium-independent VM networks. Although CD34/PAS double staining remains one of the most mainstream and widely employed approaches for VM evaluation in current literature, it is important to recognize that CD34 expression can be variable among certain endothelial populations and stem/progenitor cells^13,55–57. Consequently, it is not considered the absolute gold standard for this application. In contrast, CD31 provides a more reliable endothelial reference. Currently, the most accepted and definitive gold standard for differentiating VM structures from classical endothelial vessels is the combination of CD31 immunonegativity with PAS positivity (CD31-/PAS+)⁵⁸. Therefore, while our CD34-/PAS+ findings strongly indicate VM formation driven by SAP18, future investigations should incorporate CD31/PAS dual staining to prevent potential misinterpretation and provide the most rigorous validation of these microvascular networks.

In conclusion, this study systematically clarifies the clinical value and molecular regulatory mechanisms of VM in ESCC. First, VM serves as an independent prognostic indicator for poor outcomes in ESCC patients, and its density is closely related to TNM stage and survival time. Second, SAP18 emerges as a key molecule regulating VM formation in ESCC. High expression of SAP18 indicates a poor prognosis. Mechanistically, as evidenced by Co-IP analysis, SAP18 physically interacts with PIK3CB to activate the PI3K/AKT/mTOR/HIF-1α pathway, thereby promoting tumor proliferation, metastasis, and VM formation. Furthermore, the 4-gene VMRGs model constructed in this study demonstrates reliable prognostic predictive ability, making it a valuable tool for the clinical stratification of ESCC patients. Additionally, the application of the AKT inhibitor MK2206 effectively suppresses VM formation and tumor growth in ESCC, highlighting it as a potential candidate drug for VM-targeted therapy. Ultimately, this research not only deepens the understanding of VM regulatory mechanisms in ESCC but also provides important experimental evidence and theoretical support for prognostic evaluation and personalized treatment, demonstrating significant potential for clinical translation.

Methods

Data collection and preprocessing

RNA-seq data and corresponding clinical information of EC were acquired from the TCGA database via the University of California Santa Cruz Xena browser (UCSC-Xena) (https://xenabrowser.net/)⁵⁹. Additionally, gene expression profiles and clinical data of ESCC from GSE53625 (n = 179), GSE53624 (n = 119), and GSE53622 (n = 60), which were obtained from the Gene Expression Omnibus (GEO) database (https://www.ncbi.nlm.nih.gov/geo/)^60,61. Among these, GSE53624 was used as the training set, while GSE53622 served as the validation set. scRNA-seq data of 10 ESCC samples from 2 datasets were obtained from the GEO database (GSE188900, GSE196756). Additionally, we also collected one fresh ESCC tissue sample from a treatment-naïve patient diagnosed at the First Affiliated Hospital of Zhengzhou University, which was subjected to scRNA-seq analysis. In summary, a total of 11 scRNA-seq samples were obtained in this study. Additionally, VMRGs were retrieved from the GeneCards database (https://www.genecards.org/) using “Vasculogenic mimicry” as the search keyword. Genes with relevance scores greater than the median value were selected, resulting in a final list of 191 VMRGs (Supplementary Table S1).

Single-cell RNA-seq data processing and analysis

All analyses were conducted with Seurat (v4.4.0) in R. Preprocessing of raw UMI counts involved retaining genes detected in a minimum of three cells. Quality control thresholds were set to exclude cells with fewer than 200 expressed genes or mitochondrial gene content above 20%, in order to remove low-quality or apoptotic cells. To reduce doublets, potential multiplet events were identified and filtered using the scDblFinder function. The resulting high-quality cells were normalized via the LogNormalize method. Dimensionality reduction was performed by principal component analysis (PCA) on highly variable genes. The number of principal components (PCs) was determined using the elbow method, which identifies the “inflection point” where the standard deviation of principal components shows a sharp decline followed by a plateau, thereby balancing information retention and noise reduction^62,63. These components were then used for graph-based clustering and visualization via t-distributed Stochastic Neighbor Embedding (t-SNE). Cell clusters were identified using the FindNeighbors and FindClusters functions (resolution = 0.2). Annotation of cell types was performed by integrating reference-based annotation with SingleR and evaluating known marker genes to ensure biologically meaningful classification.

Screening VMRGs in scRNA-seq data

To delineate single-cell VMRGs, we first identified cell type-specific marker genes using the FindAllMarkers algorithm. Genes upregulated in epithelial cells were intersected with the 191 VMRGs from the GeneCards database to identify scVMRGs. The AUCell package was subsequently employed to compute gene set enrichment scores, which reflect the activity levels of scVMRGs within individual epithelial cells. Based on the distribution of these scores, epithelial cells were categorized into VM-high and VM-low subgroups using a quartile-based threshold. Marker genes distinguishing the VM-high subgroup were then identified using the FindAllMarkers function for subsequent functional investigations.

Differential expression analysis

DEGs between ESCC specimens and normal controls were identified utilizing the limma R package, applying a threshold of absolute log2-fold change greater than 1 and an adjusted P-value of less than 0.05 for statistical significance.

Screening prognostic signatures by 117 machine learning combinations

The 117 machine learning algorithms were applied by using the mime R package⁶⁴, with the GSE53624 dataset serving as the training set and the GSE53622 dataset as the validation set. These 117 algorithms were primarily constructed through permutations and combinations of 10 core methodologies. These included stepwise Cox regression, CoxBoost, Least Absolute Shrinkage and Selection Operator (Lasso), Ridge, Elastic-Net (across α values from 0.1 to 0.9) (Enet), survival-support vector machines (SVM), Supervised Principal Components (SuperPC), Generalized Boosted Regression Models (GBM), partial least squares Cox regression (plsRcox), and Random Survival Forests (RSF). Predictive performance was quantified by Harrell’s concordance index (C-index) in each validation cohort. The model exhibiting the highest average C-index across all validation sets was chosen as the final prognostic signature. The area under the curve (AUC) value, representing the predictive accuracy for patient survival, was concurrently calculated for each machine learning model. Subsequently, patients were stratified into high- and low-risk groups based on the median risk score derived from the optimal model. Meanwhile, we collected the previously published models on the prognosis of EC and compared our model with the previously published ones^65–69.

Survival analysis

Kaplan–Meier survival analysis was performed using the survfit method (Survival package), with curve visualization facilitated by Survminer in R to quantify survival differences between risk groups or clusters.

Enrichment analyses

Functional enrichment analyses were performed using the R package clusterProfiler. Specifically, Over-Representation Analysis (ORA) was conducted for both the Kyoto Encyclopedia of Genes and Genomes (KEGG) and Gene Ontology (GO) databases. Separately, Gene Set Enrichment Analysis (GSEA) was applied to the KEGG, Hallmark, and GO gene sets.

Subtype construction and principal component analysis

Molecular subtypes of ESCC were identified by consensus clustering using the R package ConsensusClusterPlus. The clustering was performed with 100 iterations, resampling 80% of the samples each time, and using k-means clustering (ClusterAlg = “pam”) with a Spearman distance metric (distance = “pearson”). The maximum number of clusters (k) evaluated was 9. The robustness of the identified subtypes was assessed via PCA using the prcomp function from the stats package. Prior to PCA, gene expression values were Z-score standardized to generate a normalized input matrix.

Immune infiltration analysis

To delineate the immune landscape among the identified clusters, we conducted an integrated computational profiling strategy. First, the relative infiltration levels of various immune cell types were quantified and compared across clusters using the CIBERSORT algorithm. Second, we employed the ESTIMATE algorithm to compute four key tumor microenvironment scores for each ESCC sample based on gene expression profiles: StromalScore, ImmuneScore, ESTIMATEScore, and resultant Tumor Purity. Additionally, we assessed the associations between cluster assignments and expression levels of immune checkpoint inhibitor (ICI) genes through correlation analysis. Lastly, the likelihood of immunotherapy response was predicted for each sample using the Tumor Immune Dysfunction and Exclusion (TIDE) computational framework (http://tide.dfci.harvard.edu/).

Drug sensitivity analysis

To evaluate the clinical potential of the identified molecular subtypes, we predicted the chemotherapeutic and targeted drug response for each ESCC patient using the oncoPredict R package. The model was trained using the Genomics of Drug Sensitivity in Cancer (GDSC2) database as a reference dataset, which provides extensive drug screening data across hundreds of cancer cell lines. By mapping the bulk RNA-seq gene expression profiles of our ESCC cohorts onto the GDSC2-derived training models, we calculated the half-maximal inhibitory concentration (IC50) for a wide range of anti-tumor agents via ridge regression. Differences in predicted IC50 values between the C1 and C2 molecular subgroups were subsequently assessed using the Wilcoxon rank-sum test, where a higher IC50 value indicated lower drug sensitivity.

Utilization of online databases

Part of data visualization and bioinformatic analyses—including immune infiltration analysis, gene mutation analysis, and functional enrichment—were performed using the BEST (Biomarker Exploration of Solid Tumors) online platform. This publicly accessible resource, established by Liu et al. in 2023, integrates multi-omics datasets from over 10,000 solid tumor samples across 27 cancer types (e.g., TCGA, CGGA, and GEO) (https://rookieutopia.hiplot.com.cn/app_direct/BEST/)⁷⁰.

Human ESCC specimens

All research involving human participants, tissues, and related data was performed in accordance with the principles of the Declaration of Helsinki. This study was approved by the Ethics Committee of the First Affiliated Hospital of Zhengzhou University, with the reference approval number 2025-KY-1380-003. A total of 150 ESCC tissues and corresponding paracancerous tissues were collected from patients admitted to the First Affiliated Hospital of Zhengzhou University between 2012 and 2019. All participating patients provided written informed consent prior to sample collection. None of the patients had received any anti-cancer treatment before the biopsy. Collected tissue samples were immediately stored at −80 °C until subsequent experimental use.

Immunohistochemistry

Immunohistochemistry (IHC) was performed as previously described¹³. Staining intensity grades were divided into levels 0 (negative), 1 (weakly positive), 2 (positive), and 3 (strongly positive), and the dyeing area score was divided into 1 (0–25%), 2 (26–50%), 3 (51–75%), and 4 (>75%). The final staining score was defined as the product of the two scores, with a score of 0–6 defined as low expression and a score of 7–12 as high expression. IHC was used to detect the expression of SAP18, CD34/PAS and VE-cadherin. For the quantification of VM, we employed the classic Weidner’s “hotspot method”⁷¹. Initially, each immunohistochemistry slide was comprehensively scanned at low magnification (40×) to visually identify the “hotspots” exhibiting the highest density of VM networks, from which five representative hotspots were selected. Subsequently, under high magnification (100×), the number of VM channels-strictly defined as PAS-positive/CD34-negative (CD34-/PAS+) networks enclosed by tumor cells was manually counted within each of the five selected hotspots. The average count across these fields was calculated to represent the VM number for that slide. To minimize subjective bias, this quantification process was independently performed by three experienced researchers who were blinded to the patients’ clinical and pathological data. The number of VM for each sample was determined by calculating the mean value of the counts from the three independent observers.

Cell culture

All experiments were performed in a laminar flow hood in the cell culture room. First, complete medium was prepared: 10% fetal bovine serum (FBS) was added to the basal medium, followed by 1% of 100× penicillin-streptomycin mixture (containing 100 μg/mL streptomycin and 100 μg/mL penicillin), and mixed thoroughly. ESCC cell lines (KYSE410, KYSE150) were cultured in DMEM high-glucose complete medium. The cell incubator was maintained at 5% CO₂ and 37 °C. Cells were observed regularly, and the medium was changed when appropriate.

Establishment of stable cell lines for SAP18 silencing and overexpression

KYSE410 and KYSE150 cells were seeded in 6-well plates at 5 × 10⁵ cells/well and cultured to ~50% confluence the next day for transfection. Transfection dosage was determined using multiplicity of infection (MOI = 10), calculated as MOI = (viral titer × volume)/cell count. SAP18 shRNA, and negative control (NC) lentivirus were mixed with 1 mL DMEM complete medium containing polybrene (5 μg/mL final concentration), then added to wells after thorough mixing. Cells were cultured in a 5% CO₂ incubator, with medium changed after 24 h. Stable transfectants were selected with puromycin (2 ug/mL) for 7–10 days until no cell death occurred, then used for subsequent experiments. Transfection efficiency was verified by qRT-PCR and Western blot. Lentiviral plasmids were synthesized by Gencefe Biotech Co., Ltd. using vector pLV-U6-shRNA-CMV-EGFP(2A)Puro. Target sequences: shNC: GCGTGATCTTCACCGACAAGA; shSAP18#1: GAGAAGACATGCCCACTGT; shSAP18#2: GATTGGCAGCACCATGTCT. The SAP18 overexpression (OE) plasmid and lentiviral packaging were commissioned and synthesized by Genepharma (Shanghai, China), and transfection was performed using the aforementioned method.

Antibodies and reagents

The antibodies used for western blotting (WB), IHC in this study included anti-SAP18 antibody (13841-1-AP, Proteintech, China), anti-VE-cadherin antibody (66804-1-Ig, Proteintech, China), anti-CD34 antibody (14486-1-AP, Proteintech, China), anti-PIK3CB (67121-1-Ig, Proteintech, China), anti-AKT antibody (#4691, CST, USA), anti-p-AKT antibody (#9275, CST, USA), anti-mTOR antibody (66888-1-Ig, Proteintech, China), anti-p-mTOR antibody (CY6571, Abways, China), anti-HIF-1α antibody (66730-1-Ig, Proteintech, China), anti-β-actin (66009-1-Ig, Proteintech, China).

qRT‑PCR

Total RNA was isolated from cultured cells with TRIzol Reagent (Invitrogen, USA). Specific primers and Power SYBR Green PCR Master Mix (Applied Biosystems) were used to amplify cDNA after reverse transcription of RNA samples by 5×HiScript®II qRT SuperMix (Vazyme Biotech, China). RNA expression levels were normalized to β-actin expression. The specific primers for SAP18 were as follows: 5’-CACTGTTGCTACGGGTCTTCA-3’ (sense); 5’-GCTGGACGGTACATTTCCCC-3’ (antisense).

Western blot

A total protein extraction buffer (GLPBIO, Montclair, CA, USA, #GK10023) containing protease inhibitors (GLPBIO, #GK10014) and phosphatase inhibitors (GLPBIO, #GK10013) was used to extract total protein. In order to separate protein samples, sodium dodecyl sulfate-polyacrylamide gel electrophoresis was used with PVDF membranes (Millipore, Billerica, MA, USA) as a transfer medium. 5% BSA was used to block membranes for 2 h and incubated them with primary and secondary antibodies. Blots were visualized with a Bio-Rad image system (Bio-Rad, USA), using β-actin as a control.

Co‑IP and mass spectrometry

The co-immunoprecipitation (Co-IP) assay was performed using the rProtein A/G Magnetic IP/Co-IP Kit (ACE Biotechnology, BK0004-02). Cells were lysed in pre-chilled Lysis/Wash Buffer (Enhanced), and the lysates were incubated with antibody-conjugated magnetic beads for 6 h at 4 °C. After washing with Lysis/Wash Buffer (Enhanced), the samples were sent for mass spectrometry analysis (Liquid Chromatography-Tandem Mass Spectrometry (LC-MS/MS), Jingjie PTM BioLab (HangZhou) Co., Inc.).

Matrigel tube formation assay

Matrigel tube formation assay assessed ESCC cells’ vessel-like structure formation (a key vasculogenic mimicry event). Matrigel (4 °C dissolved) was added to 96-well plates (60 μl/well) and polymerized at 37 °C for 1 h. Subsequently, 4 × 10⁴ ESCC cells in 100 μl DMEM/F12 were seeded and incubated at 37 °C with 5% CO₂. Tube formation was observed every 6 h. Following the observation of tube formation under an optical microscope, the structures were stained with Calcein-AM. Briefly, 50 μL of Calcein-AM diluted in serum-free medium (final concentration of 6.25 µg/mL) was added. The samples were then incubated in the dark at 37 °C with 5% CO₂ for 30 min. After incubation, the Calcein-AM solution was removed, and the samples were washed with PBS. The results were observed and captured under a fluorescence microscope. For quantification of VM, the number of complete tubular structures (defined as closed loops formed by cells connecting to one another) was manually counted. Each cell type was evaluated in triplicate wells, and the results were expressed as the mean number of tubes per well. All assays were performed in triplicate to ensure statistical reliability. In experiments using drug inhibition, cells received a 12-h treatment with MK2206 (10 uM), a potent Akt inhibitor.

CCK-8 cell viability and colony formation assays

Cell viability was assessed by seeding cells in 96-well plates. At time points 0, 1, 2, 3, 4, 5 days, 10 µL CCK-8 (DOJINDO, Japan) was added to wells, and absorbance at 450 nm was measured via spectrophotometer after 2 h. For the colony formation assay, cells were seeded in 6-well plates. After 7 days, the plates were washed, fixed, stained, and colonies counted.

Transwell migration and invasion assays

Transwell migration and invasion assays were performed using 8-µm pore Transwell inserts (Corning, #3422). For both assays, 4 × 10⁴ cells resuspended in basic medium were seeded into the upper chamber, with 10% fetal bovine serum in the lower chamber. Invasion assays additionally involved pre-coating the upper membrane with 60 µL Matrigel (Corning, #356234). After 18-h (migration) or 36-h (invasion) incubation, cells were fixed with 4% paraformaldehyde and stained with 0.1% crystal violet (Solarbio, China). Digital images of membranes were captured from three random fields per chamber.

Cell‑derived xenograft (CDX)

All animal experiments in this study were reviewed and approved by the Management and Ethics Review Committee of Henan Shuangyun Biotechnology Co., Ltd. (Approval No. SYLS2025045). The animal experiment facility operates under license number SYXK (Yu) 2025-0006. The study involved establishing subcutaneous xenograft and tail-vein injection metastasis models. Mice were randomly assigned to different experimental groups upon arrival or after model establishment to minimize bias. Investigators were blinded to the group allocation during data collection and analysis to ensure objectivity.

Prior to subcutaneous injection, 4-week-old nude mice (male) were placed in an induction chamber and anesthetized with isoflurane using a gas anesthesia machine to minimize distress. Mice were anesthetized in an induction chamber with 4–5% isoflurane; anesthesia was maintained with 1–2% isoflurane via face mask. Following anesthesia, the mice were injected subcutaneously with stably transfected KYSE150 cells (KYSE150-shNC/KYSE150-shSAP18) (5 × 10⁶ cells/mouse). After 4 weeks, the tumors were collected, fixed, sectioned, and stained. MK2206 was administered orally (p.o.) at a dosage of 120 mg/kg via intragastric gavage once every 4 days to mice bearing subcutaneous xenograft tumors. Mice were anesthetized with isoflurane inhalation, and euthanasia was performed via cervical dislocation following overdose anesthesia, in strict accordance with the guidelines of the Laboratory Animal Welfare Ethics Committee.

Lung metastasis in the nude mouse model

The KYSE150-shNC and KYSE150-shSAP18 cells were stably transduced with a lentiviral vector expressing firefly luciferase (Luc) to enable bioluminescence imaging. These luciferase-expressing cells were cultured until reaching the logarithmic growth phase. Subsequently, a single-cell suspension was prepared at a density of 2 × 10⁷ cells per milliliter in sterile phosphate-buffered saline (PBS) and kept on ice for subsequent injections. Prior to the tail-vein injection, nude mice (male) were placed in an induction chamber and anesthetized with isoflurane using a gas anesthesia machine. While under anesthesia, the mice were appropriately restrained, and their tails were cleansed with an alcohol swab to dilate the lateral tail veins. A volume of 100 μL of the single-cell suspension (containing 2 × 10⁶ cells) was slowly injected into each mouse via the tail vein using an insulin syringe. Four weeks post-injection, lung metastasis formation was monitored using the IVIS Spectrum. Approximately 5 min before imaging, each mouse received an intraperitoneal injection of D-luciferin (3 mg dissolved in PBS). The mice were then anesthetized with isoflurane and placed in the imaging chamber. Bioluminescence signals were acquired with an exposure time ranging from 1 to 5 min, and the data were analyzed using Living Image software to quantify metastatic burden.

Zebrafish xenograft model

Zebrafish xenograft models were established as previously described¹³. Tumor cells were washed twice with PBS and labeled with DiI (Beyotime, #C1036). At 48 h post-fertilization, zebrafish embryos were anesthetized with 1.2 mM tricaine, placed in a modified agarose gel mold. The ESCC cell concentration was adjusted to 5 × 10⁴ cells/μL using DMEM, and 5 nL of the cell suspension was injected into the perivitelline cavity of zebrafish. At 5 days post-injection, xenografts were fixed in neutral formalin, paraffin-embedded, and sectioned at 10 μm.

Establishment of patient-derived ESCC organoids (PDOs)

Fresh tumor tissues were obtained from ESCC patients undergoing curative surgery. The tissues were washed thoroughly in cold PBS supplemented with antibiotics (e.g., 1% Penicillin/Streptomycin/Amphotericin). Subsequently, the specimens were minced into small fragments (approximately 1–2 mm³) using surgical scalpels. The minced tissue fragments were digested in an appropriate volume of digestion buffer, typically consisting of Collagenase/Hyaluronidase or a solution of Dispose II and Collagenase IV, at 37 °C for 1–2 h with gentle agitation. The digested suspension was then filtered through a 70-μm or 100-μm cell strainer to remove undigested debris. The filtrate was centrifuged, and the resulting cell pellet was washed with advanced DMEM/F12 medium. The isolated primary cells were resuspended in a reduced-growth-factor Matrigel or similar extracellular matrix (ECM). Drops of the cell-ECM mixture were plated in a pre-warmed culture plate and allowed to polymerize at 37 °C for 15–30 min. After polymerization, the cultures were overlaid with a defined, human-specific organoid culture medium. The core components of this medium include advanced DMEM/F12, 1x B27 supplement, 1x N2 supplement, 10 mM HEPES, and GlutaMAX. Essential growth factors were added, such as human EGF (50–100 ng/mL), human FGF-10 (100–200 ng/mL), Noggin (100 ng/mL), and R-spondin-1 (500 ng/mL). To support stem cell growth and suppress differentiation, the medium was supplemented with small molecules including A83-01 (a TGF-β inhibitor, 0.5–2 μM), Y-27632 (a ROCK inhibitor, 5–10 μM), and Nicotinamide (10 mM). The culture medium was refreshed every 2–3 days. Organoids were subjected to passaging upon reaching a diameter of approximately 50 μm, typically 3–4 days post-seeding. Briefly, the culture medium was aspirated, and 1 mL of TrypLE (Gibco, #12605-028) was added per well of a 48-well plate to enzymatically digest the Matrigel. The Matrigel dome was gently dislodged using a 1 mL pipette tip, and the resulting suspension was transferred to a 15 mL conical centrifuge tube. Following a 5-min incubation in a 37 °C water bath, the mixture was triturated 15 times with a pipette. This cycle—comprising an additional 5-min incubation and 15 rounds of trituration—was repeated once. Complete dissociation into single cells was verified microscopically. Cells were subsequently pelleted by centrifugation at 300×g for 5 min at room temperature. After careful removal of the supernatant, the cell pellet was resuspended in fresh Matrigel and reseeded into a new 48-well plate. Complete culture medium was added only after the Matrigel had fully solidified. For histological analysis, organoids were harvested once they attained a diameter of approximately 100 μm. The Matrigel droplet containing the organoids was carefully scraped from the culture surface—taking care to preserve the integrity of the dome—and transferred to a 1.5 mL microcentrifuge tube. Samples were fixed overnight at room temperature in 4% neutral buffered formalin. Subsequent processing followed a standard graded dehydration and clearing protocol: immersion in 75% ethanol for 1 h; 85% ethanol for 0.5 h; two changes of 95% ethanol (0.5 h each); and three changes of absolute ethanol (20 min each). Samples were then cleared in two changes of xylene (15 min each), infiltrated with molten paraffin wax, and finally embedded in paraffin blocks for sectioning. Using CD34 and PAS double staining to evaluate the VM of organoid specimens.

Statistical analysis

All statistical analyses were conducted using R software (v. 4.4.1). For continuous variables satisfying assumptions of normality and homogeneity of variance, comparisons between two independent groups were performed using Student’s t-test, while comparisons involving three or more groups were performed using one-way analysis of variance (ANOVA). For continuous variables that violated these assumptions, differences between groups were assessed using the Wilcoxon rank-sum test (for two independent groups) or the Kruskal–Wallis test (for three or more groups). Associations involving categorical variables were analyzed using the chi-square test. Bivariate correlations were evaluated using Pearson’s method for parametric data or Spearman’s rank-order method for non-parametric data. Statistical significance was defined as a two-tailed P-value < 0.05.

Author contributions

R.Z., Q.Z., and Y.Q. were responsible for the conceptualization and design of the study. Data acquisition was performed by L.W. and J.G. Statistical analysis was carried out by L.W., J.G., and L.J.W. Laboratory experiments were conducted by L.W., J.D., and L.J.W. The first draft of the manuscript was prepared by L.W., with specific sections contributed by B.Y. and T.X. All authors reviewed, revised, and approved the final version of the manuscript.

Data availability

The single-cell RNA sequencing (scRNA-seq) data analyzed in this study are publicly available in the Gene Expression Omnibus (GEO) under the accession numbers GSE188900 (n = 7) and GSE196756 (n = 3) (https://www.ncbi.nlm.nih.gov/geo/). This study only utilized the ESCC tumor samples from these two single-cell datasets. Bulk RNA-seq datasets of TCGA can be obtained from the XENA data portal (http://xena.ucsc.edu/). In addition, Bulk RNA-seq datasets of ESCC can be obtained from GEO under the accession numbers GSE53625 (n = 179), GSE53624 (n = 119), and GSE53622 (n = 60) (https://www.ncbi.nlm.nih.gov/geo/). Additional single-cell data can be obtained from the corresponding author upon reasonable request.

Code availability

The code used in this study is available from the corresponding author upon reasonable request.

Competing interests

The authors declare no competing interests.

Supplementary information

The online version contains supplementary material available at 10.1038/s41698-026-01582-z.