The B7 family subgroup reflects tumor cell heterogeneity and patient post-operative prognosis in gallbladder cancer

Abstract

Purpose

To elucidate the biological heterogeneity of gallbladder cancer (GBC) cells and refine post-operative risk stratification by investigating the expression and downstream signaling of the B7-family immune checkpoint molecules CD276 (B7-H3), VTCN1 (B7-H4), and HHLA2 (B7-H7) at single-cell resolution.

Methods

Single-cell RNA sequencing (scRNA-seq) data from seven primary tumors delineated five epithelial subgroups via non-negative matrix factorization (NMF). The functional association between HHLA2 and RAC1/CDC42-PAK1-Cofilin signaling was validated by lentiviral manipulation, pharmacologic inhibition, and subcutaneous xenografts. A total of 188 surgically treated GBC patients were enrolled for survival modeling. Seven machine-learning survival algorithms were trained on five variables (CD276, VTCN1, HHLA2 expression, tumor size, and differentiation) and compared by C-index, ROC-AUC, and calibration curves.

Results

CD276 + and VTCN1 + epithelial cells displayed pro-proliferative profiles, whereas HHLA2 + cells exhibited high EMT and migration signatures. HHLA2 overexpression led to increased RAC1/CDC42-PAK1-Cofilin signaling activity, enhanced proliferation, invasion, and EMT in vitro, and accelerated tumor growth in vivo. These effects were reversed by RAC1, CDC42, or PAK1 inhibitors, as well as by CFL1 knockdown. NMF classified tumor epithelial cells into five functionally distinct subgroups. A gradient-boosting machine (GBM) model integrating expression of the three B7 molecules with tumor size and differentiation achieved superior discrimination and accurate calibration on both the training and validation sets.

Conclusion

B7-family expression delineates biologically distinct GBC subpopulations; HHLA2 promotes EMT via RAC1/CDC42-PAK1-Cofilin signaling. The GBM model incorporating CD276, VTCN1, and HHLA2 expression with tumor size and differentiation demonstrates potential for enhanced post-operative risk stratification, offering promising candidates for biomarker development and targeted therapeutic interventions.

Supplementary Information

The online version contains supplementary material available at https://doi.org/10.1186/s13062-026-00802-7.

Introduction

Gallbladder cancer (GBC), the most common malignant tumor in the biliary system, is characterized by latent onset, rapid progression, and an extremely poor prognosis [1]. The majority of patients are diagnosed at advanced stages, with only approximately 25% eligible for surgical intervention and a 5-year survival rate ranging from 5% to 15% [2]. Although radical resection remains the only potentially curative treatment option for GBC, its therapeutic efficacy is often limited by a high post-operative recurrence rate [3]. Consequently, early identification of post-operative patients with poor prognosis and timely intervention are critical strategies to improve patient survival outcomes [3].

Accurate post-operative survival prediction in cancer provides evidence-based guidance for formulating subsequent treatment plans, reducing overtreatment, and optimizing medical resource allocation. Tumor recurrence and metastasis are pivotal events that determine the prognosis of these patients. However, the underlying mechanisms remain highly complex, involving multiple signaling pathways, post-translational modifications, and metabolic reprogramming [4]. Specifically, CEACAM6 promotes tumor progression by activating the ERK/AKT signaling pathway, thereby enhancing tumor cell migration and suppressing cell adhesion [5]. The accumulation of intracellular acylcarnitine has been shown to facilitate metastasis through the lncBCL2L11/THOC5/JNK axis [6]. Moreover, lactylation of FBXO33 can activate epithelial–mesenchymal transition (EMT) to increase the migratory and invasive capacities of tumor cells [7]. Therefore, it is difficult to predict post-operative prognosis in patients with GBC. Tumor staging provides a useful framework for prognostic stratification in patients. The widely used American Joint Committee on Cancer (AJCC) staging system, which relies primarily on clinical-pathological criteria for prognostic evaluation, exhibits certain limitations in its application to GBC due to the anatomical complexity of the gallbladder and heterogeneous pattern of lymph node metastasis [8–10]. Thus, improving prognostic evaluation tools to enhance their accuracy and reliability is an urgent priority in clinical practice.

The emergence of immune checkpoint inhibitors has introduced new hope for GBC therapy, with anti-PD-1 therapies now established as first-line treatments for advanced GBC [11, 12]. PD-1 also exhibits a synergistic effect with TIM-3 in predicting post-operative prognosis among GBC patients [13]. Nevertheless, a considerable proportion of patients do not derive therapeutic benefits from anti-PD-1 therapy, highlighting the need for the development of novel immune checkpoint inhibitors [14]. The B7 family, a hopeful candidate for novel immune checkpoints, has been classified into three phylogenetic subgroups. Among these, subgroup III includes emerging immune checkpoint molecules such as CD276 (B7-H3), VTCN1 (B7-H4), and HHLA2 (B7-H7) [15]. In our previous research, we demonstrated that the expression of these three molecules correlated with poor prognosis in GBC and supported the stratification of post-operative patient outcomes [16]. In this study, we further investigated the roles of these B7-family molecules and refined patient risk stratification by integrating their expression data with clinical parameters, providing precise guidance for post-operative therapeutic management of GBC.

Materials and methods

Human specimens

We conducted a retrospective cohort study of GBC patients with pathological diagnoses. These patients underwent radical or palliative surgery at Shengjing Hospital Affiliated with China Medical University from January 2011 to October 2020. Patients who received pre-operative chemotherapy, had concurrent malignancies, incomplete clinical or pathological data, or were lost to follow-up were excluded. Ultimately, 188 GBC patients were included for survival analysis. Overall survival (OS) was defined as the time from GBC surgery to death or the end of follow-up. We also performed single-cell RNA sequencing (scRNA-seq) on primary GBC tumors resected from five patients between 2022 and 2025 (Supplementary data 1). Additionally, we downloaded scRNA-seq data of two GBC samples from public datasets [17]. The study was approved by the Ethics Committee of Shengjing Hospital, and informed consent was obtained from all patients.

Single-cell sequencing and data processing

Tissue specimens were minced and digested with 0.25% trypsin, then filtered through a 40 μm strainer. Red blood cell lysis buffer was added to the cell suspension and incubated at room temperature for 2–5 min. After centrifugation at 300 g and 4 °C for 5 min, the pellet was resuspended in PBS. The cell suspension (viability > 85%, 500–1500 cells/µl) was encapsulated with barcode-tagged magnetic beads in droplets via a microfluidic chip. Cell lysis and reverse transcription occurred within the droplets, followed by emulsion breaking and cDNA library construction for sequencing on the DNBSEQ platform.

For scRNA-seq data processing, raw gene expression matrices from the five newly sequenced tumors were obtained using the Cell Ranger Pipeline. Two additional publicly available datasets were downloaded from GSE201425. R software with the Seurat package was used to analyze filtered matrices from all seven samples.

Quality control was performed on each sample individually before integration. Cells expressing 200-7,000 genes and with < 20% mitochondrial UMIs were retained. To integrate data across multiple samples, we employed the Harmony algorithm. Briefly, individual Seurat objects were created for each sample and merged. Cell cycle scoring was performed using canonical S-phase and G2/M-phase gene lists to regress out cell cycle heterogeneity. Data were normalized, and 2,000 highly variable features were identified. We then scaled the data, regressing out mitochondrial gene percentage and cell cycle scores. Principal component analysis (PCA) was performed, followed by Harmony integration to correct for batch effects across the five newly generated and two publicly sourced datasets. Finally, t-distributed stochastic neighbor embedding (t-SNE) was performed on the Harmony-corrected embeddings for non-linear dimensional reduction. The resulting cell clusters were classified and annotated using known marker expression patterns of specific cell types.

DEG identification, functional enrichment, and gene set scoring

We employed the FindMarkers function from the Seurat package to detect differentially expressed genes (DEGs), with criteria of logFC > 0.25 and an adjusted p-value < 0.05. To explore the functions of these genes, we used Metascape (metascape.org) to perform enrichment analysis based on gene sets from the Gene Ontology (GO) and Kyoto Encyclopedia of Genes and Genomes (KEGG) databases. In parallel, the Ucell package was employed to compute gene set scores for individual cells.

Gene clustering and protein-protein interaction network construction

DEGs between HHLA2 + and HHLA2- epithelial cells were uploaded to GenClip3 (cismu.net/genclip3/analysis.php) for literature retrieval and functional enrichment based on keywords [18]. EMT-related genes among the DEGs were then selected and input into the STRING database (string-db.org) to construct a PPI network, using experimental data as the active interaction source and setting the minimum interaction score to 0.9 (highest confidence) to identify HHLA2-regulated EMT downstream signaling.

Cell culture and transfection

The GBC cell line GBC-SD (cat. no. CSTR: 19375.09.3101HUMTCHu16) was obtained from the Cell Bank of Type Culture Collection of the Chinese Academy of Sciences. The cells were maintained in RPMI 1640 medium (Gibco; Thermo Fisher Scientific, Inc.) containing 4.5 g/L glucose, 2 mM L-glutamine, and 10% fetal bovine serum (FBS) (Gibco; Thermo Fisher Scientific, Inc.) at 37 °C with 5% CO₂.

Lentiviruses encoding shRNA for HHLA2 silencing and non-target shRNA, as well as the HHLA2 overexpression vector (GenBank accession no. NM_007072) and empty control lentiviral vector, were designed and purchased from Shanghai GeneChem Co., Ltd. GBC-SD cells were seeded in six-well plates for transfection. After 24 h of incubation, the medium was replaced, and puromycin was used to select stable infected cell lines. Transfection efficiency was verified by RT-PCR and Western Blot.

To inhibit Cofilin expression in HHLA2-overexpressing cells, a transfection reagent mixture containing 20 pmol of CFL-1 siRNA (forward: 5′-CCUCUAUGAUGCAACCUAUTT‐3′; reverse: 5′‐AUAGGUUGCAUCAUAGAGGTT‐3′) and 2 µL of LipoGene 2000 Star was added to the stably transfected GBC-SD cell lines.

Cell proliferation assays

We performed EdU assays (cat. no. C10310-3; Guangzhou RiboBio Co., Ltd.) to evaluate cell proliferation. Cells were seeded in 24-well plates at a density of 1 × 10⁵ cells per well. Proliferation was assessed at 48 h following the manufacturer’s protocol.

Scratch assay

Cells were seeded in 6-well plates at a density of 6 × 10⁵ cells per well and cultured overnight. Once the cultures reached 85% confluency, the cell monolayer was scratched using a sterile 10-µL pipette tip. Subsequently, the cells were incubated in culture medium without FBS. Images of the scratch were captured at 0, 24, and 48 h using an Olympus IX71 inverted microscope at 100× magnification. The width of the scratch was analyzed using Olympus CellSens Dimension software.

Cell migration and invasion assays

Cell migration and invasion assays were conducted using Transwell chambers (Corning Inc., USA) according to the manufacturer’s protocol. A total of 4 × 10⁴ cells were seeded in the upper chambers, with complete medium in the lower chambers. After 48 h of culture, cells were fixed with methanol and stained with Giemsa. An Olympus DP72 microscope at 100× magnification was used to examine cells on the lower surface. Images of four random fields were captured to count the cells, and the average number of migrating cells was calculated to assess migration ability.

Pull-down assay

We performed pull-down assays using the Activation Assay Biochem kit (Cytoskeleton, Denver, CO) in accordance with the manufacturer’s protocols. Cell lysates were incubated with PAK-PBD beads at 4℃ for 1 h to detect active RAC1 and CDC42. After SDS-PAGE separation on 15% polyacrylamide gels, proteins were transferred to PVDF membranes and probed for RAC1 or CDC42.

Western blotting

Transfected GBC-SD cells were lysed in RIPA buffer, and the lysate was centrifuged at 4 °C for 15 min at 12,000 rpm to eliminate cellular debris. After electrophoresis, equal quantities of total protein were loaded onto 15% SDS-polyacrylamide gels and transferred onto PVDF membranes (Millipore). The membranes were blocked with a 5% skimmed milk solution in Tris-buffered saline containing 0.1% Tween-20 for 2 h. Subsequently, they were incubated with primary antibodies overnight at 4 °C, followed by 2 h of incubation with secondary antibodies at room temperature.

Primary antibodies utilized were rabbit anti-SMA (Abcam, ab124964), rabbit anti-Vimentin (Abcam, ab92547), rabbit anti-E-cadherin (Abcam, ab40772), rabbit anti-N-cadherin (Abcam, ab18203), rabbit anti-Collagen-I (Abcam, ab34710), rabbit anti-HHLA2 (Abcam, ab14327), rabbit anti-RAC1(Abcam, ab33186), rabbit anti-CDC42 (Abcam, ab187643), rabbit anti-PAK1 (Abcam, ab223849), and rabbit anti-Cofilin (Abcam, ab42824). The secondary antibody was peroxidase-conjugated goat anti-mouse/rabbit IgG (ZB-2305). Commercially obtained inhibitors included NSC23766 (Selleck, S8031), ZCL278 (Selleck, S7293), and IPA-3 (Selleck, S7093).

Animal studies

Five-week-old male BALB/c nude mice (weighing 16–20 g) were obtained from Beijing Weitonglihua Experimental Animal Technology Co., Ltd. Each mouse received a subcutaneous injection of 2 × 10⁶ cells into the upper flank. Tumor growth was monitored every 48 h using calipers, with volume calculated by the modified ellipsoid formula: 1/2 (length × width²). Mice were randomly divided into three groups (n = 3): control GBC-SD-HHLA2, GBC-SD-HHLA2 with NSC23766, and GBC-SD-HHLA2 with ZCL278. Each treatment group received 50 mg/kg of the respective inhibitor via intraperitoneal injection every 48 h. Mice were anesthetized with isoflurane (4% induction, 2% maintenance) and imaged using the AniView 600 multimodal imaging system (Boluteng, Guangzhou). Following imaging, mice were euthanized via CO2 inhalation.

Non-negative matrix factorization

We merged all significantly upregulated DEGs linked to CD276, VTCN1, and HHLA2, constructing a core gene set of 1,069 genes. Subsequently, we extracted the expression values of these genes from the single-cell gene expression matrix and removed genes with no expression across all samples to reduce noise. Non-negative matrix factorization (NMF) was performed on the data matrix using the Brunet algorithm from the NMF package, with a rank of 5 and 30 runs to ensure result stability, thereby generating 35 gene programs. The AUCell package was employed to score each gene program, quantifying its expression patterns across different cells. By computing the correlation matrix of AUCell scores between gene programs and applying hierarchical clustering (Ward. D2 method), we classified the gene programs and visualized the results with a correlation heatmap. Besides, we also performed NMF on the AUCell scores of each gene program to investigate the effects of different ranks. After identifying the optimal rank, epithelial cells were classified accordingly.

InferCNV

We performed CNV analysis on epithelial cells from GBC using inferCNV (version 1.18.1), with normal gallbladder epithelial cells as the reference group. To evaluate the malignancy of each epithelial cell, we calculated a CNV score for each cell. This score was defined as the mean of the squared deviations in gene expression.

Cell-cell communication analysis

We employed the CellChat package to analyze cell-cell interactions between epithelial cells and T cells in the dataset. CellChat integrates gene expression data with the CellChatDB.human database to evaluate the probability of cell–cell communication. Normalized count data were used to create CellChat objects, and preprocessing functions recommended by the package were applied to analyze individual datasets with default parameters. All ligand-receptor interaction categories from the database were included, and ligand-receptor pairs were filtered using a statistical significance threshold of p < 0.01.

Immunohistochemistry

Formalin-fixed paraffin-embedded (FFPE) specimen slides (3 μm) were prepared for immunohistochemical analysis. Sections were deparaffinized, subjected to antigen retrieval, and treated with 3% hydrogen peroxide for peroxidase blocking. After blocking with goat serum, sections were incubated with primary antibodies against CD276 (Abcam, ab105922), VTCN1 (Abcam, ab252438), and HHLA2 (Abcam, ab214327) overnight at 4 °C. Following washing with PBS, sections were incubated with the secondary antibody for 1 h. Visualization was performed with 3,3′-diaminobenzidine (DAB) for 3 min. Nuclei were stained with Harris hematoxylin. Sections were dehydrated and sealed with neutral gel.

Development and validation of the post-operative prognostic model

We employed a 5-fold cross-validation strategy to partition the original dataset into training and testing subsets. In this process, 20% of the data were randomly allocated to the testing set for model validation, while the remaining 80% were retained as the training set to build survival models. The partitioning process was carried out using the createFolds function to ensure the uniform distribution of individuals across the folds.

Using the training dataset, we constructed seven advanced survival analysis models with parameters preset or optimized via cross-validation, including a Random Survival Forest (RSF) with 1,000 trees, a node size of 5, and two variables randomly selected at each split; a Gradient Boosting Machine (GBM) with 1,000 trees, a maximum tree depth of 5, and a learning rate of 0.01; CoxBoost with the penalty parameter optimized by 3-fold cross-validation and the maximum number of boosting steps set to 200; a Survival Support Vector Machine (Survival-SVM) with a linear kernel and a regularization parameter of 0.1; XGBoost based on the Cox proportional hazards model, with a maximum tree depth of 3 and a learning rate of 0.03; SuperPC with three components, a minimum of five features, and 10-fold cross-validation for optimal threshold selection; and Partial Least Squares Regression-Cox regression (PLSR-Cox), in which 10-fold cross-validation was used to determine the number of components, and three latent components were retained [19–24]. The outcomes of interest were survival status and survival time. After model fitting, each algorithm generated an individual risk score for each patient in both the training and testing cohorts, with higher scores indicating a greater risk of the event.

Model performance was evaluated across multiple dimensions. First, by integrating the predicted values of the training and testing sets for each model, we calculated C-index values for each model using the pec::cindex function to assess the accuracy of sample risk ranking. Second, the ROC curves and their corresponding AUC values for each model were calculated using the pROC::roc function and visualized with the ggroc function. These curves illustrated the model’s sensitivity and specificity at different time points. Finally, calibration curves were plotted using the cph and calibrate functions to verify the consistency between predicted survival probabilities and actual observations.

Statistics

We conducted statistical analysis of scRNA-seq and clinical data using R software. Based on the central limit theorem, a t-test was applied to compare gene set enrichment and CNV scores between the two groups in scRNA-seq data. The statistical methods for comparing patient clinical data were provided by the gtsummary package.

Results

Analysis of the roles of CD276, VTCN1 and HHLA2 in gallbladder cancer

We performed scRNA-seq analysis on primary tumor samples from seven GBC patients to explore gene functions. After quality control, we identified various cell types: 15,945 epithelial cells, 78,044 T/NK cells, 11,890 B/plasma cells, 18,111 myeloid cells, 10,581 mesenchymal cells, 1,827 endothelial cells, 2,581 mast cells, and 4,274 undifferentiated cells (which did not express any classical cell markers) (Fig. 1A, B).To explore the functions of CD276, VTCN1, and HHLA2, we dichotomized epithelial cells into positive and negative subgroups. The negative subgroup comprised cells with 0 UMI counts (undetected), while the positive subgroup comprised cells with non-zero counts (> 0, detectable expression), and differential expression analysis was subsequently performed between these subgroups. The results showed 1,069 DEGs in CD276 + cells, 1,203 DEGs in VTCN1 + cells, and 884 DEGs in HHLA2 + cells (Fig. 1C). GO and KEGG enrichment analyses demonstrated that the upregulated DEGs in CD276 + and VTCN1 + cells were associated with protein synthesis, assembly, and transport (Fig. 1D, E). As protein synthesis is crucial for tumor cell proliferation, these cells exhibited higher mitosis scores (Fig. 1G, H). In contrast, HHLA2 + cells showed upregulated DEGs involved in epithelial cell differentiation and migration, and displayed higher mitosis and EMT scores (Fig. 1F, I, J).

Fig. 1.

Single-cell transcriptomic profiling identified distinct epithelial subpopulations and HHLA2-induced signaling in GBC. (A) t-SNE plot visualizing 8 cell types from 7 GBC samples. (B) Dotplot visualizing expression of gene signatures among identified cell types. (C) Volcano plot of DEGs between positive and negative epithelial cells. (D) GO and KEGG enrichment analysis for the DEGs between CD276 + cells and CD276- cells. (E) GO and KEGG enrichment analysis for the DEGs between VTCN1 + cells and VTCN1- cells. (F) GO and KEGG enrichment analysis for the DEGs between HHLA2 + cells and HHLA2- cells. (G) Mitosis scores between CD276- and CD276 + epithelial cells. (H) Mitosis scores between VTCN1- and VTCN1 + epithelial cells. (I) Mitosis scores between HHLA2- and HHLA2 + epithelial cells. (J) Mitosis scores between HHLA2- and HHLA2 + epithelial cells. (K) Box plot of RAC1, CDC42, PAK1 and CFL1 expression between HHLA2- and HHLA2 + epithelial cells. (L) PPI network of EMT-related genes. (M) Westen blot demonstrated the RAC1/CDC42-PAK1-Cofilin pathway in GBC-SD cell lines

Our previous study showed that HHLA2 promoted GBC cell migration, proliferation, and EMT, but its downstream signaling remains unclear. To explore this, we analyzed DEGs from HHLA2 + cells using GenClip3 and identified EMT-related genes (Supplementary data 2). Then, we constructed a PPI network with the highest confidence (0.9) using these genes and found that RAC1/CDC42-PAK1, a component of the TGF-β pathway, which is a classic signaling pathway for EMT activation, was present in this PPI [25] (Fig. 1L). HHLA2 + cells exhibited the higher expression of RAC1, CDC42, PAK1, and downstream Cofilin (CFL1) (Fig. 1K). We used the Western blot to confirm the results. As the upregulation of HHLA2 in GBC-SD cell lines, the expression of RAC1, CDC42, PAK1, and Cofilin also increased, indicating that HHLA2 may act as an upstream modulator of this pathway (Fig. 1M).

HHLA2 promotes EMT via upstream modulation of RAC1/CDC42-PAK1-Cofilin signaling

To further confirm that HHLA2 drives GBC cell proliferation, migration, and invasion via the RAC1/CDC42-PAK1-Cofilin pathway, we treated HHLA2-overexpressing GBC-SD cells with the RAC1 inhibitor NSC23766 (50µM) or the CDC42 inhibitor ZCL278 (50µM). Matrigel and Transwell assays revealed that NSC23766 and ZCL278 significantly reduced cell migration and invasion (Fig. 2A). EdU assays also demonstrated that these inhibitors suppressed cell proliferation (Fig. 2B). Moreover, scratch assays indicated that wound healing was slower in HHLA2-overexpressing cells treated with NSC23766 or ZCL278 compared to cells transfected with pCDNA-HHLA2 alone (Fig. 2C). We further investigated the impact of RAC1/CDC42 pathway inhibition on tumor development in vivo. GBC-SD cells stably overexpressing HHLA2 via lentivirus transfection were injected into nude mice to induce tumor formation. Treatment groups were administered intraperitoneal injections of NSC23766 or ZCL278 at optimal concentrations. Compared to the control group, tumor growth was significantly reduced in both treatment groups (Fig. 2D, E). These findings suggest that HHLA2 acts as an upstream regulator of RAC1 and CDC42 to promote GBC development.

Fig. 2.

Pharmacological inhibition of RAC1 and CDC42 attenuated HHLA2-mediated malignant phenotypes in GBC cells and suppressed tumor growth in vivo. (A) NSC23766 and ZCL278 inhibited HHLA2-induced migration and invasion in GBC-SD cells (B) NSC23766 and ZCL278 suppressed HHLA2-induced proliferation in GBC-SD cells (C) NSC23766 and ZCL278 restrained HHLA2-induced migratory capacity in GBC-SD cells (D) NSC23766 and ZCL278 decelerated tumor growth rate in mouse models (E) Growth curves of xenograft tumors

EMT is characterized by the downregulation of epithelial markers such as E-cadherin and the upregulation of mesenchymal markers such as N-cadherin, vimentin, α-SMA and Col-I. When cells transfected with pCDNA-HHLA2 were treated with NSC23766 or ZCL278, their expression of E-cadherin was elevated, while the expression of N-cadherin, vimentin, α-SMA and Collagen-I were reduced. HHLA2 overexpression could promote the expression of PAK1 and Cofilin, yet NSC23766 or ZCL278 could suppress this effect (Fig. 3A, B). Similarly, treatment with IPA-3 (20 µM), a PAK1 inhibitor, also increased E-cadherin expression and decreased N-cadherin, vimentin, α-SMA and Collagen-I levels, and reduced Cofilin expression in HHLA2-overexpressing cells (Fig. 3C). Silencing CFL1, a downstream molecule of this pathway, could also inhibit HHLA2 overexpression-induced EMT (Fig. 3D). Collectively, these findings strongly supported the hypothesis that HHLA2 induced EMT via the upstream modulation of RAC1/CDC42-PAK1-Cofilin signaling, thereby promoting the progression of GBC.

Fig. 3.

Disruption of the RAC1/CDC42-PAK1-Cofilin axis blocked HHLA2-driven EMT in GBC cells. (A) NSC23766 inhibited HHLA2-induced EMT and downstream signaling pathway. (B) ZCL278 inhibited HHLA2-induced EMT and downstream signaling pathway. (C) IPA3 inhibited HHLA2-induced EMT and downstream signaling pathway. (D) Siliciencing CFL1 inhibited HHLA2-induced EMT. Uncropped blot images are available in the Supplementary Material (Original data of Western Blot)

Classification of GBC cells based on CD276, VTCN1, and HHLA2

The heterogeneity of tumor epithelial cells complicates their classification in scRNA-seq analysis. Proliferation and migration capabilities serve as pivotal indicators for evaluating the malignancy of tumor cells, linked to tumor invasiveness and patients’ prognosis [26]. Our study showed that epithelial cells expressing CD276, VTCN1 and HHLA2 expressed DEGs regulating these processes. This suggested that the expression patterns of these genes could be used to classify epithelial cells and identify the functional characteristics of different subgroups.

We selected NMF to identify latent expression programs of DEGs and dissect epithelial heterogeneity. NMF decomposes the expression matrix into two non-negative matrices, yielding a parts-based representation with improved biological interpretability. Compared with traditional clustering methods such as k-means and hierarchical clustering, NMF accommodates non-negative data and can reveal overlapping or graded expression programs and biologically meaningful substructures [27]. By applying NMF to our scRNA-seq dataset, we stratified the expression program of the DEGs into two distinct functional categories: one mainly involved in protein synthesis and the other closely related to cell adhesion and various pathological processes (Fig. 4A, B). Subsequently, NMF decomposition was utilized to classify epithelial cells based on gene program scores. To determine the optimal number of subgroups, NMF was performed with ranks ranging from 2 to 10. Based on the NMF rank survey, k = 5 was selected because it provided the best balance between clustering stability and model fit, with a relatively high cophenetic coefficient, acceptable silhouette width, and diminishing improvement in RSS and residuals beyond this point. Consequently, epithelial cells were stratified into five distinct subgroups for subsequent analyses (Fig. 4C) [28].

Fig. 4.

NMF identified distinct gene programs and functional heterogeneity in GBC epithelial subpopulations. (A) Correlation matrix of gene program scores. (B) Enrichment analysis of gene program groups. (C) Evaluation of NMF factorization ranks across multiple metrics. (D) Enrichment analysis of signature genes for epithelial cell subgroups. (E) Mitosis scores of epithelial cell subgroups (F) EMT scores of epithelial cell subgroups. (G) CNV scores of epithelial cell subgroups (H) Metabolic analysis of epithelial cell subgroups. (I) Cell communications between epithelial cell subgroups and T cell Subgroups. (J) Interaction pairs between epithelial cell subgroups and T cell Subgroups

Group 1 was characterized by active energy metabolism, robust proliferation, and EMT potential. Group 2 was mainly associated with immunomodulatory functions. Group 3 exhibited high ribosomal gene expression, reflecting active protein synthesis, as well as strong proliferation capacity and migration potential. Group 4, with secretory functions and bile acid metabolism capability, had the lowest CNV score, resembling normal gallbladder epithelial cells. However, its remarkably high EMT score suggests a transitional state toward malignancy, consistent with EMT’s role in GBC transformation [29]. The fifth group was enriched for genes associated with cell adhesion and cytoskeleton organization. Its high CNV score, mitotic score, and EMT score collectively suggest the most aggressive malignant characteristics among all groups (Fig. 4D-H). Cell communication analysis revealed that group 1, 3, 4, and 5 predominantly presented antigens to exhausted T cells rather than CTLs. Moreover, the NECTIN2-TIGIT interaction was identified in the cell communication between these cells and exhausted T cells (Fig. 4I, J) (Supplementary data 3). This interaction can synergize with PD-L1/PD-1 signaling to suppress T and NK cell cytotoxicity, attenuating their tumor-killing capacity [30].

Classification of post-operative prognosis in GBC patients based on CD276, VTCN1, and HHLA2

In previous research, we established that the B7 score, incorporating CD276, VTCN1 and HHLA2, effectively stratifies the post-operative prognosis of GBC patients (Supplementary data 4). Furthermore, a model combining this score with clinical stage exhibits exceptional survival discrimination [15]. However, accurate assessment of the TNM stage for gallbladder cancer poses a significant challenge [8–10]. Therefore, we refined our model by integrating five key indicators: the expression levels of CD276, VTCN1, and HHLA2, along with the size and differentiation of the primary tumor. Using these indicators, we constructed predictive models through seven machine learning algorithms, including RSF, GBM, CoxBoost, SurvivalSVM, XGBoost, SuperPC, and PLSR-Cox. The results indicated that GBM demonstrated superior discriminative performance, achieving 1-, 3-, and 5-year AUCs of 0.826, 0.900, and 0.909 in the training cohort (Fig. 5D-F), and 0.880, 0.915, and 0.900 in the test cohort (Fig. 5J-L), with consistently high C-index values (Fig. 5B, H). The calibration curves of the GBM model in both the training and testing sets indicated a robust alignment between the predicted and observed survival rates at 1, 3, and 5 years post-surgery (Fig. 5C, I). Although RSF and XGBoost exhibited comparable performance during training (Fig. 5A-F), GBM showed more robust generalizability in the test set (Fig. 5G-L).

Fig. 5.

Performance of multiple machine learning algorithms for post-operative prognostic prediction in gallbladder cancer. (A) AUC values of various machine learning models across time points in the training set. (B) C-index of various machine learning models across time points in the training set. (C) Calibration curve of the GBM model in the training set. (D) ROC curves of various machine learning models predicting 1-year survival in the training set. (E) ROC curves of various machine learning models predicting 3-year survival in the training set. (F) ROC curves of various machine learning models predicting 5-year survival in the training set. (G) AUC values of various machine learning models across time points in the test set. (H) C-index of various machine learning models across time points in the test set. (I) Calibration curve of the GBM model in the test set. (J) ROC curves of various machine learning models predicting 1-year survival in the test set. (K) ROC curves of various machine learning models predicting 3-year survival in the test set. (L) ROC curves of various machine learning models predicting 5-year survival in the test set

These findings suggested that, within our low-dimensional framework comprising variables with established pathobiological significance, GBM effectively captured non-linear relationships and variable interactions while balancing model fit and generalization.

In conclusion, the GBM model, based on the expression of CD276, VTCN1, and HHLA2, as well as tumor size and differentiation degree, is capable of accurately stratifying the post-operative prognosis of GBC patients, positioning it as a valuable asset for clinical risk evaluation.

Discussion

Immunotherapy has significantly reshaped cancer treatment. Research and drug development targeting PD-1, PD-L1, and CTLA-4 immune checkpoints have achieved breakthrough progress and shown success in clinical treatment of multiple tumors types [31]. However, its response rate remains limited, and monotherapy targeting a single checkpoint tends to induce drug resistance, prompting researchers to explore new immune checkpoint molecules [31–33]. The B7 family has emerged as a promising candidate for novel immune checkpoints due to its role in providing co-stimulatory or co-inhibitory signals for immune cell activation [34]. Recently, the third subgroup of B7 family members, CD276, VTCN1, and HHLA2, has garnered significant attention for their regulatory roles in T cells and NK cells [34].

CD276 is an inhibitory immune checkpoint molecule that is broadly expressed across diverse tumor types and impedes T cell-mediated tumor suppression [35]. While its specific mechanism of action remains unclear due to the unidentified receptor, its established immune-suppressive and pro-tumorigenic roles have driven its entry into clinical trials [35, 36]. In lung cancer therapy, CD276-targeted antibody-drug conjugates (ADCs) have shown promising benefits [34]. Ifinatamab deruxtecan (I-DXd, also known as DS-7300a) demonstrated an objective response rate (ORR) of 52.4% in patients with small cell lung cancer in phase I/II clinical trials and has advanced to phase III trials [37]. Similarly, VTCN1, another immune-suppressive molecule with an unidentified receptor, inhibits T cell proliferation and cytotoxicity, and promotes T cell apoptosis when expressed on antigen-presenting cells [38]. The ADC targeting VTCN1, SGN-B7H4V, has achieved an ORR of 18.5% in biliary tract cancer [39]. Compared to CD276 and VTCN1, research on HHLA2 is still in its early stages. HHLA2 exerts opposing immune functions depending on its ligand interactions. HHLA2 promotes T cell responses via the HHLA2/TMIGD2 axis. Conversely, it suppresses T cell activation and proliferation through the HHLA2/KIR3DL3 axis by inhibiting T cell receptor-mediated signaling [40]. This dual function leads to inconsistent relationships between HHLA2 expression and patient prognosis across different tumor types. Our prior studies reveals that T cells expressing TMIGD2 exhibit reduced cytotoxicity, which may provide an explanation for the dominance of HHLA2’s immunosuppressive role in certain tumors [41]. Drugs targeting HHLA2 are under development, and their efficacy awaits further clinical evaluation [42].

These immune checkpoints also promote the malignant behaviors of tumor cells. CD276 activates signaling pathways such as PI3K-AKT-STAT3, JAK-STAT, HIF1α, and NF-κB, significantly enhancing tumor cell proliferation, metastasis, EMT, and cancer stemness [43–46]. VTCN1 promotes tumor progression by activating ERK1/2 and Wnt signaling and suppresses apoptosis via Bax, Bcl-2, and caspase 3 [47–50]. HHLA2 enhances EGFR-MAPK-ERK, JAK-STAT, and PI3K-AKT-mTOR pathways as well as NF-κB signaling, leading to increased cell proliferation, migration, and invasion [51–53]. Our prior studies confirmed that the expression of these three molecules correlated with poor prognosis and more aggressive clinical pathological stages in GBC patients, and that HHLA2 promoted EMT in GBC cells [16, 54]. In this study, we demonstrated that HHLA2 acted upstream of RAC1/CDC42-PAK1-Cofilin signaling to facilitate this malignant progression.

However, we acknowledge that this pathway requires further refinement. Tumor progression is not driven solely by canonical intracellular signaling, but also by coordinated programs involving extracellular matrix remodeling and metastatic niche formation [4]. Recent studies have shown that matrix metalloproteinases (MMPs), particularly MMP1 and MMP2, are closely associated with tumor invasion and metastasis. MMP1 is highly expressed in multiple tumor types, correlates with metastatic potential, and contributes to an immunosuppressive tumor microenvironment characterized by increased macrophage infiltration and impaired CD8 + T cell function [55]. In addition, JNK/C-Jun-dependent control of MMP1 expression has been shown to influence migratory and invasive behavior, although this observation was derived from a non-neoplastic trophoblast model and should therefore be interpreted cautiously in the cancer setting [56]. Consistently, MMP2-associated invasive signaling has also been linked to metastatic progression: The LRP1-SNRNP25/pJNK/37LRP/MMP2 axis promotes osteosarcoma cell migration, invasion, and distant metastasis [57], whereas ZNF24 suppresses colorectal cancer growth and metastasis through transcriptional repression of MMP2 [58]. Biomechanical cues in confined microenvironments also enhance tumor invasiveness via membrane topology changes and cytoskeletal remodeling [59]. This involves invadopodia, actin-rich membrane protrusions dedicated to extracellular matrix degradation. MMPs drive the maturation of these invasive structures, representing a critical mechanism for tumor invasion [60]. Furthermore, metastatic dissemination is also modulated by the pre-metastatic niche milieu, with microbiota-related factors, stemness-associated cues, immune cell interactions, and extracellular vesicle-mediated signaling collectively contributing to this process [61]. Therefore, the TGF-β pathway integrates multiple control mechanisms as a pivotal regulator of EMT. It is also regulated by upstream transcription factors. For example, RGS3 promotes SMAD2/3 phosphorylation through direct interaction with ARID3B, activating EMT to drive ovarian cancer cell proliferation and metastasis [62]. The RAC1/CDC42-PAK1-Cofilin axis represents a component of the non-canonical TGF-β pathway. RAC1 and CDC42 are members of the Rho GTPase that belong to the Ras superfamily of small GTP-hydrolyzing enzymes [63, 64]. As a member of the B7 family, HHLA2 has no identified domains for direct interaction with RAC1 and CDC42. The underlying mechanisms appear to be multifaceted, potentially involving transcriptional factor regulation downstream of HHLA2 as well as HHLA2-mediated remodeling of TME, and require further investigation.

The expression of B7 family molecules in tumor cells reflects not only their capacity for immune evasion but also their abilities to proliferate, migrate, and differentiate. Furthermore, these molecules exhibit synergistic interactions and co-expression within the tumor microenvironment (TME) [42]. Based on these observations, we hypothesize that downstream genes regulated by the B7 family can be leveraged to classify tumor cells. By analyzing DEGs associated with the upregulation of CD276, VTCN1, and HHLA2, we categorized tumor cells into the following clusters: cluster 1 (Metabolically Active), cluster 2 (Immunomodulatory), cluster 3 (Ribosomal-Protein Synthesis), cluster 4 (Transformational State), and cluster 5 (Adhesion-Cytoskeleton). Through cellular communication analysis, we found that, except for cluster 2, the majority of epithelial cells primarily present antigens to PD-1-high exhausted T cells and further suppress T cell function via the NECTIN2-TIGIT axis. This synergistic mechanism may explain the suboptimal therapeutic outcomes of monotherapy with anti-PD-1 in GBC. Dual immune checkpoint inhibition targeting both TIGIT and PD-1 shows promise for enhancing therapeutic outcomes in GBC treatment [65].

Considering the prognostic significance of B7 family molecule expression, we previously developed a predictive model for patients by correlating the expression of B7 family molecules with tumor clinical staging [16]. However, we found that relying on clinical pathological staging for GBC is not straightforward in clinical practice. The proximity to multiple organs complicates the T-stage determination of GBC. The balance among radical surgery, surgical trauma, and procedural complexity makes it challenging to accurately assess lymph node metastasis [8–10]. While PET-CT is the primary clinical tool for determining distant metastasis, its high cost and radiation exposure frequently lead patients to decline this examination. Since our previous research demonstrated that CD276, VTCN1, and HHLA2 expression correlated with tumor TNM staging [41], we proposed that patient survival could be predicted based on the expression of these three molecules. In this study, we integrated their expression levels with primary tumor size and tumor cell differentiation degree to construct a machine learning model for prognosis prediction, achieving satisfactory results. Given the limited risk stratification afforded by conventional staging in GBC, this model could facilitate the identification of patients requiring intensified surveillance, adjuvant therapy evaluation, or clinical trial enrollment. Notably, while currently validated as a post-operative stratification tool (using resected specimens), the GBM model holds translational potential for pre-operative risk assessment. For suspected GBC patients, immunohistochemical profiling of B7 molecules and assessment of tumor cell differentiation could theoretically be performed on ultrasound-guided percutaneous transhepatic gallbladder aspiration (PTGBA) or endoscopic ultrasound-guided fine-needle biopsy (EUS-FNB) specimens, combined with imaging assessments of tumor size to predict post-operative survival [66]. However, substantial intra-tumoral heterogeneity necessitates further investigation into the reproducibility of biomarker detection in small biopsy samples and the concordance between pre-operative and post-operative measurements before clinical implementation. Besides, due to the limited sample size, the model still requires refinements to enhance its performance. Given the low incidence and highly malignant biological behavior of GBC, developing accurate prognostic models to optimize treatment plans is essential and requires multi-center collaboration. Our study offers a potential approach to addressing this challenge in the future.

Limitations

This study is limited by the exclusive use of the GBC-SD cell line for mechanistic validation of the RAC1/CDC42-PAK1-Cofilin pathway, as well as the lack of independent patient cohort validation. These constraints may compromise the clinical generalizability and translational significance of our findings.

Electronic supplementary material

Below is the link to the electronic supplementary material.

Author contributions

C.M. designed the study and wrote the manuscript. C.M., H.H. (Huixin), Y.L., and C.Z. performed experiments. Y.Z., S.X., and C.L. contributed to patient follow-up. Y.D. and H.H. (Hangqi) built and evaluated the machine learning model. C.L. and Z.C. acquired single-cell data. Y.T. proofread and supervised the writing of the original manuscript. All authors reviewed the manuscript.

Funding

The study was supported by the Natural Science Foundation of China (81974377), Funding Project of Supporting the High-Quality Development of China Medical University by the Department of Science and Technology of Liaoning Province (2023JH2/20200128), 345 Talent Project of Shengjing Hospital (2023–2025), and the Outstanding Scientific Fund of Shengjing Hospital.

Data availability

The datasets used and analyzed during the current study are available from the corresponding author on reasonable request.

Declarations

Ethics approval and consent to participate

This study was approved by the Shengjing Hospital (2024PS710K) and was conducted respecting the Declaration of Helsinki. Each patient has signed an informed consent to participate in this research.

Consent for publication

All participants agreed to the publication of their data as part of the research results.

Competing interests

The authors declare no competing interests.