A novel signature predicts prognosis in pancreatic cancer based on tumor membrane-associated genes

Abstract

Background

Pancreatic cancer is one of the leading causes of tumor-related mortality, characterized by short patient survival times and limited treatment options. Some targeted therapies have not succeeded in improving patient prognosis. Tumor membranes possess potential target specificity, offering hope for enhancing the efficacy of immunotherapy and drug treatment.

Methods

In this study, we collected gene expression and survival data from two scRNA-seq projects and patient cohorts in TCGA, ICGC, and GEO. Differential analysis and dimensionality reduction clustering were employed to isolate tumor epithelial cells. High-expression membrane-associated genes in tumor epithelial cells were identified through PPI network analysis and functional enrichment. Subsequently, membrane-associated genes associated with patient prognosis were selected using LASSO and Cox regression to construct MaGPS, which was validated in external datasets. Potential therapeutic targets of the MaGPS signatures were identified and confirmed by integrating spatial transcriptomics, scRNA-seq, and protein expressions. In addition, drug sensitivity analysis was performed to explore potential targeted drugs associated with MaGPS.

Results

The results demonstrated the identification of a specific tumor epithelial cell cluster, c0. This cluster expressed 17 membrane-associated genes that are closely interconnected and play roles in extracellular interactions. The MaGPS model, developed based on the membrane-associated genes LMO7, APOL1, SLC2A1, C15orf48, FXYD3, and CLDN18, effectively predicted patient prognostic risk. Additionally, the expression of the six MaGPS signatures was observed to be elevated in tumors at both the protein expression and spatial transcriptomics levels. Furthermore, drug sensitivity analysis revealed that the MaGPS signature scores were significantly associated with the sensitivity to 38 different drugs, highlighting potential targeted therapies related to MaGPS.

Conclusion

The MaGPS model, based on bulk RNA-seq, scRNA-seq, and spatial transcriptomics data, effectively evaluated the prognosis of pancreatic cancer and provided valuable insights for better therapeutic targets.

Highlights

• •First to integrate bulk RNA-seq, scRNA-seq, and spatial transcriptomics to develop the prognostic MaGPS model for PDAC.
• •The MaGPS model, based on six key membrane-associated genes (LMO7, APOL1, SLC2A1, C15orf48, FXYD3, CLDN18), predicts prognostic risk in PDAC patients.
• •ClinicalImpact: The MaGPS model identifies potential therapeutic targets, offering new directions for PDAC treatment.

1. Introduction

Pancreatic cancer, particularly pancreatic ductal adenocarcinoma (PDAC), is among the most aggressive and lethal malignancies [1]. Early diagnosis of PDAC is challenging, and it is often detected at an advanced stage, resulting in a dismal prognosis with a five-year survival rate of less than 10 % [2]. Current treatment modalities for pancreatic cancer include surgery, chemotherapy, radiotherapy, and targeted immunotherapy. Surgery is viable for only a small number of early-stage patients [3]. First-line chemotherapy regimens include gemcitabine and FOLFIRINOX [4,5]. However, the response to immunotherapy is limited due to the dense stromal tissue and immunosuppressive tumor microenvironment (TME) in PDAC [6,7]. Therefore, new diagnostic and therapeutic strategies for PDAC are urgently needed.

Tumor membrane proteins play crucial roles in tumor proliferation, migration, and immune evasion. Zheng et al. reported that low-dose graphene oxide promotes tumor cell proliferation through integrin αV-mediated activation of the PI3K-AKT-mTOR signaling pathway [8]. Amin et al. demonstrated that upregulation of epithelial membrane protein 1 (EMP1) increases cancer cell migration, leading to tumor metastasis [9]. Song et al. described the overexpression of programmed death-ligand 1 (PD-L1) in tumors, highlighting the role of the PD-L1/PD-1 immune checkpoint in T-cell exhaustion, which facilitates tumor immune evasion [10]. Research targeting membrane proteins, such as PD-L1 and Epidermal growth factor receptor (EGFR), in the treatment of PDAC has shown limited progress. PD-L1 helps tumors evade the immune system by inhibiting T cell activity. Deng et al. demonstrated that glucocorticoid receptor (GR) expression is positively correlated with PD-L1, and inhibition of GR leads to the downregulation of PD-L1 on tumor cells, thereby promoting T cell infiltration and enhancing anti-tumor immunity in a PDAC mouse model [11]. EGFR inhibitors have been approved for EGFR-targeted therapy in lung adenocarcinoma patients [12]. Alan's team reported that polyclonal T cells targeting EGFR can effectively direct PDAC tumor cells, thereby increasing T cell infiltration and T cell-mediated anti-tumor effects [13]. Despite the lack of effective screening for PDAC, approximately 50 % of patients have advanced disease at diagnosis [14]. The combination of targeted therapy and immunotherapy was focused on melanoma and non-small cell lung cancer [15,16]. Combining targeted therapy, chemotherapy, or immunotherapy may currently have more potential in cancer treatment. Finding new effective targets or biomarkers is urgently needed in PDAC.

Here, by utilizing scRNA-seq, spatial transcriptomics, bulk RNA-seq, and bioinformatics analysis, we screened and validated membrane protein genes associated with prognosis. We developed a novel prognostic risk model MaGPS based on PDAC-related membrane protein genes to predict the prognosis of pancreatic cancer patients. Through validation using the Human Protein Atlas (HPA) and spatial transcriptomics, we confirmed the relationship between these membrane-related genes and PDAC at the protein and spatial levels. This MaGPS model provides new biomarkers for the prognosis of pancreatic cancer patients. In future pancreatic cancer treatments, it provides new targets for conducting preclinical experiments on new membrane protein-targeted therapies.

2. Materials and methods

2.1. Data collection

PDAC-related scRNA-seq data were obtained from GEO with accession numbers GSE197177 and GSE212966. Spatial transcriptomics data related to PDAC were acquired from GEO with accession number GSE233293. Bulk RNA-seq data for PDAC were retrieved from TCGA-PAAD, ICGC-PACA-AU, ICGC-PACA-CA, and GEO with accession number GSE224564. Sample information can be found in Supplementary Table 1.

2.2. scRNA-seq quality control and processing

For the 13 scRNA-seq samples from GSE197177 and GSE212966, cells with nCount <1000, nFeature <500, and mitochondrial percentage (pMT) > 10 % were filtered out. Following normalization of gene expression and dimensionality reduction clustering using Seurat's pipeline, batch effects were removed with Harmony [17,18]. Distinct cell subpopulations were identified based on single-cell markers, and those expressing double cell type markers were excluded. Cell markers referenced from the CellMarker2.0 [19] database included immune cells (PTPRC), T/NK cells (CD3D, CD3E, and NKG7), fibroblasts (COL1A1 and FAP), epithelial cells (EPCAM and KRT18), macrophages/monocytes (CD68, MARCO, and CD14), B cells (CD79A and MS4A1), endothelial cells (PECAM1 and VWF), mast cells (CPA3 and KIT), and plasma cells (IGLC2 and IGHA1). A total of 57,850 cells representing eight major cell types were used for downstream analysis.

2.3. Differential expression gene analysis

Differentially expressed genes (DEGs) for each cluster in the scRNA-seq data were identified using the Seurat pipeline's FindAllMarkers function, with FindMarkers being used when calculating DEGs for specific cell types. Genes with |log2 fold change| > 1 and FDR <0.05 were defined as DEGs. DEGs were visualized using volcano plots generated by the R package EnhancedVolcano. Highly expressed genes were displayed using heatmaps created with the R package pheatmap. Results can be found in Supplementary Tables 2, 3, and 4.

2.4. Protein-protein interaction network analysis

The protein-protein interaction (PPI) network was predicted using the STRING database [20]. Membrane-associated genes were input into STRING, with the species set to Homo sapiens and other parameters left at default settings. Subsequently, the enriched functional pathways within the network were visualized using the R package ggplot2.

2.5. Spatial transcriptomics analysis

For the three spatial transcriptomics datasets from GSE233293, we processed the data using the BayesSpace pipeline [21]. After data normalization, spatial spots were identified and classified into different clusters. The spatial expression patterns of membrane-associated genes were mapped using the featurePlot function. Spearman correlation coefficients for spatial expression patterns were calculated based on spatial spots.

2.6. Construction and validation of the prognostic model MaGPS

Membrane-associated genes, identified as highly expressed in PDAC epithelial cells via scRNA-seq, were used for univariate Cox regression analysis to calculate their prognostic significance in overall survival (OS) of PDAC patients in the TCGA-PAAD cohort. Genes with a significance level of p < 0.05 were identified as prognostic genes. Further, the least absolute shrinkage and selection operator (LASSO) with ten-fold cross-validation was used to select the optimal model, resulting in a list of genes with non-zero coefficients (Supplementary Table 5). The prognostic risk model was established by combining the gene mRNA expression levels with their risk coefficients. Patients were stratified into high-risk and low-risk groups based on the median risk score. The MaGPS calculation formula was developed as follows:

$$MaGPS=\sum _{i=1}^{k}riskcoefficient(i)\ast mRNAexpressionlevelofgene(i)$$

To validate the prognostic capability of MaGPS, we calculated statistical significance using the log-rank test and visualized the results with Kaplan-Meier plots. The R package timeROC was used to compute the area under the curve (AUC). The predictive performance of the model was further validated through survival analysis and AUC calculations using independent datasets from ICGC and GEO.

2.7. GSEA analysis of the model

KEGG pathway gene sets were obtained from the Molecular Signatures Database (MSigDB). For MaGPS, patients were divided into high-risk and low-risk groups, and the fold change of genes between the two groups was calculated. Genes were ranked according to fold change, and Gene Set Enrichment Analysis (GSEA) was performed using the R package fgsea. Significant pathways in both high-risk and low-risk groups were visualized using the R package ggplot2. KEGG pathway Enrichment results can be found in Supplementary Table 6.

2.8. IHC analysis of MaGPS signatures

To validate the protein expression levels of the six MaGPS signatures, we retrieved immunohistochemistry (IHC) data from the Human Protein Atlas (HPA; https://www.proteinatlas.org/) online database. We screened for IHC images of pancreatic cancer and normal pancreatic tissues corresponding to the six signatures. The results were then integrated and visualized.

2.9. Analysis of anti-tumor drug treatment efficacy assessment

To assess whether there are differences in the predicted treatment response to different drugs between MaGPS high-risk and low-risk groups, the R package pRRophetic v0.5 was used to predict mRNA expression matrices of PDAC samples from the TCGA-PAAD dataset. pRRophetic combines machine learning and statistical models to model and train on large-scale data sets to predict responses to various drugs [22]. The function pRRopheticPredict was used with the parameters set as follows: tissueType = “all”, dataset = “cgp2014”. A total of 138 drugs were evaluated, and the results were visualized using the ggboxplot function in the R package ggpubr.

2.10. Statistical analysis

Where appropriate, differences between groups were calculated using either t-tests or wilcoxon tests, depending on the data distributions. Survival analysis was conducted using the log-rank test. A p-value <0.05 was considered statistically significant.

3. Results

3.1. Study schematic and single-cell landscape of human PDAC

This study collected pancreatic cancer-related data from public databases, including two scRNA-seq projects, one spatial transcriptomics project, and four bulk RNA-seq cohorts. PDAC tumor epithelial cell subpopulations were identified through scRNA-seq analysis, and a prognostic model was constructed based on the membrane-associated genes expressed by these subpopulations. Finally, the potential therapeutic value of the model's signatures was validated through multi-omics analysis (Fig. 1A). First, we performed quality control on a total of 13 scRNA-seq samples from GSE197177 and GSE212966, including PDAC (n = 9) and adjacent normal tissue (n = 4; Fig. S1A). Cells with nCount <1000, nFeature <500, or mitochondrial gene percentage (pMT) > 10 % were filtered out (Fig. S1B). Subsequently, batch effects were removed using PCA-based harmony, and dimensionality reduction clustering was employed to obtain 19 clusters. Distinct cell types were defined based on the mRNA expression of various cell markers (Fig. S1C). Cells with non-specific expression were defined as 'ambiguous' and excluded (Fig. S1D). Eight major cell types, including cells from adjacent normal tissue (n = 15,040 cells) and PDAC (n = 42,810 cells), were identified (Fig. 1B). We observed that the infiltration proportions of immune cells, such as T/NK cells, were not low in PDAC, and the proportion of fibroblasts showed patient heterogeneity among PDAC samples (Fig. 1C). Furthermore, the expression of markers specific to the eight major cell types included T/NK cells (PTPRC, CD3D, CD3E, and NKG7), fibroblasts (COL1A1 and FAP), epithelial cells (EPCAM and KRT18), macrophages/monocytes (CD68, MARCO, and CD14), B cells (CD79A and MS4A1), endothelial cells (PECAM1 and VWF), mast cells (CPA3 and KIT), and plasma cells (IGLC2 and IGHA1) in their respective subpopulations (Fig. 1D). Despite potential discrepancies with previous reports regarding cell types [[23], [24], [25]], we obtained reliable cell subpopulations, providing a solid foundation for downstream analyses.

Fig. 1.

Workflow and single-cell landscape of human PDAC A) An overview of the study workflow. Step 1: Gene expression data were collected from GEO, TCGA, and ICGC; protein expression information was gathered from HPA; and cell markers were obtained from CellMarker2.0 website. Step 2: Single-cell level analysis was conducted. Step 3: A membrane-related gene prognostic model was established and validated. Step 4: Specific identification of the model's signatures was performed.

B) UMAP plot of the eight major cell types in adjacent tissue (15,040 cells) and PDAC (42,810 cells), with cell types distinguished by different colors.

C) Barplot depicting the cellular proportions in adjacent (n = 4) and PDAC (n = 9) tissues.

D) Dot plot illustrating the expression of representative markers of eight major cell types.

3.2. Identification of PDAC tumor epithelial cells

To identify PDAC tumor-specific epithelial cells, we subset 7093 epithelial cells and re-clustered them into six epithelial cell clusters, naming them Epi_c + cluster number, such as Epi_c0 (Fig. 2A). Compared to adjacent normal tissue, the number of epithelial cells in clusters Epi_c0, Epi_c1, and Epi_c5 was higher in PDAC tissues, whereas clusters Epi_c2, Epi_c3, and Epi_c4 were more prevalent in adjacent tissues (Fig. 2A). The cell proportion results indicated that Epi_c1 was present in the adjacent ADJ2 sample, while Epi_c0 had a high proportion in all nine PDAC samples, suggesting that Epi_c0 might be a PDAC tumor-specific epithelial cell rather than Epi_c1 (Fig. 2B). Subsequent differential analysis revealed that the top five highly expressed genes in Epi_c0 were TFF1, CEACAM5, CEACAM6, S100A4, and MUCL3 (Fig. 2C). TFF1 has been reported to be expressed in PDAC and can serve as a biomarker for the early detection of pancreatic cancer [26,27]. Notably, Epi_c5 showed high expression of proliferation-related markers such as MKI67 and was present in eight of the nine PDAC samples (except PDAC8), suggesting that Epi_c5 may be a proliferative stem-like epithelial cell (Fig. 2C). In comparison to adjacent tissues, PDAC tissues exhibited high expression of 153 genes (Fig. 2D). By intersecting the highly expressed genes in PDAC with those in Epi_c0, we identified 33 genes, including classic PDAC tumor markers like TFF1, TFF2, TFF3 and CLDN18 [28,29](Fig. 2E). These findings suggest that Epi_c0 is likely to be a PDAC tumor-specific epithelial cell.

Fig. 2.

Identification of tumor-specific epithelial cell clusters A) UMAP plot illustrating six epithelial cell clusters in adjacent (2310 cells) and PDAC (4783 cells) tissues.

B) Barplot depicting the proportions of five epithelial cell clusters in adjacent (n = 4) and PDAC (n = 9) tissues.

C) Heatmap illustrating the expression of the top 5 genes in six epithelial cell clusters.

D) Volcano plot of differentially expressed genes (DEGs) in epithelial cells between PDAC and adjacent tissues.

E) Venn plot illustrating the intersection of highly expressed genes in PDAC and those highly expressed in epithelial cell cluster c0.

3.3. Expression of 17 membrane-associated genes in tumor epithelial cells

Tumor membrane proteins play crucial roles in tumor progression. For instance, PD-L1 facilitates immune evasion of tumors by suppressing T-cell activity [30]. However, monotherapy targeting PD-1/PD-L1 blockade has shown limited efficacy in treating pancreatic cancer [31]. Here, we analyzed membrane-associated genes expressed in tumors. Initially, we obtained a set of membrane-associated genes (n = 2553 variables) from the Gavin Wright Lab's gene library [32]. Differential gene expression analysis revealed that 61 membrane-associated genes were upregulated in PDAC compared to adjacent normal tissue (Fig. 3A). Subsequently, we identified 17 membrane-associated genes that were highly expressed both in PDAC and in the Epi_c0 cluster (Fig. 3B). The 17 membrane-associated genes were input into STRING to investigate protein-protein interactions. Among them, 12 membrane proteins were found to interact with each other, forming four groups (Fig. 3C). These membrane proteins were significantly associated with pathways such as Intrinsic component of membrane, Apical part of cell, Extracellular space, Apical plasma membrane, Extracellular exosome, and Mitochondrial respiratory chain complex IV (Fig. 3D). Furthermore, spatial transcriptomic data from three PDAC samples (P1, P2, and P3) based on tissue sections were analyzed to observe the spatial localization of these 17 membrane-associated genes relative to the epithelial marker EPCAM (Figs. S2A–B). The expression patterns of most membrane-associated genes were consistent with EPCAM spatially (Fig. 3E and S2C-D). These findings suggest that the expression of the 17 membrane-associated genes holds significant potential for further research in relation to PDAC tumor epithelium.

Fig. 3.

Membrane-associated gene analysis of epithelial cell cluster c0 in PDAC

A) Volcano plot illustrating the differential expression of 2553 membrane-associated genes in PDAC compared to adjacent tissue.

B) Venn plot illustrating the intersection of highly expressed membrane-associated genes in PDAC and those highly expressed in epithelial cell cluster c0.

C) The protein-protein interaction (PPI) network of 17 membrane-associated genes via STRING.

D) Barplot illustrating the enrichment of Gene Ontology (GO) terms associated with 17 membrane proteins.

E) Spatial representation of EPCAM and 17 membrane-associated genes in PDAC P1.

3.4. Construction of the membrane-associated gene prognostic signature (MaGPS)

Given the crucial role of membrane proteins in cancer proliferation, metastasis, and immune evasion, and their high expression in tumor epithelial cells, we aimed to investigate the prognostic significance of these membrane-associated genes in PDAC patients. First, we utilized univariate Cox regression to calculate the association between the 17 membrane-associated genes and patient overall survival (OS) in the TCGA-PAAD cohort. Six genes—LMO7, APOL1, SLC2A1, C15orf48, FXYD3, and CLDN18—were found to be significantly associated with patient prognosis (Fig. 4A). Using LASSO regression with ten-fold cross-validation, these six genes were used to construct the membrane-associated gene prognostic signature (MaGPS; Fig. 4B). Samples were divided into high- and low-risk groups based on risk scores, with LMO7, APOL1, SLC2A1, and C15orf48 showing higher expression in the high-risk group (Fig. 4C). We observed that MaGPS effectively stratified patients' OS in the TCGA cohort, with a high-risk score associated with poorer prognosis (p = 0.00012; Fig. 4D). Additionally, this model demonstrated high predictive accuracy for patient 1-, 2-, and 3-year survival rates, with AUCs of 0.742, 0.676, and 0.679, respectively, in the TCGA cohort (Fig. 4E).

Fig. 4.

Prognostic model of membrane-associated genes (MaGPS) based on TCGA database

A) Forest plot of 6 membrane-associated genes significantly associated with PDAC prognosis.

B) Each independent variable's trajectory and distributions for the lambda.

C) Heatmap illustrating the expression of 6 signatures in high and low-risk groups.

D) Kaplan-Meier plot of MaGPS in the TCGA dataset. E) The area under the curve (AUC) of the 1, 2, and 3-year survival rates based on TCGA dataset.

3.5. Functional analysis and performance evaluation of the MaGPS model

Next, to evaluate the functional roles involved in MaGPS, we performed GSEA analysis on KEGG pathways. Differentially expressed genes between the high- and low-risk groups were ranked based on fold changes for GSEA in the TCGA-PAAD cohort. Pathways such as Cell Cycle, DNA Replication, P53 Signaling Pathway, Pathways in Cancer, and Spliceosome were enriched in high-risk patients, while immune-related pathways including T Cell Receptor Signaling Pathway, B Cell Receptor Signaling Pathway, and Innate Immune Response were enriched in low-risk patients (Fig. 5A). Additionally, we validated the prognostic value and performance of MaGPS in three independent cohorts (Fig. 5B–G). MaGPS effectively stratified patients in the GSE224564 cohort (p = 0.0011; Fig. 5B) and the ICGC-CA cohort (p = 0.012; Fig. 5C), while it showed a stratification trend, albeit not significant, in the ICGC-AU cohort (p = 0.059; Fig. 5D). The AUC values for 1-, 2-, and 3-years survival rates were as follows: in the GSE224564 cohort, 0.558, 0.592, and 0.631, respectively (Fig. 5E); in the ICGC-CA cohort, 0.606, 0.615, and 0.601, respectively (Fig. 5F); and in the ICGC-AU cohort, 0.637, 0.675, and 0.619, respectively (Fig. 5G). These results suggest that patients in the low-risk group have a better immune response than those in the high-risk group and demonstrate the potential prognostic significance of MaGPS for predicting patient outcomes.

Fig. 5.

Validation of the prognostic model MaGPS

A) GSEA plot depicting KEGG enrichment in high and low-risk groups based on the TCGA database.

B) Kaplan-Meier plot of MaGPS in the GSE224564 cohort.

C) Kaplan-Meier plot of MaGPS in the ICGC-CA cohort.

D) Kaplan-Meier plot of MaGPS in the ICGC-AU cohort.

E) The AUC of the 1, 2, and 3-year survival rates based on GSE224564 cohort.

F) The AUC of the 1, 2, and 3-year survival rates based on ICGC-CA cohort.

G) The AUC of the 1, 2, and 3-year survival rates based on ICGC-AU cohort.

3.6. Expression characteristics of the six MaGPS signatures

Further, we identified and validated the six signatures of the MaGPS model. These genes were highly expressed in Epi_c0 and were also present in other tumor epithelial clusters, Epi_c1 and Epi_c5 (Fig. 6A). The six signatures were highly expressed in PDAC but not in samples from adjacent normal tissue (Fig. 6B). Additionally, we found that among these signatures, CLDN18, FXYD3, and SLC2A1 showed specific expression in epithelial cells (Fig. 6C). C15orf48 was expressed in both epithelial cells and macrophages/monocytes (Fig. 6C). LMO7 and APOL1 were expressed in epithelial cells as well as in fibroblasts (Fig. 6C). Notably, APOL1 was also expressed in endothelial cells (Fig. 6C). In the tumor microenvironment, fibroblasts and macrophages are typically defined as cancer-associated fibroblasts (CAFs) and tumor-associated macrophages (TAMs), respectively, both of which play roles in promoting tumor proliferation, migration, and immune evasion. Therapeutic strategies targeting these membrane proteins might be effective not only against cancer cells but also against CAFs and TAMs. Furthermore, spatial analysis revealed that the expression of these six signatures co-localized with the epithelial marker EPCAM (Fig. 6D and S3A-B). The spatial spearman correlation between these six signatures and EPCAM ranged from 0.57 to 0.75, showing significant spatial co-localization in PDAC-P2 tissue (Fig. 6E). This indicates that the six signatures of MaGPS are spatially highly correlated with the tumor. Moreover, these signatures also exhibited spatial association with EPCAM in both P1 and P3 samples (Fig. S3C). Finally, the protein expression levels of the six signatures were identified via the Human Protein Atlas (HPA). The proteins encoded by these six signatures were highly expressed in pancreatic cancer tissues, while they were either not expressed or expressed at low levels in normal pancreatic tissues (Fig. 6F). These findings further elucidate the role of the six MaGPS membrane-associated genes in tumor single-cell analysis, spatial transcriptomics, and tissue protein expression, providing new insights for the development of novel immunotherapies and clinical decision-making.

Fig. 6.

Spatial analysis and expression profiling of MaGPS targets

A) Dotplot illustrating the expression of the 6 signatures of MaGPS across 6 epithelial clusters.

B) Dotplot illustrating the expression of the 6 signatures of MaGPS between adjacent and PDAC tissues.

C) Dotplot illustrating the expression of the 6 signatures of MaGPS across 8 major cell types.

D) Spatial plot depicting the distribution of H&E, EPCAM, and the 6 signatures of MaGPS in PDAC P2.

E) Spatial correlation between EPCAM and the 6 MaGPS signatures in PDAC P2.

F) Protein expression of the 6 MaGPS signatures in pancreatic cancer and normal pancreas via the Human Protein Atlas (HPA).

3.7. MaGPS predicted the potential benefit of anti-tumor therapy in PDAC

To explore potential targeted drugs for MaGPS-related genes, we used the Cancer Genome Project (CGP 2014) dataset to analyze the correlation between MaGPS risk groups and the sensitivity to antitumor drugs. The pRRophetic package was used to predict the IC50 values of 138 drugs for PDAC patients based on TCGA-PAAD data. The samples were divided into high and low MaGPS score groups, and differences between the two groups were compared. Significant differences in IC50 values for 38 drugs were observed between the two groups (Supplementary Table S7). Seven drugs exhibited lower IC50 values in the low-risk group, while 31 drugs showed lower IC50 values in the high-risk group. Notably, the top drugs with significantly lower IC50 values in the low-risk group included ATRA (p = 7.2e-05), AICAR (p = 0.00059), EHT-1864 (p = 0.0035), BX-795 (p = 0.0065), and CCT007093 (p = 0.0096; Fig. 7A). PDAC patients with a low MaGPS score may be more sensitive to these drugs. Moreover, drugs with lower IC50 values in the high-risk group, which may be more effective for PDAC patients with a high MaGPS signature score, included Tipifarnib (p = 7.1e-05), GSK-650394 (p = 0.00033), VX-680 (p = 0.00035), Thapsigargin (p = 0.0016), and Epothilone-B (p = 0.0019; Fig. 7B). These results demonstrated the potential of MaGPS-related genes in guiding the selection of antitumor drugs and in contributing to more tailored and effective treatment strategies for PDAC patients.

Fig. 7.

The MaGPS associated with the drug sensitivity in PDAC

A) The boxplot shows that the top drugs with significantly lower IC50 values in the low-risk group include ATRA, AICAR, EHT-1864, BX-795, and CCT007093.

B) The boxplot shows that the top drugs with significantly lower IC50 values in the high-risk group include Tipifarnib, GSK-650394, VX-680, Thapsigargin, and Epothilone-B. Statistics based on Wilcoxon test. Statistics based on Wilcoxon test. P < 0.05∗, <0.01∗∗, <0.001∗∗∗, <0.0001 ∗∗∗∗.

4. Discussion

Unlike bulk RNA-seq, scRNA-seq allows researchers to delve into PDAC and its microenvironment at the single-cell level. Numerous research teams have provided new insights and data based on scRNA-seq technology. In 2019, Peng et al. published the first scRNA-seq analysis of PDAC, revealing intratumoral heterogeneity with two distinct ductal subtypes [23]. To date, research on PDAC includes studies on the TME, tumor metastasis, and targeted therapy [24,33,34]. Pan et al. reported that targeting tumor surface membrane CD47 could enhance PDAC's response to immune checkpoint inhibitors (ICPi) by altering the TME [33]. Here, we integrated scRNA-seq, spatial transcriptomics, and bulk RNA-seq data to identify membrane-related genes specific to PDAC tumor epithelial cells, and constructed a novel prognostic risk model (MaGPS) based on these genes. Our model has been validated across multiple independent datasets, providing new insights for prognostic evaluation and targeted therapy in PDAC patients.

By analyzing 13 scRNA-seq samples from GSE197177 and GSE212966, we identified PDAC tumor epithelial cell subpopulations. Genes such as TFF1, CEACAM5, and CEACAM6 were highly expressed in the Epi_c0 cluster. The expression of TFF1 is primarily an early event in the development of pancreatic cancer, and TFF1 protein can serve as a biomarker for early-stage pancreatic cancer, detectable in urine samples [26,35]. CEACAM5/6 have been reported to be highly expressed in PDAC tumor tissues, and the development of antibodies targeting CEACAM6 and PD-1 for combined therapy has been shown to induce significant tumor regression [[36], [37], [38]]. This further supports the identification of the Epi_c0 cluster as PDAC-specific epithelial cells.

Membrane-associated genes play critical roles in tumor proliferation, migration, and immune evasion. PD-L1 helps tumors evade the immune system by inhibiting T-cell activity [39,40]. Although PD-L1 was not among the 17 membrane-associated genes identified in the Epi_c0 cluster of PDAC tumor epithelial cells, the high expression of some of these genes has been closely linked to tumor biology. Zhou et al. reported that TSPAN1 is upregulated in pancreatic cancer and promotes cancer cell proliferation [41,42]. The CD55 protein is upregulated in pancreatic cancer cell lines and its expression significantly increases in PDAC stages II-IV [43]. ECM1 is significantly elevated in PDAC and enhances cell migration and invasion [44]. Overall, the high expression of these genes may provide new targets for the diagnosis and treatment of PDAC.

We identified six prognosis-related membrane-associated genes (LMO7, APOL1, SLC2A1, C15orf48, FXYD3, and CLDN18) and constructed the MaGPS model. MaGPS demonstrated excellent stratification and high accuracy in predicting the survival of PDAC patients. LMO7 has been reported to be highly expressed in pancreatic cancer, lung cancer, and thyroid cancer, promoting cell proliferation [[45], [46], [47]]. The oncogene APOL1 is highly expressed in pancreatic cancer, promoting proliferation and inhibiting apoptosis through the activation of the NOTCH1 pathway [48]. SLC2A1 plays a role in cancer metabolism, with its high expression enhancing gastric cancer cell proliferation and inducing immune evasion and liver metastasis in colorectal cancer [49,50]. C15orf48 is involved in inflammatory responses and is associated with the prognosis of thyroid cancer and hepatocellular carcinoma [51,52]. FXYD3 is overexpressed in invasive bladder cancer [53], and its high expression in breast cancer is linked to drug resistance and persistence [54,55]. CLDN18 is commonly overexpressed in gastric and pancreatic cancers [56,57], and targeting CLDN18 could be an effective treatment strategy for these cancers [56]. These findings are consistent with our study, underscoring the clinical potential of MaGPS for prognostic assessment in PDAC patients. Spatial transcriptomics analysis of GSE233293 revealed that MaGPS genes are significantly co-localized with the epithelial marker EPCAM in PDAC tissues. Furthermore, IHC results from the HPA database confirmed the high expression of these genes in PDAC tissues, with lower expression in normal pancreatic tissues. Notably, Tipifarnib has been reported to inhibit pancreatic cancer proliferation and promote cell apoptosis [58,59]. This is consistent with our findings, where patients with a high MaGPS score exhibited lower IC50 values for tipifarnib, suggesting that tipifarnib may offer therapeutic benefits for PDAC patients with a high MaGPS signature score. Future studies should further validate these signatures as therapeutic targets and explore their application in combination therapies, such as with chemotherapy or immunotherapy.

Despite the significant findings of this study, there are some limitations, such as the limited sample size and the heterogeneity of data sources. The AUC value of the MaGPS model is between 0.55 and 0.7, which means that the predictive ability of MaGPS for patient prognosis is not high enough. Future research should expand the sample size and integrate more multi-omics data to enhance the generalizability and reliability of the model. This study lacks specific experimental validation, such as cell-based or in vivo studies targeting these genes. Future research should aim to validate these findings experimentally and explore these genes as potential therapeutic targets for pancreatic cancer.

In conclusion, this study has constructed and validated the MaGPS model, providing new biomarkers (LMO7, APOL1, SLC2A1, C15orf48, FXYD3, and CLDN18) for prognostic assessment in PDAC patients. Additionally, it has identified these six membrane-associated gene signatures as potential therapeutic targets in PDAC. These findings offer valuable insights for future clinical research and therapeutic development.

CRediT authorship contribution statement

Zhaowei Ding: Writing – original draft, Visualization, Formal analysis, Conceptualization, Methodology, Writing – review & editing. Jun Wu: Writing – original draft, Formal analysis, Writing – review & editing. Yongqing Ye: Methodology. Yunlong Zhong: Data curation. Lei Yan: Data curation. Ping Wang: Writing – review & editing, Supervision, Funding acquisition, Conceptualization.

Availability of data

scRNA-seq data can be accessed from GEO under accession numbers GSE197177 and GSE212966, and spatial transcriptomics data from GEO under accession number GSE233293. Bulk RNA-seq and clinical information can be obtained from the TCGA-PAAD cohort, ICGC's PACA-AU and PACA-CA cohorts, and GEO under accession number GSE224564. For any other reasonable requests, please contact the corresponding author.

Code availability

The relevant code used in this study for model construction and analysis will be made available upon request.

Funding

This work was supported by a grant from the Guangzhou Science and Technology Plan Project (no.202102010251 and no.2024A03J1156).

Declaration of competing interest

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.

Appendix A. Supplementary data

The following is the supplementary data to this article: