Comprehensive single-cell transcriptome landscape of metastatic colorectal cancer identifies budding-potential cells and their interactions with cancer-associated fibroblasts

Abstract

Tumor budding is positively associated with colorectal cancer (CRC) metastasis. In this study, we integrated a single-cell transcriptomic dataset of 287 CRC samples to comprehensively illustrate the transcriptomic landscape of metastatic CRC and identified a unique subcluster of tumor epithelial cells associated with tumor budding. This subcluster exhibited high mesothelin (MSLN) expression and was located at the invasive front of CRC. MSLN was confirmed to promote CRC growth and metastasis by in vitro and in vivo models. Also, POSTN⁺ fibroblasts in the CRC microenvironment showed enhanced expression of genes in epithelial-mesenchymal transition and angiogenesis signaling pathways, which wrapped around MSLN⁺ tumor budding cells in the invasive front of CRC. POSTN⁺ fibroblasts may interact with MSLN⁺ budding-potential cells through the ligand-receptor pair POSTN-ITGB5 to promote tumor metastasis. In conclusion, our findings identified the transcriptomic feature of budding-potential cells and revealed the role of crosstalk between MSLN⁺ budding-potential cells and POSTN⁺ fibroblasts in CRC metastasis, which provide new insights into targeting cancer-associated fibroblasts and tumor budding cells in CRC therapy.

Introduction

Colorectal cancer (CRC) is the third most common malignancy in the world and one of the leading causes of cancer-related deaths¹. 15% to 25% of CRC patients have distant metastasis at the time of diagnosis^2,3. Approximately 30% of patients will develop distant metastasis and recurrence after radical surgical resection, with a five-year survival rate of less than 20%⁴. Metastasis is the leading cause of death in CRC patients and also the main challenge in CRC treatment^5,6. Tumor budding has been recognized as a pathological risk factor for metastasis in CRC^7,8. It is defined as the presence of a single cell or a cluster of up to four cancer cells at the invasive front, which can predict lymph node metastasis, distant metastasis and postoperative recurrence^9–11. Tumor budding has been proposed to be histo-morphometric reflection of epithelial-mesenchymal transition (EMT)¹².

With the help of high-throughput single-cell sequencing technologies, a series of studies have identified the molecular heterogeneity of different cell subpopulations in the tumor microenvironment of CRC. For example, the crosstalk between SPP1+ macrophages and FAP+ fibroblasts leads to poorer prognosis¹³, and the interactions between stem cell-like cells and cancer-associated fibroblast (CAF) and endothelial cells promote tumor metastasis¹⁴. However, due to limited cohort size and number of captured cells, which makes it challenging to investigate rare cell subpopulations such as cells in tumor budding, there is still a lack of comprehensive understanding of molecular characteristics of tumor budding.

In order to comprehensively identify cell types associated with CRC metastasis and analyze their molecular features, we integrated eight different single-cell datasets to build a single-cell atlas that includes normal tissue, non-metastatic, and metastatic tumor tissues. We examined the heterogeneity between non-metastatic and metastatic CRC, and identified a previously unreported MSLN⁺ budding-potential cells which is strongly associated with the CRC invasive phenotype. Also, we revealed its interaction network with CTHRC1⁺POSTN⁺ CAF that promotes CRC metastasis. These findings help us to understand the mechanisms of CRC metastasis and may provide new therapeutic targets for metastatic CRC.

Results

Construction of a single-cell atlas of human metastatic CRC and identification of epithelial cell subpopulations

To construct a comprehensive human single-cell atlas of metastatic CRC, we collected eight datasets comprising 287 samples from 134 patients. This extensive single-cell atlas included 89 non-tumor normal samples and 198 tumor samples, which represented various pathological stages, including non-metastatic (AJCC stage I and II) and metastatic (AJCC stage III and IV) CRC (Fig. 1A). After quality control, a high-quality single-cell atlas comprising 562,446 cells was generated. Subpopulations were defined based on differential gene expression and classical markers, utilizing harmony to remove batch effects. Unsupervised clustering analysis led to the identification of eight distinct cell subpopulations: B cells (gene markers: MS4A1, CD79A), endothelial cells (gene markers: ENG, VWF), epithelial cells (gene markers: EPCAM, KRT18), fibroblasts (gene markers: COL3A1, THY1), mast cells (gene markers: KIT, TPSAB1), myeloid cells (gene markers: LYZ, CD68), NK/T cells (gene markers: CD3D, CD3E), and plasma cells (gene marker: MZB1) (Fig. 1B, E). The major cell types in each dataset indicated minimal batch effect (Fig. S1A, B). We observed distinct differences in the proportions of cell types in normal tissue, non-metastatic tumor tissue and metastatic tumor tissue. Specifically, normal tissues exhibited higher proportions of B cells and plasma cells, alongside lower proportions of myeloid cells. Compared with non-metastatic tumors, metastatic tumors displayed increased proportions of epithelial cells and fibroblasts, with reduced proportions of plasma cells and NK/T cells (Fig. 1C, D, Fig. S1F). These findings suggest a state of immunosuppression in metastatic CRC.

Fig. 1. Construction of a single-cell atlas of human metastatic CRC and identification of epithelial cell subpopulations.

A Schematic overview of the experimental design of this study. B UMAP plot of 562,446 single cells from 287 samples. C UMAP plot showing the metastatic status of all cells. D Bar plots showing the percentage of each major cell type in different metastatic states. E Bubble plots showing marker genes for each cell type. F UMAP plot of tumor epithelial cell subpopulations, colors represent different subclusters. G Violin plot showing differences in KRT20, EMP1 gene expression in different tumor epithelial cell subclusters. H UMAP plot showing the metastatic status of different tumor epithelial cell subclusters. I Bar plots showing the percentage of different tumor epithelial cell subclusters in different metastatic states.

To study the role of different epithelial subpopulations in metastasis, we conducted a dimensionality reduction and cluster analysis of epithelial cells from tumor tissues. We identified and annotated 11 epithelial subpopulations including stem-like cells, goblet cells (GOB), tuft cells (TUF), Epi-RPL cells, EMP1⁺KRT20⁺ differentiated cells, EMP1⁺KRT20^- undifferentiated cells, as well as Cluster 1 (C1), C2, C3, C4, and C5 (Fig. 1F). The marker gene lists of epithelial cell subclusters can be found in Table S4. Differentiated cells were identified by high levels of EMP1 and KRT20 expression. In contrast, undifferentiated cells showed high EMP1 expression and low KRT20 expression (Fig. 1G), consistent with previous study¹⁵. Cluster C3 and C4 were classified as proliferating cells, characterized by high expression of markers such as MKI67, STMN1, and TOP2A (Fig. S1C-E). Cell clusters that did not show prominent expression of specific markers were grouped as C1, C2, and C5. In addition, we examined the proportion of epithelial cell subpopulations between different metastatic states. We found that EMP1⁺KRT20- undifferentiated cells exhibited a higher proportion in metastatic tumor tissues (Fig. 1H, I), although it did not achieve statistical significance, suggesting their potential roles in tumor metastasis. To distinguish malignant cells from non-malignant cells, we used Copykat to analyze their copy number variation. The heatmap of copy number variation highlighted the differences between malignant and non-malignant cells (Fig. S2A-H). The results showed that non-malignant cells primarily comprised EMP1⁺ KRT20⁺ differentiated cells and C5 (Fig. S2I).

Metastasis-associated epithelial cell subclusters are budding-potential cells

To investigate the heterogeneity of EMP1⁺KRT20- undifferentiated cells, we conducted a subpopulation analysis and identified 11 distinct subclusters (Fig. 2A). The proportions of T7, T10, T3 and T11 subpopulations were significantly higher in metastatic CRC (Figs. 2B, C, S3A). The CNV analysis showed that the T7 and T11 subpopulations had a higher percentage of malignant cells, while the T3 and T10 subclusters consisted primarily of non-malignant cells (Fig. S3B, C).

Fig. 2. Metastasis-associated epithelial cell subcluster is related to tumor budding.

A UMAP plot of EMP1⁺KRT20- undifferentiated cells subclusters, colors represent different subclusters. B UMAP plot showing the metastatic status of different EMP1⁺KRT20^- undifferentiated cells subclusters. C Bar plots showing the percentage of each EMP1⁺KRT20^- undifferentiated cells subcluster in different metastatic states. D Violin plot showing differences in tumor budding score in different EMP1⁺KRT20^- undifferentiated cells subclusters. E Violin plot showing differences in EpiHR score in different EMP1⁺KRT20- undifferentiated cells subclusters. F Heatmap showing differentially expressed genes between different EMP1⁺KRT20- undifferentiated cells subclusters. G Heatmap showing differential pathway activation in EMP1⁺KRT20- undifferentiated cells subclusters. H Heatmap showing normalized activity of top 5 TF regulons in EMP1⁺KRT20- undifferentiated cells subclusters. I Representative IHC image of MSLN in tumor samples. Scale bar, 50 µm. J MSLN protein levels detected by IHC in 30 CRC tissues with different tumor budding grades including low budding (LB), intermediate budding (IB) and high budding (HB).

Epithelial-specific high-risk geneset (EpiHR) is associated with CRC metastasis and recurrence¹⁵. We also downloaded the differentially expressed gene list in tumor buds^15–20. The marker gene lists of tumor budding and EpiHR can be found in Table S3. We calculated EpiHR and tumor budding scores for each cell and found that T7 and T10 subpopulations had the highest EpiHR scores. Moreover, the T7 subpopulation exhibited the highest tumor budding scores (Fig. 2D, E). T10 showed elevated expression of fibroblasts markers such as COL3A1, COL1A1, and LUM. It also had low copy number variation (Figs. 2F, S3B). Doubletfinder analysis indicated that the majority of T10 and T11 subsets were doublets. Therefore, T10 and T11 subclusters were considered doublets (Fig. S3D, E). The marker gene lists of EMP1⁺KRT20^- undifferentiated cells subclusters can be found in Table S4. The T7 subpopulation showed higher expression of MSLN, SCEL, CST6, TRIM29, and KLK10 (Fig. 2F). Studies have found that MSLN activates the JNK signaling pathway to promote the metastasis of tumor cells²¹. Similarly, SCEL, TRIM29, and KLK10 have been shown to have similar effects^22–25. Also, GSVA and functional enrichment analysis revealed significant activation of several pathways in the T7 subpopulation, including epithelial-to-mesenchymal transition (EMT), angiogenesis, apical junction, and TGF-β pathways (Fig. 2G, Fig. S4A). To identify the key regulatory transcription factors in the T7 subpopulation, we performed pySCENIC analysis. The results showed that the transcription factor ZEB1, which is thought to be highly associated with epithelial-mesenchymal transition^26,27, was specifically activated in the T7 subpopulation. Additionally, other transcription factors related to metastasis, such as SOX4, ZNF281, and DLX2, also showed increased activity in this subpopulation (Fig. 2H)^28–30.

Next, we verified the expression site of MSLN, a representative gene of the T7 subcluster, in human CRC tissues using IHC staining. The results showed that MSLN exhibited positive staining in the disseminated tumor budding area at the tumor invasive front, with no expression observed in the tumor center area, indicating that the T7 subcluster was budding-potential cell (Fig. 2I). Also, by integrating the tumor budding information of 30 CRC tissues, we found that MSLN protein expression was increased in CRC with intermediate and high tumor budding compared to CRC with low tumor budding (Figs. 2J, S3F, G). To assess the relationship between tumor budding and metastasis, we retrospectively collected clinical data on 601 patients with CRC in our hospital (Table S5), and found a higher proportion of CRC in the metastatic group had high tumor budding compared to CRC in the non-metastatic group (p < 0.001) (Fig. S4B). Besides, our univariate and multivariate Cox regression analyses indicated that tumor budding was significantly associated with metastasis in CRC patients, suggesting that tumor budding is associated with metastasis (Fig. S4C, D).

MSLN expression is associated with metastasis of CRC

To further explore the relationship between the T7 subcluster and clinical pathological features, we used the CIBERSORTX deconvolution method to estimate the cell type abundance in the TCGA-CRC cohort. We found that in the TCGA-CRC cohort, patients with higher T7 subcluster abundance exhibited shorter disease-free survival (DFS) (Fig. 3A). In addition, the T7 subcluster was enriched in patients with metastatic cancer (stage III/IV, N1/N2, M1) (Fig. 3B, C). Additionally, we analyzed the association between T7 and Consensus Molecular Subtypes (CMS), revealed that T7 subcluster existed in all patients with four types of CMS, with a relatively higher proportion observed in CMS4 patients characterized by cancer-associated fibroblasts infiltration, indicating a close relationship between the T7 subcluster and CAFs (Fig. S5A, B). We then investigated the role of MSLN, a representative gene of the T7 subcluster, and found that MSLN expression was associated with poorer prognosis in patients from the TCGA-CRC cohort (Fig. 3D). Additionally, MSLN expression was higher in patients with metastatic cancer (Fig. 3E, F).

Fig. 3. MSLN promoted the metastasis of CRC.

A Kaplan-Meier curves showing the DFS in patients with high versus low T7 infiltration in the TCGA- CRC cohort. B Comparison of the proportion of T7 infiltration between non-metastatic (I + II, n = 260) and metastatic (III + IV, n = 194) tumors in the TCGA-CRC cohort. C Comparison of the proportion of T7 infiltration between N0 (n = 268), N1 (n = 105) and N2 (n = 81) stages in the TCGA-CRC cohort. D Kaplan-Meier curves showing the DFS in patients with high versus low MSLN expression in the TCGA- CRC cohort. E Comparison of MSLN expression between non-metastatic (I + II, n = 260) and metastatic (III + IV, n = 194) tumors in the TCGA-CRC cohort (F) Comparison of MSLN expression between N0 (n = 268), N1 (n = 105) and N2 (n = 81) stages in the TCGA-CRC cohort. G, H Western blotting showing the overexpression of MSLN. I, J Wound healing assays showing the effect of MSLN expression on the migratory capacity of RKO and SW480 cells (scale bar, 500 μm). K, L Transwell migration and invasion assays showing that MSLN overexpression promotes migration and invasion of RKO and SW480 cells (scale bar, 300 μm). M, N Representative bioluminescence images and fluorescence (FL) intensity analysis of SW480-luc orthotopic xenograft colon tumors after 6 weeks.

To further investigate the biological function of MSLN in CRC, we established MSLN overexpressing cell lines by lentivirus-mediated infection in colon cancer RKO and SW480 cell lines (Fig. 3G, H). Wound healing and transwell assays were used to analyze the effects of MSLN overexpression on cancer cell migration and invasion. The results showed that overexpression of MSLN significantly enhanced the migration and invasion of RKO and SW480 cells (Fig. 3I-L). Moreover, CCK8 assay also confirmed that MSLN overexpression significantly enhanced the proliferation of RKO and SW480 cell lines (Fig. S5C, D). We next established an in vivo orthotopic xenograft mouse model of CRC in BALB/c nude mice using SW480 cell line, and found that overexpression of MSLN led to a significant increase in fluorescence intensity, indicating that MSLN significantly promoted tumor progression of colorectal cancer (Fig. 3M, N). In summary, MSLN may be associated with metastasis of CRC.

CTHRC1⁺POSTN⁺ fibroblasts are associated with metastasis of CRC

In the tumor microenvironment, CAFs can influence the phenotype of tumor epithelial cells^31,32. We performed EMT scoring on all major cell types and found that fibroblasts exhibited the highest score (Fig. S6A). Hence, we explored the association between different subtypes of CAFs and MSLN⁺ budding-potential cells (T7). We extracted 20,966 fibroblasts and classified them into 13 subtypes using unsupervised clustering (Fig. 4A). Each subtype was named based on its highly expressed marker genes. The marker gene lists of fibroblasts subclusters can be found in Table S4. We found that three subclusters, including CTHRC1⁺CXCL14⁺ fibroblasts, CTHRC1⁺POSTN⁺ fibroblasts and MMP1⁺ fibroblasts were predominantly distributed in tumor tissues. In contrast, the CXCL14⁺NRG1⁺ fibroblasts, CCL13⁺ fibroblasts, OGN⁺IGF1⁺, OGN⁺PI16⁺, and S100B⁺ fibroblasts subtypes were mainly enriched in normal tissues, indicating the heterogeneity of fibroblasts across normal and tumor tissues. Moreover, there was a higher proportion of CTHRC1⁺POSTN⁺ fibroblasts in metastatic tumors compared to non-metastatic tumors, although it did not achieve statistical significance (Figs. 4B, C, S6B). To comprehensively evaluate the relationship between fibroblasts and metastasis, we conducted a differential cell abundance analysis using MiloR. Among cancer-associated fibroblasts, CTHRC1⁺POSTN⁺ fibroblasts were enriched in metastatic CRC (Fig. S6C, D). CTHRC1⁺POSTN⁺ fibroblasts showed higher expression of ECM remodeling-related genes (POSTN, CTHRC1, THBS2, COL11A1), and MMP1⁺ fibroblasts showed higher expression of matrix metalloproteinases (MMP1, MMP3) and inflammatory chemokines (CXCL1, CXCL8) (Fig. 4D).

Fig. 4. Characteristics of fibroblasts in non-metastatic and metastatic CRC.

A UMAP plot of fibroblast subclusters, colors represent different subclusters. B UMAP plot showing the metastatic status of different fibroblast subclusters. C Bar plots showing the percentage of each fibroblast subcluster in different metastatic states. D Bubble plots showing differentially expressed genes between different fibroblast subclusters. E Heatmap showing differential pathway activation in fibroblast subclusters. F Heatmap showing normalized activity of top 5 TF regulons in fibroblast subclusters. G The Kaplan-Meier curves showing the DFS in patients with high versus low CTHRC1⁺POSTN⁺ fibroblast infiltration in the TCGA-CRC cohort. H Comparison of the proportion of CTHRC1⁺POSTN⁺ fibroblast infiltration between non-metastatic (I + II) and metastatic (III + IV) tumors in the TCGA-CRC cohort. I Comparison of the proportion of CTHRC1⁺POSTN⁺ fibroblast infiltration between different N stages in the TCGA-CRC cohort.

Next, we investigated the function of different fibroblast subtypes using the hallmark gene set from Msigdb. The CTHRC1⁺CXCL14⁺ fibroblasts were enriched with interferon and WNT pathways, while the CTHRC1⁺POSTN⁺ fibroblasts showed significant activation of pathways related to EMT, apical junction, angiogenesis, and extracellular matrix organization (Figs. 4E, S6E). Meanwhile, the MMP1⁺ fibroblasts were enriched with pathways related to hypoxia, inflammatory response, and TNFα signaling pathway (Fig. 4E). Transcription factor analysis showed that CREB3L1, a transcription factor that promotes tumor metastasis by remodeling the extracellular matrix, was highly activated in CTHRC1⁺POSTN⁺ fibroblasts, which was consistent with previous studies^33,34. In addition, the transcription factor PITX2, which is associated with EMT in tumor cells, was activated in CTHRC1⁺POSTN⁺ fibroblasts (Fig. 4F)³⁵. This suggests that CTHRC1⁺POSTN⁺ fibroblasts may play a role in promoting CRC metastasis. Next, we used CIBERSORTX to infer the infiltration abundances of different fibroblast subpopulations in the TCGA-CRC cohort and found that patients with a higher proportion of CTHRC1⁺POSTN⁺ fibroblasts infiltration had a shorter DFS (Fig. 4G), which was associated with advanced stage (stage III/IV), lymph node metastasis and CMS4 (Figs. 4H, I, S6F, G). These findings suggest that CTHRC1⁺POSTN⁺ fibroblasts in the tumor microenvironment may be associated with metastasis of CRC.

Interactions between MSLN⁺ budding-potential cells and CTHRC1⁺POSTN⁺ fibroblasts are enhanced in the tumor invasive front of CRC with metastasis

Intercellular communications are critical in shaping tumor microenvironment, then we examined the differences in interactions between epithelial cells and fibroblasts between non-metastatic and metastatic CRC. In metastatic CRC, the number and strength of cellular interactions between fibroblasts and epithelial cells were increased (Fig. 5A–C). in metastatic CRC, 49 paired receptors were altered. Signaling pathways related to ECM remodeling (POSTN, THBS, FN1, COLLAGEN) were upregulated in metastatic CRC. In contrast, the intestinal digestive and absorption pathway (GUCA) and neuron-related signaling pathway (NRXN) showed decreased activity (Fig. 5D).

Fig. 5. POSTN⁺ fibroblasts and MSLN⁺ tumor budding cells are closely correlated at the tumor invasive front.

A Circle plots showing the cell interaction strengths in the non-metastatic (left) and metastatic (right) CRC, thicker edges indicate higher interaction strengths. B Circle plots showing changes in the number of interactions (left) and interaction strengths (right) among all cell types in metastatic CRC, edges colored red indicating increased interactions in metastatic CRC, edges colored blue indicating decreased interactions in metastatic CRC. C Bar plots illustrating changes in the total number of fibroblast-epithelial cell interactions and interaction strengths in non-metastatic and metastatic CRCs. D Differential expression of ligand-receptor signaling pathways in fibroblast-epithelial cell communication networks in non-metastatic and metastatic CRC. E Circle plots showing the number of interactions (left) and interaction strengths (right) among the fibroblast subclusters and MSLN⁺ tumor budding cells. F Correlations between the infiltration proportion of different EMP1⁺KRT20^- undifferentiated cells subclusters and fibroblast subclusters in the TCGA-CRC cohort. G Scatterplot demonstrating the correlation between CTHRC1⁺POSTN⁺ fibroblast and MSLN⁺ tumor budding cells. H Hematoxylin and eosin (H&E) staining of a public CRC sample with spatial transcriptome data. I Unsupervised clustering and cell type identification of the same spatial transcriptome data. J UMAP plot of spatial transcriptome spots, colors represent different subclusters. K Bubble plots showing differentially expressed genes between different clusters.

Next, we further examined the cellular communications between all fibroblast subtypes and MSLN⁺ budding-potential cells, and found that the CTHRC1⁺POSTN⁺ fibroblasts had the strongest interactions with the MSLN⁺ tumor budding cells (Fig. 5E). To determine the relationship between fibroblasts and MSLN⁺ budding-potential cells, we used CIBERSORTX to infer the infiltration abundances of different EMP1⁺KRT20^- undifferentiated cells and fibroblasts subtypes in the TCGA-CRC cohort, and the results showed a significantly positive correlation between MSLN⁺ budding-potential cells and CTHRC1⁺POSTN⁺ fibroblasts, which is consistent with the results of Cellchat analysis (Fig. 5F, G).

Since interactions between two types of cells require spatial proximity³⁶, we used spatial transcriptome datasets from the publicly available data Cancer Diversity Asia and GSE225857 to explore the spatial distribution of MSLN⁺ budding-potential cells and POSTN⁺ fibroblasts (Fig. 5H). Based on unsupervised clustering and characteristics of the spots, we classified the tumor tissue sections into 8 clusters. Cluster 1 near the central site of the tumor consisted mainly of cancer stem cells, whereas POSTN⁺ fibroblasts and MSLN⁺ epithelial cells co-localized at the same spots in the invasive front (Figs. 5H-J, S7A–E). In addition, we found high expression of the MSLN⁺ budding-potential cells’ marker genes (MSLN, SCEL, CST6, KLK10, TRIM29) in the region of the invasive front, further validating the location of MSLN⁺ budding-potential cells in the invasive front (Figs. 5K, S7F). Moreover, our high-resolution multiplexed IHC staining indicated that periostin (POSTN) surrounded the MSLN⁺ tumor buds at the invasive front (Fig. S8). This suggests a physical interaction between POSTN⁺ fibroblasts and MSLN⁺ budding-potential cells in the invasive front of CRC.

Additionally, we analyzed the immune microenvironment surrounding MSLN⁺ cells. Our results showed that immunoglobulin genes and T cells were mainly distributed in the peripheral area away from MSLN⁺ cells, which indicates that MSLN⁺ cells are located in an immunosuppressive microenvironment. Additionally, we observed colocalization between MSLN⁺ cells and SPP1⁺ macrophages, a unique macrophage subtype previously reported to be associated with immunosuppression and cancer metastasis (Fig. S7G)^13,37.

Crosstalks between POSTN⁺ fibroblasts and MSLN⁺ budding-potential cells in CRC

To further explore the crosstalk between POSTN⁺ fibroblasts and MSLN⁺ budding-potential cells, we analyzed the crucial signaling pathways between POSTN⁺ fibroblasts and MSLN⁺ budding-potential cells. We found enhanced signaling intensity of THBS, FN1, and MK signaling pathways, and POSTN⁺ fibroblasts were the major ligand source of paracrine signaling in cancer cell subpopulations (Fig. 6A–C). These signaling pathways were associated with extracellular matrix remodeling and tumor invasion^38–41, suggesting a potential invasion-promoting effect of POSTN⁺ fibroblasts on MSLN⁺ budding-potential cells. Then we analyzed critical ligand-receptor pairs involved in the communication between POSTN⁺ fibroblasts and MSLN⁺ budding-potential cells, and found that POSTN⁺ fibroblasts might interact with MSLN⁺ budding-potential cells through FN1-(ITGA3 + ITGB1), THBS2-CD47, and POSTN-ITGB5 interactions, which was in agreement with previous studies (Fig. 6D)^42–45. Also, MSLN⁺ budding-potential cells can interact with POSTN⁺ fibroblasts via MDK-LRP1, MDK-SDC2 communication network (Fig. 6D). MK signaling is essential for cancer-associated fibroblasts proliferation⁴⁶, further suggesting that cancer cells can maintain fibroblasts growth and activation via the MK signaling pathway. Periostin (POSTN) has been reported to bind to ECM and cell surface receptor ITGB5 to promote tumor cell migration and invasion⁴⁷. Therefore, we verified the association between the marker gene POSTN of POSTN⁺ fibroblasts, its receptor ITGB5 and MSLN⁺ budding-potential cells by IHC on consecutive CRC tissue slides. At the tumor invasive front, POSTN was mainly expressed in the tumor mesenchyme and wrapped around the tumor budding cells, and ITGB5 was expressed both in the tumor mesenchyme and in the tumor budding cells, while MSLN was positively expressed in the tumor budding cells, thus suggesting spatial proximity between POSTN⁺ fibroblasts, its receptor ITGB5, and MSLN⁺ tumor buds (Fig. 6E, F). To evaluate the effects of CAF-derived periostin on the invasiveness of MSLN⁺ cells, we treated SW480-MSLN cells with recombinant human Periostin (rhPOSTN) protein for 24 hours, then we evaluated the effect of POSTN on the metastatic capacity of MSLN⁺ cells using Transwell migration and invasion assays, and the results showed that rhPOSTN significantly enhanced the migration and invasion capabilities of SW480-MSLN cells (Fig. S9). These findings suggest that POSTN⁺ fibroblasts promote colorectal cancer invasion by interacting with MSLN⁺ budding-potential cells.

Fig. 6. Crosstalk between POSTN⁺ fibroblasts and MSLN⁺ tumor budding cells.

A–C Circle plots demonstrating the interactions of THBS, FN1, and MK signaling networks among the fibroblast subclusters and MSLN⁺ tumor budding cells. D Bubble plot demonstrating significant L-R pairs between CTHRC1⁺POSTN⁺ fibroblast and MSLN⁺ tumor budding cells. E, F Representative images of IHC staining for MSLN (left), ITGB5 (middle) and POSTN (right) at the invasive front region of CRC. Scale bar, 250 µm.

Discussion

In this research, we constructed a single-cell atlas of CRC by integrating single-cell transcriptomic data from 287 samples from eight datasets and identified a subpopulation of tumor cells associated with metastasis by comparing metastatic with non-metastatic CRC. More importantly, we found that this cluster was closely associated with tumor budding. By analyzing the interaction of epithelial cell subclusters with CAF subclusters, we further determined the interaction between MSLN⁺ tumor budding cells and POSTN⁺ fibroblasts, thus providing new insights into the molecular characterization of tumor budding cells and new potential therapeutic targets for the treatment of CRC metastasis.

Tumor metastasis is closely related to tumor budding cells, which grow infiltratively at the invasive margin of the tumor, enter the circulation, and seed at distant sites⁴⁸. Recent studies have shown that tumor budding reflects an aggressive growth pattern with EMT capacity^49,50. The expression of E-cadherin was significantly reduced in budding cells¹⁹. The EMT transcription factor ZEB1 and ZEB2 were significantly upregulated in tumor buds⁵¹. And the gene expression profile of tumor budding cells belongs to the EMT-like phenotype²⁰. These studies indicate that EMT plays a significant role in tumor budding. However, studies on molecular features and unique markers of tumor budding cells are still limited. Published studies have shown that metastasis of CRC arises from EMP1⁺ tumor cells¹⁵. Our analysis showed the presence of EMP1⁺KRT20⁺ differentiated cells as well as EMP1⁺KRT20- undifferentiated cells in the tumor microenvironment, consistent with previous studies¹⁵. On this basis, we further examined the subtypes of EMP1⁺KRT20^- undifferentiated cells and identified a previously unreported subcluster of tumor budding cells by tumor budding and epithelial-specific high-risk genesets. In addition, we found that this subcluster exhibited high expression of MSLN and enrichment of metastasis-related pathways such as EMT, angiogenesis, and apical junction. Importantly, we confirmed by IHC that overexpression of MSLN is specifically in tumor budding cells. Mesothelin (MSLN) is a cellular membrane-bound protein of unknown function. The Mesothelin gene encodes a 69 kDa precursor protein that is cleaved into a 31 kDa secreted fragment called megakaryocyte potentiation factor (MPF) and a 40 kDa membrane-bound protein called mesothelin (MSLN), which is a glycoprotein anchored to the plasma membrane through the glycophosphatidylinositol (GPI) structural domain^52,53. MSLN has been reported to be highly expressed in several solid tumors such as pancreatic, ovarian, and lung cancers^54–56, but reports in CRC are very limited. Our study demonstrates that overexpression of MSLN significantly enhances the proliferation, migration, and invasion of CRC cells, and reveals that MSLN enhances CRC progression in vivo orthotopic xenograft mouse model of CRC in mice. However, current research on MSLN-targeted therapy has focused on other solid tumors, such as pancreatic cancer and ovarian cancer^57–59 future studies should focus on the therapeutic efficacy of targeting MSLN⁺ tumor budding cells in CRC.

Tumor microenvironment plays a significant role in tumorigenesis and metastasis⁶⁰, in which cancer-associated fibroblasts and tumor cells have complex interactions and exhibit strong tumor-regulating effects. We determined the fibroblast profiles of normal, non-metastatic versus metastatic CRC tissues, where CTHRC1⁺POSTN⁺ fibroblasts showed significant enrichment of EMT, extracellular matrix remodeling, and angiogenesis pathways, which is in line with Chen C et al.‘s finding that POSTN⁺ CAFs promote lung cancer progression⁶¹. Also, our study suggests that POSTN⁺ fibroblasts may promote tumor budding at the invasive margin and proposes a possible mechanism for the interaction between POSTN⁺ fibroblasts and MSLN⁺ budding cells, which also provides new insights into targeting cancer-associated fibroblasts and tumor budding cells in CRC therapy. However, our current research still has limitations. We did not analyze the interaction between fibroblasts and MSLN⁺ cells using co-culture systems and conditioned media. We acknowledge that functional experiments are still needed to support our conclusions, and we may see progress in the future.

In summary, by establishing a publicly available dataset, this study has identified molecular features of tumor budding cells and highlighted the potential role of MSLN⁺ budding-potential cells together with CTHRC1⁺POSTN⁺ fibroblasts in CRC metastasis, which may provide clues for targeted therapeutic interventions of metastatic CRC in the future.

Methods

Single-cell RNA sequencing data collection and quality control

We collected eight publicly available CRC single-cell RNA sequencing datasets containing a total of 287 individual CRC primary tumor and adjacent normal mucosa samples from 134 donors, with tumor samples from patients with AJCC stage III/IV as the metastatic group, and the other tumor samples as the non-metastatic group^62–68. Details of all samples in the analysis are included in the Table S1. We used Seurat (v4.3.0) as well as the R package (v4.2.0) to process the data and perform dimensionality reduction clustering analysis. We performed quality control based on the following criteria: a gene was expressed in at least three cells, a cell contained more than 200 genes, >200 genes and <6000 genes, and less than 25% of mitochondrial genes, thus retaining 562,446 cells.

Dimensionality reduction and cluster analysis

We normalized the count matrix using the logNormalize method and determined the number of principal components (PCs) to be included in further analyses based on the generated jackstraw and Elbow plots. The FindVariableFeatures function was used to select 2000 highly variable genes with which to perform principal component analysis. To eliminate batch effects, we performed batch correction between samples and integrated the gene expression matrices of all samples using the R package harmony (v1.1.0)⁶⁹. We computed the first 30 principal components to cluster cells, used the Findneighbours function in Seurat to obtain the number of nearest neighbors, used the Findclusters function for clustering cell subtypes, and visualized cell clusters using the uniform manifold approximation and projection (UMAP) algorithm to visualize the cell clusters.

Clustering of major cell types

We clustered and annotated all cells based on known marker genes: epithelial cells (EPCAM, KRT18), B cells (CD79A, MS4A1), NK/T cells (CD3D, CD3E), fibroblasts (THY1, COL3A1), mast cells (KIT, TPSAB1), endothelial cells (PECAM1, VWF, ENG), plasma cells (MZB1) and myeloid cells (LYZ, CD68). We then performed unsupervised clustering analysis of epithelial and fibroblasts to annotate each cell subtype by manual analysis of differential genes between different clusters. Finally, we performed unsupervised clustering analysis of EMP1⁺KRT20^- epithelial cells using higher resolution to identify metastasis-associated rare subgroups using differential genes between different clusters. The R package CopyKAT (v1.0.4) was used to infer copy number variation in tumor epithelial cells by gene expression matrix⁷⁰. The parameters were set to ngene.chr = 5, win.size = 25, and KS.cut = 0.1.

TCGA bulk-RNA sequencing data collection

We used the R package TCGAbiolinks (v2.24.3) to download the TPM gene expression matrices and clinical information for TCGA-COAD and TCGA-READ from the TCGA database. We used the R package CMScaller (v2.0.1) for CMS typing of the gene expression matrices⁷¹. We set the number of iterations to 1000 with a p-value of <0.05 to run CMScaller. We included patients based on the following criteria,1. patients with OS, DFS time ≥30 days, 2. patients with complete TNM stage information, 3. patients with no CMS typing missing.

Inferring the composition of individual cell populations based on single-cell expression matrices

We used CIBERSORTX to determine the proportion of cells in each subpopulation of fibroblasts and EMP1⁺KRT20^-epithelial cells in each sample in the TCGA dataset⁷². We randomly selected 500 cells from each subpopulation as a single-cell RNA-Seq reference gene expression matrix to create the Signature Matrix. We set disable quantile normolization to analyze the RNA-Seq dataset, and all other parameters were set to default. We subsequently analyzed the correlation between different subpopulations using Spearman correlation; |Rs | > 0.3 and p < 0.05 were considered significantly correlated.

Differential expression analysis

We identified differentially expressed genes between subclusters by using the Findallmarkers function with parameters set to min.pct = 0.25 and logfc.threshold = 0.25. Genes with P < 0.05 and log2 (fold change) > 0.5 were considered differentially expressed genes.

Differential abundance analysis

Differences in cell abundance associated with metastasis status were tested using MiloR (v1.4.0). We first used the buildGraph function to construct a KNN graph based on precomputed supervised PCA with k = 30, using 30 principal components (d = 30). Next we generated neighborhoods by using makeNhoods with prop = 0.1, k = 30, d = 30, and we calculated the neighbourhoods distances using the calcNhoodDistance function. Differentially abundant neighbourhoods were detected using (SpatialFDR < 0.1, log(fold change) < 0) for metastatic neighbourhoods and (SpatialFDR < 0.1, log(fold change) > 0) for non-metastatic.

Transcription factor regulon analysis

We used pySCENIC to identify the transcription factor activities of different cell clusters, and inferred the transcription factor activities by the AUCell function in pySCENIC⁷³. We obtained differentially expressed transcription factors by comparing the transcription factor activities of each subcluster with other subclusters by Wilcoxon rank-sum test, selected the top 5 transcription factors of each subcluster and used the normalized transcription factor expression activities for heatmap visualization.

Functional enrichment analysis

Metascape (http://metascape.org/) was used to perform gene enrichment analysis of differential genes⁷⁴. Gene Ontology (GO), Kyoto Encyclopedia of Genes and Genomes (KEGG) and Reactome databases were used for enrichment analysis. To assess the expression levels of different signatures at single-cell resolution, we also performed gene set variance analysis of the Hallmark gene set by the R package GSVA (1.44.5), obtained from the R package msigdbr.

To assess the function of cell subclusters, we calculated its functional module score at the single-cell level using the AddModuleScore function in Seurat. Functional modules include tumor budding and epithelial-specific high-risk geneset (EpiHR). The gene lists can be found in Table S3. The gene list of EMT was obtained from the Hallmark gene lists that are publicly accessible at https://www.gsea-msigdb.org/gsea/msigdb/.

Cell-cell interaction analysis

To infer the interactions between epithelial cells and other major cells in the tumor microenvironment, we used Cellchat (v1.6.1) for ligand receptor analysis⁷⁵. We followed the standard Cellchat processing flow, including identifyOverExpressedGenes, identifyOverExpressedInteractions functions, and then we used the functions aggregateNet, computeCommunProbPathway and computeCommunProb to calculate the strength and probability of communication between different cell types for each L-R pair, filtering for communications involving less than 10 cells. The Netvisual function was used to visualize cellular communication networks for overall and individual signaling pathways. Bubble diagrams are used to visualize cellular communication between LR pairs of cell populations.

Analysis of spatial transcriptomics data

The CRC spatial transcriptomics data were obtained from public studies^76,77. We used Seurat (v4.3.0) to normalize, dimensionality reduce and cluster the spatial transcriptome samples, and dimensionality reduction was performed at resolution 1 with the first 30 principal components (PCs). Cluster analysis was performed using uniform manifold approximation and projection (UMAP). The major cell types of each spot cluster were annotated according to the differential genes expressed in the spots, and spatial gene expression features were visualized using the SpatialFeaturePlot function in Seurat (v4.3.0).

Survival analysis

Combining with the disease-free survival time, we finally performed a Kaplan–Meier survival analysis by R package survival (v3.3.1). We determined the optimal cutoff value using the surv_cutpoint function in the R package survminer(0.4.9), which categorizes samples into high and low groups. Log-rank p-values were calculated for statistical significance.

Cell culture and transfection

Human colon cancer cell lines SW480, RKO, and human embryonic kidney cells 293 T were purchased from the Chinese Academy of Medical Sciences (Beijing, China). All the above cells were cultured in Dulbecco’s Modified Eagle Medium (Gibco) containing 10% fetal bovine serum (FBS, NEWZERUM) and 1% antibiotics (100 U/ml penicillin and 100 mg/ml streptomycin) at 37 °C in a constant temperature incubator and 5% CO₂.

MSLN overexpression and control plasmids were synthesized by ABM and 293 T cells were transfected with MSLN overexpression and control plasmid together with packaging plasmid (psPAX2) and envelope plasmid (pMD2G) using Lipofectamine 3000 (Invitrogen) reagent, according to the manufacturer’s protocol. Lentivirus was collected, filtered and infected SW480, RKO cells and selected in puromycin-containing medium. Overexpression of MSLN was confirmed by Western blotting.

Western blotting

Protein was extracted from CRC cells using a RIPA buffer supplemented with phenylmethylsulfonyl fluoride (Applygen, Beijing, China). BCA Protein Assay Kit (Applygen, Beijing, China) was used to quantify protein concentrations. Protein lysates were separated by 10% sodium dodecyl sulfate–polyacrylamide gel electrophoresis (Yamay, Shanghai, China) and then transferred onto polyvinylidene difluoride membranes (Millipore, USA). The membranes were closed with 5% skimmed milk powder for 2 h, and then incubated with specific primary antibodies overnight at 4 °C with following dilutions: anti-Mesothelin (1:1000, 99966 T, Cell Signaling Technology), anti-GAPDH (1:10,000, GB15004-100, Servicebio), followed by washing in TBST. The membranes were then incubated with the corresponding secondary antibodies (1:10,000, AS014, Abclonal) for 2 h at room temperature and then washed 3 times (10 min each) in TBST. Protein expression levels were detected by ECL Plus (Solarbio, Beijing, China) using a bioimaging system (Bio-Rad, USA).

Cell proliferation assays

Cell proliferation was assessed using the Cell Counting Kit-8 (Lifi-ilab, Shanghai, China). Cells were inoculated in 96-well plates (5000 cells/well) and cultured in DMEM with 10% FBS for 4 days. In this experiment, 10 μl of CCK-8 reagent was added to each well and incubated at 37 °C for 2 h. Absorbance was measured spectrophotometrically at 450 nm.

Wound healing assays

RKO, SW480 cells were inoculated in 6-well plates. Once they have reached almost 100% confluence, a uniform scratch is made in the middle of the 6-well plate using a 200 μL pipette tip and then washed 3 times with PBS. The plates were replaced with serum-free DMEM at 0 and 48 h after scratching, observed for wound healing, and photographed using an inverted microscope. The cell wound healing rate was analyzed using ImageJ software.

Transwell migration and invasion assays

For the Transwell migration assay, a total of 1 × 10⁵ cells were suspended in 200 μL of serum-free medium and added to the Transwell upper chamber, and 600 μL of complete medium containing 10% FBS was added to the lower chamber and incubated for 24 h. Cells were then fixed in 4% paraformaldehyde for 15 min and stained with crystal violet for 10 min. Cells were wiped from the Transwell chamber layer with a cotton swab and photographed under a microscope to observe and count the cells. Data were analyzed using ImageJ software. Five random areas were selected for cell counting. SW480-MSLN cells were pretreated for 24 h with 500 ng/mL recombinant human Periostin (rhPOSTN, MCE, HY-P74638), then seeded (5 × 10⁴ cells/well) into the upper chambers.

For the Transwell invasion assay, the upper chamber of the Transwell was pre-encapsulated with matrix gel (BD, USA), and the subsequent experimental steps were consistent with those of the Transwell migration assay. SW480-MSLN cells were pretreated for 24 h with 500 ng/mL recombinant human Periostin (rhPOSTN), then seeded (1 × 10⁵/well) into the upper chambers.

Animal study

Four-week-old female BALB-c-nu mice were purchased from Beijing Vital River Laboratory Animal Technology (China). In order to construct a CRC orthotopic xenograft mouse model, 2 × 10⁶ MSLN overexpressing and control SW480 cells stably expressing luciferase were injected subcutaneously into the right dorsal side of mice. When the subcutaneous CRC tumors were visible to the naked eye (volume < 1.5 cm³), the tumors were detached and sliced into 1 mm³ sections and embedded in the mesentery of the mice. 6 weeks later, the mice were injected intraperitoneally with 150 mg/kg of D-luciferin (Lifi-ilab, Shanghai, China), and the fluorescence intensity was measured 10 min later using an Aniview 600 imaging system (Guangzhou, China). All animal experiments were approved by the Ethics Committee of Animal Experimentation of Peking University Third Hospital (Ethics No. A20240151).

Immunohistochemistry (IHC) staining

After obtaining the written informed consent of CRC patients, we collected tumor samples from 30 patients in Peking University Third Hospital. Relevant clinical information.is presented in Table S2. Studies involving human participants were approved by the Ethics Committee for Research in Medical Sciences (Ethics No. IRB00006761-S20250477) and were conducted in accordance with the Declaration of Helsinki. Informed consent was obtained from all participants or their guardians. According to the 2016 International Tumor Budding Consensus Conference (ITBCC) recommendations for reporting tumor budding in colorectal cancer, grading of tumor budding is recommended using the following three-tiered system used by the Japanese Society of Colon and Rectal Cancer: low grade budding (Bd 1), 0-4 buds; intermediate grade budding (Bd 2), 5-9 buds; and high-grade budding (Bd 3), 10 or more buds. The experimental specific procedures for immunohistochemical staining have been clearly described in previous studies⁷⁸. The primary antibodies used in this study included anti-Mesothelin (99966 T, Cell Signaling Technology), anti-Periostin (ab215199, Abcam), anti-Integrin β5(ab309092, Abcam).

Fluorescent multiplex immunohistochemistry (mIHC)

Paraffin-embedded CRC tissue sections of 4 μm thickness were prepared and mounted onto microscope slides. Slipping slides in xylene for dewaxing, followed by rehydration through a series of reducing ethanol solutions. Subsequently, the slides were placed in EDTA (pH 9.0) buffer for microwave treatment to perform antigen retrieval. After a 10-minute incubation period, incubate the tissue with primary antibodies, including anti-Mesothelin (99966 T, Cell Signaling Technology) and anti-Periostin (ab215199, Abcam). Then incubate with HRP-conjugated secondary antibody and tyramide signal amplification (TSA) using Opal. Subsequently, the primary and secondary antibodies were stripped together via microwave treatment. Repeat the above process for each primary antibody. Subsequently, the tissue was stained with DAPI. Finally, use the scanner to perform a full-slice scan.

Statistical analysis

Statistical analysis was performed using R version 4.2.0 and GraphPad Prism 9.5.0. Differences between the two groups were analyzed using Student’s t-test or Wilcoxon rank sum test. The proportions between different groups were assessed using unpaired two-tailed chi-square tests. All experiments were validated in at least 3 biological replicates. Data are expressed as mean ± standard deviation (SD). A p-value of less than 0.05 was considered statistically significant. ∗p < 0.05, ∗∗p < 0.01, ∗∗∗p < 0.001, ∗∗∗∗p < 0.0001.

Author contributions

Y.M. analyzed the data, performed most of the experiments and wrote the manuscript; X.Z. conceived, designed experiments, interpreted data, and wrote and reviewed the manuscript; L.J.W., Z.Y.Z, Y.F.Z. performed experiments; R.J.L., J.W.W., Z.M.Z. provided clinical samples; L.M.G. provided pathological expertise; W.F. and K.M. supervised the study. All authors read and approved the manuscript.

Data availability

The data used in this study have been summarized in Methods and all data are publicly available.

Code availability

The code that supports the findings of 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/s41698-025-01187-y.