Metastatic immune microenvironment remodeling by malignant epithelial cells drives colorectal cancer hepatic metastasis: Multi-omics insights Authors: Wenlong Zhu, Ao Dong, Rongbing Hu, Silei Zhou, Jiali Yao, Binghua Ma, Jin Ding, Yanjie Zhang Published: 2026 ABSTRACT Highlights•First identified Scissor⁺ malignant cells driving CRC HM.•Constructed a Scissor-related prognostic signature for CRC patients with HM.•Elucidated mechanisms of Scissor⁺ malignant cells worsen the microenvironment of HM.•Firet confirmed Scissor⁺ malignant cells polarize macrophages into SELENOP⁺ TAMs.Keywords: Hepatic metastatic, CRC, Multi-omics, Scissor+ epithelial cells, Prognostic signature, CD4+ Treg, SELENOP+ TAM Background Colorectal cancer (CRC) is one of the most prevalent malignant tumors. According to the 2025 GLOBOCAN data issued by the International Agency for Research on Cancer (IARC) of the World Health Organization, both the incidence and mortality rates of CRC rank third, presenting a significant challenge within the global cancer prevention and control system [1,2]. CRC frequently metastasizes to various organs, with the liver being the most common site of such spread [3]. CRC-related deaths are predominantly attributed to liver metastases, which develop in nearly 50% of patients through hematogenous dissemination. Importantly, 20–25% of such patients exhibit synchronous hepatic metastases (HM) upon initial diagnosis of the primary tumor (PT) [4]. Although surgical approaches, chemotherapy regimens, and targeted therapeutic strategies have achieved considerable improvements, individuals diagnosed with metastatic or advanced CRC consistently exhibit unfavorable clinical outcomes [5]. The tumor microenvironment (TME) constitutes a highly dynamic and intricate ecosystem that exerts pivotal effects on tumor proliferation, disease progression, patient prognosis, and therapeutic responsiveness [6]. The dismal prognosis of patients with metastatic colorectal cancer (mCRC) is largely attributable to the intricately immunosuppressive tumor microenvironment (TIME) at metastatic lesions [7]. Nevertheless, the intricate pathophysiological circumstances constrain the therapies for CRC patients with HM, and the mechanism of CRC HM remains inadequately investigated at present. Hence, elucidating the molecular mechanism of CRC HM and determining potential therapies are of paramount significance. Epithelial cells, as one of the key sources of tumor stem cells, constitute a vital component of the TME, which play an essential role in the process of tumorigenesis [8]. EMT serves as a core biological mechanism enabling tumor epithelial cells to enhance metastatic capacity. Through the process of EMT, characteristics associated with a more aggressive phenotype can be acquired by tumor epithelial cells [9,10]. Single-cell RNA sequencing (scRNA-seq) enables the dissection of heterogeneous tumors and elucidates cellular compositions, thereby clarifying the distinctions between CRC HM and PT. Previous studies have systematically clarified the heterogeneity of tumor epithelial cells and the immune landscape in CRC PT through scRNA-seq. However, research regarding the role of tumor epithelial cells in HM remains insufficient [11,12]. Recent advances in multi-omics technologies have enabled in-depth characterization of epithelial tumor heterogeneity and immune landscapes within the TME of CRC HM. Therefore, it is of utmost importance to clarify the crosstalk between malignant epithelial cells and immunosuppressive cells in CRC HM. Furthermore, developing therapeutics that disrupt such intercellular crosstalk is critical to overcoming drug resistance and alleviating local immunosuppression. Such efforts aim to enhance therapeutic efficacy and improve prognosis, aligning with the principles of precision medicine. An increasing number of evidence indicates that TME has a significant impact on the progression, treatment response and clinical outcome of CRC. Multiple immunosuppressive cells within the tumor microenvironment (TME), including tumor-associated macrophage (TAM), tumor-associated neutrophil (TAN), and cancer-associated fibroblast (CAF), drive tumor progression, invasion, and metastasis via secretion of diverse cytokines [13,14]. In a current study investigating CRC HM, it was established that ANGPTL2+ CAF and SPP1+ TAM significantly contribute to the promotion of metastases in CRC, indicating that the immune dysregulation mediated by numerous immunosuppressive cells in the TME plays a significant role in CRC HM [15]. At present, TME has been recognized as an indispensable component of malignant biology and a promising therapeutic target. Heterogeneity among intratumoral immune cells has prompted the consensus that immune signatures of the TME carry substantial prognostic and predictive relevance for CRC patients. Although the immune characteristics present within the TME can provide valuable insights for patient prognosis, the anti-tumor efficacy of immunotherapy in the majority of CRC patients remains limited. This limitation is primarily attributed to the complex crosstalk among immunosuppressive cells within the TME, which ultimately diminish the effectiveness of immunotherapies [16,17]. Furthermore, reciprocal crosstalk between malignant epithelial cells and immunosuppressive cells drives the immunosuppressive remodeling of the TME [18,19]. However, current researches on how malignant epithelial cells affect immunosuppressive cells are still insufficient. Thus, it is imperative to dissect the molecular mechanisms whereby malignant epithelial cells accelerate HM and reshape the TIME via multi-omics technologies. In this study, multi-omics technologies were utilized to investigate the characteristics of the TME in CRC patients with HM. Firstly, Scissor+ malignant epithelial cells and immunosuppressive cells were identified, while also exploring the mechanisms by which these cells exacerbate the TME in HM. Furthermore, a prognostic signature related to HM was developed based on the characteristics of Scissor+ malignant epithelial cells and validated for its superiority over 16 existing signatures. Finally, we examined the differences between pre-hepatic metastasis (PM) and HM within the TME of advanced CRC mouse models. Collectively, our findings elucidate the dynamic immune landscape in CRC patients with HM, establishing a foundational framework for deciphering stage-specific immune characteristics across disease progression. These insights may pave the way for identifying novel therapeutic targets in subsequent research. Materials and methods Data collection and preprocessing Four public datasets of scRNA-seq were obtained from the Gene Expression Omnibus (GEO) database (GSE132465 [20], GSE245552 [21], GSE158328 [22] and GSE284449 [23]). All human samples were combined into an integrated dataset, which comprises 2 normal colorectum (NC) samples, 8 PT samples and 2 HM samples. This comprehensive dataset elucidates the landscape of the TME, encompassing the progression from normal colon to CRC and HM. The clinical and pathological characteristics of the patients involved in this study are outlined in Table S1.1. Additionally, six mouse liver tissues were incorporated into another dataset comprising two normal liver (NL) samples, two PM samples and two HM samples. This dataset was generated to identify differences in the hepatic microenvironment between pre-metastatic and post-metastatic stages of CRC HM. Furthermore, bulk cohorts were obtained from The Cancer Genome Atlas (TCGA) and GEO databases to enhance the credibility of scRNA-seq analysis results, which includes TCGA-COAD, GSE17536, and GSE38832 [24,25]. The expression matrices, along with the corresponding clinical and pathological data from bulk cohorts, were utilized for immune infiltration analysis and the construction of prognostic signatures. Moreover, ST data were employed to validate the spatial positional relationships of different clusters to validate the results of scRNA-seq, among which included two datasets, GSE206552 and GSE280313 [26]. scRNA-seq data analysis ScRNA-seq analysis was performed using the Seurat package (version 4.4.0) in RStudio [27]. For quality control, cells with fewer than 300 detected features or a mitochondrial gene fraction greater than 20% were excluded. Following this standardized filtering pipeline, a dataset comprising >40,000 cells was constructed for further analyses. After quality control, standard Seurat workflows were applied, encompassing "NormalizeData", "FindVariableFeatures", "ScaleData", "RunPCA", "FindNeighbors", "FindClusters", "Harmony (Resolution=1.0)", "RunUMAP", "RunTSNE", and "FindAllMarkers" analyses. The "SingleR" package was utilized for cell type annotation, with reference to experimentally verified markers from published literature to enhance annotation accuracy and credibility [28]. Additionally, the "org.Hs.eg.db" and "ClusterGVis" R packages were utilized to investigate the representative genes and biological functions associated with each cell type [29]. Furthermore, Gene Set Variation Analysis (GSVA) was performed to investigate pathway enrichment across various cell types [30]. Finally, The "AddModuleScore" was employed to assess custom gene sets for evaluating the functional status across a variety of immune cells. The genes included in the custom gene sets are presented in Table S8.3 and Table S9.3. Ro/e analysis The "epitools" R package was utilized to calculate the Ro/e value, enabling a systematic quantification of enrichment scores for various cell types across different pathological groups. A value exceeding 1 indicates a tendency for enrichment of that specific cell type within the group [31]. CNV comparison analysis The "inferCNV" algorithm was used to assess chromosomal copy number variations among distinct epithelial clusters. Normal epithelial cells were used as reference [32]. Pseudo-time analysis Three computational tools, including "Slingshot", "Cytotrace", and "Monocle2", were applied to infer the developmental trajectories of distinct clusters derived from identical cell types [[33], [34], [35]]. Scissor analysis The Scissor algorithm was utilized to integrate bulk transcriptomic data and phenotypic information for optimized scRNA-seq data analysis. Briefly, Scissor requires three core input datasets: a single-cell expression matrix, a bulk expression matrix, and matched phenotypic information (such as clinical survival data), which were primarily acquired from TCGA database. The analytical pipeline first quantifies pairwise similarities between individual single cells and bulk samples using Pearson correlation coefficients. Subsequently, a regression model was optimized based on the correlation matrix and sample phenotypic profiles, with Cox regression applied for clinical survival outcomes. According to the signs of the estimated regression coefficients, cells with non-zero coefficients were classified as Scissor-positive (Scissor+) or Scissor-negative (Scissor-) cells, which were positively or negatively correlated with the investigated phenotype, respectively. Cells with zero coefficients were defined as background cells [36]. Cell-cell interaction analysis The "Cellchat" (version 1.1.0) and "NicheNet" (version 1.0.0) were employed to investigate the potential crosstalk between malignant epithelial cells and immunosuppressive cells, as well as to identify possible ligands and receptors involved in this interaction [37,38]. Spatial transcriptomics analysis The "Seurat" (version 4.4.0) and "hdf5r" packages were used to process and analyze ST data. Following rigorous quality control, the SpatialFeaturePlot function embedded in the Seurat package was applied to visualize the spatial expression patterns of key prognostic genes. Spots with positive expression scores were defined as high-expression regions. Additionally, the "CellChat" package (version 2.1.2) was utilized to predict potential ligand-receptor pairs based on single-cell analysis [39]. Immunological characteristics analysis The "Immuno-Oncology Biomarker Research" (IOBR) package was utilized to elucidate the immunological characteristics and responses related to immune therapy in malignant epithelial clusters, which contains eight immune infiltration analysis algorithms including "MCPcounter", "CIBERSORT", "IPS", "Quantiseq", "TIMER", "EPIC", "xCell" and "Estimate". The TCGA-COAD cohort was categorized into Scissor+ and Scissor- groups based on the expression of Scissor+ genes by single-sample gene set enrichment analysis (ssGSEA). Then, we obtained comprehensive data from the TCGA-COAD cohort, which detailed the features of "Tumor Signatures", "EMT" and "Immune Signatures" between the two groups [40]. Developing a prognostic signature using 101 machine learning algorithms Three bulk cohorts were used to construct the prognostic signature (TCGA-COAD, GSE17536 and GSE38832). The TCGA cohort, which provided comprehensive treatment information, served as the training dataset to investigate the relationship between Scissor+ genes in HM and prognosis. The remaining cohorts were employed as validation datasets. The "MIME" package, which encompasses 101 machine learning algorithms, to develop the prognostic signature while simultaneously calculating the Scissor-related risk score (SRRS) for each patient [41]. Each algorithm underwent rigorous training on the TCGA-COAD cohort and was subsequently evaluated on the GSE17536 and GSE38832 cohorts to ensure robustness and mitigate the risk of overfitting. The most reliable signature was assessed by the performance of various machine learning models across multiple cohorts. For each CRC patient, the SRRS was calculated based on the final signature. Subsequently, patients within each cohort were categorized into SRRS-high and SRRS-low groups based on the median value, which served as the threshold. The log-rank test was employed to evaluate statistical significance, with a significance level set at P ≤ 0.05. Kaplan-Meier (K-M) survival analysis was performed utilizing the "survminer" R package to assess the prognostic relevance of the SRRS [42]. Then, a nomogram model was developed using clinical data related to survival [43]. Finally, the prognostic predictive capability of the final signature was compared with 16 established CRC-related prognostic signatures. The "timeROC" package was utilized to conduct calibration curve analysis and receiver operating characteristic (ROC) curve assessments in order to evaluate the predictive accuracy of the SRRS [44]. Cell lines All cell lines employed in this study, including HT29, HCT116, SW620, NCM460 and THP-1, were procured from the American Type Culture Collection (ATCC). The five cell lines were cultured in Dulbecco’s Modified Eagle Medium (DMEM), supplemented with 10% fetal bovine serum (FBS) and 1% penicillin/streptomycin. The cultures were maintained at 37 °C in a humidified atmosphere containing 5% CO2 [[45], [46], [47], [48]]. Notably, PMA (HY-18,739, MCE) is used to induce adherence of THP-1 cells, after which the adherent THP-1 cells are stimulated with conditioned medium (CM) for cell co-culture [49]. Reverse transcription quantitative polymerase chain reaction (RT-qPCR) Total RNA was isolated from CRC cell lines (HT29, HCT116, and SW620) and the normal human colorectal cell line NCM460 using the Cellular RNA Quick Purification Kit (10606ES60; ESscience Biotech). After RNA extraction, complementary DNA (cDNA) was reverse-transcribed with the Hifair® Ⅱ 1st Strand cDNA Synthesis Kit (11119ES60; ESscience Biotech). RT-qPCR was subsequently performed to detect the expression levels of malignant epithelial genes in cultured cell lines, thereby validating the results of bioinformatic analyses. All RT-qPCR assays were conducted on the LightCycler® 480 System (Roche) using the RT-qPCR Kit (11203ES03; ESscience Biotech). Relative mRNA expression was quantified via the comparative cycle threshold (Ct) method and normalized to β-actin. The ΔCt value was calculated as follows: ΔCt = Ct (target gene) - Ct (β-actin). All experimental data were presented as the mean ± standard error of the mean (SEM) from three independent biological replicates. The primer sequences utilized in this study are detailed in Table S6. Western blotting Total cellular protein was extracted from the aforementioned cell lines using RIPA lysis buffer (PC102, Epizyme Biotech) supplemented with 1:100 diluted Protease Inhibitor Cocktail (GRF101, Epizyme Biotech). Protein concentrations were determined with a BCA protein quantification kit (Beyotime Institute of Biotechnology) following the manufacturer’s instructions. Equal quantities of total protein (25 μg per sample) were resolved by 10% SDS-PAGE and subsequently electrotransferred onto nitrocellulose (NC) membranes. After transfer, membranes were blocked with 5% non-fat dry milk (Beyotime Institute of Biotechnology) at room temperature for 1–2 h. Subsequently, the membranes were incubated overnight at 4 °C with primary antibodies as listed below: HEBP2 (1:1000, H00023593-B01P, mouse monoclonal, Thermofisher), HSPB1 (1:500, 740,230 M, mouse monoclonal, Thermofisher), ZFP36L1 (1:1000, PA5–68592, rabbit polyclonal, Thermofisher), SELENOP (1:1000, PA5–112707, mouse monoclonal, Thermofisher), GAPDH (1:10,000, ab181602, rabbit polyclonal, Abcam), and β-actin (1:10,000, ab8226, mouse monoclonal, Abcam). After primary antibody incubation, membranes were probed with corresponding secondary antibodies at room temperature for 2 h. Target protein bands were visualized using FemtoLight ECL Substrate together with matching oxidant reagent (SQ201L, Epizyme Biotech). Densitometric quantification of western blot band intensities was performed via ImageJ software (Version 1.8.0.112, National Institutes of Health, USA) [50]. Transwell migration assays Cell migratory capacity was assessed using 24-well Transwell plates fitted with 8 μm pore polycarbonate membrane inserts (Corning, Cambridge, MA, USA). Briefly, 1 × 105 cells resuspended in 200 μL serum-free medium were seeded into the upper Transwell chamber, whereas 1000 μL complete medium supplemented with 20% FBS was loaded into the lower chamber as the chemoattractant. Following incubation for 24–48 h, non-migrated cells remaining on the upper surface of the membrane were carefully wiped away with cotton swabs. Cells that had migrated across the membrane were fixed with 4% paraformaldehyde for 30 min and subsequently stained with crystal violet solution for 5 min (Solarbio, Beijing, China). For quantitative analysis, three random microscopic fields per chamber were selected, and migrated cell numbers were quantified under an optical microscope [51]. Statistical analysis All statistical analyses were performed using R software (version 4.4.0). For normally distributed continuous variables, the Student’s t-test was used to compare differences between two groups. The non-parametric Wilcoxon test was applied for continuous variables that failed to conform to a normal distribution. One-way analysis of variance (ANOVA) was utilized for multi-group comparisons of normally distributed continuous data, whereas the Kruskal-Wallis test was adopted for non-parametric multi-group comparisons. Spearman’s rank correlation analysis was conducted to evaluate the correlation between continuous variables. Kaplan-Meier (K-M) survival curves were plotted for survival analysis, and the log-rank test was used to compare survival differences between groups. A two-tailed P-value < 0.05 was defined as the threshold for statistical significance to balance type I and type II errors in all statistical tests. Results Landscapes of CRC HM-associated TME To systematically dissect the immune microenvironmental landscape of CRC HM, separate multi-omics analyses were conducted on human and mouse samples by integrating scRNA-seq, bulk cohort and ST data (Fig. 1A). From a total of 12 patient samples of CRC PT, HM and NC tissus, along with 6 mice samples of CRC HM, PM and NL tissues from several datasets (GSE132465 n = 6, GSE284449 n = 6, GSE245552 n = 4, GSE158328 n = 2), the scRNA-seq data were integrated to provide a comprehensive insight into the HM-associated TME. This analysis included eight human primary CRC tissues across TNM stages Ⅰ to Ⅳ, two hepatic metastatic tissues, and two normal colorectal tissues. Matched mouse tissues were obtained with two samples per group (Table S1.1). The dataset proportions and sample compositions of human and mice, including the TNM stages were shown in Fig. 1B Landscapes of CRC HM-associated TME (A) The study workflow. (B) Pie chart showing the proportion of datasets and the composition of human and mouse samples, along with TNM stages of tumor samples. (C) UMAP plot displaying cell distribution by each cell type. (D) UMAP plot showing marker gene expression for each cell type. (E) Heatmap illustrating tissue preferences of different cell types across Ro/e analysis. (F) Bar chart displaying the proportion of different cell types in each sample. (G) UMAP plot displaying the cell distribution in Stage IV and HM samples, while the bar chart illustrating the proportion of each cell type. After the standard processing and quality control of the raw scRNA-seq data, a total of 41,396 cells were retained for subsequent analysis (Table S1.2). Through unsupervised clustering analysis, seven distinct cell clusters were identified in the integrated dataset: T/NK cells (CD3D), myeloid cells (LYZ), epithelial cells (EPCAM), B cells (CD79A), fibroblasts (ACTA2), endothelial cells (VWF), and mast cells (TPSAB1) based on marker gene expression profiles (Fig. 1C-D). Moreover, the Ro/e analysis indicated that the NC was predominantly composed of epithelial cells and fibroblasts. In contrast, HM tissues exhibited a significant reduction in T/NK cells, B cells, endothelial cells, and MAST, alongside a notable increase in myeloid cell populations when compared to PT tissues (Fig. 1E). In addition, the immunological and source groups of seven clusters are shown in Fig. S1A-B. Meanwhile, we conducted a cell ratio analysis to identify the distribution of the seven clusters across different source groups for each sample. The results obtained from this analysis are consistent with those derived from the Ro/e analysis (Fig. 1F). Interestingly, the T/NK and myeloid cells was increased obviously in HM tissues compared to stage Ⅳ PT tissues, indicating that the existence of some clusters promoting HM in T/NK and myeloid cells (Fig. 1G). Heterogeneity of epithelial cells in CRC progression To explore the transcriptional heterogeneity of CRC during the progression from PT to HM, epithelial cells were extracted from the integrated dataset and subjected to unsupervised subclustering analysis. In total, eight discrete epithelial cell clusters were identified via UMAP dimensionality reduction, which intuitively visualized the transcriptional heterogeneity among these clusters (Fig. 2A). Meanwhile, the source groups of the eight epithelial clusters were presented in Fig. S2A. Then, a comprehensive analysis of cellular heterogeneity was conducted to identify the distinct tissue-specific distribution of each epithelial cluster. The NC group exhibited a predominant enrichment of C2 (71.4%), C6 (14.1%), and C7 (14.1%). In contrast, when compared to PT, HM demonstrated a significant expansion of C3 (53.5%vs 17.9%), C4 (22.2%vs 10.3%), C5 (20.5%vs 5.3%), and C8 (3.7%vs 1.6%) (Fig. S2C; Table S2.1). The Ro/e analysis substantiated these findings, highlighting C2, C6 and C7 as clusters enriched in NC and identifying C3, C4, C5 and C8 as clusters predominantly associated with HM (Fig. S2B). Subsequently, systematic identification of differentially expressed genes (DEGs) and enriched functional pathways across individual epithelial clusters was performed to uncover unique transcriptomic signatures of each cluster (Fig. 2B). Notably, C3 was characterized by high expression of GAS5, NME2, and RNASEK, accompanied by significant enrichment of the p53, Hippo, and Wnt signaling pathways. C4 exhibited elevated levels of HMGB1, IFITM1, and TPX2, with marked enrichment of pathways involved in cell cycle progression, DNA replication, and homologous recombination. C5 displayed upregulated expression of immunotherapy and drug resistance-related genes, including MTRNR2L12, SOX4, and JUN. This cluster was significantly enriched in the MAPK and PI3K-Akt signaling pathways and showed prominent PD-L1 expression and activated PD-1 checkpoint signaling in cancer. C8 highly expressed SPINK4, DEFA5, and TNFRSF11B, with functional enrichment in ATP-dependent chromatin remodeling, cancer-related transcriptional dysregulation, and lipid metabolism processes. Given the preferential enrichment of clusters C3, C4, C5, and C8 in HM, these epithelial subclusters exhibited highly consistent gene expression profiles (Fig. 2B) (Table S2.2). However, due to the limited representation of epithelial cells in C8, this cluster was excluded from further analysis to ensure both analytical rigor and biological relevance. The expression of marker genes in each epithelial cluster across different groups were revealed by using ST analyasis. CCL20 of C1 cluster was more enriched in PT tissues but less enriched in HM tissues. Similarly, FCGBP of C2, CLCA1 of C6 and SLC26A3 of C7 were less both in PT and HM tissues, while was enriched in NC. RNASEK of C3, IFITM1 of C4, SOX4 of C5 and TNFRSF11B of C8 were more enriched in HM compared to PT (Fig. 2C). Heterogeneity of epithelial cells in CRC progression (A) UMAP plot displaying the landscape of various epithelial clusters. (B) Heatmap showing marker genes and pathway enrichment in various epithelial clusters. (C) Spatial transcriptome analysis revealing marker gene expression in each epithelial cluster across different groups. To assess the malignant potential of each epithelial cluster, the "inferCNV" algorithm was employed to calculate and identify large-scale chromosomal copy number variations (CNVs) for each cluster. The NC-enriched C2 demonstrated lower CNV scores in comparison to other epithelial clusters enriched in PT and HM, thereby serving as a non-malignant reference for comparative analysis. Notably, C3, C4, C5 and C8 were observed elevated CNV scores relative to PT counterparts, with recurrent amplifications observed in chromosomes 1, 2, 3, 6, 7, 8, 12, 13 and 20 (Fig. S2D-E). Subsequently, GSVA analysis was employed to reveal the pathway enrichment of each epithelial cluster for differentiating their specific oncogenic pathways. Specifically, C3 exhibits significant activation of tumor immune escape mechanisms, particularly TNFA_SIGNALING_VIA_NFKB and TGF_BETA_SIGNALING. C4 is functionally linked to tumor proliferation dynamics, with prominent representation of G2M_CHECKPOINT and OXIDATIVE_PHOSPHORYLATION. C5 is associated with tumor homeostasis modulation, showing predominant involvement of APOPTOSIS and ANGIOGENESIS (Fig. S2F). These findings indicate that C3 facilitates immune evasion, C4 drives tumor proliferation, and C5 contributes to tumor regulation-collectively promoting the development of HM. Pseudo-time analysis of epithelial cells during CRC HM progression Pseudotemporal trajectory analysis based on "Slingshot" and "Monocle 2″ was implemented to characterize the differentiation dynamics of epithelial cell clusters during hepatic metastatic progression. Four distinct pseudotemporal lineages were revealed through the evolutionary dynamics of the metastatic microenvironment by "Slingshot". Due to the significant enrichment of C2 in NC, it was considered the start point of differentiation. Epithelial cell trajectory mapping indicates that C6 represents an intermediate transitional state, whereas C1, C3 and C4 exhibit terminal differentiation states (Fig. 3A). Interestingly, lineage 1 and lineage 2 disclosed that C3 and C4 were at the terminal stage of differentiation. Lineage 3 unveiled the differentiation status of the epithelial clusters in PT, while lineage 4 revealed the differentiation circumstances among NC (Fig. 3B). Furthermore, the differentiation trajectories were stratified into three distinct states by "Monocle 2″, and it was further validated the aforementioned results of the trajectory analysis. Notably, C3 and C4 exhibited strong enrichment in terminal differentiation stages (State 3) (Fig. 3C-E, Fig. S3A). This finding further substantiates the presence of a distinct epithelial differentiation trajectory in HM compared to PT. Moreover, "CytoTRACE" revealed that C3 and C4 progressively developed enhanced tumor stemness characteristics throughout the differentiation process. This finding highlights their robust differentiation capacity and increased malignant potential in comparison to other clusters. (Fig. 3F-G). Pseudo-time analysis of epithelial cells during CRC HM progression (A) UMAP plot illustrating Slingshot-based pseudo-time analysis of developmental trajectories in epithelial clusters. (B) UMAP plot of “lineage1–4″ analyzed by "Slingshot". (C-E) Trajectory analysis of epithelial clusters by "Monocle 2″. (F) Trajectory analysis of epithelial clusters by "Cytotrace". (G) Box plot displaying differentiation scores of various epithelial clusters calculated by "Cytotrace". (H) Ridgeline plot showing the distribution of epithelial clusters across Pseudo-time, accompanied by a heatmap of marker gene changes by "Monocle 2". (I) Bar plot displaying genes that correlate with the Cytotrace score. Gene enrichment analysis across all epithelial clusters identified dynamic temporal expression patterns of HM-associated genes (Fig. 3H-I). Notably, C2, C6, and C7 exhibited a progressive downregulation of their enriched genes throughout tumor progression, which aligns with the observed loss of tissue homeostasis associated with malignancy. In contrast, HM-related genes including NME2, EEF1G, and HMGB1 were sustainably upregulated in C3 and C4 throughout CRC HM, with their activation levels markedly higher than those observed in other epithelial subclusters. Thus, these findings indicate that C3 and C4 function as malignant hubs in the CRC HM. These hubs are characterized by terminal differentiation states, a progressive acquisition of tumor stem cell properties, and sustained activation of pro-metastatic genes such as NME2, EEF1G and HMGB1, collectively driving immune-evasive metastatic adaptation and contributing to microenvironmental immunosuppression. Prognostic relevance of malignant epithelial cells and impact on the TIME in CRC To explore the clinical significance of epithelial clusters in CRC HM, gene signature profiling based on specific gene sets was performed. By integrating bulk RNA-seq data from 429 CRC patients in the TCGA-COAD cohort, the "Scissor" algorithm was applied to quantify the correlation of each epithelial cluster with distinct phenotypic groups in scRNA-seq datasets. Notably, Scissor+ epithelial cells were significantly enriched in C3 and C4 (Fig. 4A-B). Consistently, Scissor+ epithelial cells were specifically enriched in C3 and C4, whereas Scissor- epithelial cells exhibited significant enrichment in C6 through Ro/e analysis. Interestingly, Scissor+ cells were significantly enriched in HM, while Scissor- cells exhibited a specific enrichment in PT and NC (Fig. 4C). This finding further accentuates the promotional role of Scissor+ epithelial cells in CRC HM. Owing to the relatively small quantity of Scissor- cells, their influence on PT can be neglected. This research merely discusses the effects of Scissor+ and Scissor- epithelial cells on HM. Prognostic relevance of Scissor+ epithelial cells and impact on the TME in CRC (A) UMAP plots showing epithelial landscape across Scissor groups. (B) Bubble plot showing cell counts of epithelial clusters across Scissor groups. (C) Heatmap displaying tissue preferences of epithelial clusters across Scissor groups via Ro/e analysis. (D) Volcano Plot displaying up-regulated and down-regulated genes in Scissor+ epithelial cells. (E) Bar plot of up-regulated pathways in Scissor+ and Scissor- epithelial cells. (F) K-M survival curve for TCGA-COAD patients stratified by he enrichment scores of Scissor+ and Scissor- epithelial cells, calculated using Z-score based on up-regulated marker genes. (G) The distribution of Tumor and EMT signatures in TCGA-COAD patients across different Scissor groups. (H) The distribution of Immune signatures in TCGA-COAD patients across different Scissor groups. * p < 0.05, ⁎⁎ p < 0.01, ⁎⁎⁎ p < 0.001, ⁎⁎⁎⁎ p < 0.0001. Then, a differential expression analysis of these epithelial cells was conducted to uncover the difference between Scissor+ cells and Scissor- epithelial cells. 491 genes were upregulated in Scissor+ epithelial cells, while 170 genes were upregulated in Scissor- epithelial cells (Table S4.1). The upregulation of NME2, EEF1G, and IFITM1 in Scissor+ epithelial cells was particularly pronounced, in relavant to oxidative phosphorylation, focal adhesion and transcriptional misregulation in cancer, indicating Scissor+ epithelial cells play pivotal role in tumor progression and metastasis. In contrast, in Scissor- epithelial cells, the upregulation of IGF2, PPBP and RPS6 is notable, which are related to ribosomes and mitophagy, and their functions are more inclined towards normal physiological conditions (Fig. 4D-E). To explore the prognostic value of Scissor+ cells in CRC hepatic metastasis, single-sample gene set enrichment analysis (ssGSEA) was performed on patients from the TCGA-COAD cohort. ssGSEA scores were calculated for each patient based on the top 100 upregulated genes. The TCGA-COAD cohort was then stratified into Scissor+ and Scissor- groups. Notably, patients in the Scissor+ group exhibited significantly poorer prognosis than those in the Scissor- group (P = 0.048), suggesting that Scissor+ epithelial cells contribute to adverse clinical outcomes in CRC patients with hepatic metastasis (Fig. 4F). Furthermore, immune infiltration analysis was conducted to elucidate the distribution of EMT, tumor signatures and immune signatures in TCGA-COAD patients across Scissor+ and Scissor- groups. Scissor+ group exhibited significant activation of EMT pathways, specifically EMT1* and EMT3*, which indicating that the Scissor+ group possesses enhanced invasiveness and metastatic potential. In addition, Scissor+ group upregulate more proliferation-related pathways, including Cell_cycle⁎⁎⁎⁎ and DNA_replication⁎⁎, which is in accordance with the previous scRNA-seq outcomes. Moreover, Scissor+ group is more favorable for the generation of tumor stroma by upregulating certain pathways, such as Fibroblasts_MCPcounter* and Endothelial_cells_MCPcounter⁎⁎. Conversely, Scissor+ group is associated with a down-regulation of certain immune signatures, such as T_cell_inflamed⁎⁎⁎⁎, Cytotoxic_lymphocytes_MCPcounter⁎⁎⁎⁎, NK_cells_MCPcounter⁎⁎⁎⁎, Plasma_cells_xCell⁎⁎⁎⁎ and B_lineage_MCPcounter⁎, which suggests that Scissor+ group would exacerbate local TIME, leading to a reduction in immune cell capable of tumor elimination (Fig. 4G-H; Table S4.2). Collectively, Scissor+ epithelial cells possess enhanced capacities for proliferation, metastasis, and tumor stroma formation. Furthermore, these cells remodel the local TIME and promote the development of an immunosuppressive microenvironment. Development and validation of a HM-related prognostic signature Considering that Scissor+ epithelial cells were correlated with unfavorable clinical outcomes in CRC patients with HM, a total of 101 machine learning algorithms were utilized to identify Scissor+ epithelial cell highly expressed genes with superior prognostic predictive performance. The TCGA-COAD (n = 429) served as the training cohort, while the GEO (GSE38832: n = 120, GSE17536: n = 177) were employed as validation cohorts to develop a prognostic signature that was refined through cross-cohort validation. Subsequently, the prognostic predictive value of the aforementioned genes were assessed by 101 machine learning algorithms to calculated the c-index for the TCGA-COAD, GSE38832, and GSE17536 cohorts (Table S5.1). The algorithms were ranked by their average c-index from the validation cohort. The predictive accuracy of each signature was assessed by calculating the mean C-index across all cohorts (Fig. 5A). When selecting the final model, the one that achieves superior performance on the validation datasets is typically selected. The data indicated that the combination of "StepCox[both] + SuperPC" exhibited optimal prognostic discrimination across various validation datasets, achieving C-index values of 0.56 for TCGA-COAD, 0.67 for GSE38832, and 0.62 for GSE17536 (Fig. 5B). The minimal partial probability of deviation in the StepCox[both] regression occurred when the penalty parameter lambda was optimized to minimize mean deviance through tenfold cross-validation. Then, three model genes were identified and their respective absolute coefficient value (|Coef|), which were ZFP36L1 (1.499993), HSPB1 (1.16527) and HEBP2 (0.7670467) (Fig. 5C). Specifically, SRRS in all cohorts was calculated using the following formula: ZFP36L1exp × 1.499993 + HSPB1exp × 1.16527 - HEBP2exp × 0.7670467. Subsequently, each cohort was devided into the SRRS-High group and the SRRS-Low group based on the median of SRRS. Both in the training and validation cohorts, the prognosis of the SRRS-High group was unfavorable (Fig. 5D-F). Furthermore, a nomogram model was constructed for the three cohorts to predict 1-, 3-, and 5-year survival in CRC patients, highlighting the clinical significance of the model (Fig. 5G-I). Evaluation of the results revealed that the AUC values of the TCGA-COAD cohort all reached more than 0.60, and the AUC values of the validation cohort are also relatively stable. Meanwhile, in the meta-analysis of univariate Cox regression, both the random effects and the fixed effects models demonstrated statistical significance (p<0.001), indicating the exceptional predictive efficacy of our model (Fig. 5J). Moreover, a comprehensive dataset encompassing 16 published signatures was curated (Table S5.2) and a comparative evaluation of our signature against these signatures was conducted to evaluate the robustness of our model. The comparative analysis indicated that the GSE17536⁎⁎, GSE38832* and TCGA-COAD* cohorts exhibited a statistically significant advantage over alternative models (Fig. 5K). Furthermore, our model achieved a top-tier ranking among the top three positions based on AUC values derived from 16 published signatures (Fig. S5A), thereby validating its predictive accuracy and reliability in HM of CRC. Development and validation of a prognostic signature for CRC HM (A) 101 machine learning algorithms calculated the c-index for the TCGA-COAD, GSE38832 and GSE17536 cohorts. The algorithms were ranked by their average c-index from the validation cohort. (B) Bar plot displaying the C-index of the final signature (StepCox[both] + SuperPC) across different cohorts. (C) Bar plot of prognostic genes and their coefficient values in the final signature. (D-F) K-M analysis illustrating the relationship between SRRS from the final signature and clinical outcomes in TCGA-COAD, GSE38832 and GSE17536 cohorts. (G-I) The clinical nomogram presenting the 1-year, 3-year and 5-year ROC curves based on the final signature and clinical parameters in the training and validation cohorts. (J) Meta-analysis of univariate Cox regression in the final signature across different cohorts. (K) Heatmap displaying the HR of the final signature compared to 16 published models in different cohorts. * p ≤ 0.05, ⁎⁎ p ≤ 0.01. Expression and significance of SRRS-related prognostic signature genes in HM ST analysis of CRC tissues was performed to characterize the spatial patterns of three SRRS-associated prognostic signature genes (Fig. 6A). EPCAM served as a canonical epithelial marker for the identification of malignant epithelial regions. Notably, the expression of ZFP36L1, HSPB1, and HEBP2 was consistently upregulated in HM tissues relative to PT and NC tissues. Meanwhile, the results of immunohistochemistry (IHC) were in accordance with those of ST analysis. Significantly higher expression of these three model genes was observed in tumor samples compared with normal tissues (Fig. 6B). Furthermore, the expression levels of SRRS-related prognostic genes, Scissor+ epithelial genes and normal epithelial gene were analyzed by RT-qPCR in the normal colorectal cellline (NCM460), CRC PT cell lines (HT29 and HCT116) and metastatic CRC cell line (SW620) respectively (Fig. 7A-I), validating their identified metastasis-associated prognostic signature. Notably, HSPB1 was highly expressed exclusively in the SW620 compared to other cell lines (Fig. 7A). Meanwhile, the expression levels of ZFP36L1 and HEBP2 were also elevated in the SW620 and PT cell lines, which may be attributed to the intricate heterogeneity present among these cell lines (Fig. 7B-C). Furthermore, RACK1, EEF1G, SEM1, and NME2, which are highly expressed in Scissor+ epithelial cells, were significantly upregulated in SW620 cells. These findings further support the notion that Scissor+ epithelial cells possess potent metastatic-promoting capacity (Fig. 7D-G). In contrast, FG16 and FCGBP, which are highly expressed in normal epithelial cells, were found to be significantly downregulated in SW620 (Fig. 7H-I). Collectively, the RT-qPCR validation results corroborated the conclusions derived from bioinformatics analyses. ST and IHC revealed a significant increase in prognostic genes in HM (A) Spatial transcriptome analysis of SRRS-related prognostic signature gene expression in various tissues. (B) IHC results of SRRS-related prognostic signature genes were obtained from the HPA database. The data in B is shown as the mean ± SEM (n = 3). The P values (* p < 0.05, ⁎⁎ p < 0.01, ⁎⁎⁎ p < 0.001) were derived from a two-tailed unpaired Student's t-test. RT-qPCR validation and significance of SRRS in prediction prognosis in HM (A-I) RT-qPCR analysis evaluated the expression levels of SRRS-related prognostic genes, Scissor+ epithelial genes, and normal epithelial gene in the normal colorectal cell line (NCM460), CRC PT cell lines (HT29 and HCT116) and metastatic CRC cell line (SW620). Data are presented as mean ± SEM (n = 3 per group) with two-tailed unpaired Student's t-test p-values. * p < 0.05, ⁎⁎ p < 0.01, ⁎⁎⁎ p < 0.001. Malignant epithelial cells enhance CRC hepatic metastasis through regulating T/NK cells To further elucidate the functional role of Scissor+ epithelial cells in CRC HM relative to stage Ⅳ primary tumor samples (Fig. 1G), we investigated cellular crosstalk between Scissor+ epithelial cells and enriched immune cells within the HM microenvironment. Based on re-clustering results, T/NK cells were partitioned into seven unique clusters. T cells in the integrated dataset were annotated as CD3+ Naive (CD3D), CD4+ Naive (IL7R), CD3+ Treg (TIGIT), CD4+ Treg (CTLA4), CD8+ Tc (GZMK), CD8+ Tex (CD8A) and NK (NKG7) based on the identified marker genes (Fig. 8A-B). Meanwhile, enrichment patterns were evaluated utilizing the Ro/e analysis, which revealed a propensity for CD8+ Tex, CD4+ Tregs and CD4+ Naive in HM compared to PT (Fig. 8C-D; Table S8). To systematically characterize the functional specialization of distinct immune clusters, seven functional hallmarks were quantified across tumor-associated immune clusters. CD3+ and CD4+ Treg cells displayed unique clustering characteristics and presented the highest scores for both immune inhibition and exhaustion, which was consistent with their intrinsic immunosuppressive properties. Compared with CD8+ Tex, CD8+Tc and NK cells exhibited significantly higher cytotoxicity and inflammatory activity scores, indicating predominant antigen-specific immune responses rather than nonspecific inflammation-mediated tumor clearance (Fig. 8E; Fig. S8A; Table S8.1). Furthermore, IL2R scores in both CD3+ Tregs and CD4+ Tregs populations peaked, a finding consistent with the results presented in Fig. 8B Notably, pathways including NK cell-mediated cytotoxicity, Th17 cell differentiation, cell adhesion molecules and T cell receptor signaling pathway were enriched in CD8+ Tex. Conversely, antigen processing and presentation, along with Th1 and Th2 cell differentiation, were functionally suppressed in CD8+ Tc, accompanied by upregulated expression of relevant genes (Fig. 8F; Table S8.2). Survival analysis further revealed that tumors enriched with CD8+ Tex-high cells were associated with a significantly worse prognosis (Fig. 8G), suggesting that cytotoxicity-driven NK cell dysfunction promoted immune evasion and disease progression in CRC. Meanwhile, pathways including Th1/Th2 cell differentiation, NF-κB and JAK-STAT signaling, cell adhesion molecules and PD-L1/PD-1 checkpoint pathway were significantly enriched in CD4+ Tregs. Concurrently, CD3+ Tregs exhibited enrichment in oxidative phosphorylation, cytokine-cytokine receptor interaction and necroptosis (Fig. 8H). These patterns suggested that CD4+ Tregs could help maintain the HM-associated TME and facilitate the conversion of CD8+ Tc into regulatory CD8+ Tex. Furthermore, a high abundance of CD4+ Tregs was also associated with poor prognosis (Fig. 8I). Together, these findings underscored the immunosuppressive features of CD8+ Tex cells in the context of CD4+ Treg involvement, implicating a mechanism by which anti-tumor immunity was suppressed through upregulation of inhibitory genes and alterations in metabolic pathways. "CellChat" analysis was subsequently conducted to map the intercellular crosstalk between Scissor+ epithelial cells and individual T/NK clusters within TME (Fig. S8B). Specifically, Scissor+ epithelial cells were observed to interact with CD8+ Tex cells via the HLA signaling pathway, while the CCL pathway appeared to mediate broader communication between Scissor+ epithelial cells and multiple T/NK clusters. Intriguingly, within the MIF signaling network, Scissor+ epithelial cells consistently functioned as the dominant signal "Sender", whereas CD4+ Naive, CD4+ Treg and CD8+ Tex cells primarily served as signal "Receiver" and "Mediator" (Fig. S8C). Notably, the MIF - (CD74 + CXCR4) axes was universally involved in interactions between Scissor+ epithelial cells and all T/NK clusters (Fig. S8B). Then, spatial co-expression patterns of MIF and (CD74 + CXCR4) were investigated in CRC HM ST data to further confirm the potential for crosstalk (Fig. 8J). Furthermore, Scissor+ epithelial cells exhibited a prominent role as a regulatory "Influencer" within the MHC-I signaling network (Fig. S8D), demonstrating extensive communication with CD8+ Tex cells across "Sender", "Receiver" and "Mediator" roles. Collectively, dynamic and bidirectional crosstalk between Scissor+ epithelial cells and CD4+ Treg/CD8+ Tex cells critically shapes the immunosuppressive tumor immune microenvironment in CRC HM. This refined cell-cell interaction landscape provides mechanistic insights into the immunosuppressive features of the CRC microenvironment and identifies promising therapeutic targets for future immunotherapeutic development. Malignant epithelial cells enhance CRC hepatic metastasis through regulating T/NK cells (A) UMAP plot depicting the T/NK cells landscape. (B) Bubble Chart showing marker gene expression in each T/NK cluster. (C) Heatmap showing tissue preferences of each T/NK cluster across groups revealed by Ro/e analysis. (D) Bar chart displaying the proportion of each T/NK cluster across different groups. (E) Violin plot displaying the "Co_stimulatory", "Inhibitory", "Exhausted" and "Cytotoxicity" scores for each T/NK cluster, calculated by "AddModuleScore". (F) Bar plot showing the up-regulated pathways in CD8+Tex and CD8+Tc. (G) K-M survival curves for TCGA-COAD patients stratified by CD8+Tex-related enrichment scores, calculated by Z-score from up-regulated marker genes. (H) Bar plot showing the up-regulated pathways in CD3+Treg and CD4+Treg. (I) K-M survival curves for TCGA-COAD patients stratified by CD4+Treg-related enrichment scores, calculated by Z-score from up-regulated marker genes. (J) ST analysis revealed the spatial co-localization of MIF-(CD74 + CXCR4) axis in CRC HM. Malignant epithelial cells promote CRC hepatic metastasis through regulating myeloid cells Considering the crucial function of Scissor+ epithelial cells in driving CRC HM, additional profiling of myeloid cell heterogeneity was performed (Fig. 9A; Table S9.1). Myeloid clusters in the integrated datasets were classified into Mono/Macro (CD163), MAST (TPSAB1), Neutrophil (S100A8) and DC (CD86) clusters based on representative marker genes (Fig. 9B). Ro/e analysis further confirmed that myeloid cells in HM tissues were dominated by Mono/Macro when compared with those in PT (Fig. 9C). Tissue preference patterns of each myeloid cluster across different sample groups were further illustrated in Fig. 9D Following re-clustering of Mono/Macro, four distinct clusters were identified (Fig. 9E). Specifically, CCL3L3+ Mono/Macro (M1) and CCL5+ Mono/Macro (M4) showed substantial genealogical overlap and close relationship. Notably, SELENOP+ Mono/Macro (M2) and FCN1+ Mono/Macro (M3) constituted the major proportion of Mono/Macro cells and were preferentially enriched in HM (Fig. 9F-G). Subsequently, "AddModuleScore" was applied to systematically assess functional specialization across the four Mono/Macro clusters in three biological processes: TAM activity, inflammatory response, antigen presentation and chemotactic cytokine signaling (Fig. 9H; Fig. S9B). M2 exhibited high "TAM" scores but low "Inflammatory" scores, suggesting that TAMs are enriched in HM without inducing robust local inflammatory response. Interestingly, an elevated "Antigen_Presentation" score in M2 coincided with putative anti-tumor effects (Fig. 9H). This may imply that antigen presentation within this contest is skewed toward CD4+ Tregs, which are abundant in HM, thereby potentially facilitating immune evasion. Thus, M2 was defined as HM-associated TAMs, namely SELENOP+ TAMs. Malignant epithelial cells promote CRC hepatic metastasis through regulating myeloid cells (A) UMAP plot depicting the myeloid cells landscape. (B) SCP heatmap showing marker gene expression in each myeloid cluster. (C) Heatmap displaying tissue preferences of each myeloid cluster across different groups revealed by Ro/e analysis. (D) Bar chart displaying the proportion of each myeloid cluster across different groups. (E) UMAP plot depicting the mononuclear macrophage landscape. (F) SCP heatmap showing marker gene expression in each mononuclear macrophage cluster. (G) Heatmap displaying tissue preferences of each mononuclear macrophage cluster across different groups revealed by Ro/e analysis. (H) Violin plot displaying the "TAM", "Inflammatory"and "Antigen_Presentation" scores for each mononuclear macrophage cluster, calculated by "AddModuleScore". (I-K) Trajectory analysis of each mononuclear macrophage cluster by "Monocle 2″. (L) Trajectory analysis of each mononuclear macrophage cluster by "Cytotrace". (M) Dot plot illustrating Scissor+ epithelial cell crosstalk with each mononuclear macrophage cluster through "CellChat". (N-O)Circle plot showing the crosstalk strength of "CCL15 - CCR1″ and "LGALS9 - HAVCR2" axes. Furthermore, the differentiation dynamics of macrophages during CRC HM progression were delineated using both "Monocle 2″ and "CytoTRACE" (Fig. 9I-L). These results revealed a pronounced tendency for M3 to transition into SELENOP+ TAMs (Fig. 9J-K; Fig. S9C; Fig. S9E). Consistent with prior observations, M3 was characterized by a low "TAM" score and a high "Inflammatory" score, and expressed VCAM, a marker gene known to promote cell proliferation and angiogenesis (Fig. 9F). Accompanying the transition from M3 to SELENOP+ TAMs, a significant enhancement in cellular stemness was observed, indicating its potential role in shaping the trajectory of CRC HM. Complementary analyses visualized the differentiation scores and spatial distribution patterns across distinct macrophage clusters. Pseudotime trajectory and genomic enrichment analyses further validated the upregulation of mononuclear macrophage signature genes, including SELENOP and APOC1, throughout this dynamic cellular transition process (Fig. S9D-E). Then, "CellChat" was performed to investigate intercellular communication within the TME, revealing intricate crosstalk between Scissor+ epithelial cells and all mononuclear macrophage clusters mediated by the CCL15-CCR1 axis (Fig. 9M-N). Moreover, the LGALS9-HAVCR2 axis was identified as specific signaling pathways mediating crosstalk between Scissor+ epithelial cells and SELENOP+ TAMs (Fig. 9O). The communication probability and interaction strengths of Scissor+ epithelial cells with each mononuclear macrophage clusters, as "Sender", "Receiver", "Mediator" and "Influencer", within the MIF and GALECTIN signaling networks were also quantified (Fig. S9F-G). Scissor+ epithelial cells emerged as the dominant participants in these signaling pathways and exerted a pivotal function in mediating intercellular crosstalk with mononuclear macrophages. These findings indicate that Scissor+ epithelial cells interact with TAMs through specific ligand-receptor pairs. TAMs, in turn, propagate Scissor+-derived oncogenic signals and contribute substantially to the reprogramming of the tumor immune microenvironment. Furthermore, "NicheNet" analysis delineated the target gene landscape of Scissor+ epithelial cells and prioritized key ligands associated with macrophage clusters (Fig. S9H; Table S9.2). The ligand-receptor network identified TNF and its molecular regulator ADAM17 as top-ranked ligands based on their strong expression profiles, together implicated in triggering inflammatory responses. Collectively, these results reveal an intricate crosstalk network between Scissor+ epithelial cells and SELENOP+ TAMs that collaboratively reshapes the TIME. Malignant epithelial cells recruit and educate macrophages into SELENOP+ TAMs via the LGALS9-HAVCR2 axis Three SRRS-related genes HSPB1, HEBP2 and ZFP36L1 had maximal expression in Scissor+ epithelial cells and SW620 (Fig. 10A). Considering the high abundance of Scissor+ epithelial cells in HM, we adopted SW620 to simulate Scissor+ epithelial cells under in vitro conditions (Fig. 7A-C). PMA stimulation was applied to induce the differentiation of THP-1 cells into macrophages. Subsequently, differentiated THP-1-derived macrophages were cultured with conditioned media collected from normal intestinal epithelial NCM460 cells and three CRC cell lines (HT29, HCT116, and SW620), including NCM460_CM, HT29_CM, HCT116_CM, and SW620_CM, to simulate the in vitro co-culture system of epithelial and stromal cells. Transwell migration assays revealed that differentiated THP-1 cells exhibited stronger migratory recruitment toward SW620_CM relative to HT29_CM and HCT116_CM (Fig. 10C). Furthermore, stimulation of THP-1 cells with SW620_CM upregulated SELENOP expression in THP-1 cells (Fig. 10B). These findings indicate that malignant epithelial cells can recruit macrophages and promote their polarization into SELENOP+ TAMs. SRRS-associated genes (HSPB1, HEBP2, ZFP36L1) were overexpressed in HT29 and HCT116 cells to acquire metastatic phenotypes, and the corresponding overexpression efficiencies are shown in Fig. S10D-F. Transwell migration assays further demonstrated that OE_HSPB1HT29_CM, OE_HEBP2HT29_CM and OE_ZFP36L1HT29_CM possessed enhanced capacity to recruit THP-1 cells, a phenotype recapitulated in HCT116 (Fig. 10D-F, Fig. S10A-C). Collectively, these findings verify that SRRS-related genes confer potent pro-metastatic activities. Malignant epithelial cells recruit and educate macrophages into SELENOP+ TAMs via the LGALS9-HAVCR2 axis (A) Western blotting assay showed that the expression of HEBP2, HSPB1 and ZFP36L1 was highly expressed in SW620. (B) Western blotting revealed that the CM from SW620 increased the SELENOP protein level in THP-1 cells. (C) Transwell assay indicated that the CM of SW620 recruited more THP-1 cells. (D-F) Transwell assay indicated that OE_HEBP2, OE_HSPB1 and OE_ZFP36L1 enabled HT29_CM to recruit more THP-1 cells. (G) Western blotting revealed that siLGALS9#3SW620_CM decreased the SELENOP protein level in THP-1 cells. (H) Transwell assay revealed that siLGALS9#3 diminished the capacity of SW620_CM to recruit THP-1 cells. (I) Western blotting showed that rhLGALS9 increased the SELENOP protein level in THP-1 cells, which was reversed by siHAVCR2#2. (J) Transwell assay revealed that rhLGALS9 significantly promoted the recruitment of THP-1 cells, while reversed by siHAVCR2#2. To further validate the role of the LGALS9-HAVCR2 axis in mediating crosstalk between malignant epithelial cells and SELENOP+ TAMs, LGALS9 was knocked down in SW620 cells via siRNA, and the knockdown efficiency is presented in Fig. S10G. Then, THP-1 cells were stimulated with siLGALS9#3SW620_CM. Notably, relative to NC, siLGALS9#3SW620_CM not only lost the ability to recruit THP-1 cells but also reduced the capacity of SW620_CM to induce macrophage polarization into SELENOP+ TAMs (Fig. 10G, H). Furthermore, recombinant human LGALS9 (rhLGALS9) robustly recruited THP-1 cells and promoted their polarization into SELENOP+ TAMs, while siHAVCR2 efficiently reversed the promoting effects induced by rhLGALS9 (Fig. 10I, J), with its knockdown efficiency demonstrated in Fig. S10H. Furthermore, spatial co-expression patterns of LGALS9 and HAVCR2 were investigated in CRC HM ST data to further confirm the potential crosstalk (Fig. S10I). Overall, the LGALS9-HAVCR2 axis plays a pivotal role in macrophage recruitment by malignant epithelial cells and promoting macrophage polarization into SELENOP+ TAMs. Hepatic microenvironment differences between PM and HM To characterize temporal shifts of the hepatic immune microenvironment during CRC metastatic progression, scRNA-seq profiling was performed on mouse specimens of HM, PM, and NL (Fig. S11A), obtained from the aforementioned datasets (GSE132465 n = 6, GSE284449 n = 6, GSE245552 n = 4, GSE158328 n = 2). Re-clustering of the integrated scRNA-seq data revealed eight distinct cell clusters (Fig. 11A). Marker gene expression profiles and cellular composition across samples were summarized in Fig. 11B For example, Cd68, C1qa and C1qb were highly expressed in mononuclear macrophages. Notably, Pecam1 showed elevated expression in mouse endothelial cells, a pattern distinct from that observed in humans. The distribution of cellular proportions across samples, especially in HM and PM, is further detailed supplementally (Fig. S11B-C; Table S11). Compared to PM, HM samples exhibited as pronounced increase in the proportions of Mono/Macro (41.87% vs 33.43%) and Neutrophil (40.26% vs 17.30%) clusters, accompanied by a marked reduction in T/NK populations (10.43% vs 33.35%) in HM. Hepatic microenvironment differences between PM and HM (A) UMAP plot depicting the hepatic microenvironment landscape. (B) Bubble Chart showing marker gene expression in each cell type. (C) Heatmap illustrating tissue preferences of various cell types across groups identified by Ro/e analysis. (D) UMAP plot showing the landscape of T/NK cells in PM and HM. (E) Heatmap illustrating tissue preferences of each T/NK cell cluster across groups identified by Ro/e analysis. (F) UMAP plot showing the landscape of mononuclear macrophages in PM and HM. (G) Heatmap illustrating tissue preferences of each mononuclear macrophage cluster across groups identified by Ro/e analysis. We next analyzed tissue-specific distribution profiles of distinct immune cell subsets. Ro/e enrichment analysis revealed that endothelial cells, T/NK cells, B cells, and epithelial cells were predominantly enriched in NL tissues, whereas Mono/Macro clusters accumulated at significantly higher proportions in PM and HM tissues. Subsequent comparative re-clustering of immune cells, with a focus on T/NK cells and mononuclear macrophages, revealed that HM tissues contained significantly fewer CD8+ Tc and NK cells, but were markedly enriched in CD4+ Tregs relative to PM samples. This shift in immune composition may underlie the loss of T/NK cell tissue tropism, particularly within the HM microenvironment (Fig. 11C-E). Among Mono/Macro clusters, M2 displayed strong tissue-specific enrichment in HM, whereas M0 and M1 were substantially reduced compared to PM (Fig. 11F-G). These findings further support the notion that M2 represent HM-associated TAMs, and that their expansion is likely driven by active recruitment rather than local proliferation. Discussion Dynamic remodeling of the TME provides critical clues for identifying actionable therapeutic vulnerabilities in CRC. However, while prior studies have mapped the cellular composition of primary tumors, systematic analyses of HM-associated TME remain largely unexplored. Interestingly, the mechanisms by which Scissor+ epithelial cells dynamically drive HM progression through dual biological and immunological reprogramming have been revealed for the first time. Scissor+ epithelial cells were mainly constituted by clusters C3 and C4. Relative to PT tissues, Scissor+ epithelial cells were markedly enriched in HM lesions and displayed elevated CNV scores, a hallmark of genomic instability. NME2, EEF1G and IFITM1 exhibited higher expression levels in Scissor+ epithelial cells, contributing to the promotion of CRC HM and resulting in unfavorable clinical outcomes. Prior investigations have linked NME2 to enhanced CRC invasion and metastasis, serving as an indicator of poor patient outcomes [[52], [53], [54]]. Furthermore, EEF1G was verified for the first time to be specifically related to the epithelial cells associated with CRC HM, while previous studies merely reported that EEF1G+ macrophages facilitated CRC HM [55]. Hence, the role of EEF1G in CRC HM has been further extended. Moreover, IFITM1, a member of the interferon-induced transmembrane protein family, has been demonstrated in prior studies to facilitate CRC HM, which aligns with our findings [56]. Pseudotime trajectory analysis further validated a unique differentiation cascade driving colorectal cancer hepatic metastasis. Along this developmental route, Scissor+ epithelial cells exhibited progressive enhancement of stemness features and aggravated malignant phenotypes. Simultaneously, the results of immune infiltration analysis further demonstrated that Scissor+ epithelial cells have extremely strong capacity of EMT, proliferation and promotion of tumor stroma generation, which lays a solid foundation for facilitating CRC HM. Conversely, Scissor+ epithelial cells act as key regulators of TME remodeling. These cells foster a permissive niche for hepatic metastatic outgrowth by suppressing antitumor effector populations, such as cytotoxic T lymphocytes, NK cells, B cells, and plasma cells. This phenomenon aligns with the previously reported "Immune Cold" characteristic observed in CRC [57]. Consequently, Scissor+ epithelial cells are defined as the core pivot of CRC HM. CRC patients with HM generally suffer from poor prognosis, mainly owing to insufficient treatment regimens and TIME-mediated immune refractoriness [7]. By characterizing HM-resident epithelial clusters correlated with unfavorable survival, our work bridges this unmet clinical need and yields translational implications for diagnostic stratification and targeted therapy. Then, highly expressed genes in Scissor+ epithelial cells were employed to develop a three-gene prognostic signature (HEBP2, HSPB1, ZFP36L1) utilizing rigorous machine learning algorithms. Our prognostic signature has successfully correlated the metastasis genes with the prognostic genes and computed that SRRS is capable of predicting the risk of CRC HM more accurately. Notably, the SRRS demonstrated superior predictive accuracy compared to 16 existing signatures, highlighting its potential for precise patient stratification. HEBP2 has been previously reported to be associated with the remodeling of the TME. Our study is the first to demonstrate that HEBP2 can facilitate HM in CRC patients and contribute to unfavorable clinical outcomes [58]. HSPB1, as a negative regulator of ferroptosis, has been reported by previous researchers to be associated with remodeling the TIME and leading to poor prognosis in patients, which is consistent with our research findings [57,59]. The role of HSPB1 in CRC HM was further emphasized by our research. ZFP36L1 has previously been reported to be capable of predicting the prognosis of patients in ovarian cancer [60]. Our research has broadened its role in CRC, serving as a marker of HM. Furthermore, the outcomes of ST, immunohistochemistry and RT-qPCR further substantiated the functions of the aforementioned three genes in CRC HM, which were in accordance with the results of scRNA-seq. Our analyses revealed that the SRRS signature exhibits robust and generalizable prognostic predictive performance across CRC cohorts. While SRRS establishes a reliable linkage between metastatic traits and clinical outcomes and deepens mechanistic understanding of CRC HM, additional proteomic validation experiments are warranted to confirm the robustness and reproducibility of its predictive capacity. Consistent with previous studies, tumor cells recruit immunosuppressive cells and reshape the tumor immune microenvironment to support localized outgrowth of HM lesions [5]. Our research findings align with this phenomenon, revealing a higher enrichment of T/NK cells and myeloid cells within the HM compared to their paired PT lesions. CD8+ Tex and CD4+ Tregs are critical drivers reshaping the TIME within CRC HM lesions. These two T clusters exhibit spatial co-localization and exert bidirectional regulatory and activating effects on one another. Intriguingly, CD8+ Tex does not possess the traditional anti-tumor capacity. In contrast, it exhibits stronger immunosuppressive and immune exhaustion function. Although the immunosuppressive function of CD4+ Treg has been widely recognized in CRC, the identification of CD8+ Tex has enabled a more comprehensive understanding of the role of CD8 in HM [61,62]. Based on this rational inference, CD4+ Treg in HM may induce the transformation of CD8+ Tc to CD8+ Tex, causing them to lose their anti-tumor capacity. Simultaneously, Scissor+ epithelial cells actively crosstalk with CD8+ Tex and CD4+ Treg via MIF signaling pathways, enhancing their immunosuppressive functions to facilitate CRC HM. Furthermore, this finding extends the role of MIF in the TME of CRC, which aligns with previous studies reporting that MIF promotes CD8+ T cell exhaustion [[63], [64], [65]]. Overall, our finding not only highlights the crucial role of Scissor+ epithelial cells in reshaping the TME but also enhances current understanding of the immune regulatory functions of Scissor+ epithelial cells in CRC HM. Myeloid cells also exert a crucial role in the remodelling of the TIME. SELENOP+ TAMs were defined HM-associated TAMs in this study due to its high expression of "TAM" and "Antigen_Presentation" score. Notably, SELENOP+ TAMs not only perform immunosuppressive functions but also exhibit antigen-presenting capabilities. It is reasonable to infer that SELENOP+ TAMs may present antigens to CD4+ Treg and CD8+ Tex, rather than to other anti-tumor effector cells [66]. Meanwhile, the transwell assay and western blotting demonstrated that Scissor+ epithelial cells engage in active crosstalk with SELENOP+ TAMs through the LGALS9-HAVCR2 axis, which promotes macrophage polarization toward the SELENOP+ TAMs and further deteriorates the TIME. Interestingly, in contrast to PM, the enrichment level of SELENOP+ TAMs in HM increased significantly, which further substantiated that SELENOP+ TAMs possesses HM specificity. Hence, the existence of HM-associated mononuclear macrophages has been confirmed by our research for the first time. Nevertheless, the current work only characterizes how Scissor+ epithelial cells reprogram macrophages into SELENOP+ TAMs, while the signaling cascades whereby SELENOP+ TAMs reciprocally modulate gene expression in Scissor+ epithelial cells to drive colorectal cancer hepatic metastasis remain uninvestigated, demanding further exploration to close this regulatory circuit. Additionally, although siRNA-mediated knockdown of HAVCR2 was performed to validate the LGALS9-HAVCR2 axis, neutralizing antibodies targeting LGALS9 and HAVCR2 should be utilized in subsequent assays to supplement our experimental data and strengthen the integrity of our functional verification. However, limitations remain: (1) The biological significance of Scissor+ cells remains elusive. In the present study, the Scissor algorithm was exclusively employed for the identification of malignant epithelial cells. However, Scissor+ epithelial cells have not been biologically characterized, and their biological properties remain undefined. Subsequent investigations are warranted to validate the core genes and key pro-metastatic pathways of Scissor+ epithelial cells, as well as to detect these cells in HM tissue sections via immunohistochemical staining; (2) Although spatial co-expression of ligand-receptor pairs mediating interactions between Scissor+ epithelial cells and immunosuppressive cells has been confirmed in spatial transcriptomics data, spatial co-expression among all three cell types remains unverified. Multicolor immunohistochemistry (mIHC) is still required to further validate the spatial co-expression of these three cell types; (3) Although our study used a mouse single-cell model to investigate changes in the hepatic immune microenvironment before and after metastasis, further demonstrating that SELENOP+ TAMs and suppressive T cells accumulate in the liver only after CRC metastasis, this model cannot prove the mechanism by which Scissor+ epithelial cells drive hepatic metastasis. Future studies will need to establish relevant organoid models to address this gap; (4) Although an overexpression model of SRRS-related genes was constructed in this study, a knockdown model was not established. Meanwhile, further experiments are still needed to validate the mechanism by which SRRS-related genes promote the secretion of MIF and LGALS9, thereby facilitating CRC HM. Future research that integrates multi-omics technologies with clinical trials is crucial for disrupting the immunosuppressive loop based on Scissor+ epithelial cells, which will advance immunotherapy strategies for CRC HM. Conclusions Overall, our research identified Scissor+ epithelial cells that can promote CRC HM, ultimately leading to unfavorable clinical outcomes in CRC patients. Based on these findings, Scissor-related prognostic signature was developed and SRRS was calculated to predict CRC patients with HM. Subsequently, the immune landscape of CRC HM has been mapped to identify immunosuppressive cells, including CD8+ Tex, CD4+ Treg and SELENOP+ TAMs. Scissor+ epithelial cells served as the primary signal senders in remodeling the TIME in CRC HM. For instance, the immunosuppressive function of CD4+ Treg and CD8+ Tex were enhanced through the MIF-(CD74 + CXCR4) axis. Simultaneously, SELENOP+ TAMs was recruited via the LGALS9-HAVCR2 axis. This complex regulatory loop not only enables immune evasion of Scissor+ epithelial cells, but also promotes the formation and deterioration of the TIME, ultimately leading to unfavorable clinical outcomes in CRC patients with HM (Fig. 12). Mechanism by which Scissor+ malignant epithelial cells enhance HM of CRC Scissor+ malignant epithelial cells potentiate the immunosuppressive activity of CD4+ Treg and CD8+ Tex via MIF and MK signaling pathways while recruiting M2 TAMs through MIF and GALECTIN signaling pathways. M2 TAMs, in turn, amplify CD4+ Treg and CD8+ Tex mediated immunosuppression via CCL and MIF signaling pathways, forming a self-reinforcing circuit. This intricate regulatory loop aggravates the TIME in CRC and enhances HM, resulting in unfavorable clinical outcomes. Abbreviations ACAT2 Acetyl-CoA Acetyltransferase 2. ANOVA Analysis of Variance. ATCC American Type Culture Collection. AUC Area Under the Curve. C1QC Complement C1qC Chain. CAF Cancer associated fibroblast. CD3D Cluster of Differentiation 3D. CD68 Cluster of Differentiation 68. CD79A CD79a Molecule. cDNA Complementary DNA. CellChat An algorithm for inferring cell-cell communication networks. CM Conditioned medium. CNV Chromosomal Copy Number Variation. Coef Coefficient value. CRC Colorectal cancer. CTLA4 Cytotoxic T-Lymphocyte Associated Protein 4. DEG Differentially Expressed Gene. DMEM Dulbecco’s Modified Eagle Medium. EMT Epithelial mesenchymal transition. EEF1G Eukaryotic Translation Elongation Factor 1 Gamma. EPCAM Epithelial Cell Adhesion Molecule. FCGBP Fc Gamma Binding Protein. FBS fetal bovine serum. FOXP3 Forkhead Box P3. GEO Gene Expression Omnibus. GSVA Gene Set Variation Analysis. HEBP2 Heme Binding Protein 2. HM Hepatic metastasis. HSPB1 Heat Shock Protein Family B Member 1. IARC International Agency for Research on Cancer. IFITM1 Interferon Induced Transmembrane Protein 1. IL7R Interleukin 7 Receptor. Infercnv An algorithm for copy number variation assessment. IOBR Immuno-Oncology Biomarker Research package. K-M Kaplan-Meier. LAG3 Lymphocyte Activating 3. LYZ Lysozyme. mCRC Metastatic colorectal cancer. mIHC Multiplex immunohistochemistry. MIF Macrophage Inflammatory Factor. MK Midkine. Monocle 2 A pseudo-time trajectory analysis algorithm. NC Normal colorectum. NicheNet An algorithm for predicting cell-cell interactions. NKG7 Natural Killer Cell Granule Protein 7. NME2 Non-Metastatic Cells 2. NME2 Nucleoside Diphosphate Kinase 2. NL Normal liver. PM Pre-hepatic metastasis. PT Primary tumor. RACK1 Receptor For Activated C Kinase 1. Ro/e Reciprocal overlap/enrichment analysis. rhLGALS9 human recombinant LGALS9. RT-qPCR Reverse Transcription Quantitative Polymerase Chain Reaction. scRNA-seq Single-cell RNA sequencing. Seurat An R package for single-cell data analysis. SEM Standard error of the mean. SEM1 SEM1 26S Proteasome Subunit. SingleR An R package for cell type annotation. Slingshot A pseudo-time algorithm. SRRS Scissor-related risk score. ssGSEA Single-sample gene set enrichment analysis. TAM Tumor associated macrophage. TAN Tumor associated neutrophil. TCGA The Cancer Genome Atlas. TEX Exhausted T cell. TIGIT T Cell Immunoreceptor With Ig and ITIM Domains. TME Tumor microenvironment. TIME Tumor immunosuppressive microenvironment. TNM Tumor Node Metastasis. TPSAB1 Tryptase Alpha/Beta 1. Treg Regulatory T cell. UMAP Uniform Manifold Approximation and Projection. VWF Von Willebrand Factor. ZFP36L1 ZFP36 Ring Finger Protein Like 1. ZG16 Zymogen Granule Protein 16. Funding statement This work was sponsored by National Natural Science Foundation of China: 82302958, 82473431, Innovation Program of Shanghai Municipal Education Commission (2023ZKZD44). CRediT authorship contribution statement Wenlong Zhu: Writing – original draft, Visualization, Methodology, Investigation, Data curation, Conceptualization. Ao Dong: Writing – original draft, Investigation, Data curation. Rongbing Hu: Visualization, Investigation. Silei Zhou: Writing – original draft, Investigation. Jiali Yao: Writing – original draft, Visualization. Binghua Ma: Writing – review & editing, Validation. Jin Ding: Writing – review & editing, Supervision. Yanjie Zhang: Writing – review & editing, Writing – original draft, Validation, Supervision, Project administration. Declaration of competing interest The authors declare that they have no competing interests. Acknowledgments The figures presented herein were created using BioRender (https://www.biorender.com/). Footnotes Supplementary material associated with this article can be found, in the online version, at doi:10.1016/j.tranon.2026.103035. Contributor Information Binghua Ma, Email: binghuama28@163.com. Jin Ding, Email: dingjin@smmu.edu.cn. Yanjie Zhang, Email: zyanjie0530@163.com. Supplementary materials Availability of data and materials All data utilized in this study were sourced from publicly accessible databases. scRNA-seq can be retrieved from GSE132465, GSE284449, GSE245552 and GSE158328, while ST data is available for download from GSE206552 and GSE280313. Additionally, bulk cohorts can be accessed through the TCGA and GEO databases. No datasets were generated during the current study. References Associated Data Supplementary Materials Data Availability Statement All data utilized in this study were sourced from publicly accessible databases. scRNA-seq can be retrieved from GSE132465, GSE284449, GSE245552 and GSE158328, while ST data is available for download from GSE206552 and GSE280313. Additionally, bulk cohorts can be accessed through the TCGA and GEO databases. No datasets were generated during the current study.