An epithelial cell fate-driven predictive model for liver metastasis risk in primary colorectal cancer through single-cell and multi-omics integration Authors: Min Yin, Xiaochen Bo, Yaoyao Li, Simin Yu, Zhijie Lin, Mei Wang, Jian Wu, Ming Zhou, Lingmin Kong, Yefei Zhu, Weiming Xiao, Yanbing Ding Published: 2026 ABSTRACT BackgroundColorectal cancer (CRC) is the third most prevalent form of cancer worldwide, with colorectal cancer liver metastases (CRLM) representing a principal cause of CRC-related mortality. However, a lack of molecular subgroups based on the differentiation states of diverse cell types in CRLM poses a significant barrier to progress in precision therapy.MethodsWe integrated single-cell RNA sequencing (scRNA-seq, GSE178318) of paired CRLM tissues to define an epithelial cell fate gene signature. To evaluate the intrinsic “malignant seed” potential, this signature was projected onto a large-scale primary reference cohort (TCGA-COAD/READ, N = 433) using consensus clustering to establish the Malignant Development Signature of CRLM (MDSCRLM). Comprehensive multi-omics analyses—encompassing somatic mutations, copy number variations, tumor mutational burden (TMB), and ATAC-seq chromatin accessibility (N = 81)—alongside immune microenvironment deconvolutions and proteomic data integration were performed. The clinical relevance of the MDSCRLM model was externally validated using immunohistochemistry in an independent clinical cohort of 45 patients. Furthermore, the biological function of HSPA1A, a core marker, was experimentally verified through in vitro assays and in vivo metastasis models.ResultsThe MDSCRLM system discriminates three distinctmetastatic risk trajectories: Cluster 1—Aggressive-Metastasis-Enhanced CRLM (AMECRLM), Cluster 2—Cell Cycle-Active CRLM (CCACRLM), and Cluster 3—Growth-Inhibited CRLM (GICRLM). Survival analyses in the primary reference cohort validated the biological plausibility and prognostic impact of this model, revealing that GICRLM is associated with the most favorable prognosis while AMECRLM confers the poorest outcome. Our findings delineate distinct patterns of cellular heterogeneity, genomic instability (TP53 mutations), and gene expression across CRLM developmental stages. Prognostic validation in the clinical cohort and functional validation demonstrating that HSPA1A knockout attenuates tumor invasion and metastasis robustly supported the robustness of the MDSCRLM risk stratification.ConclusionsWe established a novel, clinically relevant CRLM risk stratification model with strong diagnostic and prognostic potential. By identifying pre-existing malignant features and stage-specific therapeutic vulnerabilities within the primary tumor, this model bridges the gap between biological discovery and precision medicine in advanced colorectal cancer.Supplementary InformationThe online version contains supplementary material available at 10.1186/s12967-026-08375-9. Introduction Colorectal cancer (CRC) is the third most commonly diagnosed malignancy and a leading cause of cancer-related mortality worldwide, accounting for approximately 10% of all cancer cases and deaths [1, 2]. Despite advances in surgical techniques, chemotherapy, targeted therapy, and immunotherapy, the prognosis for patients with metastatic disease remains poor, largely due to the high incidence of colorectal cancer liver metastases (CRLM). Liver metastasis is the predominant cause of death in CRC patients, with only 10–20% of CRLM patients being suitable candidates for curative surgical resection. Even among these, the 5-year survival rate remains below 30% [3, 4]. These sobering statistics highlight the urgent need for improved strategies for diagnosis, risk stratification, and targeted therapy in CRLM. Molecular classification has emerged as a cornerstone of precision oncology in CRC, enabling the identification of biologically and clinically relevant subgroups that inform prognosis and therapeutic response. Early genome-wide studies distinguished CRC into hypermutated and non-hypermutated subtypes [5], and subsequent transcriptomic analyses led to the development of the Consensus Molecular Subtypes (CMS), which stratify CRC into four distinct groups based on gene expression profiles [6]. However, both inter- and intra-tumoral heterogeneity in CRC, especially at the metastatic stage, remain significant challenges. Recent multiregional and longitudinal sequencing studies have revealed dynamic changes in molecular subtypes during tumor progression and therapy, underscoring the limitations of static, bulk-tissue classifications [7, 8]. Consequently, integrating single-cell sequencing with computational deconvolution is imperative to bridge the gap between inter-tumor and intra-tumor heterogeneity, particularly for the accurate molecular subtyping of CRLM. Among the diverse cellular components of the tumor microenvironment, epithelial cells play a central role in CRC progression, invasion, and metastasis. The plasticity of epithelial cells, including their ability to undergo epithelial–mesenchymal transition (EMT) and to adopt stem-cell-like properties, is a key driver of metastatic dissemination and therapy resistance [9–11]. Moreover, the molecular and functional diversity of epithelial cell populations within CRLM has not been fully elucidated, particularly at single-cell resolution. Advances in single-cell RNA sequencing (scRNA-seq) have enabled unprecedented dissection of tumor heterogeneity, revealing distinct epithelial subpopulations and developmental trajectories [12, 13]. For example, active migration is characterized by actin-rich basal protrusions oriented in the direction of migration, which provide mechanistic support for Epithelial–Mesenchymal Transition (EMT) [14], while in intestinal cancer liver metastasis, intestinal epithelial cells acquire an L1CAM-high phenotype after loss of epithelial integrity, and L1CAM is necessary for in situ CRC development, liver metastasis, and chemotherapy resistance [13, 14]. These observations underscore that epithelial differentiation plasticity is a fundamental driver of tumor heterogeneity and progression. As the primary malignant compartment, epithelial cells act as the central hub of oncogenic signaling, orchestrating interactions with the surrounding stromal and immune microenvironment. Their dynamic transition between distinct differentiation states—from stem-like to terminally differentiated phenotypes—directly governs the tumor’s invasive capacity and therapeutic sensitivity. Furthermore, the adaptation of metastatic seeds to the liver niche relies heavily on this intrinsic plasticity to overcome colonization barriers. Consequently, deciphering these epithelial states is crucial for resolving the functional hierarchy of the tumor. Nonetheless, a comprehensive classification of CRLM subtypes based on epithelial cell differentiation states, integrating multi-omics data, is lacking. We propose a novel molecular subtype classification based on the differentiation status of epithelial cells in CRLM models, aiming to enhance understanding of the molecular characteristics of CRLM. In this study, we employed a comprehensive approach integrating multi-omics analyses, scRNA-seq, and multiplex biology to reconstruct the tumorigenic pathways of CRLM epithelial cells and establish a novel genetic subtyping and classification system. To distinguish different malignant development mechanisms, we propose a molecular categorization system—the Malignant Development Signature of CRLM (MDSCRLM)—based on epithelial cell differentiation states: Cluster 1—Aggressive-Metastasis-Enhanced CRLM (AMECRLM), Cluster 2—Cell Cycle-Active CRLM (CCACRLM), and Cluster 3—Growth-Inhibited CRLM (GICRLM). We then conducted multidimensional analyses and validation studies. In total, we identified seven distinct CRLM epithelial cell states and ten fate-determining genes. Furthermore, using the ConsensusClusterPlus algorithm, we defined a risk stratification framework based on the epithelial cell fate gene set. For each refined molecular classification, a global CRLM regulatory network was constructed, revealing intricate co-expression patterns among cell fate genes and their regulatory elements. We demonstrated the clinical relevance of our hierarchical model by applying it to CRLM patient tissue samples. Additionally, we validated the MDSCRLM classification system in a large-scale primary reference cohort and performed functional assays on AMECRLM to further support its biological and clinical rationale. To address this gap, we performed an integrative analysis of CRLM using scRNA-seq, bulk transcriptomics, and multi-omics datasets. We focused on the differentiation states and gene expression patterns of epithelial cells, which are central to metastatic behavior. Through unsupervised clustering and multi-dimensional analyses, we identified and characterized three novel predictive models of CRLM—Aggressive-Metastasis-Enhanced (AMECRLM), Cell Cycle-Active (CCACRLM), and Growth-Inhibited (GICRLM)—collectively termed the Malignant Development Signature of CRLM (MDSCRLM). We further delineated an epithelial cell fate gene set that underpins these subtypes and constructed a regulatory network integrating transcription factors, signaling pathways, immune infiltration, and proteomic features. The clinical validity and biological relevance of this classification were substantiated by survival analyses in a large-scale primary reference cohort, functional experiments, and clinical tissue validation. Our study provides new insights into the cellular and molecular landscape of CRLM, proposes a robust framework for molecular subtyping, and offers potential biomarkers and therapeutic targets for personalized management of metastatic CRC. Methods Data acquisition The raw data for scRNA-seq were obtained from the Gene Expression Omnibus (GEO) database (http://www.ncbi.nlm.nih.gov/geo/) (GEO accession: GSE178318). The data included 12 samples from 6 patients with a matched-pair design (primary tumor and synchronous liver metastasis). Multi-omics data and clinical medical information of primary colorectal cancer (CRC) samples (TCGA-COAD and TCGA-READ cohorts) were retrieved from The Cancer Genome Atlas (TCGA) database. 433 primary CRC samples with complete clinical data were utilized as a large-scale reference cohort to evaluate the prognostic relevance and molecular associations of the identified epithelial cell fate gene set. To assess the relative expression of genes, bulk RNA-seq transcriptome data were standardized to log2 (FPKM + 1). Before analysis, batch effects were corrected utilizing the limma package. (https://bioconductor.org/packages/release/bioc/html/limma.html). The lists of transcription factors were acquired from the Cistrome database (http://cistrome.org) [15].The graphical abstract was created using Biorender (https://biorender.com/). scRNA-seq and RNA-seq data processing and interpretation By following the protocol, we preprocessed the data using the 10x Genomics Chromium System (https://www.10xgenomics.com/instruments/chromium-x-series). Cellranger mkfastq from the Cell Ranger software package (v3.0, http://10xgenomics.com/) was applied to generate the fastq files which were drawn from raw scRNA-seq data. After aligning demultiplexed fastq data to the GRCh38 reference transcriptome, the gene expression matrix was identified with the cellranger count software. Seurat R toolkit (version 3.0.1; https://satijalab.org/seurat/) was applied to conduct dimensionality reduction, clustering, transcriptome analysis, and data visualization on all data [16, 17]. Following data alignment and generation of the gene expression matrix, stringent quality control (QC) was performed using the Seurat R toolkit (version 3.0.1). To filter out low-quality cells, empty droplets, and potential multiplets, cells were retained only if they expressed between 200 and 6000 genes (nFeature_RNA), and possessed a total Unique Molecular Identifier (UMI) count greater than 1000. To exclude dead or dying cells, we calculated the proportion of transcripts mapping to mitochondrial genes and strictly filtered out cells with a mitochondrial gene percentage exceeding 15%. Additionally, to eliminate artificial doublets formed during the droplet encapsulation process, the DoubletFinder R package was applied with default parameters adjusted for our specific multiplet rate, and predicted doublets were rigorously removed from the dataset. Only the remaining high-quality singlet cells were used for subsequent dimensionality reduction and clustering. Briefly, the top 2,000 highly variable genes were determined by “vst” selection and then used as input features for dimensionality reduction by means of principal component analysis (PCA). This is actually followed by the jackstraw analysis selection of the major principal components (PCs) [18]. Powerful PCs were included in graph-based Louvain clustering and Uniform Manifold Approximation and Projection (UMAP) for dimensionality reduction. Then, we identified the differentially expressed genes (DEGs) in every cluster, and also the genes whose expression was substantially different among clusters. Differential expression analysis was performed using Seurat’s “FindAllMarkers” function with default settings and an absolute log2 (fold change) larger than 0.50 (Wilcoxon rank-sum test) and a corrected p < 0.05 (Bonferroni correction). Particular cell clusters were identified as probable CRLM cell populations by analyzing the gene expression of identified cell markers using the singleR and the CellMarker database as references [19, 20]. In addition, the Cell Cycle Scoring algorithm in Seurat was employed for the analysis and measurement of the cell cycle based on previously identified cell cycle genes. Cells were projected onto UMAP space and colored based on cycle clustering for better visualization. Next, the data were proportionately divided into subsets, and only clusters required for independent analyses of CRLM epithelial cells were retained. The aforementioned analysis methods were reconducted for epithelial cells. Using Python3.6 and Seurat R Package, schematic graphics such as cellular feature plots, dot plots, and heat maps were generated [16]. RNA-seq was also conducted on CT26-sgHSPA1A cell lines and CT26-Nc cell lines. Integration of C1-H in Molecular Signatures Database (MsigDB) (https://www.gsea-msigdb.org/gsea/msigdb) and the RNA-seq results were performed for ORA analysis. Cellular communication analysis To reveal theoretical signaling interconnections between various cell populations of epithelial cells in CRLM, we used the iTALK package (iTALK: an R Package to Characterize and Illustrate Intercellular Communication) to recognize and match the differentially expressed ligand and receptor pairs, which was regarded as crucial intercellular communication. After identifying ligand-receptor pairings, the edgebundleR (https://github.com/garthtarr/edgebundleR) R package was applied for visualization. Enrichment analysis Using a Z-score limit of 1.96 and a p-value limit of 0.05, gene sets over-representation analysis (GSORA) was applied to quantify the biological enrichment or the pathway of the epithelial cell fate gene set [21]. Hallmark gene set and eight gene set subgroups (C1 to C8) were extracted from MSigDB, and the functional enrichment of the epithelial gene fate gene set was discovered using GSORA. Using the Gene Set Variation Analysis (GSVA) technique with 50 signature pathways in MSigDB and a statistical significance threshold of P value < 0.05, the comparative function of the pathway or unique signature among distinct epithelial clusters were examined. Similarly, the “gsva” R package was used to evaluate differentially regulated pathway enrichment between primary and metastatic CRLM samples [22, 23]. Utilizing the limma package, the differences and similarities between multiple cell clusters were determined. Single-cell trajectory construction Monocle2 [24] was performed to examine CRLM epithelial cell lineage pseudotime trajectories and detect the translational links between clusters of epithelial cells. For the sake of brevity, we first used the differential GeneTest tool with the default settings to screen ordering genes variably expressed among clusters. Following ordering gene selection, Monocle’s inverted graph embedding technique projected single cells into a low-dimensional space derived from expression profiles and ordered them in pseudotime trajectories. Also, we utilized Monocle2’s plot pseudotime heatmap to group DEGs for the purpose of confirming the validity of established epithelial cell lineage pseudotime and revealing the critical role of genes during the differentiation process. Consensus clustering Prognostic data, including overall survival (OS), and gene expression profiles were retrieved from the TCGA primary colorectal cancer (CRC) cohorts (TCGA-COAD and TCGA-READ). The R survival package [27] was used to perform Kaplan-Meier (K-M) and Univariate Cox (Uni Cox) regression analyses. To identify a robust prognostic epithelial cell fate gene set, we intersected the differentially expressed fate genes (identified from CRLM scRNA-seq) with genes showing significant prognostic value (K-M and Uni Cox analyses) in the TCGA primary CRC reference cohort. This final intersection gene set was used for clustering analysis. Using the “ConsensusClusterPlus” R package, we performed consensus clustering to stratify primary CRC patients in the TCGA cohort into distinct molecular subgroups based on the expression profiles of the epithelial cell fate gene set [25]. We employed Euclidean distance as our dissimilarity measure, and the Clustering was performed by partitioning around medoids to identify medoids for each cluster [22]. Average Silhouette, Elbow Method, and Gap Statistic were used to assess goodness of fit and determine the best grouping number, K. To examine the relationship between epithelial cell fate determination and multi-omics data, we conducted Principal Component Analysis [26] to determine the PCA CRLM score of the ideal k (number of clusters) according to the epithelial cell fate gene sets. Multi-omics analyses Comprehensive multi-omics data were retrieved from the TCGA primary colorectal cancer (CRC) cohorts (TCGA-COAD and TCGA-READ). These integrated datasets encompassed genomics (single nucleotide polymorphisms (SNP), microsatellite instability (MSI), tumor mutation burden (TMB), copy number variations (CNV)), transcriptomics (RNA-seq, immune gene sets, infiltrating immunocytes), proteomics (reverse-phase protein array (RPPA)), epigenetics (Assay for Transposase-Accessible Chromatin using sequencing (ATAC-seq)), and clinical information (clinical correlation, survival analysis, independent prognostic analysis). The ‘Maftools’ R package was used to evaluate and visualize the mutation patterns, identifying the distributions of TMB and producing the mutation gene waterfall plot, which displayed the type of mutations [27]. TMB was defined as the total number of somatic mutations per million bases in the genome of each CRLM sample, including insertions, deletions, and base substitutions. Based on the projected median TMB fraction, groups were devided into H-TMB and L-TMB groups. Using Kaplan–Meier analysis and the log-rank test, we merged PCA and TMB scores for each CRLM sample to assess its prognostic value. Read-depth analysis was used to detect CNVs. In order to ensure that the above variations were of high quality, we filtered them based on strand bias, read position, and quality depth. In the end, all the mutations were eventually assigned to genes based on corresponding genomic positions. To determine connection between PCA scores and TMB/MSI in CRLM, the Pearson correlation test was employed; a P value < 0.05 was considered to be significant. Single-Sample Gene Set Enrichment Analysis (ssGSEA) was used to identify 23 immune cell subgroups involved in both innate and adaptive immunity to analyze the CRLM immune infiltration environment. Using the ssGSEA package, we calculated the enrichment scores of gene signatures that serve as markers of immune cells to determine the extent of infiltration. The CIBESORT algorithm, developed for describing the cell makeup of complex tissues according to their gene expression patterns, was used to analyze the cell composition of infiltrating immune cells [28]. Central immune response markers’ expression such as PD-1 and CTLA4 was analyzed using the Wilcoxon signed-rank test. Tumor genotyping and Machine learning are used to calculate and adjust the immunophenoscore (IPS), which ranges from 0 to 10. Immunoreactivity rises in parallel with the score. In addition, The Cancer Immunome Atlas (TCIA) (https://tci.at/home) was utilized to retrieve immune cell proportion data and Immunophenoscores (IPS) for primary CRC patients in the TCGA cohort [29]. When comparing primary and metastatic CRLM tumors, we were able to distinguish between them based on their expression profiles of immune cell proportions, with a significance level of P value < 0.05. After reviewing the available research, we settled on 17 immune gene sets as representative of tumor immunity. After the ssGSEA of the 17 immune gene sets in the CRLM dataset, the GSVA package was used. We only continued processing data for genes and immune cells with a P value of less than 0.05. Then, to predict the ways in which each immune cell will communicate and recognize each other, we used a correlation analysis. Evaluation on epithelial cell fate gene set’s prognostic values The multivariate Cox model was applied to include the epithelial cell fate genes, and its efficiency was assessed using the receiver operator characteristic (ROC) curve. The following formula was then used to develop a risk score model of the fate genes involved in determining prognosis: \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\eqalign{& {\rm{Risk scorei = }} \cr & {\rm{\beta 1 \times fate}}\,{\rm{ gene 1 + \beta 2 \times }} \cr & {\rm{fate}}\,{\rm{ gene 2 + L + \beta n \times fate}}\,{\rm{ gene n}} \cr} $$\end{document} “β” meant the coefficients of fate genes in the multivariate Cox model, while “fate gene n” indicated the expression level of the Nth fate gene in each CRLM sample. Then, the median risk score for each CRLM patient was calculated, and the patients were classified into low- and high- risk groups. Using the survival package in R, we constructed K-M curves to compare patients’ survival according to their risk scores. We repeated our univariate and multivariate Cox regression analyses to identify the risk score’s independent prognostic significance, taking into account age, sex, and tumor stage to adjust the Cox model. Analysis of differential expression of bulk RNA-seq data Using the limma algorithm [27], DEGs were identified between primary CRLM samples from TCGA based on the log2 (fold change) > 1.0 or 1.0 and corrected P value < 0.05. Using the same cutoff, the limma algorithm also identified related TFs and pathways with significantly different expression levels. Global CRLM network as a regulatory network The association between epithelial cell fate genes sets and metastasis-related immune genes/TFs/pathways/related protein/immune cells was determined by using the Pearson analysis for co-analyze. Statistically significant difference was recognized when the correlation coefficient was greater than 0.600 and P was less than 0.001, and the igraph package was used to identify the components with the greatest geographic connection to regulatory networks [30]. Key fate genes as well as TFs had their gene symbols converted to their matching probes to construct positive and negative relations with each subtype and small molecule medicines, measured using a connectivity score from − 1 to 1. Then, we used a cutoff of P value<0.05 to select inhibitors that indicated a statistically significant correlation with corresponding fate genes and TFs. Patients collection This study was approved by the Institutional Review Board for Clinical Research of Affiliated Hospital of Yangzhou University (2024-YKL05-KE02) and the Institutional Review Board of Yangzhou University (YXYLL-2021-18). Written informed consent was obtained from all patients. A total of 45 tissue samples were collected at Affiliated Hospital of Yangzhou University. Histopathology and immunohistochemistry (IHC) Hematoxylin and eosin (H&E) staining was used to reveal general histology. In total, 45 tumor tissues from distant metastatic patients with CRC were fixed and embedded in formalin and paraffin. Thereafter, these samples were deparaffinized, dehydrated, rehydrated, antigen retrieved, blocked, and cut into 4-µm slices following routine procedures. Sections were incubated with anti-HSPA1A antibody (Cat#AF5466, Affinity), CCNB1(Cat#55004-1-AP, proteintech) and RPL4 (Cat#11302-1-AP, proteintech) at 4 °C overnight. Finally, all sections were labeled with polymer HRP for 40 min and counterstained with hematoxylin for 5 min at room temperature. All sections were examined by two pathologists, and cancer cells with a stained cytoplasm were defined as positive. Both cytoplasmic HSPA1A, CCNB1 and RPL4 were stained using the same antibody. The criteria for intensity scoring of tumor cells were as follows: negative (0); yellow (1); light brown (2); dark brown (3). For the negative control, a buffer was used instead of the primary antibody. Consequently, the final representative score for each sample was recorded as an integer from 0 to 3. All histopathological scoring was performed independently by two experienced pathologists who were blinded to the experimental groupings and clinical data, consistent with previously established immunohistochemical evaluation frameworks [31, 32]. To estimate the correlation between the intensity score of tumor cells and clinical characteristics of patients, the non-parametric test and Spearman correlation were applied. Knockdown cell line generation HEK293T cells were transfected using sgNC, sgHSPA1A#1 or sgHSPA1A#2, together with psPAX2 and pMD2.G, adding PEI in a ratio of 1:2. Viral supernatant was collected and used to infect CT26 cells. After 48 h of infection and subsequent culturing, puromycin was administered to successfully select infected cells. Detailed primers of plasmid construction were described in Table S1. Protein extraction and western blotting Cells lysis buffer (50mM Tris-HCl, 300mM NaCl, 1% Triton X-100, supplemented with protease inhibitor and phosphatase inhibitor cocktails) was utilised to extract total protein, and the protein was isolated through sodium dodecyl sulfate-polyacrylamide gel electrophoresis (SDS-PAGE). PBST-5% fat-free dried milk or BSA was used for blocking the pure protein extracts at room temperature for 1 h, followed by overnight incubation with anti-HSPA1A (Cat#AF5466, Affinity), anti-E-cadherin (Cat#3195, CST), anti-N-cadherin (Cat#13116, CST), anti-phospho-ERK1/2 (Cat#80031-1-RR, Proteintech), anti-ERK1/2 (Cat#11257-1-AP, Proteintech), anti-phospho-JNK (Cat#4668, CST), anti-JNK (Cat#9252, CST), anti-phospho-P38 (Cat#4511, CST), anti-P38 (Cat#9212, CST), and anti-GAPDH (Cat#ab8245, Abcam) at 4℃. Post-incubation, the samples were incubated with the relevant chemical-conjugated secondary antibodies (1:3000, Proteintech) at room temperature for 1 h. Blots were visualised using the Amersham Imager 600 (GE, USA). Migration and invasion assay Migration and invasion abilities of the stably selected CT26-sgHSPA1A#1, CT26-sgHSPA1A#2, and CT26-sgNC cell lines were evaluated using Boyden chamber assays. For the invasion assay, the upper chambers of 24-well transwell inserts (8-µm pore size) were pre-coated with Matrigel matrix (Corning Incorporated, USA) following the manufacturer’s instructions. The migration assay was performed similarly but lacked the Matrigel matrix coating. Prior to the assays, the stable cell lines were serum-starved for 24 h. Subsequently, 200 µl of cell suspension (1 × 10⁴ cells) in serum-free medium was seeded into the upper chamber. The bottom chamber was filled with 600 µl of complete medium containing 10% FBS to act as a chemoattractant. After incubation at 37 °C for 24 h (for migration) and 24 h (for invasion), the cells remaining on the upper surface of the membrane were gently removed with a cotton swab. The cells that had migrated or invaded to the lower surface of the membrane were washed with PBS, fixed with 4% paraformaldehyde for 15 min, and stained with 0.1% crystal violet for 25 min. Finally, a 100X magnification microscope was utilized to count the number of cells in three random fields per well. Wound healing assay CT26-sgHSPA1A#1, CT26-sgHSPA1A#2 and CT26-sgNC Cells were seeded in a six-well plate (1 × 105 cells per well). The cells were scratched with a pipette tip and washed with phosphate-buffered saline to remove detached cells. Cells were photographed at 0, and 48 h. The wound was calculated as follows: Relative percentage of wound healing (%) = (A0–A1)/A0 × 100, A0: initial wound area, A1: remaining area of wound at the metering point. RNA extraction, library preparation, and bulk RNA-sequencing Total RNA was extracted from the stably transfected CT26-sgNC and CT26-sgHSPA1A cell lines using the TRIzol reagent (Invitrogen) according to the manufacturer’s protocol. RNA concentration, purity, and integrity were evaluated using a NanoDrop spectrophotometer and Agilent 2100 Bioanalyzer. Subsequently, sequencing libraries were generated using the NEBNext® Ultra™ RNA Library Prep Kit for Illumina® following the manufacturer’s recommendations. The constructed libraries were sequenced on an Illumina NovaSeq 6000 platform, generating 150 bp paired-end reads. Raw sequencing data were subjected to quality control to remove adapter sequences and low-quality reads. Read counts mapping to each gene were quantified using FeatureCounts. Differential expression analysis between the CT26-sgNC control and CT26-sgHSPA1A knockout groups was performed using the DESeq2 R package. Differentially expressed genes (DEGs) were identified based on a threshold of |log2(fold change)| > 1.0 and an adjusted P-value < 0.05. The identified DEGs were visualized using a volcano plot and a hierarchical clustering heatmap, and were further subjected to KEGG pathway enrichment and Gene Set Enrichment Analysis (GSEA). In vivo tumor metastasis models All animal care and experimental procedures were approved by the Institutional Animal Care and Use Committee of Yangzhou University. Male BALB/c mice (6–8 weeks old) were housed under specific pathogen-free (SPF) conditions. To evaluate metastatic capacity, luciferase-expressing stable CT26 cell lines (CT26-sgNC, CT26-sgHSPA1A#1, and CT26-sgHSPA1A#2) were utilized. For the lung metastasis model, a total of 1 × 106 cells suspended in 100 µL of PBS were injected into the lateral tail vein of the mice (n = 5 per group). For the liver metastasis model, mice were anesthetized, and a small left flank incision was made to expose the spleen. A total of 1 × 10⁵ cells suspended in 50 µL of PBS were injected directly into the splenic subcapsule (n = 5 per group). The spleen was then returned to the abdominal cavity, and the incision was sutured. The in vivo progression of metastasis was monitored for a duration of 21 days. At the experimental endpoint, mice were intraperitoneally injected with 150 mg/kg of D-luciferin (PerkinElmer). After an incubation period of 10 min, bioluminescence signals were captured using the IVIS Spectrum In Vivo Imaging System (PerkinElmer). The luciferase signal intensity was quantified by drawing regions of interest (ROIs) encompassing the target organs, and the data were expressed as total flux (photons/second/cm²/steradian, p/s/cm²/sr) using the associated analysis software (Living Image software). Following the final imaging, all mice were humanely euthanized, and the target organs (lungs and livers) were harvested, photographed, and fixed in 4% paraformaldehyde for subsequent histopathological evaluation. Statistical analyses Differences in quantitative data between groups were analyzed using the unpaired two-tailed Student’s t-test, paired two-tailed Student’s t-test, and Mann-Whitney U test, depending on the data characteristics and study design. The unpaired two-tailed Student’s t-test was applied to compare means of two independent groups, assuming normal distribution and equal variances, while the paired version was used for comparing means within the same group at different times, assuming normal distribution of differences. The Mann-Whitney U test, a non-parametric alternative, was used to compare two independent groups without assuming normal distribution. These statistical analyses were performed using GraphPad Prism 6, IBM SPSS Statistics 20.0, and R 3.6.1 software. To run statistical analyses, the R 3.6.1 software (R Foundation for Statistical Computing, Vienna, Austria; www.r-project.org) was used. And when the two-tailed P value was less than 0.05, it was deemed statistically significant. The analysis processes and source code as well as the input data of bioinformatics analysis can be found in the supplementary material. Results Single-cell transcriptomic profiling reveals cellular composition and gene expression patterns in CRLM To characterize the cellular landscape of colorectal cancer liver metastases (CRLM), we performed scRNA-seq analysis using the GSE178318 discovery cohort, which comprises tumor tissues obtained from patients with CRLM (total cells analyzed: 140,281) [33]. Flowchart details were depicted in Fig. 1A and Figure S1. Following stringent quality control and filtering, we identified seven major cell types via unsupervised clustering and Uniform Manifold Approximation and Projection (UMAP) visualization: B cells, endothelial cells, epithelial cells, myeloid cells, natural killer (NK)/T cells, plasma cells, and stromal cells (Fig. 1B, C). Furthermore, we counted and evaluated the cells present in each sample, as well as the percentages of each cell type (Fig. 1D). Cell type identities were assigned based on canonical marker gene expression as annotated by the CellMarker database (Fig. 1E). Feature plots and heatmaps highlighted the expression of top differentially expressed genes (DEGs) in each cell type (Figure S2A). Genes such as LYZ and EPCAM were highly expressed in epithelial cells (Figure S2B). Quantitative analysis of cell type distribution revealed substantial inter-sample heterogeneity, with epithelial cells comprising a dominant fraction in certain samples (e.g., COL07_CRLM), while immune populations such as plasma cells and B cells were enriched in others (e.g., COL18_CRLM) (Figure S2B). Cell cycle analysis indicated that the majority of NK/T cells and epithelial cells were in S or G2 phases, as evidenced by cell cycle scoring and distribution histograms (Figure S2C, S2D). This finding suggests active proliferation within these populations, potentially contributing to tumor progression, modulation of the immune microenvironment, and facilitation of the epithelial–mesenchymal transition (EMT). The reliability of our cell type annotations was rigorously validated by evaluating the expression profiles of universally recognized canonical marker genes. As illustrated in the comprehensive dotplot (Figure S2E), each identified cluster exhibited highly specific marker expression: epithelial cells strongly expressed EPCAM, NK/T cells were marked by CD3D, B cells were characterized by CD79A, and myeloid cells robustly expressed CD68. Fig. 1Single-cell RNA sequencing (scRNA-seq) reveals gene expression patterns of colorectal cancer cells liver metastases. (A) The image emphasizes significant workflow steps. (B-C) UMAP map of 140,281 CRLM scRNA-seq cells, indicating clearly 15 clusters and 7 cell types (B cells, endothelial cells, epithelial cells, myeloid cells, NK/T cells, plasma cells, stromal cells). (D) Histograms depicting the quantity and proportion of each cell type inside every cluster. (E) The CellMarker database contains the gene expression levels of typical cell type markers for each cell type on UMAP plots (feature maps) Single-cell RNA sequencing (scRNA-seq) reveals gene expression patterns of colorectal cancer cells liver metastases. (A) The image emphasizes significant workflow steps. (B-C) UMAP map of 140,281 CRLM scRNA-seq cells, indicating clearly 15 clusters and 7 cell types (B cells, endothelial cells, epithelial cells, myeloid cells, NK/T cells, plasma cells, stromal cells). (D) Histograms depicting the quantity and proportion of each cell type inside every cluster. (E) The CellMarker database contains the gene expression levels of typical cell type markers for each cell type on UMAP plots (feature maps) Identification and characterization of epithelial cell subpopulations in CRLM It has been observed that CRC liver metastasis is mediated by epithelial cells. For instance, epithelial cells can change their polarity from apical-basal to anterior-posterior and promote the production of mesenchymal genes [34, 35]. Moreover, epithelial cells contribute to mesenchymal cells’ shape and migratory capacity, which in turn enhance mesenchymal invasion and subsequent metastasis [36]. Epithelial cells may transform into cancer stem cells (CSCs), and their propensity to infiltrate and metastasis would be considerably enhanced [37, 38].Given the central role of epithelial cells in CRLM progression, we performed a focused analysis on 18,192 epithelial cells extracted from the scRNA-seq dataset. UMAP dimensionality reduction revealed seven distinct epithelial subpopulations, termed as Epithelial cell 1–7 (Fig. 2A, B). Each subpopulation demonstrated unique gene expression signatures, with characteristic markers including CD44, CD24, PROM1, and MKI67 (Fig. 2C). Notably, CD24 was highly expressed across multiple epithelial clusters, consistent with its reported function in promoting angiogenesis, tumor recurrence, and metastatic potential in colorectal cancer [39]. The distribution of the top 5 DEGs for each epithelial subcluster is presented in Fig. 2D, further highlighting substantial intra-tumoral heterogeneity. Fig. 2scRNA-seq data enable the identification of CRLM clusters indicating epithelial cells. (A) An UMAP map of 18,192 malignant cells displaying seven subgroups of cells. (B) UMAP graphic illustrating which CRLM clusters each subgroup belongs to. (C) Stacked graph of epithelial cell 1–7 depicting the proportion of epithelial cells (relative to the total number of each CRLM cluster) along the y-axis. Cleveland dot plots displaying the expression levels of genes unique to epithelial cells in each cell subgroup. (D) Differential expression of the top five DEGs between malignancy clusters 1–7. (E) Each circle represents a subpopulation of cells in this network diagram of intercellular communication. The thickness of the line reflects the strength of the connection between two subpopulations, while its size of the circle indicates the strength of the interaction between the respective cell cluster and other cell clusters. iTALK analyzes the scRNA-seq data. (F) Using circular plots displaying receptor-ligand pairs for all mentioned cell subpopulations, the strongest interaction link was seen between epithelial3 and others. (G) Results of the GSEA enrichment of the signature pathway database scRNA-seq data enable the identification of CRLM clusters indicating epithelial cells. (A) An UMAP map of 18,192 malignant cells displaying seven subgroups of cells. (B) UMAP graphic illustrating which CRLM clusters each subgroup belongs to. (C) Stacked graph of epithelial cell 1–7 depicting the proportion of epithelial cells (relative to the total number of each CRLM cluster) along the y-axis. Cleveland dot plots displaying the expression levels of genes unique to epithelial cells in each cell subgroup. (D) Differential expression of the top five DEGs between malignancy clusters 1–7. (E) Each circle represents a subpopulation of cells in this network diagram of intercellular communication. The thickness of the line reflects the strength of the connection between two subpopulations, while its size of the circle indicates the strength of the interaction between the respective cell cluster and other cell clusters. iTALK analyzes the scRNA-seq data. (F) Using circular plots displaying receptor-ligand pairs for all mentioned cell subpopulations, the strongest interaction link was seen between epithelial3 and others. (G) Results of the GSEA enrichment of the signature pathway database To investigate the functional interactions among epithelial cell subpopulations, we conducted ligand–receptor interaction analysis using the iTALK algorithm. This analysis revealed that epithelial subpopulations 3 and 5 served as central hubs in intercellular communication networks, engaging in frequent ligand–receptor interactions with other epithelial subtypes (Fig. 2E and F). These subpopulations may play pivotal roles in orchestrating tumor cell signaling and microenvironmental crosstalk during metastasis. Gene Set Variation Analysis (GSVA) of hallmark pathways demonstrated that epithelial subpopulations exhibit distinct pathway enrichment profiles, DEGs were most enriched in the underlying hallmark signaling pathways, including HALLMARK_PANCREAS_BETA_CELLS, HALLMARK_KRAS_SIGNALING_DN, HALLMARK_BILE_ACID_METABOLISM, HALLMARK_ANGIOGENESIS, HALLMARK_OXIDATIVE_PHOSPHORYLATION, HALLMARK_MYC_TARGETS_V1, HALLMARK_MTORC1_SIGNALING, HALLMARK_MYC_TARGETS_V2. Specifically, pathways related to KRAS signaling and angiogenesis were predominantly enriched in subpopulations 3 and 5, suggesting their involvement in metastatic dissemination and neovascularization (Fig. 2G). Additionally, enrichment of cell cycle and DNA repair pathways in other subpopulations underscores the functional diversity of epithelial cells within CRLM. Collectively, these results delineate the heterogeneity of epithelial cell lineages in CRLM and identify key subpopulations and signaling pathways that may drive metastatic progression. Pseudotime analysis reveals epithelial cell fate trajectories and key fate-determining genes in CRLM We performed Monocle2 to undertake a pseudotime analysis depending on the transcriptional changes during proliferation and differentiation in order to rebuild a cell-developmental trajectory and gain insight into the dynamics of gene expression during the course of continuous development. Four clusters were found and given unique labels after using UMAP/Monocle 2 to illustrate the trajectory (Figure S3A). At each time point, epithelial cells diverged into differing gene expression patterns, with branches 1–3 delineating seven distinct future cellular stages (Fig. 3A, Figure S3B). Figure 3B and C showed how the reconstructed branched tree can lead to many cell fates when cells from the initial population, represented by point 2, progress through multiple intermediate subpopulations. Starting with state 2, given as the least mature stage in the pseudotime and the progenitors of other subpopulations, Epithelial Clusters 1, 2, and 3 are the most numerous. State-5 cells were composed primarily of epithelial cell4, while state-6 cells were composed primarily of epithelial cell1 (Fig. 3B, S3B). We then analyzed DEGs between the two main branches (cell fate 1 vs. 2), beginning at the initial bifurcation point (branch point 3). These DEGs were Hierarchical clustering into 7 groups (1–7) based on the expression patterns of their genes across a variety of cell types (Fig. 3D). Four distinct gene subgroups were enriched by the DEGs for cell fate 1(cluster 1, 2, 3, 4). Correspondingly, our study clearly defined cluster 5, 6, 7 for the cell fate 2 subgroups. Fig. 3Utilizing Monocle2, pseudotemporal analysis identifies epithelial cell developmental pathways and fate-associated genes. (A) Developmental trajectory of the epithelial cell depicted in pseudotime, with seven distinct cell states indicated by color. (B) Developmental trajectories of epithelial cells with pseudotime; the trajectories depict all subpopulations. (C) Subpopulations of epithelial cells in CRLM that contribute to each state are presented in the accompanying trajectories, which are labeled with the cell type. (D) Heatmap displaying branch-specific DEG expression pattern alternation at branch point 3 in different cell types Utilizing Monocle2, pseudotemporal analysis identifies epithelial cell developmental pathways and fate-associated genes. (A) Developmental trajectory of the epithelial cell depicted in pseudotime, with seven distinct cell states indicated by color. (B) Developmental trajectories of epithelial cells with pseudotime; the trajectories depict all subpopulations. (C) Subpopulations of epithelial cells in CRLM that contribute to each state are presented in the accompanying trajectories, which are labeled with the cell type. (D) Heatmap displaying branch-specific DEG expression pattern alternation at branch point 3 in different cell types Based on the visualization of significantly differentially expressed sample genes from our scRNA-Seq research, we hypothesize that RPL11 plays a crucial role across all differentiation phases due to its peak expression pattern, which tends to span the entire pseudotime trajectory. The highly expressed genes CDA, STMN1 were discovered to be enriched in the end trajectory and upregulated along pseudotime trajectories, which was abundant in state-1 cells. ENO1, IFI6, ISG15 were also upregulated in the middle trajectory (Figure S3C, S3D). Sample COL07 CRLM showed an overexpression of CDA, ENO1 and ISG15, while RPL11 was expressed in all samples (Figure S3E). Combining the prognostic data from TCGA and Monocle 2 analysis, we identified a set of epithelial fate genes. Our results successfully uncovered the transcriptome landscape throughout the commitment to epithelial cell fate. Moreover, a significant gene set (352 DEGs, P < 0.05) was used to explore genes associated with fate decisions (Fig. 3D). Additionally, gene expression patterns from CRLM RNA-seq data and clinical information provided by TCGA were used to conduct survival analysis using the K-M approach and Uni Cox regression model. The significant DEGs from K-M overlapped with those from Uni Cox by 10 genes. Eventually, 10 genes (CXCL1, CTSD, NEK9, ANGPTL4, HSPA1A, SNAI1, CALB2, CD24, FKBP4, and CCNA2) were identified as the final epithelial cell fate gene set and included for consensus clustering (Figure S4A, B). KM curves were used to compare survival probabilities of high-grade CRC patients in the TCGA cohort. (Figure S4C). These findings establish the existence of distinct epithelial cell fate trajectories in CRLM and identify a set of genes that may underlie the transition between different malignant states. The continuous clustering of CRLM depending on the epithelial cell fate gene set We used ConsensusClusterPlus’s algorithm to further group epithelial cells, and made survival analysis, immunotherapy response analysis and other analysis to determine the rationality of typing. The next problem was whether or not the identified gene set may shed light on the additional functional effects of biological programs determining survival outcomes. The regulation mechanism of medical prognosis as well as epithelial cell fate by this gene set was shown in the regulatory network (Fig. 4A), which demonstrated the extremely high connection and close correlation, either positively or negatively, between 10 epithelial cell fate genes. The linkages between 10 genes were all significant at P < 0.0001, and the entire gene collection is co-expressed in the epithelial cell spot of CRLM during differentiation. In addition, the left part of the circle (gene) was colored to show how the regulatory network controls the divergence of epithelial cell fates. Nearly half of the genes [40], including SLC6A8, SNAI1, and NUPR1, have been identified as State1 determinant genes responsible for the fate of epithelial cells of CRLM into cell state 1, whereas only two genes (orange) lead to the differentiation towards cell State 6. In addition to regulating the fate of epithelial cells, the revealed gene set was proven to be substantially linked with the prognosis of CRLM. On the basis of the survival analysis, 10 genes were divided into two groups (risk and favorable factor), correlating with the colors in the right part of the circle. The size of the circle reflected the P-value of the univariate cox multiple regression model on OS for each gene. As a result, those with elevated levels of HSPA1A, CALB2, CXCL1 and ANGPTL4 had the worse prognosis. Fig. 4Consensus clustering of CRLM TCGA cohorts utilizing genes involved in epithelial cell fate. (A) Regulatory network of 27 genes encoding epithelial cell fate. The left half of each circle represents the differentiation status of the epithelial cell fate, while the right half represents the high expression of the gene as a risk factor (red) or a protective factor (blue). The lines represent statistically significant (P < 0.0001) relationships between genes. Pink indicates a positive correlation, while blue indicates a negative correlation. P-value is inversely related to circle size (Cox test). (B)Using a consensus clustering approach based on 27 epithelial cell fate genes, the RNA-seq data of CRLM in TCGA were grouped into three clusters. On this basis, a heatmap is generated that corresponds to the k = 3 consensus matrix, with samples represented by columns and rows. Given that the consensus value shows the degree of consensus and ranges from white (0, not grouped together) to blue (1, always clustered together), k = 3 is the ideal clustering solution. (C) The CDF delta area curve demonstrates that the area under the CDF curve does not change considerably after the cluster number k = 4, but CRLM tumor cells are separated into three clusters after careful analysis. (D) Heatmap derived from differential analysis of 3 clusters of consensus clusters based on 27 gene sets encoding epithelial cell fate. (E) Kaplan-Meier survival analysis demonstrating survival probability variations across the three clusters (P = 0.01309). (F) Kaplan-Meier curves of high and low-risk CRLM patient groups based on PCA scores. (G) PCA scores for three clusters as determined by the Mann-Whitney U test. (H) Cluster heatmap of the three CRLM cell line groups based on the similarity of their expression profiles (I) The AUC of the area under the ROC curve depicted in the boxplot demonstrates that the sensitivity of CRLM clusters to the medicine varies Consensus clustering of CRLM TCGA cohorts utilizing genes involved in epithelial cell fate. (A) Regulatory network of 27 genes encoding epithelial cell fate. The left half of each circle represents the differentiation status of the epithelial cell fate, while the right half represents the high expression of the gene as a risk factor (red) or a protective factor (blue). The lines represent statistically significant (P < 0.0001) relationships between genes. Pink indicates a positive correlation, while blue indicates a negative correlation. P-value is inversely related to circle size (Cox test). (B)Using a consensus clustering approach based on 27 epithelial cell fate genes, the RNA-seq data of CRLM in TCGA were grouped into three clusters. On this basis, a heatmap is generated that corresponds to the k = 3 consensus matrix, with samples represented by columns and rows. Given that the consensus value shows the degree of consensus and ranges from white (0, not grouped together) to blue (1, always clustered together), k = 3 is the ideal clustering solution. (C) The CDF delta area curve demonstrates that the area under the CDF curve does not change considerably after the cluster number k = 4, but CRLM tumor cells are separated into three clusters after careful analysis. (D) Heatmap derived from differential analysis of 3 clusters of consensus clusters based on 27 gene sets encoding epithelial cell fate. (E) Kaplan-Meier survival analysis demonstrating survival probability variations across the three clusters (P = 0.01309). (F) Kaplan-Meier curves of high and low-risk CRLM patient groups based on PCA scores. (G) PCA scores for three clusters as determined by the Mann-Whitney U test. (H) Cluster heatmap of the three CRLM cell line groups based on the similarity of their expression profiles (I) The AUC of the area under the ROC curve depicted in the boxplot demonstrates that the sensitivity of CRLM clusters to the medicine varies We then applied consensus-based clustering to TCGA primary CRC reference cohort bulk RNA-seq data using the ConsensusClusterPlus R package based on the epithelial cell fate gene set. After estimating the intercluster variation coefficients and the average clustering consistency for each cluster, we were able to determine the optimum number of clusters. According to the CDF delta area curve, three clusters is the perfect balance for further multi-omics analysis, after which the clustering result stabilizes (Fig. 4B). To get a better sense of the scale and character of clustering, we built a consensus matrix plot in which 10 genes were divided into three clusters (Fig. 4C). Clinicopathological factors such as stage, race, gender, age, and censor status were found to be significantly correlated with cluster assignments. Cell state 1 determinant genes were highly expressed in cluster 1, whereas determinant genes for cell states 3, 4, 5, and 6 were predominantly enriched in cluster 2. State 7 determinant genes were most strongly expressed in cluster 3 (Fig. 4D). Results from Kaplan–Meier survival analysis revealed statistically significant difference in survival probability across the three clusters (P = 0.01309), with samples in clusters 3 indicating better prognoses compare to those in clusters 1 and 2 (Fig. 4E). These findings confirm that consensus clustering based on the epithelial cell fate gene set is closely associated with clinical outcomes in CRLM. On the basis of marker gene expression and functional annotation, the three predictive models were designated as: Cluster 1—Aggressive-Metastasis-Enhanced CRLM (AMECRLM); Cluster 2—Cell Cycle-Active CRLM (CCACRLM); and Cluster 3—Growth-Inhibited CRLM (GICRLM). Principal component analysis (PCA) was carried out utilizing the epithelial cell fate gene to evaluate the ability to differentiate across clusters, thereby defining substantial variations between samples from different groups (Figure S5A). Samples were stratified into high- and low-score groups based on PCA scores, revealing that patients in the high-score group had significantly improved prognosis (P = 0.004) (Fig. 4F). Moreover, integrative analysis supports PCA and survival analysis. Samples with higher PCA scores tended to gather in cluster 2, which were significantly associated with better prognosis (Fig. 4G). The Sankey diagram has shown the correlation between Clusters, PCA score, and Censor (Figure S5B). Further analysis of cell fate gene expression across clusters revealed significant differences, although genes such as SNAI1, SLC6A8, CD24, and AVPI1 were not significantly associated with clustering (Figure S5C). Expression profiles of cell fate genes in each cluster were plotted to RNA-seq gene expression data for CRLM cells collected from CCLE and GDSC data sets in order to investigate the sensitivity of three CRLM patient clusters to standard chemotherapy. In this study, 41 cell subpopulations were divided into three categories, each of which represented a distinct group of patients (Fig. 4H). Drug sensitivity analysis using ROC curves, based on gene expression responses to chemotherapeutic agents, indicated that cell subpopulations in groups 1 and 3 were more sensitive to agents such as SU-11,274, SABUTOCLAX BI-97C1, ELESCLOMOL, and CISPLATIN. In contrast, group 2 cells exhibited reduced sensitivity to all commonly used chemotherapeutics (Fig. 4I). Taken together, these results suggest that the identified CRLM subtypes may exhibit distinct drug response profiles, with important implications for personalized therapy. Establishment of genetic hallmarks and multi-dimensional omics potential analysis in CRLM Genomic and transcriptomic features for CRLM prognosis To evaluate the functional significance of the 10 fate genes, an over-representation analysis (ORA) was undertaken (Fig. 5A, P-value<0.05, adjusted). To cross-validate the prognostic robustness of the epithelial cell fate gene set, we performed a parallel screening on the parent DEGs using univariate Cox regression, which identified 17 survival-associated genes (Fig. 5B). This broader prognostic analysis corroborated the clinical significance of our core 10-gene signature. Furthermore, by integrating these prognostic data with clinical correlation analyses (Fig. 4D), we identified HSPA1A, CCNB1, and RPL4 as the most representative markers for the three subtypes, which were subsequently selected for phenotypic validation. To prevent statistical model overfitting, lasso regression with finite criteria and ten-fold cross-validation were employed to identify the top 8 genes as predictors of prognostic fate (Figure S6A, B). The entire CRLM cohort is then randomly divided into test and training cohorts in a ratio of 4:1. Applying an acceptable threshold (median risk score), black dotted boundaries separate high-risk groups from low-risk groups. Based on the survival curves, the survival probability of the subgroup with the lowest risk was considerably higher than that of the subgroup with the highest risk (Figure S6C, P < 0.001). Exploring further the presence of prognostic fate gene signatures between the low-risk and high-risk groups, significant genetic differences were found between the two groups (Fig. 5C, S6D). Using scatterplots and risk trajectories to depict the association between clinical characteristics and risk scores, CRLM patients were separated into low-risk (blue dots) and high-risk (red dots) groups, with the high-risk group having a increased mortality and shorter survival time (Figure S6E, F). In addition, the AUC Receiver Operating Characteristic (ROC) values for the three cohorts were 0.76, 0.81, and 0.70, demonstrating that the prognostic fate gene set functioned well (Figure S6G). Fig. 5Developing risk signatures associated with fate genes. (A) Functional enrichment analysis of DE fate genes in the MSigDB gene collection based on a comparison between primary and metastatic CRLM. (B) Univariate Cox regression of epithelial cell fate genes, forest plot displaying risk factors and protection factors (blue). (C) Boxplot depicting significant differences between high-risk and low-risk groups for eight prognostic signature genes. (D) Forrest plots of univariate and multivariate Cox regression analyses of four clinical variables influencing the prognosis of CRLM. (E) Bar graph illustrating the distribution of clinical characteristics. (F) GSEA reveals biological processes that are up- or down-regulated in high- and low-risk groups Developing risk signatures associated with fate genes. (A) Functional enrichment analysis of DE fate genes in the MSigDB gene collection based on a comparison between primary and metastatic CRLM. (B) Univariate Cox regression of epithelial cell fate genes, forest plot displaying risk factors and protection factors (blue). (C) Boxplot depicting significant differences between high-risk and low-risk groups for eight prognostic signature genes. (D) Forrest plots of univariate and multivariate Cox regression analyses of four clinical variables influencing the prognosis of CRLM. (E) Bar graph illustrating the distribution of clinical characteristics. (F) GSEA reveals biological processes that are up- or down-regulated in high- and low-risk groups Four variables, namely, risk score, tumor stage, sex, and age, were included in univariate and multivariate Cox model analyzes to assess whether risk score was an independent prognostic indicator for CRLM. The risk score model in CRLM patients was independent of other clinicopathological characteristics (univariate model: HR = 45.224, 95% CI (13.017-157.113), P<0.001; multivariate model: HR = 20.415, 95% CI (5.349–77.918), P<0.001) (Fig. 5D). This model described the clinical characteristics of the complete CRLM cohort (Fig. 5E). There was a correlation between high-risk populations and increased mortality and advanced tumor stages. Pathophysiological processes such as angiogenesis, cell adhesion, immunological response, epithelial-mesenchymal transition, lipid metabolism, and hypoxia-related pathways were predominantly activated in the high group of GO, KEGG, and HALLMARK databases. In contrast, the low-risk group had decreased cell cycle checkpoints, E2F, MYC, G2M, cell cycle, and biometabolism (Fig. 5F). Epithelial cell fate gene labels discovery via multi-dimensional omics potential analysis Our combined bulk/single-cell multi-omics research sheds light on the biological complexity of epithelial CRLM cells. Multi-omics analyses were performed on the TCGA dataset’s CRLM patient genomic data, investigating the correlation between multi-dimensional omics signatures and stable clustering & principal component analysis (PCA) score. High-level groups were more likely to have more mutant genes and mutation types, indicating that genetics with higher PCA scores are more possible to be mutated and that their genomes were more unstable. In addition, the frequency of harmful mutations in certain genes was significantly greater in subgroups with high or low prevalence. Significantly greater incidences of gene mutations in FBXW7 (19% vs. 9%) and TTN (53% vs. 42%) were observed in the PCA high group, indicating a better prognosis, whereas in the low group, mutations such as TP53 (61% vs. 51%) were significantly increased, which was in accordance with the findings obtained from the above PCA score survival analysis and can assist in dividing patients into good and poor prognosis groups. It had been demonstrated that the TP53-R273H mutation enhanced the stemness of colorectal cancer cells and worsened the prognosis of CRLM patients by regulating specific lncRNAs [41]. Therefore, mutations in TP53 may predict poor prognosis (Figure S7A, B). The mutation data was utilized to calculate the Tumor Mutational Burden (TMB) across the samples. Matching TMB to the PCA score of CRLM samples revealed that patients with a higher PCA score also had a greater TMB, and there was significant between Low Group and High Group (P = 0.0017) (Figure S7C). The scatter plot also showed a positive relationship between the PCA score and TMB (R = 0.24), with higher TMB seen in samples from clusters 2, 3 than cluster 1 (Figure S7D). Once we had separated all CRLM patients into those with high and low TMB levels, we used K-M analysis to look into how the two groups were connected. Notably, the discriminative power of CRLM prognosis was enhanced when PCA score was combined with TMB. Subgroups with a high PCA score and TMB had the best prognosis, while those with a low TMB and score had the worst one (Figure S7E, P = 0.003). Copy number variation (CNV), a fundamental measure of tumor heterogeneity, revealed 23 epithelial cell fate genes in TCGA primary reference cohort samples. The majority of the detected CNVs are located on certain chromosomes, such as chromosomes 4, 5, 6, 10, 14, 15, and 16, and more commonly in genes such as FKBP4, CYP2W1, ALDH1A3, HSPA1A, NUPR1, etc. On the other side, we discovered that genes such as AVP11, MAD2L1, CCNA2, DEFA6, CCNB1, and SERPINA1 commonly lost genomic areas spanning several chromosomes 7, 12, 15, 16, etc. (Figure S7F, G). We discovered no statistically significant difference between MSI and PCA score groups (Figure S7H, P = 0.45). Nonetheless, the MSI score was positively connected with the PCA score, and the unstable MSI was primarily centered in clusters 2 and 3, with each cluster’s MSI being statistically different (Figure S7I, R = 0.1, P = 0.041). The ssGSEA algorithm was then used to analyze the relative proportions of epithelial cell fate genes identified by the three clusters. To investigate the infiltration of 23 immune cell subsets in the epithelial cells microenvironment of CRLM patients, we analyzed the immune cell infiltration among the three clusters; the results revealed that the majority of immune cells exhibited significantly higher infiltration in cluster 1, whereas cluster 2 degree demonstrated lower immune infiltration, except activated CD4 T cells, neutrophils, Type 17 T.helper cells, and Type 2 T.helper cells (Figure S8A). To further study the killing efficacy of immune cells infiltrating tumor tissue, the expression of PD-L1 on T cells can be used as a useful reference point [42]. Specifically, CD274 on immune cells is connected with a favorable prognosis, although CD274 was diverse on tumor cells and did not match the criteria for prognostic indicators [43]. To determine if the observed association between PCA score, clustering, and prognosis is influenced by CD274, we evaluated the expression of PD-L1 between high and low scores using RNA-seq data. CD274 as a checkpoint PD-1 ligand was discovered to be highly and significantly expressed in high scores (P = 0.039), which had previously been associated with a favorable prognosis. This evidence enhanced the diagnostic and prognostic utility of PD-L1 in CRLM (Figure S8B). In light of the fact that tumor immunogenicity was assessed using the The Cancer Immunome Atlas (TCIA), we generated the scores of IPS vs. PD1 blockers, IPS vs. CTLA4 blockers, and IPS vs. CTLA4 and PD1 blockers to study relationship between IPS and PCA CRLM scores. Clusters 2 and 3 of CRLM patients may respond more favorably to immune checkpoint inhibitors (ICIs) than cluster 1 (Figure S8C). Immune processes inside the tumor microenvironment play a vital role in the development of many malignancies. Given that adaptive immunological resistance can inactivate numerous types of immune cells in high-grade CRLM and that innate immune resistance in CRLM patients decreases the immune response, the poor prognosis of patients with high-grade CRLM may be attributable to substantial immunotherapy resistance [43]. Subsequently, we analyzed the association between 23 immune cell types and PCA scores, and discovered positive correlations between many immune cell types and PCA scores in CRLM, as well as cellular communication between distinct immune cells, using a correlation matrix. Results may also explain why a higher grade correlates with a higher PCA score and a better prognosis (Figure S8D). Collectively, these findings suggest that the poor prognosis of the high-grade subtype (Cluster 1) may be attributable to TP53-driven aggressiveness and relative immunotherapy resistance, whereas the high-grade/high-PCA subtypes (Clusters 2 and 3) are characterized by higher TMB, CD274 expression, and better potential for immunotherapeutic intervention. Establishment of the global extensive CRLM network To further understand the molecular pathways by which epithelial cell fate genes may contribute to CRLM start and progression, an interconnected regulatory network was studied to determine the essential players and the manner in which preferred regulators interact with important genes. We used primary CRLM RNA-seq data from TCGA and focused on cell fate-determining genes because of the strong correlation between tumor differentiation and the development of metastases. Genome, transcriptome, and pathway-level analysis Heatmap and a volcano plot of DEGs revealed significant variations in expression of genes between primary and metastatic CRLM when the |log2 fold-change (FC)| > 1.0 and adjusted P-value <0.05 was defined as the criteria for differential expression analysis (Figure S9A, B). The association between the detected fate genes and metastasis was further demonstrated by using the same analysis threshold to evaluate if the 27 fate genes were also significantly differentially expressed (Figure S9C, D). The following phase involves an examination of transcription factors (TFs) and pathways that are enriched in metastasis. An absolutely enormous number of substantially differentially expressed TFs influencing DEG fate genes were found using the criteria |log2 FC| >1 and FDR < 0.05 (Figure S9E, F). GSVA was used to identify 24 metastasis-related pathways that were significantly differently expressed between main and metastatic CRLM after examining the levels of various signature pathways. Specifically, the results demonstrated a considerable overexpression of 18 important pathways. Among these, Mitotic and ADIPOGENESIS pathways revealed the most pronounced overexpression; the former has been demonstrated to be a critical driver of colorectal cancer (CRLM) development and progression [44]; while CRLM’s disorders of lipid metabolism can affect the treatment of colorectal cancer [45]. In contrast, the lowest t-value of GSVA score was obtained for “PANCREAS BETA CELLS,” indicating limited expression of pancreatic β cells susceptible to CRLM proliferation and metastasis (Figure S9G, H). Immune infiltration landscape Depending on the quantity of immune-infiltrating cells, the clinical outcomes of patients with identically histologically classified tumors may vary. Using the CIBERSORT method, we then inferred the cellular infiltration of the immunological microenvironment in CRLM. Stacked histograms depict the proportions of 22 immune cell types in the CRLM cohort, with macrophages comprising the greatest proportion of immune cells (Figure S10A). In the high-risk group, macrophage M2 and macrophage M0 exhibited more infiltration characteristics, whereas macrophage M0 and NK cells were less infiltrating in the low-risk group (Figure S10B). Simultaneously, Using the scRNA-seq cohort, we compared immune cells within primary and metastatic cancers. More macrophage M2 and monocytes infiltrate initial tumors, whereas CD8 + T cells, CD4 + memory T cells, and macrophage M0 infiltrate metastatic tumors (Figure S10C). In addition, a K-M analysis of survival was done on the CRLM cohort utilizing the immune cell score and immune function score. Considering the statistical importance of the data, we chose a panel of important immune cells relevant to survival for network construction (Figure S10D). Construction of comprehensive regulatory networks for CRLM subtypes On the basis of fate genes and metastasis-related transcription factors, pre-transcriptional and transcriptional markers for each subtype that are highly associated with metastasis and recurrence were found (Figure S11A: AMECRLM:PPARD/HSF1/SNAI1; CCACRLM:E2F7/PTTG1; GICRLM:NFIC/TEAD1/PPARG). Therefore, these genes may be potential specific therapeutic targets for different subtypes. Each isoform’s functional properties are illustrated in Figure S11B. We constructed a regulatory network to present richer differential co-expression patterns, which we refer to as the global CRLM network, based on the upstream factor (metastasis-associated TF), midstream factor (key DE fate gene), downstream factor (metastasis-associated pathway), ssGSEA (immune function), infiltrating immune cells, and RPPA (metastasis-associated protein) (Fig. 6A). Subtype 1 marker genes were found to be SLC6A8 and SNAI1, while subtype 2 marker genes were found to be CCNB1 and PTTG1, and subtype 3 marker gene was found to be NEK9. Many studies have found that specific marker genes play a significant role in CRLM growth and metastasis. In addition, each co-expression pattern of all components was depicted by a co-expression heatmap, and all components were shown to have strong relationships with one another across all networks (Fig. 6B). Using ATAC-seq analysis of samples from 81 primary CRC patients in the TCGA-COAD cohort, we further investigated the accessible chromatin landscape to identify putative regulatory areas. This analysis aims to uncover the epigenetic priming mechanisms by which critical fate genes are prematurely activated in the primary tumor, thereby establishing the molecular foundation that subsequently promotes CRLM tumor growth and metastasis. Genomic Markers Massive alterations in chromatin accessibility were seen in both cases, and these modifications were found throughout the samples’ whole genomes. In addition, we discovered that for each marker gene, alterations in chromatin accessibility (or open chromatin peak patterns) were detectable in a variety of CRLM tumor tissue types (Fig. 6C). As a result, our results can serve as a foundational basis for more in-depth studies into drug development. Fig. 6Establishing an integrated regulatory network. (A) Global CRLM network demonstrating tight gene association and coordinated expression control. The blue circles represent essential DE fate genes, the yellow circles represent essential TFs, the red circles represent pathways, the brown circles represent immunological functions, the cyan circles represent immune cells, and the green circles represent metastasis-associated proteins. (B) The co-expression heatmap identifies co-expression patterns and correlations among all subgroup components. (C) ATAC-seq tracks signals detected at several loci containing marker genes, as indicated, with open chromatin peak profiles showing corresponding chromatin areas, demonstrating that chromosomal regions are accessible and open Establishing an integrated regulatory network. (A) Global CRLM network demonstrating tight gene association and coordinated expression control. The blue circles represent essential DE fate genes, the yellow circles represent essential TFs, the red circles represent pathways, the brown circles represent immunological functions, the cyan circles represent immune cells, and the green circles represent metastasis-associated proteins. (B) The co-expression heatmap identifies co-expression patterns and correlations among all subgroup components. (C) ATAC-seq tracks signals detected at several loci containing marker genes, as indicated, with open chromatin peak profiles showing corresponding chromatin areas, demonstrating that chromosomal regions are accessible and open Immunohistochemical staining of patient samples and external validation with patient clinical information To translate our 10-gene epithelial fate discovery framework into a clinically applicable diagnostic tool, we applied a stepwise marker-reduction strategy. The final 3-marker protein panel was prioritized based on three stringent criteria. First, we selected targets with the most significant cluster-specific differential expression to maximize discriminatory power. Second, we ensured functional representativeness: HSPA1A reflects the aggressive stress-response phenotype of AMECRLM, CCNB1 mirrors the hyperactive proliferation of CCACRLM, and RPL4 aligns with the translational regulation observed in GICRLM. Finally, to guarantee translational reproducibility, we strictly filtered for markers with highly specific, commercially available clinical-grade antibodies. Consequently, this refined the 10-gene signature into a robust and practical 3-marker panel (HSPA1A, CCNB1, and RPL4) for subsequent immunohistochemical validation. Consequently, this refined the 10-gene signature into a robust and practical 3-marker panel to experimentally validate these metastatic risk trajectories in situ. To apply our transcriptomic stratification model in a clinical setting, we evaluated the protein expression of HSPA1A, CCNB1, and RPL4 using a 4-tier semi-quantitative scoring system (integer scores 0–3) in an independent cohort of 45 patients diagnosed with stage IV colorectal cancer at our center. Crucially, rather than evaluating pre-defined subgroups, this patient cohort was stratified directly based on the dominant IHC staining intensity among these three markers. Specifically, tumors exhibiting predominant HSPA1A expression were assigned to the AMECRLM trajectory (Cluster 1), those with dominant CCNB1 expression to the CCACRLM trajectory (Cluster 2), and those with predominant RPL4 expression to the GICRLM trajectory (Cluster 3). Consistent with our marker-driven stratification criteria, quantitative analysis confirmed that the expression scores of HSPA1A and RPL4 were significantly enriched in their assigned subtypes—AMECRLM and GICRLM, respectively—compared to the remaining groups (P < 0.05). Although the CCNB1 expression score showed a visible upward trend in the CCACRLM trajectory, this difference did not reach statistical significance (ns) (Fig. 7A). This lack of statistical significance for CCNB1 is likely attributable to the limited sample size within this specific intermediate subtype in our current validation cohort. Nevertheless, this direct marker-driven assignment allowed us to robustly translate the multi-omics signature into an accessible clinical workflow. Following this direct stratification, we comprehensively analyzed the clinical outcomes and pathological features of these 45 patients. Significant differences were observed in both overall survival (OS) and recurrence-free survival (RFS) among the three marker-defined clinical subtypes. Consistent with our computational projections, Cluster 3 (GICRLM) displayed notably longer survival outcomes compared to Cluster 1 (AMECRLM) (Fig. 7B and C). Further examination of clinical indicators revealed that adverse prognostic features—including high CDX2 and Ki67 expression, predominant microsatellite stability (MSS), frequent KRAS mutations, and increased tumor deposits—were most prevalent in Cluster 1. Conversely, Cluster 3 was characterized by more favorable clinical profiles: Ki67 below 50%, a balanced distribution of MSI and MSS, rare KRAS mutations, normal levels of GGT, ALT, Tbil, and globulin, and minimal tumor deposits. These distinct clinicopathological features are perfectly consistent with the observed differences in patient outcomes, firmly supporting the prognostic validity of our marker-based stratification (Fig. 7D and E). Finally, to further externally validate the core marker of our most aggressive subtype, we conducted a single-gene logistic regression analysis of HSPA1A expression in the TCGA primary CRC (COAD/READ) reference cohort. High HSPA1A expression was significantly associated with advanced TNM stage and higher pathological grade (P < 0.01) (Table S2), reinforcing its value as a robust indicator of the aggressive AMECRLM trajectory. Fig. 7Survival curve and clinical information of CRLM cohort. (A) Immunohistochemistry of three CRLM subtype markers. (B) Kaplan–Meier overall survival (OS) curves of the CRLM cohort stratified into three molecular subgroups: Subgroup 1 (blue), Subgroup 2 (red), and Subgroup 3 (green) (n = 45). (C) Kaplan–Meier relapse-free survival (RFS) curves of the CRLM cohort categorized into the same three subgroups (n = 45). (D) Heatmap illustrating clinical characteristics of the CRLM cohort alongside corresponding subgroup classification based on marker gene expression. (E) Clinical feature heatmap of the CRLM cohort organized by molecular subgroups Survival curve and clinical information of CRLM cohort. (A) Immunohistochemistry of three CRLM subtype markers. (B) Kaplan–Meier overall survival (OS) curves of the CRLM cohort stratified into three molecular subgroups: Subgroup 1 (blue), Subgroup 2 (red), and Subgroup 3 (green) (n = 45). (C) Kaplan–Meier relapse-free survival (RFS) curves of the CRLM cohort categorized into the same three subgroups (n = 45). (D) Heatmap illustrating clinical characteristics of the CRLM cohort alongside corresponding subgroup classification based on marker gene expression. (E) Clinical feature heatmap of the CRLM cohort organized by molecular subgroups Biological validation of AMECRLM molecular subtype Based on the MDSCRLM molecular classification system, we conducted knockout (KO) validation and downstream functional analyses on the poorest prognosis AMECRLM subtype defining gene HSPA1A. Using the CRISPR-Cas9 system, we knocked out the HSPA1A gene in the murine colorectal cancer CT26 cell line and verified the knockout efficiency at the protein level via Western blot (WB) analysis, the result showed a marked reduction in HSPA1A expression in KO cells compared to wild-type controls (Fig. 8A). Following KO validation, we assessed the effects on cell migration and invasion capabilities. Wound healing assays revealed a significant reduction in the migration ability of KO cells after 24 h (P < 0.001) (Fig. 8B). Consistently, Transwell and Matrigel invasion assays revealed significantly reduced invasion and migration abilities in HSPA1A-KO cells (P < 0.001) (Fig. 8C). To comprehensively elucidate the downstream molecular networks governed by HSPA1A, total RNA was extracted from CT26-sgNC and CT26-sgHSPA1A#2 cell lines for whole-transcriptome sequencing. Global transcriptomic profiling revealed that the depletion of HSPA1A induced profound gene expression shifts. As illustrated in the volcano plot (Figure S12A), a substantial number of differentially expressed genes (DEGs) were identified. Furthermore, the DEG heatmap (Figure S12B) clearly demonstrates a distinct transcriptomic segregation between the two groups. Crucially, functional annotation of these DEGs revealed a marked downregulation of transcripts associated with aggressive tumor phenotypes and cellular invasion. KEGG analysis revealed significant enrichment of the JAK-STAT, MAPK, and Rap1 signaling pathways (Fig. 8D). Gene set enrichment analysis (GSEA) further confirmed marked dysregulation of the MAPK signaling pathway in HSPA1A-deficient cells (Fig. 8E). Western blotting of MAPK pathway markers revealed significant expression differences between CT26-sgNC and CT26-sgHSPA1A#2 cells, including E-cadherin, p-ERK, p-JNK, and p-P38 (Fig. 8F). We next investigated the impact of HSPA1A knockout on tumor invasion and metastasis in vivo. Lung metastasis efficiency was evaluated via tail vein injection of CT26-sgNC and CT26-sgHSPA1A#2 cells. In vivo imaging, gross examination, H&E staining, and VEGFA IHC confirmed a marked reduction in lung metastasis in the KO group (Fig. 8G, H). Liver metastasis efficiency was evaluated via spleen injection of CT26-sgNC and CT26-sgHSPA1A#2 cells. In vivo imaging, gross observation, H&E staining, and VEGFA IHC demonstrated significantly reduced liver metastasis in KO cells (Fig. 8I, J). Collectively, these results establish HSPA1A as a critical driver of metastasis and aggressive behavior in the AMECRLM subtype, providing both mechanistic insight and a potential therapeutic target for this high-risk group. Fig. 8Biological function validation of AMECRLM. (A) Western blot validation of HSPA1A knockout efficiency in CT26 cell lines stably transduced with non-targeting control (CT26-sgNC) or HSPA1A-targeting single guide RNAs (CT26-sgHSPA1A#1 and CT26-sgHSPA1A#2). (B) Wound healing (scratch) assay comparing the migratory capacity of CT26-sgNC, CT26-sgHSPA1A#1, and CT26-sgHSPA1A#2 cells. (C) Migration and Matrigel invasion assays of CT26-sgNC, CT26-sgHSPA1A#1, and CT26-sgHSPA1A#2 cells. (D) KEGG pathway enrichment analysis comparing bulk RNA-seq gene expression profiles between the HSPA1A-knockout and control (CT26-sgNC) cells. (E) Gene Set Enrichment Analysis (GSEA) of transcriptomic differences between the knockout and control groups, highlighting the inhibition of the MAPK signaling pathway. (F) Western blot detection of downstream MAPK signaling and EMT pathway protein expressions (E-cadherin, N-cadherin, p-ERK, ERK, p-JNK, JNK, p-P38, P38) in CT26-sgNC, CT26-sgHSPA1A#1, and CT26-sgHSPA1A#2 cells. (Note: Panels A and F represent protein lysates extracted from independent biological replicates to account for slight batch-to-batch variations in stable knockout pools). (G) In vivo lung metastasis assay in BALB/c mice tail-vein injected with CT26-sgNC, CT26-sgHSPA1A#1, or CT26-sgHSPA1A#2 cells, including representative bioluminescent imaging and gross lung morphology (Scale bar = 10 mm). (H) Histological evaluation of lung metastasis in BALB/c mice, including H&E staining and immunohistochemical analysis of VEGFA and HSPA1A. (I) In vivo liver metastasis assay in BALB/c mice spleen-injected with CT26-sgNC, CT26-sgHSPA1A#1, or CT26-sgHSPA1A#2 cells, including representative bioluminescent imaging and gross liver morphology (Scale bar = 10 mm). (J) Histological evaluation of liver metastasis in BALB/c mice, including H&E staining and immunohistochemical analysis of VEGFA and HSPA1A Biological function validation of AMECRLM. (A) Western blot validation of HSPA1A knockout efficiency in CT26 cell lines stably transduced with non-targeting control (CT26-sgNC) or HSPA1A-targeting single guide RNAs (CT26-sgHSPA1A#1 and CT26-sgHSPA1A#2). (B) Wound healing (scratch) assay comparing the migratory capacity of CT26-sgNC, CT26-sgHSPA1A#1, and CT26-sgHSPA1A#2 cells. (C) Migration and Matrigel invasion assays of CT26-sgNC, CT26-sgHSPA1A#1, and CT26-sgHSPA1A#2 cells. (D) KEGG pathway enrichment analysis comparing bulk RNA-seq gene expression profiles between the HSPA1A-knockout and control (CT26-sgNC) cells. (E) Gene Set Enrichment Analysis (GSEA) of transcriptomic differences between the knockout and control groups, highlighting the inhibition of the MAPK signaling pathway. (F) Western blot detection of downstream MAPK signaling and EMT pathway protein expressions (E-cadherin, N-cadherin, p-ERK, ERK, p-JNK, JNK, p-P38, P38) in CT26-sgNC, CT26-sgHSPA1A#1, and CT26-sgHSPA1A#2 cells. (Note: Panels A and F represent protein lysates extracted from independent biological replicates to account for slight batch-to-batch variations in stable knockout pools). (G) In vivo lung metastasis assay in BALB/c mice tail-vein injected with CT26-sgNC, CT26-sgHSPA1A#1, or CT26-sgHSPA1A#2 cells, including representative bioluminescent imaging and gross lung morphology (Scale bar = 10 mm). (H) Histological evaluation of lung metastasis in BALB/c mice, including H&E staining and immunohistochemical analysis of VEGFA and HSPA1A. (I) In vivo liver metastasis assay in BALB/c mice spleen-injected with CT26-sgNC, CT26-sgHSPA1A#1, or CT26-sgHSPA1A#2 cells, including representative bioluminescent imaging and gross liver morphology (Scale bar = 10 mm). (J) Histological evaluation of liver metastasis in BALB/c mice, including H&E staining and immunohistochemical analysis of VEGFA and HSPA1A Discussion Most previous research efforts have focused on the genetic, epigenetic, and transcriptomic alterations that occur in early-stage colorectal cancer (CRC). These studies enabled the development of precise therapeutic interventions that effectively reduce mortality in early-stage CRC. However, few studies comparatively investigated stage IV CRC patients, particularly colorectal liver metastases (CRLM). To address this gap, this work uses the single-cell atlas and multi-omics data analysis changes of colon cancer liver metastases to identify potential diagnostic and therapeutic targets. The heterogeneity of CRLM, driven by divergent developmental trajectories of epithelial cells, makes the current treatment choice difficult. Although different epithelial cell states driven by different developmental fates are regarded as a major barrier to improve CRLM outcomes, so far, there is no FDA-approved treatment targeting these cellular states. To address the lack of an effective stratification system to reflect the different levels of molecular differentiation spectrum, consensus-based clustering analysis on TCGA primary CRC reference cohort is performed. We developed a new CRLM stratification model based on the results of high-throughput sequencing. Crucially, by deriving an epithelial cell fate signature from a CRLM single-cell atlas and projecting it onto the TCGA primary CRC (COAD/READ) reference cohort, our approach identifies pre-existing metastatic potential within the primary tumor. Using the epithelial cell fate gene set, we classified CRLM patients into three distinct metastatic risk trajectories. The markers of each subgroup were different, the marker of AMECRLM mainly included HSPA1A and ALDH1A3, the marker of CCACRLM mainly included PTTG1 and CCNB1, and the marker of GICRLM mainly included RPL4 and CYP2W1.These subtypes exhibited significant prognostic differences: GICRLM had the most favorable outcomes, followed by CCACRLM, while AMECRLM was associated with the poorest prognosis. Therefore, rather than suggesting that metastatic fate is definitively predetermined, our findings strongly support the “malignant-seed” hypothesis. By delineating gene expression changes along the malignancy continuum, this model provides a robust framework for the early-risk stratification of CRC patients and highlights promising stage-specific therapeutic vulnerabilities. Type I: HSPA1A /ALDH1A3 mutant CRLM (aggressive-metastatic-enhanced type CRLM) Additionally, the high expression of genes such as HSPA1A and ALDH1A3 in this cluster suggests enhanced invasiveness and metastasis potential. HSPA1A (Heat Shock Protein Family A Member 1 A), also known as Hsp70-1 or Hsp72, is a key member of the heat shock protein 70 (HSP70) family. It functions as a molecular chaperone and plays a crucial role in protein folding, stabilization, and degradation within the cell. Primarily localized in the cytoplasm and nucleus, HSPA1A participates in various biological processes, including protein quality control, cellular stress response, and DNA repair [46]. Wang X et al. reported that HSPA1A knockdown selectively suppressed the malignant progression of Arid2-deficient lung cancer [47]. He Y et al. discovered that the ILC2s in advanced HCC exhibited high expression levels of the heat shock protein 70 (HSP70) gene, including heat shock protein family A member 1 A (HSPA1A), and heat shock protein family A member 1B (HSPA1B). In ILC2, targeting ICOS and its downstream effector HSP70 can change the immunosuppressive tumor microenvironment and stop tumor growth [48]. Wu FH et al. discovered that extracellular HSPA1A, an endogenous ligand of TLR2 and TLR4, promotes tumor cell proliferation and resistance to apoptosis, hence promoting the progression of hepatocellular carcinoma [49]. Building on these findings, Wang B et al. also conducted a thorough investigation into the significant impact of the heat shock protein HSP70 family in the development and occurrence of hepatocellular carcinoma. High HSPA1A expression was negatively correlated with overall survival, tumor grade, and cancer stage in patients [50]. In summary, HSPA1A acts as a tumor promoter in the development and progression of various cancers, particularly influencing metastasis and invasion in HCC. These findings provide valuable insights into our investigation of the AMECRLM subtype. Crucially, our study extends these findings to the context of CRLM. We provide direct experimental evidence that HSPA1A serves as a functional driver of the AMECRLM subtype’s aggressive phenotype. We further performed downstream functional biological validation for the AMECRLM subtype and found that knockout of HSPA1A significantly suppressed the invasive and metastatic capabilities of tumor cells, while also inhibiting the activation of the MAPK signaling pathway. It is noteworthy that this downstream suppression of MAPK signaling (e.g., p-JNK) and EMT markers (e.g., N-cadherin) upon HSPA1A knockout operates primarily through a threshold effect rather than a strict linear correlation with residual HSPA1A protein levels. Once the chaperone activity of HSPA1A falls below a critical functional threshold, the stability of the downstream signaling cascade is severely compromised. This disruption leads to a robust and consistent collapse of the invasive phenotype, regardless of minor biochemical fluctuations between independent stable knockout cell pools, highlighting the essential bottleneck role of HSPA1A in driving the AMECRLM malignant state. Furthermore, in vivo experiments using mouse models of metastatic tumors further confirmed the critical role of HSPA1A in CRLM tumor metastasis, highlighting the clinical relevance and importance of our molecular subtyping. Similarly, ALDH1A3 is closely associated with cancer stem cell (CSC) properties, leading to chemotherapy resistance and poor prognosis [51]. ALDH1A3 has been extensively studied in cancer research. It is widely recognized as a marker of cancer stem cells and is associated with poor prognosis in various cancers, including breast cancer, pancreatic cancer, and glioblastoma [52, 53]. ALDH1A3 is typically overexpressed in cancers and promotes tumor growth, metastasis, and chemotherapy resistance by altering gene expression, cell signaling pathways, and glucose metabolism [54]. Bharadwaj AG et al. reported that ALDH1A3 regulates the plasminogen activation pathway, enhancing levels and activities of tissue plasminogen activator (tPA) and urokinase plasminogen activator (uPA), thereby promoting TNBC cell invasiveness and breast cancer progression [52]. Nie S et al. found that ALDH1A3 functionally promotes the metastasis of pancreatic ductal adenocarcinoma (PDAC) by activating the PI3K/AKT/mTOR signaling pathway and its downstream target PPARγ, thereby impacting glucose metabolism [55]. Chen Z et al. identified USP9X as a critical deubiquitinase responsible for ALDH1A3 protein stability in glioblastoma (GBM) stem cells (GSCs). Treatment with the USP9X inhibitor WP1130 induces ALDH1A3 degradation and exerts significant therapeutic effects in a MES-type GSC-derived orthotopic xenograft model [56]. These studies suggest that inhibiting ALDH1A3 can reduce the CSC-like traits in cancer cells, making them more susceptible to conventional therapies. It is also associated with epithelial-mesenchymal transition (EMT), a key process in cancer metastasis [57]. Taken together, the co-enrichment of HSPA1A (stress tolerance/metastasis driver) and ALDH1A3 (stemness/chemoresistance) provides a molecular explanation for the poor prognosis observed in the AMECRLM subtype. Specifically, HSPA1A indicates a state of enhanced “invasion-metastasis” capability. The identification of this subtype has significant clinical implications, suggesting that patients classified as AMECRLM may benefit from targeted therapies inhibiting the HSP70 family or ALDH1A3-mediated metabolic pathways, rather than standard chemotherapy alone. Type II: CCNB1/PTTG1 mutant CRLM (cell cycle-active type CRLM) This subgroup of CCACRLM subtype is characterized by hyperactivation of cell cycle-related pathways, particularly involving mutations in the CDK-RB-E2F axis. The E2F family plays a crucial role in regulating the cell cycle’s G1/S transition [58]. E2F transcriptional targets include cyclins, CDKs, checkpoint regulators, DNA repair and replication proteins, among others [59]. Elevated E2F expression and/or E2F target gene expression in tumors has been associated with poor prognosis in studies on liver and pancreatic cancer [60–62]. One reason is that oncogene-driven increased E2F activity impairs cell cycle commitment and produces aberrant cell proliferation. Additionally, prolonged E2F activity during the S phase permits cells to recover from replication stress, creating increased replication stress and decreasing genomic stability in oncogenic contexts [62, 63]. The CCNB1 gene encodes cyclin B1, a mitosis-involved regulatory protein similar to the key regulatory protein governing the G2/M transition. Upon activation, it predominantly forms maturation-promoting factor (MPF) with p34 (Cdk1) [64]. In gastric cancer and pancreatic cancer, it has been discovered that the overexpression of CCNB1 promotes tumor proliferation and metastasis via the HnRNPR-CCNB1/CENPF axis and the TP53 signaling pathway, respectively [65, 66]. PTTG1 encodes a homolog of the yeast securin protein, which inhibits separase’s ability to promote sister chromatid segregation. It is a substrate for the anaphase-promoting complex (APC) and binds to separase until APC activation. The gene product possesses in vitro transforming activity and in vivo tumorigenic activity and is highly expressed in a variety of malignancies. Similarly, it has been reported that in breast cancer and gliomas, PTTG1 can enhance tumor growth and metastasis by targeting CCNB2 and stabilizing E2F1 [67, 68]. PTTG1 has also been identified as a CRLM staging-associated gene in colorectal tumors [69]. Thus, the marker genes CCNB1 and PTTG1 play critical roles in driving cell cycle activation and proliferation in CCACRLM subtype. This suggests that patients within the CCACRLM subgroup may be particularly sensitive to therapeutic agents targeting the cell cycle machinery, such as CDK inhibitors. Type III: RPL4/CYP2W1 mutant CRLM (growth-inhibited CRLM) Key genes such as RPL4 and CYP2W1 were significantly expressed in this group. RPL4 (ribosomal protein L4) has been reported to inhibit MDM2-mediated TP53 degradation, leading to TP53 stabilization and activation, thereby inhibiting tumor growth and proliferation [70]. CYP2W1 is involved in the metabolic activation of anticancer drugs and contributes to the chemosensitivity of cancer cells [71]. In GICRLM, RPL4 was significantly overexpressed, which was significantly related to the favorable prognosis of patients. Moreover, within the tumor microenvironment of colorectal liver metastases, RPL4 may exert its beneficial effect via the MDM2–TP53 pathway. RPL4 directly binds to the central acidic domain of MDM2, which inhibits the ubiquitination and degradation of TP53 mediated by MDM2, resulting in the stabilization and activation of TP53 to reduce tumor proliferation [72]. Compared with the liver function index and common tumor index, RPL4 can better reflect the good prognosis of growth-inhibitory CRLM and provide guidance for the follow-up treatment of CRLM. Our study established a novel risk stratification framework delineating distinct metastatic risk trajectories that exhibit significant differences in prognosis and biological characteristics. Similarly, the consensus molecular subtypes (CMS) of colorectal cancer are also categorized into four subtypes: CMS1 (MSI Immune): characterized by high immune infiltration and activity, significant microsatellite instability (MSI), and robust immune response. This subtype has a good prognosis. CMS2 (Canonical): characterized by significant activation of WNT and MYC signaling pathways, exhibiting classical epithelial cell features and high chromosomal stability. CMS3 (Metabolic): characterized by metabolic reprogramming, with abnormalities in glucose metabolism and KRAS mutations. CMS4 (Mesenchymal): characterized by significant TGF-β signaling and epithelial-mesenchymal transition (EMT) features, along with significant stromal components and inflammatory response. In parallel, by projecting our CRLM-derived epithelial cell fate signature onto primary CRC cohorts, we stratified patients into three distinct metastatic risk trajectories: AMECRLM: with main markers including HSPA1A and ALDH1A3, having the worst prognosis. CCACRLM: with main markers including PTTG1 and CCNB1, having an intermediate prognosis. GICRLM: with main markers including RPL4 and CYP2W1, having the best prognosis. When compared with the consensus molecular subtypes (CMS) of primary colorectal cancer, AMECRLM shares features with CMS4, with poor prognosis and invasive characteristics; CCACRLM exhibits overlapping characteristics both CMS2 and CMS3 subtypes, exhibiting cell cycle activity and metabolic characteristics; GICRLM is similar to the CMS1 subtype, with a good prognosis and growth-inhibitory characteristics. These comparisons further emphasize the importance of individualized treatment based on molecular subtyping in the management of CRLM. Subsequent immunohistochemical analysis and external validation using clinical data revealed that the molecular markers identified in Cluster 1 and Cluster 3 exhibited high accuracy and significant clinical relevance. The current molecular diagnostic markers for CRLM mostly relied on non-specific indicators such as CEA, CA19-9, CA72-4, CA125, etc [73]., and specific cell cycle proteins such as CDX2, Ki67, MSI, KRAS, etc [74]. At present, for 50% of metastatic CRC patients, the KRAS/NRAS/BRAF wild-type monoclonal antibody, combined with chemotherapy, can prolong the median survival time by 2 to 4 months compared with chemotherapy alone. However, for the 35% to 40% of patients with KRAS or NRAS sequence mutations, effective targeted therapy was not yet available [75]. This underscores the urgent need for more precise molecular subtyping in colorectal cancer. At the same time, we did immunochemical staining on clinical samples according to the marker of three subtypes and used the clinical information of patients to do external verification to verify the correlation between molecular subtypes and clinical prognosis, so as to establish medical credibility. In view of the heterogeneity of CRLM, we provide valuable molecular markers and therapeutic targets for different molecular subtypes of CRLM. The primary innovation of our research lies in integrating single-cell transcriptomic states with multi-omics primary tumor data to prospectively identify pre-existing metastatic risk profiles. This comprehensive approach allows for a more detailed understanding of the tumor microenvironment and the identification of key regulatory pathways and genes. The different targets corresponding to different predictive models of CRLM also explain their different prognosis to some extent, paving the way for the development of new drugs in the future. For example, targeting HSPA1A in AMECRLM can inhibit metastatic progression, while modulating the activity of RPL4 in GICRLM can enhance the efficacy of existing therapies. To validate this, we performed downstream functional studies based on HSPA1A, the key marker of AMECRLM. Specific knockout of HSPA1A reduced the invasion and metastasis efficiency of its tumors at both the cellular and animal levels. To sum up, this study provided a novel framework for single-cell sequencing combined with multi-group analysis of CRLM and screened out specific markers for distinct metastatic trajectories. Compared to conventional clinical indicators, these identified markers demonstrate superior predictive accuracy, strongly supporting the “malignant-seed” hypothesis and offering immense value for the early clinical management of colorectal cancer liver metastasis. Our study has several limitations. Firstly, the sample size and potential lack of diversity in the patient cohort may limit the generalizability of our findings, necessitating validation in larger and more diverse cohorts. Secondly, a significant limitation in the current field is the absence of a large-scale, publicly available cohort of CRLM tissue with matched multi-omics layers (e.g., genomics, epigenomics, and proteomics) comparable to TCGA. Consequently, our multi-omics integration and prognostic modeling primarily leveraged the TCGA primary CRC cohort as a reference to evaluate the relevance of the epithelial fate signature, rather than a dedicated metastatic multi-omics dataset. As previously noted, this approach relies on the computational projection of CRLM-derived transcriptional states into primary tumors. While this effectively facilitates early-risk stratification and robustly supports the “malignant-seed” hypothesis, it serves as an indirect predictive tool rather than definitive biological proof of a predetermined metastatic fate. Thirdly, the translational application of our findings to clinical practice needs further investigation, including the development of diagnostic tools and therapeutic agents targeting the identified molecular markers. Lastly, the integration of single-cell RNA sequencing and multi-omics data could introduce technical biases, making robust data normalization and validation steps essential to mitigate these biases. Collectively, while our study uncovers essential mechanisms and regulators of CRLM formation and identifies potential diagnostic and prognostic markers, further research is necessary to establish the clinical relevance of these findings. Supplementary Information Below is the link to the electronic supplementary material. Supplementary Material 1 Supplementary Material 1 Supplementary Material 2 Supplementary Material 2 Supplementary Material 3 Supplementary Material 3 Supplementary Material 4 Supplementary Material 4