A novel tumor cell origin-based risk score (TCOMRS) to predict liver metastasis and guide treatment in colon adenocarcinoma

Abstract

Background

Liver metastasis (LM) is a major determinant of poor prognosis in colon adenocarcinoma (COAD), yet reliable biomarkers for identifying patients at high metastatic risk remain limited. This study aimed to develop and validate a tumor cell origin-based metastasis risk score (TCOMRS) for characterizing LM-associated transcriptional features in COAD and to identify key molecular regulators involved in liver metastatic progression.

Methods

Single-cell RNA sequencing (scRNA-seq) profiles from primary tumors and LM were integrated to identify distinct malignant-cell subpopulations and delineate their cellular origins. TCOMRS was constructed based on a one-class logistic regression (OCLR) algorithm trained with 4,512 malignant cells derived from LM, enabling assessment of the similarity between primary tumor cells and LM-associated malignant phenotypes. The clinical relevance of TCOMRS was validated across multiple independent cohorts, including The Cancer Genome Atlas (TCGA)-COAD, GSE39582, GSE38832, and GSE161158.

Results

Elevated TCOMRS was consistently associated with unfavorable overall survival, increased tumor stemness, and the presence of LM. Moreover, patients with low TCOMRS demonstrated improved responsiveness to immune checkpoint blockade therapy, suggesting potential utility in therapeutic decision-making. Mechanistically, fatty acid-binding protein 1 (FABP1) was identified as a key driver of LM, facilitating epithelial-mesenchymal transition and metastatic colonization. FABP1 expression correlated strongly with TCOMRS and metastatic propensity.

Conclusions

Our findings establish TCOMRS as a robust biomarker for risk stratification and FABP1 as a promising therapeutic target, offering valuable insights for personalized treatment strategies in COAD.

Highlight box.

Key findings

• Elevated tumor cell origin-based metastasis risk score (TCOMRS) correlates with poor survival, high stemness and higher liver metastasis risk.

• Low TCOMRS is associated with a more immune-active microenvironment and higher predicted responsiveness to immune checkpoint blockade.

• Fatty acid-binding protein 1 (FABP1) is identified as a key driver of liver metastasis by facilitating epithelial–mesenchymal transition and metastatic colonization, and is strongly correlated with TCOMRS and metastatic potential.

What is known and what is new?

• Liver metastasis remains the leading cause of mortality in colon adenocarcinoma (COAD), yet current clinical tools provide limited predictive power for identifying patients at high risk before metastasis occurs.

• TCOMRS is a newly developed cell origin-based scoring system for metastatic risk assessment. TCOMRS is associated with prognosis and liver metastasis risk and may help identify immunotherapy-related phenotypes. FABP1 is newly identified as a core driver of COAD liver metastasis.

What is the implication, and what should change now?

• TCOMRS fills the shortage of cellular-level biomarkers for COAD liver metastasis. FABP1 provides a promising target for anti-metastatic therapy. Findings promote precision and individualized management of COAD.

• TCOMRS warrants further prospective evaluation as a biomarker for early metastatic risk stratification.

• Therapy regimens should be tailored according to TCOMRS stratification.

• FABP1 warrants further investigation as a potential anti-metastatic therapeutic target.

Introduction

Advanced colon adenocarcinoma (COAD) has poor survival outcomes due to cancer stem cells (CSCs), which contribute to drug resistance, high recurrence rates, and metastasis in COAD (1). Additionally, CSCs are inherently more resistant to conventional chemotherapy, such as 5-fluorouracil (5-FU) and oxaliplatin (2,3). The liver is the most frequent site of metastasis in COAD, with approximately 50% developing liver metastases (LMs) (4-6). Despite the continuous advance of therapies, COAD with LM remains an intractable problem (7). In addition, the overall quantity of LM exerts a substantial influence on the prognosis. Therefore, further exploration of the molecular changes and the capacity to interdict LM is vital for developing effective treatments for COAD. In this context, the identification of metastasis risk and evaluation of prognosis are the first steps and would guide decision-making in subsequent examinations and the clinical management of COAD.

Over the past few years, numerous studies have developed predictors based on genomic signatures, including gene expression and somatic mutation profiles, within the primary tumor (PT) bulk to identify distant metastatic events and prognosis (8,9). However, few models have been established to quantitatively determine the LM tendency in patients with COAD. It is well known that tumor metastasis, especially occult metastases, is not easily detectable. Thus, conventional predictors trained on samples with under-detected metastatic instances may exhibit an under-estimated metastasis potential (10). Additionally, tumor cells usually comprise heterogeneous populations of subclones, each containing distinct features, making it difficult to identify specific cell types within unknown heterogeneous cell populations by using a single optimization strategy (11).

The one-class logistic regression (OCLR) machine-learning algorithm is a one-class analytical framework, which has been applied to determine independent stemness index for multiple malignancies and neuroendocrine differentiation in non-small cell lung cancer (NSCLC) based on transcriptomic data, as it performs well in the absence of a defined “negative” metastatic class and enables the identification of specific cell types in heterogeneous samples (12,13). In the study, we used the OCLR machine-learning algorithm trained on single-cell transcriptomics of liver metastatic cells to build a new model, named tumor cell origin-based metastasis risk score (TCOMRS), for determining the metastasis tendency and prognosis in COAD patients whose PTs had not yet metastasized. Our findings provide a useful strategy for quantitatively determining the tendency of LM and identifying key LM-related regulators in COAD, offering novel insights for prognosis prediction and improving the personalized therapy of COAD. We present this article in accordance with the TRIPOD, ARRIVE, and MDAR reporting checklists (available at https://tcr.amegroups.com/article/view/10.21037/tcr-2026-1049/rc).

Methods

Acquisition of colon cancer cohorts and data processing

The present study collected COAD data from diverse cohorts from two individual platforms: The Cancer Genome Atlas (TCGA) and Gene Expression Omnibus (GEO). The single-cell RNA expression profiles of LM samples of COAD were obtained from GSE225857 cohort (14). TCGA-COAD mRNA expression profiles, clinical data, proteomic information, and somatic mutation profiles were obtained from the TCGA data portal (https://portal.gdc.cancer.gov/). In addition, we retrieved and obtained COAD datasets including the GSE39582 cohort (15), GSE38832 cohort (16), GSE161158 cohort (17), GSE72970 cohort (18) and GSE104645 cohort (19), GSE6988 (20), GSE110009 (21), and GSE190609 (22). The public datasets included in this study are summarized in Table S1. The public datasets included in this study are summarized in Table S1. To reduce cross-platform systematic bias, RNA-seq transcripts per million (TPM) values were converted to log2(TPM + 1) using a pseudo-count of 1, and microarray datasets were log2-transformed followed by normalization to a 0–1 scale. Normalization was performed using the formula X’ = (X − Xmin)/(Xmax − Xmin), in which X represents the processed microarray expression value (23).

Processing of single-cell RNA sequencing (scRNA-seq) data and cell type annotations

Seurat was utilized for scRNA-seq data analysis and visualization (24). Initially, 47,892 cells from PT samples were obtained (14). Seurat v4.1.3 package in R 4.2.2 was applied for data processing and visualization. Based on the highly variable genes, principal component (PC) analysis was performed to determine the significant dimensions, termed PCs. Cell clustering and t-distributed stochastic neighbor embedding (t-SNE) algorithms were used for dimensionality reduction, with the top 20 initial PCs and cluster classifications across all cells. The cell clusters were annotated to three cell types using canonical marker genes, including tumor cells (EPCAM and SOX9), fibroblasts (COL1A1 and COL1A2), and endothelial cells (PECAM1 and CD34), as described previously (14). Marker genes for each cell cluster were identified using the FindAllMarkers function, with significance defined by an adjusted P value <0.05 and |log2 log[fold change (FC)]| >1.0 as the threshold. The cell clusters were automatically classified with the “SingleR” package based on reference scRNA-seq datasets, followed by manual verification and correction using canonical marker genes retrieved from the CellMarker database (25). The Monocle package was employed to construct a pseudotime trajectory based on the gene expression profiles (GEPs) of MC2 and MC3 (26). After dimensionality reduction and cell ordering, cells were projected onto a branched trajectory, and cells within the same branch were assigned to the same cellular state. MC2 was selected as the LM-associated malignant reference population because it contained both PT- and LM-derived malignant cells, displayed a shared PT-LM pseudotime trajectory, and showed stronger enrichment of epithelial-mesenchymal transition-related pathways than MC3.

Construction, calculation, and validation of TCOMRS

Rationale for the OCLR framework

To quantify the LM-like malignant transcriptional state in bulk COAD samples, we constructed a TCOMRS using a OCLR framework. Unlike conventional supervised classifiers that require clearly defined positive and negative training classes, OCLR learns a transcriptional weight vector from a well-defined reference population. This framework was selected because LM-derived malignant cells provided a biologically defined positive reference state, whereas a reliable negative metastatic class was difficult to define owing to intratumoral heterogeneity, intermediate malignant states, and the lack of sufficiently large cohorts with consistently annotated LM status. Therefore, LM-derived malignant cells within MC2 were used as the reference population to derive an LM-like transcriptional signature.

Feature selection and OCLR model training

Before model construction, feature genes were selected from LM-derived MC2 malignant cells. Genes with log2(TPM + 1) >0.1 in more than 95% of the reference cells were retained to exclude genes with sparse or highly inconsistent expression in the LM-associated malignant population. A total of 9,014 genes met this criterion and were included in model training. The expression matrix of the selected genes was mean-centered before model fitting. OCLR was implemented using the R package “gelnet” as previously described (27). In total, 4,512 LM-derived MC2 malignant cells were used to train the OCLR model. The resulting gene weight vector was considered to represent the LM-like malignant transcriptional signature.

Calculation of TCOMRS in bulk transcriptomic cohorts

For each bulk transcriptomic sample, genes shared by the bulk expression matrix and the OCLR weight vector were retained. TCOMRS was calculated as the Spearman correlation between the bulk expression profile and the OCLR-derived weight vector (Table S2), followed by linear normalization to a 0–1 scale. A higher TCOMRS indicated stronger transcriptomic resemblance to LM-derived MC2 malignant cells.

Clinical and external validation of TCOMRS

The clinical and biological relevance of TCOMRS was evaluated through several complementary analyses. The association between TCOMRS and metastatic behavior was evaluated in cohorts with available metastasis annotations. In TCGA-COAD and GSE190609, the distribution of metastatic samples was compared between the high- and low-TCOMRS groups. TCOMRS values were also compared between primary and metastatic tumor samples in GSE110009. Because these cohorts lacked organ-specific metastatic information, these analyses were considered evaluations of general metastatic behavior. The association between TCOMRS and LM was further assessed in GSE6988, an independent cohort with explicitly annotated LM status, by comparing TCOMRS values between patients with and without documented LM. In the TCGA-COAD cohort, the associations between TCOMRS and clinicopathological characteristics, including age, gender, and TNM classification, were assessed. In addition, the relationships of TCOMRS with consensus molecular subtype (CMS) and the mRNA-based stemness index (mRNAsi) were examined to evaluate its biological consistency with molecular features related to tumor aggressiveness and stem cell-like properties.

The prognostic relevance of TCOMRS was initially assessed in the TCGA-COAD cohort and subsequently evaluated in three independent external cohorts (GSE39582, GSE38832, and GSE161158) using survival analyses. To compare its prognostic performance with conventional supervised approaches, least absolute shrinkage and selection operator (LASSO)-Cox and survival support vector machine (survival-SVM) models were trained in the TCGA-COAD cohort using the same 9,014-gene feature space and then evaluated in the three external cohorts. Model discrimination was quantified using the concordance index (C-index).

The workflow of TCOMRS construction and validation is shown in Figure S1.

Gene set enrichment analysis (GSEA)

To characterize biological pathways associated with TCOMRS, GSEA using the Hallmark gene sets was performed to compare high- and low-TCOMRS tumors in the TCGA-COAD and GSE39582 cohorts. Epithelial-mesenchymal transition (EMT) pathway activity were further evaluated in relation to TCOMRS.

Immune microenvironment and immunotherapy-related analyses

Immune-cell and immune-function signatures were evaluated by single-sample GSEA (ssGSEA) in TCGA-COAD and GSE39582, and stromal, immune, and ESTIMATE scores were calculated using the ESTIMATE algorithm. Immune checkpoint expression and programmed death-ligand 1 (PD-L1)/CD274 were evaluated in relation to TCOMRS. SubMap analysis using a published melanoma immunotherapy dataset was used to estimate anti-programmed cell death 1 (PD-1) response in high microsatellite instability (MSI-H) COAD subgroups, and T-cell-inflamed GEP and innate anti-PD-1 resistance signature (IPRES) were assessed as immunotherapy-related signatures. Microenvironment cell populations (MCP)-counter was additionally applied to GSE39582 to independently estimate immune and stromal cell abundance.

Cell lines and culture

Human cell lines HT29, LOVO and HEK 293T were procured from China Center for Type Culture Collection and authenticated by STR profiling. Human cell lines HT29, LOVO and HEK 293T were sourced from China Center for Type Culture Collection and authenticated by STR profiling. HT29 and LOVO cells were cultured in Roswell Park Memorial Institute (RPMI)-1640 medium (Hyclone, SH30027.01, USA), and HEK 293T cells were grown in Dulbecco’s modified Eagle’s medium (DMEM)/high glucose medium (Hyclone, SH30243.01, USA), all of which were supplemented with 10% fetal bovine serum (FBS) (Natocor, SFBE-B1, Argentina) and 1% penicillin-streptomycin solution (Biological Industries, 03-031-1B, Israel). The three cell lines were incubated at 37 ℃ with 5% CO2.

Lentivirus transduction

To stably overexpress or knock out FABP1 in HT29 and LOVO cell lines, HEK 293T cells were transfected with a three-plasmid system (pLV3-CMV-FABP1-3×FLAG-CopGFP-Puro/ pLV3-U6-FABP1-sgRNA-Cas9-EGFP-Puro, psPAX2, pMD.2G) using Lipofectamine 8000 transfection reagent (Beyotime, C0533, CN). The supernatant was harvested at 48 h after transfection, which was added into HT29 and LOVO cells with polybrene. Puromycin was then added to cells with appropriate concentration to obtain cells stably expressing FABP1 or knock out FABP1.

Quantitative real-time polymerase chain reaction (qRT-PCR)

Total RNA was extracted by TRIzol reagent (Invitrogen, AM9738, USA), and reverse transcription was performed using ReverTra Ace qPCR RT kit (TOYOBO, FSQ-101, Japan) according to supplier’s instructions. qRT-PCR was carried out using Ultra SYBR mixture (Cwbio, CW0659, China) under the strict procedure. The primer sequences are as follows: FABP1, (forward) GTCCAAAGTGATCCAAAACGAA and (reverse) CGGTCACAGACTTGATGTTTTT; E-cadherin, (forward) GCCTCCTG AAAAGAGAGTGGAAG and (reverse) TGGCAGTGTCTCTCCAAATCCG; N-cadherin, (forward) CCTCCAGAGTTTACTGCCATGAC and (reverse) GTAGGAT CTCCGCCACTGATTC; Vimentin, (forward) AGGCAAAGCAGGAGTCCACTGA and (reverse)ATCTGGCGTTCCA GGGACTCAT.

Western blotting

Whole cell lysates were derived from cells lyzed by NP40 (Beyotime, P0013F, CN). Havested proteins were electrophoresed in sodium dodecyl sulfate–polyacrylamide gel electrophoresis (SDS-PAGE) gels and transferred to polyvinylidene difluoride (PVDF) membranes, which were incubated with the following primary antibodies at 4 ℃ overnight: FABP1 (Proteintech, 13626-1-AP, CN), glyceraldehyde-3-phosphate dehydrogenase (GAPDH) (Proteintech, 10494-1-AP, CN), Caspase 3 (Cell Signaling Technology, #9662, USA), Caspase 7 (Abcam, ab255818, UK).

Transwell migration assay

Cell migration capacity was evaluated by the 24-well transwell chambers (Labselect, CN). In brief, 100 μL serum-free medium containing 2×10⁴ cells were loaded in the upper chambers, and the lower chambers were filled with 600 μL culture medium with 30% FBS. The migrated cells on the lower surface of the transwell membrane were fixed with methanol and stained with crystal violet after incubating for 48 h. Digital images were captured under white light using an inverted microscope (Olympus, Japan) and analyzed by Image-J software.

Fatty acid fluorescent staining assay

Fatty acid labeling was performed using BODIPY 500/510 C1, C12 fluorescent probes (Beyotime, C2055, CN). Cells were seeded in a 24-well plate for 24 h and incubated with the fluorescent probe at a final concentration of 5 μM for 30 minutes, and then rinsed with PBS for three times. Afterwards, fixed the cells with 4% paraformaldehyde for 30 minutes and dyed with 4',6-diamidino-2-phenylindole (DAPI) (Servicebio, GDP1024, CN) for 10 minutes. Photographs were recorded by a fluorescence microscope.

Animal LM model establishment

The six-week-old female BALB/c nude mice were purchased from Vital River Laboratory Animal Technology Company (Beijing). HT29/Vector control and HT29/FABP1-OE cells were selected for the generation of a murine model of LM (n=5/cell line). Briefly, colon cancer cells were dissociated with trypsin and suspended in ice-cold PBS at 5 ×10⁷/mL. Mice were anesthetized with 1% sodium pentobarbital (50 mg/kg body weight), followed by a left abdominal laparotomy. Then, 20 μL PBS containing 1×10⁶ cells (per mouse) were injected slowly into the spleen. Since one mouse died in each group after the modeling process, the actual grouping was 4 mice per group. LM were examined macroscopically by counting visible tumor nodules for subsequent analysis and sample collection at the study endpoint (6 weeks after cell inoculation) for analysis and sampling. All mice were euthanized by cervical dislocation after anaesthetization with an intraperitoneal injection of 1% sodium pentobarbital (50 mg/kg body weight). Experiments were performed under a project license (No. ZN2023203) granted by the Experimental Animal Welfare and Ethics Committee of Zhongnan Hospital of Wuhan University, in compliance with institutional guidelines for the care and use of animals.

Statistical analysis

Student’s t-test was applied to compare two normally distributed groups, whereas the Mann-Whitney U test was used for comparisons involving non-normally distributed data. The paired t-test was performed for matched samples. For comparisons among more than two groups with normal distribution, one-way analysis of variance (ANOVA) was employed. Differential expression analysis was conducted using the R package “limma”, and Kaplan-Meier survival analysis was carried out with the R package “survival” based on the optimal cut-off value. All statistical analyses were performed with GraphPad Prism version 9.4.1 and R version 4.2.2, and a two-sided P value <0.05 was considered statistically significant.

Code availability

Data analysis and statistical processing were performed using open-source software within the R project framework (https://www.r-project.org/). Details of the major software tools and algorithms applied in this study are summarized in Table S3.

Results

Single-cell transcriptomics identified specific cancer cell clusters associated with LM of COAD

To identify the cell subpopulation driving the development and progression of LM in COAD, seurat package was applied for data processing and visualization of scRNA-seq profiles (GSE225857), a dataset containing the transcriptomes of cells from PTs and LM tumors from patients with COAD (14) (Figure 1A). Cells from the PT and LM samples were classified into 25 separate clusters using the t-SNE algorithm and divided into three cell types, fibroblasts, tumor cells and endothelial cells, after being annotated using canonical markers (25) (Figure 1B and Figure S2A,S2B). Furthermore, sub-clustering of the tumor cells was performed and 23,859 malignant cells were categorized into 10 subpopulations, termed malignant cell clusters (MCs) 0–9 (Figure 1C,1D). Notably, among all the malignant cell subpopulations, we found that only MC2 and MC3 were composed of cells derived from both PT and LM samples (Figure 1D). Thus, a trajectory analysis was performed to project MC2 and MC3 onto different branches based on pseudotime ordering. We observed that PT and LM tumor cells within MC2, rather than MC3, followed a similar differentiation trajectory and were involved in the transition processes of the four cellular states, indicating that LM tumor cells exhibited the common features of PT cells within MC2 during the dissemination of COAD (Figure 1E and Figure S2C-S2G). Moreover, we found that the EMT gene signature was enriched in MC2 compared with that in MC3 by performing gene set variation analysis using hallmark pathway gene signatures (Figure 1F), suggesting that EMT is a crucial and fundamental mechanism in the LM of COAD. Therefore, MC2 was selected for further investigation.

Figure 1.

Distinct cellular constitutions in COAD PT and LM samples delineated by scRNA-seq analysis. (A) The main organs, including PT and LM annotated in GSE225857 are illustrated by t-SNE plots. (B) The main cell types annotated by known cell lineages in PT and LM samples are illustrated by t-SNE plots. (C) The main organs, including PT and LM annotated in all malignant cells are illustrated by t-SNE plots. (D) The t-SNE plot of all malignant cells, color-coded by 10 MC clusters. (E) Trajectory analysis of MC2 from PTs (left) and LM (right) ordered and annotated by four cellular states. (F) GSVA demonstrated differences in hallmark pathways between MC2 and MC3 cells. COAD, colon adenocarcinoma; GSVA, gene set variation analysis; LM, liver metastasis; MC, malignant cell; PT, primary tumor; scRNA-seq, single-cell RNA sequencing; t-SNE, t-distributed stochastic neighbor embedding.

Construction and validation of a TCOMRS model for quantitative characterization of LM in COAD

Based on the single-cell transcriptomic profiles, we developed the TCOMRS to quantify the transcriptomic similarity of bulk tumors to LM-derived MC2 malignant cells. TCOMRS was constructed using a OCLR framework trained on 4,512 cells from LM samples, evaluating 9,014 genes (Table S2) highly expressed in MC2 [log2(TPM+1) >0.1 in >95% of cells]. TCOMRS values ranged from 0 to 1, with higher scores indicating greater similarity to the LM-associated malignant-cell transcriptional state.

To determine whether TCOMRS was associated with metastatic behavior, we first examined cohorts with available general metastatic annotations. In TCGA-COAD, the proportion of patients with metastatic disease was higher in the TCOMRS-high group than in the TCOMRS-low group (Figure S3A). Consistently, metastatic tumor samples exhibited significantly higher TCOMRS values than PT samples in GSE190609 (Figure S3B). Because these cohorts did not provide detailed information on metastatic sites, these findings support an association between elevated TCOMRS and general metastatic behavior but do not establish an association specifically with LM.

To further evaluate the metastasis-associated transcriptional phenotype captured by the OCLR-derived TCOMRS, we applied the pretrained model, without retraining, to the independent GSE110009 cohort containing primary colorectal tumor and metastatic lesion samples. TCOMRS was significantly higher in metastatic lesions than in PTs (Figure S3C), indicating that the transcriptional phenotype captured by TCOMRS was more strongly represented in metastatic lesions. However, the PT samples in this cohort served only as an external reference and were not incorporated as a negative population during model training. Moreover, because detailed organ-specific metastatic-site annotations were not consistently available, this result was interpreted as supporting an association with general metastatic behavior rather than liver-organ specificity.

To examine the association between TCOMRS and LM status more directly, we further evaluated TCOMRS in GSE6988, an independent cohort with explicitly annotated LM status. TCOMRS was significantly higher in patients with LM than in those without LM (P=0.006, Figure 2A), providing independent evidence that elevated TCOMRS is associated with LM in COAD.

Figure 2.

Development and validation of the TCOMRS. (A) Comparisons of TCOMRS values between with and without liver metastases samples in GSE6988. Data were analyzed using Student’s t-test. (B) An overview of the associations between well-known molecular and clinical signatures and the TCOMRS in TCGA-COAD. (C) The association of COAD molecular subtypes with the TCOMRS in TCGA-COAD. (D,E) The correlation between TCOMRS and Stemness index (mRNAsi) in samples of TCGA-COAD. Data were analyzed using Pearson correlation analysis. (F-I) Kaplan-Meier curves of OS according to the TCOMRS values of patients with COAD in the COAD cohorts. ***, P<0.001. COAD, colon adenocarcinoma; LM, liver metastasis; M, metastasis; N, node; ns, no significance; OS, overall survival; T, tumor; TCGA, The Cancer Genome Atlas; TCOMRS, tumor cell origin-based metastasis risk score.

We further assessed the relationship between TCOMRS and clinical as well as molecular features in TCGA-COAD. Notably, no significant association was observed between TCOMRS and sex, gender, TNM classification, or clinical stage in most patients with COAD (Figure 2B). In addition, we observed that TCOMRS showed strong associations with CMS2 and CMS4 (Figure 2C), two molecular subtypes of COAD that were previously defined to exhibit high malignant potential (28). Tumor metastasis is accompanied by progressive loss of differentiated characteristics and acquisition of stem cell-like properties (29). Accordingly, we evaluated the association between TCOMRS and the stemness index as previously described (12), and found a positive correlation between them (Figure 2D,2E), suggesting that tumors with higher metastatic potential also possess greater stem cell-like properties. As expected, we also found that a higher TCOMRS was associated with worse prognosis in the TCGA-COAD and GEO-COAD cohorts (Figure 2F-2I). These findings suggested that the TCOMRS model based on the OCLR algorithm trained on LM tumor cells of COAD could quantitatively evaluate the LM propensity of COAD at an early stage and has a high predictive value for the prognosis of patients with COAD.

Correlation of the mutational landscape and TCOMRS in COAD

Genetic variations play important roles in regulating tumor metastasis (30). We therefore compared the genomic profiles of the low and high LM risk subgroups defined by the median TCOMRS in the TCGA-COAD cohort using the R package “maftools” (31). As expected, the values of TCOMRS were negatively associated with tumor mutation burden (TMB) (Figure 3A). We found that 12 genes were mutated in more than 25% of the samples in the low LM risk subgroup, whereas only six genes met this criterion in the high LM risk subgroup (Figure 3B). Notably, ATC, TTN, TP53, and KRAS occupied the top four positions in both the subgroups. Enrichment analysis of these recurrently mutated genes revealed significant involvement of canonical oncogenic pathways, including RTK-RAS, Wnt/β-catenin, Notch, Hippo, PI3K/Akt, MYC, TGF-β, cell cycle, TP53, and NRF2 signaling (Figure 3C). These findings indicate that genomic mutation-driven oncogenic pathway dysregulation, particularly alterations in RTK-RAS signaling, may contribute to LM progression in COAD.

Figure 3.

Landscape of somatic mutations in COAD in low and high TCOMRS subgroups. (A) Correlation analysis between the TCOMRS and TMB in TCGA-COAD. Data were analyzed using Pearson correlation analysis. (B) The waterfall plot shows the mutation distribution of the top 20 most frequently mutated genes. The central panel shows the types of mutations in each COAD sample. The upper panel shows the mutation frequency of each COAD sample. The bar plots on the left and right sides show the frequency and mutation type of genes mutated in the low and high TCOMRS subgroups, respectively. (C) Oncogenic pathway analysis showing frequent genetic pathway alterations relative to TCOMRS in TCGA-COAD. (D,E) Comparisons of TCOMRS values between TP53-mutant and TP53-wild type samples, KRAS-mutant and KRAS-wild type samples as well as BRAF-mutant and BRAF-wild type samples in the TCGA and GSE39582-COAD cohort. Data were analyzed using Student’s t-test. (F) Comparisons of TCOMRS values between MSI-H and MSS/MSI-L samples in the COAD cohorts (TCGA and GSE39582). *P<0.05, **P<0.01, ***P<0.001. BRAF, B-Raf proto-oncogene, serine/threonine-protein kinase; COAD, colon adenocarcinoma; KRAS, kirsten rat sarcoma viral oncogene homolog; MSI, microsatellite instability; MSI-L, low MSI; MSI-H, high MSI; MSS, microsatellite stable; TCGA, The Cancer Genome Atlas; TCOMRS, tumor cell origin-based metastasis risk score; TMB, tumor mutation burden; TP53, tumor protein P53.

Previously, it has been shown that TP53, KRAS and BRAF are the most commonly occurring hotspot mutations and associated with metastasis of COAD (32). Therefore, we determined the correlation between these gene mutations and TCOMRS. We found that elevated TCOMRS was strongly linked to somatic alterations in the TP53, KRAS, and BRAF genes in TCGA-COAD and GEO39582 cohorts (Figure 3D,3E), indicating that dysfunction of these genes mediated by genetic variations might promote LM of COAD.

Furthermore, missense mutations accounted for the majority of genomic variants in both the low and high LM risk subgroups, representing approximately 85% of all alterations. Notably, the low LM risk subgroup harbored a significantly greater number of variants than the high LM risk subgroup (Figure S4A,S4B). Among single-nucleotide variants (SNVs), C>T was the most frequent substitution type, and all SNV events were enriched in tumors with low TCOMRS (Figure S4C,S4D). In addition, high TCOMRS was positively associated with transversion events, whereas a negative correlation was observed between TCOMRS and transition events across all SNVs (Figure S4E).

Moreover, co-occurring and mutually exclusive mutations among the top 25 most frequently mutated genes were analyzed using the “CoMEt” algorithm (33). Compared with the widespread pattern of co-occurrence, four gene pairs (MDN1-APC, KMT2D-TP53, BRAF-APC, and BRAF-KRAS) and five cases (TP53-ATM, TP53-DST, TP53-FBXW7, TP53-RYR2, and TP53-SYNE1) exhibited mutually exclusive mutations in the high and low TCOMRS subgroups, respectively (Figure S4F). These findings suggest that these gene pairs may function within the same pathways and exert redundant effects in regulating LM in COAD.

MSI status, which is likely related to high TMB, can serve as a predictive marker for prognosis of solid tumor (34). Approximately 15% of patients with COAD exhibited MSI-H and a better prognosis than patients with low MSI [microsatellite stable (MSS)/MSI-L] in COAD (35). Therefore, we evaluated the association between the TCOMRS and MSI status and found that MSS/MSI-L samples had a higher TCOMRS than that of MSI-H samples in the COAD cohort (Figure 3F).

Taken together, these results suggested that these mutations play a meaningful role in the progression of LM and the TCOMRS is highly correlated with TMB and MSI status, which may be related to the response to therapy in patients with COAD.

Association of the TCOMRS with immunotherapy response in MSI-H COAD

MSI-H COAD is linked to an immune-active tumor microenvironment, which has abundance of tumor-infiltrating lymphocytes and neoantigens, resulting in increased tumor immunogenicity and elicited antitumor immune response (36). To explore the association between TCOMRS and the tumor immune microenvironment, we performed ssGSEA to evaluate 29 immune-related pathways in the low- and high-TCOMRS subgroups (37). Higher TCOMRS was significantly associated with reduced immune cell infiltration and diminished expression of immune effector molecules in MSI-H samples from both the TCGA-COAD and GSE39582 cohorts (Figure 4A and Figure S5A). In addition, the low-TCOMRS subgroup exhibited higher stromal, immune, and ESTIMATE scores, as calculated by the ESTIMATE algorithm (38) (Figure 4B and Figure S5B). These findings indicate that MSI-H COAD patients with low TCOMRS are more likely to present an inflamed immune microenvironment. Furthermore, TCOMRS was negatively correlated with PD-L1 protein expression in the TCGA-COAD cohort (Figure 4C). Based on these observations, we infer that MSI-H COAD patients with low TCOMRS may be more responsive to immune checkpoint blockade therapy, owing to their higher immune cell infiltration and pre-existing activation of the PD-L1 pathway.

Figure 4.

Association of the TCOMRS with immunotherapy response in MSI-H of COAD. (A) Heatmap of 29 immune-related pathways scored using the ssGSEA algorithm and immune effector molecule expression in low and high TCOMRS subgroups in MSI-H samples of TCGA-COAD. (B) Comparisons of stromal, immune, and ESTIMATE scores between the low and high TCOMRS subgroups in MSI-H samples of TCGA-COAD. Data were analyzed using the Student’s t-test. (C) Correlation analysis between the TCOMRS and PD-L1 protein levels in MSI-H samples of TCGA-COAD. Data were analyzed using Spearman correlation analysis. (D,E). Submap analysis demonstrated that the low and high TCOMRS subgroups in MSI-H samples of COAD (TCGA and GSE39582) cohorts were nearly identical to PD1-response (PD1-R) group defined in the melanoma cohort. (F,G). Comparisons of the expression of ICB response-related signatures (T-cell inflamed GEP and IPRES) between the low and high TCOMRS subgroups in MSI-H samples of TCGA-COAD. Data were analyzed using Student’s t-test. *P<0.05, **P<0.01, ***P<0.001. aDC, activated dendritic cell; APC, antigen-presenting cell; CCR, C-C chemokine receptor; COAD, colon adenocarcinoma; DC, dendritic cell; HLA, human leukocyte antigen; ICB, immune checkpoint blockade; IFN, interferon; MHC, major histocompatibility complex; MSI, microsatellite instability; MSI-H, high MSI; NK, natural killer; PD1, programmed cell death 1; PD-L1, programmed death-ligand 1; ssGSEA, single-sample gene set variation analysis; TCGA, The Cancer Genome Atlas; TCOMRS, tumor cell origin-based metastasis risk score.

Furthermore, SubMap analysis was conducted using a published melanoma immunotherapy dataset to estimate the likelihood of ICB response in the low and high TCOMRS subgroups (39). The results indicated that MSI-H COAD patients in the low-TCOMRS subgroup predicted to be more likely to respond favorably to anti-PD-1 therapy (Figure 4D,4E). In addition, the T-cell-inflamed GEP and the IPRES, both of which are associated with ICB response (40,41) , were more highly enriched in the low-TCOMRS subgroup than in the high-TCOMRS subgroup across the TCGA-COAD and GSE39582 cohorts (Figure 4F,4G and Figure S5C,S5D). To independently validate the differences in immune-cell infiltration between the TCOMRS subgroups, MCP-counter was additionally applied to the GSE39582 cohort. Compared with the low-TCOMRS group, the high-TCOMRS group exhibited significantly lower estimated abundance of T cells, cytotoxic lymphocytes, B-lineage cells, NK cells, myeloid dendritic cells, and neutrophils, whereas fibroblast abundance was significantly increased (Figure S5E). These findings were concordant with the ssGSEA and ESTIMATE results and suggested that high-TCOMRS tumors were characterized by reduced antitumor immune infiltration and a relatively fibroblast-enriched microenvironment. This stromal-rich and immune-depleted phenotype may contribute to immune exclusion and reduced predicted sensitivity to immune checkpoint blockade.

TCOMRS guide decision-making regarding chemotherapeutic treatment for metastatic COAD

To further assess the clinical utility of the TCOMRS model in treatment stratification for COAD, the “oncoPredict” algorithm was applied to estimate chemotherapy response across multiple cohorts (42). Samples with higher TCOMRS values were predicted to be less sensitive to 5-FU, oxaliplatin, and irinotecan, which are widely used in COAD treatment, in the TCGA, GSE39582, GSE38832, and GSE161158 cohorts (Figure 5A-5C). Chemotherapeutic intervention coupled with targeted therapies, including bevacizumab, is the backbone of metastatic COAD treatment (43). Thus, we further investigated the predictive value of TCOMRS in response of patients with metastatic COAD to chemotherapeutic and targeted therapies in two metastatic in GSE72970 and GSE104645. We observed that patients with higher TCOMRS values were more likely to benefit from chemotherapy plus bevacizumab and showed a more favorable clinical response than those receiving chemotherapy alone (Figure 5D-5G). More interestingly, the addition of bevacizumab to chemotherapy was not found to prolong survival and improve clinical response in low LM-risk subgroup of patients with metastatic COAD (Figure 5H-5K). These results indicate that TCOMRS may function as a predictive biomarker of resistance to several chemotherapeutic agents, while the benefit of anti-angiogenic therapy appears to be greatest in the high LM risk subgroup.

Figure 5.

TCOMRS guide decision-making regarding chemotherapeutic treatment for metastatic COAD. (A-C) Comparisons of molecular drug sensitivity between the low- and high-subgroups based on IC50 values of 5-fluorouridine, Oxaliplatin and Irinotecan in TCGA, GSE39582, GSE38832 and GSE161158 cohorts by applying “oncoPredict” algorithm. (D) Kaplan-Meier curves of OS according to chemotherapy combined with bevacizumab in high TCOMRS subgroup of GSE72970 cohort. (E) Kaplan-Meier curves of PFS according to chemotherapy combined with bevacizumab in high TCOMRS subgroup of GSE72970 cohort. (F,G) Clinical response to chemotherapy combined with bevacizumab in high TCOMRS subgroup of COAD patients (GSE72970 and GSE104645). (H) Kaplan-Meier curves of OS according to chemotherapy combined with bevacizumab in low TCOMRS subgroup of GSE72970 cohort. (I) Kaplan-Meier curves of PFS according to chemotherapy combined with bevacizumab in low TCOMRS subgroup of GSE72970 cohort. (J,K) Clinical response to chemotherapy combined with bevacizumab in low TCOMRS subgroup of COAD patients (GSE72970 and GSE104645). *P<0.05, **P<0.01, ***P<0.001. COAD, colon adenocarcinoma; CR, complete response; IC50, half maximal inhibitory concentration; OS, overall survival; PD, progressive disease; PFS, progression-free survival; PR, partial response; SD, stable disease; TCGA, The Cancer Genome Atlas; TCOMRS, tumor cell origin-based metastasis risk score.

Comparison of TCOMRS with conventional supervised survival models

To evaluate whether the OCLR-derived TCOMRS provided robust prognostic information compared with conventional supervised approaches, we constructed LASSO-Cox and survival-SVM models in the TCGA-COAD cohort using the same 9,014-gene feature space and externally validated the models in GSE39582, GSE38832, and GSE161158.

TCOMRS achieved C-index values of 0.649, 0.571, and 0.612 in GSE39582, GSE38832, and GSE161158, respectively. In the corresponding cohorts, the C-index values were 0.544, 0.557, and 0.523 for LASSO-Cox and 0.529, 0.635, and 0.557 for survival-SVM (Figure S6). Although survival-SVM showed the highest C-index in GSE38832, TCOMRS achieved the highest C-index in two of the three external validation cohorts. The mean C-index across the three external cohorts was 0.611 for TCOMRS, compared with 0.541 for LASSO-Cox and 0.574 for survival-SVM.

These findings indicate that TCOMRS exhibited relatively stable prognostic performance across independent datasets.

Biological pathways associated with TCOMRS

To characterize the biological processes underlying TCOMRS, we performed GSEA using the Hallmark gene sets in the TCGA-COAD and GSE39582 cohorts. Fatty acid metabolism, oxidative phosphorylation, and interferon-related pathways were preferentially enriched in the low-TCOMRS group (Figure S7A,S7B), indicating that these transcriptional programs were relatively attenuated at the bulk-tumor level in tumors with high TCOMRS. In addition, EMT pathway activity was positively correlated with TCOMRS (Figure S7C,S7D), suggesting that acquisition of a mesenchymal and invasive phenotype represents an important biological feature associated with elevated TCOMRS.

The enrichment of inflammatory and interferon-response pathways in the low-TCOMRS group was also consistent with the more favorable response to immune checkpoint blockade observed in patients with low TCOMRS. Nevertheless, because these analyses were based on transcriptomic associations, they do not establish a causal relationship between pathway activity and therapeutic response.

FABP1-expressing colon cancer cells were associated with LM of COAD

To investigate key oncogenic regulators involved in LM pathogenesis in COAD, we analyzed differentially expressed genes between the low and high TCOMRS subgroups in the TCGA-COAD cohort (Table S4) together with marker genes of MC2 (Table S5). We found 107 genes were upregulated in high TCOMRS subgroup [P<0.05 and log(FC) ≥1 were included as the standards] and 10 signature genes (PIGR, FABP1, CEACAM5, CXCL3, CXCL1.1, ATP1b1, CEACAM6, PLCB4, TM4SF1, and C15orf48) were highly expressed in LM subpopulation within MC2 (Figure 6A). Notably, FABP1, a member of the FABP family involved in the transport and metabolism of intracellular fatty acids, was broadly and uniquely expressed in malignant cells within COAD patients with high LM risk (Figure 6B), suggesting that FABP1-expressing colon cancer cells might be associated with LM of COAD.

Figure 6.

FABP1 promoted LM by enhancing EMT phenotype of colon cancer cells. (A) Heatmap showing the top 10 signature marker genes of MC2. (B) Venn diagram of intersection of the DEGs between low and high TCOMRS subgroups in the TCGA-COAD cohort and the marker genes of LM tumor cells MC2 subpopulation (GSE225857). (C) Transwell assay of FABPI-OE and vector control of HT29 or LOVO human colon cancer cells. (D,E) Total RNA and protein from FABP1-OE and vector control of HT29 or LOVO cells were extracted and the expression levels of EMT markers (E-cadherin, N-cadherin, and vimentin) were analyzed using RT-qPCR and western blotting, respectively. (F) Transwell assay of FABP1-KO and control of HT29 or LOVO cells. (G,H) The expression levels of EMT markers were analyzed using RT-qPCR and western blotting in FABP1-KO and control of HT29 or LOVO cells. (I) FABPI-OE and vector control cells (n=4) were injected into spleen of nude mice, and liver metastases (arrows) were detected after 3 weeks. **P<0.01, ***P<0.001. COAD, colon adenocarcinoma; DEG, differentially expressed gene; EMT, epithelial-mesenchymal transition; FABP1, fatty acid-binding protein 1; KO, knockout; LM, liver metastasis; OE, overexpressing; RT-qPCR, reverse transcription quantitative polymerase chain reaction; TCGA, The Cancer Genome Atlas; TCOMRS, tumor cell origin-based metastasis risk score.

As mentioned above, the EMT-related pathway was enriched in high LM risk subgroup. Thus, we next determined the effects of FABP1 on EMT of colon cancer cells. We observed that the overexpression of FABP1 (FABP1-OE) promoted cell migration and EMT by suppressing E-cadherin expression and upregulating N-cadherin and Vimentin expression in human HT29 and LOVO colon cancer cells, indicating that FABP1 promotes EMT-induced LM in colon cancer cells (Figure 6C-6E). As expected, CRISPR/Cas9-mediated knockout of FABP1 (FABP1-KO) inhibited cell migration and EMT in colon cancer cells (Figure 6F-6H). In addition, FABP1-OE HT29 and corresponding control cells were injected into the spleens of nude mice, and we found that FABP1 promoted LM of colon cancer cells in vivo (Figure 6I). Collectively, these results suggested that FABP1 promote LM by enhancing EMT phenotype of colon cancer cells.

FABP1-mediated fatty acid oxidation supported the survival of colon cancer cells under nutrition deprivation stress

To reveal the mechanism underlying the role of FABP1 in LM of COAD cells, pathway enrichment analysis (GSEA) was performed on DEGs between the FABP1 high and low expressing cells in the LM subpopulation of MC2. Notably, fatty acid catabolism-related cancer hallmark gene sets significantly contributed to the positive side of the FABP1 expression (Figure 7A). However, the effects of FABP1 on cell proliferation and survival under nutrient stress and the underlying mechanisms remain largely unknown. Therefore, examined FABP1 expression under glucose-deprivation conditions and found that the protein and mRNA levels of FABP1 were upregulated under glucose-deprived conditions in HT29 and LOVO cells (Figure 7B,7C). Additionally, BODIPY 500/510, a fluorescent dye that stains intracellular lipids, was used to evaluate lipid content, and we found that glucose deprivation led to a decrease in lipid content in both cell types (Figure 7D). We also observed that FABP1-OE or FABP1-KO had no significant effect on the growth of HT29 and LOVO cells under normal nutrient conditions (Figure 7E,7F). However, FABP1-OE led to significantly increased cell viability and decreased apoptosis under glucose-deprived conditions in both cell lines; these effects could be antagonized by etomoxir (ETO), a widely used small-molecule inhibitor of FAO (Figure 7G-7I), indicating that FABP1 confers colon cancer cells with enhanced survival ability under nutrient stress during LM by enhancing FAO.

Figure 7.

FABP1-mediated fatty acid oxidation supported the survival of colon cancer cells. (A) GSEA of differentially expressed genes between FABP1-high and FABP1-low cells in the LM-associated MC2 subpopulation using Hallmark gene sets. (B,C) Total RNA and protein from HT29 or LOVO cells were extracted under glucose deprived conditions for 12 and 24 h, respectively, and the expression levels of FABP1 were determined by RT-qPCR and western blotting. (D) The levels of lipid content under glucose deprived conditions for 24 h or normal nutrient conditions in HT29 and LOVO cells were detected using BODIPY 500/510 dye. (E,F) Cell proliferation of FABP1-OE and FABP1-KO in HT29 or LOVO cells under normal nutrient conditions were determined using a CCK-8 assay. (G) FABP1-OE and vector control of HT29 or LOVO cells were cultured under glucose deprived or normal conditions for 24 h. Cell lysates were collected, and FABP1, caspase 3, caspase 7, and their cleaved forms were analyzed using western blotting. (H,I) FABP1-OE and vector control of HT29 or LOVO cells were cultured under glucose deprived conditions and treated with Etomoxir (ETO) for 24 h. Cell viability was evaluated using CCK-8 and FABP1, caspase 3, caspase 7, and their cleaved forms were analyzed by western blotting. (J,K) HT29 or LOVO cells were cultured under glucose deprived conditions and treated with inhibitors of the MAPK signaling pathway (U0126, SB203580, and SP600125). Total RNA (12 h) and protein (24 h) were extracted and the expressions of FABP1 were determined using RT-qPCR and western blotting. *P<0.05, **P<0.01, ***P<0.001. CCK-8, Cell Counting Kit-8; DAPI, 4',6-diamidino-2-phenylindole; FABP1, fatty acid-binding protein 1; GAPDH, glyceraldehyde-3-phosphate dehydrogenase; GSEA, gene set enrichment analysis; KO, knockout; LM, liver metastasis; NES, normalized enrichment score; OE, overexpressing; padj, adjusted P value; RT-qPCR, reverse transcription quantitative polymerase chain reaction.

Finally, we investigated how FABP1 expression was regulated by nutrient stress. The mitogen-activated protein kinase (MAPK) signaling pathway is widely recognized for its role in nutrient stress response in tumor cells (44). Here, we found that blocking the MAPK cascade with inhibitors downregulated FABP1 expression under glucose-deprived conditions (Figure 7J,7K), suggesting that FABP1 expression is regulated by the MAPK signaling pathway. Taken together, these data suggested that FABP1-mediated fatty acid oxidation supported the survival of colon cancer cells under nutritional stress, leading to LM.

Discussion

LM is one of the leading causes of death in patients with COAD (45). The evaluation of LM risk and prognosis is essential for decision-making in subsequent examinations and clinical management of COAD patients. Numerous studies have developed computational tools to predict distant metastasis events and prognosis based on genomic signatures associated with metastasis-related survival outcomes within the PT bulk, such as GEPs and somatic mutation data (8,9). However, tumor metastasis is not easily detectable, and conventional methods for the diagnosis of clinical metastasis often tend to underestimate metastatic events. Genomic profiles derived from tumor biopsy samples represent the combined characteristics of multiple cell subpopulations. By contrast, conventional approaches based on non-negative matrix factorization or related decomposition methods often fail to resolve all tumor subtypes effectively in one optimization step (46).

The OCLR algorithm, has an advantage over its binary predictor counterparts. In our study, we chose the one-class framework to build an LM predictor because it remains robust even without a defined “negative” metastatic class and identify specific cell types in heterogeneous cell populations. Application of the one-class based TCOMRS model in COAD revealed a high degree of intratumor heterogeneity with respect to LM risk and provided insight into the quantitative assessment of the LM propensity of COAD at an early stage.

The OCLR framework is particularly suitable when a biologically defined reference population is available but a reliable negative class is difficult to define. In the present study, MC2 malignant cells derived from LMs provided such a reference state, allowing us to derive an LM-associated transcriptional signature without requiring an explicitly defined negative metastatic class. The resulting TCOMRS enabled quantitative assessment of the transcriptional similarity between bulk COAD tumors and this LM-associated malignant cell state. However, because the model was trained without an explicit negative population, TCOMRS should not be interpreted as a binary classifier of LM. Instead, it may also capture broader transcriptional programs related to metastatic progression, tumor aggressiveness, or dedifferentiation.

Consistent with this interpretation, our findings suggest that TCOMRS captures a transcriptional phenotype associated with metastatic behavior in colorectal cancer. In TCGA-COAD and GSE190609 cohorts, a higher proportion of patients with metastatic disease was observed in the TCOMRS-high group. This association was further reproduced in the independent GSE110009 cohort, in which the pretrained OCLR model was applied without retraining and metastatic lesions showed significantly higher TCOMRS values than PTs. Notably, the PT samples in GSE110009 served only as an external reference and were not incorporated as a negative population during model training. Thus, the observed separation between primary and metastatic samples was not explicitly imposed by the training procedure, although this analysis does not constitute validation of a conventional binary classifier. The consistency of these findings across independent datasets suggests that the transcriptional phenotype represented by TCOMRS is not restricted to the discovery cohort and may reflect biological features associated with tumor dissemination and metastatic progression. Nevertheless, because organ-specific metastatic-site annotations were limited or unavailable in these cohorts, these results primarily support an association with general metastatic behavior rather than liver-specific metastasis.

The analysis of GSE6988 provided more direct evidence for the association between TCOMRS and LM. In this cohort, which included explicit LM annotations, TCOMRS was significantly higher in patients with LM than in those without LM. This finding supports the potential relevance of TCOMRS to hepatic metastatic disease in COAD and is consistent with its derivation from LM-associated malignant cells. Collectively, these results suggest that TCOMRS may serve as a transcriptional biomarker associated with metastatic behavior and the presence of LM in COAD. However, this association alone does not establish specificity for LM. Demonstrating such specificity would require direct comparisons among patients with LM, metastases to other organs, and nonmetastatic disease in sufficiently large and clinically well-annotated cohorts.

Several additional limitations should be acknowledged. The retrospective design, limited sample sizes, inter-cohort heterogeneity, differences in sample composition and profiling platforms, and incomplete annotation of metastatic sites may affect the generalizability of our findings. Moreover, the observed associations do not establish a causal relationship between the transcriptional programs represented by TCOMRS and metastatic progression. Larger prospective studies incorporating longitudinal samples and detailed organ-specific metastatic information are therefore needed to determine whether TCOMRS can distinguish LM from metastasis to other organs and to evaluate its potential utility in risk stratification and metastatic surveillance.

In our study, no significant association was observed between TCOMRS and clinical features, including sex, gender, TNM classification, or clinical stage, in almost all patients with COAD. Previously, COAD has been classified into four distinct molecular subtypes based on consensus gene expression: CMS1, CMS2, CMS3, and CMS4, among which CMS2 exhibited marked WNT and MYC signaling activation, and CMS4 was characterized by prominent transforming growth factor-β activation, stromal invasion and angiogenesis (47). In the study, a high TCOMRS showed a strong association with subtypes of CMS2 and CMS4, two of which distinct molecular subtypes displayed high metastatic features in COAD. Furthermore, we observed that lower TCOMRS values correlated with a better prognosis in the COAD cohorts. Stem cell-like features, characterized by self-renewal and differentiation capacity, are essential for tumor re-initiation in distant tissues during metastasis (48). Our results showed that TCOMRS exhibited a significant positive association with the stemness index in COAD.

COAD patients with a high TMB are likely to benefit from immunotherapy. In our study, different mutational and immune landscapes were observed among groups with different TCOMRS values. COAD samples with higher TCOMRS values exhibited more genomic variants, and TCOMRS was negatively associated with TMB. MSI status and MSI-related alterations are predictive markers for immunotherapy and chemotherapy (49). Patients with the MSI-H phenotype have a better prognosis than patients with MSS/MSI-L. We found that MSS/MSI-L subgroup showed a higher TCOMRS than the MSI-H subgroup, and patients with high TCOMRS were less likely to respond to ICB therapy within the MSI-H subgroup of COAD. These findings suggest that TCOMRS may serve as a potential predictor of immunotherapy efficacy in COAD patients with the MSI-H phenotype.

It has been reported that chemotherapeutic intervention coupled with targeted therapies is the backbone of metastatic COAD treatment (43). Our findings indicated that patients with higher TCOMRS were more likely to derive substantial benefits from chemotherapy combined with bevacizumab as compared with chemotherapy alone. However, the addition of bevacizumab to chemotherapy was not found to prolong survival and improve clinical response in low TCOMRS subgroup of patients with metastatic COAD. These results suggested that the COAD patients with high TCOMRS value would be preferably treated with chemotherapy combined with targeted therapies even in the absence of detectable metastasis and the TCOMRS model could guide decision-making regarding therapeutic treatment for COAD. These findings suggest that TCOMRS may help identify subgroups with differential benefit from chemotherapy combined with bevacizumab; however, prospective validation is required before TCOMRS can be used to guide treatment decisions.

Compared with the supervised LASSO-Cox and survival-SVM models, TCOMRS demonstrated relatively stable prognostic performance across the external validation cohorts, achieving the highest C-index in two of the three cohorts and the highest mean C-index across all three cohorts, although survival-SVM performed better in GSE38832. Notably, the supervised models were directly optimized using survival outcomes, whereas TCOMRS was constructed without incorporating survival information. Instead, TCOMRS quantifies the transcriptional similarity between bulk tumors and LM-associated malignant cells, thereby retaining a direct biological interpretation as a measure of the LM-like malignant state. This distinction may contribute to the biological relevance and cross-cohort transferability of TCOMRS compared with purely outcome-driven prognostic models. Nevertheless, the present comparison was based on overall survival rather than LM classification and therefore does not establish the universal superiority of OCLR over other machine-learning algorithms for predicting LM. The lack of sufficiently large cohorts with clearly annotated LM and non-LM status precluded the construction and independent validation of conventional supervised LM classifiers. Future studies using larger, prospectively collected cohorts with detailed metastatic-site information are required to directly compare OCLR with LASSO, random survival forest, and other supervised learning approaches for LM prediction.

At the pathway level, high- and low-TCOMRS tumors exhibited distinct transcriptional programs. Fatty acid metabolism, oxidative phosphorylation, inflammatory response, and interferon-related pathways were preferentially enriched in low-TCOMRS tumors, whereas EMT activity was positively associated with TCOMRS. The metabolic enrichment pattern associated with FABP1-high cells did not fully parallel the bulk TCOMRS pattern, likely reflecting the difference between a tumor-cell-specific metabolic program and an integrative bulk-tumor transcriptional score. Notably, both analyses converged on EMT-related malignant progression. These observations support a biological link between the LM-like state captured by TCOMRS and FABP1-associated tumor-cell phenotypes, while not establishing a direct causal relationship between TCOMRS and FABP1-mediated metabolism.

Our study identified a subpopulation of tumor cells expressing high levels of FABP1, as the cellular origin of LM with COAD. Previously, FABP1 was reported to be overexpressed in tumor-associated macrophages (TAMs) and facilitate the M2-like polarization of TAMs to promote the progression of primary hepatocellular carcinoma by enhancing the FAO of TAMs (50). However, the role of FABP1 in the LM of patients with COAD remains unclear.

Tumor cells exhibit high levels of glycolysis, even in the presence of normal levels of oxygen, which is referred to as the Warburg effect (51). Addiction to a high glucose supply results in the susceptibility of tumor cells to nutrient stress (52). FAO has the capacity to fuel tumor growth in conditions of nutrient stress (53). To uncover the mechanism underlying the role of FABP1 in the LM of colon cancer cells, GSEA was performed on DEGs between FABP1 high- and low-expressing cells in the LM subpopulation, and it was found that fatty acid catabolism-related cancer hallmark gene sets were significantly associated with increased FABP1 expression. Loss- and gain-of function experiments were performed to further validate the potential mechanisms of FABP1-mediated LM in colon cancer cells. We found that FABP1 was upregulated under glucose-deprived conditions in colon cancer cells and FABP1 confers colon cancer cells with enhanced survival ability under nutrient stress during LM by enhancing FAO.

The MAPK signaling pathway plays an important role in nutrient stress adaptation, and FABP1 is upregulated via activation of the MAPK signaling pathway under nutrient stress conditions. Taken together, our results suggested that FABP1-mediated fatty acid oxidation supports the survival of colon cancer cells under nutritional deprivation stress, leading to LM.

Conclusions

Overall, we developed and validated TCOMRS, a tumor cell origin-based transcriptional score that quantifies the similarity between bulk COAD tumors and an LM-associated malignant-cell state. TCOMRS was associated with prognosis, metastatic behavior, LM status, and distinct therapeutic phenotypes across multiple cohorts. Moreover, we identified FABP1 as an important LM-associated regulator that promotes EMT and supports metabolic adaptation under nutrient stress. Further prospective cohorts and clinical studies are required to establish the liver-specific predictive performance and clinical utility of TCOMRS.

Supplementary

The article’s supplementary files as

Ethical Statement: The authors are accountable for all aspects of the work in ensuring that questions related to the accuracy or integrity of any part of the work are appropriately investigated and resolved. Experiments were performed under a project license (No. ZN2023203) granted by the Experimental Animal Welfare and Ethics Committee of Zhongnan Hospital of Wuhan University, in compliance with institutional guidelines for the care and use of animals.

Data Sharing Statement

Available at https://tcr.amegroups.com/article/view/10.21037/tcr-2026-1049/dss

tcr-15-10-756-dss.pdf ^{(71.8KB, pdf)}

DOI: 10.21037/tcr-2026-1049