Gene expression profiles and matching clinical details for both CRC and normal tissue samples were acquired from the Gene Expression Omnibus (GEO) and TCGA databases. This study incorporated three reliable CRC cohorts - TCGA-COADREAD, GSE39582, and GSE103479. Microarray data from GEO were normalized using the RMA algorithm based on corresponding platform annotations. RNA-sequencing results and clinical records for TCGA samples were downloaded from the TCGA data portal. Bulk RNA-seq data related to EMT and corresponding clinical information were collected from publicly available datasets in the Clinical Proteomic Tumor Analysis Consortium (CPTAC) colon cancer database^[14].

First, DEGs were screened between CRC and normal tissues using the TCGA dataset. Subsequently, a separate screening was performed to identify DEGs between EMT samples and non-EMT samples in an independent database. When the DEGs in the two datasets overlapped, a total of 289 genes were obtained (|log₂FoldChange| > 1, P < 0.05). Univariate Cox regression analysis was then applied to these genes using the “survival” package in R software. The 35 genes with prognostic value were further included in a multivariate Cox regression framework to construct a risk scoring system. Each sample’s risk value was derived according to the equation: Risk score = h₀(t) × exp(β₁X₁ + β₂X₂ + … + β_nX_n), in which n stands for the number of model genes, β is the regression coefficient, X represents the gene expression level, and h₀(t) was produced by the “predict” tool. Based on the optimal threshold, patients in both training and validation sets were divided into high-risk and low-risk subgroups. Kaplan-Meier survival analysis was then used to compare survival outcomes between the two groups and assess the predictive performance of this eight-gene signature associated with EMT.

To explore the functional differences between the high-risk and low-risk subgroups in the TCGA cohort, we carried out gene expression difference analysis with the R package “DESeq2”. DEGs were identified using |log₂FoldChange| > 1 and P < 0.05. For Gene Ontology (GO) and Kyoto Encyclopedia of Genes and Genomes (KEGG) enrichment analysis of these DEGs, significantly enriched terms were determined using an FDR-adjusted P-value < 0.05, and the analysis was performed with the “clusterProfiler” R package. The enrichment results are presented as bar plots. In addition, Gene Set Variation Analysis (GSVA) analysis was performed using the R package “GSVA”, based on the reference gene set “c2.cp.kegg.v7.5.1.symbols.gmt” (downloaded from MSigDB), to elucidate the biological process differences between the two groups.

Risk scores derived from CRC samples were integrated with clinicopathological data (age, gender, pathological stage, and AJCC-TNM stage) for subsequent analysis. Univariate and multivariate Cox regression models were used to screen for independent prognostic factors with a statistical threshold of P < 0.05, and these significant factors were employed to construct a nomogram for individualized survival prediction. The predictive efficacy of the established nomogram was validated through 1-, 3- and 5-year receiver operating characteristic (ROC) curve analysis and calibration plots, which were performed with the R software packages of “survivalROC” and “rms”.

The drug sensitivity of each sample was predicted using the “pRRophetic” R package. This package utilizes gene expression data and corresponding drug sensitivity (IC₅₀) information from the Cancer Genome Project (CGP) as the training set to build predictive models. To assess immunotherapeutic response, the tumor immune dysfunction and exclusion (TIDE) algorithm was employed (http://tide.dfci.harvard.edu/). Moreover, the external IMvigor210 dataset was utilized to further verify the predictive performance of the EMT-associated prognostic model in a separate immunotherapy patient cohort.

This study was approved by the Biomedical Ethics Committee of Ruijin Hospital, Shanghai Jiao Tong University School of Medicine (No. 2021-28), with written informed consent obtained from all enrolled CRC patients. Human CRC cell lines HT29 and SW480, together with the normal colonic epithelial cell line NCM460, were purchased from Shanghai Shenger Biotechnology Co., Ltd., which originally sourced these cells from the American Type Culture Collection (ATCC). All cell lines were authenticated by STR profiling, performed by Suzhou Jianda Biotechnology Co., Ltd. in 2026. Routine mycoplasma testing was performed using a polymerase chain reaction (PCR)-based assay throughout the experimental period, and all lines were confirmed negative. Cells were used between passages 5 and 15 after thawing. All cell lines were cultured in RPMI-1640 medium supplemented with 10% fetal bovine serum (FBS) at 37 °C in a humidified atmosphere with 5% CO₂. The RRIDs for HT29, SW480, and NCM460 are CVCL_0320, CVCL_0546, and CVCL_0460, respectively. Oxaliplatin resistant HT29 and SW480 cells were established by exposing their respective parental cells to stepwise increasing concentrations of oxaliplatin, with the final induction concentration reaching 2 µM over a period of 6 months. The established resistant sublines were subsequently maintained in RPMI 1640 medium supplemented with 10% FBS and 2 µM oxaliplatin to preserve the resistant phenotype. Resistance was validated by assessing the half-maximal inhibitory concentration using a CCK-8 assay, with the detailed IC₅₀ values in Supplementary Table 1.

Two independent short hairpin RNA (shRNA) sequences targeting human FABP4 (shFABP4-1: ACATGATCATCAGTGTGAATG; shFABP4-2: CGAGAGTTTATGAGAGAGCAT) and a non-targeting scramble control (shNC: TTCTCCGAACGTGTCACGT) were individually cloned into the lentiviral backbone PGMLV-hU6-H-shRNA-CMV-ZsGreen1-PGK-Puro (Shanghai Genomeditech Co., Ltd., China), which co-expresses a ZsGreen1 reporter and a puromycin resistance gene for fluorescent tracking and antibiotic selection. Lentiviral transduction was performed on CRC cells to establish stable cell lines. CRC cells were seeded 24 h prior to infection to reach 50%-70% confluence on the day of transduction. Lentiviral particles were added at a multiplicity of infection (MOI) of 30 in the presence of 8 µg/mL polybrene to enhance transduction efficiency. The virus-polybrene mixture was applied to cells and incubated at 37 °C for 24 h. The medium was then replaced with fresh complete medium, and cells were further cultured for 48 h to allow transgene expression. For stable cell line generation, transduced cells were selected with 2 µg/mL puromycin. The knockdown efficiency of FABP4 was then validated via Western blotting.

Total RNA was extracted from cultured cells using the FastPure® RNA Isolation Kit V2 (Vazyme, China), and RNA concentration was determined by spectrophotometry. Reverse transcription was performed using HiScript® RT SuperMix with genomic DNA elimination reagents, with 1 µg of total RNA used per 20 µL reaction under the following conditions: 37 °C for 15 min, followed by 85 °C for 5 s. Quantitative real-time PCR was subsequently carried out using ChamQ Universal SYBR qPCR Master Mix following the manufacturer’s instructions. The thermal cycling protocol consisted of an initial denaturation at 95 °C for 30 s, followed by 40 cycles of 95 °C for 10 s and 60 °C for 30 s, with a final melt curve analysis to verify product specificity. All reactions were performed in technical triplicate. The primer sequences used in this study are listed in Supplementary Table 2.

CRC cells were lysed in RIPA buffer supplemented with protease and phosphatase inhibitor cocktails to preserve protein integrity. Protein concentrations were determined using a standard BCA protein assay kit, and equal amounts of total protein (20 µg per lane) were separated by SDS-PAGE and subsequently transferred onto PVDF membranes. The membranes were blocked with 5% non-fat milk in TBST for 1 h at room temperature to prevent nonspecific binding, followed by overnight incubation at 4 °C with primary antibodies against CD44, OCT4, SOX2, FABP4, and GAPDH (as the internal control). After thorough washing, the membranes were incubated with corresponding HRP-conjugated secondary antibodies (1:5,000) for 1 h at room temperature. Specific protein bands with immune reactivity were visualized using chemiluminescence imaging systems, and their gray values were quantitatively analyzed with ImageJ. Details of all antibodies are listed in the Supplementary Table 3.

Cell proliferative capacity was assessed using the commercial CCK-8 detection method. CRC cells were seeded into 96-well culture plates at a density of 5 × 10³ cells per well, with each well supplemented with 100 μL of RPMI-1640 culture medium. At preset experimental time points, the original culture medium was replaced with fresh medium mixed with 10% CCK-8 working solution, and cells were incubated for two hours at 37 °C under dark conditions. A microplate reader was finally used to detect the optical density value at the wavelength of 450 nm.

For the wound healing test, CRC cells were grown to full confluence in 6-well plates. A sterile pipette tip was used to create uniform scratches on the cell monolayer, followed by gentle PBS washing. The processed cells were then cultured in serum-free medium for a 48-hour incubation period. In the transwell migration experiment, 5 × 10⁴ suspended cells in serum-free medium were added to the upper inserts, while the lower compartments were filled with medium containing 20% FBS to induce cell migration. After incubation, migrated cells were fixed with 4% paraformaldehyde for 30 min and stained with 0.25% crystal violet solution for 25 min for subsequent observation and analysis.

Six-week-old male BALB/c nude mice were purchased from Phenotek Biotechnology (Shanghai) Co., Ltd. All animal experiments were approved by the Institutional Animal Care and Use Committee of Phenotek Biotechnology under protocol number (AUP-202511-02). All animal housing and experiments were conducted in strict accordance with the institutional guidelines for the care and use of laboratory animals. For subcutaneous xenograft establishment, HT29 cells (5 × 10⁵ cells per mouse in 100 µL PBS) were inoculated into the right lateral flank of each mouse. A total of 32 mice were used in this study, with 8 mice randomly assigned to each of the four experimental groups. Sample size for the primary endpoint (tumor growth) was determined by power analysis (α = 0.05, 1 - β = 0.80) based on preliminary data, indicating a minimum of 5 mice per group for the longitudinal tumor-growth cohort. An additional 3 mice per group were separately allocated as an independent cohort for histological and molecular analyses, yielding a total of 8 mice per group. In each group, five mice were designated for longitudinal tumor-volume monitoring (tumor-growth cohort). The remaining three mice per group were euthanized at predetermined intermediate time points for histological evaluation and molecular assays (tissue-analysis cohort). The two cohorts were independent, with no overlap. When tumor volumes reached approximately 100 mm³ [measured by digital caliper and calculated as volume = (length × width²)/2], mice were randomly allocated to treatment groups using a computer-generated randomization sequence. The investigator responsible for tumor measurements and data analysis was blinded to group allocation. Mice received intraperitoneal injections of either normal saline (control) or oxaliplatin (5 mg/kg) twice per week. Tumor size and body weight were recorded every two days throughout the treatment cycle. Each tumor was measured three times as technical replicates, with five mice per group as biological replicates for tumor growth analysis. Mice were monitored daily for signs of distress or morbidity. Humane endpoints included tumor volume exceeding 2,000 mm³, body weight loss > 20% of initial weight, tumor ulceration/necrosis, or moribund status; any animal reaching these endpoints was euthanized immediately. At the end of the experiment, the remaining five mice per group in the tumor-growth cohort were euthanized by cervical dislocation under isoflurane anesthesia. No animals met the exclusion criteria, and no animals were excluded from the final analysis or died unexpectedly during the experiment.

All statistical analyses and data processing were performed using R software (version 4.1.2). All in vitro experiments were performed with at least three biological replicates, each containing three technical replicates. Unless otherwise indicated, data are presented as mean ± standard deviation (SD). Comparisons between two groups were assessed using unpaired two-tailed Student’s t-test, while comparisons among three or more groups were evaluated using one-way analysis of variance (ANOVA) followed by Tukey’s post-hoc test for multiple comparisons. For tumor growth curves, two-way repeated-measures ANOVA with Bonferroni post-hoc correction was applied. Patient survival profiles were analyzed using Kaplan-Meier curves. The “timeROC” package was used to generate time-dependent ROC curves to evaluate the predictive efficiency of the prognostic model, and univariate and multivariate Cox regression analyses were utilized to screen independent prognostic factors. A two-tailed P < 0.05 was considered statistically significant. For all figures, ns indicates P ≥ 0.05, ^*P < 0.05, ^**P < 0.01, and ^***P < 0.001.

First, DEGs between CRC samples and normal tissues in the TCGA database were screened. In addition, DEGs between EMT samples and non-EMT samples from another database were selected. When the DEGs in the two datasets overlapped, a total of 289 genes were obtained [Figure 1A]. Thirty-five DEGs with prognostic value were preliminarily screened via univariate Cox regression analysis with a statistical threshold of P < 0.05, which were reserved for subsequent model building. An eight-gene signature was constructed based on multivariable Cox regression analysis [Figure 1B]. Hazard ratio (HR) values were computed for each candidate gene to assess its prognostic implications, with HR values less than 1 indicating protective effects and HR values greater than 1 suggesting adverse risk factors for CRC patients. A personalized scoring system termed the risk score was established based on the expression levels and corresponding regression coefficients of signature genes. This formula was applied to compute the individual risk score for every enrolled CRC specimen. Subsequent analyses were carried out to explore survival disparities among samples with varied risk score levels. Within the training set, a high-risk score was associated with poorer overall survival (OS) among CRC samples relative to a low-risk score [Figure 1C]. The validation sets further confirmed these observations [Figure 1D]. Collectively, these results demonstrated that the prognostic model constructed from EMT-associated genes could reliably predict survival status for CRC patients.

To further explore the underlying molecular mechanisms, GO enrichment analysis of DEGs between two groups was conducted. Results revealed top-ranked pathways, including chemotaxis, cell-matrix adhesion, focal adhesion, and extracellular matrix structural constituent [Figure 2A]. Furthermore, KEGG pathway analysis underscored a notable enrichment in the extracellular matrix (ECM)-receptor interaction, Hippo signaling, Cell cycle, and Apoptosis pathways [Figure 2B]. GSVA was then applied to dissect discrepant biological functions across the two groups. The high-risk group displayed enhanced stromal-related biological activities, such as EMT2/3 and angiogenesis [Figure 2C]. Meanwhile, GSVA results also indicated prominent enrichment of stroma-associated signaling events such as ECM-receptor interaction and focal adhesion in the high-risk group [Figure 2D]. In summary, the findings above suggested that CRC patients with different risk scores exhibited differences in prognosis, biological activity, and TME, providing new insights into CRC research.

Independent prognostic factors - including age, pathological stage, M stage, and risk score - were identified via univariate and multivariate Cox regression [Figure 3A and B]. A comprehensive prognostic nomogram integrating these variables was constructed to estimate 1-, 3-, and 5-year OS in CRC patients [Figure 3C]. Time-dependent ROC curve analysis was conducted to assess model performance, with AUC values calculated as 0.77, 0.77, and 0.78 at the three corresponding time points, thereby validating the stable and powerful predictive efficacy of this model [Figure 3D]. Meanwhile, calibration curve results showed great consistency between survival outcomes predicted by the nomogram and actual clinical observations [Figure 3E-G], indicating that this nomogram can serve as a reliable auxiliary tool for precise prognostic stratification of CRC cases.

Additional explorations were performed to clarify the correlation between calculated risk scores and drug susceptibility. For chemotherapeutic efficacy evaluation, high-risk samples presented notably higher IC₅₀ levels of 5-fluorouracil than low-risk counterparts (P = 0.024; Figure 4A), suggesting that CRC patients with elevated risk score may derive limited benefit from conventional chemotherapy. Considering the widespread clinical application and promising anti-tumor effects of immune checkpoint inhibitors (ICIs), the TIDE algorithm was adopted to explore whether this risk model could predict immunotherapeutic efficacy. Patients in the high-risk subgroup had markedly elevated TIDE scores relative to low-risk cases [Figure 4B]. In line with this result, low-risk individuals achieved a better immune treatment response rate, reaching 45.9% compared with 28.4% in the high-risk group [Figure 4C]. Moreover, Wilcoxon rank-sum tests confirmed significant correlations between risk score levels and core clinical characteristics in the IMvigor210 cohort, covering immune phenotype, metastatic status, and TCGA molecular subtype [Figure 4D-F]. All the above analytical results strongly suggested that the proposed risk signature is closely linked to both chemotherapeutic sensitivity and immunotherapeutic responsiveness in CRC.

Subsequently, the function of model-associated genes was evaluated. Among them, the expression of FABP4, GRP, ADH1B, CCL19, CALB2, and MDFIC were significant prognostic indicators [Figure 5A-F]. Specifically, elevated levels of FABP4, GRP, CALB2, and MDFIC were correlated with a poor prognosis in CRC. Conversely, reduced expression of ADH1B and CCL19 was also significantly associated with adverse patient outcomes. To verify the clinical and biological relevance of the established EMT-associated gene signature at the cellular level, we applied reverse transcription quantitative polymerase chain reaction (RT-qPCR) to detect the expression profiles of signature genes. Results showed that the expression levels of FABP4, GRP, LRP4, NOX1, CALB2, and MDFIC were up-regulated in CRC cells compared with NCM460. On the contrary, the expression levels of ADH1B and CCL19 were down-regulated [Figure 5G].

Several studies have implicated FABP4 in promoting tumorigenesis and contributing to drug resistance in other cancers, such as ovarian and breast cancer^[15,16]. To further explore the role of FABP4 in CRC, CRC samples were stratified into two groups based on the expression level of FABP4. GO enrichment analysis of DEGs between the two groups revealed top-ranked pathways including epithelial-to-mesenchymal transition, cell-matrix adhesion, tight junction, and extracellular matrix structural constituent [Figure 6A]. KEGG enrichment analysis identified Focal adhesion, Hippo signaling, and Apoptosis as significantly modulated pathways [Figure 6B]. Consistent with this, FABP4 was implicated in the regulation of critical oncogenic processes in CRC, including EMT and apoptosis. In the TIDE analysis, CRC samples with a high level of FABP4 were associated with tumor immune dysfunction, suggesting a lower therapeutic benefit and immune response to ICI treatment in these patients [Figure 6C and D]. We further examined the differential enrichment of known biological pathways between the two groups [Figure 6E]. Samples with a high level of FABP4 displayed elevated stromal functional activities, including activated EMT2/3 signaling and angiogenesis. In the IMvigor210 cohort, elevated FABP4 expression was correlated with adverse survival outcomes for tumor patients [Figure 6F]. In addition, there were significant associations between FABP4 levels and immune phenotype, treatment response, metastatic state, or TCGA subtype [Figure 6G]. To experimentally validate the bioinformatic predictions regarding the tumor immune microenvironment, immunofluorescence (IF) staining was performed on CRC tissue samples stratified by FABP4 expression levels. As shown in Supplementary Figure 1, tumors with high FABP4 expression exhibited significantly stronger α-SMA fluorescence intensity compared with their low-expression counterparts, indicating pronounced stromal activation and increased cancer-associated fibroblast infiltration. Furthermore, IF analysis of immune cell markers [Supplementary Figure 2] revealed that high-FABP4 tumors harbored a marked increase in CD8⁺ T-cell infiltration. Notably, this enhanced CD8⁺ infiltration was accompanied by concurrent upregulation of PD-L1 expression within the TME. Collectively, these findings demonstrate that FABP4 is closely associated with stromal remodeling and the establishment of an immunosuppressive PD-L1/PD-1 axis, which may have important implications for the efficacy of clinical immunotherapy in CRC.

To assess the function of FABP4 in CRC, experiments in vitro were performed. Immunohistochemical analysis of clinical specimens confirmed that FABP4 was upregulated in tumor lesions relative to paired peri-tumor normal tissues [Figure 7A and B]. FABP4 knockdown in HT29 and SW480 cells was achieved via shRNA transduction, with efficiency confirmed by RT-qPCR and western blotting at the mRNA and protein levels, respectively [Figure 7C and D]. CCK-8 results illustrated that FABP4 knockdown greatly suppressed the proliferative ability of CRC cells [Figure 7E]. Additionally, both wound healing and transwell migration assays showed that suppression of FABP4 markedly weakened the migratory capacity of CRC cells [Figure 7F-I]. All these outcomes demonstrate that FABP4 exerts critical oncogenic effects by facilitating CRC cell growth and migration.

To further investigate the effect of FABP4 on chemoresistance, CCK-8 assays were performed. Cellular assays showed that FABP4 knockdown markedly enhanced the sensitivity of CRC cells to oxaliplatin relative to shNC cells [Figure 8A and B]. Subsequent PCR and western blot detection consistently verified elevated FABP4 expression in chemotherapy-resistant CRC cell lines [Supplementary Figures 3 and 4]. Functional validation further confirmed that FABP4 knockdown effectively reversed the drug resistance phenotype and enhanced oxaliplatin responsiveness in resistant tumor cells [Supplementary Figure 5]. Mechanistic exploration via western blot assays illustrated that FABP4 downregulation suppressed the expression of stemness markers including CD44, OCT4, and SOX2, which further restored tumor cell sensitivity to oxaliplatin [Supplementary Figure 6]. To corroborate our in vitro findings, the effect of FABP4 knockdown on tumor growth and drug resistance was assessed in a subcutaneous xenograft model. Tumors derived from FABP4-depleted HT29 cells exhibited significantly reduced growth compared to controls. Importantly, the combined intervention of FABP4 knockdown and oxaliplatin administration produced the strongest anti-tumor efficacy, with the smallest tumor volume and weight observed among all experimental groups [Figure 8C-E, Supplementary Figure 7]. Body weight analysis revealed no significant differences among experimental groups [Figure 8F], indicating that FABP4 knockdown and oxaliplatin treatment were not associated with detectable toxicity in vivo. In addition, Ki67 and TUNEL staining of tumors from four groups were performed. IHC staining results showed a similar trend, with the most significant decrease in cell proliferation and increase in cell apoptosis observed in the combination of FABP4 knockdown and oxaliplatin treatment group [Figure 8G and H]. Taken together, these experimental results indicate that FABP4 suppression restrains CRC cell growth and elevates cellular responsiveness to oxaliplatin, implying its potential as an innovative therapeutic target for CRC clinical treatment.