Dietary restriction-responsive genes define prognostic subtypes and predict metastasis in colorectal cancer

Abstract

Background

Colorectal cancer (CRC) is a major cause of cancer-related death, with a poor prognosis often due to metastasis and recurrence. Dietary restriction (DR) is known to delay tumor progression and extend lifespan, but the roles of dietary restriction-responsive genes (DRRGs) in CRC remain unclear. This study aimed to identify prognostic DRRGs and explore their associations with tumor behavior and immune features.

Methods

Transcriptomic data from The Cancer Genome Atlas (TCGA) and Gene Expression Omnibus (GEO) were analyzed alongside 276 DRRGs from the GenDR database. Differentially expressed DRRGs were identified, followed by univariate Cox regression to assess prognostic relevance. A least absolute shrinkage and selection operator (LASSO)-Cox model was used to construct a prognostic signature, which was validated in external cohorts. Immune cell infiltration, functional enrichment, and unsupervised clustering were performed to evaluate the biological roles of DRRGs. Associations of the risk score with clinicopathological features, genomic alterations, and immunotherapy response (IPS) were further evaluated. Machine learning (ML) models were built to predict metastasis and recurrence using Shapley Additive exPlanations (SHAP) analysis.

Results

A total of 11 DRRGs were found to be significantly associated with CRC prognosis. A four-gene signature (RGS16, PLIN4, SLC13A2, FOXD2) effectively stratified patients into high- and low-risk groups with distinct survival outcomes. High-risk patients exhibited enrichment of extracellular matrix (ECM) and inflammatory pathways, whereas low-risk patients were associated with mitochondrial metabolism. Immune profiling revealed increased fibroblasts and myeloid cells in the high-risk group. Clustering based on DRRGs identified two molecular subtypes with different metabolic and immune features. High-risk tumors exhibited elevated tumor mutational burden (TMB) and microsatellite instability-high (MSI-H) frequency, while risk scores were inversely associated with stemness. IPS analysis further indicated that low-risk patients may derive greater benefit from CTLA-4 blockade. In metastasis prediction (GSE41258), the XGBoost model achieved an area under the receiver operating characteristic curve (AUC) of 0.855, with Matrix Gla Protein (MGP) identified as a key contributor via SHAP analysis.

Conclusions

We established a DRRG-based prognostic model for CRC and uncovered their links to metabolic regulation, immune infiltration, and metastasis. These findings highlight DRRGs as potential biomarkers and therapeutic targets and suggest that DR-mimicking strategies may benefit CRC management.

Highlight box.

Key findings

• We identified 11 dietary restriction-responsive genes (DRRGs) associated with colorectal cancer (CRC) prognosis and constructed a four-gene signature (RGS16, PLIN4, SLC13A2, FOXD2). The model stratified patients into risk groups with distinct outcomes and revealed links to metabolism and immune infiltration.

What is known and what is new?

• Dietary restriction (DR) has been shown to delay tumorigenesis across species, but the specific DR-responsive genes involved in CRC remain unclear.

• This study establishes the first DRRG-based prognostic model and demonstrates its predictive value for survival, immune features, and metastatic potential.

What is the implication, and what should change now?

• DRRG signatures may serve as biomarkers to guide prognosis and immunotherapy selection. Strategies that mimic DR-mediated gene regulation could be explored as novel therapeutic interventions in CRC.

Introduction

Colorectal cancer (CRC) is one of the most commonly diagnosed malignancies worldwide and a leading cause of cancer-related mortality (1,2). Despite advances in surgical techniques and neoadjuvant chemoradiotherapy, long-term outcomes remain suboptimal due to the high incidence of metastasis and tumor recurrence. Approximately 20% of CRC patients present with metastatic disease at diagnosis, and nearly half will eventually develop metastases over the course of their illness (3-5). Even after curative resection of advanced tumors, local tumor recurrences occur in up to 30% of cases, underscoring the aggressive nature of advanced CRC and its impact on survival (6).

Mounting evidence indicates that modifiable lifestyle factors, particularly diet, play a significant role in CRC development and progression. Among dietary interventions, dietary restriction (DR)—a sustained reduction in caloric intake or specific nutrients without malnutrition—has emerged as a robust non-pharmacological strategy for extending lifespan and delaying cancer across diverse species (7,8). In model organisms ranging from yeast and nematodes to rodents, DR consistently prolongs lifespan and inhibits spontaneous tumorigenesis (9). The potential for DR or fasting-based regimens to influence cancer outcomes in primates and humans is also supported by epidemiological and clinical studies, although results in mammals are more nuanced (10).

Mechanistically, DR exerts broad anti-cancer effects by reprogramming key cellular pathways. It significantly reduces circulating insulin and insulin-like growth factor-1 (IGF-1) levels, thereby attenuating pro-oncogenic signaling through the PI3K/AKT/mTOR and RAS/MAPK pathways (11-13). This metabolic reprogramming under DR leads to decreased cell proliferation, enhanced apoptosis of damaged cells, and lower oxidative stress and DNA damage, ultimately restraining tumor initiation and growth (10). Furthermore, DR favorably modulates the tumor immune microenvironment: it can dampen chronic inflammation and reduce immunosuppressive cell populations, while promoting cytotoxic T-cell activity and immunosurveillance against emerging tumors (14). Beyond systemic metabolic effects, recent research suggests that specific gene regulatory networks mediate the link between DR and cancer suppression. Large-scale transcriptomic studies have identified numerous dietary restriction-responsive genes (DRRGs), of which its expression is altered under caloric restriction across different organisms (15,16). Notably, many of these DRRGs are evolutionarily conserved and implicated in pathways relevant to aging and stress resistance, hinting that they may also impact tumor biology (16). However, the precise genetic mechanisms by which DR influences colorectal carcinogenesis remain insufficiently understood. In CRC, it is unclear which DRRGs are most pivotal for mediating the anti-tumor effects of DR, and how perturbations in these genes might affect tumor progression, metastasis, and the interaction with the immune microenvironment. This gap in knowledge limits our ability to translate dietary interventions or DR-mimicking therapies into effective clinical strategies for CRC patients.

In this study, we performed an integrative multi-cohort analysis to explore the roles of DRRGs in CRC development and patient outcomes. Using transcriptomic data from The Cancer Genome Atlas (TCGA) and Gene Expression Omnibus (GEO) datasets together with curated DRRGs from the GenDR database, we identified 11 prognostic DRRGs and developed a four-gene signature (RGS16, PLIN4, SLC13A2, FOXD2) that stratified patients into high- and low-risk groups with distinct survival outcomes. We further constructed an explainable prognostic model and validated its performance across independent cohorts. The risk score was positively associated with tumor progression [tumor (T), node (N), and clinical stage] and linked to genomic features including higher tumor mutational burden (TMB), increased microsatellite instability-high (MSI-H) frequency, and lower RNA-based stemness index (RNAss). To enhance interpretability, machine-learning classifiers coupled with Shapley Additive exPlanations (SHAP) were applied, highlighting Matrix Gla Protein (MGP) as a major contributor in metastasis prediction. Functional enrichment and immune infiltration analyses revealed that high-risk tumors were enriched in extracellular matrix (ECM) and inflammatory pathways with increased fibroblast and myeloid infiltration, while low-risk tumors were characterized by mitochondrial metabolism. IPS analysis suggested that low-risk patients may benefit more from CTLA-4 blockade, underscoring the potential therapeutic relevance of DRRG-based stratification. Overall, this work uncovers conserved DRRG signatures associated with CRC prognosis and metastasis and highlights their potential as biomarkers and therapeutic targets. We present this article in accordance with the TRIPOD reporting checklist (available at https://tcr.amegroups.com/article/view/10.21037/tcr-2025-1329/rc).

Methods

Data acquisition and processing

Transcriptomic profiles and corresponding clinical information for colon adenocarcinoma (COAD) were downloaded from TCGA database (https://portal.gdc.cancer.gov/). Additionally, three GEO datasets were retrieved from the GEO repository (https://www.ncbi.nlm.nih.gov/geo/) for further analysis. Detailed information on the datasets is summarized in Table 1. DRRGs were obtained from the GenDR database (http://genomics.senescence.info/diet/), and the full list is provided in table available at https://cdn.amegroups.cn/static/public/tcr-2025-1329-1.xlsx. The study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments.

Table 1. Summary of datasets and sample composition.

Datasets	Platform	Normal samples	Tumor samples	Clinical outcome
TCGA-COAD	−	39	395	−
GSE17538	GPL570	0	183	51 recurrences
GSE39582	GPL570	0	479	117 recurrences
GSE41258	GPL96	0	203	47 metastases

TCGA-COAD, The Cancer Genome Atlas-colon adenocarcinoma.

To mitigate batch effects among different datasets, the “sva” R package was employed for data normalization and integration. Differentially expressed genes (DEGs) between COAD tumor tissues and adjacent normal tissues were identified using the “limma” package (17). The threshold criteria were set as |log₂ fold change (FC)| >1 and false discovery rate (FDR) <0.05. Visualization of the DEGs was conducted using the “pheatmap”, “dplyr”, “ggplot2” and “ggrepel” R packages.

Survival analysis and genomic visualization

To evaluate the prognostic relevance of DRRGs, univariate Cox proportional hazards regression analysis was performed using overall survival (OS) as the clinical endpoint. Kaplan-Meier survival analysis was then conducted to assess survival differences between high- and low-expression groups, and survival curves were plotted accordingly.

Genes with a P value <0.05 in the univariate Cox analysis were selected for further analysis. Manhattan plots and circos plots were generated to intuitively display the genomic distribution of the prognostic DRRGs across chromosomes. Furthermore, copy number variation (CNV) analysis was conducted to determine the frequency of copy number gains and losses among the DRRGs. The genes were ranked based on CNV frequency, and bar plots were constructed to visualize the amplification and deletion events.

Construction and validation of DRRG prognostic model

To construct a robust prognostic model based on DRRGs, we applied a combination of least absolute shrinkage and selection operator (LASSO) regression and multivariate Cox proportional hazards analysis. Gene expression and survival data from TCGA-COAD, GSE17538, and GSE39582 were preprocessed by imposing a minimum survival threshold of 1 day to eliminate zero values. The dataset was randomly split into training and testing subsets (1:1 ratio) using the “createDataPartition()” function from the “caret” package, and this process was iterated 1,000 times to ensure stability of the model. The resulting prognostic signature was used to calculate individual risk scores for both training and testing sets based on the Cox model. Only models that met the following criteria were retained for further analysis: (I) P value <0.01 for survival difference in the training set; (II) the area under the receiver operating characteristic (ROC) curve (AUC) >0.65 in the training set; (III) P value <0.05 in the test set; (IV) AUC >0.60 in the test set. Samples were stratified into high- and low-risk groups using the median risk score as the cutoff. Kaplan-Meier analysis and log-rank tests were performed to evaluate survival differences between groups. To evaluate the predictive performance of the prognostic model over time, time-dependent ROC curves were plotted using the “timeROC” package. Aalen’s additive weighting approach was applied to calculate AUC at 1, 3, and 5 years. ROC curves were generated for both TCGA and GEO cohorts, and AUC values were reported to quantify sensitivity and specificity at each time point.

Nomogram and calibration

To facilitate clinical application of the prognostic model, a nomogram was constructed based on the risk score and clinical parameters using the “regplot” and “rms” R packages. A multivariate Cox proportional hazards model incorporating age, clinical features, and risk score was built to predict OS. The “regplot()” function was used to visualize the nomogram. Individual patient risk scores were calculated using the “predict()” function. Calibration curves at 1-, 3-, and 5-year time points were generated using the “calibrate()” function with 1,000 bootstrap resamplings to evaluate the predictive accuracy of the nomogram. Concordance index (C-index) values and corresponding 95% confidence intervals (CIs) were calculated from the fitted Cox model to assess discriminative ability. Expression levels of genes from the LASSO-Cox model were compared between high- and low-risk groups using boxplots generated by the “ggpubr” package.

Gene set enrichment analysis (GSEA)

To explore biological pathways associated with the risk groups, GSEA was performed using the “clusterProfiler” and “enrichplot” packages (18). Enriched pathways were identified using the “GSEA()” function with a P value of 0.05. Results were ranked by normalized enrichment score (NES). The top 5 significantly enriched pathways for each group (high- and low-risk) were visualized.

Immune cell infiltration analysis

To investigate the immune microenvironment associated with different risk groups, immune cell infiltration scores were obtained using MCP-counter (19). The infiltration matrix was normalized by Z-score transformation. Risk information was extracted from the previously defined risk model, and samples were grouped into high- and low-risk cohorts. Immune infiltration data were merged with the corresponding risk scores. For each immune cell type, violin plots combined with boxplots were generated using the “ggpubr” package to visualize differences between high- and low-risk groups. The Wilcoxon rank-sum test was used to assess statistical significance, and values above the 99th percentile were capped to reduce the effect of outliers.

Clinical correlation analysis

The association between the DRRG-derived risk score and clinicopathological characteristics was evaluated. Risk scores were integrated with clinical variables, including age (≤65 and >65 years), gender, tumor stage (I–IV), T stage (T1–4), N stage (N0–3), and M stage (M0 and M1). Boxplots were generated using the “ggpubr” R package.

Genomic and molecular feature analysis

Somatic mutation data were retrieved and analyzed using the “maftools” R package. Mutation landscapes were compared between high- and low-risk groups defined by the DRRG-derived prognostic model. TMB was calculated as the number of somatic mutations per megabase. TMB values were log₂-transformed and compared between high- and low-risk groups using the Wilcoxon rank-sum test. Distribution differences were further evaluated by Empirical Cumulative Distribution Function (ECDF) curves. Microsatellite instability (MSI) status [microsatellite stable (MSS), microsatellite instability-low (MSI-L), MSI-H] was integrated with risk scores, and intergroup differences were assessed by Kruskal-Wallis tests with pairwise comparisons. Stemness index scores (RNA-based RNAss) correlated with risk scores using Spearman correlation analysis. Scatter plots with density distributions were generated using “ggplot2” and “ggExtra”. Immunotherapy response (IPS) was predicted using the IPS retrieved from The Cancer Immunome Atlas (TCIA, https://tcia.at/).

Unsupervised clustering and functional enrichment analysis

Unsupervised consensus clustering was conducted using the “ConsensusClusterPlus” package to identify molecular subtypes. The clustering process was repeated 1,000 times to ensure stability, with a sample item consistency threshold of 0.8 and a feature consistency threshold of 1. The k-means algorithm was employed as the clustering method. DEGs between subtypes were identified using the “limma” package. Subsequently, Gene Ontology (GO) and Kyoto Encyclopedia of Genes and Genomes (KEGG) enrichment analyses were carried out to explore the functional differences between the subtypes (20,21).

SHAP model construction of recurrence and metastasis

To identify genes associated with CRC recurrence and metastasis, we constructed multiple machine learning (ML) classifiers based on different algorithms, using gene expression data from tumor samples with known recurrence or metastasis status. Sample grouping was extracted from metadata, and the data were split into training (70%) and testing (30%) sets using the “caret” package with stratified sampling. A panel of ML methods, including random forest, support vector machine (SVM), and XGBoost, was evaluated through repeated 5-fold cross-validation. ROC curves were plotted using the “pROC” package to evaluate classification performance, and the model with the highest AUC was selected for further interpretation. To explore the contribution of individual genes to the prediction of recurrence and metastasis, SHAP values were calculated using the “permshap” method. The resulting model interpretation was visualized with “shapviz” package, including bar plots (feature importance), beeswarm plots (feature effects), dependence plots, waterfall plots, and force plots for individual patients.

Statistical analysis

Statistical analysis was performed using R software version 4.3.0. Data were assumed to follow a normal distribution. For comparisons between two groups, a two-tailed unpaired Student’s t-test was applied. One-way analysis of variance (ANOVA) was used for comparisons involving more than two groups. Pearson correlation analysis was performed to assess the relationships between hub gene expression and immune cell infiltration levels. A P value <0.05 was considered statistically significant.

Results

Identification and characterization of prognostic DRRGs in CRC

Differential expression analysis was conducted on 395 tumor and 39 normal samples from the TCGA-COAD cohort. Based on the thresholds of adjusted P value <0.05 and |log₂ FC| >1, a total of 2,095 DEGs were identified, including 680 upregulated and 1,415 downregulated genes (Figure 1A). These DEGs were intersected with 276 DRRGs obtained from the GenDR database, comprising 220 positive-related and 56 negative-related genes. A total of 40 overlapping DRRGs were identified, including 29 positively and 11 negatively associated genes (Figure 1B; Table S1). To further identify DRRGs associated with CRC prognosis, univariate Cox regression analysis was performed to evaluate the association between each of the 40 DRRGs and OS. Using a significance threshold of P<0.05, a total of 11 prognostically relevant DRRGs were identified (Table 2).

Figure 1.

Characterization of prognostic DRRGs in CRC. (A) Volcano plot displaying 2,095 DEGs between tumor (n=395) and normal (n=39) samples from the TCGA-COAD cohort (adjusted P<0.05, |log₂ fold change | >1); (B) Venn diagram showing the intersection between DEGs and 276 DRRGs from the GenDR database, resulting in 40 overlapping genes; (C) box plots comparing expression levels of the 11 prognosis-related DRRGs (***, P<0.001); (D) bar plot showing the frequency of CNV events among the 11 DRRGs; (E) Manhattan plot illustrating chromosomal positions and CRC relevance of the 11 DRRGs; (F) forest plot displaying hazard ratios for the 11 DRRGs derived from univariate Cox regression analysis. CI, confidence interval; CNV, copy number variation; CRC, colorectal cancer; DEGs, differentially expressed genes; DRRGs, dietary restriction-responsive genes; FC, fold change; TCGA-COAD, The Cancer Genome Atlas-colon adenocarcinoma.

Table 2. Univariate Cox regression and survival analysis of DEGs.

Gene symbol	HR	95% CI	P value	Kaplan-Meier
RGS16	1.40	1.17–1.67	1.82×10−4	6.53×10−7
CD163	1.15	1.0–1.30	1.80×10−2	4.23×10−5
CRYAB	1.24	1.10–1.41	6.27×10−4	4.82×10−5
MGP	1.13	1.02–1.24	1.41×10−2	1.04×10−4
PLIN4	1.28	1.13–1.45	1.05×10−4	2.83×10−4
DUSP1	1.25	1.06–1.46	6.46×10−3	7.05×10−4
FOXD2	0.77	0.64–0.92	4.12×10−3	9.93×10−4
NFE2L3	0.82	0.70–0.96	1.25×10−2	1.87×10−3
SLC13A2	0.84	0.73–0.96	1.11×10−2	1.98×10−3
TSC22D3	1.27	1.03–1.56	2.70×10−2	2.83×10−3
LY6E	1.16	1.03–1.30	1.56×10−2	4.28×10−3

CI, confidence interval; DEGs, differentially expressed genes; HR, hazard ratio.

Differential expression analysis of the 11 prognosis-related DRRGs between tumor and normal tissues revealed that RGS16, NFE2L3, and LY6E were significantly upregulated in tumor samples, while CD163, CRYAB, MGP, PLIN4, DUSP1, FOXD2, SLC13A2, and TSC22D3 were markedly downregulated (Figure 1C). In addition, CNV analysis was conducted to examine genomic alterations in these DRRGs. Among them, SLC13A2 exhibited the most pronounced CNV alterations (Figure 1D). A Manhattan plot displayed the chromosomal locations of the 11 DRRGs and their relevance to CRC, with NFE2L3 on chromosome 7 showing the strongest association (Figure 1E). Forest plot illustrated the hazard ratios of these genes, indicating that NFE2L3, SLC13A2, and FOXD2 were associated with reduced CRC risk (Figure 1F).

Construction of the prognostic model based on 11 prognosis-related DRRGs

The prognostic risk model was constructed using LASSO-Cox regression based on the 11 prognosis-related DRRGs. Cross-validation identified the optimal lambda value that minimized model deviance (Figure 2A). A total of four genes—RGS16, PLIN4, SLC13A2, and FOXD2—were retained in the final model, and their corresponding coefficients are shown in Figure 2B. Based on these selected genes, a multivariate Cox regression model was established, and individual risk scores were calculated. Patients were stratified into high- and low-risk groups according to the median risk score. Kaplan-Meier survival analysis demonstrated that patients in the high-risk group had significantly worse OS than those in the low-risk group in both the training and validation cohorts (Figure 2C). Time-dependent ROC analysis for 1-year OS yielded an AUC of 0.723 in the training set, indicating favorable predictive performance (Figure 2D). Risk score distribution and survival status plots revealed that higher risk scores were associated with shorter survival time (Figure 2E). Similar results were observed in the validation cohort, confirming the robustness and generalizability of the risk model (Figure 2F-2H).

Figure 2.

Construction and validation of a DRRG-based prognostic model for colorectal cancer. (A) A LASSO coefficient profiles of the 11 prognosis-related DRRGs and cross-validation curve for optimal lambda selection; (B) four genes retained in the final model; (C) Kaplan-Meier survival curves showing significant survival differences between high- and low-risk groups in the training set; (D) time-dependent ROC curve for 1-year survival in the training cohort; (E) risk score distribution and corresponding survival status of patients in the training cohort; (F-H) Kaplan-Meier curve (F), ROC curve (G), and risk distribution plot (H) for the validation cohort. AUC, area under the curve; DRRG, dietary restriction-responsive gene; LASSO, least absolute shrinkage and selection operator; ROC, receiver operating characteristic.

Evaluation of the risk model as an independent prognostic factor

A nomogram was constructed by integrating the risk score and key clinical variables to predict 1-, 3-, and 5-year OS (Figure 3A). Calibration curves demonstrated excellent concordance between predicted and observed outcomes across all time points, with a C-index of 0.793 (95% CI: 0.729–0.858), indicating high predictive accuracy (Figure 3B). To further evaluate its clinical utility, the nomogram was compared with traditional clinical indicators using ROC curve analysis. The nomogram achieved the highest predictive performance, with an AUC of 0.865, outperforming other clinical variables (Figure 3C). The expression patterns of the four genes included in the prognostic signature were further compared between high- and low-risk groups. PLIN4 and RGS16 were significantly upregulated in the high-risk group, while FOXD2 and SLC13A2 showed higher expression in the low-risk group (Figure 3D). These findings were consistent with the results from the univariate Cox regression analysis, further supporting their prognostic relevance.

Figure 3.

Independent prognostic value and clinical applicability of the DRRG-based risk model. (A) Nomogram integrating the risk score and clinical features for predicting 1-, 3-, and 5-year overall survival; (B) calibration plots demonstrating consistency between predicted and actual survival probabilities at 1, 3, and 5 years; (C) ROC curve comparison of nomogram and clinical variables; (D) the expression patterns of the four genes in high- and low-risk groups; (E) GO terms significantly enriched in the high- and low- risk group; (F) KEGG pathways significantly enriched in the high- and low- risk group. *, P<0.05; **, P<0.01; ***, P<0.001. AUC, area under the curve; DRRG, dietary restriction-responsive gene; GO, Gene Ontology; KEGG, Kyoto Encyclopedia of Genes and Genomes; M, metastasis; N, node; OS, overall survival; ROC, receiver operating characteristic; T, tumor.

To elucidate the potential biological mechanisms associated with the prognostic signature, GO and KEGG pathway enrichment analyses were performed using GSEA. In the high-risk group, GO terms were predominantly enriched in ECM-related functions, including external encapsulating structure organization, collagen-containing ECM, and ECM structural constituents (Figure 3E). In contrast, the low-risk group showed enrichment in mitochondrial-related biological processes (BP), such as aerobic respiration, mitochondrial respiratory chain complex assembly, and ribosomal structural components (Figure 3E).

Similarly, KEGG pathway analysis revealed that the high-risk group was enriched in pathways related to immune and inflammatory responses, including cytokine-cytokine receptor interaction, ECM-receptor interaction, and cell adhesion molecules (CAMs) (Figure 3F). In contrast, the low-risk group exhibited significant enrichment in metabolic pathways, such as the citrate cycle [tricarboxylic acid (TCA) cycle], oxidative phosphorylation, and peroxisome-related metabolism (Figure 3F). These findings suggest that high-risk CRC patients are characterized by immune-ECM remodeling, while low-risk patients exhibit enhanced mitochondrial activity and metabolic homeostasis.

Immune cell infiltration analysis between high- and low-risk groups

To investigate differences in the immune microenvironment between risk subgroups, immune cell infiltration was quantified using the MCP-counter algorithm. Several immune components, including fibroblasts, monocytic lineage cells, cytotoxic lymphocytes, NK cells, myeloid dendritic cells, T cells, and endothelial cells, were significantly elevated in the high-risk group. Although CD8⁺ T cells and B lineage cells also showed an increasing trend, the differences were not statistically significant. These findings suggest that the high-risk group is characterized by a more inflamed and immunologically active tumor microenvironment (Figure 4).

Figure 4.

Immune cell infiltration analysis between high- and low-risk groups.

Correlation between risk score and clinicopathological characteristics

The relationship between the risk score and clinicopathological characteristics was examined in the TCGA cohort. No significant differences were observed with age (P=0.07), gender (P=0.15), or metastasis status (P=0.91, Figure 5A-5C). In contrast, the risk score was positively correlated with T stage, with patients in T3–4 exhibiting significantly higher scores than those in T1–2 (Figure 5D). Significant differences were also observed across N stages, showing a stepwise increase in risk score from N1 to N3 (Figure 5E). A progressive elevation in risk score was noted with advancing clinical stage, with stage III cases displaying significantly higher scores compared with stage I–II. Interestingly, a slight decrease was observed in stage IV patients (Figure 5F). This decline may be attributable to the relatively small number of stage IV cases in the TCGA cohort, as well as potential heterogeneity in treatment exposure or sample selection bias. Collectively, these findings indicate that the DRRG-based risk score is strongly associated with tumor progression, while remaining largely independent of demographic characteristics.

Figure 5.

Association of the DRRG-based risk score with clinicopathological characteristics. (A) Age (≤65 vs. >65 years); (B) gender; (C) M status; (D) T stage; (E) N stage; (F) clinical stage. *, P<0.05; **, P<0.01; ***, P<0.001; ns, not significant. DRRG, dietary restriction-responsive gene; M, metastasis; N, node; T, tumor.

Association of risk score with TMB, MSI, stemness and IPS

Distinct mutation landscapes were observed between the high- and low-risk groups (Figure 6A,6B). Both groups exhibited frequent mutations in APC, but high-risk tumors demonstrated slightly higher mutation frequencies of oncogenes such as TP53 and TTN, whereas low-risk tumors had a greater prevalence of PIK3CA and MUC16 alterations. TMB analysis showed that high-risk patients had significantly higher tumor mutation burden compared with low-risk patients (P<0.001, Figure 6C). ECDF plots further confirmed the shift toward higher TMB in the high-risk group (Figure 6D).

Figure 6.

Genomic and molecular correlates of the DRRG-based risk score. (A) Oncoplot showing the top 20 mutated genes in the high-risk group; (B) oncoplot showing the top 20 mutated genes in the low-risk group; (C) comparison of TMB between high- and low-risk groups; (D) ECDF of TMB in the two groups; (E) risk score distribution across MSI subtypes (MSS, MSI-L, MSI-H); (F) correlation between risk score and RNAss; (G) immunotherapy prediction based on the Immunophenoscore. ***, P<0.001; ns, not significant. DRRG, dietary restriction-responsive gene; ECDF, empirical cumulative distribution function; MSI, microsatellite instability; MSI-H, MSI-high; MSI-L, MSI-low; MSS, microsatellite stable; RNAss, RNA-based stemness index; TMB, tumor mutation burden.

Risk score was also associated with MSI status. Patients with MSI-H tumors exhibited significantly elevated risk scores relative to MSS or MSI-L groups (P<0.001, Figure 6E). A negative correlation was identified between risk score and stemness index (RNAss), with higher risk scores corresponding to lower stemness values (Spearman r=−0.267, P<0.001, Figure 6F).

To further explore the potential clinical utility of the DRRG-based risk model, IPS was predicted using the IPS obtained from TCIA. Four immune checkpoint therapy scenarios (CTLA-4⁻/PD-1⁻, CTLA-4⁻/PD-1⁺, CTLA-4⁺/PD-1⁻, and CTLA-4⁺/PD-1⁺) were analyzed. As shown in Figure 6G, IPS values were significantly lower in the high-risk group under CTLA-4⁻/PD-1⁻ (P<0.001) and CTLA-4⁺/PD-1⁻ (P<0.001) conditions, suggesting that low-risk patients may be more likely to benefit from CTLA-4-related blockade. In contrast, no significant differences were observed between risk groups under PD-1⁺ scenarios.

Subtype classification based on prognostic DRRGs

To further investigate the association between prognostic DRRGs and COAD, unsupervised consensus clustering was performed, stratifying patients into two distinct molecular subtypes: Cluster A (n=395) and Cluster B (n=807) (Figure 7A-7C). Kaplan-Meier analysis revealed significantly poorer OS in Cluster B compared to Cluster A (Figure 7D). Differential expression analysis between the two clusters identified 37 DEGs, among which SLC13A2 was notably downregulated (Figure 7E; Table S2).

Figure 7.

Subtype classification and functional enrichment. (A-C) Consensus clustering results based on the expression of 4 prognostic DRRG; (D) Kaplan-Meier curves comparing overall survival between the two clusters; (E) Volcano plot showing DEGs between Clusters; (F) GO enrichment analysis of DEGs; (G) KEGG pathway analysis revealed enrichment. CDF, cumulative distribution function; DEGs, differentially expressed genes; DRRG, dietary restriction-responsive gene; GO, Gene Ontology; KEGG, Kyoto Encyclopedia of Genes and Genomes.

GO and KEGG enrichment analyses were subsequently conducted to elucidate the functional roles of these DEGs. GO analysis revealed significant functional distinctions between the two molecular subtypes defined by prognostic DRRGs (Figure 7F). Enriched BP included organic anion transport, sulfate and oxalate transmembrane transport, retinol metabolism, and terpenoid biosynthetic processes. These pathways are primarily involved in maintaining metabolic homeostasis and detoxification in the intestinal epithelium. Enrichment in cellular components (CC) such as the apical part of cell, zymogen granules, Golgi lumen, brush border membrane, and apical plasma membrane highlights the functional polarization of enterocytes and secretory activity of epithelial cells. These structures are critical for nutrient absorption, protein sorting, and secretion of digestive enzymes and mucins. In terms of molecular functions (MF), the DEGs were significantly enriched in transmembrane transporter activity, particularly involving organic anions, sulfate, oxalate, and bicarbonate. These transporter functions are essential for epithelial barrier function, maintaining pH and osmotic balance, and facilitating communication between the gut environment and host tissues. KEGG pathway analysis demonstrated that these DEGs were primarily involved in fat digestion and absorption, pancreatic secretion, retinol metabolism, and biosynthesis of cofactors, suggesting metabolic reprogramming between the two DRRG-defined subtypes (Figure 7G).

SHAP-based interpretation of DRRGs in metastasis and recurrence

To evaluate the predictive value of DRRGs for CRC metastasis and recurrence, a series of ML classifiers were trained on the liver metastasis dataset (GSE41258). Multiple algorithms were compared through repeated cross-validation, and the XGBoost model demonstrated the highest classification performance with an AUC of 0.855, indicating robust predictive capability for CRC liver metastasis (Figure 8A). SHAP analysis was subsequently employed to interpret model predictions. In the SHAP bar plot, MGP exhibited the highest mean SHAP value, identifying it as the most influential feature in metastasis prediction (Figure 8B). The beeswarm plot provided a comprehensive view of gene importance and SHAP value distribution across all samples, highlighting MGP, FOXD2, and CRYAB as key contributors (Figure 8C). The dependence plots illustrated the relationship between gene expression levels and SHAP values, reflecting both main effects and interactions. Notably, higher expression of MGP was associated with increased SHAP values, suggesting that elevated MGP expression may promote metastasis. Similarly, FOXD2 and CRYAB showed substantial SHAP variation corresponding to expression changes, indicating their predictive relevance. Interaction effects among MGP, FOXD2, and CRYAB were also observed (Figure 8D). A force plot was used to visualize the contribution of individual genes for a representative sample. CRYAB showed a negative impact on metastasis risk (expression =7.56, SHAP =−0.167), while FOXD2 had a positive impact (expression =5.93, SHAP =+0.0758) (Figure 8E). The global contribution of all gene features was further summarized in a waterfall plot, where CRYAB again emerged as a key protective factor (Figure 8F).

Figure 8.

SHAP-based interpretation of the predictive model for colorectal cancer liver metastasis. (A) ROC curves comparing multiple machine learning classifiers; (B) bar plot of SHAP values showing the average contribution of each gene to the prediction output; (C) Beeswarm plot illustrating the SHAP value distribution of top contributing genes across all samples; (D) SHAP dependence plot demonstrating the relationship between gene expression and SHAP values, as well as gene-gene interactions; (E) force plot displaying the contribution of each gene to the metastasis prediction; (F) waterfall plot summarizing the overall impact of all genes. DTS, decision tree; GBM, gradient boosting machine; KNN, K-nearest neighbors; RF, random forest; ROC, receiver operating characteristic; SHAP, Shapley Additive Explanations; PLS, partial least squares; SVM, support vector machine.

Additionally, SHAP analyses were independently performed in two CRC recurrence cohorts (GSE17538 and GSE39582) as well. However, the predictive performance was notably lower in these datasets, with maximum AUC values of 0.647 and 0.590, respectively (Figures S1,S2). The relatively modest predictive accuracy in recurrence datasets therefore likely reflects dataset-specific heterogeneity and the influence of clinical confounders, such as surgical techniques, adjuvant therapy, and follow-up duration, which are not captured by gene expression profiles alone. These observations suggest that while DRRGs provide robust prognostic value in the context of metastasis, recurrence may involve additional biological and clinical determinants beyond DRRGs. Consequently, the role of DR-responsive pathways in recurrence requires further validation in prospective cohorts with standardized clinical annotations.

Discussion

In this study, we provide a comprehensive analysis of DRRGs in CRC and demonstrate their significance in tumor progression, the tumor microenvironment, and patient outcomes. Our findings highlight mechanistic links between nutritional interventions and colorectal tumor biology, showing how conserved stress-response genes shape cancer development.

We identified a subset of DRRGs whose dysregulation in CRC is significantly associated with patient prognosis. Among them, NFE2L3, a transcription factor, was markedly upregulated in CRC and associated with poor outcomes. NFE2L3 promotes tumor proliferation by integrating NF-κB signaling and cell cycle pathways (22). In mouse models, Nfe2l3 deletion reduced tumor burden by alleviating chronic inflammation and enhancing regulatory T cell infiltration (23). These results suggest that NFE2L3 fosters a pro-tumor microenvironment. Elevated NFE2L3 expression in human CRC correlates with aggressive phenotypes (24), and our findings support its role as a nutrient-sensitive oncogene. Given that DR suppresses NF-κB-mediated inflammation, its protective effects may partially involve NFE2L3 downregulation (14). In contrast, some DRRGs appear to act as tumor suppressors. FOXD2, a forkhead box transcription factor, was significantly downregulated in CRC and associated with poor prognosis. FOXD2 suppresses oncogenic programs and inhibits Wnt/β-catenin signaling, functioning as a chromatin regulator that restricts tumor growth (25). Its loss may reflect a metabolic shift away from the transcriptional program maintained under DR. Notably, FOX family genes are modulated by nutrient-sensing pathways, supporting FOXD2’s role as a DR-mimetic suppressor (10). Restoring FOXD2 activity may replicate the anti-cancer benefits of nutrient restriction. Another prognostically relevant DRRG is DUSP1, which was significantly reduced in CRC. DUSP1 deactivates MAPKs, thereby curbing inflammatory and oxidative stress responses (26). While context-dependent, its tumor-suppressive role has been widely reported. Consistent with this, low DUSP1 levels in our cohort were linked to worse outcomes. Loss of DUSP1 may relieve a key restraint on oncogenic inflammation. Thus, therapeutic strategies that enhance DUSP1 activity—potentially through DR-mimicking interventions—may provide clinical benefit in CRC.

Pathway enrichment analysis revealed that high-risk CRC tumors exhibited upregulation of genes involved in ECM organization, cell adhesion, and cytokine signaling, consistent with a pro-metastatic phenotype. Notably, RGS16, one of our model genes, was significantly elevated in high-risk tumors. Though classically involved in GPCR and chemokine regulation, RGS16 has been linked to enhanced tumor survival, invasion, and poor prognosis in CRC (27). RGS16 can enhance survival pathways and has been linked to increased migration and invasion in some contexts (28). In pancreatic cancer, loss of RGS16 actually restrained metastasis, hinting that RGS16 normally inhibits anti-migratory signals (28). In CRC, recent evidence suggests RGS16 expression is significantly higher in tumors than normal tissue and correlates with advanced stage (29). Functionally, RGS16 may promote tumor cell survival by blocking stress-induced JNK/p38 MAPK activation and facilitating a permissive niche for tumor expansion (29). Its association with ECM pathways in our analysis could reflect a role in regulating integrin or chemokine signals that orchestrate matrix deposition and remodeling. The observed association between RGS16 and integrin or chemokine pathways suggests its role in promoting a fibrotic, cytokine-rich microenvironment that facilitates tumor invasion and immune suppression.

Conversely, the low-risk group showed enrichment of mitochondrial and metabolic pathways, consistent with the metabolic adaptations induced by DR. Notably, SLC13A2, a protective gene encoding a sodium-coupled citrate transporter, was downregulated in tumors, and its higher expression correlated with improved survival. To date, no studies have specifically investigated the role of SLC13A2 in CRC, rendering our findings a novel insight into its potential tumor-suppressive function. Although its role in CRC is not fully defined, SLC13A2 may help preserve metabolic homeostasis and support oxidative metabolism—a state linked to slower tumor growth and better differentiation. DR is known to activate AMPK and sirtuins, enhancing mitochondrial respiration and reducing anabolic drive (10). Accordingly, low-risk tumors in our study may retain a calorie-restricted-like metabolic profile, characterized by active TCA cycles, reduced lactate production, and improved energy efficiency, thereby limiting tumor aggressiveness.

Moreover, PLIN4, a high-risk gene in our prognostic model, was significantly upregulated in colorectal tumors and associated with poorer survival. PLIN4 encodes a lipid droplet-associated protein involved in triglyceride storage and lipid metabolism, primarily studied in adipose and muscle tissues (30). However, its role in CRC remains unexplored. Given that aberrant lipid accumulation promotes tumor growth, oxidative stress resistance, and immune evasion, PLIN4 overexpression may indicate increased lipid storage capacity and metabolic flexibility, enabling tumor cells to survive under energy stress (31). Lipid droplets also modulate tumor-immune crosstalk by influencing inflammatory signaling and immune cell recruitment (32). The observed link between high PLIN4 expression and inflammatory pathway enrichment in our high-risk group suggests a potential role for PLIN4 in fostering a pro-tumorigenic microenvironment. These findings position PLIN4 as a putative oncogenic driver and metabolic vulnerability in CRC, meriting further mechanistic investigation.

DR also modulates tumor immunity, often enhancing anti-tumor responses. Caloric restriction reduces immunosuppressive myeloid-derived suppressor cells while increasing cytotoxic T cell activity (14). Among DRRGs, MGP emerged as the top metastasis-driving gene in our model. Traditionally associated with calcification, MGP is now implicated in immune evasion: it activates NF-κB signaling and upregulates PD-L1, leading to CD8⁺ T cell exhaustion and metastasis (33). Its inhibition suppresses liver metastasis and improves response to PD-1 blockade, suggesting that MGP links metabolic excess to immune suppression. Our findings position MGP as a promising target where dietary or metabolic interventions may synergize with immunotherapy. Overall, DRRG profiles appear to shape immune landscapes, with high-risk patterns fostering chronic inflammation and suppression, while DR-like profiles promote immune competence.

The DRRG-based risk score showed significant associations with both clinicopathological characteristics and molecular features of CRC. At the molecular level, high-risk tumors exhibited elevated TMB, which has been linked to increased genomic instability and shapes tumor immunogenicity (34). Consistently, risk scores were significantly higher in MSI-H tumors compared with MSS or MSI-L cases, in line with prior descriptions of the MSI-H/CMS1 subtype—characterized by hypermutation and dense immune infiltration—and with its proven sensitivity to PD-1 blockade (35,36).

Interestingly, a negative correlation was observed between risk score and RNAss. While high stemness has previously been associated with poor prognosis across multiple cancer types, the inverse relationship in our study may reflect the unique metabolic and immune remodeling captured by DRRGs (37). This counterintuitive finding may reflect that DRRGs capture metabolic remodeling resembling DR, where reduced glycolysis and increased metabolic efficiency can limit the maintenance of stem-like states. Recent studies of fasting and fasting-mimicking diets in CRC models have shown that dietary interventions can inhibit aerobic glycolysis and remodel cellular metabolism (38,39). Such metabolic constraints may, in turn, attenuate stem-like transcriptional programs, suggesting a possible mechanistic explanation for the reduced stemness observed in high-risk tumors. Collectively, these findings indicate that the DRRG signature integrates clinical progression, genomic instability, immune activity, and differentiation states, thereby providing a comprehensive framework for CRC prognosis.

The IPS analysis further highlighted potential clinical implications of the DRRG-based risk model. Low-risk patients consistently exhibited higher IPS values under CTLA-4⁻/PD-1⁻ and CTLA-4⁺/PD-1⁻ conditions, suggesting a greater likelihood of responding to CTLA-4-targeted immune checkpoint blockade. This finding implies that the DRRG signature not only stratifies prognosis but may also capture immunological differences relevant to treatment selection. The absence of significant differences in PD-1⁺ scenarios is consistent with previous reports that PD-1 response is strongly influenced by neoantigen load and immune microenvironment features that may not be fully represented by DRRGs (36,40). Taken together, these results suggest that DRRG-based stratification could help identify patients who are more likely to benefit from CTLA-4-oriented immunotherapeutic strategies.

The DRRG-derived gene signature developed in this study holds potential as a clinically actionable prognostic tool. Patients with a high-risk DRRG expression pattern—characterized by low levels of tumor-suppressive genes such as FOXD2 and DUSP1, and elevated levels of oncogenic or immunosuppressive genes like MGP and RGS16, may be at increased risk of recurrence and metastasis. These patients could benefit from intensified adjuvant therapies or early initiation of immune-based treatment. Moreover, several DRRGs, including MGP and RGS16, are detectable via immunohistochemistry or circulating assays, offering feasible options for biomarker-based patient stratification. Prior studies have supported their prognostic relevance in CRC, reinforcing the clinical translatability of our findings.

Our analysis also highlights DRRG-related pathways as promising therapeutic targets. Many DRRGs are regulated by nutrient-sensing mechanisms, particularly the IGF-1/insulin axis, which is suppressed under DR (13,41). Similarly, targeting the mTOR pathway, which is often hyperactive in CRC, may reproduce key aspects of DR by enhancing autophagy and reducing anabolic signaling (42,43). The observed enrichment of ECM and cytokine signaling in high-risk tumors further suggests that combining mTOR or PI3K inhibitors with anti-inflammatory agents could achieve therapeutic synergy (44).

Our findings also suggest that DRRGs contribute to shaping the tumor immune landscape. Dietary interventions themselves—such as intermittent fasting—are being tested clinically and may help restore immune competence by reducing myeloid-derived suppressor cells and enhancing effector T cell responses (39,45). Our data support these strategies, showing that DR-like transcriptional programs may sensitize tumors to immune-mediated clearance.

In summary, DRRGs represent a molecular interface between metabolism, inflammation, and tumor progression in CRC. Their dysregulation under nutrient-rich conditions appears to favor proliferation, immune evasion, and stromal activation, while restoration of DR-like states may reverse these features. Our identification of key DRRGs—such as MGP, NFE2L3, FOXD2, and DUSP1—alongside newly implicated genes like SLC13A2 and PLIN4, offers valuable mechanistic insights and potential therapeutic avenues in CRC. Despite the strengths of our integrative multi-cohort analysis, several limitations should be acknowledged. First, the study is based primarily on retrospective transcriptomic data from public databases, which may be subject to inherent biases in sample selection, clinical annotations, and batch effects despite our correction procedures. Second, while the DRRG-derived prognostic model showed robust performance in training and validation sets, prospective validation in large, independent clinical cohorts is necessary to confirm its predictive utility and generalizability. Thirdly, the mechanistic roles of key DRRGs, such as SLC13A2 and PLIN4, remain largely unexplored in CRC; functional in vitro and in vivo experiments are needed to clarify their biological relevance and potential as therapeutic targets. Finally, the TCGA cohort predominantly consists of older patients, and thus the applicability of the model to early-onset CRC remains uncertain. Given the rising incidence and distinct biology of early-onset CRC, further validation in independent cohorts enriched with younger patients is warranted. Future studies should dissect the regulatory relationships among these genes and validate strategies—dietary, pharmacologic, or genetic—that restore DRRG-mediated tumor suppression. Leveraging DR biology thus offers a compelling framework for improving CRC management through integrated metabolic and immune modulation.

Conclusions

In this study, we identified DRRGs that define prognostic subtypes and predict metastasis in CRC. The DRRG-based signature stratifies patients into distinct risk groups. These findings highlight DRRGs as potential biomarkers and therapeutic targets and support the rationale for DR-mimicking strategies in clinical management of CRC.

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. The study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments.