Prognostic risk modeling based on integrated multi-omics analysis identifies CRY2 as a key regulator in tumor immunity and patient survival in colorectal cancer

Abstract

Background

Colorectal cancer (CRC) exhibits substantial metabolic heterogeneity. This study developed a robust prognostic signature integrating ferroptosis- and lipid metabolism-related genes to investigate the role of CRY2 in CRC progression.

Methods

Transcriptomic and clinical data from the TCGA-COAD and GSE39582 cohorts were analyzed. Weighted gene co-expression network analysis (WGCNA) was performed to identify disease-associated gene modules. A machine learning framework was subsequently applied to construct and optimize the prognostic model, with the combination of forward stepwise Cox regression (StepCox) and Random Survival Forest demonstrating the best predictive performance. The tumor immune microenvironment was characterized using CIBERSORT and TIDE. The biological function of CRY2 was validated through siRNA-mediated knockdown, functional assays, and a murine xenograft model.

Results

The ferroptosis- and lipid metabolism-related RiskScore was identified as an independent predictor of overall survival (HR > 1.1, p < 0.001). Patients in the high-risk group showed poorer overall survival and an immunosuppressive tumor microenvironment (TME) characterized by increased regulatory T cells (Tregs) and Th2 cells, reduced CD4⁺ T-cell infiltration, and higher TIDE scores, suggesting immunotherapy resistance. Among the signature genes, CRY2 was identified as a key regulator associated with CRC progression. In vitro, CRY2 knockdown inhibited CRC cell proliferation and migration while inducing G1-phase cell-cycle arrest. In vivo, silencing CRY2 significantly suppressed xenograft tumor growth and reduced Ki-67 expression.

Conclusion

This study developed and validated a ferroptosis- and lipid metabolism-related prognostic signature that accurately predicts survival outcomes and immune characteristics in CRC. Furthermore, CRY2 was identified as a critical regulator of tumor growth.

Background

Colorectal cancer (CRC) remains a major global health burden, ranking as the third most commonly diagnosed malignancy and the second leading cause of cancer-related mortality worldwide [1,2]. Despite significant advances in screening strategies and multimodal treatments involving surgical resection, fluorouracil-based chemotherapy, and targeted biological agents, the prognosis of CRC patients remains highly heterogeneous [3,4], particularly among those with advanced or metastatic disease, for whom long-term survival remains poor.

The Tumor-Node-Metastasis (TNM) staging system continues to serve as the cornerstone of prognostic assessment and treatment planning in clinical practice. However, this anatomical classification system does not adequately reflect the molecular and biological heterogeneity of CRC [5], resulting in considerable variability in clinical outcomes and therapeutic responses among patients with the same stage of disease. Although molecular classification systems, such as the Consensus Molecular Subtypes (CMS), have substantially advanced our understanding of CRC biology, their widespread clinical implementation remains limited by technical complexity, cost, and challenges in standardization [6,7]. Consequently, patients with identical TNM stages often exhibit drastically different clinical trajectories and therapeutic responses. Accumulating evidence reveals that gastrointestinal carcinogenesis is heavily driven by complex epigenetic anomalies and the step-wise inactivation of pivotal tumor suppressor genes [8]. Therefore, there is an urgent and unmet need to transition toward multi-dimensional molecular stratification frameworks to optimize early screening and precision intervention strategies for CRC patients [9].

The tumor microenvironment (TME), a highly dynamic and complex ecosystem comprising malignant cells, infiltrating immune cells, cancer-associated fibroblasts, and the extracellular matrix, plays a pivotal role in the pathogenesis and progression of CRC [10,11]. Emerging evidence suggests that the interplays between tumor cells and the TME are driven, in part, by metabolic reprogramming, which shapes both tumor progression and antitumor immunity [12]. Metabolic plasticity enables CRC cells to adapt to the nutrient-deprived and hypoxic conditions of the TME [13]. Among the metabolic alterations observed in CRC, dysregulated lipid metabolism has emerged as a hallmark of tumor progression [14,15]. CRC cells exhibit extensive lipid metabolic reprogramming, characterized by enhanced fatty acid synthesis, increased uptake of exogenous lipids, and altered lipid utilization to meet the bioenergetic and biosynthetic demands of rapid proliferation [15]. Beyond serving as energy substrates and structural components of cellular membranes, lipids also function as bioactive signaling molecules that regulate cellular communication within the TME. Aberrant lipid accumulation has been shown to impair the cytotoxic activity of CD8⁺ T cells while promoting the recruitment and activation of immunosuppressive regulatory T cells and myeloid-derived suppressor cells. These changes establish an immunosuppressive microenvironment that facilitates tumor immune evasion and metastasis [16].

Metabolic plasticity allows tumor cells to adapt to the nutrient-deprived and hypoxic conditions characteristic of the TME. In CRC, this malignant phenotype is tightly orchestrated by various classical pathways; for instance, metabolic cues such as GLP-1 receptor activation can profoundly modulate CRC cellular growth and survival via the downstream PI3K/AKT/mTOR signaling axis [17]. Among the various metabolic alterations, the dysregulation of lipid metabolism. In the context of lipid metabolic dysregulation, ferroptosis has emerged as a distinct form of regulated cell death with considerable therapeutic relevance [18]. Unlike apoptosis, necrosis, or autophagy, ferroptosis is characterized by the iron-dependent accumulation of lipid peroxides to lethal levels [19,20]. Its execution is tightly coupled to lipid metabolism, as the peroxidation of polyunsaturated fatty acids (PUFAs) in membrane phospholipids represents the central biochemical event driving ferroptotic cell death [21]. Key metabolic enzymes, such as acyl-CoA synthetase long-chain family member 4 (ACSL4), promote ferroptosis by facilitating the incorporation of oxidizable PUFAs into membrane phospholipids [22]. Conversely, antioxidant defense systems, particularly the system Xc-/glutathione peroxidase 4 (GPX4) axis, protect tumor cells from ferroptosis-induced damage [23]. In CRC, pharmacological induction of ferroptosis has emerged as a promising therapeutic strategy, particularly for overcoming resistance to conventional apoptosis-inducing therapies [24,25]. However, malignant cells can evade ferroptotic cell death through extensive metabolic reprogramming, including the remodeling of lipid metabolic pathways and reinforcement of antioxidant defenses [26]. These observations suggest that the molecular interplay between lipid metabolism and ferroptosis is likely to influence tumor progression, immune regulation, and clinical outcomes. Nevertheless, the prognostic significance of genes co-regulating these interconnected pathways, as well as their collective impact on the tumor immune microenvironment, remains insufficiently understood. To address this knowledge gap, this study integrated ferroptosis- and lipid metabolism-related co-regulated genes to characterize a metabolic-immune crosstalk axis in CRC, linking metabolic dysregulation to an immunosuppressive TME and immunotherapy resistance.

The rapid development of high-throughput sequencing technologies and the availability of large-scale public genomic resources, such as The Cancer Genome Atlas (TCGA) and the Gene Expression Omnibus (GEO), have transformed cancer research by enabling comprehensive, systems-level investigations of tumor biology [27]. Among the computational approaches developed for analyzing these datasets, weighted gene co-expression network analysis (WGCNA) has emerged as a robust method for identifying biologically relevant gene modules and hub genes associated with disease phenotypes [28]. Unlike conventional differential expression analyses that evaluate genes individually, WGCNA captures coordinated transcriptional programs, providing deeper insights into the molecular networks underlying tumor progression. In parallel, the application of machine learning has substantially advanced prognostic modeling in oncology [29,30]. Conventional approaches based on a single statistical algorithm, such as Cox regression or Lasso approaches, are often susceptible to overfitting and lack reproducibility across independent cohorts. By contrast, integrative machine learning frameworks that systematically evaluate multiple feature-selection and survival prediction algorithms can improve model robustness, predictive accuracy, and generalizability.

Based on this, the present study employed an integrated multi-omics framework to develop a ferroptosis- and lipid metabolism-related prognostic signature for CRC. Differentially expressed genes identified from large-scale transcriptomic datasets were integrated with curated sets of ferroptosis and lipid metabolism-related genes. WGCNA was subsequently performed to identify clinically relevant co-expression modules. To optimize predictive performance, a comprehensive machine learning workflow evaluated 101 combinations of feature-selection and survival modeling algorithms using cross-validation, resulting in a consensus prognostic signature that was further validated in an independent external cohort. Beyond risk stratification, this study characterized the TME to elucidate the biological basis of the model, revealing that high-risk patients exhibit an immunosuppressive phenotype associated with immune exclusion and a reduced likelihood of responding to immunotherapy.

The integrative analysis identified CRY2 as a key component of the prognostic signature. Although CRY2 is classically recognized as a core regulator of the circadian clock, its role in cancer metabolism and tumor immunity remains poorly understood [31]. Functional validation through both in vitro and in vivo experiments demonstrated that CRY2 silencing inhibited cancer cell proliferation and tumor growth by inducing G1-phase cell-cycle arrest. Collectively, this study establishes a robust prognostic tool for predicting patient survival and immune characteristics while identifying CRY2 as a critical regulator linking circadian rhythm, metabolic reprogramming, and cell-cycle progression in CRC. These findings highlight CRY2 as a promising prognostic biomarker and potential therapeutic target.

Methods

Data acquisition and preprocessing

Transcriptomic profiles and corresponding clinical data for patients with COAD were obtained from the TCGA portal (https://portal.gdc.cancer.gov/) and the GEO repository (GSE39582). The TCGA-COAD cohort served as the discovery cohort for model development, whereas GSE39582 was used as an independent external validation cohort. To ensure data quality and comparability across platforms, standardized preprocessing procedures were performed. For the TCGA-COAD cohort, raw RNA sequencing counts were normalized using the variance stabilizing transformation (VST) implemented in the DESeq2 package, followed by log₂(count + 1) transformation. For the GSE39582 dataset, raw Affymetrix CEL files were background corrected, normalized, and summarized using the Robust Multi-chip Averaging (RMA) algorithm. Following data integration, batch effects arising from differences in sequencing platforms were corrected using the ComBat algorithm implemented in the sva R package, with the dataset source included as a covariate in the model. Patients lacking complete overall survival (OS) information were excluded, and only those with complete clinical and survival data were included in downstream survival analyses. The human data involved in this study were sourced from public databases (TCGA and GEO) and had previously received ethical approval from the original data collection centres. The use of these de-identified and publicly available datasets does not require additional ethical approval or informed consent. All procedures involving human data in this study were conducted in accordance with the principles of the Declaration of Helsinki (2013 edition, https://www.wma.net/policies-post/wma-declaration-of-helsinki/).

Identification of candidate genes

To identify dysregulated genes involved in the interplay between ferroptosis and lipid metabolism, differential expression analysis was performed on the TCGA-COAD cohort using the ‘DESeq2’ R package. Differentially expressed genes (DEGs) between tumor and adjacent normal tissues were identified using the thresholds of adjusted p < 0.05 and |log₂FoldChange| > 1. Ferroptosis- and lipid metabolism-related gene sets were retrieved from the GeneCards database (https://www.genecards.org/) using the search terms ‘ferroptosis’ and ‘lipid metabolism,’ respectively. All annotated human protein-coding genes associated with each term were included (Table S1). Candidate genes for subsequent network analysis were defined by the intersection between the CRC-specific DEGs and the curated ferroptosis- and lipid metabolism-related gene sets, yielding a total of 1,075 unique genes.

WGCNA

To characterize the co-expression patterns and regulatory relationships among candidate genes, WGCNA was performed using the ‘WGCNA’ R package [28]. A soft-thresholding power was selected according to the scale-free topology criterion to construct a signed scale-free network. Gene modules were identified using the dynamic tree-cut algorithm with a minimum module size of 30 genes, and modules with highly correlated eigengenes were merged using a merge cut height of 0.25. Module–trait relationships were subsequently evaluated by calculating the correlations between module eigengenes and clinical characteristics. The purple module exhibited the strongest association with patient survival and tumor staging. Therefore, it was selected for prognostic modeling.

Machine learning-based prognostic signature construction

Based on the clinically relevant genes identified from the purple module, a comprehensive machine learning framework was implemented to develop a robust prognostic signature. A ten-fold cross-validation strategy was employed to systematically evaluate 101 combinations of feature-selection methods (e.g. StepCox [forward/backward], LASSO, and Random Survival Forest) and survival prediction algorithms (e.g. CoxBoost, survival-SVM, and RSF). The predictive performance of each combination was evaluated using the time-dependent area under the receiver operating characteristic curve (time-dependent AUC) and the concordance index (C-index) for overall survival in both the training and external validation cohorts. The optimal model, consisting of forward stepwise Cox regression and Random Survival Forest, demonstrated the best overall predictive performance and was selected for subsequent model construction. Individual RiskScores were calculated as a weighted linear combination of the expression levels of the signature genes:

$$\text{Risk\hspace{0.33em}Score=}{\displaystyle \sum }_{\text{i\hspace{0.17em}=\hspace{0.17em}1}}^7\left({\beta }_{\mathrm{i}}\times {\text{Expr}}_{\mathrm{i}}\right)$$

Where β_i represents the regression coefficient of gene i, and Expr_i denotes its normalized expression level. The regression coefficients for the seven-gene signature were as follows: CRY2: 0.412; CDC42: 0.287; PDSS2: −0.319; VDAC3: 0.253; ULK1: 0.196; UBE2K: 0.174; UBE2D2: 0.158.

Validation and performance evaluation of the prognostic model

Patients in both the training and validation cohorts were stratified into high- and low-risk groups according to the median RiskScore. Overall survival was compared using Kaplan-Meier survival analysis with the log-rank test. The independent prognostic value of the RiskScore was evaluated using univariate and multivariate Cox proportional hazards regression analyses after adjustment for conventional clinicopathological covariates. To facilitate clinical translation, a prognostic nomogram incorporating the RiskScore and independent clinical prognostic factors was constructed. Model performance was assessed using calibration curves and the concordance index (C-index), while decision curve analysis (DCA) was performed to evaluate the potential clinical utility and net benefit of the prognostic model.

Analysis of the TME, immune checkpoints, and therapeutic response

The composition of the TME was characterized using the CIBERSORT algorithm to estimate the relative abundance of infiltrating immune cell populations. Associations between the RiskScore and the expression of immune checkpoint molecules, chemokines, and cytokines were evaluated. TME scores were calculated using the ESTIMATE algorithm, and differences in TME scores, immune cell infiltration, and TIDE scores were compared between high- and low-risk groups. Drug sensitivity (IC₅₀) for common chemotherapeutic agents was predicted using the ‘pRRophetic’ R package.

Functional enrichment analysis

To investigate the biological processes underlying risk stratification, Gene Set Variation Analysis (GSVA) was performed to evaluate the enrichment of Hallmark, KEGG, and Reactome pathways between the high- and low-risk groups. Functional annotation of the signature genes was conducted using Gene Ontology (GO) and KEGG enrichment analyses. In addition, Gene Set Enrichment Analysis (GSEA) was performed on the ranked whole-transcriptome dataset to identify biological pathways associated with the RiskScore.

In vitro validation

Human CRC cell lines (SW620 and SW480) were obtained from the Chinese Cell Bank (CBTCC) and cultured in high-glucose DMEM supplemented with 10% fetal bovine serum (Gibco, USA) at 37 °C with 5% CO₂. CRY2 expression was silenced using small interfering RNAs (siRNAs) synthesized by Sangon Biotech (Shanghai, China). Transient transfection was performed using Lipofectamine^™ 3000 (Thermo Fisher Scientific, USA) according to the manufacturer’s instructions. The siRNA sequences targeting CRY2 expression were as follows:

• Sense: 5′-GCACGUCGUAGCAUCAGCAUTT-3′

• Antisense: 5′-AUGCUGAUGCUACGACGUGCTT-3′

Cell proliferation assay

Cell proliferation was evaluated using a 5-ethynyl-2′-deoxyuridine (EdU) incorporation assay. Following siRNA transfection, cells were seeded into 24-well plates (1 × 10⁵ cells/well) and incubated with EdU for 2 h. Cells were then washed with PBS, fixed with 4% paraformaldehyde, permeabilized, and stained according to the manufacturer’s protocol. Nuclei were counterstained with Hoechst 33342, and EdU-positive cells were visualized by fluorescence microscopy. The proliferation rate was quantified as the percentage of EdU-positive cells.

Cell migration assay

A standardized scratch was generated in the confluent monolayer of transfected cells using a 200 μL pipette tip. Wound closure was monitored over 48 h.

Flow cytometry and Western blotting

Cell-cycle distribution was analyzed using a Cell Cycle Staining Kit (Beyotime, China) according to the manufacturer’s instructions and quantified using a BD Biosciences flow cytometer. For western blot analysis, total protein was extracted using RIPA buffer supplemented with protease/phosphatase inhibitors, and protein concentrations were determined using a BCA assay. Equal amounts of protein were separated by 10% SDS-PAGE, transferred onto PVDF membranes, and incubated overnight with primary antibodies at 4 °C following membrane blocking. After incubation with horseradish peroxidase (HRP)-conjugated secondary antibodies, protein bands were detected using enhanced chemiluminescence (ECL) reagents.

In vivo validation

To validate the biological function of CRY2 in vitro, a subcutaneous xenograft model was established using 12 5-week-old female BALB/cA-nu nude mice, which were randomly assigned to two groups (n = 6 per group). Stably transfected SW620 cells (1 × 10⁷ cells suspended in 100 μL PBS) were injected into the dorsal flank of each mouse. Tumor growth was monitored by measuring tumor length and width with digital calipers throughout the experimental period. After 21 days, mice were euthanized by carbon dioxide inhalation, and xenograft tumors excised, weighed, and photographed. The animals were obtained from Beijing Vital River Laboratory Animal Technology Co., Ltd. (License No. SCXK(Ji)2021-0011). All animal experiments were approved by the Animal Care and Use Committee (ACUC) of the School of Basic Medical Sciences, Jilin University (Approval No. 2025-105), and were conducted in accordance with the Guide for the Care and Use of Laboratory Animals (National Institutes of Health Publication No. 85-23, revised 1985).

Statistical analysis

All statistical analyses were performed using R (version 4.3.1). DEGs were identified using the DESeq2 package (version 1.42.0) with thresholds of adjusted p < 0.05 and |log₂FC| > 1. Survival differences between groups were evaluated using Kaplan-Meier curves and the log-rank test. Univariate and multivariate Cox proportional hazards regression analyses were performed to identify independent prognostic factors, with hazard ratios (HRs) and 95% confidence intervals (CIs) reported. Continuous variables were compared using the Wilcoxon rank-sum test, whereas Pearson correlation was used to evaluate module-trait associations in WGCNA. Multiple testing was controlled using the Benjamini-Hochberg false discovery rate (FDR) for differential expression and pathway enrichment analyses. Bonferroni correction was applied for pairwise comparisons of immune cell infiltration. The prognostic machine learning models were evaluated using repeated 10-fold cross-validation (five repetitions), with the time-dependent AUC and C-index as the primary performance metrics. Model robustness was further assessed via 100 bootstrap resamples, and the coefficient of variation (CV) of the C-index was calculated. All figures were generated using ggplot2 (version 3.4.4) and survminer (version 0.4.9) R packages. Unless otherwise specified, all statistical tests were two-sided, and p < 0.05 was considered statistically significant.

Results

WGCNA identifies ferroptosis- and lipid metabolism-related modules associated with prognosis

To identify transcriptomic features linking metabolic dysregulation to CRC progression, differential expression analysis was first performed on the TCGA-COAD cohort using the DESeq2 package (Figure 1A). Ferroptosis- and lipid metabolism-related gene sets were subsequently retrieved from the GeneCards database, and their intersection with the identified DEGs yielded 1,076 candidate genes for downstream analyses (Figure 1B and Table S1). WGCNA was then performed to identify biologically relevant co-expression modules (Figure 1C). A soft-thresholding power of 4 was selected to construct a scale-free co-expression network, achieving a scale-free topology fit index (R²) > 0.85 while preserving sufficient network connectivity (Figure 1D, E). This analysis identified 17 distinct gene modules. Module–trait associations were evaluated using Pearson correlation analysis, with statistical significance defined as p < 0.05. Among the identified modules, the purple module exhibited the strongest association with clinical outcomes, showing a significant negative correlation with prognosis in COAD (r = −0.14, p = 0.003; Figure 1F). Accordingly, this module was selected for subsequent prognostic modeling.

Figure 1.

Identification of ferroptosis- and lipid metabolism-related candidate genes and WGCNA module selection. (A) Differential expression analysis between tumor and adjacent normal tissues in the TCGA-COAD dataset. (B) Venn diagram showing the overlap between DEGs and ferroptosis- and lipid metabolism-related gene sets, yielding 1,076 candidate genes. (C) Sample clustering dendrogram for WGCNA. (D) Scale-free topology fit index for various soft-thresholding powers. (E) Mean connectivity across different soft-thresholding powers. (F) Module-trait relationship heatmap illustrating the correlation between gene modules and clinical traits.

Construction and validation of a machine learning-based prognostic signature

Using genes from the clinically relevant purple module, a prognostic signature was developed through a comprehensive machine learning framework. The framework evaluated 101 combinations of feature-selection and survival prediction algorithms using repeated 10-fold cross-validation. Model performance was assessed by the C-index (Figure 2A) and one-year AUC (Figure 2B) across both the training (TCGA-COAD) and validation (GSE39582) cohorts. Among all candidate models, the combination of StepCox[forward] and RSF demonstrated the best predictive performance and was selected as the final prognostic model. Individual RiskScores were subsequently calculated, and patients were stratified into high- and low-risk groups. Survival analysis revealed that patients in the high-risk group had significantly poorer prognosis than those in the low-risk group in both the training and validation cohorts (Figure 2C). During model construction, forward stepwise Cox regression was performed using the Akaike Information Criterion (AIC) as the optimization criterion. Candidate genes were sequentially incorporated into the model until no further reduction in the AIC was achieved, resulting in a final prognostic signature comprising seven genes.

Figure 2.

Construction and validation of the machine learning-based prognostic RiskScore. (A) Comparison of the C-index among 101 machine learning algorithm combinations in the training and validation cohorts. (B) Comparison of the 1-year time-dependent AUC values across different algorithm combinations. (C) Kaplan-Meier curves for overall survival (OS) of patients in high- and low-risk groups across the training and validation cohorts. (D) Univariate and multivariate Cox regression analyses evaluating the prognostic value of the RiskScore and clinicopathological factors in the training cohort. (E) Prognostic nomogram integrating the RiskScore with clinical variables. (F–G) External validation of the independent prognostic value of the RiskScore by univariate (F) and multivariate (G) Cox regression analyses (*p < 0.05, **p < 0.01, ***p < 0.001).

The RiskScore serves as an independent prognostic factor with superior clinical utility

To evaluate the independent prognostic value of the RiskScore, univariate and multivariate Cox proportional hazards regression analyses were performed incorporating the RiskScore together with conventional clinicopathological variables. In the training cohort, univariate analysis demonstrated that the RiskScore was significantly associated with overall survival (HR = 1.169, 95% CI: 1.147–1.191, p < 0.001), indicating that each one-unit increase in the RiskScore was associated with a 16.9% increase in the risk of mortality (Figure 2D). Multivariate Cox regression further confirmed the RiskScore as an independent prognostic factor (HR = 1.178, p < 0.001), with its prognostic significance maintained after adjustment for age, sex, and TNM staging, whereas TNM staging was not independently associated with survival (HR = 1.450, p = 0.152; Figure 2D). Patients stratified according to the median RiskScore exhibited markedly different survival outcomes, with a 3.2-fold difference in 5-year survival rates between the high- and low-risk groups in both the training (72.1% vs. 28.4%) and validation cohorts (68.9% vs. 31.7%), underscoring the robust risk stratification capability of the prognostic signature. To facilitate individualized prognostic assessment, a nomogram integrating the RiskScore with age, gender, and TNM staging was subsequently established (Figure 2E). The independent prognostic value of the RiskScore was further validated in the external cohort, where it remained a significant predictor of overall survival in both univariate and multivariate Cox regression analyses (HR = 1.023, p < 0.001; Figure 2F–G).

Comprehensive evaluation of the nomogram and clinicopathological characteristics

The clinicopathological characteristics associated with the RiskScore, including age, gender, TNM staging, and recurrence/metastasis status, were comprehensively visualized using heatmaps in both the training (Figure 3A) and validation (Figure 3B) cohorts, together with the distribution of patient survival status (Figure 3E). Distinct expression patterns of the seven genes constituting the RiskScore were observed between the high- and low-risk groups in both the training (Figure 3C) and validation cohorts (Figure 3D). To facilitate individualized prognostic assessment, a nomogram integrating the RiskScore with conventional clinicopathological variables was established (Figure 3F). Calibration analyses demonstrated excellent agreement between the predicted and observed probabilities of 1-, 3-, and 5-year overall survival, with the calibration curves closely approximating the ideal reference line (Figure 3G). Furthermore, DCA demonstrated that the nomogram provided greater net clinical benefit than other strategies (Figure 3H).

Figure 3.

Comprehensive evaluation of the RiskScore nomogram. (A, B) Heatmaps depicting the association of the RiskScore with clinicopathological features in the training (A) and validation (B) cohorts. (C–D) Differential expression of the seven signature genes between the high- and low-risk groups in the training (C) and validation (D) cohorts. (E) Distribution of patient survival status and RiskScores. (F) Nomogram integrating the RiskScore with clinicopathological variables for predicting 1-, 3-, and 5-year overall survival. (G) Calibration curves assessing the agreement between predicted and observed survival probabilities at 1, 3, and 5 years. (H) DCA evaluating the clinical utility of the nomogram (*p < 0.05, **p < 0.01, ***p < 0.001).

The high-risk TME exhibits distinct immunosuppressive features

The immune landscape associated with RiskScore stratification was subsequently characterized. The CIBERSORT algorithm was employed to estimate the relative abundance of 22 immune cell subsets (Figure 4A). Correlation analysis revealed that the RiskScore was positively associated with multiple pro-immunosuppressive and immune checkpoint-related molecules, including TNFSF4, TNFSF8, and VTCN1 (Figure 4B). These associations were consistently supported by several independent immune deconvolution algorithms, including xCell, TIMER, and EPIC (Figure 4C). Compared with the low-risk group, the high-risk group exhibited significantly elevated stromal scores, whereas immune scores and overall ESTIMATE scores did not differ significantly between the two groups, suggesting that RiskScore stratification is primarily associated with the stromal microenvironment (Figure 4D). Further analysis of immune cell composition revealed that the high-risk group was characterized by reduced infiltration of anti-tumor effector cells (e.g. activated CD4⁺ T cells) and increased infiltration of immunosuppressive cell populations, such as regulatory T cells and Th2 cells (Figure 4E). Notably, these immunological features were consistently reproduced in the validation cohort (Figure 4F–J).

Figure 4.

Landscape of the TME and immune infiltration in high- and low-risk groups. (A) Relative abundance of 22 immune cell types estimated using CIBERSORT. (B) Correlation between the RiskScore and the expression of immune checkpoint-related molecules. (C) Association between the RiskScore and immune cell infiltration estimated by the xCell, TIMER, and EPIC algorithms. (D) Comparison of StromalScore, ImmuneScore, and ESTIMATE Score between the high- and low-risk groups in the training cohort. (E) Differential infiltration of immune cell populations between the high- and low-risk groups in the training cohort. (F–J) Validation of immune infiltration differences in the external cohort (*p < 0.05, **p < 0.01, ***p < 0.001).

High-risk tumors are characterized by immune exclusion and therapeutic resistance

To elucidate the mechanisms underlying the immunosuppression, the expression of chemokines, cytokines, and interleukin receptors was analyzed. Distinct expression patterns were observed between the high- and low-risk groups (Figure 5A, B). Classification according to the established pan-cancer immune subtypes (C1–C6) revealed that high-risk tumors were predominantly enriched in the C1 (wound healing), C2 (IFN-γ dominant), and C4 (lymphocyte-depleted) subtypes, indicating a TME characterized by chronic inflammation accompanied by immune suppression (Figure 5C) [32]. To further assess immune escape potential, microsatellite instability (MSI) and Tumor Immune Dysfunction and Exclusion (TIDE)-related metrics were compared between the two risk groups. High-risk tumors exhibited lower Microsatellite Instability (MSI) scores but higher Exclusion, Dysfunction, and TIDE scores than low-risk tumors (Figure 5D), suggesting a more pronounced immune-evasive phenotype and a reduced likelihood of benefiting from immune checkpoint blockade. Pharmacogenomic prediction revealed that the high-risk group exhibited significantly reduced sensitivity to standard CRC chemotherapeutics. As shown in Figure 5E, the predicted IC50 value for 5-fluorouracil, oxaliplatin, paclitaxel, and sorafenib were significantly higher in the high-risk group compared to the low-risk group. By definition, a higher IC50 indicates a greater concentration of drug required to inhibit 50% of tumor cell growth, reflecting increased chemoresistance. These findings suggest that high-risk patients are less likely to derive clinical benefit from standard chemotherapy regimens.

Figure 5.

Immune escape characteristics and predicted therapeutic responses associated with the RiskScore. (A, B) Heatmaps illustrating the differential expression of chemokines, interleukin receptors, and cytokines between the high- and low-risk groups. (C) Distribution of immune subtypes (C1–C6) in the high- and low-risk groups. (D) Comparison of MSI, Exclusion, Dysfunction, and TIDE scores between the two high- and low-risk groups. (E) Predicted drug sensitivity (represented by estimated IC₅₀ values) for 5-fluorouracil, oxaliplatin, paclitaxel, and sorafenib in the high- and low-risk groups; higher IC50 values correspond to reduced drug sensitivity (*p < 0.05, **p < 0.01, ***p < 0.001).

Functional enrichment analysis reveals dysregulation of cell cycle and metabolic pathways

To investigate the biological basis of the RiskScore, comprehensive functional enrichment analyses were performed. In the TCGA-COAD cohort, KEGG and Reactome analyses demonstrated that genes associated with the high-risk group were predominantly enriched in cell cycle-related pathways, including DNA replication and mismatch repair (Figure 6A). Consistent findings were observed in the external validation cohort, where significant enrichment of ECM-receptor interaction and immune response SUMOylation was identified (Figure 6B). GO analysis further revealed significant enrichment of biological processes associated with mitochondrial translation and ribosomal function (Figure 6C). KEGG enrichment analysis of differentially expressed genes highlighted prominent alterations in cell cycle regulation, oxidative phosphorylation, and ubiquitin-mediated proteolysis pathways (Figure 6D). GSEA further confirmed the enrichment of mitochondrial and ribosome-related gene sets (Figure 6E), together with canonical KEGG pathways involved in cell cycle regulation and oxidative phosphorylation (Figure 6F). To further characterize the biological functions of the prognostic signature, functional enrichment analysis of the seven RiskScore genes was conducted. GO analysis indicated significant enrichment in cell-cell adhesion and amyloid metabolic processes (Figure 6G), whereas KEGG analysis identified calcium signaling and neuroactive ligand-receptor interaction as the principal enriched pathways (Figure 6H).

Figure 6.

Functional enrichment analyses of the RiskScore-associated genes. (A, B) KEGG and Reactome pathway enrichment analyses of genes associated with the high-risk group in the TCGA-COAD (A) and GSE39582 (B) cohorts. (C) GO enrichment analysis highlighting biological processes related to mitochondrial translation and ribosomal function. (D) Bubble plot showing significantly enriched KEGG pathways among DEGs, including cell cycle and oxidative phosphorylation. (E) GSEA of mitochondrial- and ribosome-related gene sets. (F) GSEA of significantly enriched KEGG pathways. (G–H) GO (G) and KEGG (H) enrichment analyses of the seven genes comprising the prognostic signature.

CRY2 is identified as a key prognostic gene associated with CRC progression

Among the seven signature genes, CRY2 emerged as the most influential candidate. In the StepCox model, CRY2 exhibited the largest absolute regression coefficient (β = 0.412), indicating the greatest contribution to RiskScore calculation. Furthermore, multivariate Cox regression analysis adjusting for all clinicopathological variables identified CRY2 as the only signature gene that retained independent prognostic significance (HR = 1.824, 95% CI: 1.086–3.060, p = 0.023; Figure 7A, B). Kaplan–Meier survival analysis demonstrated that patients with high CRY2 expression exhibited significantly poorer overall survival than those with low expression (Figure 7C and Figure S1). Moreover, CRY2 expression increased progressively with advancing tumor stage, suggesting a positive association with disease progression (Figure 7D). Single-cell analysis via the TISCH database revealed broad CRY2 expression across immune cell types, with relatively higher expression in T cells compared to other lineages (Figure 7E–F and Figure S2).

Figure 7.

Clinical significance and single-cell characterization of CRY2. (A, B) Univariate (A) and multivariate (B) Cox regression analyses demonstrating the independent prognostic significance of CRY2. (C) Kaplan-Meier analysis of overall survival according to CRY2 expression. (D) Association between CRY2 expression and tumor stage. (E) Single-cell expression landscape of CRY2 across immune cell populations in the TISCH database. (F) Expression of CRY2 among infiltrating immune cell subsets in the GSE108989 dataset, showing preferential enrichment in T-cell populations (*p < 0.05, **p < 0.01, ***p < 0.001).

CRY2 silencing inhibits CRC progression in vitro and in vivo

To validate the biological function of CRY2, siRNA-mediated knockdown models were established in SW480 and SW620 cells (Figure 8A). Functional assays demonstrated that CRY2 silencing markedly suppressed CRC cell proliferation, as evidenced by reduced colony formation (Figure 8B) and decreased DNA synthesis in EdU assays (Figure 8C). In addition, wound-healing assays showed that CRY2 knockdown significantly impaired cell migration (Figure 8D). Flow cytometry analysis further revealed that the antiproliferative effect of CRY2 silencing was accompanied by G1-phase cell cycle arrest (Figure 8E). To explore the molecular mechanisms by which CRY2 modulates cell cycle dynamics, we interrogated the transcriptional correlations between CRY2 and core G1/S transition regulators in the TCGA-COAD dataset. Interestingly, CRY2 expression exhibited significant negative correlations with CCND1 (R = −0.36, p < 0.05) and CDK4 (R = −0.47, p < 0.05), while demonstrating a strong positive correlation with the cyclin-dependent kinase inhibitor CDKN1A (p21; R = 0.18, p < 0.05). These data suggest that CRY2 closely interacts with the canonical p21-Cyclin D1-CDK4 signaling axis to regulate the G1 restriction point in CRC cells (Figure S3). Consistent with the in vitro findings, CRY2 knockdown significantly reduced tumor volume and weight compared to controls (Figure 8F). Immunohistochemical staining further confirmed decreased Ki-67 expression in tumors with CRY2 knockdown (Figure 8G).

Figure 8.

CRY2 knockdown suppresses CRC progression in vitro and in vivo. (A) Validation of siRNA-mediated CRY2 knockdown efficiency in SW480 and SW620 cell lines. (B) Colony formation assays assessing the proliferative capacity of control and CRY2-knockdown cells. (C) Representative images and quantification of EdU incorporation assays assessing DNA synthesis. (D) Wound-healing assays evaluating cell migration capabilities. (E) Flow cytometry analysis of cell cycle distribution. (F) Representative images of xenograft tumors derived from control and CRY2-knockdown SW620 cells in nude mice. (G) Representative immunohistochemical staining and quantitative analysis of Ki-67 expression in xenograft tumor tissues.

Discussion

CRC remains a major global health burden, accounting for substantial cancer-related morbidity and mortality despite continuous advances in surgical techniques and systemic therapies [1]. Although patients with early-stage disease can often achieve favorable outcomes following curative resection, the prognosis of advanced CRC remains unsatisfactory owing to profound molecular heterogeneity and dynamic metabolic reprogramming. Current clinical risk stratification relies primarily on the TNM staging system [5]. However, this anatomical classification does not adequately capture the biological complexity of CRC, resulting in considerable prognostic variability among patients with the same pathological stage. Consequently, there is an urgent need for robust molecular biomarkers that complement conventional clinicopathological parameters and enable more accurate prognostic prediction and therapeutic decision-making.

In this study, a prognostic signature based on ferroptosis- and lipid metabolism-related genes was established through the integration of transcriptomic datasets using WGCNA and a consensus machine learning framework. The resulting RiskScore demonstrated robust prognostic performance and independent predictive value across both the training and external validation cohorts. Furthermore, CRY2 was identified as the most influential component of the prognostic signature and subsequently validated as an independent prognostic biomarker with functional roles in colorectal cancer progression.

Metabolic reprogramming is a hallmark of cancer, enabling tumor cells to sustain survival and proliferation under physiological stress and nutrient deprivation [33,34]. In particular, the interplay between lipid metabolism and ferroptosis represents a critical metabolic vulnerability in malignant cancer cells [35]. In the present study, WGCNA was employed to identify gene modules at the intersection of these two biological processes, ensuring that the selected features were mechanistically relevant to CRC progression. A major strength of this study lies in the rigorous development of the prognostic model. Rather than relying on a single machine learning algorithm, 101 algorithmic combinations were systematically evaluated to construct a consensus RiskScore. This ensemble strategy minimizes the risk of overfitting while enhancing model robustness and generalizability across independent cohorts. The clinical applicability of this RiskScore was further supported by its consistent prognostic performance in an external validation cohort. Furthermore, integration of the RiskScore with conventional clinicopathological parameters yielded a nomogram with excellent calibration and superior clinical net benefit, suggesting that this model may complement the TNM staging system by improving individualized risk stratification and facilitating the identification of patients who may benefit from intensified postoperative management or adjuvant therapy.

From a biological perspective, the high-risk phenotype identified by the RiskScore is characterized by extensive metabolic reprogramming. Functional enrichment analysis revealed activation of pathways involved in oxidative phosphorylation and mitochondrial translation in the high-risk group. These findings are consistent with emerging evidence that aggressive CRC exploits metabolic plasticity, particularly enhanced lipid metabolism and mitochondrial bioenergetics, to sustain rapid proliferation and tumor progression [36,37]. Increased fatty acid synthesis and oxidation not only satisfy the elevated energetic and biosynthetic demands of proliferating tumor cells but also promote the generation of polyunsaturated phospholipids, theoretically increasing susceptibility to ferroptosis through lipid peroxide accumulation [38]. However, the unfavorable clinical outcomes observed in the high-risk group suggest that these tumors have acquired adaptive mechanisms to counteract ferroptotic cell death, potentially through reinforcement of antioxidant defense pathways, including the GPX4 axis. Moreover, metabolic remodeling appears to extend beyond intrinsic tumor cell metabolism to influence the TME. The enrichment of extracellular matrix (ECM)-receptor interaction pathways indicates that metabolic alterations may promote ECM remodeling, thereby facilitating tumor invasion, metastatic dissemination, and disease progression.

The efficacy of immunotherapy is profoundly influenced by the metabolic and structural characteristics of the TME. A major finding of the present study is the identification of an immune-excluded phenotype in high-risk tumors, characterized by preferential infiltration of immunosuppressive regulatory T cells (Tregs) and Th2 cells, accompanied by a marked depletion of effector CD4⁺ T cells. This immunological landscape was consistently identified across multiple immune deconvolution algorithms, including CIBERSORT, xCell, TIMER, and EPIC, despite the inherent methodological differences among these approaches. Although these computational methods rely on predefined reference signatures and may exhibit reduced accuracy in highly stromal-rich tumors or samples with substantial tumor-cell contamination, the concordance of results across independent algorithms strengthens the robustness of the observed immunosuppressive phenotype.

From an immunometabolic perspective, aberrant lipid metabolism in high-risk tumors may contribute directly to immune dysfunction by creating a nutrient-deprived and metabolically hostile microenvironment that impairs effector T-cell function while promoting the survival and functional stability of Tregs [39]. This interpretation is further supported by elevated TIDE, T-cell exclusion, and dysfunction scores, together with reduced MSI scores. These metrics indicate enhanced immune evasion and diminished responsiveness to immune checkpoint blockade (ICB) [40]. However, the predictive performance of the TIDE score requires prospective validation in patients receiving ICB, which was beyond the scope of this study. Despite this limitation, the metabolism-based RiskScore captures multiple features associated with immune exclusion and may therefore serve as a complementary biomarker for identifying patients with reduced responsiveness to ICB. These findings also suggest a potential therapeutic opportunity, whereby targeting the lipid metabolism-ferroptosis axis may alleviate metabolic barriers within the TME, promote immune-cell infiltration, and convert immunologically ‘cold’ tumors into ‘hot’ tumors, thereby enhancing sensitivity to immunotherapy.

To further elucidate the molecular basis underlying the high-risk phenotype, CRY2, a core circadian regulator, was selected for functional investigation as the most influential gene within the prognostic signature. Although circadian disruption has emerged as a hallmark of tumorigenesis, the biological role of CRY2 in CRC remains incompletely understood. This study provides integrated computational, clinical, and experimental evidence supporting its oncogenic relevance in CRC. Elevated CRY2 expression was associated with advanced tumor stage and unfavorable survival outcomes. Loss-of-function experiments demonstrated that CRY2 silencing significantly inhibited tumor cell proliferation, impaired migration, induced G1-phase cell-cycle arrest, and suppressed tumor growth in vivo. This proliferative and migratory impediment aligns with prior functional descriptions of aggressive CRC models, where the ablation of specific oncogenic non-coding RNAs or hub drivers (such as Linc01615) similarly triggered a collapse in survival, invasion, and metastatic competence [41].

These findings are consistent with the enrichment of cell-cycle-related pathways identified by GSEA in the high-risk group, suggesting that CRY2 contributes to CRC progression by promoting cell-cycle progression. Although the precise downstream mechanisms remain to be elucidated, CRY2 may regulate key cell-cycle regulators, such as Cyclin D1, CDKs, or p53 signaling, thereby linking circadian dysregulation to uncontrolled cellular proliferation.

Several limitations should also be acknowledged. First, although the prognostic performance of the RiskScore was validated in an independent external cohort, this study remains retrospective in nature, and potential selection bias cannot be excluded. Prospective, multicenter clinical studies are required to determine its clinical utility for patient stratification and therapeutic decision-making. Second, immune infiltration and immunotherapy responsiveness were inferred using computational deconvolution algorithms and transcriptome-based predictive models. While the concordance across multiple analytical approaches supports the robustness of these findings, validation in prospective cohorts receiving immune checkpoint blockade therapy and complementary spatial or single-cell profiling will be essential. Third, although CRY2 knockdown consistently suppressed malignant phenotypes in vitro and in vivo, the current study primarily employed loss-of-function models. Gain-of-function and genetic rescue experiments were not performed and will be important for establishing the causal role and dosage-dependent effects of CRY2. Moreover, the molecular mechanisms linking CRY2 to ferroptosis, lipid metabolic reprogramming, and tumor-immune interactions remain largely unresolved. Future studies integrating transcriptomic, proteomic, metabolomic, and spatial multi-omics approaches will help define the downstream regulatory network governed by CRY2 and clarify its contribution to metabolic-immune crosstalk in colorectal cancer.

Conclusion

In summary, this study establishes a biologically interpretable prognostic signature by integrating ferroptosis- and lipid metabolism-related genes and demonstrates its robust prognostic performance across independent cohorts. Beyond prognostic prediction, the findings identify CRY2 as a potential regulator of colorectal cancer progression and provide mechanistic evidence linking circadian dysregulation to cell-cycle activation within a metabolically reprogrammed tumor context. Collectively, these findings provide a theoretical foundation for developing therapeutic strategies that simultaneously target metabolic vulnerabilities and circadian regulatory pathways to improve outcomes for high-risk CRC patients.

Funding Statement

This study was supported by the Science and Technology Department of Jilin Province (Grant No. 3D52301068429) and the Department of Education of Jilin Province (Grant No. JJKH20250202BS).

Ethics approval and consent to participate

Human transcriptomic and clinical data used in this study were obtained from publicly available databases. The study protocol was reviewed and approved by the Medical Ethics Committee of the Affiliated Hospital of Jilin University (Approval No. JLU-MED-2025-872). All animal experiments were approved by the Animal Care and Use Committee (ACUC) of the School of Basic Medical Sciences, Jilin University (Approval No. 2025-105). All experimental procedures were performed in accordance with the institutional guidelines for animal care and the ‘Principles of Laboratory Animal Care’ (NIH Publication No. 85-23, revised 1985).

Disclosure statement

The authors declare that they have no competing interests.

Availability of data and materials

The datasets analyzed during the current study are publicly available from The Cancer Genome Atlas (TCGA) and the Gene Expression Omnibus (GEO). All other data generated or analyzed during this study are included in this published article and its Supplementary Materials. The code utilised in this study has been uploaded to Figshare, with the DOI: https://doi.org/10.6084/m9.figshare.31324018. These datasets are freely accessible without the need to log in.