Identification and validation of a Golgi apparatus related gene signature for prognosis prediction and immune microenvironment profiling in colorectal cancer

Abstract

Background

Colorectal cancer (CRC) is a leading cause of cancer mortality. The Golgi apparatus (GA) mediates protein glycosylation and secretion, and its dysfunction is implicated in cancer invasion and immune evasion [1]. The prognostic relevance of GA-related genes in CRC remains unclear.

Methods

Transcriptomic and clinical data from TCGA-COAD/READ and GSE39582 were analysed. A set of 1 686 GA-related genes from MSigDB was filtered by differential expression, and prognostic candidates were selected using univariate Cox, LASSO and random forest approaches to construct a multi-gene risk score. Functional enrichment, immune infiltration (ESTIMATE/CIBERSORT) and drug-sensitivity (oncoPredict) analyses were performed, and a nomogram combining the GA score with clinical covariates was built.

Results

Among 5580 differentially expressed genes, 434 were GA-related. A six-gene signature (PCSK5, RAB36, CD36, DPP7, KPNA2, HEPACAM2) stratified patients into high- and low-risk groups with significant survival differences and retained independent prognostic value after adjusting for age and stage. High-risk tumors showed enrichment of proteoglycan and G-protein-coupled receptor pathways, higher immune and stromal scores with abundant monocytes and macrophages but fewer dendritic and T cells, and were predicted to respond poorly to immune checkpoint blockade yet displayed increased sensitivity to the IGF-1R inhibitor BMS-754,807.

Conclusions

GA related gene expression delineates CRC subtypes with distinct biology, immunity and therapeutic vulnerabilities. Integrating the GA derived signature with clinical factors may help refine risk stratification and generate hypotheses regarding glycosylation-/proteoglycan-related therapeutic vulnerability and IGF-1R inhibition.

Supplementary Information

The online version contains supplementary material available at 10.1007/s12672-026-05156-0.

Introduction

Colorectal cancer (CRC) is among the most common malignancies worldwide [2–4]. Epidemiologically, it has the third highest incidence of all cancers and is the second leading cause of cancer-related mortality globally [5]. Currently, CRC can be treated with surgery, chemotherapy, radiotherapy, and biotherapy, and despite major advances in these strategies, patient prognosis remains unsatisfactory. Moreover, due to tumor heterogeneity, it is crucial to establish a systematic and precise method to stratify CRC risk [6, 7].

The Golgi apparatus (GA) is an important organelle involved in the modification, sorting, and transport of proteins and lipids [8]. Beyond these traditional functions, the GA also regulates cell signaling and immune responses through protein glycosylation and secretion. Increasing evidence indicates that cancer cells often exhibit structural abnormalities in the GA, such as fragmentation and dispersal; this dysfunction can lead to aberrant glycosylation and secretion profiles that promote tumor invasion, metastasis, and immune evasion [8, 9]. Because the GA is relatively fragile, its abnormality also presents a potential therapeutic target: for example, drugs like brefeldin A and monensin can induce apoptosis by disrupting GA-mediated protein trafficking or causing ionic imbalance, and GA-targeted anticancer strategies are being explored [8].

However, research on GA-associated molecules in the context of CRC prognosis remains very limited. Although certain Golgi proteins have been reported to contribute to CRC tumorigenesis, their prognostic value is not yet clear. For example, the Golgi membrane protein GOLM1 (GP73) was shown in a CRC cohort to drive metastasis and regulate epithelial–mesenchymal transition and secretory networks in cancer cells, suggesting its pro-malignant role in tumor progression and metastasis [10]. Additionally, GOLPH3, a classical GA-related oncoprotein/PI4P effector, links the trans-Golgi membrane to the actomyosin network to influence vesicle budding and anterograde transport; studies in multiple cancers have found it associated with adverse tumor phenotypes, and it has been reported as a predictive or prognostic marker in CRC [11, 12]. Despite the accumulation of such evidence, bioinformatic studies that systematically integrate GA-related genes (GARGs) to characterize CRC prognostic heterogeneity, immune microenvironment status, and therapy response potential are still relatively scarce. Therefore, in this study we aimed to systematically screen public multi-omics databases and construct a GA-related prognostic model, validate its robustness in independent cohorts, and further identify GA genes associated with survival to explore their biological significance and therapeutic value. Unlike previous CRC prognostic models that primarily focus on immune, metabolic, ferroptosis or autophagy pathways, our study introduces a Golgi apparatus–related perspective that captures the organelle’s central roles in protein glycosylation, secretion and membrane trafficking, thereby providing a novel framework linking glycosylation/protein secretion/microenvironment remodeling to therapeutic response [13–16].

Methods

Data acquisition

Gene-expression profiles and corresponding overall-survival data for colorectal cancer (CRC) were obtained from The Cancer Genome Atlas (TCGA) database (https://portal.gdc.cancer.gov/) and the Gene Expression Omnibus (GEO) database (https://www.ncbi.nlm.nih.gov/gds/). After quality control and availability filtering, the TCGA cohort contained 51 normal and 618 tumor samples overall; the TCGA-COAD/READ set served as the training cohort, and the GSE39582 dataset, comprising 585 tumor samples, served as the external validation cohort. Golgi apparatus (GA)–related genes were retrieved from the Molecular Signatures Database (MSigDB; URL: https://www.gsea-msigdb.org/gsea/msigdb/human/collections.jsp), yielding 1,686 GA-related genes. Machine-learning based prognostic signatures have been successfully applied in other cancers, providing methodological precedents for our approach [17, 18].

Identification of differentially expressed GA-related genes

Within TCGA-COAD/READ, we quantified expression of GA-related genes and performed differential expression analysis between normal colorectal tissue and CRC using DESeq2 (Wald test) [19]. Differentially expressed genes (DEGs) were defined by |log2 fold change| > 1 and P < 0.05 . Heatmaps were generated to visualize DEG patterns.

GO and KEGG pathway enrichment analysis

Functional enrichment was conducted with the clusterProfiler R package [20], including Gene Ontology (GO) terms—biological process (BP), cellular component (CC), molecular function (MF)—and Kyoto Encyclopedia of Genes and Genomes (KEGG) pathways. Results were visualized with ggplot2.

Construction and validation of a GA-related risk-score model

Univariate Cox proportional hazards regression (R package survival [21], function coxph) was applied in the training cohort to identify prognosis-associated genes. To reduce overfitting, we then combined LASSO-penalized Cox regression and random forest (RF) for feature selection, followed by multivariable Cox modeling. In LASSO, the optimal penalty parameter λ was chosen by 10-fold cross-validation minimizing the cross-validated partial likelihood deviance. RF was used to rank variable importance, from which the top 10 genes were retained. We took the intersection of LASSO-selected genes and the RF top-ranking genes to obtain key candidates, which were then entered into a multivariable Cox model to derive the final risk signature. An individual risk score was calculated for each sample; patients were split at the median risk score into high- and low-risk groups. Kaplan–Meier survival curves and time-dependent ROC analyses were used to evaluate prognostic performance in both the training and validation cohorts.

Nomogram development

A prognostic nomogram for 1-, 3-, and 5-year overall survival (OS) was built using the rms package based on Cox modeling. The concordance index (C-index) quantified discrimination, and calibration curves assessed agreement between predicted and observed OS at 1, 3, and 5 years.

Pathway activity and enrichment based on GSVA and GSEA

To explore functional differences between risk groups, we performed Gene Set Variation Analysis (GSVA) [22] and Gene Set Enrichment Analysis (GSEA) using KEGG gene sets from MSigDB (category c2.cp.kegg). For GSVA, RNA-seq expression values were log2-transformed and pathway activity scores were computed with the GSVA package (Gaussian kernel). The top 10 differentially enriched pathways between risk groups were depicted as heatmaps using pheatmap. For GSEA, genes were ranked by log2 fold change (high vs. low risk), and 1,000 permutations were run with clusterProfiler; pathways with FDR < 0.05 were considered significantly enriched, and enrichment curves were plotted.

Tumor microenvironment and immune-cell infiltration

We estimated tumor microenvironment status using the ESTIMATE [23] R package, obtaining ESTIMATEScore, ImmuneScore, and StromalScore per sample. Spearman’s rank correlation assessed associations between the risk score and these metrics; Wilcoxon rank-sum tests compared scores between risk groups.

Immune-cell composition was inferred using CIBERSORT [24] (LM22 signature, 1,000 permutations). Only samples with CIBERSORT P < 0.05 were retained for downstream analyses. Group differences were visualized with ggplot2/ggpubr and tested with Wilcoxon tests (two-sided, P < 0.05). The prognostic relevance of immune infiltration in CRC is further supported by the international Immunoscore validation study [25].

Immune evasion and response to immunotherapy

We applied TIDE (http://tide.dfci.harvard.edu) to estimate each sample’s TIDE score and predicted response to immune checkpoint blockade (ICB; Responder vs. Non-responder). Higher TIDE scores indicate stronger immune-evasion potential and poorer predicted ICB response. We compared mean TIDE scores and the proportion of predicted Non-responders between risk groups, and contrasted risk scores between predicted Responders and Non-responders.

Drug-sensitivity prediction

Using oncoPredict [26] and drug-response data from GDSC, we predicted IC50 values for multiple agents in TCGA-CRC samples to evaluate the association between the GA-related risk score (GARS) and therapeutic response. We compared IC50 distributions between risk groups, computed Spearman correlations between risk score and IC50, and adjusted for multiple testing using the Benjamini–Hochberg FDR (significant at FDR < 0.05). Volcano plots and boxplots were generated with ggplot2.

Statistical analysis

All analyses were performed in R (version 4.4.3). Unless otherwise specified, P < 0.05 was considered statistically significant, with multiple-testing correction applied where appropriate.

Results

Study cohorts and analysis workflow

A total of 633 CRC patients in the training cohort (TCGA-COAD/READ) and 585 in the validation cohort (GSE39582) were included. The overall study design is illustrated in Fig. 1, and the baseline clinical characteristics of both cohorts are summarized in Table 1.

Fig. 1.

Bioinformatics analysis workflow

Table 1.

Clinical features of the cohorts

Cohort	TCGA-COAD/READ	GSE39582
Number of patients	633	585
Age (mean ± SD)	66.71 ± 12.81	66.95 ± 13.17
Follow-up time (mean ± SD) (days)	813.70 ± 734.21	1760.22 ± 1179.72
Follow-up status
Alive	499 (78.83%)	385 (65.81%)
Dead	130 (20.54%)	194 (33.16%)
Gender
Male	335 (52.92%)	0 (0.00%)
Female	294 (46.45%)	0 (0.00%)
Clinical stage
Stage I	108 (17.06%)	38 (6.50%)
Stage II	229 (36.18%)	271 (46.32%)
Stage III	181 (28.59%)	210 (35.90%)
Stage IV	90 (14.22%)	60 (10.26%)
Unknown	25 (3.95%)	6 (1.03%)
T stage
T1	20 (3.16%)	12 (2.05%)
T2	109 (17.22%)	49 (8.38%)
T3	428 (67.61%)	379 (64.79%)
T4	70 (11.06%)	119 (20.34%)
Unknown	6 (0.95%)	26 (4.44%)
M stage
M0	467 (73.78%)	499 (85.30%)
M1	89 (14.06%)	61 (10.43%)
Unknown	77 (12.16%)	25 (4.27%)
N stage
N0	357 (56.40%)	314 (53.68%)
N1	151 (23.85%)	137 (23.42%)
N2	118 (18.64%)	100 (17.09%)
N3	0 (0.00%)	6 (1.03%)
Unknown	7 (1.11%)	28 (4.79%)

Differential expression and functional annotation of Golgi apparatus (GA)–related genes in CRC

From the TCGA cohort, we identified 5,580 differentially expressed genes (DEGs) between tumor and normal colorectal tissues (2,749 up-regulated and 2,831 down-regulated; |log₂FC| > 1, P < 0.05). Intersecting these DEGs with 1,686 GA-related genes curated from MSigDB yielded 434 GA-related DEGs. The volcano plot summarizes the genome-wide distribution of fold changes and significances (Fig. 2A), in which GA-related hits are embedded within the overall signal and highlight both markedly up- and down-regulated subsets. A heatmap of representative features (Fig. 2B) shows that these GA-related DEGs capture a clear tumor–normal contrast, with samples clustering by disease status and coordinated shifts in expression across Golgi trafficking and glycosylation modules. The Venn diagram (Fig. 2C) quantifies the overlap between the GA gene set and the DEG list, underscoring that a substantial fraction of Golgi-associated genes are dysregulated in CRC.

Fig. 2.

A Volcano plot showing the distribution of differentially expressed genes. Green dots indicate down-regulated genes, red dots indicate up-regulated genes, and gray dots indicate non-significant genes. B Heatmap of differentially expressed Golgi-related mRNAs. C Venn diagram showing the overlap between Golgi-related genes and differentially expressed genes. The intersection identifies 434 overlapping genes that are crucial for further analysis. D GO enrichment of Golgi-related DEGs, including biological process (BP), cellular component (CC), and molecular function (MF). E KEGG enrichment overview of GA-related DEGs. F Bar plot showing the number of enriched genes across the top KEGG pathways

To contextualize these transcriptional changes, we performed GO and KEGG enrichment analyses on the 434 GA-related DEGs. GO terms were dominated by organelle- and processing-level functions, with “Golgi apparatus subcompartment” and “glycoprotein metabolic process” among the top categories (Fig. 2D). This pattern indicates that CRC perturbs both the structural compartments of the Golgi and enzymatic pathways governing protein modification. KEGG analysis further converged on cancer-relevant signaling and matrix remodeling, with “Proteoglycans in cancer” emerging as the most significantly enriched pathway, followed by the calcium signaling pathway (Fig. 2E, F). Given the central role of proteoglycan biosynthesis and editing in receptor signaling, adhesion, and extracellular matrix dynamics, these findings suggest that GA-centered remodeling may facilitate oncogenic signaling, invasive behavior, and immune evasion in CRC [27].

Development of a GA-based prognostic model

Univariate Cox regression identified 29 GA-related DEGs associated with OS (Fig S1A). To avoid over-fitting, we performed LASSO and Random Forest analyses; the LASSO coefficient trajectories and the 10-fold cross-validation profile for λ selection are shown in (Fig. 3A, B). The top-ranked 10 genes by Random Forest are depicted in (Fig S1B). Intersecting the LASSO-selected and Random-Forest-prioritized genes yielded 8 key candidates (Fig. 3C).

Fig. 3.

A LASSO coefficient profiles for the 29 prognostic genes. B Penalty plot for the LASSO model with error bars. C Venn diagram showing the intersection between genes selected by LASSO regression and the top 10 genes selected by random forest. D Forest plot showing the multivariate Cox analysis results for the six GA genes. E Distribution of risk scores and survival status in the training cohort. F Heatmap of the six model genes in the high- and low-risk groups of the training cohort. G Kaplan–Meier survival curves for overall survival in the high- and low-risk groups of the training cohort. H Time-dependent ROC curves for predicting 1-, 3-, and 5-year overall survival in the training cohort

In the multivariable Cox model, six genes—PCSK5, RAB36, CD36, DPP7, KPNA2 and HEPACAM2—remained independent predictors and were used to construct the GA-related risk score (GARS) (Fig. 3D). The formula for GARS was:

Patients in the training cohort were classified into high- and low-risk groups according to the median GARS. The distribution of risk scores and survival status is shown in Fig. 3E. The expression heatmap of the six model genes across the two risk groups is shown in Fig. 3F. Kaplan–Meier curves showed that patients with high GARS had significantly poorer OS than those with low GARS (Fig. 3G). In the training set, the log-rank P = 1.71 × 10⁻⁷; the hazard ratio for the high- versus low-risk group was 2.75 (95% CI, 1.85–4.08). The estimated survival probabilities in the high- versus low-risk groups were 84.8% vs. 96.0% at 1 year, 69.1% vs. 88.4% at 3 years, and 48.2% vs. 75.0% at 5 years. Time-dependent ROC curves indicated that the AUCs for 1-, 3- and 5-year survival were 0.679, 0.684 and 0.697, respectively (Fig. 3H). Together, these results support the prognostic and risk-stratification value of the six-gene GARS in the training cohort.

External validation of the GARS

To assess the robustness of the GARS developed from the training cohort, the model was applied to the external GEO cohort (GSE39582). Using the same formula and median cut‑off, patients were stratified into high‑ and low‑risk groups. The distribution of risk scores and survival status in the validation cohort is shown in Fig. 4A, and the expression heatmap of the six model genes across the two risk groups is shown in Fig. 4B. Kaplan–Meier analysis demonstrated significantly poorer overall survival in the high-risk group than in the low-risk group (Fig. 4C), with a log-rank P value of 0.0063 and a hazard ratio of 1.49 (95% CI 1.12–1.99). Time-dependent ROC analysis yielded AUCs of 0.617, 0.605, and 0.611 for 1-, 3-, and 5-year overall survival, respectively (Fig. 4D). These findings indicate that the six-gene GARS retained reproducible prognostic and risk-stratification value in the external validation cohort.

Fig. 4.

A Distribution of risk scores and survival status in the validation cohort. B Heatmap of the six model genes in the high- and low-risk groups of the validation cohort. C Kaplan–Meier survival curves for overall survival in the high- and low-risk groups of the validation cohort. D Time-dependent ROC curves for predicting 1-, 3-, and 5-year overall survival in the validation cohort

Distribution of the risk score across clinical subgroups

Given the strong clinical impact of standard prognostic factors in colorectal cancer, we explicitly examined how the GA-related risk score (GARS) relates to—and performs within—key clinicopathologic variables routinely used for risk stratification, including AJCC stage, T, N, and M categories as well as age and sex [28, 29]. Consistent with anatomical disease burden, GARS varied significantly across overall stage (P = 2.9 × 10⁻⁶), T stage (P = 2.2 × 10⁻⁵), N stage (P = 3.1 × 10⁻⁸) and M stage (P = 0.0016), with a clear monotonic trend: higher stage/substage was associated with higher risk scores (Fig. 5A). By contrast, sex (P = 0.38) and age group (≤ 65 vs. > 65 years; P = 0.37) showed no difference in GARS, indicating that the signature is not merely a surrogate for demographics (Fig. 5A).

Fig. 5.

A Box plots showing distribution of risk scores clinical subgroups. B Kaplan-Meier OS curves by risk group for Stage III-IV. C, D Kaplan-Meier OS curves by risk group for patients aged ≤ 65 and > 65. E, F Kaplan-Meier OS curves by risk group for M0 and M1 stage

We next examined whether GARS retained prognostic value within clinically relevant subgroups. Kaplan–Meier analyses showed that high-risk patients had significantly worse overall survival in stage III–IV disease (Fig. 5B, P < 0.0001), in patients aged ≤ 65 years (Fig. 5C, P = 0.0021) and > 65 years (Fig. 5D, P < 0.0001), and in both the M0 (Fig. 5E, P = 0.00011) and M1 (Fig. 5F, P = 0.047) subgroups. Detailed analyses of sex, N stage, and T stage are provided in the supplementary materials (Fig S2). Together, these results indicate that GARS increases with advancing anatomic stage and, importantly, retains prognostic power across diverse clinical strata, complementing TNM rather than duplicating it. This aligns with current guidance that TNM remains the gold standard for baseline prognostication, while integrating molecular signatures can refine individualized risk.

Nomogram construction and evaluation

To account for established clinical determinants of outcome, we fitted a multivariable Cox model that simultaneously included: age (≤ 65 vs. > 65), AJCC stage (I–II vs. III–IV), T category (T1–2 vs. T3–4), N category (N0 vs. N1–2), M category (M0 vs. M1), and the continuous GA-related risk score (GARS). As shown in Fig. 6A, age (P < 0.001) and GARS (P < 0.001) were the strongest independent predictors of overall survival. Stage (P ≈ 0.017), T category (P ≈ 0.017) and N category (P ≈ 0.034) also remained independently significant, whereas M category did not reach significance (P ≈ 0.120) in this multivariable setting. The directions were consistent with clinical expectation: more advanced stage/T/N and higher GARS were associated with higher hazard, and patients > 65 years had worse survival. On this basis, we constructed a nomogram (Fig. 6B) that assigns weighted points to the above significant factors together with GARS to provide individualized probabilities of 1-, 3-, and 5-year overall survival (OS). The calibration curves in Fig. 6C show that model predictions are close to the 45° reference line at 1, 3, and 5 years, indicating good calibration (particularly in the mid-to-high probability range) and supporting its clinical utility.

Fig. 6.

A Forest plot showing multivariable Cox analysis of clinical features (age, stage, T, N, M) and risk score. B Nomogram to predict 1-,3-, and 5-year overall survival. C Calibration curves for 1-,3-, and 5-yearoverall survival

Time-dependent ROC performance of the nomogram

Time-dependent ROC analyses demonstrated consistent discrimination of the nomogram across all horizons, with AUC ≈ 0.79 at 1, 3, and 5 years (Fig. 7A-C). In contrast, single predictors performed lower: age 0.57–0.60, T category 0.55–0.59, M category 0.62–0.67, overall stage 0.67–0.72, and the risk score alone 0.67–0.71. Thus, the integrated nomogram provided an absolute AUC gain of ~ 0.07–0.24 over individual clinical factors. By combining clinical stage, TNM components, age, and GARS, the nomogram showed improved and relatively stable discrimination compared with individual clinicopathological factors. These findings suggest that GARS may contribute to research-based risk stratification, although prospective validation is required before clinical application.

Fig. 7.

Time-dependent ROC curves of the nomogram and individual clinical predictors. A, 1-year OS; B, 3-year OS; C, 5-year OS

KEGG pathway differences between risk groups by GSVA and GSEA

To further clarify the biology distinguishing the two risk strata, we next profiled pathway activity using GSVA and GSEA. GSVA highlighted the ten KEGG pathways with the largest between-group differences, with the high-risk tumors generally showing higher activity across amino-acid metabolism, lysosome, immune/receptor signaling, and infection-related processes, whereas the low-risk group displayed the reciprocal pattern (Fig. 8A). GSEA indicated that high‑risk samples were enriched in pathways related to HTLV‑I infection, antigen processing and presentation, G‑protein‑coupled receptor (GPCR) signalling, skeletal muscle‑type calcium channel signalling and Wnt regulation. Low‑risk samples were enriched in pathways associated with homologous recombination, mitochondrial complexes, mRNA nuclear export, initiation of DNA replication and initiation‑complex formation (Fig. 8B, C).

Fig. 8.

A Heatmap of GSVA showing the top 10 KEGG pathways with significant differences between different risk groups. B GSEA enrichment curve for KEGG pathways significantly enriched in the high-risk group. C GSEA enrichment curve for KEGG pathways significantly enriched in the low-risk group

Association between the risk score and the tumor microenvironment

To investigate how the tumor immune microenvironment relates to the GARS-defined risk strata, we applied ESTIMATE and CIBERSORT analyses. Using ESTIMATE, the Risk Score showed significant positive Spearman correlations with ESTIMATE Score (ρ = 0.34, P < 0.001), Immune Score (ρ = 0.27, P < 0.001), and Stromal Score (ρ = 0.36, P < 0.001) (Fig. 9A-C). Consistently, when dichotomized by the median Risk Score, the high-risk group exhibited higher ESTIMATE, Immune, and Stromal Scores than the low-risk group (P < 0.001) (Fig. 9D), indicating greater infiltration of immune and stromal components in high-risk tumors. CIBERSORT deconvolution further delineated group-specific immune compositions (Fig. 9E, F). According to the asterisk-marked comparisons, the high-risk group showed significantly higher fractions of M0, M1, and M2 macrophages, monocytes, activated NK cells, and CD8 T cells, and significantly lower fractions of activated dendritic cells and CD4 T-cell compartments (activated memory and resting memory) (Fig. 9F). Taken together, these analyses indicate that the high-risk state is associated with a more infiltrated immune microenvironment characterized by myeloid enrichment (M0/M1/M2 macrophages and monocytes) alongside increased activated NK and CD8 T-cell fractions, whereas dendritic activation and CD4 memory compartments are reduced .These features provide immunologic context for the prognostic differences between risk groups and suggest potential implications for risk-adapted surveillance and immunomodulatory strategies.

Fig. 9.

A Spearman correlation analysis of the risk score with ESTIMATE score. B Spearman correlation analysis of the risk score with Immune Score. C Spearman correlation analysis of the risk score with Stromal Score. D Box plots comparing differences in ESTIMATE score, Immune Score and Stromal Score between high- and low-risk groups. E Stacked bar chart of immune cell infiltration proportions in high- and low-risk groups according to CIBERSORT analysis; colors indicate immune cell types and the y-axis shows relative proportions. F Box plots showing immune cell proportions with significant differences between high-and low-risk groups (*P < 0.05; **P < 0.01; ***P < 0.001))

Immune evasion propensity and predicted ICB response

By TIDE, high-risk patients had significantly higher TIDE scores than low-risk patients (Wilcoxon, P < 0.001; Fig. 10A), indicating stronger immune-evasion potential. The predicted Non-responder fraction to ICB was higher in the high-risk group, whereas the Responder fraction was higher in the low-risk group (Fig. 10B). Moreover, predicted Non-responders exhibited higher risk scores than predicted Responders (P < 0.001; Fig. 10C), linking elevated risk to reduced likelihood of ICB benefit.

Fig. 10.

A Violin plot showing differences in TIDE scores between low-risk (blue) and high-risk (red) colorectal cancer groups. B Bar chart comparing the proportions of responders and non-responders in high-risk and low-risk groups. C Box plot comparing the distribution of risk scores between predicted responders and non-responders according to TIDE prediction

Drug-sensitivity analysis

Using oncoPredict with GDSC drug-response data, we assessed associations between GARS and predicted IC50 values. Volcano plots summarized Spearman correlations and significance (Fig. 11A). Several notable associations (all FDR < 0.05) emerged: the risk score correlated positively with IC50 for GDC-0810 (an estrogen-receptor degrader/antagonist), Ulixertinib (ERK1/2 inhibitor), a TAF1 inhibitor (compound ID 5496 − 1732), Dihydrorotenone, Gallibscoquinazole, and Sinularin (ρ ≈ 0.4–0.5; FDR < 0.001), suggesting higher resistance in the high-risk group to these agents. Conversely, the risk score correlated negatively with BMS-754,807 (IGF-1R inhibitor; ρ ≈ −0.3, FDR < 0.001), indicating potential greater sensitivity among high-risk patients. Representative boxplots confirmed higher predicted IC50 values in the high-risk group for the former set of drugs and lower IC50 for BMS-754,807 (all Wilcoxon P < 0.001; Fig. 11B). Collectively, these findings suggest that high-risk patients may derive limited benefit from ICB yet display distinct sensitivity/resistance patterns to specific small-molecule agents, informing individualized therapy.

Fig. 11.

A Volcano plot illustrating the results of Spearman correlation analysis between risk scores and drug sensitivity; the x-axis shows correlation coefficients and the y-axis shows -log10 (FDR). B Multiple box plots comparing predicted IC50 values of representative drugs (e.g., Ulixertinib, GDC0810, etc.) between high-risk and low-risk groups

Discussion

Colorectal cancer remains a major public health concern, and its burden is still rising, particularly among younger adults. It is the third most commonly diagnosed cancer worldwide and the second leading cause of cancer‑related deaths [30]. Although screening and improved management have reduced incidence and mortality in adults over 50, incidence rates in individuals under 50 have nearly doubled since the early 1990s, with 10–12% of new CRC cases now occurring in younger adults [30]. These trends are largely driven by modifiable lifestyle factors: Western dietary patterns characterised by high intake of red and processed meats, fat and sugar and low intake of fibre, along with obesity, metabolic syndrome, sedentary behaviour, smoking and alcohol consumption, increase CRC risk [30]. Conversely, regular physical activity and diets rich in fruits, vegetables and whole grains are associated with lower incidence and improved survival [30, 31]. The rising disease burden and the influence of modifiable risk factors underscore the need for reliable biomarkers to identify high‑risk individuals and to guide prevention and treatment strategies.

Our data demonstrate that the Golgi apparatus is a critical yet underappreciated determinant of CRC prognosis. This GA-centred signature provides a perspective complementary to existing CRC prognostic models by linking Golgi-related protein processing and secretion to tumour microenvironment remodelling. By intersecting GA‑related genes with differentially expressed genes in CRC, we identified 434 GA‑related DEGs enriched in organelle subcompartments, glycoprotein metabolism and proteoglycan pathways [28]. This enrichment underscores that GA structure and biosynthetic functions are broadly perturbed in CRC. In particular, the prominence of “Proteoglycans in cancer” suggests that altered glycosaminoglycan biosynthesis and decorin/lumican signalling contribute to tumor growth and angiogenesis, consistent with proteoglycans acting as co‑receptors and modulators of receptor tyrosine kinases [28].

The six‑gene GARS exhibited robust prognostic performance across cohorts. High‑risk patients had nearly threefold higher mortality in the training cohort and 1.5‑fold in the validation set, with time‑dependent AUCs around 0.68–0.70. Multivariate analyses confirmed that the GA score adds prognostic value beyond age and TNM stage, aligning with the view that molecular signatures can refine risk beyond anatomical staging [32]. Incorporation of the GA score into a nomogram improved discrimination (AUC ≈ 0.79), enabling individualized survival estimates. These results highlight that GA‑dependent processes have translational relevance for patient stratification.

Functional analyses revealed that high‑risk tumors activate pathways related to GPCR signalling, calcium channels and infection, whereas low‑risk tumors enrich DNA‑repair and mitochondrial functions. Notably, proteoglycan biosynthesis and glycoprotein metabolic processes were enriched in high‑risk tumors [28]. Aberrant glycosylation is a hallmark of cancer and drives invasion and immune escape [33], and recent reviews emphasise that targeting glycosylation can reshape the tumor microenvironment and augment immunotherapy [34]. Our findings suggest that high‑risk tumors rely on GA‑driven glycosylation and proteoglycan remodeling, which may create permissive niches for invasion and metastasis.

The six genes in our signature appear to converge on a coordinated malignant program centered on Golgi-associated cargo processing, vesicular trafficking, and tumour–microenvironment interactions [35]. Within this framework, PCSK5, RAB36, and CD36 collectively point to proteolytic activation, vesicle trafficking, and lipid-supported invasive remodelling, processes that are closely linked to extracellular matrix reorganization and metastatic dissemination [36–38]. In particular, RAB36 has been implicated in colorectal cancer metastasis, whereas CD36 promotes CRC invasion and metastatic spread through the MMP28/E-cadherin axis. DPP7 further connects this signature to tumour cell survival and immune regulation, as recent evidence indicates that DPP7 supports CRC proliferation, suppresses apoptosis, and attenuates NK cell-mediated cytotoxicity [39]. KPNA2 may amplify malignant progression by facilitating oncogenic nuclear transport and metastasis-associated transcriptional programs [40], whereas reduced HEPACAM2 expression may reflect loss of adhesion-associated or tumour-suppressive restraint and has been linked to more advanced tumour progression and poorer survival in CRC [41]. Thus, although not all six genes are canonical structural components of the Golgi apparatus, their combined expression pattern likely captures a Golgi-centered phenotype characterized by aberrant protein processing, trafficking-dependent invasive remodeling, and immune dysregulation in colorectal cancer.

High‑risk tumors exhibited higher ESTIMATE‑derived immune and stromal scores and greater infiltration of monocytes, macrophages and activated NK cells, yet reduced dendritic and CD4 memory T‑cell fractions. Tumor‑associated macrophages polarise into M2 phenotypes that secrete IL‑10, TGF‑β and arginase, promoting tumor progression and immunosuppression [42]. Myeloid‑derived suppressor cells accumulate in cancer, suppress T and NK cells by depleting arginine and producing inhibitory cytokines, and are associated with poor prognosis and resistance to therapy [43]. Thus, the myeloid‑dominated milieu in high‑risk CRC may underlie their higher TIDE scores and predicted non‑response to immune‑checkpoint blockade [44]. Conversely, low‑risk tumors show higher infiltration of CD8⁺ and CD4⁺ T cells and dendritic cells [45], which correlates with better prognosis and may explain their higher predicted responsiveness to immunotherapy [45]. High ImmuneScore and StromalScore in the high-risk group indicate abundant immune and stromal cell infiltration, yet this infiltration is largely composed of immunosuppressive myeloid populations. Studies show that M2-polarized macrophages are the dominant tumor-associated macrophages in CRC; they display anti-inflammatory and immunosuppressive phenotypes and suppress effector T-cell responses [46]. Likewise, DPP7-positive tumor-associated macrophages enhance fatty acid β-oxidation and drive CD8 + T-cell exhaustion, generating an immunosuppressive microenvironment and reducing PD-1 blockade efficacy [47]. Thus, high ImmuneScore and StromalScore may represent infiltration of immunosuppressive cells, which explains why TIDE predicts poor immunotherapy responses despite seemingly rich immune infiltration.

Drug‑sensitivity analysis revealed that high‑risk tumors were predicted to be less sensitive to several agents but more sensitive to BMS‑754,807, an IGF‑1R/insulin‑receptor tyrosine‑kinase inhibitor. Preclinical studies show that BMS‑754,807 exerts potent antiproliferative effects on colon cancer cell lines and induces G₂/M arrest and cell death [48]. The association of high GARS with increased BMS‑754,807 sensitivity suggests that GA‑dependent signalling may converge on IGF‑1R pathways. Given the immunosuppressive microenvironment of high‑risk tumors, combining IGF‑1R inhibitors or glycosylation modulators with myeloid‑targeted therapies (e.g., CSF1R blockade) could be explored to enhance immunotherapy responses [34]. These associations are hypothesis-generating and should not be over-interpreted; they suggest candidate compounds for future validation and may prioritize agents for preclinical investigation rather than provide definitive clinical guidance.

Our study has limitations. First, it is retrospective and relies on publicly available transcriptomic datasets; prospective, multi‑centre validation is needed. Second, functional conclusions about GA dysregulation are based on bioinformatic inference and literature; experimental studies are required to delineate how the identified genes modulate Golgi structure, glycosylation and immune interactions. Third, the immune‑cell composition estimated by CIBERSORT and TIDE predictions may not fully capture spatial heterogeneity; integration with single‑cell transcriptomic or spatial profiling would refine insights. In particular, immune-related signals derived from bulk tissue expression profiles may partly reflect variation in surrounding non-tumour components, which cannot be fully disentangled in the present analysis. Fourth, we focused on transcriptome-level data; incorporating genomic, epigenetic and proteomic alterations of GA‑related genes could yield more comprehensive models. Despite these limitations, our findings suggest that GA‑centric processes shape CRC prognosis and therapeutic vulnerabilities. Furthermore, the TCGA and GSE39582 cohorts analysed were generated more than a decade ago using earlier nucleic-acid extraction and sequencing technologies. Technical artefacts inherent to these older platforms may affect analytical accuracy, and future studies incorporating more recent bulk transcriptomic datasets, single-cell analyses, and experimental validation may further improve the robustness of these findings.

Conclusion

This study suggests that GA-related gene expression is closely associated with CRC biology. A six-gene GA signature was associated with survival, distinct immune features, and differential predicted drug sensitivity in CRC. High-risk tumors were characterised by dysregulated glycosylation- and proteoglycan-related pathways together with a myeloid-dominated immune contexture, whereas predicted sensitivity to IGF-1R inhibition may merit further investigation. These findings provide a basis for further experimental validation of GA-related mechanisms and for evaluating the potential utility of GA signatures in future CRC studies.

Supplementary Information

Below is the link to the electronic supplementary material.

Author contributions

Conceptualization: Shuaiyi Wu; Methodology: Shuaiyi Wu; Formal analysis: Shuaiyi Wu; Visualization: Shuaiyi Wu; Writing – original draft: Shuaiyi Wu; Writing – review & editing: Shuaiyi Wu; Supervision: Shuaiyi Wu.

Funding

This research received no specific funding from any public, commercial or not-for-profit sectors.

Data availability

All data used in this study are publicly accessible. RNA‑seq and clinical data for colorectal cancer were obtained from the TCGA‑COAD and TCGA‑READ projects at the Genomic Data Commons (GDC) data portal, which is maintained by the U.S. National Cancer Institute. These datasets can be accessed at https://portal.gdc.cancer.gov/projects/TCGA‑COAD and https://portal.gdc.cancer.gov/projects/TCGA‑READ. The external validation dataset GSE39582 is available through the NCBI Gene Expression Omnibus (GEO) database (https://www.ncbi.nlm.nih.gov/geo/query/acc.cgi? acc=GSE39582). The list of Golgi apparatus–related genes was downloaded from the GOCC\_GOLGI\_APPARATUS gene set in the Molecular Signatures Database (MSigDB) (https://www.gsea‑msigdb.org/gsea/msigdb/human/geneset/GOCC\_GOLGI\_APPARATUS.html). Analysis code is available from the corresponding author on reasonable request.

Declarations

Ethics approval and consent to participate

The study analysed de identified data from public databases (TCGA and GEO); therefore, no human or animal experiments were conducted, and ethical approval was not required. Clinical trial number: not applicable.

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.