A risk stratification framework based on ammonia-induced cell death signatures for prognostic assessment and therapeutic insights in colorectal cancer

Abstract

Background

Ammonia, which was traditionally regarded as a metabolic by-product, has more recently emerged as a crucial regulator of tumor metabolism and immune dysfunction. Nevertheless, the prognostic implications and therapeutic significance of ammonia-related genes in colorectal cancer (CRC) are still largely uncharted territory. This lack of exploration means that there is a significant gap in our understanding of how these genes may impact the prognosis of CRC patients and potentially serve as targets for new therapeutic strategies.

Methods

Transcriptomic and clinical data from the TCGA-COAD dataset were analyzed to pinpoint ammonia-related genes that showed differential expression. A prognostic risk signature was then developed by employing univariate Cox, LASSO, and multivariate Cox regression analyses. To assess the immune landscape characteristics and pathway alterations, techniques such as CIBERSORT, GSVA, TIDE, and TCIA were utilized. External validation of the gene expression patterns across different cell types was carried out using single-cell RNA-seq data from the GSE132465 dataset. The exploration of potential therapeutic targets was further advanced through artificial intelligence-based virtual screening on the DrugCLIP platform. Finally, the key genes were experimentally verified using RT-qPCR. This comprehensive approach aimed to understand the role of ammonia-related genes in CRC, from their identification to their potential as therapeutic targets, while validating findings at multiple levels.

Results

A three-gene signature consisting of PMM2, CLK2, and UCHL1, which is associated with ammonia-induced cellular stress, was identified. This signature was then utilized to construct a prognostic risk model. This model successfully classified CRC patients into high- and low- risk groups, with these groups showing significantly disparate survival outcomes. Through functional analyses, it was discovered that the high-risk group had an enrichment of pathways related to mitochondrial metabolism, proteostasis regulation, and oxidative stress responses. Immune analyses indicated that in the high-risk group, there was an increased infiltration of regulatory T cells and enhanced interferon - related signaling. This suggests the presence of an “inflamed but immunosuppressed” tumor microenvironment. Single-cell RNA-seq analysis further confirmed the celltype-specific expression patterns of the three genes within the tumor microenvironment. Additionally, artificial intelligence (AI)-based virtual screening identified candidate small molecules that are predicted to bind to the catalytic pocket of CLK2. This finding indicates that CLK2 has the potential to be a druggable metabolic target. Overall, these results provide valuable insights into the role of ammonia-related genes in CRC and offer potential avenues for prognostic assessment and therapeutic intervention.

Conclusions

This gene signature associated with ammonia offers a new paradigm for prognostic stratification in CRC. It mirrors the metabolic stress responses triggered by ammonia accumulation within the tumor microenvironment. These discoveries emphasize that metabolic adaptation related to ammonia could serve as a rich source of potential biomarkers. Moreover, they propose CLK2 as a highly promising therapeutic target, opening up new possibilities for more effective treatment strategies in CRC management.

Supplementary Information

The online version contains supplementary material available at 10.1007/s12672-026-05257-w.

Introduction

Colorectal cancer (CRC) is one of the most common malignancies worldwide. In 2022, there were approximately 1.9 million new cases and 0.9 million deaths. The incidence and mortality rates are particularly high in European and American countries, and by 2023 CRC had even become one of the leading contributors to the global cancer burden [1–3]. The prognosis of CRC largely depends on the stage at diagnosis: the five-year survival rate exceeds 90% for patients with localized tumors, whereas it drops to about 10% among the 20% of patients who present with distant metastases at initial diagnosis [4, 5]. The tumor microenvironment (TME) is considered to play a critical role in tumor initiation, progression, and metastasis. Recent studies have shown that the TME in CRC is characterized by immunosuppression and the accumulation of metabolic waste products. Metabolites generated by specific cells or gut microbiota within the TME may impair immune effector cell function, thereby promoting immune evasion [6]. Zhang et al. [7] identified the phenomenon of “ammonia-induced cell death (AID),” highlighting the significant metabolic and functional impact of ammonia, a metabolic waste product, on CD8 + T cells during antitumor immune responses.

Endogenous ammonia primarily originates from the metabolic byproducts of glutamine and aspartate catabolism [8, 9], while protein degradation by the gut microbiota represents another major source [10]. Studies have shown that tumor cells adapt to hypoxic environments by excessively absorbing ammonia, thereby promoting cellular proliferation, migration, and metastasis [11]. In addition, the accumulation of microbiota-derived ammonia can deplete T cells, impairing their capacity to eliminate metabolic waste and consequently increasing the risk of CRC [6]. Specific accumulation of ammonia further reshapes the TME, exacerbating T-cell exhaustion and enhancing tumor immune evasion [7]. Metabolic reprogramming is a hallmark of cancer, and recent studies have demonstrated that aberrant ammonia metabolism exerts a pivotal influence in cancers, particularly CRC [6, 12, 13]. Ammonia has thus emerged as a critical regulator of metabolic reprogramming in CRC; however, few studies have explored the relationship between ammonia-related genes and CRC, especially regarding their potential as prognostic biomarkers or predictors of individualized immunotherapy.

This study aimed to identify key ammonia-related genes and signaling pathways associated with CRC patients through bioinformatics approaches, construct a prognostic model, and further analyze the correlations with immune-related cell subsets and the potential for personalized immunotherapy, thereby providing important guidance for clinical decision-making.

Methods

Data collection

RNA-seq data and corresponding clinical information for CRC were obtained from the TCGA-COAD dataset of The Cancer Genome Atlas (TCGA; https://portal.gdc.cancer.gov), which includes 41 normal samples and 476 tumor samples. Data preprocessing and evaluation were performed using Perl (Strawberry-perl-5.30; https://www.perl.org) and R software (version 4.4).

Screening of ammonia-related genes

Ammonia - related genes were sourced from the GeneCards and Ma’ayanLab databases. In GeneCards (https://www.genecards.org/), within the Genes Search module, the keyword “ammonia” was entered, and the organism was set to Homo sapiens. Genes were ranked according to their Relevance score, and those having a Relevance score of at least 0.4 were selected, without imposing any restrictions on the gene category. In Ma’ayanLab (https://maayanlab.cloud/), human genes related to ammonia were directly retrieved using the same keyword, without applying any additional score - based filtering. Subsequently, the gene lists obtained from both databases were combined and duplicates were removed based on the official gene symbols. DEGs between CRC and normal tissues were identified using the “limma” package in R. Genes with an absolute log fold change (|logFC|) > 1.5 and a false discovery rate (FDR) < 0.05 were defined as DEGs.

Development of ammonia-based prognostic signatures

To further refine the 438 DEGs, univariate Cox regression analysis was performed with a significance threshold of p < 0.05 to identify potential ammonia-related biomarkers. To prevent overfitting, least absolute shrinkage and selection operator (LASSO) Cox regression was performed using the “glmnet” and “survival” packages in R. The optimal penalty parameter (λ) was determined by ten-fold cross-validation. To evaluate model robustness, the LASSO procedure was repeated for 1,000 random train–test splits. Genes selected by LASSO were subsequently incorporated into a multivariate Cox regression model to construct the ammonia-related gene signature.

Establishment of the ammonia-related gene prognostic signature

The ammonia-related gene risk score was calculated using the formula:

Here, n represents the number of ammonia-related mRNAs associated with CRC prognosis, and i denotes the i-th ammonia-related gene. Gene expression levels and corresponding regression coefficients are indicated by Geneexp and Coefi, respectively. Risk scores for individual patients were computed using the relevant functions within the “survival” R package. Based on the median ammonia-related gene risk score, patients with CRC were stratified into two subgroups: a high-risk group and a low-risk group. To assess the prognostic value of the ammonia-related gene signature, Kaplan–Meier (KM) survival analyses were performed to compare overall survival (OS) between the two risk groups in the TCGA cohort.

Single-cell analysis

To provide external validation and explore cell-type-specific expression patterns, single-cell RNA-seq data from the GSE132465 dataset were analyzed. The “Seurat” package was used for data normalization, dimensionality reduction, and cell-type annotation.

Validation of the independent prognostic value of the risk signature

To investigate the influence of ammonia-related genes on CRC progression, we further examined their associations with clinicopathological characteristics, including age, sex, T stage, N stage, M stage, tumor grade, and risk score. To ensure the robustness of the model, multivariate independent prognostic analyses were performed, and ROC curves for these clinical variables were generated using the “timeROC,” “survival,” and “survminer” packages.

Functional enrichment analysis

DEGs between the high- and low-risk groups were subjected to KEGG pathway enrichment analysis using the “clusterProfiler” R package. GSVA heatmaps were generated to visualize pathway activities and to compare pathway differences between the two risk groups.

Drug sensitivity analysis

The “OncoPredict” R package was used to evaluate the potential clinical applicability of ammonia-related genes in CRC treatment. OncoPredict predicts drug responses based on RNA-seq expression profiles. The calcPhenotype function was applied to estimate dose–response curves for commonly used chemotherapeutic agents using baseline tumor gene expression data from TCGA. Differences in IC50 values between risk groups were assessed using the Wilcoxon signed-rank test to identify statistically significant variations.

Immune microenvironment and immune function assessment

To quantify immune cell infiltration in CRC patients, the CIBERSORT R package was used to estimate the abundance of 22 immune cell types from gene expression data. Immune-related signatures from previous studies were incorporated, and their enrichment scores were calculated using GSVA. To systematically evaluate immune functional status within the tumor microenvironment, immune function quantification was performed based on predefined immune-activity gene sets. By integrating results from CIBERSORT and immune-function analyses, we comprehensively characterized the tumor immune microenvironment at both the immune-cell and immune-activity levels.

Immunotherapy and immune checkpoint inhibitors

To assess tumor immune-escape potential and predict response to immune checkpoint inhibitors (ICIs), the TIDE (Tumor Immune Dysfunction and Exclusion) algorithm was employed. TIDE integrates T-cell dysfunction and T-cell exclusion metrics to generate a TIDE score for each sample. To enhance predictive power for immunotherapy, we further obtained immunotherapy sensitivity scores from TCIA. Based on the Cancer Immunome Atlas framework, TCIA calculates an Immunophenoscore for each sample, which was used to estimate sensitivity to anti-PD-1 and anti-CTLA-4 therapies. To explore potential therapeutic vulnerabilities associated with the identified key genes, artificial intelligence (AI)-based virtual screening was performed using the DrugCLIP platform (https://www.drugclip.com).

Cell culture and reagents

The human normal colonic epithelial cell line NCM460 and CRC cell line HCT116 were purchased from Wuhan Boster Biological Technology Co., Ltd. Cells were cultured in DMEM (Gibco) supplemented with 10% FBS (Gibco) and 1% penicillin–streptomycin (Thermo Scientific). All cells were maintained in a humidified incubator containing 5% CO₂ at 37 °C.

RT-qPCR

Total RNA was extracted using TRIzol reagent (Invitrogen) according to the manufacturer’s protocol. One microgram of RNA was reverse-transcribed into cDNA using a commercial reverse transcription kit (e.g., HiScript or PrimeScript; manufacturer specified as required). RT-qPCR was performed on an ABI 7500 Real-Time PCR System (Applied Biosystems) using a SYBR Green–based detection kit. PCR conditions were as follows: 95 °C for 30 s, followed by 40 cycles of 95 °C for 5 s and 60 °C for 30 s. All samples were analyzed in triplicate. Relative gene expression was calculated using the 2⁻ΔΔCt method with GAPDH and β-actin as internal controls. Primers were synthesized by Yashang Biotechnology, and sequences are provided in the Table 1.

Table 1.

Primers for RT-PCR

Gene	Forward primer	Reverse primer
GAPDH	5’-TCAAGAAGGTGGTGAAGCAGG-3’	5’-TCAAAGGTGGAGGAGTGGGT-3’
CLK2	5’-AGCTACCATGTCCGTTCTCG-3’	5’-CTGCTGTTCTCCCGCTGATA-3’
PMM2	5’-AGCCAAGAAGAACGCATTGAG-3’	5’-TCCCATCCATCAGGAAAGACAT-3’
UCHL1	5’-ACGAATGCCTTTTCCGGTGA-3’	5’-GACTTCTCCTTGCTCACGCT-3’

Statistical analysis

Appropriate statistical methods were applied for quantitative analyses. Boxplots of ammonia-related gene expression were obtained from Human Protein Atlas (HPA), whereas mutation-associated survival curves were validated using the Kaplan–Meier Plotter platform [14] (accessed January 31, 2025). Differences in categorical clinical characteristics were assessed using the chi-square test. Kaplan–Meier survival curves were compared using the log-rank test. A p-value < 0.05 was considered statistically significant. Statistical analyses were conducted using R and GraphPad software. The workflow of the study is summarized in Fig. 1.

Fig. 1.

Research process flowchart

Results

Identification of ammonia-related differentially expressed genes

First, RNA-seq data from 41 normal tissue samples and 476 CRC samples were obtained from the TCGA database (Supplementary file 2: Table 2), along with clinical and prognostic information for CRC patients curated from cBioPortal (Supplementary file 3: Table 3). Subsequently, a total of 975 ammonia-related genes were integrated to form an ammonia-related gene set (Supplementary file 4: Table 4). To identify ammonia-related differentially expressed genes in the TCGA-COAD cohort, we performed differential expression analysis. As shown in the heatmap (Fig. 2A), the top 50 upregulated genes and the top 50 downregulated genes were selected for visualization. Based on the absolute value of the log fold change, a total of 438 DEGs were identified (Fig. 2B).

Fig. 2.

Identification of ammonia-related differentially expressed genes (DEGs) in colorectal cancer. A Heatmap visualization of the top 50 upregulated and top 50 downregulated ammonia-related genes in CRC tissues compared with normal tissues. The color scale from blue to red indicates low to high gene expression levels. B Volcano plot displaying the distribution of all ammonia-related genes. Red dots represent significantly upregulated genes, blue dots represent significantly downregulated genes, and gray dots represent genes with no significant difference (|logFC| > 1.5, FDR < 0.05)

Construction and validation of the ammonia-related risk model

DEGs in combination with patient survival data, we performed univariate Cox regression analysis with a significance threshold of p < 0.05, identifying 41 ammonia-related genes significantly associated with prognosis. Among them, nine genes with HR < 1 functioned as protective factors, whereas 32 genes with HR > 1 indicated increased risk (Table 1). These 41 genes were subsequently subjected to LASSO Cox regression with ten-fold cross-validation to determine the optimal penalty parameter (λ) (Fig. 3A, B). The LASSO analysis reduced the candidate gene set to five genes (Supplementary file 6: Table 5), which were further incorporated into multivariate Cox regression analysis, ultimately yielding a prognostic risk signature composed of three ammonia-related genes. The risk score was calculated as follows:

Fig. 3.

Construction and validation of the ammonia-related prognostic risk signature. A LASSO coefficient profiles of the candidate ammonia-related genes. B Selection of the optimal parameter (λ) in the LASSO model using 10-fold cross-validation. C, D Univariate (C) and multivariate (D) Cox regression analyses evaluating the independence of the risk score and clinical characteristics (Age, M stage, N stage, T stage, Gender, Stage) as prognostic factors. E A nomogram for predicting 1-, 3-, and 5-year overall survival (OS) probabilities in CRC patients, incorporating the risk score and other clinical parameters. F Calibration curves comparing the predicted survival probabilities with the observed survival rates at 1, 3, and 5 years. G Time-dependent ROC curves assessing the predictive accuracy of the prognostic model for 1-, 3-, and 5-year survival

To further evaluate the reliability of the risk score as an independent prognostic feature, we integrated the ammonia-related risk model with additional clinical parameters. Univariate (Fig. 3C) and multivariate (Fig. 3D) Cox regression analyses demonstrated that the risk signature possessed independent prognostic value comparable to established clinical variables, including age, T stage, tumor stage, and risk score. A nomogram incorporating these independent predictors was constructed (Fig. 3E), and calibration curves (Fig. 3F) confirmed its predictive accuracy. The ROC analysis showed that the AUC values for 1-, 3-, and 5-year survival all exceeded 0.6 (Fig. 3G), indicating acceptable discriminatory performance.

The entire cohort was then randomly divided into a training set and a testing set (Table 2) for internal validation. Randomization based on the three clinical subgroups demonstrated no significant baseline differences among the groups (p > 0.05, Table 2). Using the median risk score within each dataset, patients in the training, testing, and overall cohorts were categorized into high- and low-risk groups. As the risk score increased, the mortality rate of CRC patients correspondingly rose (Fig. 4A, B). PMM2 was highly expressed in the low-risk group, whereas CLK2 and UCHL1 were upregulated in the high-risk group (Fig. 4C). OS was significantly poorer in the high-risk group compared with the low-risk group (Fig. 4D). These findings were further confirmed in the testing cohort (Fig. 4E–H) and the validation cohort (Fig. 4I–L).

Table 2.

Integrated clinical data grouping validation for the total, training and validation cohorts

Covariates	Type	Total	Test	Train	Pvalue
Age	<=65	146(39.46%)	74(40.44%)	72(38.5%)	0.7839
Age	>65	224(60.54%)	109(59.56%)	115(61.5%)
M	0	317(85.68%)	158(86.34%)	159(85.03%)	0.8323
M	1	53(14.32%)	25(13.66%)	28(14.97%)
N	0	224(60.54%)	113(61.75%)	111(59.36%)	0.8255
N	1	83(22.43%)	41(22.4%)	42(22.46%)
N	2	63(17.03%)	29(15.85%)	34(18.18%)
T	1	8(2.16%)	4(2.19%)	4(2.14%)	0.9738
T	2	64(17.3%)	31(16.94%)	33(17.65%)
T	3	253(68.38%)	127(69.4%)	126(67.38%)
T	4	45(12.16%)	21(11.48%)	24(12.83%)
Gender	0	177(47.84%)	88(48.09%)	89(47.59%)	1
Gender	1	193(52.16%)	95(51.91%)	98(52.41%)
Stage	1	65(17.57%)	31(16.94%)	34(18.18%)	0.4513
Stage	2	151(40.81%)	79(43.17%)	72(38.5%)
Stage	3	97(26.22%)	50(27.32%)	47(25.13%)
Stage	4	57(15.41%)	23(12.57%)	34(18.18%)

Fig. 4.

Prognostic value analysis of the risk signature in the training, testing, and entire cohorts. A, E, I Distribution of risk scores and survival status of patients in the entire (A), training (E), and testing (I) cohorts. Red dots indicate deceased patients, and blue dots indicate surviving patients. B, F, J Risk score curves for the training (B), testing (F), and entire (J) cohorts. Patients are ranked by increasing risk score. C, G, K Heatmaps showing the expression profiles of the three signature genes (PMM2, CLK2, UCHL1) in the high- and low-risk groups across the training (C), testing (G), and entire (K) cohorts. D, H, L Kaplan-Meier survival curves comparing overall survival (OS) between high-risk (red) and low-risk (blue) groups in the training (D), testing (H), and entire (L) cohorts. Log-rank test p-values are shown

Pathway differences and drug sensitivity between risk groups

To identify the key pathways underlying the differences between the risk groups, we performed GSVA-based pathway enrichment analysis. The results revealed that the low-risk group was strongly enriched for ammonia-detoxification pathways, including KEGG_ALANINE_ASPARTATE_AND_GLUTAMATE_METABOLISM and KEGG_ARGININE_AND_PROLINE_METABOLISM, whereas the high-risk group showed marked enrichment in pathways such as KEGG_NEUROACTIVE_LIGAND_RECEPTOR_INTERACTION and KEGG_GLYCOSAMINOGLYCAN_BIOSYNTHESIS_CHONDROITIN_SULFATE (Fig. 5A). To further explore the cellular consequences of ammonia-associated metabolic dysregulation, we performed a focused GSVA analysis of pathways related to ammonia-induced cellular stress responses. The high-risk group exhibited enrichment of pathways associated with mitochondrial metabolism (oxidative phosphorylation), protein quality control (ubiquitin-mediated proteolysis and proteasome), and oxidative stress regulation (peroxisome) (Fig. 5B). We further evaluated drug sensitivity between the two groups and found that Alpelisib and MK-2206 exhibited greater predicted sensitivity in the low-risk cohort (Fig. 5C, D), while Gefitinib and Erlotinib were more sensitive in the high-risk cohort (Fig. 5E, F). These findings provide theoretical support for risk-based therapeutic decision-making in CRC.

Fig. 5.

Analysis of biological pathways and drug sensitivity associated with the risk signature. A GSVA heatmap illustrating the differentially enriched KEGG pathways between the high-risk and low-risk groups. Red indicates activation, and blue indicates suppression. B GSVA analysis reveals enrichment of ammonia-associated cellular stress pathways across risk groups. C–F Boxplots comparing the estimated IC50 values of chemotherapeutic and targeted drugs between low- and high-risk groups: Alpelisib (C), MK-2206 (D), Gefitinib (E), and Erlotinib (F). Statistical significance was determined by the Wilcoxon test

Immune correlation and immunotherapy response

To examine the relationship between risk groups and immune cell infiltration, we quantified immune cell proportions across patients. High-risk tumors demonstrated significantly elevated infiltration of naïve B cells, regulatory T cells (Tregs), and activated NK cells, whereas plasma cells, resting and activated CD4 memory T cells, and resting dendritic cells were more enriched in the low-risk group (Fig. 6A). Correlation analysis between risk score and immune functional subtypes further revealed that the high-risk group exhibited stronger activation in four immune functions: HLA expression, macrophage activity, Type I interferon response, and Type II interferon response (Fig. 6B). TIDE analysis showed significant differences in immunophenoscores between high- and low-risk groups (Fig. 6C). However, TCIA results indicated that the low-risk group was not sensitive to CTLA-4 or PD-1 immune checkpoint inhibitors, and unfortunately, the findings did not provide evidence supporting the effectiveness of CTLA-4 or PD-1 blockade in the high-risk group either (Fig. 6D–G). To further explore potential therapeutic implications of the identified key genes, AI-based virtual screening was performed using the DrugCLIP platform. The analysis identified multiple candidate small molecules predicted to bind the catalytic pocket of CLK2 (Fig. 6H) with strong docking scores ( < − 10.0, Table 3), suggesting that CLK2 may represent a potential druggable target associated with the ammonia-related risk model.

Fig. 6.

Immune landscape characterization and immunotherapy response prediction. A Comparison of the infiltration fractions of 22 immune cell types between high- and low-risk groups derived from CIBERSORT analysis. B Differences in immune-related functional scores between the two risk groups. Asterisks indicate statistical significance (*p < 0.05, **p < 0.01, ***p < 0.001). C Violin plot comparing the Tumor Immune Dysfunction and Exclusion (TIDE) scores between high- and low-risk groups. D–G Violin plots showing the distribution of Immunophenoscores (IPS) from The Cancer Immunome Atlas (TCIA) to predict response to CTLA-4 and PD-1 blockade therapies in high- and low-risk patients. H AI-based virtual screening identifies candidate small molecules targeting the catalytic pocket of CLK2

Table 3.

Predicted binding pocket of CLK2 identified by DrugCLIP (TOP 10)

Mol ID	Pocket ID	DCScore	Docking Score	Source
Z449591574	2	5.63	-10.78	REAL
ZINC001240181259	2	4.6	-10.64	ZINC
Z1502749297	2	5	-10.48	REAL
Z968582716	2	5.42	-10.48	REAL
Z1029946934	2	4.69	-10.47	REAL
ZINC000036360247	1	6.36	-10.37	ZINC
ZINC000013119367	2	5.03	-10.26	ZINC
ZINC001175030928	2	5.09	-10.2	ZINC
ZINC001236583167	1	6.91	-10.19	ZINC
ZINC001174542441	2	4.89	-10.17	ZINC

Validation of key genes

Among the three prognostic genes—PMM2, CLK2, and UCHL1—identified in the ammonia - related risk model, multiple experimental methods consistently verified their differential expression in CRC. Data from the Human Protein Atlas (Fig. 7A–F) and RT - qPCR (Fig. 8A–C) indicated that CLK2 was significantly overexpressed in CRC cells and tissues compared to normal controls. In contrast, PMM2 and UCHL1 were notably underexpressed. Moreover, single - gene survival analyses (Fig. 9A–C) and independent validation using the Kaplan–Meier database (Fig. 9D–F) consistently showed the prognostic significance of these genes. Overall, these results experimentally support the differential expression of the identified key genes in CRC and emphasize their potential biological importance in prognostic prediction.

Fig. 7.

Validation of protein expression levels of key genes using the Human Protein Atlas (HPA). Representative immunohistochemistry (IHC) staining images of PMM2 (A, B), CLK2 (C, D), and UCHL1 (E, F) in normal colon tissues (Left panels: A, C, E) and colorectal cancer tissues (Right panels: B, D, F). CLK2 show stronger staining in tumor tissues, while PMM2 and UCHL1 shows weaker staining in tumor tissues compared to normal controls

Fig. 8.

Experimental validation of key genes in CRC cell lines. A–C Relative mRNA expression levels of PMM2 (A), CLK2 (B), and UCHL1 (C) in normal colonic epithelial cells (NCM460) and CRC cells (HCT116) detected by RT-qPCR. Data are presented as mean ± SD. (*p < 0.05, ***p < 0.001)

Fig. 9.

Survival analysis of the three key ammonia-related genes. A–C Kaplan-Meier overall survival curves for PMM2 (A), CLK2 (B), and UCHL1 (C) based on the study cohort. D–F Validation of the prognostic value of PMM2 (D), CLK2 (E), and UCHL1 (F) using the Kaplan-Meier Plotter database. High expression of CLK2 correlates with poor prognosis, whereas high expression of PMM2 and UCHL1 correlates with better prognosis in this external dataset validation

Single-cell RNA-seq validation of the three-gene signature

To externally validate our findings and explore the cellular distribution of the identified genes, we further investigated the three - gene signature using the single - cell RNA - seq dataset GSE132465. Through UMAP (Uniform Manifold Approximation and Projection) analysis, we identified the major cell populations present in the dataset. These included epithelial cells, stromal cells, myeloid cells, T cells, B cells, and mast cells (as shown in Supplementary Figure S1A).We also observed that the relative proportions of these cell types varied between tumor and normal tissues (Supplementary Figure S1B). By conducting dot plot analysis, we were able to reveal the distinct cell - type - specific expression patterns of PMM2, CLK2, and UCHL1 within the tumor microenvironment (Supplementary Figure S1C). Significantly, differential expression of these genes was detected between tumor and normal epithelial cells, which validates the robustness of the proposed three - gene signature (Supplementary Figure S1D).

Discussion

AID arises chiefly from the accumulation of ammonia inside cells, which disrupts the functioning of mitochondria and other organelles. Due to its fundamental differences from earlier identified cell-death programs, AID has attracted more scientific attention [15]. The incidence of CRC, the leading cancer among men in Western nations, is on the rise, demanding more accurate and effective treatment strategies aligned with precision oncology [1, 16]. Tumor cells utilize metabolic plasticity to adapt to local microenvironment changes, such as hypoxia, enabling survival, invasion, and metastasis. This metabolic adaptability leads to significant metabolic diversity [17]. Abnormal metabolic reprogramming is acknowledged as a key factor in CRC progression and metastasis, involving the disruption of various metabolic pathways [18]. Ammonia is recognized as a crucial metabolic signaling molecule; when released by cancer cells into the surrounding microenvironment, it can influence the tumor microenvironment and encourage tumor growth [19, 20]. Ammonia, beyond its role as a by-product of energy metabolism, can also serve as a signaling mediator that controls mitochondrial release and the transfer of mitochondria between cells [21]. Beyond the discovery of molecular biomarkers, the advent of interdisciplinary approaches holds promise for further refining prognostic assessment in CRC. The recent progress in digital medicine has shown that artificial intelligence, in conjunction with computational pathology, can glean clinically relevant features from histopathological images. These technologies can also integrate multi - modal clinical data, thereby enabling more accurate risk stratification and providing better decision - making support for medical professionals [22, 23]. Meanwhile, mathematical modeling frameworks have increasingly been applied to characterize tumor dynamics, simulate tumor–microenvironment interactions, and predict therapeutic responses in cancer systems [24, 25]. Combining these computational and quantitative methods with molecular signatures, like the ammonia - related gene signature discovered in this research, could potentially lead to more accurate prognostic predictions. This integration may also streamline the development of more targeted and personalized treatment strategies for CRC patients. By leveraging the power of computational tools to analyze molecular data, medical professionals can gain a more comprehensive understanding of each patient’s unique cancer profile, enabling them to tailor treatment plans that are likely to be more effective and less prone to adverse effects.

In this research, we discovered and confirmed genes related to ammonia that are crucial for predicting CRC outcomes. We identified 438 differentially expressed genes linked to ammonia-related metabolic functions by analyzing large-scale transcriptomic data from the TCGA-COAD cohort. Following the selection of 41 genes linked to prognosis using univariate Cox regression, we utilized LASSO and multivariate Cox regression to develop a risk model based on AID, which includes three key genes: PMM2, CLK2, and UCHL1. This risk model effectively categorized patients into high-risk and low-risk groups, demonstrating a strong stratification ability. Survival analyses indicated a notable difference in overall survival between these groups, and the model showed strong predictive performance. We performed pathway enrichment (GSVA) and functional analyses across risk groups to elucidate the mechanisms behind the risk differences. The low-risk group showed a significant enrichment in pathways related to ammonia detoxification, such as alanine, aspartate, and glutamate metabolism, along with arginine and proline metabolism, indicating that low-risk tumors have a maintained ability to clear ammonia. Conversely, the high-risk group showed activation in pathways like neuroactive ligand–receptor interaction and glycosaminoglycan biosynthesis, yet lacked enrichment in standard ammonia-detoxification pathways, suggesting increased metabolic activity.

Our findings also indicated that the ammonia-related risk score was strongly linked to the immune landscape. The high-risk group showed a complex pattern with both inflammatory activation and immunosuppression, including Type I/II interferon response. The responses of Type I/II interferons were boosted, leading to increased infiltration of activated NK cells, but a significant buildup of Tregs might weaken antitumor immunity. For clinical application, we evaluated drug sensitivity with the OncoPredict algorithm and discovered that the low-risk group was more responsive to Alpelisib and MK-2206, while the high-risk group had a higher predicted response to Gefitinib and Erlotinib. In terms of immunotherapy, while TCIA didn’t offer direct proof of enhanced sensitivity to CTLA-4 or PD-1 blockade in the high-risk group, there was a notable difference in TIDE scores between the groups. Intriguingly, the high-risk group showed both augmented Treg (regulatory T-cell) infiltration and heightened Type I/II interferon responses. Generally, interferon signaling is linked to “hot” tumors, which are more likely to respond favorably to immune checkpoint blockade. However, the risk model did not foresee an increased sensitivity to PD-1 or CTLA-4 inhibition. This seemingly contradictory finding might signify a metabolically subdued immune microenvironment.It has been reported that ammonia accumulation can hamper T-cell effector functions and foster immunosuppressive conditions within the tumor microenvironment. In such circumstances, interferon signaling could imply chronic inflammatory activation rather than an effective antitumor immune response. The concurrent presence of interferon responses and increased Treg infiltration suggests that infiltrating T cells may experience functional exhaustion under metabolic stress. Consequently, the ammonia-associated signature may mirror an “inflamed yet suppressed” tumor microenvironment, where immune cell infiltration takes place, but the antitumor immunity remains functionally compromised. These findings, coupled with the augmented interferon-response signatures noted in the high-risk group, indicate that the risk model encompasses immune-related biological features. Despite the lack of significant differences in the response to immune checkpoint inhibitors, AI-driven virtual screening on the DrugCLIP platform identified several candidate compounds. These are predicted to bind to the kinase pocket of CLK2 with a high degree of affinity, suggesting that CLK2 could potentially serve as a therapeutic target.

Several prognostic signatures have been developed for CRC based on immune-related genes or metabolic pathways. Immune-based models mainly focus on the composition and functional status of tumor-infiltrating immune cells and have demonstrated predictive value for patient survival and response to immunotherapy [26]. Meanwhile, metabolism-related signatures have highlighted the critical role of metabolic reprogramming, including glucose and lipid metabolism, in CRC progression and prognosis [27]. More recently, integrative models combining immune and metabolic genes have been proposed to better capture the interaction between tumor metabolism and the immune microenvironment [28]. Unlike these pre - existing methods, the current study is uniquely centered on ammonia - associated biological processes. These processes are an area within tumor metabolism that has not been thoroughly investigated. As a result, the ammonia - related gene signature discovered in this research offers an additional viewpoint to the existing prognostic models that are based on the immune system or metabolism. This signature might also uncover specific metabolic weaknesses that are related to the development of CRC, filling a gap in our understanding of the disease’s metabolic landscape.

Our AID-based prognostic model highlighted PMM2, UCHL1, and CLK2 as crucial genes. Consistent with public datasets and earlier research, these genes were upregulated in CRC tissues and correlated with lymph-node metastasis, aggressive phenotypes, and poor outcomes [29–31]. Significantly, the three genes identified in our prognostic model should not be construed as direct enzymatic regulators of ammonia metabolism or ammonia - induced cell death. Rather, they probably signify transcriptional characteristics linked to the more extensive metabolic stress environment brought about by ammonia accumulation in the tumor microenvironment. Consequently, the signature mirrors downstream adaptive reactions to ammonia - related metabolic disruptions, rather than being direct elements of the ammonia metabolic pathway. PMM2 encodes phosphomannomutase 2, which is a crucial enzyme in glycoprotein biosynthesis and protein glycosylation. Glycosylation is of vital importance for maintaining proper protein folding and endoplasmic reticulum (ER) homeostasis. It has been reported that the accumulation of ammonia can disrupt protein-folding mechanisms and trigger ER stress, ultimately resulting in cellular proteotoxic stress responses. In this regard, the glycosylation processes associated with PMM2 might represent adaptive reactions to ammonia-related metabolic disruptions, rather than functioning as a direct regulator of ammonia metabolism. Thus, the protective role of PMM2 identified in our prognostic model could suggest that an increased capacity for glycoprotein processing enables tumor cells to alleviate the metabolic stress induced by ammonia. PMM2, an enzyme crucial for N-glycosylation, is essential for energy metabolism [29]. New findings highlight PMM2 as a component of a two-gene signature linked to glycolysis (P4HA1–PMM2) that predicts CRC survival, with immunohistochemistry verifying its increased expression in CRC tissues [32]. While this study didn’t measure intratumoral ammonia levels directly, GSVA analysis showed a notable enrichment of nitrogen-compound metabolism and lysosomal pathways in high-risk samples, indirectly indicating pathological ammonia accumulation. Research has previously indicated that high ammonia concentrations can lead to lysosome alkalinization and result in cytotoxic effects [15]. Therefore, elevated PMM2 levels might boost glycoprotein production and stabilize lysosomal membranes, aiding CRC cells in preserving balance during nutrient scarcity or stress caused by ammonia.

UCHL1, a ubiquitin carboxyl - terminal hydrolase, is involved in regulating ubiquitin - dependent protein degradation and cellular proteostasis. Ammonia toxicity has been linked to mitochondrial dysfunction, oxidative stress, and proteotoxic stress. These situations typically necessitate an up - regulation of the ubiquitin - proteasome system’s activity to uphold protein quality control. In this context, UCHL1 may be engaged in preserving proteostasis when the cell is under ammonia - related metabolic stress. Consequently, it indirectly mirrors the cell’s response to the accumulation of ammonia within the tumor microenvironment. UCHL1, an enzyme that removes ubiquitin, promotes tumor growth in CRC by stabilizing important signaling proteins. Bringing back UCHL1 decreases the breakdown of β-catenin. thus maintaining the activation of the Wnt/β-catenin/TCF pathway and encouraging proliferation, migration, and metastasis [31]. UCHL1 prevents the VHL-mediated breakdown of HIF-1α, thereby stabilizing hypoxia signaling pathways. In mouse models, abnormal expression of UCHL1 promotes distant metastasis, while blocking the UCHL1–HIF-1 interaction reduces metastatic colonization [24]. In clinical settings, UCHL1 expression is positively associated with HIF-1α, and elevated UCHL1 levels are indicative of a poor prognosis in breast, lung, and various other cancers [33]. In clinical settings, UCHL1 expression is positively associated with HIF-1α, and elevated UCHL1 levels are indicative of a poor prognosis in breast, lung, and various other cancers [34]. We hypothesize that in CRC, the overexpression of UCHL1 promotes aerobic glycolysis and glutamine metabolism, allowing tumor cells to endure in environments that are low in nutrients, oxygen-deficient, and high in ammonia.

CLK2, or CDC - like kinase 2, is a serine/threonine kinase responsible for regulating RNA splicing. It has been associated with metabolic signaling pathways, specifically the AKT - mediated metabolic adaptation. Prior research indicates that CLK2 is involved in cellular responses to metabolic stress through the modulation of transcriptome remodeling and activation of signaling pathways. Given that ammonia accumulation is known to disrupt cellular metabolic homeostasis and trigger adaptive signaling networks, CLK2 likely plays a role in transcriptional reprogramming as a response to ammonia - related metabolic stress. However, it does not directly partake in ammonia metabolism. CLK2, a kinase akin to CDC that targets serine/threonine, was significantly elevated in CRC and associated with poor chemotherapy responses and lower survival rates [30]. In chemoresistant CRC cells, CLK2 moves from the nucleus to the mitochondria, aiding in enhanced oxidative phosphorylation and inhibiting ferroptosis, a cell death process dependent on ROS [30]. Within the nucleus, CLK2 boosts the transcriptional activity of β-catenin and preserves the characteristics of cancer stem cells [35]. Since ammonia accumulation has been associated with mitochondrial dysfunction and metabolic stress within the CRC microenvironment, tumor cells may require adaptive metabolic mechanisms to tolerate ammonia-associated toxicity [6]. Given that CLK2 has been linked to the regulation of mitochondrial oxidative phosphorylation, it is plausible that CLK2 contributes to cellular adaptation to ammonia-induced metabolic stress by modulating mitochondrial metabolic flux. In this context, CLK2 may help tumor cells maintain mitochondrial energy metabolism and cope with the metabolic challenges imposed by ammonia accumulation [6, 9]. By enhancing mitochondrial oxidative processes and reducing ROS accumulation, CLK2 allows CRC cells to escape apoptosis and resist damage from drugs. Inhibiting CLK2 leads to impaired mitochondrial function, increased lipid peroxidation, and elevated ROS, and resensitizes resistant CRC cells to 5-FU and oxaliplatin [30]. Despite the potential of CLK2 inhibitors to oppose ABC transporters in lung cancer [36], CRC research indicates that dual targeting of CLK2 and OXPHOS results in synergistic cytotoxicity, suggesting that CRC cells with elevated CLK2 levels are metabolically reliant on oxidative phosphorylation.

In conclusion, we developed a strong risk stratification model using ammonia-induced cell death markers, offering a new perspective for prognostic evaluation in CRC. The three-gene signature identified in this study should be interpreted as an ammonia-associated transcriptional feature reflecting metabolic stress within the tumor microenvironment, rather than direct enzymatic regulators of ammonia metabolism. Notably, within this framework, PMM2, CLK2 and UCHL1 were validated as critical independent prognostic factors. The abnormal increase in their levels acts as a key mechanism connecting metabolic reprogramming to tumor growth, allowing CRC cells to endure metabolic stress, like ammonia buildup, by coordinating mitochondrial function and hypoxia signaling. As a result, this model provides an accurate method for forecasting patient survival and response to immunotherapy, while also emphasizing the therapeutic possibilities of targeting CLK2 and UCHL1. Future approaches should aim to leverage the metabolic weaknesses controlled by these main factors to defeat drug resistance and enhance precision cancer treatment in CRC.

Supplementary Information

Below is the link to the electronic supplementary material.

Author contributions

S.S.: Conceptualization, Writing – Original Draft, Figure Preparation.Y.Z.: Investigation, Pathological Assessment.Z.L.: Software, Formal Analysis, Experimental Validation.X.R.: Supervision, Writing – Review & Editing, Final Approval.All authors read and approved the final manuscript.

Funding

This research received no external funding.

Data availability

The RNA sequencing data and corresponding clinical information analyzed in this study were obtained from The Cancer Genome Atlas (TCGA) colorectal adenocarcinoma cohort (TCGA-COAD project), hosted by the Genomic Data Commons (GDC). https://portal.gdc.cancer.gov/projects/TCGA-COAD. Additional clinical and survival data were retrieved from cBioPortal for Cancer Genomics. https://www.cbioportal.org. Immunohistochemistry images were obtained from the Human Protein Atlas. https://www.proteinatlas.org. Immunotherapy response prediction data were accessed from TCIA and the TIDE platform.No new sequencing datasets were generated in this study.

Declarations

Ethics approval and consent to participate

This study utilized publicly available datasets and established human cell lines. No human participants or animal experiments were involved. The human cell lines (NCM460 and HCT116) were purchased from commercial sources and handled in accordance with institutional laboratory guidelines. Therefore, formal ethical approval was not required. This study was based exclusively on publicly available datasets and did not involve human participants or animal experiments requiring ethical approval. This study did not involve human subjects or patient recruitment.

Consent for publication

Not applicable. The manuscript does not contain any identifiable individual data.

Competing interests

The authors declare no competing interests.