Development and validation of a machine learning-based prognostic model using mitochondrial dysfunction-related genes for colorectal cancer patients

Introduction

Colorectal cancer (CRC) is the most prevalent form of cancer in the gastrointestinal tract. The most common type is colon adenocarcinoma, which originates in the glandular epithelium of the colon (1,2). According to the International Agency for Research on Cancer (IARC) in 2022, CRC is the third primary diagnosed cancer and the second most common cause of cancer-related deaths, with 1.92 million new cases and 904,000 deaths (3). Given its high morbidity and mortality rates, it is significant to discover new biomarkers and develop novel models for individualized therapy.

Mitochondria are crucial organelles that modulate metabolism, energy production, proliferation, and apoptosis (4,5). Mitochondrial dysfunction is a recognized characteristic of cancer (6). These changes often include mitochondrial abnormalities, insufficient mitochondrial copy numbers, accumulation of reactive oxygen species (ROS), aberrant energetic metabolism, and imbalance in biogenesis and mitophagy (7,8). Mitochondrial dysfunction can stimulate the proliferation and invasion of various cancer cells (9,10). Nonetheless, the function of mitochondria-related genes (MRGs) remains elusive. Given the involvement of MRGs in CRC development, it is vital to uncover new markers and develop prognostic models based on MRGs for CRC patients.

So far, many clinical features have been identified as prognostic factors for CRC, including age, sex, tumor stage, smoking, and drinking. However, these factors often lack sufficient sensitivity to reliably differentiate between patients with high and low survival rates or to predict the response to treatment. In contrast, prognostic models based on RNA sequencing (RNA-seq) expression profiles show great potential as more accurate predictive tools. Therefore, there is an urgent need in the genomic era to identify biomarkers that can simultaneously forecast overall survival (OS) and inform therapeutic strategies for CRC patients. Unfortunately, although multiple studies have constructed prognostic signatures based on messenger RNAs (mRNAs) to predict patients’ survival in various cancers (11-13), they either fail to reveal the prognostic characteristics of CRC or they focus on specific metabolomics.

Machine learning (ML) broadly means fitting predictive models to data or filtering informative groupings within data, which helps humans learn patterns from complex data to predict future behavioral outcomes and trends (14). ML aims to screen diagnostic biomarkers for tumors by combining with transcriptome sequencing. Previous research commonly used a single ML algorithm or two integrated ML algorithms to optimize variables (15). However, a single or only two integrated ML algorithms may miss important potential genes, while integrated ML methods have more advantages in variable screening and model construction (16). Integrating ML with gene expression of tumor cells is promising for tumor diagnosis, prognosis evaluation, and personalized therapy (17,18). In this paper, we analyzed MRG expression in CRC, aiming to develop an ML-based optimal diagnostic and prognostic model. The main part of this study is presented in Figure 1. We present this article in accordance with the TRIPOD reporting checklist (available at https://tcr.amegroups.com/article/view/10.21037/tcr-2025-1148/rc).

Figure 1 Key components of the research. DE-MRGs, differentially expressed mitochondria-related genes; GEO, Gene Expression Omnibus; GSE, Gene Set Enrichment; TCGA-COAD, The Cancer Genome Atlas Colon Adenocarcinoma.

Methods

Data collection

RNA-seq data and clinical information on CRC were obtained from The Cancer Genome Atlas Program (TCGA) database (https://portal.gdc.cancer.gov/). The GSE39582, GSE38832, and GSE17536 datasets and clinical data were acquired from the Gene Expression Omnibus (GEO) database (https://www.ncbi.nlm.nih.gov/geo/). MRGs were retrieved from the Genecards database (https://www.genecards.org/). The study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments.

Differential expression gene analysis

For TCGA RNA-seq data, genes with log₂ |fold change (FC)| >2 and adjusted P<0.05 were considered differentially expressed genes (DEGs). Intersections between DEGs and MRGs were then plotted as a Venn diagram. Using heatmaps, MRG expression in CRC was associated with patient clinical information.

Protein-protein interaction (PPI) network

PPI networks were constructed using STRING database 12.0 (https://cn.string-db.org/) and visualized with Cytoscape. PPI networks of the top 50% genes were shown using the STRING database and Cytoscape.

Functional enrichment analysis

The Gene Ontology (GO) system and Kyoto Encyclopedia of Genes and Genomes (KEGG) provide structured and standardized nomenclatures for describing gene functions and related biological pathways. The clusterProfiler is a well-established R package in the field of bioinformatics, designed for in-depth investigation of high-throughput biological datasets to reveal functional and pathway correlations. To determine the differences in biological functions and pathways, the identified differentially expressed MRGs (DE-MRGs) were analyzed for GO and KEGG enrichment using the clusterProfiler package. P<0.05 was set as the significance level. Through the Gene Set Enrichment Analysis (GSEA) website (https://www.gsea-msigdb.org), the functional pathways enriched in CRC were plotted as a heatmap using the GSEABase and GSVA packages.

Univariate Cox regression

Univariate Cox regression was leveraged to ascertain the independent prognostic value of DE-MRGs and clinicopathologic features in the survival time and status of CRC patients. P<0.05 implied statistical significance.

Prognostic risk modeling based on ML

Firstly, clinical information from the TCGA database was utilized for constructing our diagnostic model. Three distinct ML methodologies were employed to predict CRC progression. The support vector machine (SVM) is a well-known supervised tool in the ML field, mainly used for classification and regression. The recursive feature elimination (RFE) algorithm is designed to identify the most representative genes from the accumulated database. The SVM-RFE combines the basic principles of SVM with the iterative selection process of RFE to determine the key gene subset. Random forest (RF), as an ensemble ML algorithm, is widely used in feature screening and classification for model construction in biomedical research because of its excellent performance in processing high-dimensional data and nonlinear relationships. Logistic regression, as a classical probabilistic model, has the advantages of high predictive efficacy and interpretability in clinical diagnosis. Gene crossovers identified by all computational methods were preserved.

Immunohistochemistry (IHC)

IHC data from CRC patients were sourced from the Human Protein Atlas (HPA) (www.proteinatlas.org). The staining results for normal colon or rectum tissue and CRC tissues were downloaded. TPM2, GSTM1, CYP11A1, and LEP expression levels were detected using antibodies HPA047089, HPA048652, HPA016436, and CAB010490, respectively.

Risk assessment model

Least absolute shrinkage and selection operator (LASSO) regression is a variant of linear regression that can perform feature selection and model regularization simultaneously. This balance between data adherence and penalization can eliminate insignificant variables and keep only meaningful predictors. LASSO regression can identify genes with notable correlations with CRC utilizing the glmnet package. Individuals were assigned to high- and low-risk groups as per the median risk score. OS was estimated with the Kaplan-Meier (KM) method. The survival status was presented utilizing the pheatmap package. Receiver operating characteristic (ROC) curves were plotted, and the area under the curve (AUC) was computed. Clinicopathologic factors were combined with risk scores to construct nomograms to predict the 1/3/5-year OS rates in CRC patients. Line plots were constructed to forecast 1/3/5-year OS rates in CRC patients, considering clinicopathologic factors. The rms, regplot, and survival packages were adopted to construct column plots. The TCGA cohort (n=473) served as the training set, while three independent GEO datasets (GSE39582/GSE38832/GSE17536) functioned as validation sets. The expression metrics of selected genes were further validated in the GSE39582, GSE38832, and GSE17536 datasets. The TCGA cohort (n=473) served as the training set, while three independent GEO datasets (GSE39582/GSE38832/GSE17536) functioned as validation sets. The TCGA datasets incorporated OS and clinical features, including age, sex, stage, and T, N, M. The RNA-seq and OS of GSE39582 (n=544), GSE38832 (n=122), and GSE17536 (n=177) datasets were retrieved to validate the prognostic outcomes.

Immune infiltration analysis

The Estimate package v4.2.2 was employed to evaluate the stromal, immune, and estimate scores in the tumor microenvironment (TME) of CRC. The enrichment score of each immune cell was calculated according to marker genes to analyze the immune infiltration of 22 immune cells.

Immunoprognostic score (IPS)

IPS was calculated using The Cancer Immunome Database (https://tcia.at/patients) to predict the sensitivity of immunotherapy.

Drug sensitivity analysis

Drug response was predicted using the pRRophetic package, which integrated the Genomics of Drug Sensitivity in Cancer (GDSC) 2016 dataset and pre-trained elastic net regression models. For each drug in the GDSC library, the half maximal inhibitory concentration (IC50) values were calculated with extreme outliers (>99th percentile) truncated to enhance robustness. Patients were stratified into high- and low-risk groups. Differential drug sensitivity between groups was determined with the Wilcoxon rank-sum test (P<0.05). Significant drugs were visualized via boxplots using the ggpubr package.

Statistical analyses

Statistical analyses were performed in R software (4.2.2). Variables in the normal distribution were analyzed with the t-test. P<0.05 implied statistical significance.

Results

Identification of DE-MRGs in CRC

A total of 473 CRC samples and 41 normal tissue samples were acquired from the TCGA database. The differential analysis using the limma package revealed 2,657 DEGs (316 upregulated and 2,341 downregulated) (Figure 2A). These DEGs were intersected with 2,700 MRGs, resulting in 316 DE-MRGs (Figure 2B). Based on the STRING database and Cytoscape, a PPI network was mapped. The darker color indicated the greater neighboring nodes. The top 50% of genes with the most neighboring nodes are manifested in Figure 2C. The expression of DE-MRGs was presented in a heatmap (Figure 2D).

Figure 2 Identification of MRGs in CRC. (A) Volcano plot of the DEGs between CRC tissues and normal tissues. (B) Venn diagram of the MRGs and DEGs. (C) PPI networks show interactions between proteins encoded by the top 50% genes associated with the prognosis of mitochondrial dysfunction in CRC. (D) The heatmap depicted the expression patterns of DE-MRGs between CRC tissues and normal tissues. CRC, colorectal cancer; DE-MRGs, differentially expressed mitochondria-related genes; DEGs, differential expression genes; MRGs, mitochondria-related genes; N, lymph node; PPI, protein-protein interaction; T, tumor.

Functional enrichment analysis of DE-MRGs

KEGG pathway analysis revealed key pathways closely associated with mitochondrial dysfunction in CRC, with energy metabolism imbalance and cellular stress regulation as the core mechanisms (Figure 3A). Skeletal remodeling of cancer cells is a key step in invasion and metastasis. Mitochondria regulate cytoskeletal stability through calcium signaling, and insufficient energy supply may exacerbate skeletal disorders. Other significantly enriched pathways included adenosine monophosphate-activated protein kinase (AMPK), peroxisome proliferator-activated receptor (PPAR), and fatty acid degradation, suggesting that abnormal mitochondrial energy metabolism may be a driving factor for CRC progression. GO enrichment analysis demonstrated strong associations of mitochondrial dysfunction with oxidative stress imbalance, abnormal muscle contraction regulation, and membrane remodeling in CRC (Figure 3B). Significantly enriched entries focused on reactive oxygen metabolism [biological process (BP)], mitochondrial outer membrane structure [cellular component (CC)], and REDOX coenzyme binding [molecular function (MF)], suggesting that mitochondrial breakdown drives metabolic reprogramming in tumors. Based on TCGA, GSEA was conducted on the gene sets of “ADIPOGENESIS, ALLOGRAFT_REJECTION, HEME_METABOLISM, MYOGENESIS, PANCREAS_BETA_CELLS”. MRGs were greatly downregulated (Figure 3C), and a “HALLMARK MYC” signaling pathway was upregulated (Figure 3D) in CRC tissues.

Figure 3 Functional enrichment analysis of DE-MRGs. (A) KEGG enrichment analyses. (B) GO enrichment analyses. (C,D) GSEA enrichment analyses. The top 5 hallmark signaling pathways were downregulated, and one hallmark signaling pathway was upregulated in CRC tissues compared with normal tissues. The significance criteria were nominal P<0.05. BP, biological process; CC, cellular component; CRC, colorectal cancer; DE-MRGs, differentially expressed mitochondria-related genes; GO, Gene Ontology; GSEA, Gene Set Enrichment Analysis; KEGG, Kyoto Encyclopedia of Genes and Genomes; MF, molecular function.

Validation of prognostic signature

Clinicopathological variables containing survival time and status were combined for univariate Cox analyses (Figure 4A). Subsequently, it refined the 316 DE-MRGs to 25 genes. The SVM-RFE algorithm was leveraged for filtering DE-FRGs (Figure 4B). Then, feature importance was identified using RF (Figure 4C,4D). Ultimately, 18 genes were discerned as the optimal feature genes using logistic regression (Figure 4E). An intersection of outcomes from SVM-RFE, RF, and logistic regression highlighted 18 pivotal genes, depicted in a Venn diagram (Figure 4F).

Figure 4 Validation of prognostic signature. (A) Univariate COX regression analysis. (B) The curve of change in the predicted true and error value of each gene in the SVM-RFE algorithm. (C,D) The identification of feature importance based on random forests. (E) Logistic regression. (F) Venn diagram demonstrates the intersection of diagnostic markers obtained from the three algorithms. CI, confidence interval; HR, hazard ratio; OR, odds ratio; RMSE, root-mean-square error; SVM-RFE, support vector machine-recursive feature elimination.

Construction and validation of the prognostic risk model

A risk model was developed to ascertain the impact of DE-MRGs on CRC prognosis. 18 DE-MRGs identified as prognostic genes in the previous step were examined through LASSO regression (Figure 5A), resulting in 7 genes (Figure 5B). Risk score =0.034960948 × TPM2 + 0.009724934 × GSTM1 + 0.055765073 × CYP11A1 + 0.084320257 × SCN4A + 0.046370742 × LEP − 0.069588099 × PPARGC1A − 0.049269033 × NRG1.

Figure 5 Construction and validation of the prognostic risk model. (A,B) LASSO regression analysis of selected prognostic genes. (C) The KM curves of high- and low-risk groups in two datasets. (D) The ROC curves of 1/3/5-year in the TAGA database. (E,F) The risk curve, survival status plots, and heatmap of the prognostic risk model in the TAGA database. (G-I) Survival difference analysis between the distinct risk groups in the GSE39582, GSE38832, and GSE17536 sets. AUC, area under the curve; KM, Kaplan-Meier; LASSO, least absolute shrinkage and selection operator; ROC, receiver operating characteristic; TAGA, The Cancer Genome Atlas Program.

KM survival analysis (Figure 5C) illustrated that high-risk cohorts had worse prognoses in the training set. ROC curves (Figure 5D) demonstrated that AUC values for the 1/3/5-year OS were 0.689, 0.642, and 0.707 in the training set, respectively. Based on the risk score and survival status, mortality increased with increasing scores (Figure 5E). The heatmap illustrated that TPM2, GSTM1, CYP11A1, SCN4A, and LEP were upregulated, while PPARGC1A and NRG1 were downregulated in the high-risk group (Figure 5F). In the validation sets GSE39582 (n=544), GSE38832 (n=122), and GSE17536 (n=177), survival difference analysis showed that low-risk individuals had higher OS rates (Figure 5G,5I).

Immunohistochemical results

The above results were tested by downloading immunohistochemical staining results from the HPA database for normal colon or rectum tissue and CRC tissue. The results indicated that TPM2, GSTM1, CYP11A1, and LEP were highly expressed in CRC tissue (Figure 6).

Figure 6 IHC results of normal colon or rectum tissue and CRC tissue were downloaded from the HPA database. TPM2, GSTM1, CYP11A1, and LEP were detected using antibodies HPA047089, HPA048652, HPA016436, and CAB010490, respectively. CRC, colorectal cancer; HPA, Human Protein Atlas; IHC, immunohistochemistry.

Independent predictors and nomogram construction

To validate the prognosis, we included some clinical features, such as age, sex, stage, tumor (T), lymph node (N), metastasis (M), and risk score. Univariate and multivariate Cox analyses showed that the risk scoring model had good prognostic ability, and clinical characteristics were independent prognostic factors for CRC prognosis (Figure 7A,7B). Then a nomogram was developed based on these factors (Figure 7C,7D). The OS rate decreased as the risk score increased. The slope of the calibration curve was close to 1, suggesting that the predictions were true and reliable.

Figure 7 Independent prediction of risk models and nomogram construction. (A,B) Univariate and multivariate COX regression analysis of training sets. (C) Nomogram survival prediction of CRC patients with risk scores. (D) Calibration curves assessing 1/3/5-year OS consistency. **, P<0.01; ***, P<0.001; CI, confidence interval; CRC, colorectal cancer; HR, hazard ratio; M, metastasis; N, lymph node; OS, overall survival; T, tumor.

Immune infiltration analysis

An abundance of immune cell expression demonstrated 22 immune cells (Figure 8A). The proportions of resting mast cells, M0 macrophages, M2 macrophages, and regulatory T cells (Tregs) were greatly higher, while the proportions of activated dendritic cells, plasma cells, activated mast cells, and resting CD4 memory T cells were considerably lower in the high-risk group (Figure 8B). In summary, mitochondrial dysfunction reshapes the immune TME of CRC through metabolic reprogramming, leading to the enrichment of M2 macrophages and T-regulatory cells (Tregs) while suppressing the activity of anti-tumor immune cells. Targeting the mitochondrial-immune axis may provide a precision therapeutic approach for high-risk patients. The ESTIMATE algorithm showed that the high-risk group exhibited higher StromalScore, ImmuneScore, and EstimatedScore (Figure 8C). IPS analysis indicated that the low-risk group had higher IPS values (Figure 8D-8G), suggesting that low-risk patients may respond well to immunotherapy and may achieve better efficacy.

Figure 8 Immune infiltration analysis. (A) The abundance of 22 immune cells in high- and low-risk groups. (B) Discrepancies in the proportion of immune cells in the two groups. (C) StromalScore, ImmuneScore, and EstimatedScore between high- and low-risk groups. (D-G) The IPS analysis of the high- and low-risk groups. “ns” indicates not significant; **, P<0.01; ***, P<0.001; ****, P<0.0001. IPS, immunoprognostic score; TME, tumor microenvironment.

Drug sensitivity analysis

A total of 59 drugs showed substantially different IC50 values. The IC50 values for the top 12 drugs are displayed in Figure 9. The low-risk group demonstrated high IC50 values. Ponatinib (AP24534), all-trans retinoic acid (ATRA), SB505124 (TGF-betaR inhibitor), HG-6-64-1 [B-Raf activating factor (BRAF) inhibitor], elesclomol, shikonin, and vorinostat were associated with CRC treatment. Therefore, risk scores could forecast sensitivity to these drugs for CRC patients, and drug inhibitors might be more effective in high-risk samples.

Figure 9 The discrepancies in drug sensitivity between high- and low-risk groups. (A-L) The lower the log (IC50) value, the more sensitive the groups of patients were to the drug. IC50, half maximal inhibitory concentration.

Discussion

CRC shows high morbidity and mortality rates in the gastrointestinal tract worldwide (19). Now, many genetic markers have been established as prognostic biomarkers for CRC. For example, Huang et al. developed the hypoxia- and lactate metabolism-related molecular subtyping and prognostic signature, which exhibited favorable performance in predicting CRC prognosis (20). Mitochondrial dysfunction is a key attribute of cancer and is crucial for driving tumor progression, emerging as a critical target for cancer prevention and management. Prior studies have revealed basic mitochondrial functional differences between tumor and normal cells (6). Evidence also revealed a link between mitochondrial dysfunction and treatment efficacy (21). The prognostic models based on MRGs have exhibited favorable prediction power in bladder cancer (22), breast cancer (23), diffuse large B-cell lymphoma (24), and hepatocellular carcinoma (25). Despite growing evidence linking mitochondrial dysfunction to CRC aggressiveness (26), the prognostic potential of MRGs remains underexplored.

So far, many ML models have been developed to assist in the diagnosis and prognosis of CRC. For instance, several studies have focused on the role of m6A-related lncRNAs within the immune TME (27); surgical candidates using population-level data (28); non-invasive prediction of key mutations such as BRAF via radiomics (29); and prognostic models based on circulating tumor cell genes (30). While these approaches offer valuable insights into tumor behaviors, metastasis, and treatment responses, they also highlight a critical gap: the underexplored role of fundamental biological processes, such as mitochondrial dysfunction, in systemic prognostic modeling using ML. Therefore, this study developed an ML-based prognostic model using MRGs to stratify CRC patients and guide personalized therapeutic strategies. A total of 316 DE-MRGs were identified in CRC, and functional enrichment analyses revealed that mitochondrial dysfunction impacted CRC progression. DE-MRGs were significantly enriched in energy metabolism imbalance (e.g., AMPK and PPAR pathways) and fatty acid degradation, suggesting that mitochondrial dysfunction may drive tumor progression through metabolic reprogramming. Other enriched pathways included calcium signaling and oxidative phosphorylation (OXPHOS), which were associated with mitochondrial membrane potential dysfunction and ROS accumulation. The enriched pathways align with the Warburg effect (aerobic glycolysis) in CRC (31-33), suggesting mitochondrial dysfunction as a key driver of metabolic switching. ROS-related pathway enrichment supports mitochondrial damage-induced genomic instability, while apoptosis inhibition may enhance tumor cell survival. In previous studies, Liu et al. found a potential regulator that could induce ferroptosis resistance through the ROS/AMPK pathway in CRC cells (34). Cui et al. also pointed out that key pathways, including AMPK, link ferroptosis to CRC pathogenesis (35). Liang et al. concluded that a prospective target may modulate Erastin-induced ferroptosis through the AMPK pathway (36). Wang et al. discussed the roles of ROS and ferroptosis-associated genes in the early diagnosis and prognosis of CRC (37). The metabolic pathways enriched by DE-MRGs show close correlations with ferroptosis. Our study provides insights for future research on ferroptosis and mitochondrial dysfunction in CRC treatment, which could be further developed and combined with targeted therapy to optimize clinical efficacy.

ML has significantly transformed the landscape of biomarker identification by facilitating the integration of high-dimensional data and the development of predictive models. It offers several advantages in screening tumor biomarkers by integrating transcriptome data. Through integrative ML approaches (SVM-RFE, RF, and logistic regression) and LASSO, we identified a 7-gene prognostic signature (TPM2, GSTM1, CYP11A1, SCN4A, LEP, PPARGC1A, NRG1). The risk model based on the prognostic signature effectively stratified patients, with high-risk patients showing worse OS rates. Validation in independent GEO datasets (GSE39582, GSE38832, GSE17536) confirmed the model’s robustness (AUC 0.642–0.707 for 1/3/5-year OS).

These scientific findings hold significant insights and application potential for CRC management. First, these 7 MRGs show pivotal roles in early diagnosis, risk stratification, and prognostic evaluation of CRC. Second, a novel prognostic model may offer a more precise and reliable method of CRC evaluation and help healthcare staff better evaluate CRC progression and treatment regimens. Third, our findings offer references for the molecular mechanisms of CRC and are expected to reveal the potential roles of these MRGs in CRC progression. In conclusion, this study offers new methods of CRC prognostic assessment and valuable insights into precision medicine and personalized therapy. Prospective multicenter studies would prompt the generalizability of the findings and translate the findings into clinical practice through collaborative multicenter efforts.

Most of these 7 MRGs are closely associated with the development and prognosis of CRC and other tumors. TPM2 is a tumor suppressor gene that is frequently silenced by promoter hypermethylation in CRC, leading to reduced expression levels. TPM2 downregulation activates the RhoA signaling pathway to enhance tumor cell invasion and metastasis (38). TPM2 loss leads to dysregulated epithelial-mesenchymal transition and cytoskeletal dynamics, thus promoting CRC progression (39). CYP11A1, involved in cholesterol metabolism and steroidogenesis, influences CRC progression through bile acid metabolism and immune modulation. In the TME, CYP11A1-mediated steroidogenesis in T cells could suppress antitumor immunity and facilitate immune escape (40). LEP, an adipokine regulating energy balance, is overexpressed in CRC and correlates with advanced tumor stage and poor prognoses (41). LEP and LEPR levels correlate with the CRC TME. LEP levels are enhanced in CRC patients, indicating their potential to block CRC progression (42). NRG1 gene fusions are recognized as oncogenic drivers in CRC, pancreatic, gallbladder, and bladder cancers (43). Although the roles of GSTM1, SCN4A, and PPARGC1A in CRC remain unreported, our findings infer that they serve as potential markers for CRC prognosis.

Notably, the high-risk group showed an immunosuppressive microenvironment, characterized by high proportions of M2 macrophages and Tregs but reduced cytotoxic immune cell infiltration. This aligns with prior evidence that mitochondrial dysfunction promotes immune evasion via metabolic reprogramming. The combined enrichment of M2 macrophages and Tregs in CRC is attributed to multifaceted mechanisms, including cytokine feedback loops (44,45), STAT3-dependent signaling (46), metabolic competition (47), and checkpoint synergy (48). Targeting these pathways, such as combining STAT3 inhibitors with programmed death-1 (PD-1) blockade, may reverse immunosuppression and offer novel immunotherapeutic strategies for CRC.

Drug sensitivity analysis further revealed that high-risk individuals might respond better to immunotherapy and targeted therapy. Among them, ponatinib (AP24534) (49), ATRA (50,51), SB505124 (TGF-betaR inhibitor) (52), shikonin (53), vorinostat (54), HG-6-64-1 (BRAF inhibitor) (55), and elesclomol (56) are associated with CRC treatment, but they have not all been widely applied in clinical practice and lack experimental verification. Therefore, risk scores could predict sensitivity to these drugs for CRC patients, where drug inhibitors might be more effective against organisms or cell lines in high-risk samples, providing a bright prospect for the drug treatment of CRC.

Limitation

However, there are certain limitations in this study. Firstly, the model was trained on TCGA and GEO datasets, which may not fully represent real-world heterogeneity. Prospective multicenter validation is needed for validation. Most patients are either White, Black, or Latinx. While our multi-cohort validation demonstrates model consistency across different Western populations, we recognize the need for Asian-specific validation in future studies. Due to the limitations of the public database, the model does not account for potential confounding factors, such as lifestyle, comorbidities, or treatment adherence. Secondly, the model’s moderate AUC values (0.642–0.707) reflect the diagnostic and prognostic values of mitochondrial dysfunction in CRC. Though statistically significant, technical improvements in ML algorithms could enhance performance.

More assays are warranted to validate the properties of these MRGs and their patterns during CRC progression.

Conclusions

A prognostic model of 7 MRGs (TPM2, GSTM1, CYP11A1, SCN4A, LEP, PPARGC1A, NRG1) in CRC was constructed, with strong power in predicting OS. The risk score was closely linked to immune cell infiltration, immune function, and drug sensitivity. A novel diagnostic model based on the 7 MRGs was developed. Our findings may enhance accuracy and reliability in CRC diagnosis and prognosis, thus advancing personalized and effective interventions. Future prospective validation is required to confirm its clinical utility.

Acknowledgments

None.

Footnote

Reporting Checklist: The authors have completed the TRIPOD reporting checklist. Available at https://tcr.amegroups.com/article/view/10.21037/tcr-2025-1148/rc

Peer Review File: Available at https://tcr.amegroups.com/article/view/10.21037/tcr-2025-1148/prf

Funding: None.

Conflicts of Interest: All authors have completed the ICMJE uniform disclosure form (available at https://tcr.amegroups.com/article/view/10.21037/tcr-2025-1148/coif). The authors have no conflicts of interest to declare.

Ethical Statement: The authors are accountable for all aspects of the work in ensuring that questions related to the accuracy or integrity of any part of the work are appropriately investigated and resolved. The study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments.

Open Access Statement: This is an Open Access article distributed in accordance with the Creative Commons Attribution-NonCommercial-NoDerivs 4.0 International License (CC BY-NC-ND 4.0), which permits the non-commercial replication and distribution of the article with the strict proviso that no changes or edits are made and the original work is properly cited (including links to both the formal publication through the relevant DOI and the license). See: https://creativecommons.org/licenses/by-nc-nd/4.0/.

References

1. Jung G, Hernández-Illán E, Moreira L, et al. Epigenetics of colorectal cancer: biomarker and therapeutic potential. Nat Rev Gastroenterol Hepatol 2020;17:111-30. [Crossref] [PubMed]
2. Eng C, Yoshino T, Ruíz-García E, et al. Colorectal cancer. Lancet 2024;404:294-310. [Crossref] [PubMed]
3. Bray F, Laversanne M, Sung H, et al. Global cancer statistics 2022: GLOBOCAN estimates of incidence and mortality worldwide for 36 cancers in 185 countries. CA Cancer J Clin 2024;74:229-63. [Crossref] [PubMed]
4. Bock FJ, Tait SWG. Mitochondria as multifaceted regulators of cell death. Nat Rev Mol Cell Biol 2020;21:85-100. [Crossref] [PubMed]
5. Giacomello M, Pyakurel A, Glytsou C, et al. The cell biology of mitochondrial membrane dynamics. Nat Rev Mol Cell Biol 2020;21:204-24. [Crossref] [PubMed]
6. Zong Y, Li H, Liao P, et al. Mitochondrial dysfunction: mechanisms and advances in therapy. Signal Transduct Target Ther 2024;9:124. [Crossref] [PubMed]
7. Han M, Bushong EA, Segawa M, et al. Spatial mapping of mitochondrial networks and bioenergetics in lung cancer. Nature 2023;615:712-9. [Crossref] [PubMed]
8. Bao X, Zhang J, Huang G, et al. The crosstalk between HIFs and mitochondrial dysfunctions in cancer development. Cell Death Dis 2021;12:215. [Crossref] [PubMed]
9. Ikeda H, Kawase K, Nishi T, et al. Immune evasion through mitochondrial transfer in the tumour microenvironment. Nature 2025;638:225-36. [Crossref] [PubMed]
10. Mahmood M, Liu EM, Shergold AL, et al. Mitochondrial DNA mutations drive aerobic glycolysis to enhance checkpoint blockade response in melanoma. Nat Cancer 2024;5:659-72. [Crossref] [PubMed]
11. Sun M, Ma Y, Zhou F, et al. A nine-gene signature with potential targets for predicting the prognosis of patients with esophageal cancer. Transl Cancer Res 2025;14:4305-20. [Crossref] [PubMed]
12. Wang D, Wang R, Zhang Y, et al. Stemness-hypoxia genes PPARGC1A, PFKFB3, and SLC2A5 are associated with prognosis and tumor immune microenvironment in hepatocellular carcinoma. Transl Cancer Res 2025;14:4009-23. [Crossref] [PubMed]
13. Wang Z, Fu Y, Zhang H, et al. Diagnostic potential of the B9D2 gene in colorectal cancer based on whole blood gene expression data and machine learning. Discov Oncol 2025;16:1554. [Crossref] [PubMed]
14. Greener JG, Kandathil SM, Moffat L, et al. A guide to machine learning for biologists. Nat Rev Mol Cell Biol 2022;23:40-55. [Crossref] [PubMed]
15. Huang S, Cai N, Pacheco PP, et al. Applications of Support Vector Machine (SVM) Learning in Cancer Genomics. Cancer Genomics Proteomics 2018;15:41-51. [Crossref] [PubMed]
16. Zhang L, Liu Y, Wang K, et al. Integration of machine learning to identify diagnostic genes in leukocytes for acute myocardial infarction patients. J Transl Med 2023;21:761. [Crossref] [PubMed]
17. Choi RY, Coyner AS, Kalpathy-Cramer J, et al. Introduction to Machine Learning, Neural Networks, and Deep Learning. Transl Vis Sci Technol 2020;9:14. [Crossref] [PubMed]
18. Rauschert S, Raubenheimer K, Melton PE, et al. Machine learning and clinical epigenetics: a review of challenges for diagnosis and classification. Clin Epigenetics 2020;12:51. [Crossref] [PubMed]
19. Sung H, Ferlay J, Siegel RL, et al. Global Cancer Statistics 2020: GLOBOCAN Estimates of Incidence and Mortality Worldwide for 36 Cancers in 185 Countries. CA Cancer J Clin 2021;71:209-49. [Crossref] [PubMed]
20. Huang A, Sun Z, Hong H, et al. Novel hypoxia- and lactate metabolism-related molecular subtyping and prognostic signature for colorectal cancer. J Transl Med 2024;22:587. [Crossref] [PubMed]
21. Chiu HY, Tay EXY, Ong DST, et al. Mitochondrial Dysfunction at the Center of Cancer Therapy. Antioxid Redox Signal 2020;32:309-30. [Crossref] [PubMed]
22. Li J, Wang Z, Wang T. Machine-learning prediction of a novel diagnostic model using mitochondria-related genes for patients with bladder cancer. Sci Rep 2024;14:9282. [Crossref] [PubMed]
23. Yan LR, Wang A, Lv Z, et al. Mitochondria-related core genes and TF-miRNA-hub mrDEGs network in breast cancer. Biosci Rep 2021;41:BSR20203481. [Crossref] [PubMed]
24. Zhou ZZ, Lu JC, Guo SB, et al. A Mitochondria-Related Signature in Diffuse Large B-Cell Lymphoma: Prognosis, Immune and Therapeutic Features. Cancer Med 2025;14:e70602. [Crossref] [PubMed]
25. Wei Y, Ma L, Peng Q, et al. Establishing an oxidative stress mitochondria-related prognostic model in hepatocellular carcinoma based on multi-omics characteristics and machine learning computational framework. Discov Oncol 2024;15:287. [Crossref] [PubMed]
26. Feng WQ, Zhang YC, Xu ZQ, et al. IL-17A-mediated mitochondrial dysfunction induces pyroptosis in colorectal cancer cells and promotes CD8 + T-cell tumour infiltration. J Transl Med 2023;21:335. [Crossref] [PubMed]
27. Lan TH, Li W, Xu F, et al. Comprehensive analysis of m6A-related lncRNAs as prognosis factors in the immune microenvironment of colorectal cancer. J Gastrointest Oncol 2025;16:986-1000. [Crossref] [PubMed]
28. Jin CW, Lv SY, Yang C, et al. Most eligible candidates for primary tumor resection among metastatic colorectal cancer patients: a SEER-based population analysis. Transl Cancer Res 2025;14:4381-98. [Crossref] [PubMed]
29. Kang W, Ye X, Deng W, et al. Predictive value of CT-based imaging model for BRAF gene mutation in patients with colorectal cancer: a retrospective study. Transl Cancer Res 2025;14:3400-9. [Crossref] [PubMed]
30. Zheng Y, Yang M, Yi H, et al. Developing a prognostic risk model based on circulating tumor cell genes to predict prognosis and provide potential therapeutic strategies in colorectal cancer. Transl Cancer Res 2025;14:3096-112. [Crossref] [PubMed]
31. Zhong X, He X, Wang Y, et al. Warburg effect in colorectal cancer: the emerging roles in tumor microenvironment and therapeutic implications. J Hematol Oncol 2022;15:160. [Crossref] [PubMed]
32. Jing Z, Liu Q, He X, et al. NCAPD3 enhances Warburg effect through c-myc and E2F1 and promotes the occurrence and progression of colorectal cancer. J Exp Clin Cancer Res 2022;41:198. [Crossref] [PubMed]
33. Kasprzak A. Insulin-Like Growth Factor 1 (IGF-1) Signaling in Glucose Metabolism in Colorectal Cancer. Int J Mol Sci 2021;22:6434. [Crossref] [PubMed]
34. Liu MY, Li HM, Wang XY, et al. TIGAR drives colorectal cancer ferroptosis resistance through ROS/AMPK/SCD1 pathway. Free Radic Biol Med 2022;182:219-31. [Crossref] [PubMed]
35. Cui W, Hao M, Yang X, et al. Gut microbial metabolism in ferroptosis and colorectal cancer. Trends Cell Biol 2025;35:341-51. [Crossref] [PubMed]
36. Liang X, Zhang H, Shang W, et al. PPP2CA Inhibition Promotes Ferroptosis Sensitivity Through AMPK/SCD1 Pathway in Colorectal Cancer. Dig Dis Sci 2024;69:2083-95. [Crossref] [PubMed]
37. Wang Y, Zhang Z, Sun W, et al. Ferroptosis in colorectal cancer: Potential mechanisms and effective therapeutic targets. Biomed Pharmacother 2022;153:113524. [Crossref] [PubMed]
38. Cui J, Cai Y, Hu Y, et al. Epigenetic silencing of TPM2 contributes to colorectal cancer progression upon RhoA activation. Tumour Biol 2016;37:12477-83. [Crossref] [PubMed]
39. Zhou Y, Bian S, Zhou X, et al. Single-Cell Multiomics Sequencing Reveals Prevalent Genomic Alterations in Tumor Stromal Cells of Human Colorectal Cancer. Cancer Cell 2020;38:818-828.e5. [Crossref] [PubMed]
40. Mahata B, Pramanik J, van der Weyden L, et al. Tumors induce de novo steroid biosynthesis in T cells to evade immunity. Nat Commun 2020;11:3588. [Crossref] [PubMed]
41. Socol CT, Chira A, Martinez-Sanchez MA, et al. Leptin Signaling in Obesity and Colorectal Cancer. Int J Mol Sci 2022;23:4713. [Crossref] [PubMed]
42. Mhaidat NM, Alzoubi KH, Kubas MA, et al. High levels of leptin and non-high molecular weight-adiponectin in patients with colorectal cancer: Association with chemotherapy and common genetic polymorphisms. Biomed Rep 2021;14:13. [Crossref] [PubMed]
43. Jonna S, Feldman RA, Swensen J, et al. Detection of NRG1 Gene Fusions in Solid Tumors. Clin Cancer Res 2019;25:4966-72. [Crossref] [PubMed]
44. Schlechter BL, Stebbing J. CCR5 and CCL5 in metastatic colorectal cancer. J Immunother Cancer 2024;12:e008722. [Crossref] [PubMed]
45. Li Y, Zhang C, Jiang A, et al. Potential anti-tumor effects of regulatory T cells in the tumor microenvironment: a review. J Transl Med 2024;22:293. [Crossref] [PubMed]
46. Huang L, Zhao Y, Shan M, et al. Targeting crosstalk of STAT3 between tumor-associated M2 macrophages and Tregs in colorectal cancer. Cancer Biol Ther 2023;24:2226418. [Crossref] [PubMed]
47. Singer M, Zhang Z, Dayyani F, et al. Modulation of Tumor-Associated Macrophages to Overcome Immune Suppression in the Hepatocellular Carcinoma Microenvironment. Cancers (Basel) 2024;17:66. [Crossref] [PubMed]
48. Singh M, Morris VK, Bandey IN, et al. Advancements in combining targeted therapy and immunotherapy for colorectal cancer. Trends Cancer 2024;10:598-609. [Crossref] [PubMed]
49. Tan FH, Putoczki TL, Lou J, et al. Ponatinib Inhibits Multiple Signaling Pathways Involved in STAT3 Signaling and Attenuates Colorectal Tumor Growth. Cancers (Basel) 2018;10:526. [Crossref] [PubMed]
50. Liu W, Song Y, Zhang C, et al. The protective role of all-transretinoic acid (ATRA) against colorectal cancer development is achieved via increasing miR-3666 expression and decreasing E2F7 expression. Biomed Pharmacother 2018;104:94-101. [Crossref] [PubMed]
51. Hunsu VO, Facey COB, Fields JZ, et al. Retinoids as Chemo-Preventive and Molecular-Targeted Anti-Cancer Therapies. Int J Mol Sci 2021;22:7731. [Crossref] [PubMed]
52. Li K, Liu W, Yu H, et al. 68Ga-FAPI PET imaging monitors response to combined TGF-βR inhibition and immunotherapy in metastatic colorectal cancer. J Clin Invest 2024;134:e170490. [Crossref] [PubMed]
53. Zhang N, Peng F, Wang Y, et al. Shikonin induces colorectal carcinoma cells apoptosis and autophagy by targeting galectin-1/JNK signaling axis. Int J Biol Sci 2020;16:147-61. [Crossref] [PubMed]
54. Di Gennaro E, Piro G, Chianese MI, et al. Vorinostat synergises with capecitabine through upregulation of thymidine phosphorylase. Br J Cancer 2010;103:1680-91. [Crossref] [PubMed]
55. Tian J, Chen JH, Chao SX, et al. Combined PD-1, BRAF and MEK inhibition in BRAF(V600E) colorectal cancer: a phase 2 trial. Nat Med 2023;29:458-66. [Crossref] [PubMed]
56. Gao W, Huang Z, Duan J, et al. Elesclomol induces copper-dependent ferroptosis in colorectal cancer cells via degradation of ATP7A. Mol Oncol 2021;15:3527-44. [Crossref] [PubMed]