A prognostic model based on nucleotide metabolism genes in osteosarcoma Authors: Songli Ju, Lu Tao, Xuyan Li, Nijiao Huang, Xixian Ke Published: 2026 ABSTRACT AbstractOsteosarcoma typically arises during adolescence, posing a significant challenge. Despite comprehensive treatment strategies encompassing surgery, radiation therapy, and chemotherapy, which can notably enhance long-term survival rates among osteosarcoma patients, the 5-year survival rate for metastatic cases remains discouragingly low. Consequently, early diagnosis and prompt intervention are paramount in improving the prognosis of patients afflicted with this condition. Metabolic reprogramming holds paramount significance in the initiation and progression of tumors. In this meticulous investigation, we devised a risk prediction model that encompasses seven pivotal nucleotide metabolism-related genes: MYC, MUC1, IMPDH1, SAMHD1, NUDT13, UCK2, and NUDT16. This model was formulated leveraging six advanced machine learning algorithms. The results demonstrated that the risk prediction model exhibited robust prognostic predictive capability. Notably, patients identified with a high-risk phenotype exhibited a significantly lower long-term survival rate, coupled with elevated expression of immunosuppressive genes, highlighting the importance of metabolic reprogramming in influencing both survival outcomes and immune status. The multivariate Cox regression analysis confirmed that our model serves as an independent prognostic indicator, significantly impacting the long-term prognosis of osteosarcoma patients. Subsequently, we developed and validated a nomogram, which accurately predicts 1-, 3-, and 5-year survival rates for these patients. Furthermore, we compared chemosensitivity between high- and low-risk groups, gaining valuable insights into potential therapeutic differences. In conclusion, this model demonstrates superior prognostic predictive capability and holds promise in guiding chemotherapy treatment strategies for osteosarcoma patients, thereby enhancing treatment outcomes.Keywords: Osteosarcoma, Nucleotide metabolism genes, Risk model, Prognosis Introduction Osteosarcoma is a form of malignant bone tumor that predominantly affects teenagers, frequently arising in the femur, tibia, and humerus, characterized by rapid growth at the dry epiphyseal end of bones [1]. When combined with surgery and chemotherapy, the 5-year survival rate for patients with non-metastatic osteosarcoma can rise to an impressive 70% [2]. However, osteosarcoma has a high propensity for bone and lung metastasis. Once metastasis occurs, even with the administration of chemotherapy, neoadjuvant chemotherapy, radiotherapy, and immunotherapy, the 5-year survival rate for patients drops significantly, to less than 20% [3–5]. Metabolic reprogramming is a pivotal characteristic of tumor cells, with nucleotide metabolism standing as a crucial metabolic pathway among many substances. This metabolic form is intricately linked to the unchecked proliferation of tumor cells, as well as phenomena such as chemotherapy resistance, immune evasion, and metastasis within tumors [6]. At the genetic landscape, unraveling the intricate regulatory mechanisms of nucleotide metabolism-related genes within tumors represents a cornerstone for the development of innovative therapeutic targets. Increasingly, studies have illuminated the crucial function of these genes in orchestrating the onset and progression of malignancies, underscoring their paramount significance in both cancer research and the design of targeted therapies. For instance, CSN6 nucleotides have been found to initiate and foster the development of colon cancer, subsequently contributing to chemotherapy resistance in this malignancy. The precise mechanism involves CSN6 inhibiting the ubiquitination and degradation of the DDX5 protein, mediated by β-TrCP. This inhibition promotes the stabilization of PHGDH mRNA within the DDX pathway, which is a pivotalenzyme in the serine synthesis pathway. Consequently, this enhancement facilitates the synthesis of purines and pyrimidines, key components in nucleotide metabolism, thereby promoting tumor progression and drug resistance [7]. In HCC, CLOCK’s nuclear export acetylates PRPS, boosting de novo nucleotide synthesis and liver tumor growth [8]. Furthermore, recent studies have underscored the potential of nucleotide metabolism-related genes as biomarkers for constructing predictive models of tumor prognosis. For instance, Li et al., employed Lasso-Cox regression to develop a prognostic model comprising six nucleotide metabolism-related genes, specifically tailored for liver cancer patients. This model, upon validation, accurately identified patients with a higher risk of adverse outcomes, thereby demonstrating the clinical value of these genes in refining prognosis assessment and guiding personalized therapeutic strategies [9]. The deregulated expression of nucleotide metabolism-related genes plays a pivotal role in modulating the malignant phenotype of tumors, influencing their growth, invasion, and metastasis. In recent years, there has been a surge in the development of predictive models leveraging these genes to forecast the prognosis of cancer patients, underscoring their clinical significance. Nonetheless, despite this progress, a genetic risk model specifically focusing on nucleotide metabolism-related genes remains elusive in the context of osteosarcoma, highlighting an urgent need for further research in this area.Drawing upon the nucleotide metabolism-associated genes documented in the literature, this study conducted a comparative analysis of six machine learning algorithms, ultimately selecting the SpuerPC algorithm, which exhibited the highest AUC value, to devise a predictive risk model. This model encompasses seven pivotal nucleotide metabolism-related genes, enabling a more accurate prognosis stratification of patients into high- and low-risk groups. Furthermore, the study delves into the expression patterns of immune checkpoint genes and the infiltration profiles of immune cells in these distinct risk cohorts, providing valuable insights that can inform tailored chemotherapy strategies for high-risk patients. Methods Sample information database A comprehensive screening of the literature has yielded a total of 120 genes that are intricately linked to nucleotide metabolism [6, 10]. The GEO database was utilized to procure the GSE16088 dataset, comprising 14 osteosarcoma patients and 6 healthy individuals, for the development of a prognostic model, fourteen patients were used to construct the prognostic model. The six normal control samples in this dataset were only used to screen differentially expressed genes and did not participate in the subsequent model training. Subsequently, the GSE39055 dataset, encompassing 42 osteosarcoma patients, was downloaded for validation purposes. Both of these datasets encompassed comprehensive transcriptome profiles along with corresponding clinical information of the samples. Differential gene screening and prognostic model construction The “limma” package was employed to meticulously analyze the differential expression patterns of 120 nucleotide metabolism genes within the GSE16088 dataset. Specifically, we applied stringent criteria of logFC (logarithmic fold change) ≥ 1 and statistical significance at p < 0.05 to filter and identify genes that were differentially expressed between osteosarcoma patients and healthy individuals. This approach allowed for the precise discrimination of genes that may play pivotal roles in the pathogenesis or progression of osteosarcoma. Utilizing univariate Cox regression, alongside the “bootstrap” and “Boruta” algorithms, we meticulously filtered through differentially expressed genes to identify those that hold significant prognostic value for osteosarcoma patients. The specific approach was as follows: Based on the 34 differentially expressed genes selected from 120 nucleotide metabolism-related genes, we conducted feature selection through a three-step method to construct a robust prognostic model. First, we performed univariate Cox proportional hazards regression analysis on these 34 genes in the training set (GSE16088, n = 14) to screen out 12 candidate genes significantly associated with overall survival of patients (p < 0.05). To evaluate the robustness of these 12 candidate genes under sampling fluctuations in the training set, we used Bootstrap resampling. In each iteration-generated sample, we re-conducted univariate Cox regression analysis and recorded the number of times each gene was selected (p < 0.05). We filtered the genes based on their selection frequency in the resampling process, retaining those that could repeatedly validate their prognostic significance in multiple iterations. Through this robustness filtering, we ultimately obtained 9 genes for subsequent analysis. Finally, we applied the Boruta algorithm to perform the final feature selection on these 9 genes. Boruta is a “full-relevance” feature selection method based on random forests, which creates “shadow features” and conducts statistical tests to identify all features truly related to survival outcomes. This algorithm ensures that the genes finally selected all have reliable and non-random predictive capabilities [11]. After this step, we ultimately determined 7 key genes for constructing the prognostic model: MYC, MUC1, IMPDH1, SAMHD1, NUDT13, UCK2, and NUDT16. Six advanced machine learning algorithms—lasso, Elastic Net (Enet), Ridge, Cox partial least squares regression (plsRcox), CoxBoost, and supervised principal component analysis (SuperPC)—were subsequently applied. Each algorithm underwent rigorous evaluation based on the C-index, Integrated Brier Score (IBS), and Area Under the Curve (AUC) metrics to assess their predictive performance. The algorithm that demonstrated the highest C-index and AUC, coupled with the lowest IBS, was carefully selected as the optimal choice for constructing the prognostic model, ensuring the utmost precision and reliability in predicting outcomes. Accuracy test of prognostic model The GSE16088 dataset served as the training set, while the GSE39055 dataset functioned as the validation set. To establish an optimal cut-off point for the dataset, we leveraged the “surv_cutpoint()” function from the “survminer” package. The specific approach is to use the “surv_cutpoint()” function in the “survminer” package, with “minprop = 0.3” to determine the optimal cut-off value. The “surv_cutpoint” function traverses all possible cut-off values, evaluates the differences between the two groups based on the “maximal selection rank statistics” method, and identifies the point that maximizes the difference between the survival curves of the two groups after grouping (i.e., the point with the smallest log-rank test P value) as the optimal cut-off value. Group strictly using the unique cutoff value derived from the training set to ensure the independence of validation. Based on this cut-off, patients within the datasets were categorized into high-risk and low-risk groups. Subsequently, we employed the “ggplot2”, “ggpubr”, and “pheatmap” packages to generate insightful visualizations: a risk score area plot showcasing patient risk stratification, a survival time distribution plot illustrating patient outcomes, and a gene expression heatmap highlighting the differential expression patterns across the two datasets. Utilizing the “survminer” package, we conducted a thorough survival analysis of the high-risk and low-risk groups within both datasets. When evaluating the predictive accuracy of the risk model, we conducted the calculation through time-dependent ROC analysis. The endpoint event used in this analysis was the ‘overall survival’ of the patients. Specifically, we assessed the accuracy of the model’s risk score in predicting the 1-year, 3-year, and 5-year overall survival rates of the patients (i.e., whether the patients were still alive at the corresponding time points). The area under the ROC curve (AUC) was used to quantify the model’s discriminatory ability at different time points. This analysis was conducted separately in the training set and the independent validation set. Nomogram construction Employing the “glmnet” and “survminer” packages, we performed both univariate and multivariate Cox regression analyses to rigorously test the independence of our prognostic models. Based on the findings from the multivariate Cox regression, we crafted a precise nomogram plot. This nomogram serves as a valuable tool for predicting the 1-, 3-, and 5-year survival rates of osteosarcoma patients, offering clinicians and researchers a straightforward and intuitive method for assessing patient prognosis. To validate the predictive accuracy of the nomogram plot, we utilized the Receiver Operating Characteristic (ROC) curve. This comprehensive evaluation ensures that the nomogram provides reliable and accurate predictions, thereby reinforcing its clinical value in assessing the prognosis of osteosarcoma patients. KEGG and GO enrichment analysis Utilizing the “limma” package, we delved into the differentially expressed genes (DEGs) between the high-risk and low-risk groups, applying the criteria of |logFC|≥1 and p < 0.05. The results were visually represented in both volcano plots and heatmaps, offering a comprehensive view of the gene expression patterns that distinguish these two risk groups. To further understand the biological significance of the differentially expressed genes (DEGs), we conducted enrichment analyses using the KEGG database (https://www.genome.jp/kegg/) and the GO database (http://geneontology.org/). This approach allowed us to identify the enriched KEGG pathways and GO terms associated with the DEGs, thereby providing valuable insights into their potential roles and interactions within biological systems. Immune landscape analysis According to the immune checkpoint genes reported in the literature [12], To gain deeper insights into the role of immune checkpoint genes in osteosarcoma prognosis, we conducted a differential expression analysis between the high-risk and low-risk groups, focusing specifically on these genes. Following the identification of significantly differentially expressed immune checkpoint genes, we employed Pearson correlation analysis to explore the relationship between these genes and the risk factors. This approach allowed us to unravel potential associations between the immune checkpoint genes and patient prognosis, shedding light on their potential significance in osteosarcoma progression. To delve deeper into the link between our risk score and the immune microenvironment, we leveraged the CIBERSORT immune infiltration analysis algorithm along with Pearson correlation analysis. Chemosensitivity analysis By integrating drug information from the GDSC database with our risk assessment, we utilized the R package “oncoPredict” to conduct a thorough analysis of chemotherapeutic drug sensitivity between the high-risk and low-risk groups of osteosarcoma patients. Statistical analysis All analyses were performed with the use of R, version 4.0.2, and a p value of less than 0.05 was considered to indicate statistical significance. Results Model construction A total of 120 genes related to nucleotide metabolism were screened from the literature [6, 10], In osteosarcoma, 27 genes were up-regulated and 7 genes were down-regulated (Fig. 1A). The expression profiles of 34 differentially expressed genes are shown in Fig. 1B. Univariate COX regression screened out 12 nucleotide metabolism genes in GSE16088 dataset that significantly affected prognosis: MUC1, UCK2, NT5C1B, ENTPD3, NUDT16, ME1, NUDT1, IMPDH1, MY, NUDT13, ADK and SAMHD1 (Fig. 1C). To evaluate the prognostic robustness of 12 candidate genes, this study employed the Bootstrap resampling method for analysis. The results showed that 9 of these genes consistently (p < 0.05) correlated significantly with patient survival across multiple resampling iterations, indicating that their prognostic signals have high reproducibility and stability. These 9 genes were retained for subsequent feature selection and modeling (Fig. 1D). Finally, Boruta algorithm was used to screen out 7 nucleotide metabolism genes related to osteosarcoma survival, namely MYC, MUC1, IMPDH1, SAMHD1, NUDT13, UCK2 and NUDT16 (Fig. 1E). Using these seven signature genes as features, six distinct machine learning algorithms were employed for modeling: Lasso, Elastic Net (Enet), Ridge Regression, Cox Partial Least Squares Regression (plsRcox), CoxBoost, and Supervised Principal Component Analysis (SuperPC). These algorithms aimed to refine and optimize the predictive models for prognosis. The model based on SuperPC had the best C-index (0.971), the smallest IBS (0.119) and the largest AUC (AUC-1 year = 0.844, AUC-3 year = 0.813, AUC-5 year = 0.860). Therefore, SuperPC was selected to construct a risk model related to nucleotide metabolism (Fig. 1F). A risk prognosis model was constructed. A, B Differentially expressed genes related to nucleotide metabolism in osteosarcoma patients and normal individuals; C twelve genes related to nucleotide metabolism were identified by univariate Cox regression analysis; D bootsrap screened 9 genes related to the prognosis of osteosarcoma; E Boruta screening identified 7 genes related to nucleotide metabolism in osteosarcoma; F Six machine learning algorithms were used to construct prognostic models Accuracy test of prediction model The GSE16088 dataset served as the training set, while the GSE39055 dataset, comprising 42 osteosarcoma patients, was utilized as the validation set to assess the performance and generalizability of the models. Using the predict function to predict the risk score of each sample in the training set (GSE16088), the optimal cut-off value for the risk score was determined to be 0.492. Based on this, the patients in the training set were divided into high-risk and low-risk groups. Subsequently, this same cut-off value (0.492) was applied to the independent validation set (GSE39055), and the patients in the validation set were grouped according to their risk, and the risk factor triple plot revealed that patients in the high-risk group in the training set and test set had higher mortality, and MYC, IMPDH1, NUDT16, and UCK2 were highly expressed in both cohorts (Fig. 2A, B). Survival analysis unequivocally demonstrated that patients categorized into the high-risk group exhibited significantly reduced survival rates compared to those in the low-risk group, both in the training set (Fig. 2C) and the test set (Fig. 2D). This observation underscores the prognostic value of the risk stratification model employed in the study. The ROC curve analysis revealed impressive AUC values for the training set, with 0.857, 0.719, and 0.719 for 1-year, 3-year, and 5-year survival predictions, respectively. Similarly, the test set demonstrated promising AUC values of 0.799, 0.620, and 0.698 for the same time points (Fig. 2E, F). These results indicate that the model exhibits a relatively high degree of accuracy in predicting the prognosis of osteosarcoma patients, underscoring its potential clinical utility. Accuracy test of prognostic model. A, B Number of deaths and risk genes in high and low risk groups of patients in training set and test set; C, D Survival analysis of training set and test set revealed the survival rates of patients in high and low risk groups; E, F The ROC curve revealed the accuracy of the risk model in predicting the 1-, 3- and 5-year prognosis of osteosarcoma patients Construct nomogram model The outcomes of both univariate and multivariate Cox regression analyses conclusively demonstrated that the risk model served as a potent poor prognostic factor for osteosarcoma patients. To evaluate the independent prognostic value of the nucleotide metabolism-related gene risk score constructed in this study, we conducted a series of analyses. Firstly, we used the Cox proportional hazards regression model to assess the impact of the risk score and other key clinical pathological factors (including gender, race, metastasis status, and primary tumor site) on the overall survival of patients with osteosarcoma. The results of the univariate analysis (Fig. 3A) showed that the risk score was a significant risk factor for the prognosis of patients (HR = 3.2, 95% CI [1.8–5.7], p < 0.001). Subsequently, we incorporated all of these factors into a multivariate Cox regression model. The results (Fig. 3B) showed that even after adjusting for the metastatic status, gender, race, and tumor location, the risk score remained an independent and significant predictor of overall survival in osteosarcoma patients (HR = 3.1, 95% CI [1.7–5.6], p < 0.001). This result indicates that the prognostic information provided by this risk model is independent of the aforementioned basic clinical variables and has unique predictive value. Based on the above results, we integrated the variables with independent predictive significance from the multi-factor analysis—the risk score and the metastasis status—to construct a clinically practical prognostic nomogram (Nomogram) for the intuitive prediction of the overall survival rate of osteosarcoma patients at 1 year, 3 years, and 5 years (Fig. 3C). To evaluate the predictive efficacy of this nomogram, we plotted the time-dependent receiver operating characteristic curve. The results showed that this nomogram demonstrated good discriminatory ability in predicting the survival rates of patients at 1 year, 3 years, and 5 years, with the area under the curve (AUC) being 0.955, 0.770, and 0.776 respectively (Fig. 3D). Construct nomogram model. A Univariate Cox regression was used to test the independence of risk models; B multivariate Cox regression was used to test the independence of risk models; C a Nomogram model was constructed to predict the 1-, 3- and 5-year survival rates of osteosarcoma patients; D The accuracy of the nomogram model was tested by ROC curve In conclusion, the risk score constructed in this study is an independent adverse prognostic factor for osteosarcoma patients. The nomogram constructed based on the metastasis status shows high accuracy in predicting the short-term and long-term survival rates of patients, and is expected to provide a quantitative tool for individualized prognostic assessment of osteosarcoma patients. Enrichment analysis of risk model-related differential gene pathways By leveraging the transcriptome data from patients categorized into high-risk and low-risk groups based on the risk model, we conducted a thorough analysis that identified 74 differentially expressed genes between the two cohorts (Fig. 4A, B). This analysis provided valuable insights into the molecular underpinnings that distinguish these risk groups. The KEGG enrichment analysis revealed that the differentially expressed genes were significantly enriched in various pathways, including osteoclast differentiation, viral protein interaction with cytokines and cytokine receptors, the NOD-like receptor signaling pathway, neutrophil extracellular trap formation, the cytosolic DNA-sensing pathway, B cell receptor signaling pathway, cytokine-cytokine receptor interaction, among others, as depicted in Fig. 4C. The GO enrichment analysis yielded a total of 79 GO annotations, with the Sankey bubble plot highlighting the top 15 annotations sorted by their gen.ratio. These annotations encompass key biological processes such as the external side of the plasma membrane, cytokine-mediated signaling pathways, responses to lipopolysaccharide and bacterial-origin molecules, negative regulation of leukocyte proliferation, cellular responses to tumor necrosis factor, regulation of leukocyte proliferation, and leukocyte proliferation itself. Notably, the majority of these annotations are intimately tied to immune responses, as evident from their association with negative regulation of immune system processes (Fig. 4D). Enrichment analysis of differential gene pathways related to risk model. A, B Volcano map and heat map revealed risk model-related differentially expressed genes C KEGG enrichment analysis of differentially expressed genes; D GO enrichment analysis of differentially expressed genes Immune landscape analysis To explore the potential immunological mechanisms underlying the prognostic risk score, we analyzed the expression differences of immune checkpoint genes between the high-risk and low-risk groups. A total of 8 significantly differentially expressed immune checkpoint genes were identified (Fig. 5A). Among them, TIGIT, LAG3, CXCL10, etc. showed significantly upregulated expression in the high-risk group (p < 0.01). Further immune cell infiltration analysis (CIBERSORT) revealed that the risk score of the high-risk group was significantly positively correlated with the infiltration level of M2-type macrophages (r = 0.62, p < 0.001) (Fig. 5B, C), while it was weakly or insignificantly negatively correlated with activated CD4+ T cells, etc. These results collectively suggest that patients with high-risk scores of osteosarcoma may have a more significant immunosuppressive tumor microenvironment, characterized by upregulated inhibitory immune checkpoint molecules and enrichment of M2-type tumor-associated macrophages. Immune landscape analysis. A Risk model-based gene differential expression analysis of immune checkpoint; B correlation analysis between immune checkpoint genes and risk models; C correlation analysis of risk model with risk model genes and immune infiltrating cells Drug sensitivity analysis The IC50 values of GSK269962A, Ribociclib, AZD8055, PF-4708671, SB216763, and AZD6482 were all observed to be more sensitive in the high-risk group (Fig. 6), suggesting that these drugs hold greater potential as chemotherapeutic agents specifically tailored for the treatment of patients within this risk category. Sensitivity analysis of chemotherapy drugs. A–F GSK269962A, Ribociclib, AZD8055, PF-4708671, SB216763 and AZD6482 were more sensitive in high-risk osteosarcoma patients than in low-risk osteosarcoma patients Discussions The metastasis and recurrence of osteosarcoma are the main reasons for poor clinical prognosis. It is crucial to develop a predictive tool that can accurately assess the risk of metastasis and recurrence and is independent of traditional clinical factors. In this study, by integrating multiple machine learning algorithms, a prognostic risk model based on 7 nucleotide metabolism-related genes (MYC, MUC1, IMPDH1, SAMHD1, UCK2, NUDT13, NUDT16) was constructed and validated. This model not only demonstrated robust prognostic stratification ability in an independent cohort, but more importantly, multivariate Cox regression analysis confirmed its predictive value independent of available clinical information such as metastasis status, suggesting that it may capture the intrinsic biological processes driving the malignant progression of osteosarcoma, providing a new molecular dimension for prognosis assessment. This study is the first to incorporate genes such as IMPDH1, NUDT13, and NUDT16 into the prognosis model for osteosarcoma, expanding the understanding in this field. It is worth noting that these 7 genes do not act independently; they jointly outline an active nucleotide metabolism network. MYC, as a core oncogene, can globally drive the expression of genes related to nucleotide synthesis [13]. IMPDH1 and UCK2 are the key rate-limiting enzymes in the purine and pyrimidine salvage synthesis pathways, and their upregulation directly meets the huge demand of tumor cells for nucleotide precursors for rapid proliferatio [14, 15]. NUDT13 and NUDT16 act as hydrolases and may affect genomic stability and cellular stress responses by regulating the dynamic balance of intracellular nucleotide metabolism pools or signaling molecules (such as ADP-radiolabeled [16, 17]. MUC1 and SAMHD1 may indirectly support metabolic reprogramming by regulating intracellular signal transduction and dNTP homeostasis [18–20]. Although this study focused on the transcriptional level, the preliminary bioinformatics analysis suggested that the dysregulation of some genes (such as MYC) might be accompanied by DNA copy number variations. Future integration of multi-omics data will help clarify the genomic basis of their expression regulation. This gene characteristic centered on nucleotide metabolism explains the proliferation and survival advantages of high-risk tumors from the perspective of metabolic adaptation. Our analysis further reveals that high-risk scores are closely associated with an immunosuppressive tumor microenvironment. In terms of cell composition, the high-risk group significantly enriches M2-type tumor-associated macrophages, which are well-known as key mediators of immunosuppression and tumor progression [21]. In terms of molecular phenotypes, multiple inhibitory immune checkpoint molecules, such as TIGIT and LAG3, are synergistically upregulated in the high-risk group. These molecules directly weaken the anti-tumor immune response by inhibiting the functions of T cells and NK cells [22]. Additionally, the increase in the proportion of follicular helper T cells in the high-risk group has also been reported to be associated with poor prognosis of osteosarcoma, and the specific mechanism deserves in-depth exploration [23]. These findings consistently point to the possibility that high-risk osteosarcoma may achieve immune evasion by shaping an immunosuppressive ecosystem characterized by M2 macrophage infiltration, upregulation of inhibitory checkpoints, and changes in specific T cell subsets, which may be another key mechanism for its poorer prognosis. The above findings provide implicit clues for the treatment strategies for high-risk patients. Firstly, the drug sensitivity analysis indicates that high-risk patients may be more sensitive to targeted drugs such as Ribociclib (CDK4/6 inhibitor) and AZD8055 (mTOR inhibitor), which offers new ideas for overcoming the chemotherapy resistance of this type of patients. Secondly, the more significant implication lies in that the immunosuppressive microenvironment (high expression of TIGIT/LAG3, enrichment of M2 macrophages) suggested by the high-risk score strongly implies that these patients may be potential beneficiaries of immune checkpoint inhibitors (such as anti-TIGIT and anti-LAG-3 antibodies) or macrophage reprogramming therapies. Therefore, our model has the dual potential of prognostic prediction and treatment guidance. This study constructed a prognosis model for osteosarcoma from the unique perspective of “nucleotide metabolism”. Compared with other biomarker models reported in recent years, this model has its own characteristics in terms of perspective, performance and mechanism explanation. For instance, the 4-gene model constructed by Lv et al. based on aging-related genes demonstrated comparable predictive performance in the TARGET cohort (with an 5-year AUC of approximately 0.724), and explored its association with the immune cycle and single-cell phenotype in depth [24]. Yang et al. constructed a model with 6 molecular features based on copper death-related lncRNAs, achieving high predictive accuracy in the TCGA cohort (with an 5-year AUC of 0.812), and verified the expression of key gene [25]. In terms of the robustness of predictive performance, although direct comparison of AUC values is limited due to the different independent validation sets used in various studies (such as TARGET, TCGA, GEO subsets), this model still demonstrated robust discriminatory ability in a completely independent GSE39055 cohort (n = 42) (5-year AUC = 0.698), and the risk score was an independent prognostic factor in multivariate analysis, which collectively proved the reliability of its predictive efficacy. In terms of potential clinical applicability, the genes included in this model, such as IMPDH1 and NUDT13, were reported for the first time to be related to the prognosis of osteosarcoma, providing new candidate molecules. More importantly, this model contains only 7 genes, similar to the aforementioned models, and all belong to refined feature sets. This high simplicity is its potential advantage for clinical translation, facilitating the development of an economical and efficient detection tool. In conclusion, these models based on different biological pathways (metabolism, aging, cell death) complement each other and jointly enrich the molecular classification system of osteosarcoma. The nucleotide metabolism model in this study adds a novel and robust dimension to this system. In the future, through prospective, large-scale multicenter cohorts for “head-to-head” comparisons and integration, it will be the key to determining the optimal clinical translation path for each model. This study has several limitations, which also point out the direction for future work. Firstly, due to the characteristics of retrospective public data, this study has deficiencies in sample size, depth of clinical information, and multi-omics integration. The training set sample size is small, which may affect the statistical power assessment of some genes with weak effects (such as NT5C1B), although we have ensured the robustness of the final model genes through Bootstrap and Boruta algorithms. At the same time, the lack of detailed clinical covariates (such as AJCC stage, specific treatment plan) and paired genomic information (such as mutations, copy number) in the existing data limits our ability to more comprehensively evaluate the clinical applicability of the model and deeply analyze its molecular basis. Secondly, the conclusions of this study are mainly based on bioinformatics analysis. The specific functions of key genes (such as IMPDH1, NUDT13) in osteosarcoma, as well as the “metabolism-immune” regulatory axis suggested, need to be confirmed through subsequent in vitro and in vivo functional experiments. Therefore, validating the universality of this model in a larger-scale, prospective, multi-center cohort study, and conducting in-depth functional and mechanistic research, is an indispensable next step in translating this discovery into clinical application. In conclusion, we have successfully constructed and validated a bone cancer prognosis risk model based on genes related to nucleotide metabolism. This model is an independent and robust prognostic prediction tool. The underlying genetic characteristics are associated with the metabolic reprogramming of the tumor and the shaping of the immunosuppressive microenvironment, providing new ideas for identifying high-risk patients and revealing potential therapeutic targets (such as metabolic enzymes like IMPDH1, UCK2, etc., and immune checkpoints like TIGIT, LAG3, etc.), and has certain clinical translational potential. Author contributions Songli Ju: Writing—original draft. Tao Lu: writing—original draft. Xuyan Li: software. Nijiao Huang: methodology. Xixian Ke: writing—review & editing, project administration. Funding Not applicable. Data availability The datasets generated and/or analyzed during the current study are available from the corresponding author on reasonable request. Declarations Ethics approval and consent to participate Not applicable. Patient consent for publication Not applicable. Competing interests The authors declare no competing interests. Footnotes Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. Songli Ju and Lu Tao contributed equally to this work as first authors. Songli Ju and Xixian Ke contributed equally to this work. Contributor Information Songli Ju, Email: 9695170@qq.com. Xixian Ke, Email: kexixian@zmu.edu.cn, Email: 763289051@qq.com. References Associated Data Data Availability Statement The datasets generated and/or analyzed during the current study are available from the corresponding author on reasonable request.