Osteosarcoma (OS) is the most common primary malignancy originating from bone and soft tissues in children and adolescents, and it ranks third in frequency among adults. Incidence varies by ethnicity and sex, with higher rates in African Americans compared to other groups. Despite advances in surgical resection and multi-agent chemotherapy, roughly 30% of patients present with metastatic disease at diagnosis, and overall 5-year survival for metastatic OS remains below 30%. Identifying robust molecular prognostic biomarkers that can stratify patients into meaningful risk groups at diagnosis is therefore a high clinical priority.
The apoptosis angle: Apoptosis, the process of programmed cell death triggered by DNA damage and immune activation, is frequently dysregulated in cancer cells. Alterations in apoptotic signaling allow tumor cells to evade cell death and develop resistance to cytotoxic chemotherapy. Osteosarcoma is no exception: aberrant apoptotic pathways have been documented to accelerate disease progression and contribute to poor treatment response. The research community has increasingly focused on apoptosis-related genes as candidate prognostic markers because their expression can reflect fundamental tumor biology.
Study rationale: This 2024 study published in Scientific Reports set out to build a prognostic model for OS by applying multiple machine learning algorithms to identify which apoptosis-related differentially expressed genes carry the strongest prognostic signal. The authors then narrowed this list to a two-gene signature (SERPINE2 and CPT1B), validated it using an independent dataset and immunohistochemistry, and explored how these genes interact with immune cell populations in the tumor microenvironment. A secondary analysis of routine blood count data from over 20,000 individuals was used to corroborate the immune findings at the population level.
The study spans bioinformatics, machine learning, tissue-level protein validation, and clinical blood data, making it an unusually multi-layered approach for a prognostic biomarker study in a rare bone cancer.
The study assembled gene expression data from three separate sources. Primary OS expression profiles for 88 osteosarcoma cases were downloaded from the UCSC Xena platform, which aggregates genomic and clinical data from public repositories. Normal control expression profiles for 396 skeletal muscle tissue samples were drawn from the Genotype-Tissue Expression (GTEx) project. Raw data from both sources were standardized using normalization and log2 transformation in R to reduce cross-sample variability before any downstream analysis. A list of 4,675 apoptosis-related genes was compiled from the Gene Set Enrichment Analysis (GSEA) database to focus the investigation.
Validation dataset: An independent validation cohort was sourced from the Gene Expression Omnibus (GEO) under accession GSE21257, comprising gene expression profiles from 53 OS samples. This external dataset was held out for confirming that the prognostic model generalizes beyond the training data, providing a meaningful check against overfitting.
Differential expression and AI screening: A differential expression analysis spanning 54,751 genes identified 1,197 significantly dysregulated genes between OS tissue and normal skeletal muscle, visualized using volcano plots and heatmaps. The overlap between these differentially expressed genes and the 4,675 apoptosis-related genes yielded 278 candidate genes for downstream analysis. These 278 candidates were then submitted to four AI-based screening methods: Support Vector Machine (SVM), Random Forest (RF), Generalized Linear Model (GLM), and Extreme Gradient Boosting (XGBoost). Each method independently ranked genes by their discriminative importance, and the intersection of genes flagged by all four methods formed the final candidate set for prognostic model development.
Prognostic model construction: From the AI-screened candidates, univariate Cox regression identified 34 genes significantly associated with OS prognosis (p-value less than 0.05). LASSO (Least Absolute Shrinkage and Selection Operator) regression was applied to this 34-gene list to shrink the model and select 20 genes with the strongest independent prognostic weight. A final multivariate Cox regression then distilled the model to two genes, SERPINE2 and CPT1B, as the minimal sufficient signature. Survival curves were generated using Kaplan-Meier analysis with patients split into high-risk and low-risk groups by median gene expression.
Before building the prognostic model, the authors performed Gene Ontology (GO) and Kyoto Encyclopedia of Genes and Genomes (KEGG) enrichment analyses on the 278 apoptosis-related differentially expressed genes. These tools map gene lists onto known biological processes and signaling pathways, revealing which cellular functions are most disrupted in OS compared to normal tissue.
GO enrichment findings: The GO analysis highlighted enrichment in processes central to tumor biology. Collagen fibril organization and extracellular matrix (ECM) assembly emerged as top biological processes, consistent with the heavily stromal biology of bone tumors where remodeling of the ECM scaffold plays a key role in invasion and metastasis. Additionally, nuclear and organelle processes related to cell cycle control and apoptotic signaling were enriched, suggesting that the dysregulated genes participate in multiple hallmarks of cancer simultaneously rather than just one pathway.
KEGG pathway enrichment: KEGG analysis revealed significant enrichment in pathways with well-established roles in cancer biology. ECM-receptor interaction, cell cycle regulation, and apoptosis pathways were all highlighted. The identification of cell cycle-related pathways alongside apoptosis pathways is mechanistically logical because these two processes are tightly coupled: cells with cell cycle dysregulation frequently also have impaired apoptotic checkpoints, allowing damaged cells to proliferate rather than self-destruct. The co-enrichment of these pathways reinforces the biological plausibility of the candidate gene list as a reflection of genuine OS tumor biology rather than statistical noise.
These pathway findings provided the biological framing for why apoptosis-related genes are clinically relevant in OS: the genes that are most dysregulated in OS tissue cluster around processes that directly control survival, proliferation, and invasion. Identifying which specific genes within these pathways carry independent prognostic weight is the mission the AI screening step was designed to accomplish.
The multivariate Cox regression analysis converged on two genes as the final prognostic signature. Multivariate hazard ratios for both genes were statistically significant (p-values of 0.000431 for SERPINE2 and 0.005952 for CPT1B). SERPINE2 carried a hazard ratio of 1.60 (95% CI: 1.23 to 2.07), indicating that higher SERPINE2 expression is associated with approximately 60% increased hazard of death in OS patients. CPT1B showed a far larger hazard ratio of 21.73 (95% CI: 2.42 to 194.94), though the wide confidence interval reflects the small sample size and warrants caution in interpreting the point estimate.
SERPINE2 biology: SERPINE2, encoding Serpin Family E Member 2, is a serine protease inhibitor. Prior literature has linked elevated SERPINE2 expression to poor overall survival in lung adenocarcinoma and to elevated serum levels in papillary thyroid cancer. This study adds OS to that list. SERPINE2 is upregulated in OS tissue relative to adjacent non-tumorous tissue, as confirmed by immunohistochemistry (positive staining area in OS: 72%, versus 22% in paracancerous tissue). Its role in protease inhibition may influence ECM remodeling and tumor invasion, connecting it to the pathway enrichment results.
CPT1B biology: CPT1B (Carnitine Palmitoyltransferase 1B) encodes the rate-limiting enzyme for mitochondrial fatty acid beta-oxidation. In OS, CPT1B is downregulated in tumor tissue: the IHC positive area was 31% in OS versus 68% in surrounding non-tumorous tissue. This inverse expression pattern, where lower CPT1B in tumor tissue correlates with worse prognosis, is consistent with findings in bladder cancer and castration-resistant prostate cancer where CPT1B has been identified as a metabolic tumor suppressor. Reduced fatty acid oxidation capacity may force tumor cells toward alternative metabolic programs that support more aggressive growth.
Survival analysis: Kaplan-Meier curves for ten apoptosis-related genes showed significant separation between high- and low-expression groups, with the integrated two-gene prognostic model achieving a significant distinction in survival outcomes between high-risk and low-risk groups (p less than 0.05). Validation in the independent GSE21257 cohort of 53 samples confirmed the model's discriminative ability, providing a critical check against dataset-specific overfitting.
After constructing the two-gene prognostic model, the authors submitted it to three complementary forms of reliability testing: ROC curve analysis, risk group comparison, and calibration plot assessment. Each approach captures a different dimension of model performance and together they provide a more complete picture of clinical utility than any single metric alone.
ROC curve performance: Receiver operating characteristic (ROC) curves were constructed for 1-year, 3-year, and 5-year survival prediction. The area under the curve (AUC) consistently exceeded 0.65 across all time points in both the training dataset and the GSE21257 validation set. While AUC values above 0.65 are modest compared to the near-perfect AUC values sometimes reported for larger, better-powered studies, they represent meaningful improvement over random chance (AUC 0.50) and are noteworthy given the small sample sizes (88 training cases, 53 validation cases) typical for rare pediatric bone cancers.
Risk group stratification: When patients were divided into high-risk and low-risk groups based on the model's risk score, the high-risk group showed statistically significantly higher mean expression of SERPINE2 and lower mean expression of CPT1B relative to the low-risk group (p less than 0.05 for both). This internal consistency confirms that the model's risk scores align with the individual gene expression patterns in a biologically coherent direction rather than simply reflecting statistical artifact.
Calibration plot: A calibration curve, which plots predicted survival probabilities against observed outcomes, showed that the model slightly overestimated survival probability at early time points but converged with observed values by the endpoint. This pattern is common in prognostic models and indicates that the model provides reasonable absolute risk estimates rather than only relative risk ranking. The calibration check is a more stringent test of a model's clinical utility than AUC alone because it assesses whether the predicted probabilities can be trusted as estimates of actual patient risk.
A key strength of this study is its investigation of how the two prognostic genes relate to immune cell populations in the tumor microenvironment (TME). The TME has become a central focus in oncology because the composition and functional state of immune cells infiltrating a tumor influence both natural disease progression and responsiveness to immunotherapy. The authors used CIBERSORT, a deconvolution algorithm that uses gene expression matrices and linear support vector regression to infer the proportions of 22 immune cell types from bulk RNA data, to characterize immune infiltration across the OS samples.
SERPINE2 and immune correlations: Analysis revealed a significant positive correlation between SERPINE2 expression and the presence of memory B cells (R=0.22, p less than 0.05). Memory B cells are a component of adaptive immunity that can contribute to anti-tumor immune responses, particularly in cancers where tertiary lymphoid structures form. A positive correlation between a tumor-promoting gene (high SERPINE2 associates with worse prognosis) and memory B cells is counterintuitive and may reflect a complex, context-dependent immune relationship rather than a simple causal link. The authors interpret this as a potential immunological interaction worthy of further mechanistic investigation.
CPT1B and CD8+ T cell correlation: CPT1B expression showed a substantial positive correlation with CD8+ T cells (R=0.3, p less than 0.001). CD8+ cytotoxic T cells are the primary effectors of anti-tumor immunity and are central to the efficacy of checkpoint inhibitor immunotherapy. The finding that lower CPT1B expression (which characterizes the high-risk group) correlates with lower CD8+ T cell infiltration suggests that CPT1B downregulation in OS may be accompanied by immune evasion, and this combination could explain the poor prognosis associated with the high-risk group. This intersection of metabolic reprogramming and immune exclusion is an increasingly recognized axis in aggressive tumors.
These immune correlations connect the molecular biomarker findings to the broader context of cancer immunology and suggest that targeting the SERPINE2-memory B cell or CPT1B-CD8+ T cell axes could be relevant for future immunotherapy strategies in osteosarcoma, though the small sample size means these associations must be interpreted as hypothesis-generating rather than definitive.
The bioinformatic findings were taken to the bench with two forms of independent clinical validation. The first used immunohistochemistry (IHC) on surgical pathology specimens. Tissue samples were obtained during OS surgeries at the First Clinical Affiliated Hospital of Guangxi Medical University. Formalin-fixed, paraffin-embedded sections were stained using antibodies specific to SERPINE2 (Proteintech catalog number 66203-1-Ig) and CPT1B (ABclonal item number A6796), with antigen retrieval and blocking performed according to standard protocols. Slides were imaged using an inverted microscope and independently evaluated by pathologists blinded to the molecular data.
IHC results for SERPINE2: SERPINE2 protein showed markedly elevated expression within OS tissue compared to adjacent non-cancerous tissue. The positive staining area was 72% in OS versus 22% in paracancerous tissue, a roughly 3.3-fold enrichment at the protein level that directly corroborates the upregulated mRNA expression detected in the bioinformatic analysis. This concordance between RNA and protein expression strengthens confidence that SERPINE2 overexpression in OS is a genuine biological feature rather than a platform-specific artifact.
IHC results for CPT1B: CPT1B showed the opposite pattern, with a positive staining area of 68% in surrounding non-tumorous tissue versus only 31% in OS tissue. The down-regulation of CPT1B in the tumor itself again matches the bioinformatic data, and is consistent with published findings in other cancer types where reduced fatty acid oxidation enzyme expression correlates with aggressive phenotype. H&E staining further confirmed the tissue architecture differences, with OS cell nuclei appearing tightly packed and darkly stained in contrast to the more dispersed, lightly stained nuclei in control tissue.
Routine blood count analysis: A large-scale analysis of routine blood parameters compared 20,679 individuals without OS (data from January 2012 to January 2022 at the First Affiliated Hospital of Guangxi Medical University) with 437 individuals diagnosed with OS. Statistical examination of these datasets revealed a significant disparity in lymphocyte counts and percentages, with healthy controls exhibiting higher lymphocyte values than OS patients (p less than 0.001). This real-world clinical blood data corroborates the CIBERSORT prediction of reduced immune cell representation in OS patients and adds a readily measurable clinical variable to the prognostic picture.
The authors explicitly acknowledge several limitations that should frame how the findings are interpreted and what studies are needed before this signature could influence clinical practice. The most fundamental constraint is sample size: the primary training dataset contains only 88 OS cases, and the validation set contains 53 cases. Osteosarcoma is a relatively rare cancer, which makes building large datasets challenging, but sample sizes in this range limit statistical power, increase the risk of inflated hazard ratio estimates (the wide confidence interval for CPT1B's HR of 21.73 is a clear indicator of this), and reduce the reliability of multivariate models that contain multiple covariates.
Suboptimal clinical data utilization: The study did not integrate detailed clinical variables such as tumor grade, site, response to neoadjuvant chemotherapy (which is the strongest established prognostic factor in OS), surgical margins, or presence of skip lesions. In OS, histological response to preoperative chemotherapy as assessed by necrosis percentage is the gold-standard prognostic indicator, and a prognostic model that does not account for this variable cannot be directly compared against clinical practice benchmarks. Future work should incorporate these clinical variables to determine whether the molecular signature provides independent prognostic value beyond what established clinical factors already capture.
Insufficient laboratory validation: While IHC was performed on a small surgical cohort, functional studies demonstrating the mechanistic roles of SERPINE2 and CPT1B in OS biology are still needed. Cell line and animal model experiments that knock down or overexpress these genes would strengthen the causal claims and potentially reveal therapeutic intervention points. The immune cell correlations from CIBERSORT, while interesting, are inferred from bulk RNA data and would benefit from orthogonal validation using multiplex immunofluorescence or single-cell RNA sequencing on tumor tissue.
Prospective validation needed: All datasets used in this study are retrospective. Prospective cohort studies and, ultimately, randomized trials incorporating biomarker-based risk stratification are required to demonstrate that using SERPINE2 and CPT1B expression levels to guide treatment decisions leads to improved patient outcomes. The correlation of this signature with standard-of-care imaging and pathological response measures should also be evaluated to understand its potential niche in the clinical pathway.