Osteosarcoma (OSA) is the most common primary malignant bone tumor and disproportionately strikes two age groups: adolescents during peak skeletal growth and adults over 60. Despite decades of refinement in surgical technique and the adoption of the MAP chemotherapy regimen (high-dose methotrexate, adriamycin, and cisplatin), the 5-year overall survival rate has remained stubbornly below 70% since the 1980s. The absence of meaningful survival gains over 40 years points to a fundamental problem: clinicians lack reliable tools to identify which patients are at high risk for chemoresistance and poor outcomes before treatment begins.
Current staging limitations: The Musculoskeletal Tumor Society (MSTS) staging system and the Huvos grading system (which assesses tumor necrosis rate in surgically resected specimens, with grades I/II indicating poor response and grades III/IV indicating good response) provide some prognostic guidance, but neither has enabled individualized treatment adjustments that actually improve survival. Attempts to adjust chemotherapy based on Huvos grading have consistently failed to boost outcomes, leading to its diminished clinical use. MSTS stage III and axial primary tumor location are recognized adverse pre-treatment indicators, yet the standard MAP regimen is applied uniformly regardless of these factors.
The multiomics opportunity: A prior study used multiomics data to cluster OSA patients into four subtypes, identifying the MYC-driven cluster, characterized by MYC amplification, activated mTOR signaling, and suppressed immune responses, as the most aggressive subtype with the worst prognosis. However, the high cost of collecting transcriptomic, genomic, and epigenomic data simultaneously limits the clinical scalability of full multiomics-based stratification. Emerging research suggested that simplified gene expression signatures could predict OSA outcomes cost-effectively, but most published signatures have been derived from small, single-cohort datasets with limited generalizability.
This study addresses these gaps by deploying a framework of 101 machine-learning algorithm combinations applied to whole-transcriptome profiles from 254 OSA biopsy samples. The goal was to develop a robust, reproducible prognostic index, termed the Artificial Intelligence-Derived Prognostic Index (AIDPI), that outperforms existing signatures, predicts chemotherapy response, and guides a precision treatment strategy for the high-risk subgroup it identifies.
The AIDPI was developed through a systematic, multi-stage computational workflow designed to maximize generalizability. The authors curated public datasets with two strict criteria: OSA diagnosis confirmed by histology and complete overall survival (OS) data with RNA sequencing or microarray detection performed on fresh-frozen biopsy samples. The primary training set was GEO-OSA, formed by merging GSE21257 (n = 53) and GSE16091 (n = 34) after applying the Rank-In algorithm to mitigate inter-dataset batch effects. The TARGET-OSA dataset (n = 85), sourced from the Therapeutically Applicable Research to Generate Effective Treatments (TARGET) program via the Genomic Data Commons portal, served as the independent validation set. GSE33382 (n = 82) provided a second-tier independent test set. An additional meta-OSA cohort was constructed by integrating GSE21257, GSE16091, GSE33382, and TARGET-OSA for 254 total samples used in combined analyses.
Algorithm framework: Ten classical machine-learning algorithms were used: LASSO (least absolute shrinkage and selection operator), gradient boosting machine (GBM), random survival forest (RSF), partial least squares regression for Cox (plsRcox), stepwise Cox (StepCox, both forward-backward and backward-only directions), supervised principal components (SuperPC), ridge regression, survival support vector machine (Survival-SVM), CoxBoost, and elastic network (Enet). The first step used RSF, LASSO, CoxBoost, and two StepCox variants for dimensionality reduction and variable screening. These were then combined with the remaining algorithms, producing 101 unique pairings in total.
Model selection process: First, univariate Cox regression was applied to both the training set and validation set independently. Genes with consistent prognostic hazard ratios (p less than 0.01 in both cohorts, with HR consistently above 1 or below 1) were designated Consistent Prognostic Genes (CPGs), yielding 18 candidates. Each of the 101 algorithm combinations then selected genes from these CPGs and fitted a prognostic model in the training set. Harrell's concordance index (C-index) was calculated for each model across the training set, validation set, and independent test set. The model with the highest average C-index across all three cohorts was automatically selected as optimal. The final risk score calculated from this model is the AIDPI.
Validation architecture: The AIDPI's predictive performance was assessed via time-dependent receiver operating characteristic (tROC) analysis using the timeROC package for 1-, 3-, and 5-year OS. An optimal AIDPI threshold was determined in the training set using the surv_cutpoint function, and patients in each subsequent cohort were dichotomized into low- and high-AIDPI groups. Kaplan-Meier survival analysis confirmed group separation. Predictive performance was then compared against 53 previously published OSA prognostic signatures using the compareC package to test statistical significance of C-index differences. A Shiny application was deployed online to allow real-time AIDPI computation, and a dynamic interactive nomogram was published using the DynNom package.
The combination of CoxBoost and GBM emerged as the optimal model with an average C-index of 0.817 across three cohorts, the highest of all 101 combinations. This combination addressed a key technical challenge: GBM alone produced a C-index of 0.941 in the training set but dropped to 0.775 in the validation set and 0.695 in the independent test set, suggesting significant overfitting. CoxBoost alone achieved a more modest average C-index of 0.768. By using CoxBoost first for dimensionality reduction (selecting 12 prognostic genes from the original 18 CPGs) and then applying GBM to fit the final model, the combined approach achieved both accuracy and generalizability.
Time-dependent AUC values: In the training set (GEO-OSA), tROC analysis yielded AUCs of 0.981, 0.995, and 0.988 for 1-, 3-, and 5-year OS, respectively. In the validation set (TARGET-OSA), AUCs were 0.817, 0.772, and 0.776. For the independent test set (GSE33382-OSA), AUCs were 0.886, 0.767, and 0.849. Kaplan-Meier analyses consistently showed worse outcomes for high-AIDPI patients in the training set (p less than 0.0001), validation set (p less than 0.0001), and independent test set (p = 0.0025), as well as across all additional cohorts including the combined meta-OSA cohort.
Comparative benchmarking: Of 68 previously published OSA prognostic signatures, only 53 could be tested due to gene naming changes and gene absences in microarray datasets. A heatmap displayed C-index results for all 53 signatures and the AIDPI across six cohorts. Only two signatures consistently achieved statistical significance across all cohorts: the AIDPI and the signature published by Xu et al. Direct C-index comparisons demonstrated AIDPI's statistically significant superiority over Xu et al.'s signature in three cohorts (Z test, p less than 0.05 and p less than 0.0001 in multiple pairings).
Nomogram for clinical integration: Multivariate Cox regression in the meta-OSA cohort identified AIDPI, MSTS stage, Huvos grade, and primary tumor site as independent prognostic factors. Because Huvos grade had over 25% missing values, it was excluded from the final nomogram, which incorporated AIDPI, age, MSTS stage, and primary tumor site. Calibration curves confirmed robust agreement between predicted and observed survival probabilities. The nomogram achieved AUCs of 0.938, 0.903, and 0.904 for 1-, 3-, and 5-year OS, surpassing all individual clinical parameters by AUC comparison. Decision curve analysis confirmed broader net clinical benefit compared to isolated clinical factors. An interactive online nomogram application was deployed for real-time survival estimation.
To understand why high-AIDPI patients do poorly, the study systematically interrogated transcriptomic, genomic, and epigenomic data from the TARGET-OSA cohort. A heatmap of the 12 AIDPI genes revealed that 7 were markedly upregulated in high-AIDPI tumors, with 5 of these inversely correlated with immune score, a composite metric of immune cell infiltration. The remaining 5 AIDPI genes were downregulated, and 3 showed positive correlation with immune score. This pattern already suggests that high-AIDPI tumors simultaneously activate oncogenic programs and suppress anti-tumor immunity.
Gene Set Enrichment Analysis (GSEA): The most significantly enriched gene sets in high-AIDPI tumors were MYC targets V2, MYC targets V1, cholesterol homeostasis, and mTORC1 signaling. In contrast, gene sets for apoptosis and specific immune responses were negatively enriched, meaning they were suppressed in high-AIDPI tumors relative to low-AIDPI tumors. KEGG enrichment analysis of differentially expressed genes (DEGs) highlighted PI3K-Akt signaling, cytokine-cytokine receptor interaction, osteoclast differentiation, focal adhesion, and extracellular matrix (ECM)-receptor interaction as significantly dysregulated pathways, with an enrichment map showing tight interconnection among the PI3K-Akt, focal adhesion, and ECM-receptor networks.
Genomic and epigenomic origins of dysregulation: Notably, the AIDPI genes themselves lacked recurrent somatic mutations, and tumor mutation burden did not differ significantly between low- and high-AIDPI groups. Instead, copy number analysis identified increased copy numbers for MCAM, MYC, and SQLE in the high-AIDPI group, suggesting that genomic amplification rather than point mutation drives their overexpression. DNA methylome analysis did not reveal significant shifts in average methylation levels between the two groups, but empirical Bayes GSEA based on epigenomic data identified focal adhesion as the most significantly dysregulated pathway, mirroring the transcriptomic findings.
Immune cell infiltration: Methylome-based cell fraction estimation using EpiDISH revealed reduced CD4+ T cells, monocytes, and neutrophils in the high-AIDPI group, alongside an increase in fibroblasts. These findings collectively suggest that the poor prognosis of high-AIDPI OSA stems from a convergence of MYC-driven oncogenic signaling, cholesterol and mTOR pathway activation, and an immunosuppressive, fibroblast-enriched tumor microenvironment, all rooted in DNA copy number amplifications at the genomic level.
To determine which cell types within the OSA tumor microenvironment are responsible for the transcriptomic differences between low- and high-AIDPI patients, and to narrow the search for druggable therapeutic targets to OSA tumor cells specifically, the study analyzed a scRNA-seq dataset from six OSA biopsy samples (BioProject PRJNA681896). Cells were analyzed using Seurat, with quality filtering removing cells with fewer than 300 expressed genes or over 10% mitochondrial gene content. Doublets were removed using DoubletFinder, batch effects were mitigated using Harmony, and cell types were annotated using the scGate and infercna packages.
Cell type annotation: Nine primary cell clusters were identified: OSA cells, B cells, endothelial cells, myeloid cells, NK cells, osteoclasts, plasma cells, non-tumor stromal cells, and T cells. OSA cells were distinguished from normal cells using inferred copy number alteration (CNA) signals. Predicted OSA cells showed distinct chromosomal amplifications and deletions absent from reference and normal cell populations. Marker gene expression was confirmed using ACP5 (osteoclasts), VWF (endothelial cells), and COL1A1 (stromal cells).
Target identification pipeline: Positively expressed genes (PEGs) were mapped for each of the nine cell clusters. Comparing the DEGs between low- and high-AIDPI groups (from bulk RNA-seq) against the PEGs of each cluster showed that only 8% of DEGs were predominantly expressed in OSA cells rather than in stromal or immune cells. Cross-referencing the 12 AIDPI genes with both the bulk RNA-seq DEGs and the OSA cell-specific PEGs revealed three shared genes. When the proteins encoded by these three genes were queried against the canSAR database (a drug discovery platform), only MYC and SQLE possessed druggable protein structures. MYC is nuclear and has no well-defined protein pockets, making pharmacological inhibition challenging. SQLE, a cytosolic metabolic enzyme, is directly accessible to small-molecule inhibitors.
SQLE overexpression in OSA: SQLE mRNA was markedly elevated in OSA tissues compared to adjacent normal tissues, and in OSA cell lines compared to osteoblasts (OB) and mesenchymal stem cells (MSC), the putative progenitors of OSA. Critically, tumors with Huvos grade I/II (poor chemotherapy response) had higher SQLE expression than Huvos grade III/IV tumors. Methotrexate-resistant Saos2 cells (Saos2/MTX) similarly expressed more SQLE than their parental cell line, establishing a direct mechanistic link between SQLE and chemoresistance. A tissue microarray (TMA) of 77 human OSA samples confirmed that high SQLE protein expression was significantly associated with worse OS (Kaplan-Meier analysis, p = 0.037). Multivariate Cox regression in meta-OSA confirmed SQLE mRNA as an independent mortality risk factor even after adjustment for age, MSTS stage, and primary tumor site.
To functionally validate SQLE as a therapeutic target, shRNA constructs targeting SQLE mRNA (shSQLE) were designed, screened, and the optimal construct was used to generate stable knockdown cell lines in MNNG/HOS and U2OS cells, two OSA lines chosen for their high endogenous SQLE protein expression. Scramble sequences (shControl) served as the experimental control. Knockdown efficiency was verified by both qRT-PCR and Western blotting.
In vitro phenotypic effects: SQLE knockdown inhibited all major malignant phenotypes tested. CCK-8 and clonogenic assays showed significantly reduced proliferation and colony formation. Transwell migration and Matrigel invasion assays confirmed impaired motility. Annexin V-APC/7-AAD flow cytometry demonstrated increased apoptosis both in the absence of drug treatment and after adriamycin (ADM) exposure. Critically, shSQLE cells showed enhanced sensitivity to all three MAP regimen agents (methotrexate, adriamycin, cisplatin), as measured by decreased IC50 values in the knockdown group.
In vivo xenograft results: Nude mice (n = 5 per group) were injected subcutaneously with U2OS cells expressing shSQLE or shControl. Tumor volume was monitored every 3 days for 26 days. SQLE silencing significantly inhibited tumor growth, reduced tumor weight, and lowered tumor cholesterol content. IHC staining confirmed reduced SQLE protein and fewer Ki-67-positive proliferating cells in tumors derived from shSQLE cells, validating the anti-tumor effect in a living system.
Mechanistic pathway analysis: RNA-seq on shSQLE versus shControl U2OS cells showed GSEA enrichment of cholesterol homeostasis downregulation as the top biological process affected. Intracellular cholesterol levels were indeed significantly reduced in both MNNG and U2OS cells after SQLE knockdown. KEGG enrichment of DEGs identified PI3K-Akt signaling as the most significantly affected pathway, with the highest gene ratio. Pharmacogenomic analysis using GDSC datasets revealed positive correlations between SQLE mRNA and sensitivity to PI3K/mTOR antagonists, particularly GSK1059615. Western blotting confirmed that SQLE silencing reduced phosphorylation of focal adhesion kinase (FAK), PI3K, Akt, and mTOR. Restoring cholesterol by supplementing cells with 100 micrograms per milliliter rescued all these phosphorylation patterns and counteracted the suppressive effects of SQLE knockdown on proliferation, colony formation, migration, and invasion, confirming that cholesterol reduction is the upstream driver of FAK/PI3K/Akt/mTOR inhibition.
Having established SQLE as a functional target via genetic silencing, the study assessed pharmacological SQLE inhibition as a translatable approach. Fungal SQLE inhibitors such as naftifine and terbinafine have shown anti-tumor activity in various cancer lines without affecting healthy fibroblasts. However, both were designed for fungal SQLE and may have limited efficacy and bioavailability against the human ortholog. FR194738 is a potent, selective mammalian SQLE inhibitor with demonstrated preclinical efficacy in prostate cancer. Sensitivity testing in MNNG and U2OS cells showed markedly lower IC50 and GR50 values for FR194738 compared to naftifine and terbinafine, establishing it as the superior pharmacological agent for targeting human OSA cells.
Cell-line selectivity: Four OSA cell lines were tested at 72 hours of FR194738 exposure at 4 micromolar concentration. MNNG and U2OS cells showed significant decreases in growth rate and cell viability. MG63 and 143B cells did not respond appreciably, correlating with their lower SQLE protein expression detected by Western blotting. This selectivity indicates that FR194738's anti-OSA activity is SQLE-dependent and suggests it would spare cells with low SQLE expression, a property relevant to safety in normal tissues. FR194738 treatment reduced intracellular cholesterol levels in responsive cell lines, consistent with its mechanism of action as an SQLE inhibitor.
In vivo efficacy: MNNG cells in Matrigel were implanted subcutaneously into nude mice (5 cells per mouse, n = 5 per group). After tumors reached approximately 60 mm3 (day 9), mice received either daily intraperitoneal injection of 10% DMSO (vehicle) or 100 mg/kg FR194738. Tumors were analyzed on day 16 post-injection. FR194738 significantly stunted tumor growth and reduced tumor cholesterol content without inducing weight loss, a key safety indicator. Average tumor weight comparisons yielded a tumor growth inhibition (TGI) rate of 59.67%. IHC analyses confirmed reduced Ki-67 staining (proliferation marker) and increased cleaved caspase-3 (apoptosis marker) in FR194738-treated tumors.
Synergy with MAP agents: Given that SQLE silencing suppressed FAK/PI3K/Akt/mTOR signaling, and that this pathway's inhibition increases OSA cell sensitivity to MTX, ADM, and DDP, the study tested whether FR194738 would synergize with first-line chemotherapy. Drug combination experiments in MNNG and U2OS cells with various pairings were analyzed using four synergy models: Highest Single Agent, Loewe Additivity, Bliss Independence, and Zero Interaction Potency. Most synergy scores indicated potent synergistic effects. Combination sensitivity scores (CSS) and synergy scores for all tested combinations identified the FR194738 plus DDP (cisplatin) pairing as the best combination across both cell lines, consistent with a prior report linking SQLE overexpression to DDP resistance in head-and-neck squamous cell carcinoma. These results suggest FR194738 has the potential to overcome DDP resistance specifically in high-AIDPI, high-SQLE OSA patients.
The study makes a strong computational and experimental case for AIDPI-guided stratification and SQLE-targeted therapy in high-risk OSA, but the authors identify several critical limitations that constrain the translational timeline. The first and most important is retrospective design. All patient cohorts used for AIDPI development and validation are retrospective, meaning that patients were not prospectively selected, randomized, or treated based on AIDPI scores. Prospective clinical trials are needed to confirm that AIDPI-guided treatment allocation actually improves survival outcomes compared to standard uniform therapy.
Biological unknowns regarding SQLE inhibition: The study conclusively demonstrates that SQLE silencing reduces intracellular cholesterol and deactivates FAK/PI3K/Akt/mTOR signaling. However, SQLE inhibition also causes squalene accumulation upstream (since SQLE oxidizes squalene to 2,3-epoxy squalene in the cholesterol synthesis pathway). In neuroendocrine tumors, squalene accumulation has been shown to have independent anti-tumor effects. Whether squalene accumulation contributes to the anti-OSA activity of FR194738, alongside or separately from cholesterol depletion, remains unaddressed in the current study and represents a mechanistic question with potential implications for drug combination strategies.
Immune microenvironment and checkpoint blockade synergy: The study identified an inverse correlation between SQLE mRNA expression and CD4+ T cell infiltration in the TARGET-OSA dataset, suggesting that SQLE overexpression actively suppresses immune infiltration. A separate study in glioblastoma has shown that SQLE inhibition combined with immune checkpoint blockades achieves improved outcomes. The authors acknowledge that exploring FR194738's effects on OSA immune responses, and its potential synergy with PD-1/PD-L1 blockade in the OSA tumor microenvironment, is a necessary next step that the current work does not cover.
Preclinical model gaps and proposed clinical pathway: All in vivo experiments used xenograft models, which are immunocompromised nude mice and therefore cannot model immune-mediated effects. Rigorous evaluation in OSA patient-derived xenograft (PDX) models, which preserve patient tumor heterogeneity, and transgenic mouse models with intact immune systems would strengthen the translational case for FR194738. For clinical translation, the proposed workflow involves collecting fresh-frozen biopsy RNA-seq from newly diagnosed OSA patients, harmonizing transcriptomes with the meta-OSA cohort using the published pipeline, and computing the AIDPI via the online Shiny application. Low-AIDPI patients would proceed with standard MAP chemotherapy, while high-AIDPI patients would be candidates for MAP plus SQLE inhibitor combination therapy, potentially specifically with FR194738 plus cisplatin given the observed synergy data.