Integrating machine learning and spatiotemporal transcriptomics to build a diagnostic model for osteosarcoma metastasis and to decipher the role of necroptosis genes

OBJECTIVE: To develop a diagnostic model for osteosarcoma metastasis and elucidate the spatiotemporal role of necroptosis genes. METHODS: Multi-omics data (TCGA/GEO, single‑cell, spatial transcriptomics) were integrated using 113 machine learning algorithms. Core genes were identified, followed by prognostic analysis, drug screening, molecular docking, and spatiotemporal mapping. Experimental validation included RT‑qPCR in four osteosarcoma cell lines and independent transcriptomic sequencing. RESULTS: The glmBoost+Ridge model selected seven core genes (e.g., FOS, MYC), achieving AUCs of 0.861 (training) and 0.873 (validation). A combined risk score predicted prognosis (AUC = 0.790). Pseudotime analysis showed MYC up‑regulation and TNFRSF21 loss during metastasis; spatially, MYC localized to invasive fronts while TNFRSF21‑deficient regions formed immune‑exempt zones. Experimental data confirmed MYC overexpression/TNFRSF21 underexpression in metastatic cells and revealed a strong negative correlation (r = -0.931). Calcitriol and Sulindac were identified as potential therapeutic candidates. CONCLUSION: This study provides a diagnostic model for osteosarcoma metastasis and proposes that MYC/TNFRSF21 drive metastasis via a "necroptosis‑immune exemption" axis, suggesting new therapeutic strategies.