Construction and clinical validation of a machine-learning-based consensus prognostic signature (MLPS) for...

Osteosarcoma represents the most prevalent primary malignant bone tumor in adolescents and young adults, with limited improvement in survival outcomes observed over recent decades. The development of personalized therapies remains challenging due to substantial molecular heterogeneity and the absence of robust prognostic models. In this study, we developed a machine learning-based consensus prognostic signature (MLPS) by integrating multi-cohort transcriptomic datasets and applying ten distinct machine learning algorithms. The resulting MLPS, comprising 11 prognostic genes, reliably stratified patients into high- and low-risk groups with significantly different survival outcomes and outperformed existing models (C-index = 0.862). The application of functional enrichment analyses revealed immune activation and inflammatory signaling in the low-risk group, consistent with a "hot tumor" phenotype. In contrast, the high-risk group exhibited upregulation of proliferative oncogenic pathways. Of particular significance is the observation that MLPS was able to suggest differential responses to chemotherapy and immunotherapy, thereby under-scoring its clinical utility. Single-cell analysis revealed that LGR4 expression is largely confined to malignant cell clusters. In vitro LGR4 knockdown significantly reduced osteosarcoma cell proliferation, migration, and PI3K-AKT-mTOR pathway activity, confirming its oncogenic role. These findings demonstrate the value of MLPS as a robust tool for prognostic assessment and individualized therapeutic decision-making in osteosarcoma.