Patient-balanced single-cell analysis resolves moderately recurrent and patient-restricted malignant programmes in osteosarcoma with matched spatial support

Cell-weighted analyses can allow high-yield specimens to dominate malignant-state discovery in osteosarcoma. We asked which malignant programmes recur across patients and which remain restricted to patient subsets. We integrated 276,787 single cells from 31 specimens across three cohorts and applied patient-balanced consensus non-negative matrix factorisation (NMF) in a 14-patient in-house cohort. We assessed recurrence by independent decomposition of two external osteosarcoma cohorts and examined biological context using patient-paired expression, public regulatory data, lung-metastasis paired RNA and chromatin-accessibility, three matched Visium sections and patient-resolved communication models. Seven malignant programmes were resolved. Patient balancing separated a moderately recurrent F3 ECM-mesenchymal/perivascular programme across external cohorts from an F4 chondro-hypoxia programme that was structurally stable but concentrated in four patients and strongly recovered in only one external cohort. Patient-paired analysis linked F4 to cartilage development, hypoxia and glycolysis, while SOX9 expression and public regulatory datasets provided cross-context support without proving direct regulation. F3 and F4 occupied coherent chromatin-defined neighbourhoods and patient-variable spatial domains. Patient-resolved communication models further prioritised an endothelial DLL4 / JAG2 -NOTCH context for F3. Patient-balanced analysis separated a pooled extracellular-matrix/chondro-matrix (ECM-CM) signal into a moderately recurrent F3 ECM-mesenchymal/perivascular programme across external cohorts and a patient-subset F4 chondro-hypoxia programme. The SOX9-associated and endothelial-NOTCH results are computational, hypothesis-generating findings rather than validated causal mechanisms or clinical biomarkers. Baseline programme activity nevertheless provides a rational basis for selecting models in which to test these focused hypotheses.