Fusion evidence supports computational attribution of neoplastic and microenvironmental states in tenosynovial giant cell tumor

Tenosynovial giant cell tumor (TGCT) contains rare neoplastic cells with colony-stimulating factor 1 (CSF1) rearrangements within a macrophage-rich lesion, complicating attribution of overlapping tumor and microenvironmental expression programs. We reconstructed patient-specific fusion junctions from public long-read RNA sequences from three TGCT patients, requiring concordant short-read alignment, exact junction k-mers, barcodes, and unique molecular identifiers. Fusion-supported, state-concordant cells provided operational training anchors. Analyses included leave-one-patient-out (LOPO) program transfer, external cohort projection, and a conditional linear profile-mixture simulation. A separate 19-library cohort underwent caller-guided, source-restricted sequence auditing. Across three biologically independent patients, 590 cross-platform consensus junction events identified 165 fusion-supported cells, including 111 state-concordant training anchors. Descriptive LOPO areas under the curve were 0.683, 0.803, and 0.939, not independent validation estimates. The 30-gene program had higher point estimates in fibroblast lineages but lower estimates in whole-synovium cohorts; none of five projection contrasts remained significant after joint correction. The CSF1 receptor (CSF1R) myeloid core was directionally elevated in all three external cohorts, with Benjamini-Hochberg-significant support only in GSE3698. The matrix axis remained provisional. Joint non-negative least-squares estimation improved recovery under the simulation assumptions. Fusion evidence supports computational cell-state attribution in TGCT, but biological replication remains limited to three patients. The program is not established as a whole-tissue biomarker, and the simulation does not establish clinical deconvolution. Independent tissue and clinical validation are necessary.