BackgroundThe Oxford Classic (OxC) prognostic signature classifies high-grade serous ovarian cancer (HGSOC) into five transcriptional programs, with epithelial-to-mesenchymal transition (EMT) marking poor prognosis. While successful in bulk transcriptomics, the spatial organisation of these programs within the tumour microenvironment remains unexplored.MethodsWe developed the Signature-guided Zero-inflated Beta Variational Autoencoder (Sig-ZIB-VAE), a deep learning deconvolution method tailored for spatial transcriptomics data, and applied it to a large-scale HGSOC cohort comprising 94 tumours to quantify spatial cellular organisation. Prognostic significance was assessed using penalised Cox proportional hazards regression integrating clinical, molecular, and spatial features.ResultsHere we show that EMT cells form dense homotypic clusters broadly depleted from stromal and immune neighbourhoods, yet maintain selective monocyte co-localisation at cluster boundaries. EMT-high tumours display enhanced spatial reorganisation characterised by increased clustering and connectivity, forming locally concentrated mesenchymal-rich domains. Survival analysis confirms EMT-high status as an adverse prognostic factor.ConclusionsCritically, spatial metrics of immune cell organisation-particularly monocyte connectivity and clustering-provide substantially stronger prognostic discrimination than EMT proportion alone, demonstrating that tumour microenvironment architecture supersedes cellular composition in determining clinical outcomes in HGSOC.