EPCO-43. A MULTI-MODAL SINGLE-CELL AND SPATIAL TRANSCRIPTOMIC ATLAS REVEALS GENE-PROGRAM AND NEIGHBOURHOOD DRIVERS OF MENINGIOMA MALIGNISATION

Abstract BACKGROUND Meningiomas range from indolent to highly aggressive, yet current grading captures only part of its biological diversity. METHODS We built an integrated atlas by profiling 302 133 nuclei from 47 meningiomas with scnuc-RNA-seq and mapping 37 samples with Visium spatial transcriptomics. Cell types were annotated, and a knowledge-graph-restricted “topic” model (SPECTRA) distilled tumour RNA profiles into recurrent gene-programmes across samples. Spatial architecture was captured by a graph-attention network that merges per-spot cell-type estimates with tokenised CONCH-pre-trained-VisualTransformer histology embeddings via cross-modal attention RESULTS Thirty-nine robust gene programmes emerged, reflecting hallmark processes such as proliferation, invasion and extracellular-matrix remodelling, immune evasion and metabolic adaptation. Tumour cells varied along five independent axes representing invasiveness, growth rate, capsule formation and brain-barrier phenotype. Aggregate programme scores traced a smooth benign → atypical → anaplastic trajectory that aligned with methylation classes and highlighting discrepancies of staining only WHO classification to capture transcriptional variance in Meningioma. Aligning topics usage with malignant transformation suggests a trajectore of hyaluronan-rich stroma and macrophage “shield” formation, followed by a phase of rapid proliferation driven by RAS–MAPK, Wnt/β-catenin and YAP/integrin signalling during malignisation. Associating per cell program activity with copy-number changes uncovered that hypoxia/TGF-β modules were enriched in 1p-loss tumours (odds ratio 3.2, p = 0.004) and DNA-repair modules in 6p-gain tumours (odds ratio 2.7, p = 0.01). Using a GAT for multimodal representation of meningioma neighborhoods captured significant biological effects allowing us to predict from neighborhood embeddings DNA-methylation classes with 93 % accuracy (AUROC 0.94). Pinpointing in malignisation neighbourhoods rich in macrophages, endothelial cells and exhausted T cells, marked by ECM-remodelling and axon-guidance signals. CONCLUSIONS Combining single-cell sequencing, spatial transcriptomics and image-aware machine learning clarifies the molecular and architectural logic of meningioma progression. The identified gene programmes and spatial niches outperform morphology-only grading, link to specific chromosomal events and can be recognised on routine H&E slides, offering practical biomarkers and new therapeutic entry points for high-risk meningioma.