Large-scale integration of single-nuclei and spatial transcriptomics from the ETOP BEAT-meso trial reveals clinically relevant heterogeneity in malignant pleural mesothelioma

Malignant pleural mesothelioma (MPM) is a rare, aggressive cancer of the lung lining, predominantly caused by asbestos exposure (1,2). Despite advancing therapies, 5-year survival remains poor at 10-20% (3,4). A major challenge in managing MPM is its extensive heterogeneity, which contributes to variable treatment response. Comprehensive characterisation of this heterogeneity may help guide more effective therapies. Meeting this need requires large-scale, multimodal datasets that capture the cellular, spatial and molecular landscape of MPM. We analysed FFPE tissues from 159 patients enrolled in the BEAT-meso trial (5), generating the largest multimodal MPM dataset from a single clinical trial. The dataset includes paired single-nuclei FLEX RNA-seq (snRNA-seq; 612,587 cells), spatial Xenium transcriptomics (37,949,307 cells), H&E images, and matched clinical data. We built a high-resolution snRNA-seq atlas through extensive preprocessing, reference integration, and hierarchical cell type annotations. Clinical variables were used to stratify the cohort and assess variation in cellular composition, pathway activity, and molecular patterns. Xenium data were used to contextualise the snRNA-seq-derived annotations, and transcriptomic data of malignant cells were utilised to identify clinically associated programmes. We used foundation models to learn multimodal representations across histology and transcriptomic data, and studied relationships across modalities and their associations with clinical variables.Histology-based stratification revealed differences in cell-type composition, pathway activation, and checkpoint signalling between epithelioid and non-epithelioid tumours. Foundation-model analysis identified patients with sarcomatoid-like molecular signatures, revealing heterogeneity beyond standard classification. Transcriptional programme analysis further refined malignant cell states across histologies. Integration with spatial transcriptomics confirmed the presence and localisation of all snRNA-seq-derived cell types and enabled identification of tertiary lymphoid structures. We present the most extensive multimodal resource for MPM, integrating single-nuclei, spatial and clinical data from 159 BEAT-meso patients. This multimodal framework refines molecular and histological characterisation of MPM, highlighting features associated with aggressive disease, and provides the cancer and computational biology communities with a scalable reference for benchmarking, training next-generation models, and accelerating biomarker and therapeutic discovery.