Spatially resolved cell-cell architecture in WHO grade 4 IDH- wildtype glioblastoma revealed by CosMx 6K discovery panel

Background: 
 IDH -wildtype glioblastoma is well-known for its intratumoral heterogeneity, with malignant, immune, and stromal populations coexisting in shared microenvironment that may shape hyper-progression and therapeutic resistance. These could be highly localized while spot-resolution spatial assays probably mask the underlying cellular architecture. Targeted single-cell spatial transcriptomics, like CosMx SMI now provide confident cells segmentation for delicate patient-specific interaction networks across GBM. Methods :
GBM tissues were profiled using the CosMx 6K Human Discovery Panel. Standardized preprocessing, including RNA QC, batch harmonization, and multi-reference cell typing were completed with Seurat v5. Cells underwent stepwise cleanup, which are the removal of AtoMx-flagged cells and those below the first quartile of number of detected genes and total transcript counts and further exclusion of fields of view (FOVs) retaining <50% of cells after cell filter. SCTransform, PCA and UMAP were applied on raw counts using AtoMx-recommended parameters with Giotto v4. Spatial networks and neighborhood were built using radius-based and k-nearest neighbor criteria. Co-localization was evaluated with Ripley's K-function. L-R networks were inferred with CellChat v2. All computation and analysis were with reference to AtoMx manual v2.1 (MAN-10162-10) under R 4.4.1 located at HPCF2, CPOS, HKU. Results: 
After the stringent filtering, 297,125 confident cells across 413 FOVs in 23 TMA cores were retained for studying cell-cell communication. Pearson correlation of 0.99 indicated that cell density did not bias per-core analyses. Radius-based spatial network revealed extensive intratumoral heterogeneity between and within patient. Of notes, Patient #025 showed a marked shift, with a stem-like/mesenchymal-like cluster dominating one core and differentiated-like cells predominating in another, given both cores extracted from the same FFPE block resected at primary diagnosis. k-nn modeling showed reproducible tumor-tumor and tumor-vascular adjacency, while myeloid and lymphoid cells remained spatially diffuse across the cores, which was further confirmed significantly with Ripley's K function. Next, highly heterogeneous L-R pairs were inferred across patients, with a mean of 546 highly variable ligand-receptor pairs per core. A set of canonical GBM interaction pathways, like NCAM1-, CD99-, and JAG1-mediated signaling was conserved across cores. Apart from the homotypic network within tumoral and vasculature, there are extensive patient-specific L-R, like CNTN1-NOTCH1 networking in patient # 011. Conclusions: 
 By resolving cell-cell interaction, CosMx exposed distinct ligand-target dependencies across patients, which is a basis for designing bespoke neoantigen-directed vaccines or ligand-target bispecific antibody approaches.