Point-Process Modelling of Cell-Type Interaction Structure in CosMx Glioma Data for Hypothesis-Driven Tumor Inference from scRNA-seq

Abstract

Glioblastoma (GBM) exhibits marked spatial heterogeneity that is lost after tissue dissociation for single-cell RNA sequencing. Here, we developed a spatial-statistical framework to characterize tumor-microenvironment organization and derive molecularly predictable spatial phenotypes from CosMx data. Eight GBM specimens comprising 2,427,362 cells, 15 cell populations, and four malignant states were analyzed, revealing substantial inter-patient differences in cellular composition and density. Fine finite-element meshes best preserved continuous spatial intensity and supported whole-tissue Log-Gaussian Cox Process modelling. Immune populations, including macrophages, microglia, monocytes, cDCs, neutrophils, and T cells, showed spatial attraction toward malignant states. These models generated cell-level measures of conditional intensity and tumor-proximity probability. Transcript abundances and metabolic pathway scores served as the sole predictive features to model these spatial phenotypes. Transcriptomic representations outperformed metabolic pathway scores. Leave-one-patient-out validation reduced predictive performance, and patient-wise harmonization failed to restore cross-patient transportability. Spatial outcome distributions varied between specimens, indicating that molecular-spatial relationships are strongly conditioned by patient-specific tumor architecture. Predicted interaction scores stratified transcriptional subclusters within cell types in external non-spatial single-cell datasets. This framework links spatial point-process phenotypes to molecular states and provides a basis for inferring spatial context from non-spatial single-cell data. Generalizing these models requires larger and more diverse spatial cohorts.