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

Glioblastoma (GBM) presents significant spatial differences that are lost when tissue is broken down for single-cell RNA sequencing. Researchers have created a spatial-statistical model to understand tumor organization and generate cellular phenotypes based on the data. The model, applied to eight GBM samples containing 2.4 million cells, 15 cell types, and four malignant conditions, highlights significant variability in patient-specific cell compositions and densities.

Fine mesh models that maintain continuous spatial intensity proved effective for whole-tissue Log-Gaussian Cox Process modeling. By utilizing these models, researchers identified spatial patterns in immune cell populations, such as macrophages, microglia, monocytes, cDCs, neutrophils, and T cells, which tend to cluster around malignant states. The models produced precise metrics to assess cell-level interactions and proximity to tumors.

The analysis relied solely on transcript abundances and metabolic pathway scores as predictive features for spatial phenotypes. Subsequent testing showed that transcriptomic data provided better predictive capabilities than metabolic pathway scores. In a leave-one-patient-out validation, model performance decreased when patient-specific information was excluded. Additionally, attempts to harmonize patient data failed to maintain cross-patient applicability.

Spatial distributions of tumor outcomes varied across specimens, indicating that molecular-spatial relationships are heavily influenced by the unique architecture of each patient's tumor. The spatial interaction scores were able to classify transcriptional subclusters within individual cell types when applied to external single-cell data that did not include spatial information.

Ultimately, this framework bridges the gap between spatial point-process phenotypes and molecular states, offering a potential method for deducing spatial context from non-spatial single-cell data. To further improve these models, larger and more diverse spatial cohorts will be necessary.