P04.05.B DEEP LEARNING ANALYSIS OF SPATIALLY RESOLVED SINGLE-CELL TRANSCRIPTOMICS UNCOVERS CELLULAR REORGANIZATION IN GLIOBLASTOMA AT RECURRENCE

Abstract BACKGROUND Recent studies using single-cell RNA sequencing (scRNA-seq) have provided insights into the cellular composition and molecular states in glioblastoma (GBM). However, these studies have not identified significant differences between primary and recurrent GBMs. We hypothesize that other factors, such as spatial organization and cellular interactions, may provide key information driving recurrence in GBM. Spatial transcriptomics techniques, such as MERFISH (Multiplexed Error-Robust Fluorescence In Situ Hybridization), provide single-cell spatial resolution of tissues, enabling the identification of distinct molecular features. MERFISH, in particular, allows for the 3D reconstruction of gene expression signatures in consecutive tissue slices, providing a detailed spatial map of the tumor cells and surrounding environment. This technology is invaluable in studying longitudinal GBMs, where changes in tumor organization, immune cell interactions, and mechanisms of therapeutic resistance remain poorly understood. MATERIAL AND METHODS We analyzed 102 tumor samples consisting of primary and recurrent GBMs and normal brain tissue using Visium and MERSCOPE technologies. We used supervised artificial intelligence (AI)-based learning, in-silico perturbation and explainable-AI to identify distinct spatial and cell-type-specific transcriptomic reorganization associated with GBM recurrence. RESULTS We developed a spatial transcriptomics atlas of GBM at sub-cellular resolution. Our analysis revealed distinct cellular neighborhoods in the primary and recurrent GBM tumors, with changes leading to the dynamic spatial reorganization of GBM at recurrence. The 3D spatial images acquired provided a comprehensive view of the tumor cells interacting with the tumor microenvironment in situ. Our explainable-AI models successfully identified distinct spatial transcriptomic features associated with recurrence in GBM. The identified spatially distinct cellular neighborhoods reveal dynamic shift in cellular organization following in silico perturbation. CONCLUSION This study provides a high-resolution spatial map of longitudinal GBM that offers new insights into the dynamic cellular reorganization of GBM at recurrence. Our findings highlight the role of cellular reorganization during tumor recurrence in GBM, a potential avenue for targeted therapeutic intervention.