TMIC-68. DEEP LEARNING OF THE SPATIAL IMMUNE ENGAGEMENT IN GLIOBLASTOMA UNDER TEMOZOLOMIDE TREATMENT

Abstract Effective immune synapses are a prerequisite for desirable immune surveillance and control of tumour growth. Cell-to-cell interactions form the basis of immune synapses, resulting in either immunoprotected or immunosuppressed manifestations in the tumour brain that could explain glioblastoma (GBM) recurrence. Using state-of-the-art multiplex imaging, spatial statistics, and spatially aware deep learning tools in 14 patient-matching tissues from primary and recurrent GBM, we detected spatial immune networks (SINets). SINets across 105 regions of interest comprising 2.3 million cells show stronger spatial engagement between myeloid and lymphoid lineages at recurrence. SINets’ architecture in primary tumours is associated with relapse timing. Patients, where the SINets are more modular, recur sooner than patients with denser SiNets. In addition, this association presents sexual dimorphism in our cohort (N=3 females, and N=4 males). Phenotype-to-phenotype spatial coexistence also contributes to enhancing phenotype prediction. A spatial-aware deep learning model trained with single-cell data from primary tumours reaches an average balanced accuracy of 0.76 in primary tumour cellular communities unseen during training, compared to a spatial-agnostic model that reached an average balanced accuracy of 0.69. In summary, we detected a distinct spatial engagement in glioblastoma undergoing chemo-radiation treatment. The spatial patterns of coexistence represent ecological niches that contribute to defining cell occurrence and may underpin effective immune synapses. Overall, the spatiotemporal elements provide a newer approach for dissecting glioblastoma’s complexity and discovering opportunities for stratifying patients by harnessing spatial information encoded in multicellular niches.