Abstract 913: Automated reasoning artificial intelligence (AI) to model the cell-cell biochemical and cellular interactions in the tumor microenvironment (TME) of oral cancer (OC)

Abstract Our goal was to study the spatial complexity of HPV negative OC by integrating scRNAseq and spatial transcriptomics from 11 samples including 5 from never smokers never drinkers (NSND). We used Turing Biosystems proprietary graph database engine and automated reasoning AI to analyze and model the cell-cell biochemical and cellular interactions in the TME. The tumors H&E images, Visium 10X, and single cell 10X data were pre-processed using standard methods. QC-based genes and cells filtering, gene expression normalization, dimensionality reduction, clustering of Visium spots, and identification of spatially variable genes were performed. We generated multiple networks from tumor data and prior biological knowledge to integrate different layers of information: single-cell and spatial transcriptomics data, cell-cell interactions networks, cell neighboring, spatial correlation networks. We first used a novel network-based technique to identify the spatial distribution of any cell subtypes (e.g. CAF, T cells), by integrating differentially expressed genes from scRNAseq, prior knowledge and prior data (e.g. TCGA). We used a multilayer network approach to integrate the different networks of data and an automated reasoning AI analysis which allows to interpret the results based on a reasoning on biological and clinical knowledge in opposition to machine learning methods based on statistical patterns. We identified CAFs/T cells/IFN-g to be strongly correlated within the tumor islets interacting with the stroma in NSND while they were more correlated within the stromal regions in smokers. We also identified 2 or 3 subtypes of tumor islets areas based on their functional states (e.g., immune pathways activated or metabolic differences), allowing us to unravel a more detailed map of the intratumoral heterogeneity when compared to histopathology annotations. In order to get some insights into the functional impact of the spatial distribution of cells, the tumors were then represented as multilayer networks to identify all possible cell-cell interactions in the TME. This was followed by a spatial simulation of these interactions using logic (e.g., inhibition, activation) and correlative (e.g., gene expression, imaging features correlations) interactions in the networks to identify spatial interactions. From that we analysed all the possible known protein-protein and metabolite-protein interactions between the cell types analysed and mapped the tumors with the gene expressions corresponding to the pairs of interacting proteins (or enzymes linked to metabolites). We found an interaction between CAF and Tregs via CXCL12 (CAFs) and CXCR4 (Tregs). In conclusion, automated reasoning AI integrating scRNAseq and spatial transcriptomics allows an in-depth analysis of cells spatial distribution and its functional impact in OC. Citation Format: Karène Mahtouk, Maxime Vincent, Yannick Le Meitour, Béatrice Vanbervliet-Defrance, Sonia Canjura, Cyril Degletagne, Laurie Tonon, Lucas Michon, Jebrane Bouaoud, Aude Excoffier, Philippe Zrounba, Rémy Boutonnet, Adam Amara, Pierre Saintigny. Automated reasoning artificial intelligence (AI) to model the cell-cell biochemical and cellular interactions in the tumor microenvironment (TME) of oral cancer (OC) [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2024; Part 1 (Regular Abstracts); 2024 Apr 5-10; San Diego, CA. Philadelphia (PA): AACR; Cancer Res 2024;84(6_Suppl):Abstract nr 913.