Inferring Spatial Organization Of Tumor Microenvironment From Single-Cell Rna Sequencing Data Using Graph Embedding.

Abstract Spatial heterogeneity of diverse cellular components in the tumor microenvironment (TME) plays a critical role in the reprogramming of tumor initiation, growth, invasion, metastasis, and response to therapies. Systematic knowledge of TME spatial organization with regards to immune infiltration and tumor resource distribution is of high clinical significance. High throughput single-cell RNA sequencing (scRNA-seq) has become a revolutionary approach for studying cell composition and the development of TME. However, the spatial information of cells is lost as the tissue must be dissociated before the sequencing is performed. While various spatial techniques are emerging, their applicability is still rather limited. To address this challenge computationally, we develop a novel de novo framework to reconstruct TME spatial organization from scRNA-seq data. We hypothesized that cell spatial organization in a microenvironment is mainly determined by cell identity and interactions between different cells. In particular, the spatial organization of structural cells and immune cells follow different mechanisms. Neighboring structural cells, which share similar whole transcriptome profiles, form a scaffold of the TME; immune cells, whose activities are influenced by the structural cells, migrate in the scaffold to interact with structural cells and exert their functions. The algorithm models the scaffold of structural cells using adaptive nearest neighbor graph by taking the cell density estimation into the consideration, where the nearest neighbor graph was further augmented by inserting immune cells into the appropriate locations of the scaffold according to the LR similarities. To reconstruct 3D spatial organization while preserving the cell topology represented by the graph, we employed a graph embedding strategy to minimize the discrepancy between the graph topology and the embedded 3D space. We evaluated the framework on two diffuse intrinsic pontine gliomas (DIPG) samples from a mouse model with coupled scRNA-seq and spatial transcriptome (ST, 10x Visium platform) data. The predicted spatial organization successfully separates the major cell types. The T-cell infiltrated tumor, verified by the T-cell spatial spots of the ST image, is well recapitulated. We deconvoluted the ST data by integrating the scRNA-seq data using SPOTlight. The neighborhood enrichment distributions of predicted spatial organization and the spot deconvoluted ST data show high consistency as measured by Kullback-Leibler divergence. We found heterogeneous neighborhood composition of CD8+ T-cells, indicating diverse clonality and functions with respect to their locations in the TME. Citation Format: Liang Ding, Hao Shi, Koon-Kiu Yan, Yogesh Dhungana, Sivaraman Natarajan, John Easton, Jason Chiang, Christopher L. Tinkle, Xiaoyan Zhu, Liming Cai, Suzanne J. Baker, Hongbo Chi, Jiyang Yu. Inferring spatial organization of tumor microenvironment from single-cell RNA sequencing data using graph embedding [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2021; 2021 Apr 10-15 and May 17-21. Philadelphia (PA): AACR; Cancer Res 2021;81(13_Suppl):Abstract nr 237.