SCIITensor: A tensor decomposition based algorithm to construct actionable TME modules with spatially resolved intercellular communications

Abstract

Advanced spatial transcriptomics (ST) technology has paved the way for elucidating the spatial architecture of the tumor microenvironment (TME) from multiple perspectives. However, available tools only focus on the static molecular and cellular composition of the TME when analyzing the high-throughput ST data, neglecting to uncover the in-depth spatial co-variation of intercellular communications arising from heterogeneous spatial TMEs. Here, we introduce SCIITensor, which decomposes TME modules from the perspective of spatially resolved intercellular communication by spatially quantifying the cellular and molecular interaction intensities between proximal cells within each domain. It then constructs a three-dimensional matrix, formulating the task as a matrix decomposition problem, and identifies biologically relevant spatial interactions and TME patterns using Non-Negative Tucker Decomposition (NTD). We evaluated the performance of SCIITensor on liver cancer datasets obtained from multiple ST platforms. At the research setting of a single-sample investigation, SCIITensor precisely identified a functional TME module indicating a tumor boundary structure specific domain with co-variant interaction contexts, which were involved in construction of immunosuppressive TME. Moreover, we also proved that SCIITensor was able to construct TME meta-modules across multiple samples and to further identify an immune-infiltration associated and sample-common meta-module. We demonstrate that SCIITensor is applicable for dissecting TME modules from a new perspective by constructing spatial interaction contexts using ST datasets of individual and multiple samples, providing new insights into tumor research and potential therapeutic targets.