Classification of Tumor Spatial Transcriptomics Data and Methods

Abstract

Spatial transcriptomics technologies combine gene expression profiles with the spatial positions of tissues, providing an unprecedented perspective for dissecting the heterogeneity of the tumor microenvironment. However, tumor tissues are highly heterogeneous, and their discontinuous, sparse, and blurred-boundary features pose unique challenges to spatial domain identification. This paper systematically sorts out the distribution characteristics of current tumor spatial transcriptomics data in terms of acquisition platforms, cancer types, and data quality, and points out that data imbalance imposes substantial constraints on algorithm selection. Furthermore, according to the richness of input information, 21 mainstream computational methods are divided into three levels and six subclasses. The inherent assumptions and applicable boundaries of each type of method are explained, and their known applications in tumor research are summarized. On this basis, this paper constructs a decision framework based on data characteristics, and verifies the effectiveness of the framework through empirical analysis of seven representative methods on 10x Visium breast cancer data. The results show that the effectiveness of a method depends on the matching degree between its inherent assumptions and data characteristics, rather than the advancement of the method itself. This paper provides a systematic reference for tumor researchers to select appropriate methods according to their own data conditions.