A Systematic Review of Clustering Methods for Spatial Multi-omics: A Unified Analysis of Model Architecture and Parameter Sensitivity

Abstract

While preserving the spatial coordinates of tissues, spatial multi-omics technologies jointly measure gene expression, protein abundance and chromatin accessibility, offering complementary molecular perspectives for deciphering tissue microenvironment heterogeneity. Nevertheless, the interwoven problems of high-dimensional sparsity, heterogeneous modality distribution and spatial autocorrelation require spatial domain identification to not only align cross-modal information, but also balance modality-specific signals and tissue spatial continuity. Following the logical thread of “data characteristics-core challenges-integration strategies-empirical boundaries”, this paper first defines the task and data properties of spatial multi-omics clustering. Key issues are then summarized as cross-modal heterogeneity, information complementarity and spatial consistency. Representative approaches including graph-structure modelling, attention and contrastive learning, probabilistic factor decomposition, ensemble consensus and biological prior knowledge are compared comprehensively from the architectural, statistical and semantic levels. On this basis, this study takes Leiden clustering as a reproducible baseline, carries out 72 groups of parameter evaluations on MouseBrain and HumanTonsil datasets, and systematically compares the effects of PCA dimension, neighbour number k and resolution on clustering granularity, consistency with reference labels and spatial continuity. The results show that higher resolution is accompanied by an increase in cluster number and a decline in weighted Moran’s I; the optimal configurations for ARI, NMI and spatial continuity do not overlap in the MouseBrain dataset. These findings reveal clear parameter-selection sensitivity of basic clustering workflows and provide empirical references for subsequent comparisons of complex integration models under unified data, pre-processing and evaluation standards.