SOMaC：一个可解释、可泛化的框架，用于整合与聚类多模态空间组学以解析癌症中的细胞状态和免疫生态位

Spatial omics technologies, including spatial transcriptomics and multiplex spatial proteomics, provide unprecedented opportunities to characterize cellular organization, microenvironmental interactions, and functional states within intact tumor tissues. However, integrating these heterogeneous modalities and resolving spatially coherent cell states and immune niches remain major computational challenges. Existing approaches typically analyze individual modalities in isolation, rely on expression-based feature spaces that overlook higher-order molecular structure, and provide limited interpretability for biological discovery. We present SOMaC, an interpretable and generalizable framework for integrating and clustering multimodal spatial omics data to resolve cellular and microenvironmental architecture in cancer tissues. SOMaC extends the OmicsMap paradigm by transforming each cell or spot into an image-like molecular interaction map that encodes pairwise gene or protein correlations, capturing regulatory structure beyond conventional expression matrices. These biologically grounded representations are jointly integrated with spatial topology using a convolutional autoencoder, graph neural network, and learnable-center clustering module, enabling end-to-end optimization of molecular-spatial embeddings for domain discovery. Across 13 benchmark datasets spanning nine human and mouse tissues profiled by 10x Visium, Xenium, Slide-seqV2, Stereo-seq, imaging mass cytometry (IMC), and CODEX, SOMaC consistently outperformed leading methods such as GraphST and SEDR. SOMaC achieved significantly higher clustering quality, with an average Silhouette score of 0.21 (vs. 0.11 for GraphST) and more than a two-fold improvement in the Calinski-Harabasz Index, indicating tighter and more separable spatial domains. On multiplex proteomics datasets, including human colon cancer IMC, SOMaC demonstrated strong cross-modality generalization, achieving a Calinski-Harabasz Index of 40,541 and a Davies-Bouldin Index of 3.03. Notably, SOMaC resolved fine-scale tumor-immune interfaces, stromal barriers, and functional immune niches that were poorly captured by existing methods. By jointly modeling molecular interaction structure and spatial architecture across transcriptomic and proteomic platforms, SOMaC provides a unified and interpretable framework for dissecting tumor ecosystems. This approach enables high-resolution discovery of spatially organized cell states, regulatory programs, and microenvironmental niches, offering new insights into cancer biology and therapeutic targeting.