Recent breakthroughs in spatial multi-omics enable simultaneous profiling of different modalities while preserving tissue architecture, providing unprecedented opportunities to explore tissue complexity. However, due to the sparse and noisy nature of the data, interpreting these complex tissue structures and cellular communication remains challenging. We present MultiSP, a deep learning framework that enhances data representation through efficient spatial and feature similarity fusion, modality-specific probabilistic generative modeling, and cross-modality adversarial learning. Applied to various spatial multi-omics datasets, it outperforms existing methods in capturing biologically interpretable spatial domains. MultiSP also denoises spatial data, uncovers modality-specific spatial variations, and reveals gene regulation mechanisms. In the tumor microenvironment, it unravels fine-resolution cellular distribution maps, such as spatially neighboring macrophage-enriched sub-regions with distinct prognosis outcomes. Additionally, MultiSP facilitates the inference of spatially multimodal cell-cell communication. Together, MultiSP serves as a powerful framework for uncovering spatially multimodal heterogeneity and communication by integrating complementary information from multiple modalities.
Mo et al. (Sun,) studied this question.