Deciphering cell-type composition is critical for charting single-cell spatial maps and cellular atlases of organisms. Most Spatial Transcriptomics (ST) data lack single-cell resolution, computational deconvolution methods have emerged to characterize the composition and spatial heterogeneity of different cell types within spots and tissues. To date, various cell-type deconvolution methods have been developed, each exhibiting its own distinct advantages. To conduct a comprehensive review of these methods, we first provide a formal description of the deconvolution problem. Then, we analyze the advantages and pitfalls of these methods based on their mathematical models. We further discuss related downstream analyses, potential applications, and future directions. In summary, our review aims to guide researchers in gaining an in-depth understanding of the spatial deconvolution problem, enabling them to make informed choices in spatial analysis and advance research on related fields, such as the developmental biology, tumor microenvironment, disease pathology, and clinical treatments.

Kirsten rat sarcoma viral oncogene homolog (namely KRAS) is a key biomarker for prognostic analysis and targeted therapy of colorectal cancer. Recently, the advancement of machine learning, especially deep learning, has greatly promoted the development of KRAS mutation detection from tumor phenotype data, such as pathology slides or radiology images. However, there are still two major problems in existing studies: inadequate single-modal feature learning and lack of multimodal phenotypic feature fusion. In this paper, we propose a Disentangled Representation-based Multimodal Fusion framework integrating Pathomics and Radiomics (DRMF-PaRa) for KRAS mutation detection. Specifically, the DRMF-PaRa model consists of three parts: (1) the pathomics learning module, which introduces a tissue-guided Transformer model to extract more comprehensive and targeted pathological features; (2) the radiomics learning module, which captures the generic hand-crafted radiomics features and the task-specific deep radiomics features; (3) the disentangled representation-based multimodal fusion module, which learns factorized subspaces for each modality and provides a holistic view of the two heterogeneous phenotypic features. The proposed model is developed and evaluated on a multi modality dataset of 111 colorectal cancer patients with whole slide images and contrast-enhanced CT. The experimental results demonstrate the superiority of the proposed DRMF-PaRa model with an accuracy of 0.876 and an AUC of 0.865 for KRAS mutation detection.