Tumor metastasis is the leading cause of death in cancer patients, with epithelial–mesenchymal transition (EMT) playing a key role. To systematically elucidate the cellular mechanisms and molecular networks through which EMT drives metastasis across cancers, this study integrated transcriptomic data from over 1.2 million single cells across 265 samples representing 12 primary epithelial cancers, constructing a comprehensive pan-cancer single-cell atlas covering diverse stages and metastatic states. By analyzing the metastatic features and interaction networks of malignant epithelial cells and cancer-associated fibroblasts (CAFs), we identified cancer-specific metastasis-related gene sets. Based on these genes, multiple machine learning algorithms were applied to build cancer-specific and cross-cancer metastasis prediction models, leading to the development of the metastasis prediction score (MPS) and global metastasis prediction score (GMPS). Both scores showed excellent predictive performance in independent test and external validation cohorts. MPS exhibited higher cancer specificity, whereas GMPS showed stronger cross-cancer generalization. Moreover, elevated MPS and GMPS reflected immunosuppressive tumor microenvironment features and were significantly associated with poor prognosis across multiple cancer types. Finally, through a drug repositioning framework, we identified several potential anti-metastatic compounds targeting the metastasis network, among which Fostamatinib demonstrated broad-spectrum therapeutic potential against metastasis in multiple cancers.