Analysis of tumors using single-cell and spatial modalities is critical to advance our understanding of cancer. The growth of technologies that enable these studies provides an increasing number of single cell datasets. Integrating such data across studies will increase the impact of individual studies and speed cancer research. Most existing integration approaches are tailored to transcriptomic data and assume large sets of shared features, an assumption that fails for lower-dimensional proteomic measurements. Here, we developed CellFuse, a deep learning-based integration framework that unifies antibody-based proteomic datasets including high-dimensional cytometry, CITE-seq, and spatial proteomics data. Leveraging supervised contrastive learning, CellFuse learned a shared embedding space that enabled accurate cross-modality cell type prediction and robust label transfer across tumor samples and experimental conditions. Applied to datasets spanning peripheral blood, bone marrow, and lymphoma, CellFuse consistently outperformed existing approaches in recovering clinically relevant populations, including rare malignant and immune subsets. In solid tumors, it reconstructed spatially resolved microenvironments, capturing interactions between malignant, stromal, and immune cells that correlated with treatment response. By enabling scalable, modality-agnostic integration, CellFuse provides a powerful tool to uncover prognostic cell states and delineate the architecture of the tumor-immune ecosystem with translational relevance, driving cancer discoveries.
Koladiya et al. (2026) studied this question.