SPRINT enables spatial proteomic profile reconstruction through multimodal fusion of histology and transcriptomics

Spatially resolved protein expression is essential for characterizing tissue organization and immune microenvironments, yet existing computational inference methods predominantly rely on either histological images or gene expression alone, limiting their ability to recover proteins that are weakly associated with a single modality. Here we present SPRINT, a multimodal framework that reconstructs spatial proteomic profiles by jointly leveraging spatial gene expression and accompanying H&E-stained histological images. SPRINT combines resolution-adaptive image encoders with token-level self-attention to fuse morphological and transcriptional information. Across human tonsil, mouse spleen and human brain tissues, SPRINT reconstructed protein-specific spatial patterns across sections, donors and spatially disjoint regions. Its compatibility with different image resolutions, transcriptomic gene and assay configurations broadens its applicability across spatial transcriptomics datasets. Ablation analyses showed that multimodal fusion compensated for proteins with weak associations to either tissue morphology or transcript abundance, thereby expanding the range of immune-microenvironment proteins that could be reliably inferred. SPRINT further supported cross-species protein reconstruction, and the reconstructed profiles improved the recovery of spatial tissue domains. The framework was also extendable beyond proteomics, reconstructing spatial metabolite distributions in an independent tissue section when trained directly on metabolomic measurements. Together, these results establish SPRINT as a general framework for completing spatial molecular profiles across tissue sections, datasets, image-resolution and species, while enabling downstream characterization of tissue organization.