Coordinated Enhancement of Molecular and Spatial Fidelity for Biological Inference in Spatial Transcriptomics

Abstract

Understanding tissue organization requires accurate molecular quantification and faithful spatial localization. Spatial omics, particularly spatial transcriptomics (ST), face measurement trade-offs that limit simultaneous achievement of molecular and spatial fidelity across spatial resolution, detection sensitivity, molecular-tissue alignment and cell-level fidelity. Existing computational methods typically target individual aspects and optimize them in isolation, despite their interdependence. Here we present FOCUS, a multimodal generative foundation model for coordinated ST enhancement. FOCUS integrates native ST with histology images, cell masks, scRNA-seq references and spatial co-expression priors through task-specific modules, with cross-task coordination enabling joint correction of co-occurring limitations. FOCUS improved the performance of individual enhancement tasks and provided additional gains with coordinated mitigation of coupled limitations. Across immunology, tissue development and oncology, FOCUS recovered biologically consistent patterns obscured in native ST or isolated enhancement, enabling clearer inference of neutrophil-tumor interactions, spatial osteocyte differentiation and primary-metastatic HNSCC programs. Together, FOCUS establishes coordinated enhancement of molecular and spatial fidelity as a strategy for downstream biological inference from ST.