BRIDGE: An Interpretable Framework for Transcriptome-Scale Completion of Panel-Limited Spatial Transcriptomics

Abstract

Imaging-based spatial transcriptomics provides single-cell resolution but is restricted to targeted gene panels, leaving most transcriptional variation unexplored. Existing imputation approaches often rely on latent-space alignment, which can distort biological structure and capture limited within-type heterogeneity. We present BRIDGE, an interpretable framework that expands panel-limited spatial transcriptomics to transcriptome-level resolution through anchor-guided linear calibration and multi-scale neighborhood-based prediction. BRIDGE learns a global calibration matrix that directly adjusts shared-gene expression between technologies and provides explicit gene-level interpretability. Using the calibrated reference, BRIDGE predicts unmeasured genes by integrating multi-scale local neighborhoods to achieve both accuracy and robustness. Across seven datasets from CosMx, MERFISH, Xenium and multiple tissues and species, BRIDGE exceeds or matches existing methods on gene-level and cell-level accuracy. The calibration matrix offers clear biological interpretation, and the completed transcriptomes recover fine-scale spatial patterns such as cortical lamination and support refined characterization of cell-state heterogeneity, including B-cell states and Stromal subtypes in human breast cancer. BRIDGE provides a robust and interpretable solution for extending imaging-based spatial transcriptomics to transcriptome-scale analysis and enables deeper investigation of microenvironment-dependent cellular programs.