SpaceRec: Reference-anchored low-rank prediction of high-resolution spatial expression from histology

Spatial transcriptomics measures gene expression at sparse, multicellular spots, making histology-guided high-resolution reconstruction fundamentally underdetermined. We introduce SpaceRec, a section-specific framework that regularizes this problem through a reference-anchored factorization. RCTD constrains which cell states are present and in what proportions, H&E proposes where they occur, single-cell mean profiles anchor what they express, and measured Visium counts supervise reconstruction. SpaceRec represents expression with a rank-$C$ reference backbone and a bounded, morphology-dependent correction that captures residual platform and within-state variation without unconstrained departure from the reference. Evaluated against registered Xenium and Visium HD measurements in breast and colorectal cancer, SpaceRec consistently improves high-resolution expression recovery over dense and nucleus-centered baselines, including in unmeasured inter-spot regions. It also recovers coherent tumor, stromal, myoepithelial, and immune organization. These results show that explicit biological factorization provides an interpretable and effective basis for reconstructing dense spatial expression from H&E and spot-level measurements.