Composition‐Aware Cross‐Sectional Integration for Spatial Transcriptomics

Spatially resolved transcriptomics (SRT) can profile contiguous tissue sections at near‐cellular resolution. However, building multi‐slice maps remains difficult because each spot mixes multiple cell types and section‐specific batch effects disrupt shared structure. Most pipelines handle cell‐type deconvolution, domain detection, and cross‐section integration separately, cluster on gene‐level variance rather than cellular composition, and rely on post‐hoc alignment that misplaces domain boundaries. To overcome these limitations, we present FUSION, a fast, unified, spatial integration and cell‐type composition method that leverages single‐cell RNA references to unify multiple tasks. FUSION models sequencing read as arising from a latent topic that captures the transcriptomic program of a reference cell type and aggregates read‐level topic probabilities into spot‐wise compositional embeddings. Clustering directly in this embedding space provides interpretable, composition–driven domains and aligns homogeneous regions across slices. Through a computationally efficient inference scheme augmented with a distance‐aware and sparse self‐attention kernel, FUSION minimizes runtime and maintains low memory usage. To align embeddings across sections without erasing biological signal, the FUSION pipeline incorporates a Wasserstein GAN module. Across four datasets spanning 10x Visium, ST, Slide‐seq, and Stereo‐seq, FUSION improves cell‐type deconvolution, sharpens domain boundaries, and reconstructs consistent multi‐slice architectures, outperforming existing tools while scaling to larger studies.