CrossBranch: cross-domain cell-type deconvolution with dual-branch representation learning

Abstract Accurate estimation of cell-type composition from mixed omics data is essential for understanding tissue heterogeneity and disease mechanisms. However, existing deconvolution methods are often affected by discrepancies between reference single-cell data and target bulk, proteomic, or spatial omics measurements. This study aims to develop a robust and biologically informed framework for cross-domain cell-type deconvolution. We present CrossBranch, a dual-branch representation learning framework that integrates gene-level and pathway-level information. CrossBranch generates labeled simulated mixtures from single-cell references and jointly encodes simulated and target data through a gene-expression branch and a pathway-informed branch. A prediction head is trained using simulated mixtures with known cell-type proportions, while latent-space alignment reduces distribution discrepancies between simulated and target data. For spatial transcriptomics data, a neighboring-spot-based spatial consistency loss is further incorporated. Across bulk RNA-seq, proteomics, and spatial transcriptomics benchmarks, CrossBranch consistently achieves competitive deconvolution performance compared with existing statistical and deep learning methods. Ablation analyses confirm the contributions of pathway-level representation, cross-domain alignment, and spatial neighborhood modeling. Applications to prostate, colorectal, and pancreatic cancers further demonstrate that CrossBranch can identify tumor-associated cellular changes, survival-associated cell-type patterns, malignant epithelial localization, fibroblast–endothelial co-localization, and compartment-specific spatial organization in tumor microenvironments. CrossBranch provides a unified cross-domain deconvolution framework that improves cell-type composition inference across diverse omics modalities and supports biologically meaningful interpretation of disease microenvironments.