DuaST: an integrated deep learning framework for spatial transcriptomics with cross-branch interaction.

Spatial transcriptomics (ST) enables the joint characterization of gene expression and spatial information, indicating the need for generalizable computational methods to exploit these data. However, integrating spatial and non-spatial information remains a challenge. In this study, we propose DuaST, an integrated dual-branch learning framework for ST. Specifically, DuaST designs a spatially aware branch and a non-spatial branch to separately model neighborhood dependencies and topology-agnostic features. To integrate these complementary representations, DuaST employs a synergistic combination of local-global contrastive learning, adversarial alignment, and attention-based fusion. These mechanisms reinforce cross-branch interactions and enable the learning of biologically meaningful representations. Beyond pattern modeling, DuaST identifies spatially variable genes (SVGs) by reconstructing gene expression with a learnable weight matrix that integrates both spatial and non-spatial dependencies. Furthermore, DuaST extends seamlessly from single- to multi-omics analyses, offering a unified framework that leverages supplementary omics to enhance biological insight. The experimental results indicate that DuaST achieves superior performance in several tasks, such as spatial domain identification, SVGs detection, and multi-omics integration. Ablation studies further demonstrate the overall effectiveness of the model design in capturing spatial and non-spatial representations.