Geometric-aware Deep Learning for Deciphering Tissue Structure from Spatially Resolved Transcriptomics

SpatialGEO is a geometric-aware deep learning framework designed to dissect complex tissue structures from Spatially Resolved Transcriptomics (SRT) data across diverse platforms and resolutions, ranging from spot-level to single-cell and sub-cellular scales. The framework initially employs a dual-encoder architecture to extract embeddings, utilizing an Autoencoder (AE) to capture intrinsic gene expression features and an Enhanced Graph Autoencoder (EGAE) to reinforce spatial graph representations by simultaneously reconstructing gene attributes and spatial adjacency. To further refine these representations, SpatialGEO incorporates a geometric-aware latent embedding generation module that dynamically fuses cross-modal potentials. By leveraging geometric graph learning to characterize continuous spatial relationships between spots, this module effectively overcomes the limitations of binary adjacency graphs in modeling spatial continuity. The entire framework is optimized via a triplet self-supervised strategy, which unifies the learning objectives of the AE, EGAE, and the fusion module. This unified paradigm enhances the alignment between modalities and promotes efficient information integration, yielding robust embeddings that support a wide range of downstream analysis tasks.