Analyzing transcriptomic data with topology guided neural geometric diffusion

We present a topology- and geometry-aware framework for learning on dynamic cellular graphs from scRNA-seq and spatial transcriptomics (TANGO). Unlike conventional approaches that treat neighborhood graphs as fixed, TANGO explicitly models local topological rewiring during development. Directional geometric diffusion incorporates modality-specific developmental signals, while the optimal transport-based distance quantifies changes in local neighborhood structure across developmental stages. This provides a complementary, topology-native measure of developmental progression alongside RNA velocity and trajectory inference. Experiments across molecular and spatial datasets demonstrate the potential of local topology as an informative signal for characterizing cellular differentiation and tissue organization.