Generative cerebral vasculature visualization using spatial transcriptomic data

The brain is sustained by an intricate vascular network that provides a continuous supply of nutrients and oxygen essential for its function. Understanding the structural variability and region-specific function of this neurovascular system is essential for interpreting neurological alterations in disease models. Leveraging the spatial mRNA-guided generative model Tera-MIND (simulating Tera-scale Mouse braINs using a patch-based and boundary-aware Diffusion model), we propose to predict the spatial organization of brain vasculature by the co-expression patterns of Cldn5 and Acta2 genes and their learned attention map at cell-level resolution. These findings demonstrate the capability of generative AI models trained on single-cell spatial transcriptomic data to reconstruct biologically meaningful higher-order structures. They highlight the potential of such models as in silico systems - GenAI-based simulations that generate realistic representations of biological architecture from spatial molecular data - for high-throughput exploration of vascular function and its dysregulation in neurological disorders. Taken together, our approach repurposes existing spatial transcriptomics datasets to derive new spatial insights into vascular organization, without the need for additional tissue processing. The brain’s vascular network is complex and functionally essential. Accurate reconstruction of this network at single-cell transcriptomic resolution is key to understanding neurovascular disorders. This study applies spatial transcriptomic and histological data to generate representations of brain vasculature using Tera-MIND, providing a framework that offers a scalable approach for understanding vascular organization.