Semi-Independent Transcriptomic, Morphological, Connectivity Dimensions of the Mouse Brain Revealed by CoT-Mining 258 Million Multimodal Associations

We address a fundamental question in brain cell typing: to what extent do transcriptomic, morphological, and connectivity features of neurons correspond to one another, and are they redundant or largely independent? To this end, we constructed NeuroXiv2, a large unified multimodal mouse-brain neuron database, integrating neuronal morphology, connectivity, and transcriptomic profiles across 1,385 hierarchical brain regions, 182,483 reconstructed neurons, and 5.24 million cells with expression data for 1,122 genes. We next built the NeuroXiv2 knowledge graph, encoding 258.4 million cross-modal associations among nodes spanning brain regions, neurons, microenvironmental subregions, and transcriptomic cell types. We then developed AI-Powered Open Mining with Chain-of-Thought (AIPOM-CoT) reasoning, an agent that integrates multimodal information and outperforms single-modality analyses in cell-type separation. Using this approach, we achieved a brain-wide, single-neuron resolution quantification demonstrating that neuronal morphology, brain connectivity, and transcriptomic profiles are neither completely redundant or fully independent. Indeed they form three semi-independent dimensions. Further, we showed that combining any two substantially improves predictive power, however none is fully deterministic. We also applied AIPOM-CoT to identify homogeneous Car3-positive neuronal subclasses jointly defined by anatomy, connectivity, and molecular signatures. Overall, the conceptual advances and methodological innovations presented here demonstrate the value of the NeuroXiv2 database, the power of AI-driven analysis, and the broader need for scalable, integrative cross-modal discovery in neuroscience.