Interrogating spinal cord injury pathophysiology from a single-cell perspective

Spinal cord injury (SCI) consists of an immediate mechanical insult followed by a secondary injury cascade involving inflammatory infiltration, excitotoxicity, vascular breakdown, demyelination, and the formation of a fibrotic core bordered by reactive astrocytes. These processes unfold across distinct spatial compartments and rapidly reshape the cellular landscape, but bulk approaches have been unable to resolve which cell types initiate, sustain, or constrain these events. A single-cell perspective is uniquely positioned to address this challenge: by resolving transcriptional and regulatory states at cellular resolution, it can disentangle the heterogeneous populations that orchestrate degeneration and repair and reveal the coordinated logic underlying the evolution of the injured tissue. In this thesis, I establish a single-cell framework for decoding the multicellular organization of SCI through multimodal atlas construction, robust analytical methodology, and cross-species comparison. I first introduce single-nucleus, multiomic, and spatial transcriptomic atlases that comprehensively map the injured spinal cord across diverse biological contexts. These resources reveal two core principles governing injury progression: the timely reestablishment of the tripartite neuroprotective barrier, and the coordinated activation of transcriptomic programs that structure the lesion microenvironment. The atlases further uncover conserved and divergent injury-induced states, identify neuronal subpopulations primed for circuit reorganization, and clarify how interactions among cellular populations determine whether tissue stabilizes or degenerates. Constructing these atlases highlighted key analytical limitations in emerging single-cell modalities. To address them, I developed a benchmark study that establishes the statistical principles governing differential accessibility analysis in scATAC-seq, demonstrating that pseudobulk methods offer the most biologically faithful results. In parallel, I introduced Vespucci, a machine-learning framework for comparative spatial transcriptomics that identifies perturbation-responsive regions and gene programs with high sensitivity while controlling for replicate-level variability. I finally extend this framework across evolution by constructing a cross-species single-nucleus atlas spanning non-regenerating and naturally regenerating vertebrates. These comparisons reveal that regenerative capacity is governed not by anatomical divergence, but by the timing and coordination with which conserved molecular programs such as tripartite barrier restoration and intrinsic neuronal growth pathways are deployed. In non-regenerating mammals, including non-human primates, these programs are activated more slowly and incompletely. Targeted gene delivery that accelerates these conserved responses partially reinstates regenerative logic and improves repair in non-regenerating models. Looking ahead, the conceptual and technical foundat