Multi-Scale Self-Supervised Pretraining on Neuronal Morphology Transfers Across Laboratories and Imaging Modalities

Self-supervised pretraining has produced representations of neural activity that transfer across sessions, animals, and recording modalities. Whether the same holds for neuronal morphology, the structure on which cell types are defined, remains unclear, in part because existing morphology encoders sub-sample each reconstruction to a fixed node budget and encode it as a flat graph, discarding the hierarchy of nodes, branches, and subtrees that distinguishes one cell type from another. Here we leverage a corpus of brain-wide dendritic reconstructions registered to the Allen Mouse Brain Common Coordinate Framework. We train TopoDINO, a self-supervised model that represents each reconstruction at all three scales simultaneously, with a learned codebook per scale trained by self-distillation and no node sub-sampled. TopoDINO outperforms prior self-supervised morphology encoders pretrained on the same corpus on every held-out task. We find that the frozen representation transfers under two distribution shifts absent from pretraining: from dendrites to axon-inclusive morphologies, where it organizes neurons by anatomical division and projection target, and across laboratories, tracing protocols, and from light to electron microscopy, where it separates excitatory from inhibitory neurons without fine-tuning. Retaining the full hierarchy of a reconstruction, rather than reducing it to a tractable graph, is what makes this transfer possible, and it places morphology alongside neural activity as a modality whose pretrained representations can be judged by transfer rather than by in-distribution accuracy. The electron-microscopy volume in which the representation transfers also contains co-registered neural activity, raising the possibility of relating a neuron's morphological representation directly to its recorded function.