bioRxiv2025
Tracking single neurons across days in high-density extracellular recordings is essential for establishing neural mechanisms of learning, memory, and post-injury recovery. However, in weeks-long recordings, identifying cross-day matches among thousands of units is confounded by changes in spike waveforms, unit turnover, and representational drift in neural responses. We introduce DANT (Density-based Across-day Neuron Tracking), an unsupervised framework that jointly estimates probe motion and neuron identity by alternating between density-based clustering in feature space and probe-motion correction inferred from provisional matches. Estimated drift is used to re-register spike waveforms across sessions, after which clustering is recomputed; this iterative loop continues until the set of matches stabilizes. In parallel, DANT learns a decision boundary from match and non-match assignments derived from the clustering results, enabling it to reject low-similarity candidate pairs. Applied to weeks-long Neuropixels recordings from the cortex and striatum in freely moving rats during stable behavior and task switching, DANT substantially increases match yield and reduces false negatives while maintaining a low false-positive rate relative to existing approaches. Together, these results indicate that DANT provides a general, unsupervised solution for longitudinal tracking in chronic, high-density recordings. The bigger pictureModern high-density probes such as Neuropixels record hundreds of units per session. When chronically implanted, many of these units can be stably recorded over extended periods, opening the door to tracking how large neural populations change over days to weeks, a question central to learning, memory, internal states, and recovery from brain injury. Yet, small shifts in probe position, changes in spike waveforms, and drifts in firing properties mean that linking units across sessions is challenging. DANT (Density-based Across-day Neuron Tracking) addresses this problem by providing an unsupervised framework that combines density-based clustering and iterative motion correction to identify cross-day matches with minimal subjective decisions. Applied to chronic Neuropixels recordings from freely moving, task-performing rats, DANT recovers longer and more complete chains of putative same neurons than existing approaches, while keeping the false-positive rate low. By making cross-day tracking more sensitive, scalable, and reproducible, DANT enables a wider range of neuroscience studies to examine the longitudinal behavior of single neurons and neural populations.