CNS*2026 Halifax

Introduction
Advances in transcriptomic, morphological, and electrophysiological techniques have led to more comprehensive characterization of cortical cell types. However, in the absence of “ground truth” labels, it is difficult to classify single units into neuron types based on extracellular action potentials (EAPs) recorded from high-density electrode arrays. Previously published classifiers rely on extracted features, machine learning (ML) algorithms, and dimensionality reduction approaches that produce a range of EAP type classes (Haynes, 2024; Jia, 2019; Lee, 2021). We extend these approaches by developing an unsupervised ML framework that relies on spatiotemporal patterns of multi-channel waveforms.


Methods
We developed an unsupervised ML framework to analyze spiking events from extracellular recordings. For each event, a spatiotemporal kernel waveform is constructed with a 1.5 ms temporal window that is centered on a spike time, based on waveforms from channels 200 um above and below the channel with the largest peak amplitude. We analyzed thousands of spiking events and applied KMeans clustering on two different representations of the data: 1) extracted spatiotemporal features and 2) vectorized multi-channel waveforms. We applied our framework to extracellular recordings in the auditory cortex of an awake common marmoset (Callithrix jacchus). We compare these results to previously published studies that cluster EAPs from rodent (Jia, 2019).


Results
Both the feature-based and the vectorized data representations produced well-defined separation of clusters of spiking events recorded from marmoset auditory cortex. For the feature based approach, EAP duration and vertical spread across channels were the most distinguishing features across clusters. Clustering based on vectorized multi-channel waveforms resulted in greater variability in feature distributions, cortical depth profiles, and spatiotemporal patterns across cluster groups.


Discussion
Together, these findings suggest the presence of diverse neural subtypes in the marmoset auditory cortex that exhibit distinct extracellular signatures. Our data-driven framework demonstrates that unsupervised machine learning approaches reveal physiologically meaningful variation in EAP waveforms, offering a scalable approach to comparative analysis of cortical organization and function. Moreover, this framework is applied to extracellular recordings from other species to identify distinct waveform features and identify which aspects of  spatiotemporal patterns are conserved across species.


References

Haynes, V. R., Zhou, Y., & Crook, S. M. (2024). Discovering optimal features for neuron-type identification from extracellular recordings. Frontiers in Neuroinformatics, 18, 1303993. https://doi.org/10.3389/fninf.2024.1303993
Jia, X., Siegle, J. H., Bennett, C., Gale, S. D., Denman, D. J., Koch, C., & Olsen, S. R. (2019). High-density extracellular probes reveal dendritic backpropagation and facilitate neuron classification. Journal of Neurophysiology, 121(5), 1831–1847. https://doi.org/10.1152/jn.00680.2018
Lee, E. K., Balasubramanian, H., Tsolias, A., Anakwe, S. U., Medalla, M., Shenoy, K. V., & Chandrasekaran, C. (2021). Non-linear dimensionality reduction on extracellular waveforms reveals cell type diversity in premotor cortex. eLife, 10, e67490. https://doi.org/10.7554/eLife.67490


Acknowledgement
This research is support by the NIDCD (R01DC019278).