Multi-modal composition of physiological signals to delineate candidate cell types in-vivo

Precise identification of in-vivo cell types using electrophysiological signals would allow for the simultaneous monitoring of different cell populations during behavior. Current approaches typically classify waveforms into broad- and narrow-spiking. However, these methods rarely augment analysis of waveforms with other physiological properties (e.g., inter-spike intervals) or validate against “ground truth” obtained through optotagging. We need a principled and validated approach to combine multiple physiological properties to delineate cell types in-vivo in an unsupervised manner.

Our approach, “PhysMAP”, inspired by advances in multi-omics, uses a weighted k-nearest neighbors (WNN) approach to combine physiological properties (e.g., waveform shape, inter-spike interval [ISI], and peri-stimulus time histogram [PSTH]) in order to generate shared latent structure as a high-dimensional graph. We validated PhysMAP using optotagged electrophysiological datasets from mice. In a mouse juxtacellular S1 dataset, cell types were clearly separated by waveform and more modestly separated in ISI and PSTH. We next tested if a multi-modal representation better separated types from one another using two approaches: First, we measured classification accuracy of cell types with a multi-modal representation. We found that—across the board—balanced accuracy was higher for the multi-modal compared to the features of the waveform in isolation. Results were similar for a second in-vivo mouse extracellular dataset from auditory cortex. Second, we performed Louvain clustering on UMAP graphs for either individual modalities or the multi-modal representation and assessed the similarity between the clusters and optotagged cell classes (using the modified adjusted rand index [MARI]). Across Louvain resolutions, MARI was higher for the combined vs. individual modalities.

Thus, PhysMAP provides a principled way for unsupervised identification of cell types from electrophysiology. Long term, multi-modal composition is a path towards a physiological database to “lookup” cell types in-vivo.