Rapid ganglion and amacrine cell type classification using temporal cross-correlation in the macaque retina

Abstract

Purpose : Recent advances in multi-electrode array (MEA) recordings and analysis have enabled studying the functional properties of over two dozen novel polyaxonal amacrine cell (PAC) and retinal ganglion cell (RGC) types in the macaque retina. While several methods exist for classifying cell types within a recording, robust real-time classification during an experiment remains a challenge due to variability between retinas and the long visual stimulus duration required. One approach leverages the observation that RGCs of the major types exhibit distinctive temporal spiking correlations within and across types. Here, we demonstrate unique correlation patterns between novel and known types and apply these findings to identify novel cell types during an experiment.

Methods : MEA recordings were performed ex vivo in the peripheral macaque monkey retina. The spike-triggered average (STA) of each cell was calculated from its responses to spatiotemporal noise and used to summarize its spatial, temporal, and chromatic properties. Cell types were identified by clustering these properties. Cross-correlation histograms (CCHs) of spike trains between each cell pair were calculated. An average CCH shape was determined for each cell type pair using neighboring cell pairs.

Results : The form of the CCH for each cell type pair was highly consistent across retinas. Because ON and OFF parasol and midget cells were easily identified using other methods, they could be easily leveraged for classification of novel cell types using correlations alone. The average CCHs for each cell with its parasol and midget cell neighbors was calculated in 17 recordings. Then, for each cell in a held-out recording, the correlation coefficients were computed between its CCHs with parasol and midget cells and the average CCHs. Using this leave-one-out cross-validation over 18 recordings, a support vector machine trained on these data yielded 89% accuracy in identifying 5398 cells of 5 novel types. In some cases, the correlations also revealed likely functional relationships between cell types, such as homotypic gap junction coupling between neighboring putative recursive bistratified RGCs and heterotypic coupling between these cells and A1 amacrine cells.

Conclusions : Correlation analysis permits consistent, rapid classification of RGCs across different retinas and requires relatively little data to classify novel cell types.

This abstract was presented at the 2024 ARVO Annual Meeting, held in Seattle, WA, May 5-9, 2024.