Subunit Nonlinearities In Primate Retinal Ganglion Cells At Single Cone Resolution

Abstract

Purpose: : Our goal is to understand retinal ganglion cell (RGC) receptive fields (RFs) at the resolution of individual cone photoreceptors. A fundamental property of RGCs is the presence of nonlinear subunits in the RF that reflect the converging rectified input of bipolar cells. We used measured RF subunits at the resolution of single cones, and generated a model of RGC light response that captures deviations from linear spatial summation introduced by subunits.

Methods: : Recordings were performed from RGCs in the isolated peripheral primate retina using a 512-electrode recording system. ON and OFF midget and parasol RGCs were identified based on their light response properties, density, and mosaic organization of RFs. Fine-grained white noise visual stimuli and reverse correlation analysis were used to measure the linear RFs of RGCs at the resolution of single cones. STC analysis was used to identify and characterize nonlinear spatial summation of cone inputs within the RFs.

Results: : The eigenvectors obtained from STC analysis displayed spatial structure indicative of RF subunits. Parasol (midget) cells exhibited 15-25 (8-14) significant eigenvectors, in approximate agreement with the number of converging bipolar cells. These eigenvectors were used in a multi-dimensional generalization of the linear-nonlinear-Poisson (LNP) model of light response to predict the responses of RGCs to contrast-reversing gratings. The spatial filter given by each eigenvector, and the filter given by the spike-triggered average stimulus, was convolved with the stimulus and passed through a static nonlinearity fitted to data. The output of these filters were weighted and summed, providing the input to a Poisson spike generator. This model, derived from the RGC response to white noise stimuli, predicted that parasol cells exhibit a larger second harmonic response to contrast-reversing gratings than midget cells, in agreement with experimental data.

Conclusions: : A generalization of the LNP model predicts the nonlinear spatial summation observed in midget and parasol RGCs at single-cell resolution. This model captures both the linear and nonlinear computations performed by the retinal circuitry to generate RGC RFs, and suggests that a complete circuit map describing signal flow from cones to bipolar cells to RGCs is within reach.

Keywords: ganglion cells • computational modeling • photoreceptors