Neural data science for studying visual computations in neuronal populations at single-cell resolution

This thesis presents applications of neural data analysis to investigate early visual processing and assess the efficacy of vision restoration approaches in neuronal populations at single-cell resolution. Recent advancements in electrophysiological recording techniques have enabled the recording of extracellular activity from hundreds of neurons in parallel. This work examines the concerted computations performed by populations of visual neurons by characterizing their neural response functions utilizing various visual stimuli and analysis methods.
The first part of this thesis addresses the challenge of characterizing the neural response functions of functionally specialized visual neurons. Functional specialization is a fundamental principle in visual processing. While neurons in early visual processing stages respond to unstructured contrast changes, neurons in later stages encode more complex, structured stimuli. A systematic, unbiased approach for estimating the neural response function of a sensory neuron is provided by white noise (WN) analysis, where the neural response is correlated with an unstructured, uncorrelated WN stimulus to estimate the linear white noise receptive field (WN-RF). However, this approach is ineffective for functionally specialized sensory neurons. To overcome this limitation, we developed the reverse correlation against stimulus elements (RCASE) method, which utilizes structured stimuli characterized by uncorrelated stimulus parameters. Here, the neural response is correlated with a nonlinear reparameterization of the stimulus. Using this method, we efficiently mapped nonlinear receptive fields (RFs) of nearly all recorded retinal ganglion cells (RGCs) in the mouse retina by utilizing a stimulus consisting of randomly parameterized moving objects that elicited robust responses across the recorded population.
We applied this approach one synapse downstream to the nucleus of the optic tract (NOT), a subcortical structure receiving direct retinal input. We found that functional specialization was more pronounced in the NOT than in the retina, rendering WN inefficient in estimating WN-RFs for a majority of the recorded population. However, utilizing RCASE allowed us to estimate localized nonlinear RFs in the majority of NOT neurons. The identified RFs were located above and along the horizontal meridian of the visual space. The NOT is necessary for the horizontal optokinetic reflex (HOKR), a visual behavior that stabilizes the retinal image during large horizontal visual shifts. Psychophysical experiments demonstrated that the HOKR was triggered in the same visual regions as predicted by the locations of the RFs, thereby establishing a link between neural activity and behavior. Additionally, unsupervised clustering of functional responses identified distinct functional cell types within the mouse NOT, providing the first comprehensive functional characterization of this area.
We further present the results of a large-scale collaborative study on an optogenetic vision restoration approach aimed at resensitizing non-functional, dormant cone photoreceptors in the human retina. We developed a functional assay and a corresponding data processing pipeline for large-scale functional screening of gene therapy vectors. Through this approach, we identified a candidate vector that effectively restored light responses in previously non-light-sensitive human retinal explants. The functional properties of RGCs in treated human retinal explants were comparable to those in intrinsically photosensitive human retinal explants. We identified five functional cell types in the treated condition that matched those previously characterized in intrinsically photosensitive human retinal explants. These findings led to the identification of a promising gene therapy vector for restoring high-resolution vision in blind patients who retain dormant cone photoreceptors.
Finally, we performed the first functional characterization of human foveal RGCs to date. Through unsupervised clustering of light-evoked activity of human foveal RGCs, we identified five functional cell types, four of which matched to midget and parasol cells, the most abundant RGC types in the human retina. The same methodology enabled the functional identification of midget and parasol cells in the periphery of the primate retina. These findings contributed to a study investigating cell-type-specific intraretinal RGC axonal propagation speeds at different retinal locations in the human retina.
In summary, this thesis presents several applications for studying early visual processing by utilizing existing methods and developing novel approaches in neural data science, both for basic research and translational applications in vision restoration.