Individual variability of neural computations in the primate retina

Abstract

Variation in the neural code between individuals contributes to making each person unique. Using ~100 neural population recordings from major ganglion cell types in the macaque retina, we develop an interpretable computational representation of individual variability using machine learning. This representation preserves invariances, such as asymmetries between ON and OFF cells, while capturing individual variation and covariation in properties such as nonlinearity, temporal dynamics, and spatial receptive field size. The similarity of these properties across cell types was dependent on the similarity of their synaptic connections. Surprisingly, male retinas exhibited higher firing rates and faster temporal integration than female retinas. By exploiting data from previously recorded macaque retinas, a new macaque retina (and crucially, a human retina) could be efficiently characterized. Simulations indicated that combining a vast dataset of healthy macaque recordings with behavioral feedback could be used to identify the neural code and improve retinal implants for treating blindness.

Removing Inter-Experimental Variability from Functional Data in Systems Neuroscience

“…One solution to this could be to learn a shared embedding space [see 51], from which domain effects are removed, but distinct encoders f i and decoders g i for different trials i. In another setting, where different stimuli are presented between experiments, one might resort to an approach like Shah et al [3]. Nevertheless, we do acknowledge that the data in our applications consists of ex vivo retinal recordings which have little to no attentional effects or task-dependent noise correlations like they would be present in in vivo cortical data.…”

“…Other studies have suggested models of neural function that integrate data across experiments. Shah et al [3] build encoding models to predict the responses of retinal ganglion cells across different experiments [see also 13] and compare it to covariates such as the gender of an animal. Sorochynsky et al [14] propose a way to measure noise correlations in each recording and integrate those into models of neural populations of a specific cell type.…”

“…Both the advance of large-scale data acquisition techniques in experimental neuroscience as well as the development of more efficient and powerful data analysis methods allow collecting and analyzing datasets of increasing size; and hence the discovery of more subtle variations in neural function between cell types [e.g. [1][2][3][4]. However, as data acquisition is often an incremental process, it has become common practice to pool data from multiple experiments.…”

“…adapting the algorithm to work both on the training as well as some (usually unseen) test domain [11]. In single-cell genomics, a number of different studies have proposed methods for removing inter-experimental variability (see Section 2), but related works in systems neuroscience are lacking, despite the recognized need for such approaches [3,5]. Here, we contribute to closing this gap as follows:…”

“…sample handling resulting in small differences in tissue quality, or temperature fluctuations affecting the rates of biochemical processes) but potentially also unforeseen biological ones (e.g. subtle genetic variations) [3,5,6]. Such variability due to external factors, here referred to as inter-experimental variability, can confound and obscure the biological signals of interest.…”