Discovery of many new cell types in the retina using real-time imaging

In a paper entitled The functional diversity of retinal ganglion cells in the mouse, the authors provide a new and startling classification of retinal ganglion cells (RGCs) by recording their real-time activity in response to visual input, revealing many new types of cells. It is generally thought that each type of retinal ganglion cell provides a complete and separate representation of visual information from the entire retina, each type making up a sort of “channel” of visual information. For example, some channels might convey information about edges while others convey information about one of the primary colors we perceive (red, green, blue) ­. To thoroughly investigate these cells, the authors built upon decades of classification work and overcame experimental limitations involved in recording from live cells. The result of their work is an increase in the known number of output channels from the retina to the brain --where each channel carries information from the entire visual field--from fifteen to more than thirty.

Newly discovered cells include those that have direction and motion sensitivity, meaning that they activate in response to objects moving in particular directions. Beyond discovering these new types of output channels, the sheer number of channels identified may represent a pivot in the history of cell type classification: having as many as thirty separate channels suggests that the animal is processing some features of its environment very early on in visual processing, meaning perhaps that some of these channels trigger specific behaviors. Another possibility is that some of these cells types are evolutionary leftovers and, as was previously thought, only a few truly contribute to the bulk of mammalian visual perception.

Up until recently, RGC classification methods were hotly contested, mainly due to a lack of data, experimental challenges, and our relatively small knowledge of how various retinal channels are processed further downstream. As a result there was no serious estimate of how many cell types there are in the mammalian retina. RGCs, which perform the last layer of visual processing in the retina, are now classified by their shape, connectivity pattern, genetic markers, and retinal coverage mosaic. In addition, RGC’s vary substantially in their patterns of neural activity: some RGC cell-types are very sensitive to visual features like edges, while others prefer moving bars. (This is very similar to what has been observed in cortex in the seminal studies by Hubel and Wiesel.)

In this study, the authors utilized a common optical imaging technique called two-photon calcium imaging in a substantial patch of ventral mouse retina to differentiate retinal ganglion cell types based on their varying responses to visual stimuli, such as receptive field size and response to moving bars. They then combined this information from these responses with a classification of the unique set of proteins expressed in those cells, which they obtained by genetic labeling and antibody staining against RGC cell surface proteins. To determine a cell’s morphology, they filled the cells with dyes. They consolidated all of these separate types of classification data by assigning each of the various features a number and creating a vector of these values for each cell. The authors then used an unsupervised feature clustering algorithm involving PCA and a mixture of gaussians model to classify the cell types while avoiding human bias. The result was the classification of 32 types of retinal ganglion cells, at least 10 more than were previously known to exist.

The novelty in this work lies not only in the multi-factorial classification approach the authors employed, but also the clustering methods they used to reliably parse apart the various cell types. Furthermore, this work emphasizes the fact that the way the retina processes information is likely very dissimilar to the way a digital camera does, and more similar to the way a programmed artificial neural net does. The implications of that realization are that there may well be some specific cell types that participate in reflex behavior in the mouse. This could be further studied by studying the responses of these cells to visual stimuli that more accurately represent behaviorally relevant visual stimuli in the mammalian environment.