Periodic Reporting for period 1 - DEEPRETINA (A perturbative approach to model retinal processing of natural scenes)
Berichtszeitraum: 2022-10-01 bis 2025-03-31
A major goal of sensory neuroscience is to understand how sensory neurons process natural scenes. Models built from the responses of sensory neurons to simple stimuli do not generalize to predict how complex, natural scene are processed. Even as early as in the retina, this issue is not solved. Deep network models have been proposed to predict the responses of visual neurons to natural stimuli. However, they are still far from being a realistic model of the visual system. First, the sensitivity to perturbations of the stimulus can thus be very different for a deep network model and for our visual system. Second, it is not clear how the model components can be related to actual mechanisms in the brain
Our purpose is to understand how the retina processes natural scenes. We will follow an interdisciplinary approach where we will build realistic deep network models of retinal processing and test them in experiments. We will develop deep network models that can predict ganglion cell responses to natural stimuli, and map the components of these models to specific cell types in the retinal network.
Our project is original because it will use two novel methods, that will be key to achieve our goal. The first one is a novel approach to characterize retinal function, where we will probe the selectivity of the retina to perturbations of natural stimuli. The second one is a novel tool based on 2 photon holographic stimulation to decompose the retinal circuit. They are tailored to address the specific issues of deep networks.
Our purpose is to understand how the retina processes natural scenes. We will follow an interdisciplinary approach where we will build realistic deep network models of retinal processing and test them in experiments. We will develop deep network models that can predict ganglion cell responses to natural stimuli, and map the components of these models to specific cell types in the retinal network.
Our project is original because it will use two novel methods, that will be key to achieve our goal. The first one is a novel approach to characterize retinal function, where we will probe the selectivity of the retina to perturbations of natural stimuli. The second one is a novel tool based on 2 photon holographic stimulation to decompose the retinal circuit. They are tailored to address the specific issues of deep networks.
One of our aims (aim 1) is to apply a novel, perturbative method to understand how ganglion cells process natural images. We have developed this method and show that it led to novel insights on how ganglion cells process natural images, highlighting that the same cell can sometimes be sensitive to a light increase and sometimes to a light decrease. We found that deep networks trained on retinal data could explain these results, and led to new predictions that were tested and verified. We further explored how eye optics impacted this processing. This work has been published (Goldin et al, 2022; Goethals et al, 2024).
Another aim (aim 3) was to understand better how ganglion cells encode stimuli with complex temporal dynamics, and the role of amacrine interneurons in this non-linear processing. We have studied a well-known, specific phenomenon whose mechanism was still unclear: the so-called Omitted Stimulus Response (OSR), i.e. the fact that some ganglion cells respond specifically to omitted flashes in a periodic sequence. We have shown that these specific responses allow ganglion cells to encode for how surprising the stimulus is (Destopovic et al, 2024). Deep networks have been proposed to explain this phenomenon, but without an experimental confirmation so far. We have shown that glycinergic amacrine cells are necessary for the OSR. This goes against the proposed models so far, and we have proposed a new one that explain our experimental results, produces new experimental predictions that we verified. This work has been published (Ebert et al, 2024).
We have also tested a strategy to express optogenetic proteins in a specific type of amacrine cells, which was part of the aim 1. We probed the functional impact of this strategy and demonstrated it has a potential clinical impact for vision restoration (Khabou et al, 2023).
In a collaboration we have also tested how ganglion cell types are affected by NO, a peptide released by specific amacrine cell types (Gonschorek et al, 2024). This is in line with the strategy to estimate the impact of several amacrine cell types on retinal processing in aim 3.
Another aim (aim 3) was to understand better how ganglion cells encode stimuli with complex temporal dynamics, and the role of amacrine interneurons in this non-linear processing. We have studied a well-known, specific phenomenon whose mechanism was still unclear: the so-called Omitted Stimulus Response (OSR), i.e. the fact that some ganglion cells respond specifically to omitted flashes in a periodic sequence. We have shown that these specific responses allow ganglion cells to encode for how surprising the stimulus is (Destopovic et al, 2024). Deep networks have been proposed to explain this phenomenon, but without an experimental confirmation so far. We have shown that glycinergic amacrine cells are necessary for the OSR. This goes against the proposed models so far, and we have proposed a new one that explain our experimental results, produces new experimental predictions that we verified. This work has been published (Ebert et al, 2024).
We have also tested a strategy to express optogenetic proteins in a specific type of amacrine cells, which was part of the aim 1. We probed the functional impact of this strategy and demonstrated it has a potential clinical impact for vision restoration (Khabou et al, 2023).
In a collaboration we have also tested how ganglion cell types are affected by NO, a peptide released by specific amacrine cell types (Gonschorek et al, 2024). This is in line with the strategy to estimate the impact of several amacrine cell types on retinal processing in aim 3.
We have further explored how the retina processes after our first publication (Goldin et al 2022), to understand how the optics of the eye impact how a natural image is processed by the retina. Eye growth is regulated by the visual input. Many studies suggest that the retina can detect if a visual image is focused in front or behind the back of the eye, and modulate eye growth to bring it back to focus. How can the retina distinguish between these two types of defocus? We simulated how eye optics transform natural images and recorded how the isolated retina responds to different types of simulated defocus. We found that some ganglion cell types could distinguish between an image focussed in front or behind the retina, by estimating spatial contrast. Aberrations in the eye optics made spatial contrast, but not luminance, a reliable cue to distinguish these two types of defocus. Our results suggest a mechanism for how the retina can estimate the sign of defocus. This work has been published as a pre-print (Goethals et al, 2024).
This result has a potential industrial and clinical impact in the field of myopia mitigation. Myopia is an excess of eye growth. Current projectinos predict that by 2050, half of the worldwide population will be myopic. For strong myopia (projected to be concerning almost a billion patient), this will lead to severe consequences (e.g. retinal detachment). One therapeutical strategy to avoid this is to wear new types of glasses that can slow down the growth of the eye, by transforming the image received by the retina to simulate an image focused in front, such that the retina is “tricked” into slowing down eye growth. Several glass designs have been commercialized with moderate efficiency, but it is unclear why they even partially work. Our paper provides a explanatory framework and we can predict which glass design should perform best. We also explain why near sight vision is a factor increasing myopia prevalence.
A patent has been filed related to these results (name: “Using local spatial contrast in retinal images to predict myopia control lenses efficacy”).
This result has a potential industrial and clinical impact in the field of myopia mitigation. Myopia is an excess of eye growth. Current projectinos predict that by 2050, half of the worldwide population will be myopic. For strong myopia (projected to be concerning almost a billion patient), this will lead to severe consequences (e.g. retinal detachment). One therapeutical strategy to avoid this is to wear new types of glasses that can slow down the growth of the eye, by transforming the image received by the retina to simulate an image focused in front, such that the retina is “tricked” into slowing down eye growth. Several glass designs have been commercialized with moderate efficiency, but it is unclear why they even partially work. Our paper provides a explanatory framework and we can predict which glass design should perform best. We also explain why near sight vision is a factor increasing myopia prevalence.
A patent has been filed related to these results (name: “Using local spatial contrast in retinal images to predict myopia control lenses efficacy”).