Comprehensive wiring diagrams from electron microscopy (EM) are a powerful tool to understand the inner workings of the brain. The retina is an easily accessible part of the brain that performs complex visual computations. Its thin, layered structure offers a unique opportunity to decipher neural cell types and map their connectivity. A major obstacle has been the limited size of existing retinal EM datasets, which could not resolve rare cell types and neurons with large dendritic arbors. Here, we describe Eyewire II, a large-scale EM dataset covering nearly 1 mm2 of the adult mouse retina - roughly 10-100 times larger than previous retinal EM volumes. Human proofreading of an automated reconstruction has so far yielded more than 8,000 bipolar cells, 13,000 amacrine cells, and 4,000 retinal ganglion cells. Automated detection is complete for synaptic ribbons and in progress for conventional synapses. Prior to EM imaging, visual responses to diverse stimuli - including natural movies - were recorded in a subset of neurons using two-photon Ca2+ imaging, enabling direct alignment of morphological and functional cell type identity. To enable high-throughput, automatic cell typing, we devised a human-in-the-loop approach that combines deep learning with human expert annotations. As a proof-of-principle, we show that morphological features of reconstructed bipolar cells are sufficient to recover all 15 known bipolar cell types with regular, non-overlapping mosaics. Together, these data and tools establish Eyewire II as a shared resource for the field of retina research. Already now, more than 30 laboratories worldwide are contributing proofreading, expert annotations, and software tools, advancing Eyewire II towards a complete cell type catalog and synaptic wiring diagram of a mammalian retina.

The superior colliculus integrates retinal input to drive rapid, adaptive visual behavior, yet how the functional diversity of retinal ganglion cell types is represented in superior colliculus remains poorly understood. Using chronic two-photon calcium imaging of retinal ganglion cell axonal boutons in awake mice, we recorded over 200,000 boutons across superficial superior colliculus layers - a scale that enabled systematic comparison with large-scale ex vivo retinal datasets. This revealed that the superior colliculus receives a near-complete sampling of retinal ganglion cell functional diversity. Functionally distinct response types were organized in systematic laminar gradients: not only response properties such as direction selectivity and contrast suppression, but retinal response types themselves varied systematically with depth. To probe how this organized input encodes natural scenes, we trained a "digital twin" deep network model on natural movie responses and validated its generalization to parametric stimuli, including cell type identification. Leveraging this model to generate predicted responses to looming stimuli, we identify a discrete subset of retinal response types tuned for collision detection at low angular thresholds - a specialization embedded within a broader, non-specialized retinal population. The digital twin is made publicly available as a community resource. Together, these findings provide a comprehensive functional map of retinal drive to the superior colliculus and an in silico platform for linking retinal cell types to behaviorally relevant superior colliculus computations.

We present the ALL-GCL dataset, a large-scale resource of functional two-photon Ca2+-imaging recordings with rich metadata information from more than 80,000 cells in the ganglion cell layer (GCL) of the ex vivo mouse retina. Collected over nine years across more than 155 experimental sessions, the dataset provides recordings of light-evoked responses to various stimuli, ranging from a shared set of core stimuli to natural movies. To enable cell-type-specific analyses, cells are probabilistically assigned to 46 previously characterised functional groups, including retinal ganglion cells and displaced amacrine cells. Further, we assessed the influence of experimental and biological factors on the functional responses. Classifier-based analyses identified measurable signatures associated with acquisition conditions and recording sessions, providing a quantitative characterisation of dataset structure and potential sources of batch effects. The ALL-GCL dataset offers a comprehensive and standardised reference for studying retinal computation at scale. It supports population-level analyses, computational modelling, and the development of machine learning approaches for biological time-series data. Future releases will expand the dataset with additional mouse lines and light stimuli, creating a growing resource for the vision science community.

Over the last two decades, models of visual processing in the retina have increased in scale and predictive power, driven by advances in recording technologies and deep learning. Current models span a spectrum from functional, high-performance architectures to detailed biophysical simulations. Here, we survey the state-of-the-art across this spectrum, with a focus on the key developments needed to leverage these models to improve our understanding of retinal processing. We argue for a collective effort to build a scientific ecosystem around a core of shared datasets, modular composable models, and standardised benchmarks. We propose that such an infrastructure would accelerate model-driven discovery and enable the community to focus its efforts on open scientific questions such as the role of neural variability, adaptation, and cell-type diversity in retinal coding.

ABSTRACT Across sensory systems, neurons often encode stimulus changes of opposite sign as On and Off signals, a fundamental strategy for representing deviations from baseline. In vertebrate vision, this computation is classically attributed to an early, hardwired split in the retina, with segregated pathways subsequently propagated through downstream circuits. Here, we challenge this view. Combining comparative transcriptomics, pharmacology, genetics, in vivo imaging and electrophysiology in zebrafish, with validation in mouse retina, we show that On-Off coding is a latent and intrinsic property of visual circuits. Bipolar cells frequently co-express receptor systems of opposing polarity and can generate mixed responses that are normally suppressed by inhibitory circuitry. Disrupting inhibition or selectively blocking receptor pathways unmasks robust On-Off signalling. In addition, mixed On-Off responses emerge naturally as effective light input increases, including during development and at higher light levels. Across processing stages, including retinal ganglion cells and central neurons, inhibition repeatedly enforces apparent polarity segregation. Thus, polarity splitting is a distributed, dynamically regulated computation rather than a fixed circuit feature. HIGHLIGHTS Mixed On-Off signals are latent across visual processing stages Inhibition actively sculpts polarity, not just refines it Polarity splitting is a dynamic, distributed and reusable computation Conserved mechanism revealed in zebrafish and mouse

Neural system identification approaches use empirical data to fit the stimulus-response functions of neurons. Augmented by deep neural networks, such models have achieved high predictive performance and allow to perform in silico experiments to test hypotheses. Yet, many of these methods ignore common features in visual systems, such as inhibitory interactions between neurons, which are essential for nonlinear neural computation. Here, we incorporate inhibition as an inductive bias into a deep model for neural prediction and investigate the influence of inhibition on the learned transfer functions. To this end, we employ difference-of-Gaussian (subtraction) and within-channel divisive normalization (division), which have been proposed to relate inhibition to neural processing, in deep networks for predicting visual responses. We observe that incorporating such operations maintains the predictive performance and encourages the learning of biologically plausible kernels reminiscent of neural representation in early vision. Additionally, our in silico experiments demonstrate that implementing either subtractive or divisive operation benefits the learning of surround suppression but not cross-orientation inhibition. Interestingly, while division increases the sparsity of activation and reduces the sparsity of weights, subtraction has the reverse effect.

Vision first evolved in the water, where the spectral content of light informs about viewing distance. However, whether and how aquatic visual systems exploit this "fact of physics" remains unknown. Here, we show that zebrafish use "color" information to suppress responses to the visual background. For this, zebrafish divide their intact ancestral cone complement into two opposing systems: PR1/4 ("red/UV cones") versus PR2/3 ("green/blue cones"). Of these, the achromatic PR1 and PR4, which are retained in mammals, are necessary and sufficient for vision. By contrast, the color-opponent PR2 and PR3, which are lost in mammals, are neither necessary nor sufficient for vision. Instead, they form an "auxiliary" system that spectrally suppresses the "core" drive from PR1 and PR4. Our insights challenge the long-held notion that vertebrate cone diversity primarily serves color vision and further hint at terrestrialization, not nocturnalization, as the leading driver for visual circuit reorganization in mammals.

We propose a standardized naming system for vertebrate visual photoreceptors (i.e., rods and cones) that reflects our current understanding of their evolutionary history. Vertebrate photoreceptors have been studied for well over a century, but a fixed nomenclature for referring to orthologous cell types across diverse species has been lacking. Instead, photoreceptors have been variably - and often confusingly - named according to morphology, presence/absence of ‘rhodopsin,’ spectral sensitivity, chromophore usage, and/or the gene family of the opsin(s) they express. Here, we propose a unified nomenclature for vertebrate rods and cones that aligns with the naming systems of other retinal cell classes and that is based on the photoreceptor’s putative ancestral derivation. This classification is informed by the functional, anatomical, developmental and molecular identities of the neuron as a whole, including the expression of deeply conserved transcription factors required for development. The proposed names will be applicable across all vertebrates and indicative of the widest-possible range of properties, including their postsynaptic wiring, and hence will allude to their common and species-specific roles in vision. Furthermore, the naming system is open-ended to accommodate the future discovery of as-yet unknown photoreceptor types.

The state of our brain shapes what we see, but how early in the visual system does this start? A new study in PLOS Biology shows that brain state-dependent release of histamine modulates the very first stage of vision in the retina.

In the mouse retina, sustained ON alpha (sONα) retinal ganglion cells (RGCs) have different dendritic and receptive field sizes along the nasotemporal axis, with temporal sONα RGCs likely playing a role in visually guided hunting. Thus, we hypothesized that this cell type also exhibits regional adaptations in dendritic signal processing and that these adaptations are advantageous for prey capture. Here, we measured dendritic signals from individual sONα RGCs at different retinal locations. We measured both postsynaptic Ca2+ signals at dendrites and presynaptic glutamate signals from bipolar cells (BCs). We found that temporal sONα RGCs exhibit, in addition to sustained-ON signals with only weak surrounds, signals with strong surround suppression, which were not present in nasal sONα RGCs. This difference was also present in the presynaptic inputs from BCs. Last, using population models in an encoder-decoder paradigm, we showed that these adaptations might be beneficial for detecting crickets in hunting behavior.

Studying the retina plays a crucial role in understanding how the visual world is translated into the brains language. As a stand-alone neural circuit with easily controllable input, the retina provides a unique opportunity to develop a complete and quantitatively precise model of a computational module in the brain. However, decades of data and models remain fragmente across labs and approaches. To address this, we have launched an open-source retina modelling platform on a shared GitHub repository, aiming to provide a unified data and modelling framework across species, recording techniques, stimulus conditions, and use cases. Our initial release consists of a Python package, openretina, a modelling framework based on PyTorch, which we designed for optimal accessibility and extensibility. The package includes different variations on a basic Core + Readout model architecture, easily adaptable dataloaders, integration with modern deep learning libraries, and methods for performing in-silico experiments and analyses on the models. We illustrate the versatility of the package by providing dataloaders and pre-trained models for data from several laboratories and studies across species. With this starter pack in place, openretina can be used within minutes. Through step-by-step examples, we here provide retina researchers of diverse backgrounds a hands-on introduction to modelling, including using models as tools for visualising retinal computations, generating and testing hypotheses, and guiding experimental design. ### Competing Interest Statement The authors have declared no competing interest.

Visual processing starts in the outer retina where photoreceptors transform light into electrochemical signals. These signals are modulated by inhibition from horizontal cells and sent to the inner retina via excitatory bipolar cells. The outer retina is thought to play an important role in contrast invariant coding of visual information, but how the different cell types implement this computation together remains incompletely understood. To understand the role of each cell type, we developed a fully-differentiable biophysical model of a circular patch of mouse outer retina. The model includes 200 cone photoreceptors with a realistic phototransduction cascade and ribbon synapses as well as horizontal and bipolar cells, all with cell-type specific ion channels. Going beyond decades of work constraining biophysical models of neurons only by experimental data, we used a dual approach, constraining some parameters of the model with available measurements and others by a visual task: (1) We fit the parameters of the cone models to whole cell patch-clamp measurements of photocurrents and two-photon glutamate imaging measurements of synaptic release. (2) We then trained the spatiotemporal outer retina model with photoreceptors and the other cell types to perform a visual classification task with varying contrast and luminance levels. We found that our outer retina model could learn to solve the classification task despite contrast and luminance variance in the stimuli. Testing different cell type compositions and connectivity patterns, we found that feedback from horizontal cells did not further improve task performance beyond that of excitatory photoreceptors and bipolar cells. This is surprising given that horizontal cells are positioned to mediate communication across cones and that they add to the model's number of trainable parameters. Finally, we found that our model generalized better to out of distribution contrast levels than a linear classifier. Our work shows how the nonlinearities found in the outer retina can accomplish contrast invariant classification and teases apart the contributions of different cell types.

Retinitis pigmentosa is a hereditary disease-causing progressive degeneration of rod and cone photoreceptors, with no effective therapies. Using rd10 mice, which mirror the human condition, we examined its disease progression. Rods deteriorate by postnatal day (P) 45, followed by cone degeneration, with most photoreceptors lost by P180. Until then, retinal ganglion cells (RGCs) remain light-responsive under photopic conditions, despite extensive outer retinal remodelling. However, it is still unknown if distinct functional RGC types alter their activity or are even lost during disease progression. Here, we asked if and how the response diversity of functional RGC types changes with rd10 disease progression. At P30, we identified all functional wild-type RGC types also in rd10 retinae, suggesting that at this early degenerative stage, the full breadth of retinal output is still present. Remarkably, we found that the fractions of functional types changed throughout progressing degeneration between rd10 and wild-type: responses of RGCs with 'Off'-components ('Off' and 'On-Off' RGCs) were more vulnerable than 'On'-cells, with 'Fast On' types being the most resilient. Notably, direction-selective RGCs appeared to be more vulnerable than orientation-selective RGCs. In summary, we found differences in resilience of response types (from resilient to vulnerable): 'Uncertain' > 'Fast On' > 'Slow On' > 'On-Off' > 'Off'. Taken together, our results suggest that rd10 photoreceptor degeneration has heterogeneous effects on functional RGC types, with distinct sets of types losing their characteristic light responses earlier than others. This differential susceptibility of RGC circuits may be of relevance for future neuroprotective therapeutic strategies. KEY POINTS: Retinitis pigmentosa is a hereditary disease causing progressive degeneration of rod and cone photoreceptors, with no effective therapies; it can be investigated using mutant mouse models, like rd10, that mirror the human condition. Recent studies found that retinal ganglion cells (RGCs) in rd10 remain light-responsive, despite extensive loss of photoreceptors and outer retinal remodelling; however, specific RGC types still may change or even lose their functional response profile during early degeneration. Using two-photon calcium imaging, we assessed if and how the light-evoked activity of RGCs, and, hence, the retinal output to the brain, differs during the disease progression in rd10 compared to wild-type mice. We found differences in the resilience of functional RGC types: generally, 'On'-types were more resilient than 'On-Off' or especially 'Off' types. Our data suggest that interventions may be more effective in the 'On' pathways, which turned out to be more resilient in rd10.

Neuromodulators have major influences on the regulation of neural circuit activity across the nervous system. Nitric oxide (NO) has been shown to be a prominent neuromodulator in many circuits and has been extensively studied in the retina. Here, it has been associated with the regulation of light adaptation, gain control, and gap junctional coupling, but its effect on the retinal output, specifically on the different types of retinal ganglion cells (RGCs), is still poorly understood. In this study, we used two-photon Ca2+imaging and multi-electrode array (MEA) recordings to measure light-evoked activity of RGCs in the ganglion cell layer in theex vivomouse retina. This approach allowed us to investigate the neuromodulatory effects of NO on a cell type-level. Our findings reveal that NO selectively modulates the suppression of temporal responses in a distinct subset of contrast-suppressed RGC types, increasing their activity without altering the spatial properties of their receptive fields. Given that under photopic conditions, NO release is triggered by quick changes in light levels, we propose that these RGC types signal fast contrast changes to higher visual regions. Remarkably, we found that about one-third of the RGC types, recorded using two-photon Ca2+imaging, exhibited consistent, cell type-specific adaptational response changes throughout an experiment, independent of NO. By employing a sequential-recording paradigm, we could disentangle those additional adaptational response changes from drug-induced modulations. Taken together, our research highlights the selective neuromodulatory effects of NO on RGCs and emphasizes the need of considering non-pharmacological activity changes, like adaptation, in such study designs.

Vertebrate photoreceptors have been studied for well over a century, but a fixed nomenclature for referring to orthologous cell types across diverse species has been lacking. Instead, photoreceptors have been variably-and often confusingly-named according to morphology, presence/absence of 'rhodopsin', spectral sensitivity, chromophore usage, and/or the gene family of the opsin(s) they express. Here, we propose a unified nomenclature for vertebrate rods and cones that aligns with the naming systems of other retinal cell classes and that is based on the photoreceptor type's putative evolutionary history. This classification is informed by the functional, anatomical, developmental, and molecular identities of the neuron as a whole, including the expression of deeply conserved transcription factors required for development. The proposed names will be applicable across all vertebrates and indicative of the widest possible range of properties, including their postsynaptic wiring, and hence will allude to their common and species-specific roles in vision. Furthermore, the naming system is open-ended to accommodate the future discovery of as-yet unknown photoreceptor types.

Vision first evolved in the water, where light becomes increasingly monochromatic with viewing distance. The presence of spectrally broad (′white′) light is therefore the exclusive remit of the visual foreground. However, if and how aquatic visual systems exploit this ′white effect′ as an inductive bias, for example to judge distance, remains unknown. By combining two-photon imaging with hyperspectral stimulation, genetic cone-type ablation, and behaviour, we here show that zebrafish suppress neural responses to the visual background by contrasting ′greyscale′ and ′colour′ circuits that emerge at the first synapse of vision. To do so, zebrafish use an early retinal architecture that fundamentally differs from that of mammals: Rather than combining cone signals to drive the retinal output leading to behaviour, zebrafish vision is built around competing ancestral cone systems: Red/UV versus green/blue. Of these, the non-opponent red and UV cones, which are retained in mammals, are necessary and sufficient for vision. By contrast, the colour opponent green and blue cones, which are lost in mammals, form a net-suppressive ′auxiliary′ system that shape the ′core′ drive from red and UV cones. Our insights challenge the long–held notions that cones act in concert to drive visual behaviour, and that their spectral diversity primarily serves colour vision. Instead, we posit that vertebrate vision is ancestrally built upon opposing cone systems that emerged to exploit the strong spectral interactions of light with water. This alternative view points at terrestrialisation, not nocturnalisation, as the leading driver for visual circuit reorganisation in mammals. ### Competing Interest Statement The authors have declared no competing interest.

Neural population responses in sensory systems are driven by external physical stimuli. This stimulus-response relationship is typically characterized by receptive fields, which have been estimated by neural system identification approaches. Such models usually require a large amount of training data, yet, the recording time for animal experiments is limited, giving rise to epistemic uncertainty for the learned neural transfer functions. While deep neural network models have demonstrated excellent power on neural prediction, they usually do not provide the uncertainty of the resulting neural representations and derived statistics, such as most exciting inputs (MEIs), from in silico experiments. Here, we present a Bayesian system identification approach to predict neural responses to visual stimuli, and explore whether explicitly modeling network weight variability can be beneficial for identifying neural response properties. To this end, we use variational inference to estimate the posterior distribution of each model weight given the training data. Tests with different neural datasets demonstrate that this method can achieve higher or comparable performance on neural prediction, with a much higher data efficiency compared to Monte Carlo dropout methods and traditional models using point estimates of the model parameters. At the same time, our variational method provides us with an effectively infinite ensemble, avoiding the idiosyncrasy of any single model, to generate MEIs. This allows us to estimate the uncertainty of stimulus-response function, which we have found to be negatively correlated with the predictive performance at model level and may serve to evaluate models. Furthermore, our approach enables us to identify response properties with credible intervals and to determine whether the inferred features are meaningful by performing statistical tests on MEIs. Finally, in silico experiments show that our model generates stimuli driving neuronal activity significantly better than traditional models in the limited-data regime.

Identifying cell types and understanding their functional properties is crucial for unraveling the mechanisms underlying perception and cognition. In the retina, functional types can be identified by carefully selected stimuli, but this requires expert domain knowledge and biases the procedure towards previously known cell types. In the visual cortex, it is still unknown what functional types exist and how to identify them. Thus, for unbiased identification of the functional cell types in retina and visual cortex, new approaches are needed. Here we propose an optimization-based clustering approach using deep predictive models to obtain functional clusters of neurons using Most Discriminative Stimuli (MDS). Our approach alternates between stimulus optimization with cluster reassignment akin to an expectation-maximization algorithm. The algorithm recovers functional clusters in mouse retina, marmoset retina and macaque visual area V4. This demonstrates that our approach can successfully find discriminative stimuli across species, stages of the visual system and recording techniques. The resulting most discriminative stimuli can be used to assign functional cell types fast and on the fly, without the need to train complex predictive models or show a large natural scene dataset, paving the way for experiments that were previously limited by experimental time. Crucially, MDS are interpretable: they visualize the distinctive stimulus patterns that most unambiguously identify a specific type of neuron.