Synaptic Organization of the Primate Retina: Toward a Human Foveal Connectome

Abstract

The great diversity of retinal cell types across mammalian species, including humans, is well recognized. Recent transcriptomic profiling has not only confirmed this diversity but also identified homologous cell types across species, revealing molecular conservation of cell type identity. Yet retinal circuits are functionally specialized for distinct visual behaviors across species, raising a key question: If cell types are conserved, then what drives this functional specialization? The answer requires mapping the precise synaptic connections among all cell types—information that has long been difficult to acquire. Here we review, specifically for primates, how artificial intelligence–powered connectomics directly addresses this fundamental question by creating complete retinal connectomes that capture the synaptic connections formed by all cell types within a retinal volume. Recent data from a human foveal connectome suggest that while retinal cell types may be molecularly conserved across human and nonhuman primates, differences in synaptic connectivity drive species-specific functional specialization. Thus, synaptic connectivity, not just cell type identity, drives evolutionary adaptation.

1. INTRODUCTION

A connectome—a large-scale comprehensive map of neurons and their connections at synaptic resolution—represents a potential shift in how we approach the study of neural circuitry. In the retina of macaques, application of serial section transmission electron microscopy at limited scale has beautifully served to reveal individual cell types and their local connections and test functional hypotheses about the visual process (e.g., Calkins & Sterling 1996; Calkins et al. 1994, 1998; Saha et al. 2023; Tsukamoto & Omi 2015, 2016). By contrast, comprehensive petabyte-scale connectomes can now capture complete synaptic networks within entire neural systems (Azevedo et al. 2024, Ding et al. 2025) or even whole brains (Abbott et al. 2020, Dorkenwald et al. 2024). Thus, a connectome promises to reveal circuit motifs and wiring strategies that have remained largely inaccessible with previous smaller-scale approaches and can potentially catalyze new hypotheses about neural computation or distinguish among competing neural mechanisms (Denk et al. 2012, Seung 2024).

The rationale for applying connectomics to the primate visual pathway—or any neural system—is that neural computations can be more fully understood by mapping the complete pattern of synaptic connections among all participating cell types (Denk et al. 2012). One might argue that because neural systems are repetitive and statistical, accounting for every cell and synapse is an effort without reward. However, there are at least two reasons to reject this view. First, even for well-studied systems such as the retina—with more than 100 identified cell types (Baden et al. 2016, Bae et al. 2018, Goetz et al. 2022, Li et al. 2024)—we lack comprehensive wiring diagrams. Without knowing how most cells connect, we cannot relate circuit structure to function or understand the computational logic of visual processing. Thus, while molecular profiling has revealed remarkable cell type diversity throughout the nervous system, knowing which cell types exist is insufficient. To understand how diverse neural functions arise from a conserved toolkit of cell types, we need synaptic connectivity data. This is because functional diversity emerges largely through differences in circuit wiring—variation that is invisible to molecular methods but revealed by connectomics. Second, nervous systems have evolved independently across species to support distinct behavioral repertoires, meaning there is no canonical model system (Baden 2024). Connectomes can reveal species-specific circuit adaptations not apparent from cell type catalogs alone. They can also reveal convergent neural strategies across distantly related species (Barsotti et al. 2021, Kim et al. 2023, Schlegel et al. 2024)—computational principles that cannot be inferred from studying any single model organism in isolation.

In this context, the connectomic approach complements large-scale recording methods that employ high-density electrode arrays (International Brain Laboratory et al. 2025; Steinmetz et al. 2018, 2021). A central goal of high-density extracellular recordings—now nearly achieved (Ye et al. 2025)—is to sample neural activity densely enough to identify individual cell types and map their spatial distributions across multiple brain regions simultaneously. This parallels connectomics in its ambition to achieve both synaptic resolution and large spatial scale. However, the two approaches offer complementary strengths. Large-scale electrophysiology captures neural dynamics across brain regions during behavior, revealing how distributed populations coordinate in real time. In contrast, a complete connectome provides an exhaustive morphological census: every cell and every synaptic connection within a defined neural volume. This structural blueprint enables precise inference about information flow and computational motifs that constrain and shape the activity patterns observed during behavior. Together, these approaches bridge microcircuit organization and systems-level computation, linking the anatomical substrate of neural circuits to species-specific behavioral repertoires and cognitive functions.

The retina, as the most accessible part of the primate central nervous system (CNS) with well-characterized cell types, measurable functional outputs, and direct relevance to human vision, provides an ideal model system for pursuing connectomic goals. Moreover, the foveal retina is spatially compressed by the need to densely sample visual space, creating a unique opportunity in the vertebrate CNS to derive a connectome from functionally specialized but miniaturized and tightly packed circuitry.

In this review, we first briefly discuss the technical advances that have made whole-brain connectomes a realistic goal for vertebrate nervous systems, which was recently achieved for the Drosophila brain (Dorkenwald et al. 2024) and reviewed in detail elsewhere (Lee et al. 2019, Popovych et al. 2024). We then turn to the retina, briefly reviewing the pioneering studies that advanced our understanding of retinal circuits using electron microscopy reconstruction and established the retina as a model system for connectomics. Finally, we examine more recent work on the primate and human retina using targeted reconstructions from small volumes and consider the potential impact of using artificial intelligence (AI) tools to create the first comprehensive connectome for the human foveal retina—a key structure in the human visual system.

2. METHODOLOGICAL REVOLUTION: ARTIFICIAL INTELLIGENCE–POWERED LARGE-SCALE CONNECTOMICS

Recent advances in applying convolutional neural networks (CNNs) (i.e., deep learning) to volume electron microscopy images have transformed the construction of large-scale connectomes from an aspirational goal to a practical reality marked by accelerated progress. Artificial intelligence (AI)-based segmentation and related approaches have revolutionized three critical aspects of connectomic reconstruction that previously limited the scale of analysis: (a) more precise image alignment across thousands of sections, (b) 3D cellular segmentation of densely packed neuropil, and (c) automated synapse identification (Buhmann et al. 2021, Januszewski et al. 2018, Lee et al. 2019, Popovych et al. 2024, Turner et al. 2020). CNNs now perform these tasks with speed and accuracy approaching or exceeding those of human experts (Lee et al. 2017) (Figure 1). As a result, we can now generate realistic, complete circuit diagrams for substantial volumes of neural tissue, enabling sophisticated computational modeling that incorporates actual rather than hypothetical connectivity patterns (e.g., Lin et al. 2024). These complete circuit diagrams are leading to new hypotheses about how specific synaptic motifs give rise to function (e.g., Christenson et al. 2024).

Figure 1.

CNN-based image processing pipeline for connectome derivation. (a) A single raw EM image from a human foveal volume. The tissue was imaged at 5 nm X–Y resolution with 50 nm section thickness using serial block-face scanning electron microscopy. The input is derived from a finely aligned image stack forming a 3D tissue volume. This image, sampled from the outer retina at the level of cone pedicles, shows LM cone pedicle 13, which is tracked throughout subsequent processing steps. (b) CNNs detect neuronal boundaries. The boundary map is oversegmented into small supervoxels that define the minimal dimensions for subsequent error correction. (c) Supervoxels are agglomerated into 3D segments using statistical properties of the boundary map. Agglomerated structures provide the substrate for error correction [performed manually using efficient tools such as the Connectome Annotation and Versioning Engine (CAVE) infrastructure (Dorkenwald et al. 2025)] and ultimately yield refined segmentation of all neural and nonneural objects in the volume. (d) Synaptic structures are also predicted using CNNs; here, synaptic ribbons (red) are predicted for LM cone pedicle 13 and shown as red meshes visible through the partially transparent pedicle. (e) CNNs predict synaptic partner relations based on training with ground-truth images demarcating synaptic connections. Predicted synaptic connections are shown by yellow connecting lines arising from each ribbon. (f, g) Two representations of the synaptic output from this cone pedicle, which formed 210 synapses with diverse bipolar cell types and H1 horizontal cells, are shown. (f) 3D views of the complete network of H1 horizontal cells (12 cells) that formed synaptic connections with LM cone pedicle 13, shown with the pedicle above, viewed from the outer retina looking inward (top) and from H1 cell bodies toward photoreceptors (bottom). (g) This connectivity graph shows complete connectivity of LM cone pedicle 13 with output to 12 H1 horizontal cells (green) and diverse cone bipolar cell types. Numbers on lines connecting cells indicate the number of synaptic connections. Scale bars: 2 μm (a–d) and 5 μm (f). Abbreviations: CNN, convolutional neural network; DB1–5, ON and OFF diffuse bipolar cell types; EM, electron microscopy; FMB, flat midget bipolar (OFF) cell; IMB, invaginating midget bipolar (ON) cell; OPL, outer plexiform layer.

While all three advances are important for building a connectome, reliable and automated synapse detection is fundamentally critical given the orders of magnitude greater number of synapses than neurons (Figure 1d,e). Interestingly, some early studies using manual reconstruction methods took a seemingly backward step by enhancing the staining of membrane boundaries to facilitate dendritic reconstruction while sacrificing the ability to visualize synaptic specializations (Ding et al. 2016; Helmstaedter et al. 2011, 2013). However, the realization that CNNs could excel at predicting synapses when given appropriate training data (Azevedo et al. 2024, Buhmann et al. 2021, Turner et al. 2020) has led to a greater emphasis on achieving the best possible subcellular ultrastructure quality (Phelps et al. 2021, Zheng et al. 2018). CNNs can recognize subtle features of synaptic structures revealed at nanometer-scale resolution, including presynaptic ribbons and vesicle clouds (Kim et al. 2026), postsynaptic densities, and synaptic cleft characteristics, achieving expert-level performance with minimal false negatives (i.e., missed synapses) or false positives (i.e., incorrect synaptic assignment) (MICrONS Consortium 2025). This automation has increased the volumetric scale of feasible reconstructions by orders of magnitude, making it possible to predict millions of synapses with high reliability (Shapson-Coe et al. 2024).

Similarly, larger-volume data sets are increasingly unwieldy for manual cell type annotation and include many partial cells that intersect the edges of a volume. CNNs can now aid in cell classification by recognizing distinct features of dendritic morphology (including partial cells) and even nuclear ultrastructure, accelerating identification of distinct but low-density cell types (Dorkenwald et al. 2023, Elabbady et al. 2025).

Finally, the creation of large-scale connectomic data sets motivated the development of sophisticated collaborative proofreading tools and workflows for validating and continuously refining automated reconstructions at scale. The Connectome Annotation and Versioning Engine (CAVE) infrastructure enables distributed teams of researchers to collaboratively review segmentation results, correct errors systematically, and continuously improve reconstructions through iterative refinement (Dorkenwald et al. 2025). This system maintains sophisticated version control, allowing connectomes to be dynamically updated as corrections are made while preserving the complete history of changes. The emerging next step in this process will be to leverage the proofreading edits logged in CAVE to train networks for automated error detection and correction, further reducing the need for manual proofreading (Celii et al. 2025). The impact of these technological advances is powerfully exemplified by recent landmark data sets. A complete connectome of the adult Drosophila brain and ventral nerve cord has been produced, revealing unexpected circuit motifs and providing insights into sensory integration and motor control (Azevedo et al. 2024; Dorkenwald et al. 2022, 2024; Lesser et al. 2024; Lin et al. 2024; Sturner et al. 2025). The MICrONS Consortium generated the largest mammalian cortical connectome to date; a cubic millimeter of mouse visual cortex containing more than 200,000 cells and 500 million synapses was mapped, revealing principles of cortical organization (Ding et al. 2025, MICrONS Consortium 2025). The H01 data set provided the first large-scale reconstruction of human cortical tissue from the temporal lobe (150 million synapses), opening an unprecedented window into unique features of human neural circuitry (Shapson-Coe et al. 2024).

CNNs provide the ability to align, segment, and predict synaptic connections in large volumes of nervous tissue. However, image acquisition for these large volumes remains a major bottleneck—scanning with a single electron beam at high pixel density can take months for even small volumes in the terabyte range. Serial block-face scanning electron microscopy imaging, which scans the block-face directly after each section, eliminating section distortions, is feasible only for limited volume sizes due to the long imaging times required (Denk & Horstmann 2004). With the growing interest in larger volumes in the millimeter (petabyte) range, high-throughput methodologies have been introduced. For scanning electron microscopy, the standout advancement in acquisition speed was Zeiss’s development of the multibeam scanning electron microscope, which scans large areas with 61–91 beams in parallel (Eberle et al. 2015). However, due to its expense and resulting scarcity, published data sets using multibeam scanning electron microscopy have been limited (e.g., Shapson-Coe et al. 2024). A less expensive approach—automated transmission electron microscopy (e.g., GridTape)—relies on fast automated imaging using single or multiple electron microscopes running in parallel (Ding et al. 2025, MICrONS Consortium 2025, Phelps et al. 2021). Notably, both scanning electron microscopy and transmission electron microscopy approaches also require automated sectioning to acquire thousands of serial sections with minimal loss (Baena et al. 2019, Kasthuri et al. 2015). For details on the benefits and drawbacks of these and other related methods, see Briggman & Bock (2012) and Peddie et al. (2022).

3. GETTING TO A RETINAL CONNECTOME

The application of electron microscopy to retinal circuit reconstruction has a rich history spanning over four decades, with the retina as a crucible for each generation of connectomic technology. This history reveals not only steady technical progress but also evolving conceptual frameworks for understanding neural circuits—from simple parallel pathways to complex interconnected networks. The path from manual reconstruction of individual cells to automated mapping of complete circuits reflects a fundamental shift in how we conceptualize neural organization—from sampling to completeness, and from hypothesis-driven to discovery-driven approaches. Several comprehensive reviews provide additional context (Marc et al. 2013, 2014b; Sawant et al. 2023; Sigulinsky et al. 2024).

The 1980s opened with Peter Sterling and colleagues, who applied novel photographic approaches and minicomputers to reconstruct retinal and cortical circuits from hundreds of serial electron micrographs. They clearly recognized the fundamental need to comprehensively determine “microcircuitry,” stating: “To analyze a particular region of the vertebrate brain will ultimately require a comprehensive account of its ‘microcircuitry’. The detailed three-dimensional geometry of individual cells must be fully documented if one is to understand the electrical transformations performed by a neuron on its inputs. The detailed patterns of synaptic connections between adjacent neurons must also be fully documented if one is to understand how the neurons interact” (Stevens et al. 1980, p. 266).

Working with cat retina, these studies identified previously unrecognized cell type diversity (McGuire et al. 1984). An emphasis was placed on interconnected networks, such as ganglion cells that receive inputs from multiple bipolar cell types, raising fundamental questions about why such convergent circuitry is needed in the retina and how it gives rise to the spatial and temporal structure of the receptive field (Freed & Sterling 1988, McGuire et al. 1986, Sterling 1983). These first connectomic reconstructions initiated a movement away from simple notions of retinal circuitry as a few well-defined parallel pathways (Lennie 1980) toward recognition of the extremely complex networks that characterize mammalian retinas (Masland & Raviola 2000).

Sterling’s group extended their approach to the macaque central retina at a time when the physiology of retinal cell types was beginning to be examined in detail by applying methods to maintain the primate retina in vitro (Dacey 1989, Dacey & Lee 1994, Dacey et al. 1996, Rodieck 1988, Rodieck & Watanabe 1993, Watanabe & Rodieck 1989). The reconstructions began to challenge hypotheses about the synaptic basis for trichromatic color coding in primate retinal circuitry (Calkins & Sterling 1996, 1999; Calkins et al. 1994, 1998; Rodieck 1991). For example, the prevailing view at the time, based on physiological recordings, was that the midget pathway mediated all chromatic coding at the retinal level through cone type–specific connections to the receptive field center and surround (Kaplan et al. 1990, Reid & Shapley 1992). Electron microscopy reconstruction, by contrast, argued against cone type–selective pathways (Calkins & Sterling 1996, 1999). There is more recent evidence both for (Field et al. 2010, Lee et al. 2012, Martin et al. 2001) and against (Martin et al. 2011, Wool et al. 2018) cone type–selective or biased connections in the midget circuit. However, the underlying synaptic connections and mechanisms remain to be directly determined in a complete midget pathway connectome (Kim et al. 2024).

Greater automation was added with a terabyte-scale volume of rabbit retina (Anderson et al. 2011), emphasizing the unexpected complexity of retinal circuits previously thought to be well understood (Marc et al. 2014a). Thus, the AII amacrine cell, intensively studied with conventional electron microscopy methods and noted for its basic role in transmitting rod signals to cone bipolar cells, was shown to connect to 28 distinct retinal cell types. The connectomics of the mouse retina demonstrated the tremendous value of collaborative reconstruction and proofreading (Kim et al. 2014). For both the retina (Briggman et al. 2011) and visual cortex (Bock et al. 2011), connectomic reconstruction linked distinctive synaptic motifs directly to physiological properties. CNNs (LeCun et al. 2015) eventually enabled automated segmentation of larger tissue volumes (Turaga et al. 2010), setting both the technical and conceptual paths forward.

What is the current state of the retinal connectome? As noted above, large-scale connectomes have become available for the mouse neocortex (MICrONS Consortium 2025) and the Drosophila brain and ventral nerve cord (Azevedo et al. 2024, Dorkenwald et al. 2024) through the application of deep learning and collaborative proofreading. By contrast, retinal neurobiology, despite early recognition of the connectome’s potential, has focused largely on using volume electron microscopy to address circuit-level questions with sparse, manual reconstructions of specific cell types (for a recent review, see Sawant et al. 2023), gradually adding data points but leaving fundamental questions about complete circuit connectivity unanswered.

In the following sections, we focus on the primate foveal retina as a unique and tractable target for achieving a complete connectome of an important component of the human CNS. We first examine recent results from sparse reconstructions that have begun to reveal species-specific features of primate retinal circuitry. We then turn to the first comprehensive human foveal connectome, which directly addresses long-standing questions about the origins of human visual pathways and the synaptic connectivity underlying trichromatic vision.

4. THE FOVEA AS A TARGET FOR CONNECTOMIC RECONSTRUCTION

The foveal retina is a highly specialized CNS structure that, among the mammals, is found only in day-active primates (Bringmann et al. 2018). Foveating and fixational eye movements deliver and maintain the image at the foveal locus, where the highest spatial resolution and best color vision are achieved (Provis et al. 2013, Rucci & Poletti 2015). The fovea’s unique anatomical organization—characterized by its distinctive pit-like depression, densely packed cone photoreceptor population, laterally displaced inner retinal neurons, distinct avascular zone, and highly specialized neural circuits—fundamentally distinguishes it from both the peripheral primate retina and the retinas of nonprimate mammals (Baden et al. 2020). Understanding this specialization is therefore crucial for understanding human vision and how neural circuits can be evolutionarily optimized for specific computational tasks. Moreover, the fovea’s small size and high neuronal density make it an ideal target for comprehensive connectomic reconstruction, offering a manageable volume that nonetheless contains all major retinal cell types and circuits necessary for high-acuity vision.

4.1. Foveal Circuit Specialization and Outstanding Questions

The mammalian retina employs diverse cell types to build parallel processing pathways each tuned to extract distinct visual features optimized for species-specific behaviors (Hsiang et al. 2024, Kerschensteiner 2022, Riccitelli et al. 2025). The primate fovea, however, dramatically deviates from this conserved plan through an extreme circuit specialization. Most notably, cones are specialized for dense image sampling, and neural circuits are miniaturized to achieve matching sampling densities. The primate fovea’s defining feature—the midget (parvocellular) pathway—forms “private lines” whereby individual cones connect through single midget bipolar cells to single midget ganglion cells (Calkins et al. 1994, Jusuf et al. 2006, Kolb & Dekorver 1991, Zhang et al. 2020). This one-to-one connectivity from cone to ganglion cell is unique among mammalian retinal circuits. The result is that spatial acuity at the foveal center is limited only by the physical spacing of cones in the photoreceptor array (Rossi & Roorda 2010).

Trichromatic vision adds further complexity. Most primates possess long (L), middle (M), and short (S) wavelength-sensitive cone types, contrasting with dichromatic mammals (Jacobs 2008). The precise manner by which these cone signals are combined by retinal synapses to initiate color processing remains controversial and of great interest, notably because of human psychophysical measurements that predict distinct color-coding channels (Conway et al. 2023, Stockman & Brainard 2010).

These specializations raise fundamental questions that connectomics can now address. Does miniaturization simply pack identical circuits into smaller volumes, or have unique connectivity patterns evolved for the fovea’s unique computational demands? Recent transcriptomic profiling suggests that the fovea produces fewer parallel visual pathways than the peripheral retina (Peng et al. 2019, Yan et al. 2020), but the synaptic basis for this reduction and its functional consequences remain unknown. What can the connectome add to our understanding of the origin of color-coding pathways in the fovea? Conversely, the absence of rod photoreceptors in the fovea implies circuit adaptations whose details have not been systematically addressed. Answering these questions requires moving beyond sparse, hypothesis-driven reconstructions to comprehensive connectomes that capture complete wiring patterns.

4.2. Circuit Expectations for Human Color Vision: The Cardinal Axes

Classic human psychophysical experiments together with physiological studies of the macaque visual pathway established a single, highly influential view of early-stage processing of chromatic signals in the primate trichromatic retina (Derrington et al. 1984, Krauskopf et al. 1982, Lennie & D’Zmura 1988) that became linked over time to two distinct retinal pathways or color subsystems (Mollon 1999). First, in the more evolutionarily conserved subsystem (Haverkamp et al. 2005), a selective pathway from S wavelength-sensitive cones through S cone–selective bipolar cells to a distinct bistratified ganglion cell type showing S versus LM cone opponency initiates a blue–yellow axis of color vision. This synaptic connectivity has been demonstrated using connectomics in human (Kim et al. 2023, 2024), macaque (Calkins et al. 1998, Patterson et al. 2019, Wool et al. 2019) and marmoset (Ghosh & Grünert 1999, Kim et al. 2023) foveal retina.

Second, in the phylogenetically more recent pathway, the appearance of L and M wavelength-sensitive cones in the primate lineage was utilized by a system originally established for high spatial resolution: the private-line midget circuit. In the foveal retina, each L and M cone makes a private connection to ON and OFF midget bipolar cells, which in turn connect privately to ON and OFF midget ganglion cells, establishing pure L and M cone receptive field centers. The surround created by horizontal cell negative feedback is nonselectively connected to both L and M cone types but nevertheless, by differential weighting of cone inputs across the center and surround, establishes an L versus M cone opponent pathway (Field et al. 2010, Lennie et al. 1991, Wool et al. 2018). Such a pathway is critical for human color vision because the red–green channel defined psychophysically requires the creation of L versus M cone opponency (Stockman & Brainard 2010).

Connectomic reconstruction has revealed an additional S cone pathway that complicates this picture. A morphologically distinct S-OFF midget pathway exists in the macaque retina (Klug et al. 2003, Patterson et al. 2019, Wool et al. 2019), where each S cone synapses on an OFF midget bipolar cell, which is in turn presynaptic to an OFF midget ganglion cell, forming a private-line S-OFF pathway. The chromatic tuning of this pathway in the foveal retina is not well studied, but the expectation is that it would show an S-OFF, LM-ON response and contribute to the blue–yellow axis of color vision. This midget S-OFF pathway’s presence in the human retina has now been directly demonstrated by connectomic reconstruction (Kim et al. 2023).

A key property of the canonical model is the complete segregation of S cone synaptic connections from the midget pathways that transmit L versus M cone opponent signals. This segregation is the essential feature of the cardinal axes, where one chromatic axis is S versus LM and the orthogonal axis is L versus M. As we discuss below, connectomic reconstruction of the human foveal retina suggests significant departures from that predicted by the cardinal axes—at all synaptic levels in the retina—raising the question of how this unexpected circuitry fits with the classical view of early-stage color processing in human vision.

4.3. Photoreceptor Coupling: A Human-Specific Trichromatic Network

The first synapse in the visual pathway is electrical and bidirectional, formed by gap junctions between photoreceptor axon terminals (rod spherules and cone pedicles). These gap junctions are present at telodendritic extensions that interconnect the cone pedicles, and in the central retina where cone synapses are packed tightly; they also occur when the pedicles contact each other (Tsukamoto et al. 1992). Previous physiological studies in the macaque retina found that gap junctions mediated a large conductance between L and M cones but that S cones were electrically isolated from this coupled network (Hornstein et al. 2004). The lack of gap junction–mediated conductance between S and L or M cones is consistent with immunohistochemical labeling of gap junction proteins (O’Brien et al. 2012) and with the lack of a physical connection between these cone types (Wool et al. 2019).

However, observations of human cone axon terminals that suggested the presence of S–LM cone gap junctions were inconsistent with this result (Kolb et al. 1997). It is possible that limited data from the human retina did not represent the true picture as shown for the macaque. However, recent analysis of the human foveal retinal connectome confirmed these previous observations (Kim et al. 2023). In addition, a complete human connectome with systematic mapping of cone–cone contacts for more than 300 cone pedicles revealed extensive LM–LM and S–LM cone contacts in the volume (Kim et al. 2026) (Figure 2).

Figure 2.

Species-specific photoreceptor coupling patterns reveal human-specific S–LM cone connections. (a) Human: array of photoreceptor axon terminals (cone pedicles, rod spherules) viewing the synaptic face of the terminals (296 LM cones, gold; 17 S cones, blue; 24 rods, brown). (b) Six cone pedicles from within the white dotted box in panel a. (b, left) The pedicles contact each other via telodendria (arrows) within the circled areas. A pedicle identified as an S cone (blue, white asterisk) contacts neighboring LM cone 230. (b, right) Electron microscopy micrographs showing telodendritic contacts (open white arrowheads) between LM cones (top) and an S cone with an LM cone (bottom). (c) Marmoset: five cone pedicles with an S cone pedicle (purple) at the center. The S cone pedicle completely lacks telodendria and any contacts with neighboring LM cones (yellow/gold/beige). (d) Macaque: an S cone pedicle (blue) neighboring three LM cone pedicles (yellow/gold). LM cones connect to each other via these telodendria. The S cone contact with neighboring LM cones is rare, with either zero contacts or one contact. The human fovea uniquely shows S–LM cone coupling absent in other primates. Panels a and b adapted with permission from Kim et al. (2026) (CC BY-NC-ND 4.0); panels c and d adapted with permission from Kim et al. (2023) (CC BY-NC-ND 4.0).

What is the functional significance of this S–LM cone coupling in the human foveal retina? The lack of S–LM cone coupling found for the nonhuman primate retina is also observed in ground squirrel, where paired recordings also show M–M cone coupling but no S–M cone coupling (Li & DeVries 2004). Electrical coupling between cones averages their signals, reducing uncorrelated noise and improving signal-to-noise ratios (DeVries et al. 2002). However, coupling also dilutes chromatic contrast by mixing signals from cones with different spectral sensitivities. In this view, in circuits specialized for pure color vision, such as the S cone pathway, maintaining chromatic purity is critical: S cones must remain electrically isolated from LM cones to preserve strong chromatic signals for downstream opponent processing in ganglion cells. By contrast, the midget pathway processes both chromatic and achromatic spatial information. Here, LM cone coupling enhances spatial vision at the cost of modest chromatic dilution—a favorable trade-off given that achromatic contrast sensitivity drives high-acuity spatial vision (Hsu et al. 2000).

The discovery of S–LM cone coupling in the human fovea thus presents a puzzle: Why would the human retina adopt a connectivity pattern that degrades chromatic purity? Biophysical modeling offers insight. Given current knowledge of gap junction conductances (Srinivas et al. 1999) and their distribution within cone–cone contacts (Tsukamoto et al. 1992), together with the number and surface area of those contacts derived from the human foveal data set, it was shown that even for the modest conductances found, the S–LM cone coupling impacts the spectral tuning of both S and LM cone synaptic voltage responses (Kim et al. 2026), suggesting that the role of S–LM coupling could be to alter the spectral tuning of the cone light response, adding short wavelength sensitivity to LM cone output and vice versa for the S cone output, as suggested by previous models (Hsu et al. 2000).

4.4. Horizontal Cells: Another S Cone Pathway

The second synaptic site that could impact color vision is the connection between cones and horizontal cell interneurons. Horizontal cells provide lateral negative feedback to cones, contribute to the receptive field surround of cones, bipolar cells, and ganglion cells, and are essential for center–surround receptive field organization (Chapot et al. 2017, Morikawa et al. 2024, Rogerson et al. 2017, Thoreson & Mangel 2012). In the macaque retina, the horizontal cell–mediated surround is critical for L versus M cone opponency in midget ganglion cells (Crook et al. 2011). In both macaque and marmoset, two horizontal cell types with distinct cone connectivity, H1 and H2 cells, are recognized. H1 cells contact LM cones nearly exclusively while avoiding input from S cones, whereas H2 cells densely contact S cones and receive reduced input from LM cones (Chan & Grünert 1998, Dacey et al. 1996, Goodchild et al. 1996). This selective connectivity appears to maintain the L versus M cardinal axis segregation at the first synaptic level with H1 cells as the source of the surround in LM cones.

The human fovea reveals a striking departure from this pattern. First, H1 cells comprise the great majority (~96%) of horizontal cells at this foveal location (Kim et al. 2026). Based on the foveal connectome, H1 cells receive 43% of their synaptic input from S cone pedicles—nearly comparable to H2 cells, which receive ~57% of their synaptic input from S cones. Thus, H1 cells, the numerically dominant horizontal cell of the human fovea, integrate signals from all three cone types rather than selectively sampling LM cones, as observed in the retinal periphery of nonhuman primates. This connectivity suggests that H1 cells may have receptive fields that receive signals from all three cone classes, which would in turn produce surrounds with this property in the midget bipolar and midget ganglion cells. If confirmed, this would represent a substantial S cone component in the human midget receptive field surround, inconsistent with the midget circuit in the macaque and its strict L versus M opponency.

4.5. Human Bipolar Cells Lack Cone Type Specificity

Perhaps the most compelling evidence for human-specific chromatic processing comes from the systematic patterns of connections between cones and bipolar cells revealed by recent human foveal connectomics. These connections violate the prevailing view from nonhuman primate data for segregation of S cone signals from the L versus M chromatic pathway (Calkins et al. 1994, Herr et al. 2003, Klug et al. 2003, Wool et al. 2018). In macaques, S cones make all synaptic output at invaginating contacts with the S cone–selective, S-ON bipolar cell type, but in the human foveal retina, some of these synapses are made with ON midget bipolar cells—cells that normally connect exclusively to single L or M cones. Conversely, LM cones can be presynaptic to S-ON (blue cone) bipolar cells (Kim et al. 2023) that canonically should receive input only from S cones to maintain blue–yellow channel purity. These connectivity patterns provide another source for noncardinal interactions where S cones could contribute to the excitatory receptive field center of the L versus M-ON midget pathway, and LM cones could contribute to the S-ON receptive field center response of the small bistratified blue–yellow pathway. These findings fundamentally challenge the canonical view that cardinal color axes are established through strict cone type selectivity at the first retinal synapse.

4.6. Ganglion Cells: Convergent Chromatic Outputs

The culmination of S–LM integration in the human foveal retina manifests in specialized ganglion cell types that receive converging connections from traditionally separate chromatic pathways, creating output channels to the brain that may encode chromatic information via circuits that are not yet well understood. Two circuit motifs exemplify this integration, both representing significant departures from canonical nonhuman primate visual pathways.

4.6.1. Large bistratified ganglion cells.

Large bistratified ganglion cells, which costratify with small bistratified ganglion cells and also show an S-ON light response, can be distinguished in the macaque retina by their much larger, more sparsely branched dendritic fields, at least in the retinal periphery (Dacey et al. 2003). This cell type has also been recognized anatomically in both the human and the marmoset peripheral retina (Masri et al. 2019, Peterson & Dacey 2000). These distinctive ganglion cells have now been observed in the human foveal retina and were found to receive convergent excitatory inputs from both midget and blue cone bipolar cell (BB) types (Kim et al. 2024). Systematic reconstruction reveals that the total excitatory synaptic input to the large bistratified ganglion cell is diverse, arising from S-ON bipolar cells (16%), ON midget bipolar cells (30%), OFF midget bipolar cells (11%), and other diffuse bipolar types (43%) (Kim et al. 2024) (Figure 3). This connectivity pattern directly merges the classical S cone pathway (mediated via S-ON bipolars) with the red–green pathway (via midget bipolars) at the level of a single visual pathway.

Figure 3.

Convergent ON and OFF midget bipolar and S-ON bipolar inputs to large bistratified ganglion cells. (a) Horizontal view of the dendritic trees of an SBGC (gold) and a partial LBGC (red). Blue balls mark S-ON bipolar synapses to the LBGC (50 synapses from 12 BBs), revealing sparse convergent inputs across the large dendritic field. (b) Vertical view (90° rotation of panel a) showing bistratification in the IPL. The LBGC receives convergent inputs from three bipolar cell types. IMBs (green balls; 43 synapses from 31 IMBs) stratify deep in the IPL, colocalizing with BB synapses. FMBs (yellow balls; 33 synapses from 16 FMBs) target outer in the outer IPL. An additional 60 synapses from diverse ON and OFF diffuse bipolar types provide further convergent inputs. This multi-pathway convergence creates LBGCs that integrate S-ON chromatic signals with L versus M opponent midget bipolar inputs. Abbreviations: BB, blue cone bipolar cell; FMB, flat midget bipolar (OFF) cell; GCL, ganglion cell layer; IMB, invaginating midget bipolar (ON) cell; IPL, inner plexiform layer; LBGC, large bistratified ganglion cell; SBGC, small bistratified ganglion cell. Figure adapted with permission from Kim et al. (2024) (CC BY-NC-ND 4.0).

The convergence of blue cone and midget bipolar inputs to the large bistratified ganglion cell raises the question of whether they also exhibit an S-ON response as well as L versus M color opponency. A complete connectome of this cell population is now available in the human foveal connectome (Figure 4). Realistic biophysical modeling of these cells could be used to address whether the receptive field could be tuned to noncardinal color directions. Again, these synaptic connections in the human fovea do not support the view, established from work in macaque, that S cone signals are kept separate from L versus M pathways until higher levels of the visual system (Li et al. 2022).

Figure 4.

Human midget, parasol, small bistratified, and large bistratified ganglion cells. (a–d) Vertical views showing stratification patterns in the IPL. (a) Cluster of inner/ON (red-violet) and outer/OFF (yellow-green) midget ganglion cells; dendrites stratify narrowly in inner/ON and outer/OFF halves of the IPL. (b) Inner/ON (red-violet) and outer/OFF (yellow-green) parasol ganglion cells; dendrites stratify broadly near the center of the IPL. (c) Small bistratified ganglion cells extend dendrites across the ON-OFF subdivision of the IPL with compact, lobulated arbors. (d) Large bistratified ganglion cells show large dendritic fields and broad stratification with sparse branching. The white arrowhead indicates branching processes arising from the cell body (of unknown significance). (e–h) Horizontal views showing dendritic mosaic organization of the same cell types. (e) Midget cell mosaic: Outer/OFF (yellow-green, top) and inner/ON (red-violet, bottom) show dense tiling with minimal dendritic overlap. (f) Parasol cell mosaic: Outer/OFF dendrites (multiple colors showing individual cells) demonstrate minimal overlap and uniform coverage. The inner/ON parasol mosaic is not shown. (g) Small bistratified ganglion cell mosaic: Three cells (different shades of blue) show minimal dendritic tree overlap and relatively uniform coverage. (h) Large bistratified ganglion cell mosaic: Three cells (dark blue, light blue, purple) show much larger, irregularly shaped, and more sparsely branching dendritic trees compared to the small bistratified ganglion cells. Abbreviations: GCL, ganglion cell layer; IPL, inner plexiform layer. Figure adapted with permission from Kim et al. (2026) (CC BY-NC-ND 4.0).

4.6.2. ON midget ganglion cells with S-ON bipolar input.

A second circuit motif that shows this potential S + L-versus-M convergent pattern has been observed in a subset of ON midget ganglion cells. Some proportion of these cells receives direct synaptic input from S-ON or blue cone bipolar cells—sometimes through multiple ribbon synapses—alongside their dominant midget bipolar input from a single L or M cone (Kim et al. 2023, 2024) (Figure 5). As for the large bistratified ganglion cells, adding a minor S cone contribution to the receptive field center of L versus M midget ganglion cells could generate noncardinal chromatic tuning, much like that shown previously for the S-OFF midget pathway (Wool et al. 2019). This also raises the question of whether the high-density midget population is subdivided into chromatically distinct subpopulations by synaptic connections that modify the chromatic response. This question can be addressed directly in future analysis of the human foveal connectome, where the synaptic connections of all midget ganglion cells can now be characterized.

Figure 5.

Divergent S-ON bipolar output creates noncardinal chromatic pathways in the human fovea. (a) A single S-ON bipolar cell (BB, blue) presynaptic to two neighboring IMGCs [IMGC 1 (violet) and IMGC 2 (gold)]. Ribbon synapses (red balls) from this BB comprise 15.9% and 19.1% of the total bipolar input to IMGCs 1 and 2, respectively; the remaining input derives from LM cone midget bipolars. (b) Reconstruction of a single BB synaptic output (ribbon synapse locations are indicated by colored balls within the axon terminal) exemplifies the main findings. Most synapses are presynaptic to the SBGC (gold balls; 69 synapses; 78.4% of total), with smaller numbers distributing to the LBGC (red balls; 12 synapses; 13.6% of total) and the IMGC (green balls; 7 synapses; 8% of total). Not every BB contacts all three GC types, and the majority of IMGCs do not receive BB input, but when this synapse occurs, it provides a significant fraction of the total input to a given midget GC. (c–e) Hypothesized receptive field spatiochromatic 2D Gaussian profiles for the two noncardinal chromatic pathways, compared to that for the SBGC. (c) The SBGC shows S-ON versus LM-OFF cone opponency. (d) LBGC: The receptive field center combines moderate S-ON input with L versus M opponency derived from midget bipolar input. L versus M opponency assumes surround cancellation of the weaker cone input to the receptive field surround. (e) IMGC: SL versus M or SM versus L opponency, assuming small S-ON input combined with L versus M cone opponency. Abbreviations: BB, blue cone bipolar cell; GC, ganglion cell; IMGC, inner/ON midget GC; LBGC, large bistratified GC; SBGC, small bistratified GC. Figure adapted with permission from Kim et al. (2024) (CC BY-NC-ND 4.0).

4.7. Functional Significance of Noncardinal Synaptic Interactions

This systematic S–LM integration across all retinal levels—beginning with S–LM cone coupling and extending to the convergence of blue cone and midget bipolar signals at the ganglion cell level—reveals that the S cone pathway isolation observed in nonhuman primates does not strictly apply to the human fovea retina. These findings suggest that while human opponent color channels may be initiated by comparable anatomical pathways (midget, small bistratified, and large bistratified types), the balance of cone inputs to each opponent mechanism may be human specific and distinct from that of nonhuman primates (at least the macaque). Rather than maintaining independent cardinal channels that combine only at cortical levels (e.g., De Valois et al. 2000), the human retina shows extensive chromatic pathway interaction at the earliest stages of the visual process.

The functional consequences of this organization remain unclear, but these initial anatomical findings may help resolve ongoing controversies about the neural bases of human trichromatic vision (see also Mollon & Danilova 2026). It has long been recognized that chromatic tuning along the cardinal axes does not fit the chromatic tuning of the classically defined and long-standing red–green and blue–yellow appearance channels of human color vision because S cones contribute to the red–green channel (De Valois et al. 1997). One solution to this problem is to assume that mixing of the two channels occurs at a later stage of processing (Solomon & Lennie 2007). Alternatively, additional chromatic channels have been proposed, while the usefulness of the classical color channels themselves continues to be debated (Conway et al. 2023, Danilova & Mollon 2012). A retinal basis for noncardinal mechanisms in the nonhuman primate retina has been proposed (Godat et al. 2024), but its anatomical basis remains elusive.

Data from the human foveal connectome provide an alternative interpretation: Rather than being a generic feature of nonhuman primates, S cone contributions to the L versus M channel may arise specifically from the wiring of the human foveal retina. The synaptic interactions between S and LM opponent pathways—at the photoreceptor, horizontal cell, bipolar cell, and ganglion cell levels—create multiple origins for possible noncardinal chromatic responses. These data suggest that more complete circuit diagrams have the potential to substantially revise much of our thinking about how retinal circuitry contributes to human color vision. Understanding how these human-specific synaptic patterns translate into chromatic perception will require integration of connectomic data with psychophysical measurements and computational modeling, which is now facilitated by the availability of complete human foveal circuits.

5. IDENTITY AND NUMBER OF VISUAL PATHWAYS ARISING FROM THE HUMAN FOVEA

A striking pattern emerges from comparing visual pathway diversity across species: Converging evidence reveals a clear mismatch in the number of visual pathways (i.e., ganglion cell types) in the mouse versus the nonhuman primate. Calcium imaging of cell populations, morphological reconstructions, and transcriptomic profiling (Goetz et al. 2022, Kim et al. 2022, Peng et al. 2019, Yan et al. 2020) of retinal neurons all converge on the same conclusion: The mouse retina contains more than twice as many visual pathways (~42 types) as the monkey or the human retina (~20 types each). It might be argued that the less accessible primate ganglion cell types may be only partially identified. However, for both the mouse and the monkey, the anatomical and transcriptomic data are in close alignment, suggesting that for both species we have converged on a nearly complete picture of visual pathway diversity.

Even more striking is that for the human foveal retina, the number of visual pathways is reduced further, by nearly half (Yan et al. 2020), with both molecular profiling and the connectome of the human fovea suggesting 12 visual pathways present at this distinctive location (Kim et al. 2026). In addition, the connectome reveals that only six types (ON and OFF midget and parasol cells as well as the small and large bistratified ganglion cells of the S cone pathway) account for more than 97% of the total ganglion cells in the human fovea (Figure 4). Indeed, the midget ganglion cells alone account for nearly 90% of the total ganglion cells.

What explains this dramatic reduction in ganglion cell types from mouse to primate? The difference in pathway diversity and gene expression suggests that humans and mice have evolved fundamentally different solutions to analyzing the visual scene. Great emphasis has been placed on the behaviorally relevant feature-detection properties of diverse mouse ganglion cell types (Hsiang et al. 2024, Kerschensteiner 2022, Riccitelli et al. 2025). By contrast, in primates, considerable emphasis has been placed on the dual and more general role of the midget pathway in both form and color perception (Crook et al. 2014, Lennie & Movshon 2005, Martin et al. 2011). While there remains interest in the discovery of additional feature-detecting visual pathways in the primate retina (Kling et al. 2024), here we explore how the foveal specialization reflects fundamental differences in visual strategies between species.

5.1. Feature Detectors Versus Fixational Eye Movements

Integrative transcriptomic analysis of retinas from a wide range of vertebrates has extended our understanding of how evolutionary change shapes visual pathway structure and function. By identifying orthologous cell types across species—orthotypes that share related transcriptomic profiles—it becomes possible to trace the evolutionary history of homologous cell populations and understand how they have diverged or been conserved across lineages (Hahn et al. 2023).

One long-standing question concerns the evolutionary origins of primate midget and parasol ganglion cells—the numerically dominant retinal output neurons that have been intensively studied. Specifically, which ganglion cell types in other mammals represent the primitive forms from which midget and parasol cells evolved? This question first emerged when cat retinal ganglion cells were characterized both physiologically and anatomically. In the cat, X cells and Y cells—defined by their distinct physiological response properties—were identified as corresponding to the anatomically defined beta and alpha ganglion cell types, respectively (Cleland et al. 1975; Peichl & Wässle 1979, 1981). Beta/X and alpha/Y cells were subsequently linked to the midget and parasol ganglion cells of the primate based on similarities in morphology, physiology, and central projections (Dreher et al. 1976, Leventhal et al. 1981, Sherman et al. 1976, Wässle & Boycott 1991). However, the proposed homology proved controversial for a number of reasons (Crook et al. 2014, Shapley & Perry 1986) and centered around the proposal that the specialization of the midget pathway for color vision reflected a primate specialization rather than an elaboration of an ancestral circuit.

Strikingly, the mouse retina lacks a fovea and an area of high cell density like those the monkey or the cat, and there are no obvious high-density ganglion cell types that might be candidate homologues to ON and OFF midget and parasol ganglion cells. However, orthotype analysis now shows that these primate ganglion cell types are homologous to sustained and transient alpha cells, respectively, in the mouse (Hahn et al. 2023). Consistent with this interpretation, the ON and OFF sustained alpha cell types share both depth of stratification [near the inner plexiform layer (IPL) borders] and sustained spike discharge with midget ganglion cells. Similarly, ON and OFF transient alpha cells share stratification (near the IPL center) and physiological properties with the parasol cells. Unexpected in this result is the extremely large dendritic tree and consequent low spatial density (~2% of total ganglion cells) of the sustained alpha/midget ortholog.

Thus, the mouse retina appears to sacrifice receptive field spatial resolution (Prusky & Douglas 2004), opting instead for an expansion in ganglion cell types that act as feature detectors that do not require detailed inspection of a visual image but instead rely on trigger features linked to motor output—with the mouse eye apparently closer to the frog’s eye (Lettvin et al. 1959) in strategies for using vision to survive (Kerschensteiner 2022, Riccitelli et al. 2025). By contrast, the midget circuit represents an evolutionary extreme where the spatial resolution afforded by tightly packed photoreceptors is preserved through private-line connectivity linking each cone to a single ganglion cell.

In this context, it is critical to recognize that primates not only explore a visual scene by making saccadic eye movements but also make small fixational eye movements (microsaccades) to extract information at peak resolution (Rucci & Poletti 2015). These microsaccades are dynamically modulated during highly precise manipulation (e.g., threading a needle) around regions of interest (Ko et al. 2010, Poletti & Rucci 2016, Rucci & Poletti 2015). Thus, the incessantly mobile fovea and related retinal circuits are designed for the inspection, manipulation, interpretation, and understanding of an object, the converse of a feature detector that triggers motor responses based on specific image features without detailed scene analysis.

5.2. Low-Density Visual Pathways: Completing the Connectome

A number of recent studies using small volumes from the macaque retina have made sparse, targeted reconstructions to address specific questions about the circuitry of low-density, large-field ganglion cell types whose functional significance is not well understood. These studies are extremely labor intensive and require significant manual or semiautomated tracing to reconstruct even a single neuron and its synaptic partners with fidelity. Despite this effort, only a few examples of each cell type can be captured, and conclusions about population-level connectivity patterns or circuit variability are necessarily limited. Nevertheless, the circuits discovered by these targeted reconstructions have revealed unexpected synaptic partnerships, identified previously unknown presynaptic cell types, and generated specific hypotheses about visual computations. In doing so, they demonstrate the kind of mechanistic insight that becomes possible with connectomic data and point toward the far richer understanding that will emerge when large-scale, densely segmented connectomes enable systematic analysis of complete circuit motifs across all instances of every cell type within entire retinal volumes.

5.2.1. Direction selectivity.

The circuitry that underlies direction selectivity in the primate retina has started to come into view. Two direction-selective ganglion cell types have been recognized physiologically, an ON-OFF type and an ON type, linked respectively to the recursive bistratified and monostratified cells identified anatomically (Kim et al. 2022, Wang et al. 2023). The well-studied starburst amacrine cell type found in nonprimates (Vaney et al. 2012, Wei 2018) is also present in the primate retina (Kim et al. 2022, Rodieck 1989, Rodieck & Marshak 1992), and given its crucial role in direction selectivity in nonprimate mammals, its synaptic connectivity in primates is a major question. Results suggest that for the presumptive ON recursive monostratified cell, asymmetric connectivity to individual starburst dendrites could underlie direction selectivity utilizing the same mechanism and connectivity suggested for nonprimate mammals (Patterson et al. 2022). An important next step in completing the connectome for the primate retina will be a detailed analysis of the synaptic connectivity of the starburst network, which is also present in the human foveal connectome (Kim et al. 2026). It will be critical to determine the identity of all ganglion cell types with input from starburst amacrine cells and the spatial distribution of the starburst connections. Moreover, the connectome will permit more extensive characterization of the roles other amacrine cell types play in direction selectivity (Ding et al. 2016, Huang et al. 2019, Kim et al. 2022).

5.2.2. Intrinsically photosensitive retinal ganglion cells.

Recent reconstructions have identified the intrinsically photosensitive retinal ganglion cells (ipRGC) of the primate retina in retinal volumes of macaques, revealing some surprising results about their connectivity. In human and nonhuman primates, ipRGCs comprise two main types based on their stratification pattern: outer-stratifying (M1) cells and inner-stratifying (M2) cells. Both types feature very large and sparsely branching dendritic trees, making them potentially identifiable in electron microscopy volumes with reasonable certainty. Despite their stratification differences, both the M1 and M2 cell types show ON-type light responses, as well as an S-OFF color opponency (Dacey et al. 2005).

The circuitry underlying these responses differs between types. For M1 (outer-stratifying) cells, reconstructions suggest that an amacrine cell type with selective input from S-ON bipolar cells provides a sign-inverting inhibitory input, thus accounting for the S-OFF opponent response component (Patterson et al. 2020b). However, a puzzling finding emerged: Reconstructions from four displaced M1 cells revealed diverse bipolar cell inputs—three received input from predominantly OFF types (DB1, DB2, and OFF midget), while one received input from predominantly ON types (ON midget, DB5, and DB6)—but lacked the expected S cone amacrine cell connections (Bordt et al. 2022). This finding would suggest that displaced M1 cells are physiologically distinct from their counterparts in the ganglion cell layer.

For M2 (inner-stratifying) cells, reconstructions showed abundant direct input from S-ON bipolar cells, which would predict an S-ON opponent response (Patterson et al. 2020c), yet previously recorded M2 cells did not show such S-ON responses but rather S-OFF responses (Dacey et al. 2005). This mismatch between anatomical predictions and physiological recordings remains unresolved.

Taken together, these results from sparse reconstructions reveal inconsistencies that highlight the limitations of studying small, selected samples. A large-scale connectome will be essential for systematic identification of all ipRGC subtypes in the context of all other ganglion cell types, enabling resolution of these circuit–function discrepancies.

5.2.3. Broad thorny and smooth monostratified cells.

Two low-density types, the broad thorny and smooth monostratified ganglion cells, have been identified by their distinct morphology and stratification depth in the IPL; reconstructions reveal distinct and unusual connectivity patterns. The broad thorny type, an achromatic ON-OFF cell with fine, densely branched, thorn-laden dendrites extending broadly across the center of the IPL (Puller et al. 2015), surprisingly receives almost all of its synaptic input from amacrine cells (Bordt et al. 2021). The large excitatory ON-OFF synaptic currents recorded in these cells likely derive from excitatory amacrine cells, such as starburst amacrines that release acetylcholine (Lee et al. 2010), or from the small-field glycinergic type, immunoreactive for vesicular glutamate transporter 3 (VGluT3), that also releases glutamate (Marshak 2016).

By contrast, the smooth monostratified ganglion cells form mirror-symmetric ON and OFF responding pairs that costratify with parasol cells near the center of the IPL and show transient achromatic light responses (Crook et al. 2008). Unlike parasol cells, their receptive fields are much larger, and sensitivity is not homogeneous across the receptive field due to gaps in the large, sparsely branching dendritic tree (Rhoades et al. 2019). Parasol cells derive ~80% of their synaptic input from amacrine cells (Patterson et al. 2020a), while smooth monostratified ganglion cells receive a similar percentage of input from amacrine cells (Patterson et al. 2024). The two types can be distinguished, however, by the density and clustering of bipolar inputs, suggesting an explanation for areas of increased sensitivity within the large receptive fields (Patterson et al. 2024). Despite these labor-intensive reconstructions, the functional significance of these pathways remains enigmatic. Their function may be better understood in large-scale, complete connectomes where the effects of retinal location can be systematically analyzed, and more complete wiring diagrams can suggest new hypotheses about function testable with direct recordings and accurate biophysical models.

SUMMARY POINTS.

1. Artificial intelligence–powered connectomics has made large-scale, complete neural circuit reconstruction feasible through convolutional neural networks for image alignment, 3D segmentation, synapse detection, and cell type annotation.

2. Collaborative proofreading platforms such as the Connectome Annotation and Versioning Engine (CAVE) enable distributed teams to validate and refine automated reconstructions, making community-scale connectome projects practical, efficient, and fast.

3. The first human foveal connectome reveals both conservation and species-specific specializations in retinal circuits, with molecular cell type identity largely preserved but synaptic connectivity patterns showing unexpected human-specific features.

4. The human foveal retina shows dramatically reduced pathway diversity compared to that of the mouse, with only ~12 ganglion cell types (versus ~42 in mouse) and midget cells alone comprising ~90% of foveal output.

5. Human foveal circuits show extensive S–LM cone integration at multiple synaptic levels—photoreceptor coupling, horizontal cells, bipolar cells, and ganglion cells—challenging the canonical model of segregated cardinal color axes and suggesting human-specific mechanisms for chromatic processing.

6. Synaptic connectivity, not just cell type identity, drives functional specialization: While retinal cell types are conserved across primates, species-specific differences in synaptic connectivity appear to underlie distinct visual capabilities.

FUTURE ISSUES.

1. Large-scale primate foveal connectomes are now technically feasible and urgently needed to address fundamental questions about retinal circuit organization and evolution.

2. How do visual pathways compare across primate species in their synaptic connectivity? Systematic comparison of human, macaque, and marmoset connectomes would reveal precisely how and where synaptic organization is conserved or modified across species.

3. How does synaptic connectivity vary with retinal eccentricity? Extending connectomic analysis from the fovea to a large-scale macular connectome would reveal how visual pathways adapt to changing photoreceptor densities, cone-to-rod ratios, and spatial resolution requirements.

4. What is the functional impact of human-specific connectivity patterns? Physiological recordings from identified cell types in connectomic volumes, combined with biophysically realistic computational modeling, are essential to determine how S–LM integration affects color and spatial processing in human vision.

5. Beyond the dominant midget and parasol pathways, do low-density ganglion cell types show conserved or species-specific connectivity? Complete connectomes can reveal the full complement of these sparse pathways and determine whether they maintain consistent circuits across species or show evolutionary plasticity.

6. Diverse amacrine cell type synaptic connectivity can best be understood with large-scale connectomes. For example, the role of starburst amacrine cells in direction selectivity in the primate retina, as well as other functions, now requires a complete starburst connectome.

Glossary

Connectome

complete map of all neural connections in a nervous system or part of a nervous system at synaptic resolution

Transmission electron microscopy

electron microscopy technique where an electron beam passes through ultrathin tissue sections, providing the highest-resolution images for synapse identification

Electron microscopy

imaging technique using electron beams to achieve nanometer resolution, revealing subcellular structures including synapses and organelles

Convolutional neural network (CNN)

deep learning architecture using layered filters to automatically detect patterns; used for image alignment, segmentation, and synapse detection

Artificial intelligence (AI)-based segmentation

use of convolutional neural networks to automatically identify and trace individual neurons and synapses in electron microscopyimage volumes

Connectome Annotation Versioning Engine (CAVE)

collaborative platform enabling distributed teams to proofread, correct, and refine automated connectome reconstructions

Scanning electron microscopy

uses an electron microscope that scans the specimen surface with the electron beam; block-face scanning electron microscopy scans the block-face surface after each section for 3D reconstruction

Multibeam scanning electron microscopy

uses an advanced scanning electron microscope using 61–91 parallel electron beams simultaneously, dramatically accelerating large-volume tissue imaging for petascale connectomes

Midget pathway

mediates both high achromatic resolution and cone opponency via private-line circuits; proposed orthotype to the mouse sustained alpha cell

Trichromatic vision

color vision in primates based on three cone photoreceptor types with peak sensitivities to short (~420 nm), middle (~530 nm), and long (~560 nm) wavelengths

Fovea

central retinal region in primates with the highest cone density and spatial resolution; specialized for detailed inspection and color vision

S-OFF midget pathway

pathway where S cones form private-line connections through OFF midget bipolar to OFF midget ganglion cells; expected S-OFF, LM-ON chromatic response

Cardinal axes

classical model proposing complete segregation of two cone-opponent channels: the S versus LM (blue–yellow) and L versus M (red–green) pathways

Gap junction

electrical synapse between neurons allowing bidirectional signal flow; gap junctions between photoreceptors increase the signal-to-noise ratio

Cone coupling

electrical connection between cone photoreceptors through gap junctions; the human fovea shows unique S–LM cone coupling not found in macaque or marmoset

S–LM integration

human-specific convergence of S cone and LM cone signals at the photoreceptor, horizontal cell, bipolar cell, and ganglion cell levels; challenges cardinal axes segregation

Large bistratified ganglion cell

bistratified ganglion cell integrating traditionally separate chromatic pathways via S-ON (blue cone) bipolar, ON midget, and OFF midget bipolar inputs; creates a human-specific noncardinal color channel

Blue cone bipolar cell (BB)

specialized bipolar cell type selectively contacting S cones, initiating the blue–yellow color pathway by transmitting S cone signals to bistratified cells

Parasol cells

magnocellular visual pathway; ON and OFF types convey motion and luminance information with high temporal resolution to the brain; proposed ortholog to mouse transient alpha cells

Orthotype

cell types sharing conserved gene expression profiles across species, identified through transcriptomic analysis; reveals homology despite morphological or functional variation