Neural Circuits Underlying Multifeature Extraction in the Retina

Abstract

Classic ON–OFF direction-selective ganglion cells (DSGCs) that encode the four cardinal directions were recently shown to also be orientation-selective. To clarify the mechanisms underlying orientation selectivity, we employed a variety of electrophysiological, optogenetic, and gene knock-out strategies to test the relative contributions of glutamate, GABA, and acetylcholine (ACh) input that are known to drive DSGCs, in male and female mouse retinas. Extracellular spike recordings revealed that DSGCs respond preferentially to either vertical or horizontal bars, those that are perpendicular to their preferred–null motion axes. By contrast, the glutamate input to all four DSGC types measured using whole-cell patch-clamp techniques was found to be tuned along the vertical axis. Tuned glutamatergic excitation was heavily reliant on type 5A bipolar cells, which appear to be electrically coupled via connexin 36 containing gap junctions to the vertically oriented processes of wide-field amacrine cells. Vertically tuned inputs are transformed by the GABAergic/cholinergic “starburst” amacrine cells (SACs), which are critical components of the direction-selective circuit, into distinct patterns of inhibition and excitation. Feed-forward SAC inhibition appears to “veto” preferred orientation glutamate excitation in dorsal/ventral (but not nasal/temporal) coding DSGCs “flipping” their orientation tuning by 90° and accounts for the apparent mismatch between glutamate input tuning and the DSGC's spiking response. Together, these results reveal how two distinct synaptic motifs interact to generate complex feature selectivity, shedding light on the intricate circuitry that underlies visual processing in the retina.

Significance Statement

The classical work of Hubel and Wiesel (1959) demonstrated that neurons in the cat visual cortex are often selective for multiple stimulus features, such as direction and orientation. Here, we show that direction-selective ganglion cells (DSGCs) in the mouse retina are also selective for stimulus orientation, suggesting that multifeature extraction may begin earlier in the visual system than previously envisioned. Using a combination of patch-clamp, cell-specific genetic KO, and optogenetic strategies, we show that multifeature coding relies on distinct mechanisms in the nasal/temporal and dorsal/ventral coding DSGC.

Introduction

Detecting the orientation and direction of moving objects is critical for navigation through the visual environment. It is not surprising that these features of the visual scene are extracted at multiple levels in the visual system, starting in the retina itself. Classic single-unit recordings in the retina revealed that distinct orientation-selective (OS) and direction-selective (DS) ganglion cells (OSGCs and DSGCs, respectively) convey feature information in parallel to higher visual processing areas (Barlow and Levick, 1965; Levick, 1967; Caldwell et al., 1978). However, in a recent study, classic ON–OFF DSGCs that encode the cardinal directions were reported to also exhibit robust orientation selectivity (Hanson et al., 2023), similar to direction-selective neurons in the cortex and fly visual system (Hubel and Wiesel, 1959; Gizzi et al., 1990; Fisher et al., 2015). While the synaptic mechanisms underlying retinal direction selectivity have been well described (Fig. 1A,B; reviewed by Vaney et al., 2012; Mauss et al., 2017; Murphy-Baum et al., 2021), the mechanisms that shape orientation selectivity in DSGCs are only beginning to surface (Fig. 1C; Hanson et al., 2023).

Figure 1.

Synaptic mechanisms underlying direction and orientation selectivity in the mammalian retina. A, A schematic highlighting the ON component of the direction-selective circuit in cross section. A well-defined population of glutamatergic ON bipolar cells (BC5A-C and BC7; magenta) bridge photoreceptors to output DSGC. In addition, DSGCs receive spatially offset GABAergic/cholinergic signals from starburst amacrine cells that shape their direction-selective responses. B, SACs are endowed with radiating dendrites, each that responds best to motion in a particular direction (indicated by color). Importantly, SAC dendrites pointing in a particular direction preferentially make synapses with one of four types of DSGCs (nasal/temporal/dorsal/ventral coding), which encodes the opposite direction. For example, the temporal coding DSGC cell coding rightward motion receives input from leftward pointing SAC dendrites (blue; Briggman et al., 2011). This asymmetric wiring also predicts that SAC inhibition to each DSGC would be orientation-selective (e.g., T-DSGCs would receive horizontally tuned inhibition; dotted oval), but the degree to which SACs are activated by static stimuli is not clear. C, At least one BC type, the BC5A, has recently been shown to be selective for vertically oriented features and appears to shape the orientation tuning properties of N/T DSGCs (not requiring SAC input). Orientation selectivity in BC5A has been hypothesized to arise through Cx36-mediated gap junctions made with the vertically oriented processes of wide-field amacrine type 5A (which also receive synaptic input from BC5A; Hanson et al., 2023). The first goal of this study was to test whether any of the other types of BCs in the circuit are tuned along the horizontal axis through an analogous coupling mechanism.

In the mammalian retina, selectivity for direction and orientation appears to originate in morphologically distinct types of amacrine cells. Orientation-selective amacrine cells appear to be a diverse group, which typically span large areas of the retina (up to ∼1 mm) and are thus considered wide-field amacrine cells. Some wide-field amacrine cell types are endowed with oriented dendritic arbors that serve as an ideal substrate for orientation selectivity, while others may acquire their selectivity through network interactions (Mariani, 1983; Bloomfield, 1991, 1994; Murphy-Baum and Rowland Taylor, 2015; Antinucci and Hindges, 2018). Orientation-selective wide-field amacrine cells also appear to shape the response properties of other downstream neurons including ganglion and bipolar cells (Venkataramani and Taylor, 2010, 2016; Nath and Schwartz, 2016, 2017; Antinucci et al., 2016; Antinucci and Hindges, 2018; Johnston et al., 2019; Hanson et al., 2023). Interestingly, wide-field amacrine cells may convey information via GABAergic inhibition and/or gap junctions (Völgyi et al., 2005; Ackert et al., 2009; Hoggarth et al., 2015; Nath and Schwartz, 2017; Roy and Field, 2019; Hanson et al., 2023).

In contrast to the diverse set of wide-field amacrine cells that underlie orientation selectivity, direction selectivity appears to originate from a single, well-defined mirror symmetrical population of ON and OFF GABAergic/cholinergic “starburst” amacrine cells (SACs) with medium-sized dendritic fields (∼0.2 mm in diameter; Fig. 1A). The radiating dendrites of SACs have the unique ability to compute the direction of moving objects (Euler et al., 2002); through specific asymmetric connectivity patterns (Fig. 1B), SACs mediate distinct patterns of excitation/inhibition that alone are sufficient to shape cardinal directional selectivity across the four ON–OFF DSGC types (Briggman et al., 2011; Yonehara et al., 2011; Sethuramanujam et al., 2016; reviewed by Murphy-Baum et al., 2021). While the mechanisms of direction selectivity are well established, how they are modulated by the wide-field amacrine cell circuitry to generate orientation selectivity remains less well understood.

In theory, orientation selectivity observed in ON–OFF DSGCs could arise from the asymmetric properties of the direction-selective circuit alone, as originally envisioned for direction-selective neurons in the visual cortex (Hubel and Wiesel, 1959). However, pharmacologically blocking SAC output abolishes direction but not orientation selectivity (Hanson et al., 2023). Instead, orientation selectivity appears to arise from tuned glutamate inputs, in part arising from type 5A bipolar cells (BC5A). BC5As appear to make connexin 36 (Cx36) containing gap junctions with vertically oriented processes of a specific wide-field amacrine cell (WF5A; Fig. 1C, left; Hanson et al., 2023) and are thus thought to be selective for vertical features. The WF5As also receive BC5A input throughout their long unbranching processes (hence their name) leading to the hypothesis that the BC5A-WF5A circuit represents the minimal circuit for generating vertical orientation selectivity in the mouse retina.

The extent to which the proposed mechanism for orientation selectivity described above generalizes across the nasal/temporal and dorsal/ventral populations of ON–OFF DSGCs (which we refer to as N/T and D/V DSGCs, hereafter) remains unclear. To clarify this, we first evaluated the direction and orientation stimulus-generated spiking properties of the four types of ON–OFF DSGCs in the mouse retina and employed a variety of optogenetic and gene knock-out (KO) strategies to test how each of the three neurotransmitter systems [glutamate, GABA, and acetylcholine (ACh)] that are known to drive DSGCs (Fig. 1A) participate in shaping orientation selectivity. We found that N/T and D/V DSGCs respond more robustly to vertical and horizontal bars, respectively (i.e., bars that are orthogonal to their preferred–null motion axis); however, their orientation selectivity appears to rely on distinct synaptic mechanisms: the former relies on tuned bipolar cell excitation, while the latter relies on tuned SAC inhibition. This organization enables a common set of bipolar cells to drive multiple features and serves as an excellent example of the efficient circuit wiring schemes found in the retina.

Materials and Methods

Animals

All procedures were carried out in accordance with the Canadian Council on Animal Care (CCAC) and approved by the University of Victoria's Animal Care Committee. The majority of the recordings from ON–OFF DSGCs were obtained from retina harvested from nontransgenic C57BL/6J (RRID, IMSR_JAX:000664) mice of either sex. DSGCs were identified by their directional tuning properties. To investigate the role of electrical coupling, the Kcng4-cre mouse line [B6.129(SJL)-Kcng4^{tm1.1(cre)Jrs}/J; RRID, IMSR_JAX:029414] was crossed with Cx36-floxed conditional knock out mouse line (provided by Dr. David Paul, Harvard, Yao et al., 2018). Both single-floxed and double-floxed Cx36 transgenic mice were used for experiments as control and complete knock-out models. For optogenetic experiments, Cre-dependent ChR2 (Ai32; RRID, IMSR_JAX:024109) was crossed to either Kcng4-cre to target BC5A or ChAT-IRES-Cre (Δneo) (RRID, IMSR_JAX:031661) to target SACs. Mice of both sexes between P21 (postnatal day 21) and P140 were used for experiments. Mice were housed and maintained on a 12 h light/dark cycle.

Retinal preparation

Mice were dark-adapted for 45–60 min before carrying out the experiments. The animals were anesthetized with isoflurane and killed by cervical dislocation. Retina dissections were performed in Ringer's solution [(in mM) 110 NaCl, 2.5 KCl, 1 CaCl₂, 1.6 MgCl₂, 10 dextrose, and 22 NaHCO₃] under dim red light. Isolated retinae were laid photoreceptor side down on a 0.45 µm membrane nitrocellulose filter (Millipore) with a pre-cut window, through which images were focused onto the retina. Visualization under infrared illumination utilized a Spot RT3 charge-coupled device camera (Diagnostic Instruments) attached to an upright Olympus BX51 WI fluorescent microscope, equipped with a 60× water immersion lens (Olympus Canada).

The dissected retina was continuously perfused with Ringer's solution bubbled with carbogen (95% O₂, 5% CO₂) and warmed to 35–37°C. Perfusion rates were maintained at ∼3 ml/min. During the dissections, a small incision was made based on the scleral landmarks (Wei et al., 2010) to identify the orientation of the retina. In the experiments performed across different days, the retina was also mounted differently in various orientations to ensure that the orientation selectivity observed was not due to a bias in the experimental setup. Responses were then normalized to the dorsal/ventral axis of the retina. All pharmacology was administered through bath application. DCG-IV was purchased from Tocris, and L-AP4 and UBP-310 were purchased from Abcam. Unless otherwise noted, all other reagents were purchased from Sigma-Aldrich Canada.

Physiological recordings

Extracellular spike recordings were made by loose patch (25–50 MΩ) using a 6–10 MΩ electrode filled with Ringer's solution. Cells were considered to be ON–OFF DSGCs if they exhibited a strong directional preference; that is, they had a direction selectivity index (DSI; calculated based on the vector sum of responses measured in eight directions) greater than 0.3 and responded to the onset and offset of a static flash.

Based on the extracellular spike recordings, identified cells were then patched for whole-cell recordings using a 5–7 MΩ electrode. For voltage-clamp recordings, electrodes contained (in mM) 112.5 CsCH₃SO₃, 9.7 KCl, 1 MgCl₂, 1.5 EGTA, and 10 HEPES; the pH was adjusted to 7.4 with CsOH. Recordings were made with a MultiClamp 700B amplifier (Molecular Devices). Signals were digitized at 10 kHz (PCI-6036E acquisition board, National 9 Instruments) and acquired using custom software written in the LabVIEW environment.

Light stimulation

Visual stimuli were presented with a digital light processing projector (Texas Instruments; refresh rate 75 Hz) controlled with custom STIMGEN software written by Dr. Murphy-Baum (University of Victoria) based on the Psychophysics Toolbox extension for Matlab. The ambient background intensity, measured with a calibrated spectrophotometer (USB2000, Ocean Optics), was reported as rhodopsin photoisomerizations per second (R*/s; derived from the absorption spectrum of mouse photoreceptors; Lyubarsky et al., 1999). Visual stimuli utilized intensities that were expected to favor rod responses (less than 13 R*/s; Farrow et al., 2013; Grimes et al., 2014) unless otherwise noted. Neutral density filters were used to control the stimulus light intensity.

Light stimuli, projected from below the preparation, were focused with a substage condenser onto the photoreceptors and centered over the soma or the measured receptive field of the targeted neuron. The visual stimuli were positive contrast ranging between 20 and 1,000%. For determining the directionality of the DSGCs, light-evoked activity was measured for 100 µm × 600 µm bars moving across the visual field in eight different directions. For measuring the orientation selectivity of the identified DSGCs, light-evoked activity was measured for 100 µm × 600 µm bars flashing in eight different orientations.

Analysis of physiological data

Physiological data were analyzed using custom software written by Dr. Benjamin Murphy-Baum in Matlab (MathWorks) or Igor Pro (Wavemetrics). The spike rate was estimated by filtering the light-evoked spike train using a convolution with a Gaussian kernel with a fixed width, σ = 25 ms. Responses were averaged over multiple trials (2–5). Either peak amplitude or integrated postsynaptic currents were used to quantify light-evoked synaptic current responses, but no qualitative differences were observed between these two methods. Population data have been expressed as mean ± SEM and are indicated in the figure legend along with the number of samples. Student's t test was used to compare values under different conditions, and the differences were considered significant when p ≤ 0.05.

DSI and orientation selectivity index (OSI) were calculated as the amplitude of the vector sum:

$$\mathrm{DSI}={\displaystyle \frac{\sum R(\theta )e^{\mathrm{i}\theta }}{\sum R(\theta )}\mathrm{OSI}={\displaystyle \frac{\sum R(\theta )e^{2\mathrm{i}\theta }}{\sum R(\theta )}}}$$

where $R(\theta )$ is the response for $\theta$ direction or orientation calculated from the peak spike rate or amplitude across eight directions or orientations. The preferred angle was calculated from the resultant (DSI) or half phase of the resultant (OSI) of the vector sum. DSI and OSI ranged from 0 to 1, with 0 indicating a perfectly symmetrical response and 1 indicating a response in only one of eight directions or orientations presented. PD (preferred) refers to the direction which elicits the largest response and ND (null) refers to the opposite direction. Direction and orientation tuning widths were estimated using a Gaussian function fit to a linear tuning curve.

Results

Orientation selectivity across the four types of ON–OFF DSGCs

In the general population of ON–OFF DSGCs, the preferred orientation of static stimuli was found to be orthogonal to the preferred–null motion axis (Hanson et al., 2023), and thus it is expected that N/T and D/V DSGCs would be selective for vertical and horizontal features, respectively. As ON–OFF DSGCs had not been assigned to nasal/temporal/dorsal/ventral coding types in the previous study, the first step was to confirm this expectation. To do so we identified ON–OFF DSGCs in an in vitro retinal whole-mount preparation based on the directional tuning properties of their spikes, which we measured using extracellular recordings (Fig. 2A,B; see Materials and Methods). The dorsal pole of the retina was notched before being placed in the recording chamber (Fig. 2A), which allowed us to estimate the DSGC's directional preference relative to the veridical retinal axes. As expected, the preferred directions of ON–OFF DSGCs clustered around the cardinal directions (Fig. 2C; Sabbah et al., 2017; Tiriac et al., 2022). For simplicity, we assigned DSGCs to each subtype based on the cardinal direction that was closest to its preferred direction (i.e., within ±45°); when we checked the tuning properties of nasal-preferring DSGCs labeled in the TRHR GFP+ transgenic mouse line, they were all nasal-coding, as previously described (Rivlin-Etzion et al., 2011).

Figure 2.

All four ON–OFF DSGC types respond best to static bars that are oriented perpendicular to their preferred–null motion axis. A, A picture of the whole-mount retinal preparation. The retina was notched (white arrow), to keep track of retinal orientation. B, ON–OFF DSGCs were identified based on their spiking properties. The typical tuning patterns of the ON and OFF responses of an example N DSGC are shown in polar form (arrows indicate the normalized vector sums for ON and OFF responses; DSI). Responses of the N DSGC to static bars (8 orientations; 100 µm × 600 µm bar) shown below demonstrate that the N DSGC responds best to vertical bars. C, DSGCs were identified as nasal (N), temporal (T), dorsal (D), or ventral (V) DSGCs when their “preferred” directions fell within 90° of a particular cardinal direction. D, A scatter plot showing the relationship between the degree of orientation tuning (OSI) versus the angle relative to the DSGC preferred direction (PD). Note DSGCs appear to respond best when the orientation of the bar is 90° offset relative to the PD (the red point indicates the average OSI and preferred orientation across the population). E, The average responses to bars moving in eight directions (red) or presented in eight orientations (black) are shown for each of the four types of ON–OFF DSGCs (the x-axis range for static stimuli is 0–180°, but the data have been reflected for visual symmetry). Note that the preferred orientation is always offset 90° relative to the DSGC's PD, regardless of cell type; that is, N/T DSGCs prefer vertical bars while D/V DSGCs prefer horizontal bars.

When we stimulated DSGCs with stationary bars (100 µm × 600 µm bars; each one of eight different orientations), we found all four subtypes displayed orientation-selective spiking responses (Fig. 2E). On average, the peak spike rates of responses to bars of a preferred orientation were roughly twice as strong as responses evoked by bars orthogonal to it. The distribution of the OSI (calculated as the vector sum of responses measured across eight stimulus orientations) in DSGCs was similar to that measured for OSGCs (Fig. 2D; Nath and Schwartz, 2016). In these experiments, we took care to center the stimulus over the DSGC's receptive field to ensure that the observed selectivity for orientation did not arise artificially.

Notably, in almost every instance, the preferred orientation of static stimuli was orthogonal to the DSGC's preferred–null axis (defined by moving stimuli). In Figure 2E, the normalized peak spike rates for responses to static and moving bars are plotted together. For the sake of visual symmetry, the responses to static bars have been reflected and plotted on the same axis (i.e., for moving stimuli the angle axis range is 0–360°, while for static stimuli it is 0–180°). Importantly, the peak of the orientation tuning curve was predominantly 90° offset from the direction tuning curve (Fig. 2D,E). These results confirm that ON–OFF DSGCs exhibit either vertical or horizontal preference for static bars, orthogonal to their preferred axis of motion, extending the preliminary findings on unassigned ON–OFF DSGCs (Hanson et al., 2023).

In the retina, orientation selectivity relies on two principal mechanisms: asymmetric morphology and/or tuned synaptic inputs (Mariani, 1983; Bloomfield, 1991; Murphy-Baum and Rowland Taylor, 2015; Nath and Schwartz, 2016, 2017; Venkataramani and Taylor, 2016; Hanson et al., 2023). Indeed, ventral coding DSGCs are known to have highly asymmetric dendritic trees that generally point away from the soma toward the ventral pole of the retina (Kay et al., 2011; Trenholm et al., 2011). However, their preferred orientations lie along the horizontal axis, which is orthogonal to their vertically oriented dendritic arbors. Additionally, the other types of ON–OFF DSGCs usually exhibit more symmetrical dendritic trees (Trenholm et al., 2011; Bae et al., 2018); thus, it appears more likely the main factors that shape orientation selectivity reside in the presynaptic circuitry rather than DSGC dendritic morphology. For this reason, we next analyzed the tuning properties of the primary glutamate inputs, as well as the feed-forward GABAergic and cholinergic inputs mediated by motion-sensitive SACs, which together shape the response properties of ON–OFF DSGCs (Murphy-Baum et al., 2021).

Glutamatergic input to all four DSGC types is tuned along the vertical axis

Previously, BC5As were shown to be tuned for vertical bars, making them one of the likely sources of tuned excitation that drives vertical orientation selectivity in the N/T DSGCs (Hanson et al., 2023). However, several other BCs, including BC5B-C and BC7, are likely to drive ON responses in SACs and DSGC, and we wondered if any of these were tuned along the horizontal axis and shaped the horizontal orientation selectivity in D/V DSGCs (Figs. 1C, 2E). While connections from specific BC types appear to be stereotypically organized across the dendritic trees of DSGCs and SACs (Helmstaedter et al., 2013; Ding et al., 2016; Greene et al., 2016; Matsumoto et al., 2019; Srivastava et al., 2022), whether each type is matched to different DSGC subtypes has not been previously investigated.

Contrary to expectation, glutamate inputs to all four types of DSGCs were predominantly tuned along the vertical axis (Fig. 3A,B). In these experiments, glutamatergic excitatory postsynaptic currents (EPSCs) were isolated by pharmacologically blocking the feed-forward SAC GABAergic/cholinergic input, using a combination of the broad-spectrum nicotinic receptor antagonist 100 µM hexamethonium (Hex) and GABA_A&C receptor antagonists (5 µM gabazine and 100 µM TPMPA, respectively). Since a large component of the BC input is mediated by NMDA receptors, EPSCs were measured at a holding potential of −30 mV, to alleviate the Mg²⁺ block of NMDA receptors (Poleg-Polsky and Diamond, 2016; Sethuramanujam et al., 2017).

Figure 3.

All ON–OFF DSGC types receive vertically tuned glutamatergic inputs that in large part arise from BC5A. A, Examples of light bar-evoked glutamatergic EPSCs measured in DSGCs in the presence of GABA and ACh receptor antagonists (5 µM gabazine + 100 µM TPMPA + 100 µM Hex). EPSCs were recorded in wild-type DSGCs (black) or in DSGCs in transgenic mouse lines in which BC5As lack Cx36 (orange; Kcng4^Cre × Gjd2^fl/fl) or express ChR2 (blue; Kcng4^Cre × Ai32), or in which BC5As lack Cx36 and express ChR2 (dark blue; Kcng4^Cre × Gjd2^fl/fl × Ai32). The ChR2 responses were recorded in the added presence of photoreceptor blockers (20 µM L-AP4 + 10 µM UBP-310) using 10,000-fold higher stimulus intensities. B, The peak amplitudes of the light-evoked EPSCs measured in the four conditions in A are plotted against the orientation of the bar. The tuning properties of the D/V or N/T DSGCs were similar and thus averaged. C, Similar to A, except the center of the DSGC's receptive field was masked (400-µm-diameter mask). These responses reveal excitation from the surround. D, The photoreceptor and ChR2-mediated EPSCs shown in C are plotted against bar orientation.

When stimuli were presented along the vertical axis, glutamatergic EPSCs could be evoked several hundreds of microns away from the DSGC's central receptive field. As illustrated in Figure 3C, EPSCs were evoked by oriented stimuli even when the DSGC's receptive field center was occluded using a mask (400–900 µm in diameter). While the extent to which excitation spreads is likely to be exaggerated under these conditions due to the presence of inhibitory receptor blockers, these results indicate the presence of a vertically oriented wide-field element in the circuit (e.g., the WF5As; Fig. 1C). Furthermore, the finding that excitation is vertically tuned in both N/T and D/V DSGC types argues against the idea of a complementary set of BCs for horizontal orientation selectivity (Fig. 1C).

Experiments using two BC5A-specific genetic manipulations revealed that a large part of the vertically tuned excitation requires proper BC5A functioning. First, we analyzed DSGC EPSCs in mice in which Cx36 was knocked out in BC5As (Kcng4^Cre × Gjd2^fl/fl; Hanson et al., 2023). In this mouse line, BC5As are expected to be uncoupled from the WF5As and from the AII amacrine cells that mediate the rod response. In the BC5A Cx36 KO mice, we found it more difficult to identify DSGCs blindly. However, in the cells that exhibited clear directional tuning preferences, we found the EPSCs were reduced in amplitude and were no longer tuned for orientation. This was true for all four types of DSGCs. Furthermore, EPSCs could no longer be evoked from the far surround in any type of DSGCs, as we observed in the wild-type retina (Fig. 3A–D). However, these results should be interpreted with caution, as knocking out Cx36 early in development could potentially have more widespread effects on the circuitry than we currently appreciate.

Complementary evidence that BC5As play a major role in mediating vertically tuned excitation came from using a gain-of-function approach. Here, we expressed the light-sensitive neuromodulator channelrhodopsin2 (ChR2) in BC5As (Kcng4^Cre × Ai32) and tested the ability of optogenetically stimulated BC5As to drive DSGCs alone. Since BC5As provide most inputs to the vertically oriented WF5A processes, it might be expected that stimulating BC5As alone would be sufficient to drive orientation-selective inputs to DGSCs. In these experiments, after functionally identifying ON–OFF DSGCs, we blocked photoreceptor output [using the metabotropic glutamate receptor 6 agonist L-AP4 (20 µM), and the kainite receptor antagonist UBP-310 (10 µM)] and increased stimulus intensity 10,000-fold (10⁷ R*/s) to directly activate BC5As expressing ChR2 (Fig. 3).

Under photoreceptor blockade, we found the EPSCs were larger than those measured in wild-type conditions, indicating that the ChR2 stimulation effectively activates BC5A (Fig. 3A). Notably, the ChR2-evoked EPSCs were tuned along the vertical axis (Fig. 3A,B). When we used a mask to occlude the DSGC's receptive field center, the orientation selectivity was further increased (Fig. 3C,D), suggesting that the degree of tuning with full bars was underestimated due to saturation effects. Finally, vertically tuned and distal glutamatergic excitation was not observed in triple transgenics in which Cx36 was deleted (Kcng4^Cre × Ai32 × Gjd2^fl/fl; Fig. 3C,D). Taken together with the results from the Cx36 KO experiments, these findings strongly support a primary role for BC5As in mediating vertical excitation to all DSGC types.

Feed-forward cholinergic excitation is also tuned along the vertical axis

Bipolar cells are thought to be shared between SACs and DSGCs (Sethuramanujam et al., 2017), possibly through “dyads” that contain a single presynaptic “ribbon” laden with vesicles positioned across the junction of two postsynaptic processes (Yu et al., 2023). However, if BC input to SACs is tuned, then we might expect the feed-forward cholinergic excitation and GABAergic inhibition to DSGCs to also be tuned.

To test if cholinergic excitation was tuned, we examined the effects of the ACh nicotinic receptor antagonist (Hex) on EPSCs measured in DSGCs (V_HOLD ∼ −60 mV; in the presence of GABA receptor antagonists), which are mediated by a combination of cholinergic/glutamatergic synapses (Sethuramanujam et al., 2016). Since Hex has been shown to have minor presynaptic effects, we reasoned the Hex-sensitive component would provide a reasonable way to measure the tuning properties of cholinergic input (Lee et al., 2010; Sethuramanujam et al., 2016; Hellmer et al., 2021). We found that the Hex-sensitive component of the EPSCs (i.e., control-Hex) was tuned along the vertical axis (Fig. 4A,B). The tuning of the cholinergic component was slightly sharper compared to the tuning of the glutamatergic EPSCs alone (measured after the blockade of cholinergic inputs), likely due to a thresholding effect. Thus, the tuned glutamate excitation appears to drive vertically tuned feed-forward cholinergic excitation through the SAC network.

Figure 4.

Feed-forward cholinergic excitation is vertically tuned. A, Mixed glutamatergic/cholinergic EPSCs measured in control (photoreceptor-driven responses, GABA receptors blocked; black), or in the added presence of a broad-spectrum cholinergic antagonist (100 µM Hex; blue). Note the Hex-sensitive component (red) is tuned along the vertical axis. B, The average peak amplitudes of the EPSCs measured in the conditions shown in A are plotted against bar orientation (n = 13 DSGCs). C, SAC-ChR2-evoked EPSCs (ChAT^Cre × Ai32 mice) recorded in DSGCs in the added presence of photoreceptor blockers (20 µM L-AP4 + 10 µM UBP-310) using 10,000-fold higher stimulus intensities. D, The average peak amplitude of the ChR2-evoked EPSCs is plotted against the stimulus orientation. The SAC network alone provides untuned cholinergic excitation to DSGCs (n = 9 DSGCs).

To confirm that cholinergic tuning relies on the specific properties of the BC circuits, we next examined the properties of cholinergic connections between SACs and DSGCs in isolation using optogenetics. In these experiments, to remove/bypass photoreceptor/BC input, we expressed ChR2 in the SACs (ChAT^Cre × Ai32; Sethuramanujam et al., 2016) and examined EPSCs in DSGCs evoked by oriented bars, after blocking photoreceptors (L-AP4 + UBP-310). Under these conditions, EPSCs were not tuned for orientation, indicating there was no asymmetry intrinsic to the cholinergic circuit (Fig. 4C,D). Thus, we conclude that under control conditions where responses are driven by photoreceptors, the tuned glutamate excitation is effectively translated by BCs into tuned cholinergic excitation through the SAC network.

Inhibitory inputs are differentially tuned in dorsal/ventral and nasal/temporal coding DSGCs

The findings that both glutamatergic and cholinergic excitatory inputs to all DSGC types were tuned along the vertical axis are puzzling in the case of D/V DSGCs, as these cells spike more robustly to stationary horizontal bars. Given that cholinergic excitation is tuned along the vertical axis, it might be expected that inhibitory SAC GABA release would be modulated in the same way. However, potentiating SAC output along the vertical axis would be expected to differentially impact the N/T and D/V DSGCs, owing to differences in the SAC-DSGC wiring asymmetries (Fig. 1B). To directly test this hypothesis, we monitored inhibitory postsynaptic currents (IPSCs) in voltage-clamped DSGCs (V_HOLD ∼ 0 mV; reversal for excitation), which are dominated by GABAergic SAC output (Sethuramanujam et al., 2016).

As expected from the asymmetric SAC wiring, bars oriented along the DSGC's preferred–null motion axis always evoked the largest IPSCs (Fig. 5A,C). However, the degree of orientation tuning was significantly greater in D/V than that in N/T DSGCs (Fig. 5A,C; D/V DSGCs OSI, 0.18 ± 4, n = 17; N/T DSGCs OSI, 0.10 ± 4, n = 11; p = 0.012). Such differences in orientation-selective tuning of the IPSCs are consistent with the idea that the BC5A/WF5A circuit and asymmetric SAC-DSGC circuit intersect in different ways across the population (Fig. 5A). Specifically, for D/V DSGCs the asymmetries in the BC5A/WF5A and SAC circuits both lie along the vertical axis and thus would increase OSI. By contrast, for N/T DSGCs these pathways are orthogonal to each other, which would be expected to decrease OSI (also see Fig. 7).

Figure 5.

Differential tuning of inhibitory inputs across D/V and N/T DSGCs. A, Example IPSCs (V_HOLD ∼0 mV) evoked by oriented bars, recorded in wild-type D/V DSGCs (left) and N/T DSGCs (right). The diagrams show that for D/V DSGCs, the BC5A/WF5A and SAC asymmetries are parallel, while for N/T cells these circuits are orthogonal to each other. Note tuning is stronger in D/V versus N/T DSGCs. B, SAC-only responses evoked using ChR2 (ChAT^Cre × Ai32 mice). The ChR2 responses were recorded in the added presence of photoreceptor blockers (20 µM L-AP4 + 10 µM UBP-310) using 10,000-fold higher light intensities. C, The peak amplitude of the light-evoked IPSCs measured in the D/V (black line) versus N/T (gray) DSGC populations. ** denotes a statistically weaker tuning in N/T DSGCs compared to D/V DSGCs. D, Similar to C, but for SAC ChR2-evoked inhibitory currents. In contrast to photoreceptor-driven responses, the tuning of the ChR2-evoked responses in D/V (dark blue) and N/T coding cells (light blue) is nearly identical. The dotted lines in C and D illustrate the directional tuning properties of the inhibitory currents evoked by moving stimuli (ND denotes the null direction).

Figure 7.

Distinct patterns of inhibition and excitation shape orientation selectivity in dorsal/ventral and nasal/temporal coding DSGCs. A, Top view schematic of the N/T and D/V DSGC circuits highlighting the elements that shape multiple feature selectivity in DSGCs. Direction selectivity originates in the GABAergic/cholinergic starburst amacrine cells (blue), which are spatially offset along the horizontal or vertical axes (as indicated) and provide mixed feed-forward excitatory and inhibitory inputs to DSGCs (black). Orientation selectivity originates in the vertically oriented electrically coupled WF5A/BC5A network (red; also see Fig. 1). This excitatory pathway provides vertically tuned excitation to SACs and DSGCs. Note that the vertically tuned excitatory pathway is orthogonal or parallel to the asymmetries of the SAC circuit for N/T and D/V DSGCs, respectively. B, Average IPSCs evoked by vertical and horizontal bars for N/T (left; n = 11) and D/V DSGCs (right; n = 17). Orientation selectivity of inhibition depends both on the orientation of the SAC wiring as well as on tuned glutamate input. In the case of N/T DSGCs, these factors are not aligned resulting in a reduced orientation selectivity (due to increased inhibition in the nonpreferred orientation; indicated by the vertical arrow on the left). For D/V, the tuned excitation is aligned with the SAC wiring asymmetries and the preferred orientation responses are boosted (vertical arrow on the right), which sharpens the orientation tuning curve for inhibition. C, Average mixed ACh and glutamate receptor-mediated EPSCs evoked by vertical and horizontal bars for N/T (left; n = 6) and D/V DSGCs (right; n = 7). Note that ACh spreads beyond the confines of the synapse and is insensitive to the SAC-DSGC wiring asymmetries (Sethuramanujam et al., 2021). Thus, cholinergic signals help sharpen the excitation along the vertical axis. Note that EPSCs do not reflect the contribution of NMDA receptors, which amplify AMPA and nACh receptor-mediated excitatory signals. D, Spiking responses from example N/T (left) and D/V DSGCs (right) are related to the E/I ratios (average E/I ratios are indicated). By comparing the tuning properties of inhibition and excitation to the DSGC's spiking responses, we infer that orientation selectivity arises through distinct mechanisms in N/T versus D/V DSGCs. For N/T coding DSGCs, vertical bars trigger a maximal excitation (ACh and glutamate; C) and minimum inhibition (B), resulting in robust spike output (D). Horizontal bars trigger a relatively weaker excitation and stronger inhibition, resulting in fewer spikes. Thus, tuned glutamate excitation appears to be the main factor that shapes feature selectivity. By contrast, for D/V coding DSGCs, vertical bars trigger a maximal excitation (C) and maximal inhibition (B), which appear to cancel resulting in little or no spike output (D). Horizontal bars trigger relatively weaker patterns of excitation and inhibition yet generate robust spiking. Thus, tuned SAC inhibition appears critical for feature selectivity.

Next, we examined the properties of inhibitory connections between SACs and DSGCs using an optogenetic approach (ChAT^Cre × Ai32; Fig. 5B,D). When the SAC network alone was stimulated (photoreceptor synapses blocked with L-AP4 + UBP-310), the largest IPSCs were evoked when oriented stimuli were placed along the DSGC's preferred–null axes for motion. However, the IPSCs appeared to be tuned to the same extent in D/V and N/T DSGCs (Fig. 5D; D/V DSGCs OSI, 0.15 ± 2, n = 6; N/T DSGCs OSI, 0.17 ± 2, n = 3; p = 0.69). These results indicate that the wiring asymmetries between SACs and D/V versus N/T DSGCs are similar. Thus, we attribute differences in orientation selectivity observed in these cells under control conditions to arise from vertically tuned BC excitation.

Finally, to test directly whether SACs shape the DSGC's response to oriented bars, we next examined the impact of blocking their output pharmacologically. To do so, we used 5 µM DCG-IV, which is a metabotropic glutamate receptor 2 (mgluR2) agonist that is shown to block SAC output (Sethuramanujam et al., 2018). This is expected to be a relatively selective manipulation in the retina, as mGluR2 receptors are almost exclusively expressed by SACs (Yoshida et al., 2001). As noted previously, the bath application of DCG-IV abolished direction selectivity in DSGCs but did not significantly impact orientation selectivity in N/T coding DSGCs (Fig. 6; Hanson et al., 2023). However, in the case of D/V DSGCs, not only was direction selectivity abolished, but the orientation selectivity was rotated by 90° (Fig. 6A,C). The same D/V DSGCs that preferred horizontal bars in control conditions switched to preferring vertical bars after SAC output was blocked. Hence, SACs appear to play a major role in shaping orientation selectivity in D/V but not N/T DSGCs.

Figure 6.

Starburst amacrine cells are critically required to shape orientation selectivity in dorsal/ventral but not nasal/temporal coding DSGCs. A, Direction tuning curves of D/V DSGCs measured in control (solid line) or after the application of the mGluR2 agonist (5 µM DCG-IV; dotted line), which inhibits SAC output. As expected, SACs are required for the generation of direction selectivity. B, Same as in A but for N/T DSGCs. C, Orientation tuning curves of D/V DSGCs measured in control (solid line) or after the application of the mGluR2 agonist (5 µM DCG-IV; dotted line). Here, blocking SACs “flips” the tuning curve, as in the drug condition the spiking responses are driven mainly by vertically tuned glutamate input. D, Orientation tuning curves of N/T DSGCs measured in control (solid line) or after the application of a mGluR2 agonist (5 µM DCG-IV; dotted line) shows that responses remain tuned along the vertical axis, indicating SAC is not critically required for their orientation selectivity.

Discussion

Direction-selective neurons in diverse species including monkeys, cats, mice, and flies often have a common characteristic: they respond better to static/stationary bar stimuli that are oriented orthogonal to their preferred–null motion axis (Hubel and Wiesel, 1959; Gizzi et al., 1990; Fisher et al., 2015). While initial studies suggested that such selectivity for orientation was a natural outcome of the circuit asymmetries required to generate direction selectivity, more recent studies have indicated that distinct circuit motifs are involved in extracting these features (Weiler et al., 2022; Hanson et al., 2023). In a preceding study, we proposed that for nasal coding, ON–OFF DSGCs in the retina, direction, and orientation selectivity rely on the unique computational properties of distinct neurons, namely, the GABAergic/cholinergic SACs and wide-field amacrine cell (WF5A), respectively. Here, we consolidate and extend the results of our previous study in several important ways. First, we demonstrate that although the D/V and N/T DSGCs exhibit selectivity for horizontal and vertical orientations, respectively, they share vertically tuned glutamate inputs mediated by the putative Cx36-containing wide-field circuit (BC5A-WF5A). Second, tuned glutamate excitation interacts with the SAC circuitry predictably, differentially impacting the tuning of inhibition in D/V and N/T DSGCs, giving rise to distinct mechanisms for generating vertical and horizontal orientation selectivity in these respective cells. Thus, although SACs are conventionally considered motion sensors, our data suggests that they may play an important role in shaping orientation-selective responses evoked by static stimuli, especially in the case of D/V DSGCs.

Postsynaptic mechanisms underlying orientation selectivity

Our results indicate that the excitatory inputs (glutamate and ACh) to all types of DSGCs are tuned along the vertical axis, while the tuning of the inhibitory input (GABA) is always oriented along the DSGC's preferred–null motion axis (Fig. 5C). By comparing the tuning properties of the synaptic inputs with the DSGC's physiological spiking responses (Fig. 7), we infer that multiple distinct strategies are used for extracting orientation information, as observed for other OSGCs (reviewed by Antinucci et al., 2018).

For N/T DSGCs, inhibition and excitation are tuned in a complementary manner; that is, maximum inhibition is evoked during presentations of horizontal bars, and maximum excitation is evoked by vertical bars (Fig. 7B,C). The tuning of the N/T DSGC's spiking output roughly follows that of excitation, suggesting inhibition plays a minor role (Fig. 7D; Hanson et al., 2023). By contrast, for D/V DSGCs, inhibition and excitation are co-tuned (Fig. 7B,C). Vertical bars evoke maximal inhibition and excitation, but inhibition appears dominant, as vertical bars evoke little or no spikes at all in D/V DSGCs. Inhibition evoked by vertical bars is likely boosted, through the feed-forward SAC network, which is driven by a shared vertically tuned source of glutamate excitation (Fig. 7A). Conversely, horizontal bars evoke relatively weak inhibition and excitation in D/V DSGCs, which are sufficient to drive robust spiking in D/V DSGCs. Together, these results indicate that orientation selectivity in N/T DSGCs is shaped primarily through glutamate excitation, while orientation selectivity in D/V DSGCs critically requires SAC inhibitory input.

It is important to note that the tuning properties of the glutamate and ACh inputs were measured under inhibitory receptor blockade, and could potentially be distorted. This prevents the construction of a detailed computational model, similar to the ones that describe the formation of direction selectivity (Tukker et al., 2004). However, the proposed postsynaptic integration scheme underlying orientation selectivity was tested by more selectively blocking SAC output using DCG. Blocking SAC output did not change the tuning properties of N/T DSGCs (Hanson et al., 2023); but for D/V DSGCs, the preferred angle shifted by ∼90°, “flipping” the cell tuning curve from preferring horizontal to vertical bars. These results provide strong support for the idea that distinct synaptic mechanisms underlie orientation selectivity in N/T and D/V DSGCs.

Presynaptic mechanisms underlying orientation selectivity

While our synaptic input measurements provide a general framework of how orientation selectivity is generated at the level of DSGCs, how each of the three transmitter systems acquires their distinct tuning properties is less clear. Below we discuss the distinct mechanisms that are involved in shaping the tuning properties of glutamate, ACh, and GABA inputs to DSGCs.

Our results directly demonstrated that glutamate inputs to all four DSGC types are vertically tuned, refuting the simple hypothesis that the vertical and horizontal tuning observed in the different types of DSGCs arise from distinct patterns of excitation. It should also be noted that many of our experiments were performed in the presence of inhibitory receptor blockers, which could have abolished potential GABA-dependent horizontal tuning of BC terminals (Johnston et al., 2019). However, interpreting the directional tuning properties of BCs based on EPSCs in the absence of inhibitory blockers is difficult due to space-clamp issues (Poleg-Polsky and Diamond, 2011), and thus we felt it was necessary to block GABAergic inhibition in our experiments.

How bipolar cells acquire vertical tuning is only beginning to be understood. Recently BC5As were shown to receive a profuse synaptic input (30/40 synapses/terminal) from the vertically oriented processes of specific wide-field amacrine cells (WF5A; Hanson et al., 2023). As the orientation of the WF5A process matched the functional tuning properties of BC5A, these cells were hypothesized to be electrically coupled through Cx36 containing gap junctions (Hanson et al., 2023). Our current study provides three additional lines of support for the BC5A-WF5A coupling mechanism for orientation tuning. First, we found that the vertical tuning of EPSCs in all DSGCs does not rely on inhibition, consistent with the pharmacological profile of BC5A tuning properties (Hanson et al., 2023). Second, when BC5As were directly stimulated using ChR2 (photoreceptors and inhibitory synapses blocked) glutamatergic EPSCs could be evoked along the vertical axis at considerable distances from the DSGC’s receptive field center (Fig. 3). As WF5As also receive a dominant glutamate input from BC5A, the most plausible explanation for distal excitation is that it is mediated by the reciprocal BC5A-WF5A glutamate/gap junction connections. It is also possible that BC-BC coupling plays a role, although this would require asymmetrical electrical connections to be made selectively along the vertical axis and spanning a long distance, which remains to be demonstrated.

Third, EPSC tuning was absent in all DSGCs in the transgenic mouse line in which Cx36 is deleted in BC5As (and possibly a few other cell types; Duan et al., 2014; Greene et al., 2016). We found that the EPSC amplitude was much reduced in mutant mice in which Cx36 was deleted from BC5As, compared to their wild-type littermates, which is puzzling, given that BC5B, BC5C, and BC7 are also likely to contribute to the DSGC's ON EPSC. The constitutive deletion of Cx36 may affect the proper development of the circuit, leading to reduced responses. Alternatively, it is also possible that nonlinear synergistic interactions between BC5 types mediated via Cx36-containing gap junctions are required for their robust activation (Kuo et al., 2016; Tsukamoto and Omi, 2017). Future studies in which Cx36 is deleted in other BC types will address this question. Nevertheless, given that the EPSCs in the Cx36 KO were untuned, this indicates that the other BC types in the circuit are untuned. If they were tuned, then the residual EPSCs measured in the BC5A-specific Cx36 KO mouse would be expected to exhibit similar tuning properties.

Since SACs and DSGCs are driven by an overlapping set of BCs, it is expected that their output would also be vertically tuned. Indeed, when we examined the cholinergic component of the EPSCs, we found it to be tuned along the vertical axis. Thus, the SAC's output appears to reflect the vertical tuning properties of their BC input. Vertically tuned excitation was not observed in optogenetic experiments in which we stimulated the SAC network in relative isolation suggesting that the cholinergic wiring is functionally symmetrical (Lee et al., 2010; Yonehara et al., 2011; Sethuramanujam et al., 2016).

As SACs co-release ACh and GABA, at first pass, it might be expected that they share similar tuning characteristics. However, ACh spreads beyond the confines of the wrap-around synapse and can activate multiple DSGC dendrites in the vicinity, and thus the patterns of cholinergic excitation are insensitive to the asymmetry in the SAC-DSGC “hard-wiring” (Sethuramanujam et al., 2021). By contrast, GABAergic signals are synaptic and strongly influenced by SACs-DSGC wiring asymmetries. Specifically, the strongest inhibition was always evoked by bars that aligned with the DGSC's preferred–null motion axes, along which maximal SAC-DSGC contacts are made (Briggman et al., 2011).

We also observed that the IPSCs were more weakly tuned for orientation in N/T versus D/V DSGCs. This difference in tuning could not be explained by subtle differences in connectivity since mapping the connections functionally by directly stimulating SAC cells revealed that the tuning properties were identical between N/T and D/V DSGCs. If we consider SAC GABAergic output was boosted for vertical bars, then OS inhibition in D/V DSGCs would increase (as vertical bars align with their preferred–null axis), while OS inhibition in N/T DSGCs would be reduced (vertical bars boost orthogonal responses). Thus, the gross tuning patterns of the inhibitory currents in all DSGCs can be best understood by considering the known asymmetric wiring patterns of the SACs and DSGCs and vertically tuned excitation. Regardless of the precise mechanism, the stronger IPSC tuning observed in D/V DSGC suggests that it plays a more important role in shaping its selectivity compared to N/T DSGCs.

Conclusions

A large number of studies have investigated the spatiotemporal tuning properties of glutamate inputs to DSGCs and SACs (Taylor and Vaney, 2002; Fried et al., 2005; Poleg-Polsky and Diamond, 2011; Yonehara et al., 2013; Park et al., 2014; Sethuramanujam et al., 2016; Matsumoto et al., 2021; Gaynes et al., 2022; Srivastava et al., 2022; Strauss et al., 2022). However, most studies examined BC function in the context of direction selectivity and have generally overlooked the pathway that mediates excitation from the far surround (BC5A-WF5A circuit; Hanson et al., 2023). While the results from the current study demonstrate that lateral excitation shapes responses to static-oriented bars, either directly or indirectly through SACs, the extent to which this pathway is recruited by moving stimuli remains an open question. Intuitively, it would be expected that wide-field modulation of excitation would influence the DSGC's tuning properties in a contextual manner (e.g., tuning may depend on stimulus orientation and or velocity). How a potentially adaptive tuning curve could be advantageous over one that is generally believed to be relatively invariant across stimulus parameters is an exciting topic for future investigations.