Compartmentalized pooling generates orientation selectivity in wide-field amacrine cells Authors: Wanyu Lei, Damon A. Clark, Jonathan B. Demb Published: 2024 ABSTRACT SignificanceOrientation detection is a fundamental computation performed by the visual system. In the retina, a specialized group of giant, nonspiking interneurons, known as B/K wide-field amacrine cells, encode orientation within their dendrites. However, the underlying mechanism remains unresolved. Here, we combine two-photon calcium imaging, electrophysiological recording, and computational modeling to reveal the biophysical mechanism of dendritic integration that generates orientation selectivity in these cells. This work sheds light on the visual feature computation carried out by enigmatic wide-field amacrine cells in the retina and advances our understanding of the strategies employed by the visual system to detect orientation. Orientation selectivity represents a fundamental computation in the visual system, as classically described by the receptive field properties of neurons in the primary visual cortex (V1) (1). Orientation-selective (OS) neurons are found at subsequent stages of cortical processing (2–5) and at precortical stages as well (6–8). Indeed, orientation selectivity at both cortical and precortical levels can be linked to OS neurons in the retina (6–8), which includes certain types of output neurons, retinal ganglion cells (RGCs) (9–15), as well as retinal interneurons, bipolar, and amacrine cells (16–22). Understanding the mechanisms for orientation selectivity in individual cell types, at each stage of the visual system, promises to reveal both unique and common strategies for a basic visual computation across diverse neuronal circuits (23). To detect orientation, a neuron must sample anisotropically across two dimensions of visual space. Specifically, a neuron’s OS receptive field must differ from its perpendicular (i.e., 90° rotated) counterpart to respond differently to two orthogonal orientations. OS receptive fields could be constructed using several designs, with varying degrees of integration along the preferred orientation (PO) axis and varying ratios of excitation-to-inhibition along the orthogonal, null orientation (NO) axis. For example, an OS receptive field could be elongated along the PO axis either with or without center-surround antagonism along the NO axis (17, 24) (SI Appendix, Fig. S1 A and B). Alternatively, an OS receptive field could be relatively confined along the PO axis, with either even-symmetric or odd-symmetric antagonism along the NO axis (19, 25) (SI Appendix, Fig. S1 C and D). The geometric diversity of OS receptive fields underscores the importance of characterizing the receptive field along both PO and NO axes in order to reveal the computational strategy for generating orientation selectivity. Here, we investigated the mechanism for orientation selectivity where it appears most peripherally in the visual system: retinal interneurons. We studied orientation selectivity in the largest interneurons in the mouse retina, wide-field amacrine cells (WACs), and targeted specific types of WACs using an intersectional genetic strategy relying on the coexpression of the transcription factor Bhlhe22 (B) and the Kappa Opioid Receptor (K) (22). The B/K WACs do not fire action potentials, and their dendrites contain both synaptic inputs and outputs. These dendrites are long, straight, and unbranching, extending up to 1 mm from the soma, with OS tuning aligned with their physical orientations (22). To understand the mechanism underlying the orientation selectivity in B/K WACs, we used simultaneous two-photon calcium imaging and patch clamp recording to characterize the receptive field of B/K WAC dendrites along the PO and NO axes and identified the computational strategy that generates orientation selectivity. We further illuminated the underlying biophysics by recording the excitation and inhibition converging onto B/K WACs and by computational modeling. Our results reveal how B/K WAC dendrites act as orientation detectors and broaden our understanding of how dendritic mechanisms contribute to sensory computations (26, 27). Results Orientation Selectivity of B/K WACs at the Single-Cell Level. B/K WACs were targeted using an intersectional genetic strategy with transgenic lines that express two recombinases (Bhlhe22-Flpo, KOR-Cre) crossed with a dual (Flp + Cre) reporter (dual-tdTomato; see SI Appendix, Methods). B/K WACs comprise ON- and OFF-types that stratify dendrites in either ON- or OFF-levels of the inner plexiform layer (IPL) (Fig. 1A) (22). Population calcium imaging of B/K dendrites revealed OS dendritic segments in both ON and OFF levels of the IPL (22). However, due to the dense labeling, it was not clear whether the orientation selectivity was confined to certain locations along each dendrite. To better understand the OS calcium responses, we loaded a high signal-to-noise ratio inorganic calcium indicator, Cal-520 (28), into individual B/K WACs in the ganglion cell layer (GCL) and studied light-evoked dendritic calcium responses at the single-cell level with two-photon imaging (Fig. 1B). B/K WACs exhibit orientation selectivity along the dendrites. (A) Schematic of ON and OFF B/K WAC dendrite stratification, relative to the ChAT bands (Left), and typical dendritic tree morphology (Right), data from ref. 22. B/K WACs studied here had cell bodies in the GCL. INL, inner nuclear layer. (B) Schematic of the single-cell dendritic calcium imaging method (Top) and an example field-of-view containing a dendritic ROI (Bottom). (C) Calcium responses evoked by drifting gratings across eight different directions for ON (Top) and OFF (Bottom) B/K WAC dendritic ROIs. Data are presented as mean ± SEM (shaded area) across ROIs, after aligning the PO to the horizontal axis. Responses indicate percentage changes in fluorescence from baseline (dF/F). (D) OSI of dendritic ROIs at different distances from the soma. (E) OSI of ROIs from ON and OFF B/K WAC dendrites in (D). Group data (mean ± SEM) are superimposed with individual data points. OSI: ON WACs, 0.64 ± 0.05, n = 15 ROIs (8 cells, 3 mice); OFF WACs, 0.9 ± 0.03, n = 16 ROIs (9 cells, 8 mice). ***P < 0.001, unpaired t test. (F) Polar plots of the mean grating-evoked dF/F in C for ON and OFF B/K WAC dendritic ROIs. Responses are normalized to the maximum dF/F. (G) Schematic illustrating the physical orientation and PO of the dendritic ROI. The physical orientation is defined as the counterclockwise angle (θ) between the nasal–temporal axis (black dashed line) and the physical dendrite (blue line). The PO is calculated as the counterclockwise angle (θ’) between the nasal–temporal axis and the orientation of the vector sum (yellow arrows, yellow dashed line) of responses in the orientation space (green arrows). (H) The difference between the physical orientation and PO of the ON and OFF B/K WAC dendrites. Group data (mean ± SEM) are superimposed with individual data points: ON B/K WAC, −1.96 ± 3.16°, 15 ROIs (8 cells, 3 mice), one-sample t test comparing to zero, P = 0.54; OFF B/K WAC, 0.57 ± 0.83°, n = 15 ROIs (8 cells, 3 mice), one-sample t test comparing to zero, P = 0.50. The ON B/K WACs in the GCL typically extend 7-13 dendrites in a roughly circular pattern, while the OFF B/K WACs extend 4-6 dendrites in an oriented pattern (Fig. 1A) (22). To assess their dendritic orientation selectivity, we recorded calcium responses in dendritic regions of interest (ROIs) at different distances from the soma, while presenting drifting grating stimuli in multiple directions (4.5 cycles/mm, 2 Hz, 440 µm in diameter, centered on the ROI). All responsive dendritic ROIs exhibited orientation selectivity regardless of their locations on the dendrite, which was quantified with an orientation selectivity index (OSI; see SI Appendix, Methods) (Fig. 1 C and D). The ROIs in the OFF cells demonstrated slightly sharper orientation tuning with a higher OSI compared to the ROIs in the ON cells (Fig. 1 E and F). In both cases, the PO aligned with the physical orientation of the dendrite (Fig. 1 G and H), demonstrating a strong structure–function alignment, consistent with results from population imaging (22). B/K WAC Dendritic Receptive Field Extends along the PO Axis Yet Terminates at the Soma. To probe the mechanisms underlying the dendritic orientation selectivity of the B/K WACs, we started by mapping the receptive field along the PO axis. Specifically, we measured the integration length along the dendrite, which defines the degree of signal pooling. We imaged dendritic Ca2+ signals at an ROI location ~220 µm from the soma, aligned with stimulus location 5 (Fig. 2A), while simultaneously recording somatic voltage. Responses were elicited by drifting grating patches (4.5 cycles/mm, 2 Hz, 220-µm diameter) presented at eleven stimulus locations (Fig. 2A). The centers of the grating patches were positioned along the PO axis of the recorded dendrite, starting on the opposite side as the dendritic ROI, 220-µm from the soma (location 1), and continuing onto the same side as the ROI, 880-µm from the soma (location 11; 110-µm intervals). The integration length of each ROI was calculated as the difference between the most distant locations with significant grating-evoked Ca2+ responses (i.e., >2 SD above the noise level; see SI Appendix, Methods). Average integration lengths were ~340 µm for ON B/K WACs and ~570 µm for OFF B/K WACs (Fig. 2 B and D), which both exceed the ~100-µm width of the receptive field of a presynaptic (excitatory) bipolar cell (29–31). Thus, individual ON and OFF B/K WAC dendrites integrate inputs from multiple bipolar cells. Dendritic receptive fields of B/K WACs are elongated yet compartmentalized along the PO axis. (A) Simultaneous recording of dendritic calcium (green traces), fixed at stimulus location 5, and somatic voltage (blue traces) responses evoked by a local drifting grating patch at eleven stimulus locations along the dendrite for ON and OFF B/K WACs. Data are presented as mean ± SEM (shaded area) across ROIs. (B) Integration length of dendritic calcium fluorescence (Fden) shown in (A) for ON B/K WACs. Group data (mean ± SEM) are superimposed with individual data points. Integration length: 339 ± 50 µm, n = 10 ROIs (7 cells, 4 mice). (C) Compartmentalization index of dendritic calcium (Fden) and somatic voltage (Vsoma) shown in (A) for ON B/K WACs. Group data (mean ± SEM) are superimposed with individual data points. Compartmentalization index: Fden, 0.82 ± 0.08, n = 10 ROIs (7 cells, 4 mice); Vsoma, 0.04 ± 0.12, n = 10 ROIs (7 cells, 4 mice). ***P < 0.001, paired t test. The compartmentalization index was computed separately for each measurement (Results); simultaneous measurements are connected by a line. (D) Same as (B) for OFF B/K WACs. Integration length: 568 ± 62 µm, n = 12 ROIs (11 cells, 9 mice). (E) Same as (C) for OFF B/K WACs. Compartmentalized index: Fden, 0.55 ± 0.11, n = 12 ROIs (11 cells, 9 mice); Vsoma, 0.14 ± 0.09, n = 12 ROIs (11 cells, 9 mice). *P < 0.05, paired t test. Notably, the dendritic Ca2+ response integrated along the physical extent of the imaged dendrite but stopped at the soma. By contrast, the somatic voltage response integrated more narrowly and roughly symmetrically about the soma (Fig. 2A). To quantify the degree of compartmentalization of each signal within the imaged dendrite, we analyzed responses to stimuli at regions equidistant to the recording location. For the dendritic ROI (location 5) we compared responses that were two to four steps away on either side of the ROI (i.e., locations 1 to 3 vs. locations 7 to 9). For the voltage recorded at the soma (location 3) we compared responses that were one to two steps away on either side of the soma (i.e., locations 1 to 2 vs. locations 4 to 5). In each case, a compartmentalization index (SI Appendix, Methods), varying between 0 (no compartmentalization) and 1 (complete compartmentalization), showed significantly higher values for the dendritic ROI as compared to the somatic voltage (Fig. 2 C and E). These data suggest that the dendritic Ca2+ signal is compartmentalized within each dendrite, operating independently of the somatic voltage, whereas somatic voltage integrates symmetrically from all dendrites, as expected. The compartmentalization of the dendritic Ca2+ signal ensures that B/K WAC dendrites with differing orientation preferences remain independent of one another, preserving the fidelity of orientation selectivity within individual dendrites. This mechanism seems particularly crucial for ON B/K WACs, which commonly extend dendrites at orthogonal orientations (Fig. 1A). The results above demonstrate that B/K WAC dendrites exhibit a spatially confined yet extended receptive field along the dendrite, which we term “compartmentalized pooling.” B/K WAC Dendritic Receptive Field along the NO Axis Exhibits Center-Surround Antagonism. Next, we investigated the B/K WAC dendritic receptive field structure along the NO axis. Specifically, we tested whether the receptive field along the dimension perpendicular to the physical dendrite was either purely excitatory or a mix of excitatory and inhibitory components (SI Appendix, Fig. S1 A and B). We presented drifting gratings of various spatial frequencies (1.1 to 18.2 cycles/mm) in the PO, centered on the dendritic ROI (Fig. 3A), and measured receptive field properties in the frequency domain for both dendritic Ca2+ and somatic voltage responses. Consistent with the findings above, the dendritic Ca2+ responses and somatic voltage responses showed distinct tuning, indicating functional independence between these cellular compartments. For example, ON B/K WACs showed the strongest Ca2+ response for a PO grating at the middle spatial frequency but the largest somatic depolarization to a PO grating at the lowest spatial frequency. We proceeded to characterize the tuning properties of the dendritic signals. In both ON and OFF B/K WACs, the evoked dendritic Ca2+ signal exhibited band-pass spatial frequency tuning (Fig. 3 B and G), consistent with an antagonistic center-surround structure of the receptive field along the NO axis. Spatial frequency tuning profiles of B/K WAC dendrites reveal center-surround antagonism along the NO axis. (A) Simultaneous recording of dendritic calcium (dF/F; green traces) and somatic voltage (blue traces) evoked by gratings of different spatial frequencies (1.1, 2.3, 4.5, 9.1, and 18.2 cycles/mm; half-cycle width: 440, 220, 110, 55, and 22.5 µm) at the PO and the NO for ON and OFF B/K WACs. Data are presented as mean (line) ± SEM (shaded area) across ROIs. (B) Spatial frequency tuning profile of mean calcium responses (dF/F) of ON B/K WACs. A DoG model was fitted to the mean tuning curve for PO gratings. Data are presented as mean ± SEM (shaded area). *P < 0.05, **P < 0.01, ***P < 0.001, paired t test. (C) Spatial frequency tuning profile of the maximum voltage relative to the resting membrane potential (dVmax) of ON B/K WACs. Data are presented as mean ± SEM (shaded area). *P < 0.05, **P < 0.01, ***P < 0.001, paired t test. (D) Profiles of the fitted center and surround receptive fields along the NO axis for ON B/K WACs. The DoG model was fitted with the mean tuning curve for PO gratings shown in (B). (E) FWHM of the fitted center and surround receptive fields of individual dendrite ROIs in ON B/K WACs. Group data (mean ± SEM) are superimposed with individual data points. FWHM: center, 78 ± 12 µm; surround, 162 ± 18 µm, n = 8 ROIs (7 cells, 4 mice). *P < 0.05, paired t test. (F) The strength ratio of the center weight relative to the surround weight in the DoG model of ON B/K WACs. Group data (mean ± SEM) are superimposed with individual data points. Strength ratio: 0.92 ± 0.05, n = 8 ROIs. (G) Same as (B) for OFF B/K WACs. (H) Same as (C) for OFF B/K WACs. (I) Same as (D) for OFF B/K WACs. (J) Same as (E) for OFF B/K WACs. FWHM: center, 68 ± 9 µm; surround, 174 ± 30 µm, n = 10 ROIs (8 cells, 7 mice). *P < 0.05, paired t test. (K) Same as (F) for OFF B/K WACs. Strength ratio: 1.10 ± 0.07, n = 10 ROIs. To estimate the spatial extent of the center and surround receptive fields, we fitted the spatial-frequency tuning curve of the dendritic Ca2+ signal to a difference-of-Gaussians (DoG) model (Fig. 3 D and I). For the ON B/K WACs, the full-width-at-half-maximum (FWHM) of the center was 78 ± 12 µm (mean ± SEM) and that of the surround was 162 ± 18 µm (Fig. 3E). The ratio of the center and surround Gaussian weights was close to one (Fig. 3F). For OFF B/K WACs, the FWHM values for the center and surround were 68 ± 9 µm and 174 ± 30 µm, respectively (Fig. 3E), and the center-to-surround weight ratio was similarly balanced (Fig. 3K). Notably, the FWHM of the receptive field center for both ON and OFF B/K WACs matched the spatial scale of a bipolar cell (~100 µm) (29–31), which is much smaller than the spatial extent of the integration fields (~300 to 600 µm) of B/K WACs along the PO axis (Fig. 2). Combined with our prior measurements, these data suggest that both ON and OFF B/K WACs possess elongated center-surround antagonistic dendritic receptive fields. To further characterize the receptive field structure, we leveraged the stable somatic voltage recordings from OFF B/K WACs, which have an overall oriented dendritic tree (Fig. 1A) and corresponding PO. We used a 1-dimensional white noise stimulus composed of bars elongated in the PO and mapped the receptive field of the OFF B/K WAC in the spatial domain using system identification methods (Fig. 4 B–D). The stimulus consisted of three 110- µm by 440-µm binary flickering bars, covering the “center” (C) and the flanking “surround” (S1, S2) (Fig. 4A). The center and surround linear receptive field kernels were extracted using regularized linear regression. The extracted kernels showed center-surround antagonism, with two symmetric surround lobes of similar strength and kinetics (Fig. 4 E and F). These data support the three-lobe model consisting of a symmetric flanking surround, consistent with the DoG model and distinct from a two-lobe model observed in some V1 cells (25) (Fig. 4G). Additionally, the surround kernels were significantly slower than the center kernel (Fig. 4F), suggesting a delayed amacrine cell and/or horizontal cell contribution to the surround. Center-surround receptive field structure of OFF B/K WACs. (A) Schematic illustrating the 3-bar white-noise receptive field mapping of the OFF B/K WAC, with the somatic voltage recorded. (B) Spatiotemporal profiles of the extracted center (C) and surround kernels (S1, S2) of an exemplary OFF B/K WAC. (C) Performance of the linear kernel model. The model prediction and the raw data of the exemplary OFF B/K WAC in (B) are superimposed. (D) Temporal profiles of the center and surround kernels. Data are presented as mean ± SEM (shaded area). (E) Peak strength (a.u.) of the center and surround kernels. Group data (mean ± SEM) are superimposed with individual data points. Peak strength: C, 0.39 ± 0.08, n = 5 cells (3 mice); S1: 0.06 ± 0.01, n = 5; S2: 0.06 ± 0.04, n = 5. n.s. P > 0.05, *P < 0.05, **P < 0.01, paired t test. (F) Time-to-peak of the center and surround kernels. Group data (mean ± SEM) are superimposed with individual data points. Time-to-peak: C, 96.7 ± 2.4 ms, n = 5; S1: 112.7 ± 2.7 ms, n = 5 cells (3 mice); S2: 111.3 ± 4.0 ms, n = 5. n.s. P > 0.05, *P < 0.05, **P < 0.01, paired t test. (G) The extracted kernels support the three-lobe receptive field model with symmetric flanking antagonistic surrounds (Top), and falsify the two-lobe receptive field model with single-sided antagonism (Bottom). Spatial Frequency Tuning of Orientation Selectivity in B/K WAC Dendrites. To determine the robustness of orientation selectivity across different spatial frequencies, we also recorded dendritic Ca2+ and somatic voltage responses to NO gratings (Fig. 3A). Both ON and OFF B/K WAC dendrites exhibited orientation sensitivity over 3-4 octaves of spatial frequencies (Fig. 3 B and G). Notably, NO gratings failed to evoke a dendritic Ca2+ response, even at the low spatial frequency (Fig. 3 A, B, and G). Despite the lack of dendritic Ca2+ responses to NO gratings, ON B/K WACs exhibited modulated somatic voltage responses at lower spatial frequencies (Fig. 3 A and C). In particular, the depolarization evoked by the lowest- frequency grating (1.1 cycles/mm) in NO was comparable to that evoked by the middle-frequency gratings (2.3 to 4.5 cycles/mm) in PO. This discrepancy between somatic voltage response and the dendritic Ca2+ responses reinforces the functional independence of the two compartments. For OFF B/K WACs, there were prominent hyperpolarizations to NO gratings (Fig. 3 A and H), consistent with a direct inhibitory synaptic input, which we tested further below. Compartmentalized Pooling Establishes Orientation Selectivity in B/K WAC Dendrites. Receptive field mapping suggests that B/K WAC dendrites possess a center-surround receptive field extending along the dendrite and terminating at the soma (Figs. 2–4). Using the measured parameters, we constructed a phenomenological receptive field model for the dendritic Ca2+ response. The voltage output of an R0I at a vertical dendrite was modeled as the convolution between the receptive field and the visual stimulus (drifting gratings at various spatial frequencies). A voltage threshold was applied to model the nonlinear activation of voltage-gated Ca2+ channels. To understand how the key receptive field properties contribute to orientation selectivity, we manipulated various aspects of the receptive field structure in silico (Fig. 5A), as direct experimental alterations were less feasible. Compartmentalized pooling generates orientation selectivity in B/K WAC dendrites. (A) Simulated responses of receptive field (RF) models for an ROI at a vertical dendrite of a simplified WAC with three dendrites. Control model: “B/K dendritic RF”; In silico manipulated RF models: “no pooling,” “no antagonistic surround,” “no electrotonic isolation.” Left: RFs of different models. Red indicates positive weight; blue indicates negative weight. Right: Model responses to NO and PO gratings at optimal spatial frequency (4.5 cycles/mm) and low spatial frequency (1.1 cycles/mm) are shown as blue traces. The dashed gray line represents resting membrane potential, and the teal line represents the voltage-gated calcium channel (Cav) activation threshold. (B) Simulated spatial frequency tuning curves of the OSI for different RF models shown in (A). The spatial frequencies used in (A) are indicated by dashed lines. To assess the role of the extended pooling along the dendrite, we decreased the integration length so that the dendrite gathered input from just one excitatory input (i.e., bipolar cell) with a circular receptive field. This extremely narrow integration is biologically plausible, as it has been observed in the dendrites of the A17 amacrine cell (32). Here, the limited integration led to a small membrane depolarization that barely reached the activation threshold of the voltage-gated Ca2+ channel (Fig. 5A), and, of course, even the subthreshold depolarization of the circular receptive field structure lacked orientation selectivity across spatial frequencies (Fig. 5B). These findings underscore the importance of extended dendritic pooling in generating orientation selectivity in B/K WAC dendrites. We next eliminated the inhibitory surround Gaussian in the DoG model, to assess its role in orientation selectivity. The elongated excitatory center itself generated orientation selectivity. However, the center-only model had diminished selectivity at low spatial frequencies, which was primarily explained by an increased Ca2+ response to the NO grating (Fig. 5 A and B). Thus, the inhibitory surround sharpens orientation selectivity at lower frequencies by suppressing null-orientation responses. To assess the role of dendritic compartmentalization, we incorporated a receptive field driven by orthogonal (horizontal) dendrites on either side of the soma into the vertical dendrite’s receptive field (Fig. 5A), which simulated a condition where dendrites lack electrotonic isolation. This manipulation disrupted orientation selectivity, due to “leakage” depolarization from the orthogonal dendrites, which themselves have orthogonal orientation selectivity (Fig. 5 A and B). This suggests compartmentalization is essential for generating orientation selectivity in B/K WAC dendrites, particularly for ON B/K WACs, which possess orthogonally oriented dendrites. But even for OFF B/K WACs, dendrites on one side of the soma commonly span 45 degrees, and the absence of compartmentalization could likewise reduce orientation tuning. Taken together, compartmentalized pooling along the dendrite is the key to generating orientation selectivity in B/K WAC dendrites. The antagonistic surround is not a prerequisite for this selectivity, but it sharpens the tuning at lower spatial frequencies. EI Coordination Underlies Orientation Selectivity in OFF B/K WACs. To gain insight into synaptic mechanisms underlying the orientation selectivity of B/K WACs, we performed whole-cell voltage clamp recordings in OFF B/K WACs (Fig. 6), since their somatic voltage responses were OS (Fig. 3). Due to the oriented morphology of OFF B/K WACs, the somatic excitatory postsynaptic currents (EPSCs) and inhibitory PSCs (IPSCs) could provide a qualitative reflection of the synaptic inputs at the dendrite. We defined the PO of an OFF B/K WAC as its average dendritic orientation and assumed the NO is orthogonal to the PO. Coordination of excitation and inhibition underlying orientation selectivity in OFF B/K WACs. (A) EPSCs and IPSCs of OFF B/K WACs evoked by gratings across different spatial frequencies in the PO and the NO. Data are presented as mean ± SEM (shaded area) across cells. (B) Definition of peak EPSC and EPSC amplitude based on the peak and trough locations of the F1 component (green line) of the EPSC trace. (C) Spatial frequency tuning profiles of the peak EPSC. Data are presented as mean ± SEM (shaded area), n = 4 cells (3 mice). *P < 0.05, **P < 0.01, ***P < 0.001, paired t test. (D) Spatial frequency tuning profiles of EPSC amplitude. Data are presented as mean ± SEM (shaded area), n = 4 cells (3 mice). At several spatial frequencies, amplitudes were larger for the PO. *P < 0.05, **P < 0.01, ***P < 0.001, paired t test. (E) Definition of phasic and tonic IPSCs based on peak and trough locations of the F1 component (green line) of the IPSC trace. (F) Spatial frequency tuning profiles of tonic IPSCs. Data are presented as mean ± SEM (shaded area), n = 6 cells (5 mice). *P < 0.05, **P < 0.01, ***P < 0.001, paired t test. (G) Spatial frequency tuning profiles of phasic IPSCs. Data are presented as mean ± SEM (shaded area), n = 6 cells (5 mice). *P < 0.05, **P < 0.01, ***P < 0.001, paired t test. (H) Definition of the phase difference between the IPSC (outward current) and EPSC (inward current) evoked by a PO grating, calculated as the phase shift between their F1 components (green lines). (I) Phase difference between EPSC and IPSC across three major spatial frequencies. Group data (mean ± SEM) are superimposed with individual data points, n = 4 cells (3 mice). At medium to low spatial frequencies (1.1 to 4.5 cycles/mm), the PO gratings elicited small EPSCs with periodic modulation. Both the peak (inward) and peak-to-trough amplitude of the EPSC exhibited orientation selectivity aligned with the orientation of the cell’s dendritic tree. (Fig. 6 B–D), consistent with the tuning of the dendritic Ca2+ response (Fig. 3G). By comparison, the IPSCs were much larger than the EPSCs and could be decomposed into two components: a tonic IPSC, which was a sustained outward current, and a phasic IPSC, which was a periodic outward current riding on top of the tonic current (Fig. 6E). The tonic IPSC evoked by the NO grating was consistent with the tonic hyperpolarization observed in the somatic voltage (Fig. 3 A and H). Interestingly, the tonic IPSC evoked by the NO grating was larger than that evoked by the PO grating (Fig. 6F), indicating an orientation selectivity of the IPSC that was orthogonal to that of the EPSC. Moreover, the phasic IPSC evoked by the PO grating was slightly larger than that evoked by the NO grating revealing an orientation selectivity aligned with that of the EPSC (Fig. 6G). Notably, the phasic IPSCs and EPSCs elicited by the PO grating were out-of-phase (Fig. 6 H and I), which broadens the dynamic range for voltage responses to PO gratings. Glycinergic Crossover Inhibition Is Not Required for Generating Orientation Selectivity in B/K WACs. One possible mechanism underlying the out-of-phase relationship between EPSCs and phasic IPSCs evoked by PO gratings (Fig. 6 H and I) is that excitation and inhibition are mediated by distinct OFF and ON bipolar cell pathways, respectively. The dark bars of the gratings would evoke excitatory inputs, while the bright bars would evoke inhibitory inputs. Inhibition from the ON (OFF) pathway to OFF (ON) WACs would represent crossover inhibition, a circuit motif commonly mediated by narrow-field glycinergic amacrine cells (33–37). To examine the presence of crossover inhibition in B/K WACs, we recorded IPSCs as well as EPSCs evoked by a contrast-reversing spot (440-µm diameter, 2 Hz) centered at the soma (Fig. 7A). The contrast-reversing spot stimulus provides a clear distinction between the response to positive and negative contrasts. In ON B/K WACs, negative contrast induced pronounced IPSCs, while positive contrast induced EPSCs (Fig. 7B). The IPSCs were significantly blocked by strychnine, whereas the EPSCs were not significantly affected (Fig. 7C). Similarly, for OFF B/K WACs, positive contrast induced pronounced IPSCs, while negative contrast induced small EPSCs. The IPSCs were blocked by strychnine, whereas the EPSCs were not significantly affected (Fig. 7 D and E). These results provide direct evidence that inhibition onto B/K WACs is primarily mediated by glycinergic crossover inhibition. Orientation selectivity of B/K WAC dendrites persists in the absence of glycinergic crossover inhibition. (A) Schematic of the recording setup for measuring EPSCs and IPSCs evoked by a contrast-reversing spot from the ON B/K WAC (Top) and the OFF B/K WAC (Bottom). (B) Traces of IPSCs (Middle) and EPSCs (Bottom) evoked by a contrast-reversing spot (contrast shown at Top) of ON B/K WACs in control and in the presence of 1 µm Strychnine. Data are presented as mean ± SEM (shaded area) across cells. (C) Group data of IPSCs (Top) and EPSCs (Bottom) shown in (B). Data from control (CT) and strychnine (ST) conditions are presented as mean ± SEM, superimposed with individual data points. IPSCs: CT, 135 ± 20 pA, n = 5 cells (5 mice); ST, 18 ± 3 pA, n = 5 cells (5 mice). **P < 0.01, paired t test. EPSCs: CT, −30 ± 8 pA, n = 5; ST, −36 ± 11 pA, n = 5. P = 0.3, paired t test. (D) Same as (B) for OFF B/K WACs. (E) Same as (C) for OFF B/K WACs. IPSCs: CT, 190 ± 62 pA, n = 6 cells (6 mice); ST, 20 ± 8 pA, n = 6 cells (6 mice). *P < 0.05, paired t test. EPSCs: CT, −8 ± 3 pA, n = 6 cells (6 mice); ST, −3 ± 2 pA, n = 6 cells (6 mice). P = 0.09, paired t test. (F) GCaMP7s responses of example dendritic ROIs in the ON B/K WAC dendritic layer to drifting gratings in the control condition (Left; data from ref. 22) and in the presence of 1 µm Strychnine (Right). ROIs are outlined on the correlated image, with the corresponding raw fluorescence traces (in arbitrary units, a.u.), response polar plots, and values of the OSI shown on the Right. (G) OSI of ROIs from the ON B/K dendritic layer in the control condition (CT) and in the presence of 1 µm Strychnine (ST). Group data (mean ± SEM) are superimposed with individual data points. CT: 0.71 ± 0.03, n = 32 ROIs (6 regions, 6 retinas, 4 mice); ST: 0.58 ± 0.04, n = 26 ROIs (6 regions, 5 retinas, 4 mice). *P < 0.05, unpaired t test. (H) Statistics of the angle difference between the physical orientation and the PO of ROIs with OSI > 0.5 from the ON B/K dendritic layer in the control condition (CT) and in the presence of 1 µm Strychnine (ST). Group data (mean ± SEM) are superimposed with individual data points. CT: 1.78 ± 2.76°, n = 28 ROIs (6 regions, 6 retinas, 4 mice); ST: −4.8 ± 4.56°, n = 18 ROIs (6 regions, 5 retinas, 4 mice). P = 0.20, unpaired t test. (I) Same as F for example ROIs in the OFF B/K WAC dendritic layer. (J) Same as (G) for ROIs in the OFF B/K WAC dendritic layer. CT: 0.84 ± 0.03, n = 41 ROIs (6 regions, 6 retinas, 4 mice); ST: 0.73 ± 0.04, n = 37 ROIs (12 regions, 5 retinas, 4 mice). *P < 0.05, unpaired t test. (K) Same as (H) for ROIs in the OFF B/K WAC dendritic layer. CT: 1.94 ± 1.94°, n = 39 ROIs (6 regions, 6 retinas, 4 mice); ST: −2.13 ± 2.56 °, n = 33 ROIs (12 regions, 5 retinas, 4 mice). *P < 0.05, unpaired t test. To determine whether glycinergic inhibition contributes to orientation selectivity in B/K WACs, we examined the dendritic calcium responses in primary layers of ON and OFF B/K WAC dendrites with population calcium imaging using a dual (Flp + Cre)-dependent GCaMP7s reporter. We compared calcium responses to gratings drifting in eight directions (4.5 cycles/mm, 2 Hz, 880-µm diameter) acquired in a control condition (22) with responses acquired in the presence of strychnine (Fig. 7 F–K). Imaging was performed in separate tissues in the control and strychnine conditions to avoid effects of repeated imaging. Orientation selectivity persisted in both ON and OFF B/K WAC dendrites in the presence of strychnine and were only modestly reduced compared to control recordings (Fig. 7 F, G, I and J). Further, the PO in strychnine remained aligned with the physical dendrite (Fig. 7 H and K). These results suggest that although B/K WACs receive prominent glycinergic crossover inhibition, it is not required for generating orientation selectivity for these drifting grating stimuli. Biophysical Underpinnings of Compartmentalized Pooling. Orientation selectivity persists in B/K WAC dendrites after blocking the primary inhibitory input (Fig. 7 F–K), suggesting that compartmentalized pooling can generate orientation selectivity solely from excitatory inputs. Consistent with this interpretation, B/K WACs receive glutamatergic ribbon synapses from bipolar cells on their dendrites (22) and EPSCs showed orientation selectivity (Fig. 6 A–D). We next asked how a B/K WAC dendrite could integrate information from multiple bipolar cells while remaining electrotonically isolated from the soma and neighboring dendrites, the key features of compartmentalized pooling. To understand how a B/K WAC dendrite could implement compartmentalized pooling, we constructed a minimal biophysical model consisting of a soma and two identical dendritic cables with excitatory inputs (Fig. 8A). We then performed a numerical simulation recapitulating the experiment above probing the somatic voltage and dendritic ROI Ca2+ response to grating patches at various locations (Fig. 2). Specifically, depolarizing current, mimicking bipolar cell input, was applied to the model dendritic cable at different locations, and we calculated the membrane potential of the dendrite (220 µm from the soma). To assess whether the model could implement compartmentalized pooling, two metrics were evaluated: compartmentalization of the dendrite from the soma; and integration length along the dendrite. Implementation of compartmentalized pooling by a minimal biophysical model. (A) A minimal biophysical model with feasible parameters such that (s.t.) the key conditions of compartmentalized pooling are satisfied. Left: architecture of the model, consisting of a soma and two passive dendritic cables with excitatory inputs from bipolar cells. The free parameters in the model are the specific membrane resistance of the soma (Rs) and that of the dendrites (Rd). Right: schematic of compartmentalized pooling. The feasible parameter regime must satisfy two conditions: electrotonic isolation between neighboring dendrites and an extended dendritic integration length. (B) Parameter sweeps to identify the feasible parameter regime for compartmentalized pooling. Left: the feasible regime [the region marked with “C” (“Compartmentalized”)] that meets the electrotonic isolation constraint, characterized by the cross-compartmental depolarization remaining below the calcium threshold (Vthreshold; white dashed line). The infeasible regime is marked with “non-C” (“noncompartmentalized”). Right: the feasible regime (marked with an asterisk) with extended integration length and compartmentalization lies in the region where Rd is much larger than Rs, i.e., Rd ≫ Rs. (C) Voltage prediction by the minimal model with parameters in the feasible regime for compartmentalized pooling (* in B). ROI voltage (VROI) was calculated for 50 pA depolarizing current injection at points along the dendrite (xinput). xinput denotes the distance between the soma and the current injection location. xROI denotes the distance between the soma and the dendritic ROI. Vthreshold denotes the voltage threshold for calcium influx (marked with dashed line), which is used for calculating integration length in (D). (D) Prediction of the integration length by the minimal model with parameters in the feasible regime for compartmentalized pooling (* in B). The integration length for each ROI is defined as the distance along the dendrite where VROI > Vthreshold. The Inset shows examples of two dendritic receptive fields with different integration lengths. Key parameters of the model included the somatic specific membrane resistance Rs and the dendritic specific membrane resistance Rd, which were both varied within a biologically plausible range (1 to 10 kΩ cm2) (38). We fixed the parameters for input current (50 pA) and axial resistance (100 Ω cm) and also added a threshold for Ca2+ entry (10 mV positive to the resting potential; Vthreshold). To find the ranges of Rs and Rd enabling compartmentalized pooling, we performed two parameter sweeps. First, we identified the regime for dendritic compartmentalization, where depolarizing current injection at the soma failed to evoke a Ca2+ response (i.e., <10 mV depolarization) at the dendritic ROI, which included a broad region of the two-dimensional parameter space (Fig. 8 B, Left). Second, we identified the regime with extended spatial pooling that was within the compartmentalization regime. Extended spatial integration (~550 µm) required a high Rd. The feasible regime with both extended pooling and compartmentalization lay within a corner of the parameter space, where Rs is low and Rd is much larger than Rs, i.e., Rd ≫ Rs (Fig. 8 B, Right). This result can be understood analytically from cable theory: The high membrane resistance of the dendrite enhances its space constant, while the leaky soma shunts the cross-compartmental depolarization (SI Appendix). Minimal Biophysical Model Predicts Heterogeneity of Integration Length along the Dendrite. In our experiment and simulation (Figs. 2A and 8B), the integration length was measured at a dendritic ROI 220 µm from the soma. We further analyzed the model to test how the integration length changes at different dendritic locations. Setting the model parameters (Rs and Rd) within the feasible regime for compartmentalized pooling (* in Fig. 8B), we calculated the dendritic ROI voltage (VROI) evoked by 50 pA depolarizing current injection at positions (xinput) along the dendrite (Fig. 8C). The integration length was defined as the dendritic extent over which the voltage exceeds the threshold for calcium influx (Vthreshold). Moving from the soma to the dendritic tip, the integration length first increased sharply, reached a peak, and then mildly decreased (Fig. 8D). To explain this trend, we provide an exact analytical solution of the minimal model in SI Appendix. The variation in integration length along the dendrite would lead to heterogeneous dendritic receptive fields (Fig. 8 D, Inset). Discussion In this study, we identified the key mechanism—compartmentalized pooling—that generates orientation selectivity within the dendrites of B/K WACs. Additional factors, including center-surround antagonism and glycinergic crossover inhibition, refine orientation selectivity but are not required for its generation. Accordingly, we illuminated the minimal biophysical constraints required for generating orientation selectivity by constructing a model for compartmentalized pooling within the dendrites. As a classical strategy for generating orientation selectivity, spiking cortical neurons integrated a row of presynaptic neurons with circular receptive fields (1). Here, we showed how a similar operation is conducted in a compartmentalized manner within the dendrites of a nonspiking OS interneuron in the retina, illustrating how the visual system adopts both circuit-specialized and universal computational primitives to accomplish the fundamental task of orientation detection. Dendritic Compartmentalization Efficiently Preserves the Fidelity of Orientation Detection in B/K WAC Dendrites. Neurons across a range of brain areas use dendritic compartmentalization to accomplish computational goals. For instance, hippocampal CA1 pyramidal neurons utilize compartmentalized dendritic plasticity to mediate information storage of synaptic input features (39); layer 2/3 pyramidal neurons of the somatosensory and visual cortex use compartmentalized dendrites to discriminate spatiotemporal patterns of synaptic inputs (40); and M1 layer 5 pyramidal neurons form independent amplification units in their tuft dendrites to represent different motor sequences through combinatorial recruitment (41). In the retina, amacrine cells use compartmentalization to facilitate parallel processing and increase computational capacity (32, 38, 42–44). For example, the A17 amacrine cell generates hundreds of independent negative feedback circuits at synapses with individual rod bipolar cells (32); and starburst amacrine cells generate multiple independent direction-selective release sites with a full range of direction preferences (43–45). B/K WAC dendrites likewise function as independent computational units, as compartmentalization prevents crosstalk between neighboring dendrites with distinct physical orientations. Without compartmentalization, the depolarization in a B/K WAC dendrite would travel to neighboring dendrites and disrupt orientation selectivity. In the mouse retina, the only other WAC with confirmed orientation selectivity is unipolar, with dendrites narrowly extending in a single orientation toward the dorsal retina (12). The unipolar WAC could generate orientation selectivity for a single orientation, without requiring dendritic compartmentalization. However, encoding a wide range of orientations with this strategy would apparently require multiple mosaics of unipolar WACs, each with a different physical orientation. Encoding multiple orientations with B/K WACs may be a more efficient strategy, using multiple independent dendrites with varying orientations connected to the same soma, and thereby reducing costs related to cell number and overall wiring (46). Morphology Must Orchestrate with Biophysical Constraints to Facilitate Orientation Selectivity. In B/K WACs, the oriented and unbranched dendrite sets the structural basis for extended pooling, facilitating the convergence of spatially offset bipolar cells in one dimension. This design demonstrates how morphology promotes a structural basis for neural computation (12, 16, 47–49). However, the elongated dendritic geometry is insufficient for generating orientation selectivity and must synergize with other factors. For example, the excitatory inputs along the dendrite should have temporal kinetics that are location invariant. As a counterexample, the individual dendritic branch of a starburst amacrine cell (SAC) is elongated and oriented; however, it displays direction selectivity instead of orientation selectivity, which is explained, according to one model, by differential temporal encoding of excitatory inputs along the dendrite: sustained inputs at proximal locations and transient inputs at distal locations, which depend on different types of bipolar cells (48, 50, 51), stratifying at distinct depths of the inner plexiform layer (29–31). Alternative models, however, implicate other mechanisms for direction selectivity (52, 53), including the spatial segregation of synaptic inputs and outputs (54, 55), voltage-dependent shunting by somatic potassium channels [Kv3; (56, 57)] and mGluR2-dependent modulation of dendritic Ca2+ channels (43, 56). In any case, there must be some differences between SAC and B/K WAC dendrites that render them selective for direction vs. orientation. As another mechanism to attain orientation selectivity, the integration length along the B/K WAC dendrite must be long enough to pool excitatory inputs from multiple bipolar cells, each with a presumed circular receptive field. The length scale of dendritic spatial integration varies across cell types (58). In an extreme case, individual dendritic varicosities of A17 amacrine cells, spaced ~20 µm apart, form independent local feedback circuits for parallel computation. This is attributed to the extremely short effective length constant of voltage-gated calcium channel signaling (~13 µm) (32). The B/K WAC dendrites, lying on the other end of the spectrum, have an effective length constant for calcium signaling on the scale of several hundreds of microns (Fig. 2), which fulfills the extended integration length constraint. Network Motifs Underlying Orientation Tuning of B/K WAC Dendrites. Circuit motifs represent stereotyped patterns of functional connectivity, serving as fundamental computational primitives in the brain (59–65). Functionally dissecting complex neuronal networks into motifs presents an effective approach to understanding how the brain performs diverse computational tasks on the algorithmic level. In this study, we revealed three network motifs underlying the orientation selectivity of the B/K WAC dendrites. The first motif is feedforward excitation, consistent with bipolar cell input (Figs. 3, 5, and 6) (22). It is unlikely that B/K WACs directly inherit orientation selectivity from bipolar cells, which primarily possess concentric center-surround receptive fields (29–31) and should lack orientation selectivity. An exception in the mouse retina, type BC5A, shows tuning to vertical orientation (18), which cannot explain variable orientation tuning across B/K WAC dendrites (22). As the excitatory receptive field size along the dendritic axis of B/K WACs exceeds that of individual bipolar cells (Fig. 2), we conclude that B/K WACs pool non-OS feedforward excitation from bipolar cells along individual dendrites to generate dendritic OS. The second motif is lateral inhibition, which generates center-surround antagonism (Figs. 3 and 4). The center-surround receptive field structure is a canonical computational primitive in sensory systems. In the retina, this receptive field structure increases sensitivity to edges (66), modulates the nonlinearity of spatial integration of the center (67), generates radial motion sensitivity, and facilitates novel object detection (31, 52). Here, center-surround antagonism shapes band-pass orientation tuning (Fig. 3) and sharpens OS at low spatial frequencies (Fig. 5). For OFF B/K WACs, tonic inhibition tuned to the NO (Fig. 3) may arise from lateral inhibition mechanisms with orthogonal OS, resembling cross-orientation inhibition in the cortex (68, 69). The third motif is crossover inhibition, or push–pull inhibition, where opposite contrast polarities drive excitation and inhibition (33, 35–37). Although both ON and OFF B/K WACs receive substantial glycinergic crossover inhibition (Fig. 7 A–E), their orientation selectivity persisted and reduced only modestly with the blockade of glycine receptors (Fig. 7 F–K). Moreover, crossover inhibition probably does not contribute to the compartmentalization of dendrites when activated by PO gratings, because compartmentalization would seem to depend on tonic conduction (e.g., leak channels) as opposed to modulated conductance. However, crossover inhibition may play a more significant role in orientation tuning under other stimulus conditions. For instance, mutual inhibition between ON starburst amacrine cells contributes to direction selectivity in response to a moving bar stimulus on a noisy background but not a homogeneous background (70–72). Additionally, push–pull inhibition has been proposed to account for contrast-invariant orientation tuning of simple cells in the visual cortex (73). Therefore, it would be informative to understand the role of crossover inhibition in B/K WACs across a wider range of stimulus conditions. Role of B/K WACs in Visual Processing. The connections of B/K WACs within the retina are not fully understood, but they seem likely to participate in orientation tuning of OS RGCs. Four types of OS RGCs have been studied with whole-cell recording in the mouse retina: ON and OFF types that are tuned to either horizontal (hOS RGCs) or vertical orientations (vOS RGCs) (11, 12). The dendritic arbors of these RGCs are either bistratified (ON OS RGCs) or monostratified (OFF OS RGCs). They apparently costratify with the ON or OFF B/K WACs (22, 23), and thus they could make connections that generate orientation tuning. For example, if a network of B/K WAC dendrites at all angles except for those within ~30 degrees of the horizontal axis connected with an OS RGC, that OS RGC would become selective for horizontal orientations, because the responses to other orientations would be inhibited by the GABA release from the B/K WAC network. Indeed, orientation-tuned GABAergic inhibition seems to explain OS RGC receptive fields in the rabbit retina (13, 14, 74). For mouse OS RGCs, the mechanism for orientation tuning seems to be more diverse. For example, the OFF vOS RGCs appear to gain orientation tuning via depolarizing input through gap junctions with vertically tuned amacrine cells, other than B/K WACs, with unipolar dendritic trees that extend dorsally (12); whereas direct inhibitory input to these OFF vOS RGCs lacks orientation tuning. For ON OS RGCs, synaptic excitation is tuned to the PO (i.e., aligned with the tuning of the firing rate). Inhibition is tuned to the orthogonal orientation in ON hOS RGCs, similar to the case for rabbit cells, but the tuning is more variable in the ON vOS RGCs. Thus, the clearest putative role for B/K WACs would be selective wiring with ON hOS RGCs to generate inhibition that is tuned orthogonally to excitation. This circuit would need to be confirmed with optogenetic experiments and connectomic analysis (22). There are potentially additional types of OS RGCs in the mouse, identified by calcium imaging experiments and by reconstructions from scanning block face electron microscopy (75–77); and the synaptic mechanisms for these cells would need to be studied further to determine whether orientation-tuned inhibition contributes to their receptive fields. To the extent that B/K WACs drive tuning in OS RGCs, the wide range of B/K WAC dendritic angles could, through selective convergence, generate different degrees of orientation tuning and also allow the tuning to vary across the retina (78). If B/K WACs did drive orientation selectivity in any type of OS RGC, their influence would extend deep into the visual system, since OS RGCs directly contribute to orientation tuning of downstream neurons in LGN, SC, and V1 (6–8). B/K WACs could potentially play broader roles in information processing within the retina. For example, there are WAC inputs to bipolar cell terminals, which could contribute to the tuning of glutamate release onto OS and DS RGCs depending on the degree of convergence (i.e., the range of angles of connected WAC dendrites) and the nature of the connection (i.e., electrical vs. chemical inhibitory synapse) (18, 79, 80). Based on optogenetic experiments, B/K WACs also connect with non-OS RGCs, including Alpha RGC, potentially via convergence of dendrites at all angles (22). In this case, the net effect of B/K WAC input would not be orientation tuning in the postsynaptic neuron. However, there may be some advantage for a network of inhibitory interneurons to sample natural images with orientation-tuned filters, even if their collective role in RGC function is not to generate orientation tuning. Additional experiments would be required to determine the precise role of B/K WACs in non-OS RGCs, using cell inactivation techniques. Furthermore, our minimal model predicts heterogeneous receptive fields along the dendrites of B/K WACs (Fig. 8D). It would be interesting to investigate the presence of heterogeneity in dendritic tuning properties as well as corresponding increases in coding capacity and efficiency at the population level (81–83). Materials and Methods All animal procedures were approved by the Institutional Animal Care and Use Committees at Yale University and complied with NIH guidelines. B/K WACs were genetically labeled with either a fluorescent reporter or calcium indicator using an intersectional genetic strategy described previously (22). Animals were maintained on a 12/12-h light/dark cycle and studied at 2 to 6 mo, including both sexes. A mouse was dark-adapted for 1 h prior to retinal dissection. Calcium imaging and targeted patch-clamp recording of fluorescent cells were performed using a custom-built two-photon microscope (22). The minimal biophysical model was implemented using NEURON version 8.2.6 in Python 3.12 and analyzed using Mathematica 14.1. Detailed materials and methods are described in SI Appendix. Supplementary Material Appendix 01 (PDF)