Ganglion cells in larval zebrafish retina integrate inputs from multiple cone types

Abstract

We recently showed the presence of seven physiological cone opsins—R1 (575 nm), R2 (556 nm), G1 (460 nm), G3 (480 nm), B1 (415 nm), B2 (440 nm), and UV (358 nm)—in electroretinogram (ERG) recordings of larval zebrafish (Danio rerio) retina. Larval ganglion cells (GCs) are generally thought to integrate only four cone opsin signals (red, green, blue, and UV). We address the question as to whether they may integrate seven cone spectral signals. Here we examined the 127 possible combinations of seven cone signals to find the optimal representation, as based on impulse discharge data sets from GC axons in the larval optic nerve. We recorded four varieties of light-response waveform, sustained-ON, transient-ON, ON-OFF, and OFF, based on the time course of mean discharge rates to all stimulus wavelengths combined. Modeling of GC responses revealed that each received 1–6 cone opsin signals, with a mean of 3.8 ± 1.3 cone signals/GC. Most onset or offset responses were opponent (ON, 80%; OFF, 100%). The most common cone signals were UV (93%), R2 (50%), G3 (55%), and G1 (60%). Seventy-three percent of cone opsin signals were excitatory, and 27% were inhibitory. UV signals favored excitation, whereas G3 and B2 signals favored inhibition. R1/R2, G1/G3, and B1/B2 opsin signals were selectively associated along a nonsynergistic/opponent axis. Overall, these results suggest that larval zebrafish GC spectral responses are complex and use inputs from the seven expressed opsins.

NEW & NOTEWORTHY Ganglion cells in larval zebrafish retina have complex spectral responses driven by seven different cone opsin types. UV cone inputs are significant and excitatory to ganglion cells, whereas green and blue cone inputs favor inhibition. Most dramatic are the pentachromatic cells. These responses were identified at 5–6 days after fertilization, reflecting an impressive level of color processing not seen in older fish or mammals.

INTRODUCTION

As the output neurons of the retina, ganglion cells (GCs) extract and integrate the essential parameters of the visual stimulus. GC light responses are diverse and complex, with chromatic, temporal, and spatial properties. White light stimuli identify transient (ON-OFF, ON, OFF) and sustained (ON, OFF) response types (1–5), whereas spectrally distinct stimuli identify both nonopponent and color-opponent responses. Nonopponent GCs are spectrally monophasic and respond similarly to all spectral wavelengths, whereas color-opponent GCs are alternately excited or inhibited by different spectral regions (6). Spectrally biphasic color-opponent GC responses are present in fish (7–11) and mammals (12–15). In fish and turtle, spectral responses further indicate inputs to GCs from more than two cone types (16), with spectrally triphasic responses present in both genera (17–19) and tetraphasic responses also reported in turtle (20). The spectrally distinct responses of GCs result from circuitry throughout the retina involving horizontal cell feedback, direct inputs from bipolar cells connected to specific cone types, and/or inhibitory feedback from amacrine cells (21–25).

Zebrafish GCs form at ∼32 h postfertilization (hpf), followed by horizontal and photoreceptor cells (26); a small number of GC axons leave the eye to form the optic nerve a few hours later at ∼34–36 hpf (26–29). By hatching [70–74 hpf (26)], the optic nerve has innervated the tectum (29), light responses are first recorded (4), and visually guided behaviors begin (30, 31). Stable light responses, recorded at 4 days postfertilization (dpf), identify ON, OFF, and ON-OFF GC types (4); these responses are also present in 6–8 dpf larvae (2). Morphological studies with larvae ≤ 6 dpf identify different GC dendritic stratification patterns in the inner plexiform layer (IPL), including mono- and multistratified arbors (5, 32) with as many as 15 different stratification patterns identified in the earlier study. Similar characterizations suggest that 11 anatomical types of GC are present in adult zebrafish (33).

The zebrafish retina has rich potential for color vision. One of the advantages for cone color vision studies in larval zebrafish is that rod signals are very weak (34, 35). Four morphological cone types [red (R), green (G), blue (B), ultraviolet (UV)] and one rod are present. Molecular studies show different opsins expressed in these photoreceptor types: rods and B and UV cones express a single opsin each, whereas R cones express one of two opsin genes (LWS1 or LWS2) and G cones have four copies of the green opsin gene (RH2-1, RH2-2, RH2-3, RH2-4) from which to choose (36, 37). Electrophysiological studies indicate the existence of a second B cone opsin pigment (38). The chromophore bound to the opsin in zebrafish photoreceptors is vitamin A1 (39), although A2-based pigments are made if adult zebrafish are exposed to thyroxin (40). Postsynaptic to photoreceptors, horizontal cell spectral responses include contributions from all cone types (41), as do amacrine cell responses (42), and morphological examination of zebrafish bipolar cells indicates type-specific connections with cones (43, 44). As yet there is no information on spectral/chromatic properties of GC impulse discharges in larval zebrafish. Thus, the purpose of this study was to determine how GCs in larval zebrafish retina process wavelength information. In particular, we sought to determine whether GC light discharges reflect cone opsin diversity in a variety of new spectral patterns or if these interactions resemble the more limited cone interactions found in adult zebrafish horizontal cells (41), the red-green opponent spectral patterns in adult goldfish GCs (7, 45) or the UV-dominated patterns in trout molts (16).

METHODS

Animals and Tissue Preparation

Wild-type (EK strain) zebrafish larvae (5–6 dpf) were obtained from in-house spawning of adult animals. Larvae were transferred to a microscope slide and then picked up with a wedge of black 0.45-µm nitrocellulose filter paper. With long 37-mm insect pins, larvae were decapitated and the eyes removed. Eyes were placed cornea side down on the filter paper, so the optic nerve head was visible, and the tissue was stabilized with a drop of 0.8% low-melting-point agarose (warmed to 37–40°C). The agarose solution covered the sides of each eye only, leaving the optic nerve head uncovered. The filter paper + eye was transferred to the recording chamber and superfused with oxygenated MEM solution equilibrated with 95% O₂-5% CO₂. The perfusion needle was placed near the eye with a perfusion rate of 0.05 mL/min. Each eye was visualized and positioned with a microscope IR camera (Teledyne QImaging, Surrey, BC, Canada) and Metamorph imaging software (Molecular Devices, San Jose, CA). The camera was mounted on an Olympus BX51WI compound microscope. Thin-walled patch-clamp electrodes pulled to the desired ∼3-µm tip diameter with a Flaming-Brown P-87 micropipette puller (Sutter Instrument, Novato, CA) and filled with a 0.5 M NaCl solution were used to make spike recordings.

All procedures performed in these studies were in accordance with the guidelines and regulations of the Institutional Animal Care and Use Committee (IACUC) of the National Institute of Neurological Disorders and Stroke (NINDS) at the National Institutes of Health. The experimental protocols were approved by the NINDS IACUC (ASP-1227, 1307).

Physiological Recordings

For single-unit spike recordings and electroretinograms (ERGs), electrode tips were inserted either directly into the optic nerve head at the back of the eye or through the cornea with a micromanipulator (Sutter Instrument, Novato, CA) (46). The tissue was stimulated with nine wavelengths ranging from 330 nm to 650 nm (20-nm half-width interference filters, 40-nm increments; Chroma Technology, Bellows Falls, VT) over seven intensities (UV-compliant neutral-density filters), 0.5-log unit increments covering 3 log units (Andover Corp., Salem, NH). Stimulus irradiance was measured in the plane of electrode insertion with a calibrated photodiode (Newport Corp., Irvine, CA). The source was a 150-W OFR xenon arc with two optical channels both gated by Uniblitz shutters. The stimulus beam passed through three Sutter filter wheels and a UV-visible-compliant liquid light guide (Sutter Instruments), through the epifluorescence port of the BX51WI upright microscope (Olympus), through a ×10 UPlanFLN/0.3 microscope objective (Olympus–Life Science Solutions), and onto the tissue. The second optical channel passed through hand-inserted filters and an infrared-compliant liquid light guide (Newport Corp.) providing infrared side illumination for visualization.

ERGs were recorded first, immediately after contacting the tissue, to assess the viability and stability of the preparation. Electrodes were then advanced into the optic nerve until GC spike responses were encountered. These were digitized at 10,000 Hz and digitally filtered through a 200-Hz high-pass filter (2). When spikes were first encountered, electrode placement was adjusted slightly to maximize the response and to isolate the discharges of a single fiber. Impulse discharges were recorded with an Axopatch 200B amplifier, a Digidata 1440A, and pCLAMP 10 software (Molecular Devices). Data were analyzed with Origin and Origin LabTalk scripts (various versions, OriginLab Corp., Northampton, MA).

Three spectral protocols were used for data acquisition. The first (A4) was a “search” protocol consisting of 1-s flashes at a selected wavelength (typically 570 nm) repeated every 3 s. When A4 was a recorded protocol, spike traces in response to 10 such stimuli were combined to determine response dynamics (ON, OFF, ON-OFF). The second spectral protocol (C6) consisted of a nine-wavelength quick spectral scan (in time order 650–330 nm) repeated three times for a total of 27 impulse traces. Flash duration was 0.7 s, and flash interval was 3 s. Irradiances ranged between 3.6 (330 nm) and 5.2 (650 nm) log(quanta·µm⁻²·s⁻¹). The third spectral protocol (Cs2) consisted of seven irradiances at each of 10 wavelengths, with each wavelength-irradiance combination repeated four times, for a total of 280 stimuli, each evoking an impulse trace. The flash duration was 0.3 s with intervals varying from 2.5 to 6 s, the longer intervals with brighter stimulus irradiances.

Data Analysis and Spectral Modeling

Impulse traces were digitized by thresholding and converted into “rate-meter” or “spike-rate” records to visualize response temporal dynamics. After high-pass filtering (200 Hz) and smoothing [Savitsky–Golay, 10 points (1 ms), order 3], the standard deviation (SD) of the 20,000-time point (2 s) trace was calculated. Any point deviating by >3 SDs from the mean was scored an “event” and set to “1,” while other points were set to “0.” Event records from three (protocol C6) or four (protocol Cs2) stimulus presentations were summed. To generate the spike-rate trace, the mean number of events in a 50-ms running bin was calculated and graphed. For spectral modeling, the number of events occurring in a time window following light onset, or following light offset, were counted. From this count, the number of events (adjusted for the different time intervals considered) occurring in the 0.4 s before stimulation were subtracted. For ON signals, all signals during the light ON phase were counted. For OFF signals, all events for 600 or 800 ms following the stimulus were counted, with the shorter time window corresponding to the shorter stimuli (protocol Cs2).

Amplitude-wavelength-irradiance data sets were fit to each of the 127 spectral models in a seven-cone spectral scheme to determine which combination(s) of seven cone opsins excited or inhibited each GC. Zebrafish have eight cone opsin genes (47), at least seven of which may be expressed (37, 48–52) and/or functional (38) in 5–6 dpf larvae. To determine which opsin(s) provided signals to larval GCs, all 127 models comprising all the potential combinations of the seven physiological opsins were tested on each cell. The nonlinear curve fit provides each fit parameter with a standard error of the mean (SE). Using the parameter value and the SE, a t test gave the probability that the parameter differed from “0.” Only models where all the opsin signals were significantly different from “0” [t test, P ≤ 0.05, based on fit values for maximal or saturation number of events (Vm_i) mean and SE] were considered “valid.” Of these valid models, the one with either the largest r² fit value, or that one together with others that were indistinguishable from it (F test, P ≥ 0.95), were considered to best represent the GC cone signal sources. Each model fit determines whether a particular cone within the modeled cone subset is excitatory, inhibitory, or not significant. Thus, as a group the 127 models support 3⁷ (2,187) potential combinations of excitation, inhibition, or noncontribution from the seven-cone scheme.

The 127 multicone spectral models in the spectral scheme were generated from sums of Hill functions (53).

$$V={\sum }_{i=1}^{7}V{\mathrm{m}}_i\times \frac{I}{I+\frac{k_i}{A(\mathrm{wlma}{\mathrm{x}}_i,\mathrm{wl})}}$$ (1)

V (Eq. 1) is the net discharge of events for a stimulus of irradiance I and wavelength wl. Vm_i is the maximal or saturation number of events contributed to the signal from a cone expressing opsin_i with peak wavelength wlmax_i and semisaturation irradiance k_i. The Vm_i value may be positive or negative, signifying excitation or inhibition, respectively. A(wlmax_i, wl) is the normalized absorbance function for opsin_i. The opsin absorbance maxima, and cone semisaturation irradiances used in the models, appear in Table 1. The absorbance function is based on the Dartnall nomogram concept that opsin shapes are invariant when represented on a reciprocal wavelength axis (54). The opsin shapes are given by order 8 polynomials. These polynomials derive from spectral, suction electrode recordings of individual cones in giant danio (55). A single polynomial (56) serves for all opsins except UV. The narrower UV opsin polynomial function is from Ref. 55. Polynomial coefficients used are summarized in Ref. 53.

Table 1.

Properties of cones used in the spectral model

Index (i)	Opsin	Cone Type	Gene	Peak Wavelength, nm	Relative Semisaturation, log(k)
1	UV	UV	SWS1	358	0.00
2	B1	Blue	SWS2	415	0.00
3	B2	Blue	Not known	440	0.50
4	G1	Green	RH2-1/RH2-2	460	0.00
5	G3	Green	RH2-3	480	1.00
6	R2	Red	LWS2	556	1.00
7	R1	Red	LWS1	575	1.00

k, Semisaturation irradiance. The opsins listed correspond to the seven physiological opsins identified in zebrafish retina. Two opsin types were reported in red cones (R1, R2), two in green cones (G1, G3), two in blue cones (B1, B2), and one physiological opsin observed in ultraviolet (UV) cones. i refers to cone number in the model (as in Eq. 1).

RESULTS

Absorption Spectrum of Larval Sclera, Choroid, and Pigment Epithelium

All but 1 of the 43 larval GC recordings were from the optic nerve with spectral stimulation passing through the sclera, choroid, and pigment epithelium. Stimulus quantification requires measurement of how much energy these layers absorb at each stimulus wavelength. We estimate this by comparing ERG b wave sensitivities as measured transcorneally to those recorded within the optic nerve (Fig. 1). ERG b waves were measured in the same set of eyes used for GC electrophysiology. The b wave stimulus protocol was identical to Cs2, with 280 stimuli delivered at nine wavelengths and seven irradiances each. Maximum ERG b wave amplitudes averaged 167.3 ± 21.6 µV in transcorneal recordings (n = 8 Cs2 protocols, 8 eyes, mean ± SE) and 204.6 ± 24.9 µV from within the optic nerve (n = 21 Cs2 protocols, 16 eyes, mean ± SE). These amplitudes did not differ [t(37) = −0.87, P = 0.39]. Amplitude-normalized irradiance-response series at 490 nm are superimposed in Fig. 1A. Optic nerve b waves were slightly more prolonged than vitreal b waves. In this example, the optic nerve b waves appear slightly more sensitive.

Figure 1.

Absorption of spectral stimuli by larval sclera, choroid, and pigment epithelium. Electroretinogram (ERG) b wave sensitivity is used to estimate stimulus attenuation with transscleral stimuli. A: recordings within the optic nerve, using transscleral stimulation (blue), and within the vitreous, using corneal stimulation (red), are normalized to the maximal amplitude at each recording configuration and superimposed. The quantal irradiance, given on left of each response trace in log(quanta·µm⁻²·s⁻¹), is calibrated at the eye surface. B: maximum-amplitude-normalized data sets at 490 nm for 8 vitreal and 21 optic nerve b wave data sets are combined and globally fit by Hill functions. The semisaturation irradiance k is fit separately, but the saturation amplitude Vm and exponent n are constrained to be the same. Overall, vitreal b waves were 0.18 log units more sensitive than optic nerve b waves at 490 nm. C: the change in log(k) across stimulus wavelengths is fit by a 4th-order polynomial. In B and C error bars are SEs.

Irradiance-response data sets were normalized and combined at each wavelength. Hill functions (Eq. 2) were simultaneously fit to vitreal and optic nerve data sets with the constraint that Vm, the saturation amplitude, and n, the slope, be global parameters and only k, the semisaturation irradiance, be allowed to differ between vitreal and optic nerve fits at each wavelength.

$$V=\frac{V\mathrm{m}\times I^n}{I^n+k^n}$$ (2)

Global 490-nm irradiance-response fits appear in Fig. 1B. The log(k) values show that for all 490-nm data sets optic nerve b waves, obtained with transscleral stimuli, are 0.18 log units less sensitive at 490 nm than corneal b waves, obtained with transcorneal stimuli. The inferred scleral/pigment epithelial/choroidal absorption at all stimulus wavelengths appears in Fig. 1C, where the data are represented by a fourth-order polynomial fit. This trend line is used to correct irradiance values at the cones for scleral stimulation. The trend suggests some opaqueness in the UV but overall little transscleral impediment to the transmission of the stimulus.

Spectral Discharges of Larval Ganglion Cells

On penetration of the optic nerve, single fibers with impulse discharges stably responsive to the A4 test stimulus were encountered. Then, to quickly access the spectral properties, the GC was stimulated by the C6 protocol: three identical sweeps across nine wavelengths from long to short. The GC of Fig. 2 is maximally excited at 610 nm and 370 nm but appears inhibited at 530 nm and 490 nm. One might anecdotally refer to this as an R- and UV-ON, G-OFF type cell, but the details about which of the seven spectrally identifiable larval opsin types contribute are difficult to determine by inspection. The peak long wavelength excitability at 610 nm is not the same as either R1 (575 nm) or R2 (556 nm) LWS opsin types, suggesting that the red peak results from an interaction of cone opsins, though it is not clear which cone opsins are interacting. Similarly, although the strong excitation at 370 nm is easy to identify as being driven by the UV (358 nm) cone, the cone types involved in the nearby excitation at 450 nm and 410 nm are difficult to pinpoint by inspection. These are the reasons we decided to evaluate GC spectral patterns using a process that examines all the potential combinations of cone signals to determine which input patterns is/are most likely to be generating each response spectral profile. The modeling results for this cell appear later (see Fig. 4).

Figure 2.

Spectral response data set for C6 stimulus protocol. This 5 days postfertilization (dpf) ganglion cell is most sensitive to long (610 nm)- and short (370–330 nm)-wavelength stimuli, while appearing inhibited in the midspectrum (530–490 nm). The spikes are shown after 200-Hz high-pass filtering and Savitsky–Golay smoothing (10 points, 1 ms, order 3). The rectangular traces indicate light stimulation. The column headers give the wavelength and log of neutral-density attenuation of the beam used in the C6 stimulus protocol. The modeling of this data set appears in Fig. 4.

Figure 4.

Fitting the sample C6 impulse data to a spectral model. A: impulse events (shown in Fig. 2) are detected, and the running averages of events in a 50-ms time bin are displayed as normalized event rates. Wavelengths are color coded and shown above each stimulus trace. The number below each wavelength is the irradiance in log(quanta·µm⁻²·s⁻¹) (where hν = quanta). C6 stimulus protocol. B: the mean waveform for all 9 wavelengths shows increased activity throughout the 700-ms stimulus. C: the best-fitting model (112) uses 5 cone signals to represent the data set. + and − indicate significance of excitatory or inhibitory signals. respectively, with the number of symbols representing t test significance (GraphPad convention). D: event-rate amplitudes for constant quantal stimulation across wavelengths [log(quanta·µm⁻²·s⁻¹) = 4.60] for the best-fitting model (112). No other model had a comparable residual variance (F test, P ≥ 0.95).

Model Assumptions

The spectral scheme is based on literature values of cone opsin peaks, spectral shapes, and relative sensitivities of different cone types. The scheme (Eq. 1) generates 127 models resulting from the different cone combinations implicit in the overall seven-opsin spectral summation (38). We evaluated all of these to determine which combinations of cones best represent GC spectral properties. The choice of seven cone spectral types was based on molecular (37, 47), microspectrophotometric (57, 58), and electrophysiological (36, 38) evidence showing that multiple cone opsin types generate functionally and biochemically distinguishable cone types in larval zebrafish. By modeling raw, single-unit impulse discharges from larval GCs, we sought to identify the responsible cone signals and whether there were common patterns of cone combinations. The modeling determined the weight/level of input from each cone opsin type. Since a visual pigment does not respond to only one wavelength but responds to wavelengths flanking the λ_max, each model looked at the progression of light responses across all stimulating wavelengths. Models with any cone Vmax value indistinguishable from zero were rejected (t test, P > 0.05). Accepted models were those with the largest value of r² and those with F values that were indistinguishable from it (F test, P ≥ 0.95).

The spectral scheme assumes that linear additions and subtractions of saturable cone signals capture the bulk of circuitry processing of spectral signals as they converge on GCs. ON-OFF responses involve, by definition, a further temporal nonlinearity. We therefore modeled ON and OFF signals separately. Each of the 127 seven-opsin models tested was fit to either 280-point (Cs2) or 27-point (C6) data sets, with degrees of freedom being 2–8 less, depending on how many opsin signals were included in the best-fitting model.

Model Noise Rejection

Unlike ERG signals from the whole retina (53) or microelectrode signals from distal retinal neurons (41), GCs are spontaneously active. The spontaneous noise signal might be modeled as false spectral peaks. To assess this, we constructed 26 data sets with randomly timed spikes for both the C6 and Cs2 stimulus protocols. Either 50 or 100 impulses were placed in each response trace. These were 100 µV in amplitude and 300 µs in duration superimposed on a 10-µV Gaussian white noise baseline. The 127 seven-cone models were fit to each of these data sets. For C6, 95% of r² values were <0.302, and for Cs2, 95% of r² values were <0.024 (Fig. 3). In random C6 data sets, 0.12% of models fit (4 of 3,327) were “valid fits,” with all the fit cone signals > 0 (t test, P ≤ 0.05). In random Cs2 data sets, 0.69% of models fit were valid (23 of 3,328). Creation of a false spectral model was rare, as it required that noise alter multiple data points both in wavelength and in irradiance in a specific manner. This accounts for the low frequency of valid models from random data sets. Therefore, in analysis of GC data sets, we chose to exclude all C6 models with r² ≤ 0.30 and Cs2 models with r² ≤ 0.03 (vertical dashed lines, Fig. 3). This effectively eliminated the insertion of “false opsins” generated by noise.

Figure 3.

Discriminating cone signals from noise. For both C6 (27 stimuli, left) and Cs2 (280 stimuli, right) protocols, 26 random impulse data sets were generated and fit to the set of 127 models that test all combinations of 7 cones. The cumulative frequencies of resultant r² values (blue open triangles) are compared to values for “valid” spectral models of ganglion cell (GC) data sets (valid GC fits, open red circles) and to the “best” r² values for each GC. Models with all cone Vmax values significantly different from “0” are “valid” models (t test). “Best” models are those with the greatest r² value and those with r² values indistinguishable from the best value [best GC fits, filled red circles (F tests, P ≥ 0.95)]. Ninety-five percent of random fits lie to the left of the dashed vertical lines. Only fits of GC data sets to the right of this line are considered to be candidate GC models.

The cumulative frequency distributions for r² in random data set models and in both “valid” and “best” GC data set models are compared in Fig. 3. The “best” GC models are those with the greatest r² values and include all those with residual variances that are indistinguishable from the greatest r² value (F test, P ≥ 0.95). A substantial fraction (43%) of “best” models in C6 data sets were rejected as potentially noise contaminated, whereas few “best” models in Cs2 data sets (2%) were rejected. Among 26 GC data sets with a model meeting the threshold r² criteria, 670 models were “valid.” Of these, 40 were “best” models (F test, P ≥ 0.95), 387 were worse fits (F test, P ≤ 0.05), and 243 were indeterminate (F test, 0.05 < P < 0.95).

Sample Model Fits

The C6 impulse data from the cell of Fig. 2 were best fit by a five-cone model, model 112, of the 127 models fit (Fig. 4). Model 124 came the closest to being indistinguishable from model 112 (F_112,124 = 0.976, P = 0.942); however, model 124 used six cones to achieve this fit and so was also rejected as not being parsimonious with parameters. Model 112 represents this GC as being significantly excited by R2, B2, and UV cones, while being inhibited by R1 and G1 cones (Fig. 4C). Spectrally (Fig. 4D) the greatest model amplitude is in the UV, peaking at 359 nm, with the next largest peak in the red (579 nm). Excitation in the midspectrum was low. The modeled amplitude spectrum closely resembles the event-rate histograms (Fig. 1A). There is a sustained discharge pattern of the mean response across all wavelengths (Fig. 1B).

An example of a Cs2 data set is modeled in Fig. 5. In Cs2 there are responses to seven different stimulus irradiances at each wavelength. Three of these, each matched across the spectrum as representing responses to equal quanta stimulation, are displayed in Fig. 5A. There is a large UV excitation at 370 nm and 330 nm and a further broad spectral peak at long wavelengths. An activity minimum occurs in the midspectrum. The best-fitting model (model 120, Fig. 5B) employs six of the seven cone types, four excitatory (R2, G1, B1, and UV) and two inhibitory (G3 and B2). Irradiance response curves generated by model 120 at two excitatory wavelengths are compared to amplitude measurements in Fig. 5C. Model 120 interpolates a constant-quanta response-amplitude plot (Fig. 5D). The modeled plot resembles the event-rate histograms (Fig. 5A). There is a prominent excitatory peak in the UV at 368 nm and a broader, lower-amplitude excitation at long wavelengths, peaking at 548 nm. A lesser-amplitude response occurs in midspectrum, with a trough at 480 nm. There were no other models with indistinguishable residual variances (F test, P ≥ 0.95). The next best fit was model 121, also a six-cone model but substituting R1 excitation for R2 excitation (F_120,121 = 0.976, P = 0.838).

Figure 5.

Fitting a Cs2 data set to a model. A: for clarity, response traces to only 3 of the 7 irradiances presented at each wavelength in the Cs2 protocol are shown. These are equal quantal across all wavelengths. The responses are color coded, with black being the dimmest, blue intermediate, and red the brightest. Numbers on right of the stimulus traces give the irradiances in log(quanta·µm⁻²·s⁻¹). B: the best-fitting model (120) uses 6 of the 7 cones, 4 excitatory (R2, G1, B1, and UV) and 2 inhibitory (G3 and B2). No other indistinguishable models (F test, P ≥ 0.95) were found. +++, ++++ and − − −, − − − − (P ≤ 0.001, P ≤ 0.0001) are t test probability codes for excitatory and inhibitory signal amplitudes, respectively. ns, Nonsignificant. C: model 120 fits to Cs2 irradiance response data at 2 wavelengths (370 nm and 570 nm). D: model 120 interpolates response amplitudes across the spectrum for constant quantal stimulation at 4.00 log(quanta·µm⁻²·s⁻¹).

Nomenclature

Spectral data sets were collected from a total of 43 GCs, 26 of which were stable and met the above criteria with r² > 0.30 (C6, 27 stimuli) or r² > 0.03 (Cs2, 280 stimuli). If the best model included both excitatory and inhibitory inputs from spectrally distinct cones, the GC was considered “spectrally opponent”; otherwise the cell was classified as “nonopponent.” Chromatic complexity was defined by the number of opsin signals contributing to the GC response. For example, a cell with four signals including both excitatory and inhibitory inputs would be termed opponent and tetrachromatic. The modeled equal-quantum spectral curves were characterized by excitatory peaks and inhibitory troughs. The number of peaks and troughs determined the number of “phases” in the spectrum. A cell with one peak and one trough was termed “spectrally biphasic.” The number of cone opsin inputs could differ from the number of spectral phases produced. There are five spectral peaks and troughs in the pentachromatic GC of Fig. 4D. It is pentaphasic. There are three spectral peaks and troughs in the hexachromatic GC of Fig. 5D. It is triphasic.

Waveform Classification

Four ganglion cell types were identified based on mean waveform to all protocol stimuli: sustained-ON (n = 23; Fig. 6A), transient-ON (n = 7; Fig. 6B), ON-OFF (n = 12; Fig. 6C), and OFF (n = 1). We classify a GC as “sustained” if the duration of mean ON excitation to all wavelengths exceeds 300 ms, the duration of Cs2 stimuli; otherwise the GC is termed “transient.” The GCs of Figs. 2, 4, and 5 are sustained types. The prevalence of each waveform appears in Fig. 6D, Within each of these types, both spectrally opponent and nonopponent responses were observed, and the fraction of opponent types was similar among all waveforms, except for the single OFF type, whose spectral model fits did not meet the acceptance criteria.

Figure 6.

Light evokes different discharge kinetics in larval ganglion cells. A–C: mean firing rate waveforms are generated from all wavelengths and irradiances in C6 (27 stimuli) spectral protocols. These are subjective groupings of impulse-discharge kinetics. A: sustained-ON. The impulse discharge is elevated throughout the light step. B: transient-ON. The impulse discharge is complete within 300 ms and is followed by maintained suppression of activity for the remainder of the light pulse. C: ON-OFF. Separate discharge peaks occur at stimulus onset and offset. Inhibition appears during the later phase of stimulation. The ON-OFF dynamic may either be wavelength independent or arise from color opponency. D: frequency of occurrence of the sustained-ON, transient-ON, ON-OFF, and OFF temporal patterns. Purely OFF responses are rare under these stimulus conditions. WT, wild type.

Onset Excitation and Inhibition by Multiple Cone Signals

In 5–6 dpf larval retinas, cone photoreceptors are modeled as expressing one of seven possible opsin types (Table 1): R1 or R2 (in red cones), G1 or G3 (green cones), B1 or B2 (blue cones), and UV (UV cones). Among the 26 accepted GC data sets, the 40 “best” or “indistinguishable from best” models found 153 cone signals at stimulus onset: 111 (73%) were excitatory, and 42 (27%) were inhibitory. On average there were 3.8 ± 1.3 (SD) cone signals/model, ∼2.8 excitatory and ∼1.0 inhibitory, with the number of different cone signals ranging from 1 to 6 (Fig. 7A). The modal number of different cone signals was 4. The distribution was not described by a random selection among all 127 models [χ²(6,40) = 19.27, P ≤ 0.01; Fig. 7A]. Notably, the seven cone types did not contribute in the same way to GC excitation or inhibition. Inputs from UV cone opsins were excitatory in 93% of models, significantly greater than the 40% mean for all cone types [χ²(2,40) = 46.83, P ≤ 10⁻¹⁰; Fig. 7B]. In contrast, inputs from B2 and G3 cone opsins tended to be more inhibitory than excitatory [χ²(2,40)_B2 = 14.73, P ≤ 0.001; χ²(2,40)_G3 = 8.16, P ≤ 0.05; Fig. 7B]. These differences in overall excitatory/inhibitory tendencies for cone signals were also observed in individual GC data sets (Fig. 2).

Figure 7.

Cone input patterns in onset responses. A: the number of cone signals found for ON responses in “best” ganglion cell (GC) models varied from 1 to 6, with a peaked distribution centered at 3.8 (dark bars). The distribution differed from blind choices among the 127 models [hatched bars, χ²(6,40) = 19.27, P ≤ 0.01]. Models are numbered from 1 to 127, with the group of single-cone models on left and the single-membered 7-cone group on right. The “blind” choice distribution is proportional to the number of models in each group. B: overall patterns of excitation and inhibition found at stimulus onset differed among spectral types of cone. The fraction of excitatory (red) or inhibitory (blue) signals found for the 40 “best” onset models is illustrated for each cone type. For all cones, the fraction of modeled excitation was 0.40 (dashed red line). The fraction of modeled inhibition was 0.15 (dashed blue line). The significance of the ways individual cone types differed from these overall means is indicated by asterisks (χ² tests, see text). Vu358 (UV) cone signals were significantly different, being 93% excitatory, whereas Vg480 (G3) green cones and Vb440 (B2) blue cones favored inhibition. C: in sustained-ON (Sust-ON), transient-ON (Trans-ON), and ON-OFF GC types, modeled opponency was the most common pattern, and nonopponency was rare. “Undetermined” models are those that did not meet the criterion r² values. D: for ON-OFF GCs where opponency could be determined for both ON and OFF signals, it was common that both phases were color opponent, but the type of opponency, or nonopponency, differed. WT, wild type.

Nonopponent Responses

Nonopponent GCs were excited at stimulus onset by inputs from 1–3 cone-opsin with a mean of 1.6 ± 0.9 cones (SD, 8 best or indistinguishable from best models). These GCs, as a group, were most excited by UV cones (Vu358, 6/8 models), and by R1 cones (Vr575, 4/8 models). Additional excitation from G3, G1, or B1 cones occurred only in single examples. Nonopponency did not depend on response time course. Sustained-ON, transient-ON, and the onset responses of ON-OFF cells could be nonopponent (Fig. 7, C and D). The sustained-ON discharges of a nonopponent, ON-transient GC (Fig. 8A) are prominent at short wavelengths (330 nm, 370 nm) and small, yet discernable, at longer wavelengths. The best-fitting model (Fig. 8B, model 55, r² = 0.8248) indicates large-amplitude and very significant excitation from UV cones (Vu358), together with lower-amplitude yet significant excitation from both R2 and G3 cones (Vr575, Vg480). The residual variance of one other three-cone model (Fig. 8C, model 52, r² = 0.8234) was indistinguishable from the best model (F_55,52 = 0.992, P ≥ 0.95). The model 52 interpretation was similar to model 55, except that G1 excitation substituted for G3 (Vg460 for Vg480). The two modeled GC spectral sensitivities (Fig. 8D) are nearly indistinguishable, as are the values for peak and trough wavelengths.

Figure 8.

Excitation patterns in a nonopponent, trichromatic ganglion cell (GC). A: UV stimuli (370 nm) evoke maximal spike rates. Wavelengths are color coded and shown above each stimulus trace. The number below each wavelength is the irradiance in log(quanta·µm⁻²·s⁻¹) (where hν = quanta). B: the best model was model 55, showing high-amplitude and very significant excitation from UV cones (Vu358) and less significant, lower-amplitude excitation from R1 and G3 cones (Vr575 and Vg480). C: although the r² value of model 52 was less than model 55, the residual variance was indistinguishable (F test). In model 52, G1 cone excitation substitutes for G3 excitation. Both models 55 and 52 are trichromatic and, lacking inhibition, nonopponent. In B and C, saturation amplitudes for cone signals are given as a fraction of the maximum light-evoked spike-discharge in the data set. ++++, ++, and + give significance of excitatory signals (GraphPad convention). D: the fractional light-evoked discharge amplitudes for model 55 (black line) and model 52 (magenta line) at constant-quantal [4.4 log(quanta·µm⁻²·s⁻¹)] stimulation across the wavelength spectrum are nearly identical, as are the spectral peaks and troughs (+symbols). Four C6 data sets from a single GC are combined.

Color-Opponent Responses

Modeling suggests that many subsets of the seven-opsin scheme could contribute to GC color opponency (Fig. 4C, Fig. 5B). Individual color-opponent GC signals were composed of a mixture of excitatory and inhibitory inputs from three to six different cone opsin types with a mean of 4.4 ± 0.7 cones (SD, 32 best or indistinguishable from best models). Overall, 80% of GC spectral models (32 of 40) were color opponent at response onset (Fig. 7C). Cells receiving inputs from four or more cone opsin types were always color opponent. Sustained-ON, transient-ON, and both the ON and OFF phases of ON-OFF GCs were color opponent (Fig. 7, C and D).

In addition to the characteristic mean frequencies of excitation, inhibition, or absence for each of the seven-cone opsin signals (Fig. 7B), some opsin pairs were selectively associated. Of the 21 possible pairings of seven cones, contingency tables with expectation values based on nonpreferential frequencies of response pairing identified three cone pairs whose frequencies lay outside of expectation values (Fig. 9). These cone pairs were R1 versus R2 (Vr575, Vr556), G3 versus G1 (Vg480, Vg460), and B2 versus B1 (Vb440, Vb415). In all three pairs, synergistic actions (both excitatory or both inhibitory) were absent and opponent activity (one excitatory, the other inhibitory) was more common than expectation values. The overall chances for these distributions (Fig. 9) were ${\chi }_{\mathrm{R}1,\mathrm{R}2}^2$(3,40) = 11.8, P ≤ 0.01; ${\chi }_{\mathrm{G3},\mathrm{G1}}^2$(3,40) = 21.5, P ≤ 0.0001; ${\chi }_{\mathrm{B2},\mathrm{B1}}^2$(3,40) = 27.1, P ≤ 0.00001. The most associated pairs of synergistic (nonopponent) excitatory inputs were UV and R1 (n = 16) and UV and G1 (n = 15). These were not preferential associations, as excitatory UV signals occurred in 93% of GCs [${\chi }_{\mathrm{R}1,\mathrm{UV}}^2$(3,40) = 0.25, P = 0.97; ${\chi }_{\mathrm{G}1,\mathrm{UV}}^2$(3,40) = 0.74, P = 0.86].

Figure 9.

Selective association among cone pairs. In a 7-cone scheme there are 21 unique pairs. Three of these showed association patterns that differed from what would be expected. The expectation value is based on nonpreferential pairing according to the frequency that cone-type excitation or inhibition were found in the overall population of ganglion cells (GCs). Selective associations were R1-R2 (Vr575 and Vr556) (A), G1-G3 (Vg460 and Vg480) (B), and B1-B2 (Vb440 and Vb415) (C). No pairing of dual excitatory or dual inhibitory signals (synergistic) were found in any of these pairs, and excitatory-inhibitory pairing (opponent) exceeded expectation values. The probabilities of these distributions (P) are calculated from χ² tests (df = 3, n = 40). Both best-fit models and models indistinguishable from best fit are counted.

Transient-ON Ganglion Cells

In best models of transient-ON GCs (n = 5), four (80%) were opponent, combining four or five cone signals at light onset, and one was nonopponent, utilizing only a single cone. Transient-ON GCs did not differ from sustained-ON GCs in the numbers of onset cone signals [χ²(4,5) = 2.1, P = 0.72]. The mean transient-ON peak time was at 134 ± 11 ms after stimulus onset (SE, n = 12). In the color-opponent ON-transient GC shown in Fig. 10, 650-nm and 610-nm stimuli evoked excitatory discharges followed by a poststimulus discharge suppression. A 370-nm stimulus also evoked an excitatory discharge for all irradiances illustrated. Response waveforms (Fig. 10A) are transient, as the ON discharges are completed within the 300-ms Cs2 stimulus. The best-fitting cone model was pentachromatic (Fig. 10B, model 117). The next best fitting model (model 118) was indistinguishable (F_117,118 = 0.996, P = 0.97). Both models are pentachromatic, sharing four cone signals, but in the fifth, B1 (Vb415) substitutes for UV (Vu358) in model 118. The spectral shapes, peaks, and troughs are similar.

Figure 10.

Transient-ON spectral patterns in an opponent ganglion cell (GC). A: red (650 nm, 610 nm) and UV (370 nm) stimuli evoked maximal spike rates in response to the Cs2 spectral protocol. Wavelengths appear above each nested set of 3 spike-rate discharge records, which are in response to different irradiances. The irradiances are selected from 7 total to give matching irradiances at each wavelength. These are equal quantal responses. The quantal brightness is given in log(hν·µm⁻²·s⁻¹) to the right of each stimulus trace, with black corresponding to the dimmest, red to the brightest, and blue intermediate. B: model 117 fit the full, 280-point, GC data set best. R1 (Vr575), G1 (Vg460), and UV (Vu358) were modeled as excitatory and R2 (Vr556) and G3 (Vg480) as inhibitory. + and − signify excitation and inhibition, respectively, with the number of each symbol denoting the level of significance (as in Figs. 4 and 5). ns, Not significant. C: the residual variance for model 118 was indistinguishable from model 117 (F_117,118 = 0.996, P = 0.97). Both models are pentachromatic and color opponent. Model 118 differs from model 117 only in the substitution of B1 excitation (Vb415) for UV excitation (Vu358). D: constant-quantal spectral curves for models 117 and 118. The curves are similar in shape and in peak and trough positions.

Sustained-ON Ganglion Cells

Sustained-ON GCs are defined by ON-excitatory discharge kinetics extending beyond 300 ms (the stimulus duration in the Cs2 protocol). The peak discharge in sustained-ON GCs occurred 134 ± 9 ms after stimulus onset (SE, n = 23). ON-sustained ganglion cells were 53% (23/43) of the GCs recorded (Fig. 6D). Of sustained-ON GCs with accepted models, 85% were color opponent (22 of 26 accepted models; Fig. 7C). Color-opponent sustained-ON GCs received inputs from three to six cone opsin signals; nonopponent sustained-ON GCs received inputs from one or three cone opsin signals. The individual examples of sustained-ON GCs in Figs. 4 and 5 are color opponent, with mixtures of excitation and inhibition from five and six cones, respectively. Three cone signals combine in the nonopponent sustained-ON GC of Fig. 8.

ON-OFF Ganglion Cells

It is unusual to think of ON-OFF retinal neurons as color opponent. ON-OFF amacrine cells in adult zebrafish are red cone monochromatic (42). However, larval ON-OFF GCs may be color opponent in both ON and OFF phases of the response. There were nine ON-OFF GCs with accepted spectral models for the ON response, and seven were also successfully modeled for the OFF response. For the ON phase one to five cone signals contributed and for the OFF phases two to six cone signals were combined in the best model fits. The mean number of cone opsin signals was 3.1 ± 1.6 (SD) for the ON phase. For the OFF phase 3.9 ± 1.6 cones were detected. For the ON phase, 67% of best models were color opponent (6 of 9). In the OFF phase, 100% of accepted models were color opponent.

An ON-OFF GC example (Fig. 11) is trichromatic (Fig. 11C) and transient (Fig. 11B) at stimulus ON and hexachromatic (Fig. 11C′) and sustained (Fig. 11B) at OFF. For both ON and OFF discharges, UV wavelengths were the most effective (Fig. 11, A and B). The OFF discharge, in addition, is excited by long wavelengths (650–570 nm) (Fig. 11B). Spontaneous activity is inhibited by midspectral wavelengths (490–530 nm; Fig. 11, A and B). Model 36 best fits the ON discharge pattern. This combines excitation from UV (Vu358) and G1 (Vg460) cones with inhibition by G3 cones (Vg480), a trichromatic opponent pattern. (Fig. 11, C and D). There were no indistinguishable from best fit models. Model 64 was next best (F_36,64 = 0.989, P = 0.92) and, in addition to greater residual variance, used four cones. Model 125 best fit the spectrum of the OFF discharge. This combined six cone signals, including all those found in the ON discharge plus three additional signals. These additional signals are R1 (Vr575) excitation and both R2 (Vr556) and G1 (Vg460) inhibition. R1 contributes long-wavelength excitation to the OFF discharge, not seen in the ON discharge (Fig. 11D′). For the OFF discharge, there were no other models with residual variance that was indistinguishable from best fit. The closest was model 124 (F_125,124 = 0.991, P = 0.943), also a six-cone model, making a substitution B1 (Vb415) for B2 (Vb440) inhibition.

Figure 11.

ON-OFF opponent ganglion cell spectral patterns. A: in this cell both ON and OFF discharges are prominent particularly with UV excitation (370 nm, 330 nm). At long wavelengths (650 nm, 610 nm, 570 nm) only an OFF discharge is apparent (at ∼500 ms), with a suggestion of impulse silencing during stimulation. Wavelengths appear above each nested set of 3 spike-rate discharge records, which are in response to dim, intermediate, and bright irradiances. The irradiances, selected from 7 total at each wavelength, are matched for comparable quantal stimulation. These are equal quantal responses across wavelengths. The quantal brightness of each monochromatic stimulus is given in log(hν·µm⁻²·s⁻¹) on left of each stimulus trace. B: mean response waveforms for all stimuli at long (650–570 nm), middle (530–490 nm), and short (370–330 nm) wavelengths. At middle wavelengths there is discharge quieting, at long wavelengths an OFF response, and at short wavelength an ON-OFF pattern. ON (C) and OFF (C′) responses are fit by different spectral models, both opponent but employing different numbers of cone opsin signals. All onset model cone signals are included in the offset model, but further cone signals are added. + and − are the probabilities of nonzero excitatory and inhibitory cone opsin signals, following the GraphPad convention for significance (t tests). Modeled discharge amplitude spectra for constant quantal irradiances at ON (D) and OFF (D′): the response onset model (model 36) is trichromatic but generates 4 spectral phases. Long-wavelength excitation is absent. The offset model (model 125) is hexachromatic and pentaphasic. The stimulus protocol is Cs2 (280 stimuli).

DISCUSSION

All cone types in zebrafish have differentiated by 4 dpf, with R cones more abundant than G cones and UV cones more abundant than B cones (59). Our results indicate that 1–2 days later (at 5–6 dpf), the retina is processing multispectral inputs that begin with light absorption by cone photoreceptors expressing one of at least seven different opsins (37, 38, 47). In adults, these opsin genes are distributed among the four morphological types of cone photoreceptors: the principal member of the double cone (R), the accessory member of the double cone (G), long single cones (B), and short single cones (UV) (60). In larvae, differences in cone morphology are less clear, but mosaic patterns of opsin antigenicity also suggest four distinct cone opsin types (48). However, antigenicity does not discriminate among opsin types arising from gene duplications. This suggests that several morphological cone types express more than one opsin. Our model is based on the possibility that each of seven larval cone opsins takes physiological ownership of a single cone (52, 61). In the scheme we present, R cones are either R1 or R2 expressing, G cones either G1 or G3, and B cones either B1 or B2 (38). UV cones express only one opsin gene (37, 38, 47). Many studies, including our own past work, have modeled zebrafish retina as a four-opsin system. Some adult horizontal cells were modeled as tetrachromatic, with combinations of signals from R1, G3, B1, and UV cones (41). All 15 cone combinations in this four-opsin adult scheme are included in the extended seven-opsin scheme used here, where they compete with additional combinations resulting from the inclusion of three additional opsins. Model 1 (UV cones only), one of the adult models, occurred three times; model 55 (R1, G3, and UV cones), another adult model, was selected twice. This accounts for 5 best fits among the 23 where models were used more than once.

Different opsin genes produce overlapping opsin absorbance spectra, distinct peak absorbance wavelengths, and different intrinsic cone sensitivities (Table 1). It has been suggested that changes in either specific amino acid residues or protein folding are able to confer the different peak absorbance wavelengths to similar opsins (62) but sensitivities depend mainly on the transduction cascade (63). Patterns of opsin gene expression differ between larval and adult cones (47). Our analysis does not exclude the possibility that, as a group, larval GCs make use of not just four but nearly all cone opsin genes in generating spectral responses. Individually, our data are consistent with the idea that each GC receives inputs from one to six different cone opsin types, indicating that signaling pathways may be present for all opsin types.

The number and diversity of GC morphologies in larval zebrafish retina are not known. Time-lapse imaging of larval GC dendrites in the IPL identified 15 stratification types (32), which is slightly more than the 11 morphological types in adults (33). The clustering of narrowly versus broadly stratified dendritic patterns in strike zone versus peripheral regions of larval retinas has recently been studied (5). As observed with other retinal neurons, the spectral diversity of larval GCs is likely larger than their morphological diversity. Most GCs (95%) were modeled as receiving signals from one to five cone opsin types in response to light onset. Onset responses evoked by stimulation of the different cone opsins may be relayed through the entire retina via distinct channels and recombined in various patterns by GCs. These channels are initially conveyed from outer retina, contributing to the diversity of horizontal cell spectral responses (41) and bipolar cell connectivity patterns (43), to inner retina, resulting in spectrally multichromatic and multiphasic amacrine cells (42). We suspect that some of the preferential spectral clustering at the GC level, of both individual cone signal types and cone signal pairs, represents bipolar cell spectral types. Calcium imaging of larval bipolar terminals revealed strong clustering for UV, monochromatic, excitatory terminals, both in the strike zone and in peripheral retina. Long-wavelength excitatory terminals were also dense, particularly in peripheral retina (64). These may represent UV and red cone bipolar pathways to larval GCs. Five color-opponent clusters of bipolar terminals were also reported, providing a potential substrate for opponent clustering of opsin pairs at the GC level (64).

ON and OFF chromatic and achromatic larval ganglion cells have been observed with two-photon calcium imaging of islet2b:mGCaMP6fgCamp transgenics (5). Steps of simulated white light revealed several of what we classify as sustained-ON types and an occasional OFF type. Inspections of kernel responses to binary tetrachromatic stimuli (“natural white”) were the principal classification method, which yielded a predominantly ON pattern, though not to the extent found in the present results. ON-OFF types are not detectable with first-order kernels. Cluster analysis of calcium imaging in optic tectum revealed OFF and ON-OFF clusters (but not ON) (65). Classification into ON, OFF, and ON-OFF categories may be strongly influenced by stimulation protocols. In the tectal study, the stimulus included an OFF-favoring stimulus (dimming from a steady background), which was not used here or in the retinal imaging study (5).

In the retinal imaging study (5), cells were about equally split between chromatic (different polarities to different colors) and achromatic (like responses to all colors). UV-ON responses were particularly prominent in the strike zone. With tetrachromatic, sequential, sinusoidal red, green, blue, and UV LED stimulation, calcium imaging of GC axon terminals in tectum and other brain areas revealed a rich variety of chromatic patterns (66). These patterns were not targeted to particular brain areas. Despite differences in stimulation protocols, and calcium imaging versus spike recording, all studies agree on a very wide variety of spectral properties in larval GCs.

Our analysis did not attempt to bin GCs into just a few spectral types but rather to examine all the 127 potential combinations of seven-cone signals. Interestingly, we did not find that our cohort of larval GCs was limited to just a few spectral combinations, nor did we find that models were selected at random from the 127 possibilities. Model 76 was used five times in 40 fits (P < 0.0001, binomial distribution). In this model R2 and UV cones were excitatory, whereas G3 and G1 were mutually opponent, in either direction. Models 1 and 86 appeared three times each (P < 0.01). Model 1 represents UV excitation only. Model 86 is like model 76, except that excitation by R1 cones is substituted for R2. This gives a total of 8 of 40 models where either R1 or R2 and UV cones were excitatory, while G1 and G3 formed an opponent pair. Such results further allow us to reject the hypothesis of indiscriminate sampling of cone combinations. In zebrafish larvae there are preferential cone opsin input patterns to GCs, and there are many more patterns than commonly supposed.

Old world primate color physiology is represented on three axes, blue-yellow, red-green, and luminance. The blue-yellow axis is often considered to be driven by blue-yellow opponent neurons (12, 67), the red-green axis by red-green opponent neurons (13), and the luminosity axis by achromatic luminosity (e.g., white detecting) neurons (13). Zebrafish larvae do not follow this pattern. Blue-yellow opponency is replaced by neurons that are excited by both UV and long-wavelength cones. There does not appear to be a luminosity neuron, that is, a cell excited by all cones. Color-opponent ganglion cells are rarely dichromatic as they are in old world primates but involve three to five cones in multiple opponent patterns. The present results suggest that the zebrafish visual system operates in a very different color space from old world primates or mammals. The strong GC excitation by UV light in zebrafish larvae bears a resemblance to salmonid juveniles (16, 68) and turtles (19, 68).

Conclusions

There are as many as seven distinct opsin signals visible at the GC level in larval zebrafish, and one to five of them typically combine as excitation and inhibition to provide a large range of detectors for spectral textures. These combinations occur as the different cone signals are passed through spectrally distinct horizontal, bipolar, and amacrine cells presynaptic to GCs. Although excitation by UV, R1, and R2 cone types was typical, as was inhibition by G3 and B2 cones, there was a high diversity and complexity in GC spectral responses. Thus, there appears to be a greater variety of spectral patterns and interactions in larval retina compared to the more limited cone interactions found in adult zebrafish horizontal cells (41) and the red-green opponent spectral patterns in adult goldfish GCs (7, 45). This suggests that larval zebrafish have fine-tuned spectral detection abilities with which to discriminate aspects of the visual scene.

GRANTS

The research was supported by the Intramural Program of the National Institute of Neurological Disorders and Stroke, National Institutes of Health.

DISCLOSURES

No conflicts of interest, financial or otherwise, are declared by the authors.

AUTHOR CONTRIBUTIONS

V.P.C. and R.N. conceived and designed research; performed experiments; analyzed data; interpreted results of experiments; prepared figures; drafted manuscript; edited and revised manuscript; and approved final version of manuscript.