A heterogeneous population code at the first synapse of vision

Abstract

Vision begins when photoreceptors convert fluctuations in light intensity into temporal patterns of glutamate release that drive the retinal network. The input-output relation at this first stage has not been studied in vivo so it is not known how it operates across a photoreceptor population. Using glutamate imaging in zebrafish, we find that individual type 1 cones (PR1; ancestral red cones), which dominate daylight vision in non-avian vertebrates, encode visual stimuli with high reliability and time-precision but routinely vary in sensitivity to luminance, contrast and frequency across the population. Variations in input-output relations are generated by feedback from the horizontal cell network that effectively decorrelate feature representation. A model capturing how zebrafish sample their visual environment indicates that heterogenous cone outputs expand the dynamic range of the retina to improve the coding of natural scenes. Moreover, we find that different kinetic release components are used to encode distinct stimulus features in parallel: sustained release linearly encodes low amplitude light and dark contrasts, but transient release encodes large amplitude dark contrasts. This study reveals an unexpected degree of functional heterogeneity within a population of cones and illustrates how separation of different visual features begins in the first synapse in vision.

It is currently unknown how uniform is signalling at the first synapse of vision in vivo. Here, the authors show neighbouring PR1 (red) cones differ in contrast sensitivity and timing due to locally stochastic horizontal-cell feedback, thereby extending total dynamic range.

Introduction

The light-encoding properties of photoreceptors underpin the performance limits of vision^1,2. While we have a detailed understanding of the processes that convert light into a photocurrent^3,4, we have much less understanding of the output signal by which photoreceptors drive the retinal circuit - the synaptic release of glutamate^5–7. Here we use larval zebrafish to make an in vivo investigation of the way in which cone photoreceptors encode visual stimuli.

Spatial correlations in natural visual scenes^8,9 cause most stimuli to be encoded by the simultaneous activity of several cones. It is therefore important to understand how the input-output relation varies across a population of same-type photoreceptors^10,11. A similar relation across the population would indicate a high degree of redundancy in the signal transmitted at the first stage of vision but it is unclear if this is the case^7,10,12 because individual cones can be re-tuned by cone-intrinsic factors^6,7,13, their surrounding circuitry^6,14,15 and/or the local neuromodulator environment¹⁶. We have investigated how a population of cones encodes visual stimuli by monitoring glutamate release from synaptic pedicles in vivo. We focus on type 1 cones^17,18 (PR1, previously: ancestral red cones¹⁹), which likely represent the ancestral general-purpose greyscale system of the vertebrate eye^19–22. More than 90% of cones in the human eye, comprising both red/L and green/M variants²³, are of this highly conserved photoreceptor type^17,18,24,25.

We find that while individual PR1 generate synaptic outputs that are exceptionally reliable and precise in time, the population as a whole is heterogenous in terms of sensitivity to luminance and contrast. For example, while some cones exhibited approximately linear contrast-response functions, others were strongly rectifying, signalling negative contrasts more strongly than positive. Similarly, while glutamate release rates from some cones reliably followed temporal contrast up to 20 Hz, other cones ceased to respond above 8 Hz. Blocking inhibitory feedback from horizontal cells causes all red cones to transmit signals at a lower frequency and become strongly biased to negative contrasts, demonstrating that the outer retinal circuitry plays a key role in determining the input-output relation of the first neurons in vision. A model of bipolar cells summing inputs from cones, as zebrafish sample their visual environment, indicates that heterogeneity in cone signals expands the retina's dynamic range, improving the coding of natural scenes.

Results

Isolating the synaptic output of individual PR1

Cones drive the retinal circuit through synaptic pedicles that are invaginated by post-synaptic bipolar cells and horizontal cells to create extracellular compartments tightly sealed against the surrounding circuit^5,26 (Fig. 1A, B). In larval zebrafish, there is only one such invagination per cone¹⁴, and we exploited this anatomical arrangement to use the dendritic tips of horizontal cells expressing the fluorescent glutamate sensor SFiGluSnFR²⁷ as glutamate antennae. A spatially isolated SFiGluSnFR hotspot therefore provided a glutamate read-out from an individual cone pedicle^6,7, each typically comprising some 2-3 ribbons with a total surface area of 0.2–0.5 µm² (Supplemental Figure S1). Applying line scans at 1 kHz typically allowed us to sample the outputs of 2–5 cone pedicles simultaneously (Supplemental Video 1).

Fig. 1. In vivo monitoring of light driven glutamate release from zebrafish ‘red cones’ PR1.

A, B Dendritic tips of horizontal cells invaginate the synaptic pedicles of cones, and we used SFiGluSnFR-expression in HCs as postsynaptic glutamate antennas of cone release (see also ref. ⁶). C In vivo spectral tuning of larval zebrafish cone types (PR1, PR2, PR3, PR4¹⁷ (previously: ancestral red, green, blue, UV¹⁹) with 590 nm stimulus wavelength overlaid (modified from ref. ⁹²). D Maximum intensity projection of the field of view (FOV) from a typical two-photon recording, focused on a portion of the nasal outer retina of a 6 dpf larval zebrafish expressing SFiGluSnFR in HCs (image representative of >50 FOVs). Red line indicates approximate positioning of line scan (l.s.) in (E). E Kymograph of the line scan indicated in (D) during widefield visual stimulation. Right: the fluorescence signals over time (fluorescence profile, solid line) from two neighbouring cone pedicles (labelled PR1 and PR2) were extracted based on the two Gaussians that best approximated their spatial profile (dashed line) (based on ref. ²⁸). F, G Fluorescence traces from (E) for two different parts of the stimulus sequence, as indicated, and H, the corresponding mean responses to 98 off-steps of light ( ± sd, shading). I, J Deconvolved versions of fluorescence traces shown in (F, G Methods). K Superposition of three amplitude-normalised responses from the PR1 in (E) illustrates their highly stereotyped time-courses. Source data are provided as a Source Data file.

To isolate responses of PR1, we used wide-field stimuli of amber light ( ~ 590 nm) because in larval zebrafish, only PR1 exhibits excitatory responses to light decrements at this wavelength¹⁴ (Fig. 1C). PR2¹⁷ (ancestral green cones¹⁹) are also sensitive to amber light, but they are red-opponent and therefore exhibit inhibitory responses to the same light decrements¹⁴. The reliability of this strategy for identifying PR1 was confirmed by co-labelling PR1 with tdTomato (Supplemental Fig. S2).

Examples of synaptic responses are shown in Fig. 1D–H. At the start of this experiment, a light flash was applied, bright enough to reduce the fluorescence signal in the pedicle labelled PR1 to close to background fluorescence levels while also suppressing the noise caused by vesicle fusion in the dark, indicating complete, or almost complete, suppression of glutamate release (Fig. 1F). From this bright background, off-steps were applied by turning off the light for 40 ms. This pedicle was identified as belonging to a PR1 because it exhibited large amplitude glutamate transients in responses to the off-steps (Fig. 1G, H). In contrast, the neighbouring pedicle labelled PR2 exhibited transient suppressions of glutamate release, indicating that it belonged to a PR2. Beyond these two basic response types, other pedicles did not generate reliable SFiGluSnFR signals, indicating that they belonged PR3 or PR4¹⁷ (ancestral blue/UV, respectively; not shown).

A notable feature of the glutamate transients generated by PR1 pedicles was that when the amplitude varied, the shape did not (Fig. 1K). This property allowed us to assess relative changes in the rate of glutamate release by deconvolving the SFiGluSnFR signal using the temporal kernel derived from the average response to brief off-steps (Fig. 1I, J). A similar approach has been used to detect individual vesicle release events at the synapse of retinal bipolar cells²⁸. In cones, however, it was not possible to reliably disambiguate the signals from single vesicles. Instead, we interpret our cone data as reflecting the summed signals from multiple vesicles released from multiple ribbons⁷. Below, we use this technique to investigate how cone synapses encode light intensity and contrast.

Variable sensitivity to light at the synaptic output of individual cones

The usual approach to investigating the light sensitivity of photoreceptors, for instance, when measuring photocurrents, would be to apply brief flashes of fixed duration but different intensities on top of a dark background²⁹. We did not use this approach because the dark signal at the cone synapse is noisy (e.g., Fig. 1F, I) and the response to light is a decrease in the SFiGluSnFR signal caused by an interruption in the continuous release of vesicles. Measuring a decreasing signal from a high and noisy baseline degrades the signal-to-noise ratio compared to measuring an increasing signal from a low and stable baseline. We therefore carried out these measurements using off-steps, where the stimulus strength was the number of photons missing when the light was turned off, which was directly proportional to the duration. A sequence of off-steps 5-100 ms in duration were presented in a pseudorandom order. Electrophysiological measurements in vitro show consistent responses between cones³⁰ but imaging the final synaptic output in vivo revealed that PR1 had significantly different sensitivities to light, as illustrated by the two nearby pedicles in Fig. 2A–D (see also Supplemental Video 2). While cone 1 reliably responded to 10 ms flashes and saturated at 20 ms, cone 2 only started to respond to flashes of ~18 ms, and this response did not saturate until the flash duration was 50 ms (Fig. 2C, D).

Fig. 2. Distinct off-step-sensitivities in neighbouring PR1 in vivo.

A–D Simultaneously recorded glutamate release (A, C, D) from two neighbouring cones (B) in response to a pseudorandom sequence of 100% contrast off-steps of varying duration (5 to 100 ms) (representative of >50 FOVs) in a cx55.5:SFiGluSnFR larval zebrafish. A and C show raw deconvolved signals at two timescales as indicated, while D shows their stimulus-sorted means ( ± sd, shading). E Stimulus response functions of the two example cones from A–D are plotted as mean ± sd (error bars, 25 repeats per stimulus condition), with sigmoidal fits (red). F Amplitude-normalised fits of contrast-response functions of n = 70 cones from 12 fish and G summary of their corresponding inflection points d_1/2 (see also arrowheads in E, Methods). Hartigan Dip test statistic = 0.046, indicating bimodality. However, fit inflection points strongly correlated with other metrics of heterogeneity (Supplemental Fig. S3), and Principal Component Analysis across these metrics did not reveal robust substructure in the overall population of PR1 responses. Source data are provided as a Source Data file.

The relationship between flash duration (d) and relative rate of glutamate release (R) could be described as a sigmoid function,

$$R=\left(R_{\max }/\left(1+\exp \left(\left(d_{\frac{1}{2}}-\mathrm{d}\right)/\mathrm{k}\right)\right)\right)$$ 1

where R_max is the saturating response, d_1/2 is the off-step duration generating the half-maximal response and k is a constant (Fig. 2E). Importantly, the range of flash durations used fall within the integration period of cones where flash duration and intensity are related linearly³¹. The stimulus-response functions for cone 1 and 2 differed both in terms of R_max, reflecting synaptic gain, and d_1/2, reflecting light sensitivity. Qualitatively similar differences were routinely observed in simultaneously recorded neighbouring PR1, indicating that response heterogeneity reflected biological rather than experimental variation. A series of fits to the stimulus-response function of 70 cones from 12 fish, normalized to the maximum response, is shown in Fig. 2F, and the distribution of d_1/2 in Fig. 2G. The d_1/2 value ranged from 10 to 41 ms, and the mean value of d_1/2 was 21 ± 6.6 ms ( ± sd). These results demonstrate a substantial degree of response heterogeneity across PR1 in vivo.

High reliability and temporal precision of the synaptic signal

A key property of any neural signal is its reliability. Neural information is degraded by noise that causes responses to vary when a stimulus is repeated, and synapses are a major source of such noise because of the stochasticity of the presynaptic processes that control vesicle fusion^2,32,33. At many central synapses these processes obey Poisson statistics³⁴ with responses to a repeated stimulus showing a coefficient of variation (CV = sd/mean) of 1. Although it has sometimes been assumed that ribbon synapses of cones also obey Poisson statistics³⁵, more recent work indicates that this is not the case in rods³⁶.

To explore the reliability of the cone output, we delivered many off-steps (usually 98) of a fixed duration (40 ms) (Fig. 3A). For individual PR1, the amplitudes of responses to this repeated stimulus were consistently found to be normally distributed (Fig. 3B). The scatter plot in Fig. 3C shows the relation between the standard deviation and mean of the responses output from 32 cones: in all cases the CV was far below 1 (dashed line) and averaged 0.13 ± 0.05 F’s^-1 (mean ± sd; Fig. 3D). Re-expressing this metric as the signal-to-noise ratio (SNR = mean²/variance) yields an average of 93 for the output of one pedicle. The SNR will depend on the square root of the number of ribbons if these are affected by independent sources of noise. With this assumption and taking an average of two ribbons per pedicle we can place a lower limit of SNR = 66 for a single ribbon synapse.

Fig. 3. The cone output is reliable and temporally precise.

A, B Glutamate responses of one example cone to 98 identical 40 ms duration off-steps and (B) distribution of response amplitudes with Gaussian fit superimposed (solid black line) (mean = 6.2 ± 0.9 ( ± sd)) in a cx55.5:SFiGluSnFR larval zebrafish. C Relationships of mean versus standard deviation of response amplitudes from n = 32 cones in 13 fish systematically fall below the equivalence line where the mean equals the standard deviation, as would be expected from a Poisson release process (dashed). D The coefficient of variation of the data shown in C (two-sided Pearson correlation, r(30) = −0.46, p = 0.008). E–G Overlay of all 98 responses from (A) demonstrates a high degree of temporal precision, here quantified as time-to-threshold (F), and temporal jitter (G, standard deviation of time-to-threshold; F,G violin pots show kernel density estimation, with mean as filled black diamond). H Temporal jitter was stable for stimulus durations above 30 ms, plotted as mean ± sd (error bars; n = 70 cones, 12 fish). Source data are provided as a Source Data file.

Responses of sensory neurons vary in their timing as well as their amplitude. To investigate this aspect of the signal transmitted from cones we began by measuring the time-to-half peak (t_1/2) of responses to off-steps, as shown in Fig. 3E. The mean value of t_1/2 after a 40 ms off-step was 29 ± 6.5 ms across a sample of 32 PR1 (Fig. 3F). The temporal precision of these responses was measured as the standard deviation in t_1/2 at each cone output (Fig. 3G). Some cones displayed temporal jitter in the order of 1-2 ms, and the average was 3.1 ± 0.8 ms. Similar levels of temporal precision were observed for flashes of 30 ms and longer, but jitter increased as off-steps became shorter (Fig. 3H). A temporal precision of a few milliseconds can also be observed in the responses of bipolar cells and ganglion cells to high-contrast stimuli^37,38.

Together, the results in Figs. 1–3 demonstrate that while the ribbon synapses of PR1 transmit the first visual signal with extreme reliability and temporal precision (Fig. 3), the luminance sensitivity varies substantially across the population (Fig. 2). Similar variations were observed across different average levels of luminance (Supplemental Fig. S4).

Variations in sensitivity to contrast

Natural scenes rarely contain an off-step to darkness. The major task of the cone array is to encode continuous fluctuations in light intensity around an average luminance (Fig. 4) that varies much more slowly than the frequencies driving most vision (Fig. 5). The fact that intensity varies up and down is critical because neurons are rarely linear – changes in opposite directions are often encoded unequally. Synapses are key sites for such rectification. We therefore asked how far the first synapse in vision rectifies and found that the answer depended on timescale (Fig. 4). On short timescales, PR1 encode dark-transitions, but not light-transitions, with a transient burst of glutamate release. However, after this transient phase, a more sustained component of release allowed some cones to linearly represent both positive and negative contrasts. We reach the above conclusions after presenting 0.5 s steps from a photopic background ( ± 10-100% contrast, 1 Hz, 50% duty cycle). Figure 4A–C shows representative responses and their basic quantification from three PR1. Consistent with previous measurements of ribbon-mediated release at photoreceptor synapses^7,39,40, responses consistently displayed both transient and sustained components (Fig. 4D–F). Cone 1 responded to both negative and positive contrasts up to 50% in an approximately linear manner, cone 2 only responded to strong negative contrasts, and cone 3 mainly responded to positive contrasts.

Fig. 4. Differential encoding of positive and negative contrasts.

A, B Glutamate release from one example cone to a pseudorandom sequence of 500 ms light and off-steps from an intermediate baseline shown at two different timescales as indicated, in a cx55.5:SFiGluSnFR larval zebrafish. C cone 1, as (A, B), arranged by contrast (mean ± sd (shading)), alongside two further cones processed in the same way (cones 2 and 3). Note that the three example cones exhibit different types of light/dark biases. D Expansion and superposition of cone 1’s mean 100% positive and negative contrast responses as shown in C (dashed and solid lines, respectively) with transient (T) and sustained (S) response components annotated. E, F Contrast response functions of transient (E) and sustained (F) release components of n = 97 PR1 from 30 fish with mean superimposed (red). G Distribution of each cone’s dark-light-index (DLI, Methods) of transient and sustained components and linear fit (red). H–J A simple model illustrates how slow linear (here: sustained) and fast non-linear (here: transient) encoding can highlight very different aspects of the same natural scenes (Methods), here showing translational movement through a typical shallow underwater scene from a zebrafish natural habitat in India (from ref. ⁴³). Each RGB-input-pixel (H) was converted into an 8-bit brightness signal over time (I, top), and then filtered to mimic a strongly dark-biased transient cone response (I, middle), or a linear sustained response (I, bottom). The results are summarised in (J) for two movie frames as indicated. Note that the sustained representation recapitulates most detail in the input, while the transient representation separates out the high-contrast dark edges, which are largely concentrated in the foreground. See also Supplemental Video 2. Source data are provided as a Source Data file.

Fig. 5. Heterogeneous encoding of temporal contrast.

A Release from two example cones driven by a chirp stimulus, decelerating over 20 s from 20 Hz at 100% contrast. B Histogram of spectral centroids of glutamate responses to the stimulus shown in (A) from n = 49 PR1, 18 fish. C, D release from two further examples cones in response to continuous 20 Hz sinusoidal modulation at 100% contrast, with an expansion of the traces shown in (D), and population histogram of detected release events from the population data (bottom, n = 49 cones, 18 fish). E Distributions of detected responses during the stimulus phase for cone 3 (top), cone 4 (middle), and for the population data (bottom) which have a vector strength of 0.98, 0.41, and 0.61, respectively. Source data are provided as a Source Data file.

A range of contrast-response functions were observed in our sample of 97 cones. The most consistent were generated by the transient output and this was always biased to negative contrasts (Off; Fig. 4E). The sustained component tended to be more linear around low contrasts before saturating in some cases (On; Fig. 4F). Bipolar cells post-synaptic to cones also vary in contrast-sensitivity but saturate more obviously above 50%^41,42. In larval zebrafish, the contrast gain measured as the slope of the temporal contrast-response function is at a maximum around 30% in both On and Off bipolar cells⁴². To quantify the linearity of each response component, we computed a dark-light index (DLI) following ref. ¹⁰:

$$DLI=\frac{R^{Light}-R^{Dark}}{R^{Light}+R^{Dark}}$$ 2

where R^light and R^dark are the corresponding areas under the curve for the contrast response functions of transient or sustained responses (Methods). Accordingly, DLI ranged from -1 (responding exclusively to negative contrasts) to 1 (exclusively positive contrasts), while 0 indicated a balanced (linear) contrast-dependence. This revealed that both transient and sustained DLI varied greatly across cones, but while the sustained component spanned almost the entire possible coding range, including strongly positively rectified cones (e.g. cone 3), the transient components were always dark-biased (Fig. 4G). Nevertheless, across cones, the DLI of transient and sustained components covaried strongly (Pearson’s correlation coefficient; r = 0.78, p < 0.001), suggesting a single mechanism controlling rectification. Strong positive correlations were found between a number of functional metrics that varied between cones (Supplemental Fig. S3A–E). Moreover, PR1 heterogeneity was robustly observed in different eye regions, including nasal, dorsal, and the acute zone, with no significant difference in the distribution of DLI values between retinal regions (Supplemental Fig. S5A, C).

What might be the consequences of such kinetically unbalanced encoding of positive and negative contrasts? To explore this question, we used stimuli based on natural scenes and a simple model that captured these basic properties of cone release. Starting with an underwater video of translatory optic flow taken in the natural habitat of zebrafish (from ref. ⁴³) (Fig. 4H), we processed each pixel’s intensity time-series to mimic transient and sustained release components of cones (Supplemental Video 3, Fig. 4I, J, Methods): to mimic fast and dark-biased transient release we differentiated and then negatively rectified each intensity sequence while to mimic sustained and linear release the sequence was smoothed in time. This procedure demonstrated that an approximately linear sustained component would yield a time-blurred but otherwise relatively faithful representation of the scene (Fig. 4J, compare first two sets of panels). By contrast, the transient component delivered a very different parallel representation that was strongly biased to features creating negative contrast in the foreground (Fig. 4J, 3^rd set of panels – see also Discussion).

Are the variations in PR1 rectification apparent in immediately neighbouring cones? If it were, this would allow a postsynaptic bipolar cell integrating signals from cones with different sensitivities to encode a wider range of contrasts. The dendrites of bipolar cells in larval zebrafish typically span ~20 µm⁴⁴, which corresponds to ~3-4 PR1 in a row⁴⁵. To sample such a group of cones we scanned the laser in a line and compared the dark-light index values (DLi). In the example in Supplemental Fig. S6, neighbouring PR1 within a 5 µm distance exhibited DLI values ranging from −0.34 to −0.87. Moreover, we found the no systematic relationship between inter-cone distance and DLI difference, and the average difference in DLI between neighbouring cones was 0.18 (Supplemental Fig. S6E and F). Similar local heterogeneity in DLI was observed across most recordings (Supplemental Fig. S6G). We conclude that the observed cone heterogeneity is a local property of the cone array that routinely occurs at a spatial scale relevant to integration by immediate downstream circuits.

Variations in frequency-dependent output of individual cones

The decomposition of visual signals through retinal ganglion cells implementing different temporal filters has generally been thought to depend on inhibitory interactions between bipolar cells and amacrine cells in the inner retina^46–48. The results in Fig. 4 demonstrate that these different filters are implemented as early as cone synapses, differentially tuned to emphasise either transient or sustained visual signals. The striking emphasis on the transient signalling of negative contrasts rather than positive at the output of PR1 (Fig. 4E) may contribute to the observation that downstream retinal Off-circuits tend to transmit higher frequencies than on^49–51.

To characterise the temporal filters at PR1 synapses, we measured responses to a chirp stimulus in which the frequency of a full-field sinusoid (100% contrast) was swept from 20 Hz to 1 Hz over a period of 20 s. Figure 5A shows two example cones with different responses to this chirp. Cone 1 tracked the stimulus throughout the frequency range, while cone 2 responded preferentially at lower frequencies. As before, different frequency-responses were routinely observed amongst neighbouring PR1 recorded simultaneously. The frequency responses were characterised as the spectral centroid, the frequency where the centre of mass of the spectrum is located (Methods). Cones 1 and 2 had spectral centroids of 11.2 Hz and 7.5 Hz, respectively, and across 49 cones this central frequency ranged from ~6 to 11 Hz (mean 9.2 ± 1.1) (Fig. 5B).

To assess the variability in the timing of responses, we next presented a 20 Hz sinusoidal stimulus at 100% contrast for 30 s (Fig. 5C). PR1 again responded heterogeneously: while some were nearly perfectly phase-locked to the stimulus (e.g. cone 3 in Fig. 5D) others appeared to respond more stochastically (cone 4). As a measure of the temporal precision, we calculated the vector strength³⁸, a metric which varies from a value 1 for perfect synchronisation to zero for random response timing. Cones 3 and 4 had a vector strength of 0.98 and 0.41, respectively. Across a population of 49 cones from 6 fish, the vector strength at 20 Hz ranged from 0.17 to 0.98 and averaged 0.61 ± 0.25 ( ± sd) (Fig. 5E).

A key process during ongoing synaptic activity is the replenishment of the readily releasable pool (RRP) at the active zone. The ribbon structure of cones is thought to organize this refilling process to maintain vesicle release at rates required for vision³⁹. Figure 5A shows that refilling operated to allow cone 1 to maintain almost maximal responses at stimulus frequencies beyond 15 Hz while cone 2 became unreliable at frequencies above ~7 Hz and did not respond at all to stimuli beyond 12 Hz. To investigate refilling more directly, we used a paired-pulse protocol in which off-steps of 40 ms were delivered with interstimulus intervals (ISI) between 2 and 100 ms (Fig. 6A, B). A 40 ms step was used because it was just long enough to release the RRP involved in the transient phase of release (Figs. 2D, 4E). Overlaying the averaged trace for each ISI pair demonstrates that the peak release rate elicited by the second pulse was lower than that elicited by the first, but gradually recovered with increasing recovery interval (Fig. 6C).

Fig. 6. PR1 exhibit rapid but heterogeneous recovery from paired pulse depression.

A Glutamate release from an example cone in a cx55.5:SFiGluSnFR larval zebrafish during a protocol in which pairs of 40 ms off-steps were delivered, separated by an interstimulus interval (ISI) between 2 and 100 ms in a pseudo-randomised order. Each ISI condition was repeated 25 times. B Expansion of the recording showing variations in response to the second pulse at different ISIs. C Mean responses for each ISI condition overlaid for this example cone. D–F Procedure for accurately estimating the amplitude of the second response in a pair. The decay of the response to the last 40 ms pulse was fitted with a Gaussian (red line, A) to create a scalable decay template (E) for each cone. The template was scaled to the first peak in each response pair, and the peak of the second response measured relative to this baseline, as shown by blue line in (F). G The mean paired pulse ratio for each condition is plotted as a function of ISI for each cone. The average recovery time-course for the population is overlaid in red. Note that in some cones, the second stimulus generated a larger response than the first at short ISIs. H Histogram of time to 63% recovery measured across cones. Note the wide variation. Cones in which potentiation occurred at short ISIs ( < 20 ms) contribute to the initial peak. Source data are provided as a Source Data file.

The amount of refilling at a given interval was calculated as the paired pulse ratio (PPR; response 2/response 1), but it was necessary to apply a correction to the second response in a pair when the ISI was short, and the first response had not recovered fully to baseline (Fig. 6D–F). This was done by using the decay phase of the response to a single 40 ms pulse as a template that was scaled to the first response in each pair; the amplitude of the second response was then measured relative to a baseline provided by the scaled template, as shown by the blue line in Fig. 6F. The PPR as a function of ISI for 40 cones is shown in Fig. 6G with the average recovery time-course for the whole population overlaid in red. Notably, there were a few cones in which the second stimulus generated a larger response than the first at short ISIs. The time for peak release rate during the second pulse to recover to 63% of the first varied widely, as shown by the histogram in Fig. 6H. A distinct subset of cones showed strong recovery at ISIs of 10 ms or less, demonstrating that the refilling process was indeed capable of following frequencies of 20 kHz, as observed in cone 1 in Fig. 5A. The remaining cones showed much slower refilling, with 63% recovery at intervals up to 100 ms. This heterogeneity in the dynamics of refilling likely contributes to wide variations in temporal filtering and jitter across cones (Fig. 5).

Together, the experiments in Figs. 4–6 illustrate how PR1 encode visual stimuli in a highly heterogeneous manner. The simple picture of a given cone type as representing a single temporal filter driving downstream circuits does not hold.

Heterogenous output of PR1 depends on horizontal cell feedback

What is the source of the variability in the output of different cone synapses? Do these variations reflect functional differences intrinsic to each cone or differences in the feedback signals from the network in which they are embedded? We reasoned that the relative contributions from these two possible sources might be disambiguated by pharmacologically isolating cones from the rest of the retinal circuit. This was achieved by blocking the cone drive to horizontal cells using the AMPA receptor antagonist CNQX injected into the eye (Methods, estimated final concentration: 50 µM)^6,52, a manipulation expected to hyperpolarise horizontal cells and reduce their negative feedback to cones¹⁵.

A direct comparison of the output from one cone before and after inhibiting negative feedback is shown in Fig. 7A, where the stimulus is the same series of positive and negative contrast steps used in Fig. 4. With feedback intact, the sustained component of the response was approximately linear, but after injecting CNQX the output was strongly rectified to negative contrasts. A similar pattern was observed in the contrast-response functions of 26 cones in 17 fish (Fig.7B, where for simplicity the DLI is the sum of the transient and sustained release components). These results immediately suggest a possible mechanism for the varying degrees of rectification in the contrast-response functions of PR1: differences in the strength of negative feedback from horizontal cells cause the synapses of different cones to operate around different set-points, defined by the baseline rate of vesicle release at a given average luminance (see also ref. ⁵³).

Fig. 7. Cone heterogeneity requires intact outer retinal circuitry.

A Mean responses ( ± sd, shading) to contrast flashes (cf. Fig. 4A) of one example cone before (black) and after (red) intraocular injection of CNQX to block horizontal cell feedback in a cx55.5:SFiGluSnFR larval zebrafish. B This manipulation resulted in a systematic and significant drop in cone DLI values from -0.22 ± 0.27 to -0.63 ± 0.20 (p < 0.001 (0.29 × 10^-6), two-tailed Wilcoxon rank sum, n = 26 cones in 17 fish) and value distribution (Chi-squared value = 16.1, critical value = 7.8, p < 0.01 (0.0011)). C Superposition of mean contrast responses from (A) to illustrate how dynamic range and baseline (solid black lines) were extracted and D comparison of baseline across experimental conditions; control: 0.24 ± 0.11, CNQX: 0.10 ± 0.06 (p < 0.001 (0.13 × 10^-7), two-tailed Wilcoxon Rank Sum, n = 26 cones in 17 fish) and a significant decrease in heterogeneity (Chi-squared value = 8.8, critical value = 5.9, p < 0.05 (0.012)). E Strong correlations between cone DLI and baseline value across both conditions. Control (black) two-sided Pearson correlation coefficient r(24) = 0.77, p_adj < 0001 (0.12 × 10^-4); CNQX (red) r(24) = 0.89, p_adj < 0.001 (0.33 ×10^-8), both conditions combined: r(48) = 0.85, p_adj < 0.001 (0.78 × 10^-15; Bonferroni corrected p-values). F–H Glutamate release from an example PR1 driven by chirp stimulus (cf. Fig. 5A) before (top, black) and after (red, bottom) CNQX injection (F) and expansion of the same data as indicated (G). Spectral centroids before (left, black) and after (right, red) CNQX injection, with a significant decrease (H, p < 0.05 (0.039), n = 25 cones, 11 fish, two-tailed Wilcoxon Rank Sum) but no significant change in heterogeneity (Chi-squared value = 7.3, critical value = 12.6, n.s. (p = 0.30). Source data are provided as a Source Data file.

To explore this idea, we estimated each cones’ full coding range based on the minimum release rate at the highest light intensity and the maximum rate in darkness (Fig.7C). We then measured the baseline release rate as a fraction of this full range. Blocking feedback from horizontal cells lowered this set-point (Fig. 7D; p < 0.001, Wilcoxon Rank Sum). Crucially, the degree of rectification in cone output was strongly correlated with the baseline measured with or without block of negative feedback (Fig. 7E; Pearson’s correlation coefficient r = 0.77 in control and 0.89 in CNQX, both at p < 0.001). Removal of horizontal cell feedback always shifted an individual cone’s synapse towards a lower baseline and a correspondingly lower DLI as well as reducing the variability in these properties across the population (Fig. 7C, E; variance of DLI values for control = 0.07 and in CNQX = 0.04; variance of baseline values for control = 0.01 and in CNQX = 0.003). Qualitatively similar results were also observed when blocking horizontal cell feedback using HEPES instead of CNQX (Supplemental Fig. S3G–L). Blocking horizontal cell feedback also reduced preferred temporal frequencies (Fig. 7F–H; control = 6.2 ± 2.1 Hz, CNQX = 5.0 ± 1.7 Hz, Wilcoxon Rank Sum p < 0.05, n = 25 cones, 11 fish).

Together, these observations demonstrate that feedback received from horizontal cells determines the set-point of a cone synapse and, therefore, how it encodes positive versus negative contrasts. We conclude that a major source of heterogeneity in cone outputs is differences in their interactions with horizontal cells in the outer plexiform layer.

Possible benefits of cone-heterogeneity for encoding natural contrasts

Variations in the input-output relation of the first synapse in vision are expected to impact all downstream processing. To explore what those impacts might be, we set up a data-driven model of bipolar cells⁵⁴ that sum cone inputs. We drove the model with a naturalistic contrast series and compared responses to homogeneous and heterogeneous populations of cone synapses (Fig. 8). These responses were obtained by repeatedly stimulating individual PR1 with a 15 s naturalistic contrast sequence extracted from the zebrafish natural environment (Fig. 8A–C). This sequence was synthesised by combining available video data from the zebrafish natural environment⁴⁵ (Fig. 8A, B) with data on eye- and body-movements from free-swimming zebrafish⁵⁵ (Fig. 8C, D; see Methods). The results indicate that heterogeneous sampling can improve representation of naturalistic contrast series (Fig. 8).

Fig. 8. Data-driven model of heterogeneous versus homogeneous contrast encoding in natural scenes.

A–D A 180° fisheye natural scene movie recorded in India (A, from Ref. ⁴⁵) was converted into a luminance signal over time projected onto a 2D plane (B, Methods). Next, a representative azimuth-only combined eye and body trajectory was extracted from a freely swimming zebrafish larvae (C, from ref. ⁵⁵) and mapped onto the natural scene move at a fixed elevation of −10° (dashed line in B) to extract a naturalistic contrast series (D). E Responses of two example cones to the contrast sequence from (D) (mean superimposed on 5 repeats). F top, Correlations between the example cone mean response and the stimulus, computed over a 1 second duration sliding window as indicated. Grey and blue-filled areas highlight instances where the responses of cone 1 or cone 2, respectively, correlated more strongly with the stimulus. (F, middle), as top, each of the n = 12 cones recorded (homogeneous encoding), and (F, bottom), for 1000 averages of 5 randomly sampled repeats across different cones (heterogeneous encoding). G Comparison of the average performance of the homogeneous and heterogeneous populations of correlations from (F). Source data are provided as a Source Data file.

Five repeats of the contrast series were presented to a total of n = 20 PR1 from 8 fish. As before, different PR1 encoded the sequence in a highly heterogeneous manner, here illustrated by the two example cones shown in Fig. 8E. Despite this variation across different cones, the responses of both cones individually were highly reproducible, as expected from the high levels of response reliability and time precision (cf. Figs. 2, 3). To explore how these two cones differed in their encoding of a common stimulus, we next calculated each cone’s mean response across the five repeats and compared each to the original stimulus. We reasoned that the mean of five responses from a single cone is equivalent to the response of a hypothetical downstream neuron that averages individual responses from five functionally identical cones. In other words, the two mean responses mimic two different variants of a scenario in which the cone array is functionally homogeneous. To compare how accurately these two homogenous cone-scenarios represented the contrast sequence, we calculated the correlation coefficient between each mean and the stimulus over a 1 s sliding window (Fig. 8F, top traces, Methods). Neither cone’s output was systematically superior to the other; during some phases of the sequence cone 1 correlated more strongly with the stimulus than cone 2, but in other phases this was reversed. It appears difficult to define an optimal set point for an individual cone.

Next, we tested bipolar cell responses with heterogenous cone inputs. Using the same five stimulus repeats from twelve recorded cones, we compared performance with i) a population of 12 bipolar cells, each summing across the five trials recorded from the same cone (as above; homogeneous), and ii) a population of 1,000 bipolar cells, each randomly summing any five trials from the full set of 60 trials available (heterogeneous). The individual performances of all modelled bipolar cells are shown in Fig. 8F middle/bottom. We then took the mean performance of either population as an indication of their central tendency and computed their relative differences (Fig. 8G, Methods) such that a difference of 0% indicates that both strategies, on average, perform equally well, while positive or negative percentages indicate correspondingly superior performance of the heterogeneous and homogeneous strategy, respectively. According to this simple comparison metric, the heterogeneous model systematically outperformed the homogeneous model by ~8% on average, and by up to ~30% depending on the detail in the stimulus. Moreover, the estimated performance boost was positive in >97% of time epochs lasting 1 s.

This model provides a simple example of how a functionally heterogeneous array of cone synapses may improve the coding of a naturalistic contrast series in downstream neurons.

Discussion

This study demonstrates that the synapses of PR1 encode visual stimuli with exceptional reliability and time precision individually but are heterogeneous as a population in terms of sensitivity to luminance, contrast and frequency (Figs. 1–6). These functional variations are tightly linked to differences in the set-point around which the synapses of individual cones modulate glutamate release, which is in turn determined by interactions with horizontal cells in the outer retina (Fig. 7). A model of cone convergence on to bipolar cells indicates that this heterogeneity has the potential to improve the coding of naturalistic stimuli (Fig. 8). We also find that cones differentially encode different aspects of natural scenes via their transient versus sustained release components (Fig. 4). These findings reveal an unexpected degree of functional heterogeneity in the input to the retinal circuit and support the idea that photoreceptors interacting with their surrounding circuits can serve as different feature channels^19,20.

Reliability of the cone synapse

Synapses inject noise into neural circuits because of the stochasticity of the processes that control vesicle fusion^56,57 and in the retina this noise reduces the reliability of visual computations^33,58. Once information is lost it can never be regained, so it is important that the first synapse in vision minimises the noise introduced into the visual signal. Here we provide the first estimate of the reliability of the first synapse in vision in vivo and find that the variability is 10-fold less than expected for the Poisson process thought to operate at most synapses in the brain. The SNR of ~90 per synapse for a 40 ms stimulus can be put into context by comparing it with the SNR of the total excitatory synaptic current that a mammalian alpha retinal ganglion cell receives from ~500 bipolar cell inputs, which is ~50–100^59,60.

It has long been recognised that envisioning the ribbon synapses of rod photoreceptors as Poisson machines cannot easily account for the reliability with which single photons are detected and it has been suggested that a clocking mechanism might regularise the intervals between release of individual vesicles⁶¹. An alternative mechanism for making the synaptic output less variable at a given average release rate is the process of multivesicular release, where multiple vesicles are released as one synaptic event^62,63. Electrophysiology in retinal wholemounts provides evidence for a combination of these mechanisms operating in rods under voltage-clamp stimulation³⁶, although it is still not clear how they determine responses to light. The technique we have used to isolate the output from individual cones has the advantage of providing in vivo measurements of glutamate release with a temporal resolution close to 1 kHz, which will allow the synaptic coding of visual information to be examined using, for instance, natural scenes. It is now clear that ribbon synapses in rods, cones and bipolar cells do not encode visual information by modulating the rate of a Poisson process^28,36,64 but it remains to be seen which structural or functional features underly this improved reliability³⁹.

Stimulus representation across the cone array

Sensory systems deal with the statistical structure of the physical world with neurons of limited encoding capability: incoming information tends to be linear and high bit-depth, but sensory neurons are usually nonlinear and operate at lower bit-depth^65,66. Consequently, populations of sensory neurons preferentially transmit information about aspects of the stimulus that are most likely to be useful for the animal^67,68. Our in vivo measurements of the cone drive to the retinal network provide several insights into the strategies used to meet the demands of natural scenes.

Expanding the coding range of natural contrasts

Within the cone population a large degree of redundancy is expected because of spatial correlations in natural scenes⁶⁹ but we have shown that this redundancy is reduced by local variations in the input-output relation (Figs. 1–4), with the potential to expand the effective coding range of the visual system (Fig. 8). The presence of mechanistically distinct but conceptually related strategies in the early visual systems of mice¹⁰ and flies^12,70 suggests that locally heterogeneous sampling of the outside world is a general principle across convergently evolved visual systems. However, how possible improvements in overall signal representation scale with spatial scale, and in turn how this scale maps onto the operational spatial bandwidth of the system, remains important to explore in the future. Similarly, whether and how functional cone impacts more complex spatiotemporal receptive field properties of downstream retinal neurons, such as directionally selective circuits⁷¹, remains unclear.

Encoding positive and negative contrasts

Natural scenes contain both positive and negative contrasts, but non-linearities in photoreceptors, and their synapses in particular, cause them to preferentially transmit information about one over the other. One important nonlinearity in cones occurs when synaptic release sites are well-stocked^39,40 causing a sudden decrease in intensity to open calcium channels and release many vesicles from the RRP in one transient^7,72,73. However, the opposite is not the case: release does not cease in an equally transient manner if light intensity increases (Fig. 4D), and there is a strong off-bias in transient release (Fig. 4E, G) which is probably inevitable for most cones^7,10,74. Yet the sustained component of release which depends on refilling of the RRP from the reserve pool was notably more linear (Fig. 4F, G). It appears that a single cone can, in effect, multiplex two different types of information, with large and transient release events encoding a high-contrast dark-transition, while slower sustained release encodes both light and dark transitions in a more balanced manner over longer timescales. Such a compound code could be readily read out by postsynaptic circuits, for example, based on the kinetics of postsynaptic receptors^5,54.

The specific encoding of high-contrast dark events by the transient release component is an example of feature selectivity at the first synapse of vision¹⁹. In fish, this feature will tend to occur in the foreground because contrast underwater rapidly deteriorates with viewing distance^19,21,75. This dependence of contrast on distance largely disappears in air²⁰, but terrestrial species might also the transient release component use a similar strategy to disambiguate in-focus versus to out-of-focus visual structure⁷⁶. The disproportionate representation of large negative contrasts in cones may be linked to the observation that Off-circuits tend to dominate over On-circuits in detecting several elementary aspects of the visual scene, including fast temporal contrasts and spectrally broad achromatic signals^50,51,77,78. Various types of off-biases in retinal coding have been linked to statistical off-biases in some^43,79 but not all^6,9,10,80 natural scenes.

Local and global heterogeneity

The local heterogeneity observed between the light responses of neighbouring PR1 is largely driven by HC feedback, in line with previous work⁸¹. PR1 are contacted by H1-type HCs only¹⁴. Two non-mutually exclusive mechanisms might underlie how H1 differentially impacts neighbouring cones to generate heterogeneous responses to full-field light stimulation. First, cell-to-cell variability within the H1 population¹⁴ might locally propagate onto cones if each is connected to a distinct subset of H1 neurons. Alternatively, or in addition, possible synapse-by-synapse variability at each HC-cone contact might impart functional heterogeneity onto the PR1 population (see also ref. ⁵³). Beyond HCs, the persistence of a small degree of heterogeneity in the absence of HC inputs suggests that other factors also contribute. One source of variation might be the synaptic ultrastructure of cones (Supplemental Fig. S1). A second possibility is that the resting potential varies between cones because of differences in the density of conductances such as calcium, calcium-activated potassium, or HCN channels.

Previous work has shown that zebrafish PR4 (ancestral UV cones) are also functionally heterogeneous. However, unlike PR1, PR4 were globally heterogeneous but showed no obvious heterogeneity locally^6,7; These intriguing differences might relate to the very different roles served by the PR1 and PR4 systems in zebrafish vision^19–21. PR4 are approximately linear in the acute zone, but highly non-linear nasally. The finding that HC feedback facilitates PR1 response speeds and the encoding of higher frequencies also contrasts with the observation that PR4 becomes more sustained when HCs are blocked⁶. A similar scenario is observed in mice, where blocking HCs increased low-frequency signals⁸². These findings suggest that kinetic interactions between HCs and photoreceptors are species and cone-type-specific.

Spatial processing

In this study we used widefield stimuli modulated in time and intensity but not space. This approach probes the most common use of a cone, where its centre and surround are driven concurrently. Nonetheless, in the future it will be important to investigate how the centre-surround receptive field^14,52,83 affects the encoding of spatial pattens. Here, our results from cones mirror those from previous work on bipolar cells where inhibitory interactions speed up, linearise and decorrelate visual feature representation at a population level⁴⁷. Recordings of cone activity in the pharmacological absence of horizontal cell feedback (Fig. 7) approximately mimic spatially restricted stimulation of the centre and causes cones to become more non-linear and dark biased (Fig. 7B), suggesting that the same will happen in cones located near an edge. Cones in the intact network might therefore be expected to encode spatial contrast in a dark-biased manner – an effect that would be emphasised by the pre-existing dark-bias in transient release (Fig. 4G).

Methods

Experimental model

Animals

All procedures were carried out in accordance with the UK Animals (Scientific Procedures) Act 1986, under the UK Home Office guidelines, and approved by the University of Sussex Animal Welfare and Ethical Review Board. Zebrafish larvae (Danio rerio) were housed in petri dishes (maximum 50 animals per dish) at 28 °C under a standard 14:10 day/night light cycle (lights on at 8 am, lights off at 10 pm, with a 15-minute luminance ramp up or down accordingly). Animals were grown in E2 medium (Westerfield, 2000) with the addition of 200 μM 1-phenyl-2-thiourea (PTU, Sigma, P7629) from 1 day post fertilisation (dpf) to prevent further melanogenesis (Karlsson, Von Hofsten and Olsson, 2001).

The following previously published transgenic lines of zebrafish were used: Tg(Cx55:5:nlsTrpR,-tUAS:SFiGluSnFR) for expression of SFiGluSnFR in horizontal cells (Yoshimatsu et al., 2021), Tg(thrb:TdTomato) for expression of TdTomato in PR1 (Suzuki et al., 2013). Adults were outcrossed with casper for minimised pigmentation (White et al., 2008) and crystal for minimised pigmentation and melanin (Antinucci and Hindges, 2016), and offspring were raised in E2 solution with the addition of 0.1 mM 1-phenyl-2-thiourea (PTU, Sigma-Aldrich, P7629) to prevent further melanogenesis.

For all two-photon imaging experiments, zebrafish larvae were 6-7 dpf larvae. Imaging experiments were carried out in a temperature-controlled room ( ~ 20 °C) between 1 and 6 pm. Zebrafish larvae were immobilised in 2% low-melting-point agarose (Fisher Scientific, Cat: BP1360-100), positioned to lie on their side in an imaging chamber on top of a glass coverslip (thickness 0). The imaging chamber was filled with E2 fish medium, submerging the agarose-fixed larvae. Eye movements were further prevented by injection of α-bungarotoxin (1 nL of 2 mg/ml; Tocris, Cat: 2133) into the ocular muscles behind the eye.

Two-photon glutamate imaging

A Scientifica two-photon imaging system was used, equipped with a mode-locked Ti:Sapphire laser (Chameleon II, Coherent) tuned to 915 nm for excitation of the SFiGluSnFR reporter. Two-photon excitation was delivered through the objective (20X water-immersion, XLUMPlanFL, numerical aperture 0.95, Olympus). Emission of the SFiGluSnFR signal was captured above and below the sample, through the objective and condenser (Oil-immersion, numerical aperture 1.4, Olympus). The emitted signal was filtered through GFP filters (HQ 525/50, Chroma Technology) and detected by above- and sub-stage GaAsP photomultiplier tubes (PMTs, H7422P-40, Hamamatsu). The signal from the above- and sub-stage PMTs passed through a current-to-voltage converter before being summed with a custom-built summing amplifier and digitised. Image acquisition was controlled through ScanImage for Windows (3.8, MatLab). Functional recordings of the SFiGluSnFR signal were taken as 1 kHz line scans (128 X 1 pixels per frame, 1 ms per line). The dendritic tips of horizontal cells expressing SFiGluSnFR invaginate the photoreceptor pedicle (alongside bipolar cells (not labelled)) and act as an antenna for photoreceptor glutamate release. Line scans were positioned over the terminals of 3-5 cones per scan. Laser activation of photoreceptors was minimised by focusing the line scan across the pedicle, thus avoiding the photosensitive outer segment.

For consistency, and unless stated otherwise, recordings were taken from the nasal retina, approximately aligned with the outward-facing visual horizon (discussed in ref. ⁴⁵). Moreover, unlike functional measurements, biosensor expression levels were judged to be relatively homogeneous, based on immunolabelling against GFP in a subset of experiments (not shown). Any observed functional heterogeneity is therefore unlikely to be related to possible heterogeneities in biosensor expression and associated buffering effects⁸⁴.

Light stimulation

Full-field visual stimulation was delivered from an amber light emitting diode (LED, 590 nm) through a lightguide (both ThorLabs) positioned approximately 1 cm from the zebrafish retina. The light was filtered through a 590/10 nm bandpass filter (Thorlabs) to maximally activate PR1, whilst minimally activating PR2, and to fall outside of the activation range of PR3 and PR4. At full power, the output of the LED at the sample plane was ~660 µW, sufficient to almost suppress glutamate release completely for tens of seconds. Stimulation protocols were driven through IGOR pro 6.3 for Mac (Wavemetrics), and the microscope was synchronised to visual stimulation.

Pre-processing

Movies were imported as tiff files in IGOR Pro 8.0 for Windows (Wavemetrics) and analysed using a suite of analysis routines; SARFIA⁸⁵, GlueSniffer Analysis Package²⁸ (James et al., 2019), and custom written procedures.

The GlueSniffer Analysis Package was used to analyse line scans and isolate SFiGluSnFR signals to individual cone terminals. Briefly, the spatial profile of a line scan was measured by making a temporal average of the fluorescence signal along the line scan. The average fluorescence profile was fit with a sum of Gaussians, where each Gaussian component corresponds to a ROI. For each ROI, the change in fluorescence over time was extracted, weighted towards the pixels in the centre of the spatial profile, facilitating significant denoising. The relative change in fluorescence, or DF/F, was calculated using the most frequent value (that is, the baseline of the recording). Any drift in the resulting DF/F trace was baseline corrected where required, using a linear correction. Recordings with strong photobleaching or low signal:noise were discarded. Additionally, recordings were discarded where an initial 100% contrast, 2 s off-step evoked a response with an amplitude of less than 3 standard deviations of the baseline.

The possibility of conflating SFiGluSnFR signals from multiple cone terminals was limited experimentally by taking line scans at spatially separated ROIs, avoiding planes of focus where multiple ROIs were crowded together. Second, the small point-spread function (PSF) of the microscope reduced conflation of signals. The microscope used for these experiments has a PSF of 0.7 μm in x and y, meaning the synapses of cones >1 μm apart could be easily distinguished. The PSF in z was 2.2. μm. The average cone terminal width was approximately 2-3 μm (Supplemental Fig. S1). Moreover, during imaging the centre of the cone terminals of interest were approximately identified by the strongest SFiGluSnFR signal, and line scans were taken at this imaging plane to minimise the conflation of signal from terminals above and below the plane of focus was minimised.

Next, DF/F traces of the SFiGluSnFR signal were deconvolved to limit the decay element of the SFiGluSnFR fluorescent reporter from the recordings and produce traces that are proportional to the rate of glutamate release, or DF/F per second (F’s^-1), using a Wiener filter²⁸. The decay of the SFiGluSnFR reporter was estimated by fitting transient responses with a kernel. Transients at most cone terminals could be described with a kernel with a decay of 0.06 s. By filtering the decay of the fluorescent reporter from the signal, it was possible to recover an estimate of the glutamate signal.

Dark-light index

The dark-light index (DLI) was calculated from the cone’s contrast-response function; the area under the curve (AUC) for the negative contrast flashes was measured, giving value b, and the AUC for the positive contrast flashes was measured, giving value a. To calculate an index value, both numbers must be positive; therefore, the value was multiplied by -1. Following this, if the 'a' value was negative (i.e., because the contrast response curve sat above zero for the positive contrast flashes due to a noisy signal), the 'a' value was assumed to be zero. The dark-light index was calculated as (a-b)/(a + b), producing a DLI value between -1 and 1. Where the DLI = -1, the cone exclusively responded to negative contrast flashes, where the DLI = 0 the cone responded to both positive and negative contrast flashes equally, and where DLI = 1 the cone responded exclusively to positive contrast flashes.

Spectral centroid

Glutamate responses to the 20 Hz chirp stimulus were analysed to identify the spectral centroid, that is, the centre of mass of the frequencies in the glutamate responses, following:

$$Spectralcentroid=\frac{{\sum }_{k=1}^{N}kF\left[k\right]}{{\sum }_{k=1}^{N}F\left[k\right]}$$ 3

Briefly, the Fourier transform (FT) of the response trace was weighted by the power of the frequencies. The sum of this weighted FT was divided by the sum of the FT to produce the spectral centroid.

Vector strength

Glutamate responses to the 20 Hz sinusoid stimulus were analysed to measure the phase locking of responses over repeats of the stimulus phase. Responses were detected by differentiating the glutamate records and using the built-in peak-finding function in Igor Pro 8.0 for Windows, PeakFinder, to detect the rise of the response transient. The mean + 1 sd of the glutamate signal during exposure to mean-light levels (between 6 and 8 s of stimulus) was applied as the threshold for detecting a response. Response times were converted to a phase angle, ${\mathrm{\theta }}_i$, i.e., where they occur within a stimulus period. Briefly, where one stimulus phase is equal to 360 degrees, the time of the rise of each response was converted to the degree at which it occurred in the stimulus phase, and then to radians. The vector strength was calculated, as described elsewhere⁸⁶, as follows:

$$Vectorstrength=\sqrt{\frac{{\sum }_{j=1}^N\sin \left({\mathrm{\theta }}_i\right)+{\sum }_{j=1}^N\cos \left({\mathrm{\theta }}_i\right)}{N}}$$ 4

The vector strength for each cone recording is the square root of the sum of the sine and cosine of the vector values for each response. The vector strength may vary from 0 to 1, with 1 implying perfect synchronisation of responses with the stimulus.

Statistics

Statistical analysis was carried out using in-built functions of IGOR Pro 8.0 for Windows (Wavemetrics). When data were not normally distributed, nonparametric tests were applied. Tests were two-sided and significance was defined as p < 0.05. Where required, post-hoc tests were used to correct for multiple groups. If an experiment required the delivery of more than one visual stimulus protocol, the order of the stimuli was randomised. Data collection and analyses were not carried out blind as PR1 responses were specifically selected for during experiments and analysis. Recordings were excluded from analysis if a low signal-to-noise ratio of the deconvolved SFiGluSnFR signal impeded accurate detection of light responses to off-steps of light.

Pharmacology

For some experiments, HCs' light responses were blocked with an injection of cyanquixaline (CNQX, Tocris, Cat: 1045, final concentration of approximately 50µM) in artificial cerebrospinal fluid (aCSF). CNQX is a synthesised non-NMDA (N-Methyl-D-Aspartate) receptor antagonist that blocks AMPA- and kainate-receptors, effectively blocking HC light responses. CNQX was injected intravitreally into the eye, through the cornea adjacent to the lens (usually temporal to the lens).

Beyond blocking cone input into HCs, CNQX additionally affects downstream retinal circuitry, which in turn could modulate the neuromodulator milieu of cones, for example via dopaminergic interplexiform cells (DA-IPCs)^87,88. However, this is unlikely to cause major off-target effects on PR1 and associated synaptic pathways, because (I) DA-IPCs are thought to be primarily driven by the brain and (ii) previous work points to the rod system, rather than cones, as the major target for dopaminergic modulation in the zebrafish retina⁸⁹.

Immunohistochemistry and confocal imaging

The transgenic line expressing TdTomato in PR1 (Tg(thrb:TdTomato)) was outcrossed with the crystal line, and the resulting larvae were screened for TdTomato expression and a crystal phenotype (clear, non-pigmented eyes). The positive larvae were euthanised by tricaine methanesulfonate (MS222, Sigma Aldrich) overdose and fixed in 4% paraformaldehyde (PFA, Agar Scientific, AGR1026) in 1 X PBS on a rocker for 30 minutes exactly at room temperature. After three washes in 1 X PBS, whole eyes were enucleated, and the cornea was removed by using the tip of a 30 G needle. The dissected eyes were incubated in 0.1% Triton X-100 (Sigma, X100) made up in 1 x PBS for 15 minutes at room temperature. In fresh 0.1% Triton X-100, primary antibody was added; anti-1D4 (Santa Cruz, mouse, sc-57432) at 1:20. Samples were incubated at 4 °C for 4 days. Samples were washed five times in 0.1% TritonX-100 in 1X PBS and treated with DAPI nuclear dye (Invitrogen, 33342) at 1:2000, and secondary antibody; Donkey-anti-mouse Dylight 647 (ThermoFisher, A32787) at 1:200. After one day incubation at 4 °C, samples were washed in 0.1% Triton X-100 in 1X PBS twice. Samples were mounted in 1.5% agarose in PBS on a coverslip. The eyes were manually oriented with the lens facing down, and the PBS was replaced with mounting media (Vectorlabs, Vectashield, H-1000) for imaging.

Confocal image stacks were taken on an LSM880 (Zeiss) using a 40X oil immersion objective (plan apochromat 40X/1.3 oil DIC UV-IR M27), or a 63x oil immersion objective (HC PL APO CS2, Leica). Contrast, brightness, and pseudo-colour were adjusted for display in Fiji (NIH). Quantification of soma length, outer segment length, and terminal width were performed using custom scripts in IGOR Pro 8.0 (Wavemetrics) after manually marking the inner and outer locations of each organelle of interest.

Electron microscopy and analysis

Serial blockface scanning electron microscopy (SEM) volumes were gratefully received from Prof Rachel Wong (University of Washington, USA), as previously part-analysed and published in ref. ¹⁴. The data consist of 652 vertical slices through the OPL of the Acute Zone region of the retina of a 6 dpf wild-type larval zebrafish. There is 50 nm between each layer, and the xy pixel dimension is 5 nm.

Using the TrakEM2 plugin in FIJI, the data set was manually aligned, and structures of interest (e.g., photoreceptor outer segment, mitochondria, nucleus, terminal, and HC and BC nuclei and dendritic projections) were manually traced. From here, 3D reconstructions of the traced structures were generated, and basic measurements e.g. volumes and surface areas, were calculated using in-built functions of FIJI.

Modelling bipolar cell responses

A representative stimulus segment was extracted by combining a previously published⁴⁵ video-segment (60 Hz) showing an underwater scene of a typical zebrafish natural habitat in Northern India with a typical azimuth-only eye and body movement trajectory measured in a free swimming zebrafish⁵⁵ as shown in Fig. 8A–D. To this end, the video was converted to 8-bit greyscale by averaging the RGB components, followed by reading out of the single pixel brightness values over time (1 pixel per frame) as dictated by the eye-body movement trajectory. For simplicity, we did not include eye movements in elevation and kept the entire sequence at -10 degrees relative to the visual horizon, approximately aligned with the position sampled by the nasal PR1 during physiology. We presented five repeats of this stimulus sequence as a 60 Hz widefield stimulus to a total of 20 cones and extracted their glutamate responses. Twelve out of these 20 cones passed a quality criterion of 0.4 (see definitions in ref. ⁹⁰) and were used for the presented analysis.

Data Exclusions, reproducibility, sample sizes

Glutamate recordings were excluded from analysis if a low signal to noise ratio of the deconvolved SFiGluSnFR signal impeded accurate detection of light responses to off-steps of light. In addition, where applicable, a threshold of mean + 1 sd of the glutamate signal during exposure to bright light was applied, and responses not surpassing this threshold were discarded to reduce the false detection of noise as a response. Reproducibility of the experimental findings was verified by repeating experiments across different batches of larval zebrafish and observing similar effects; by repeating experiments on a different two-photon microscope and observing similar effects; and by collaborating with researchers to verify data collection, data extraction, and data analysis methods. All attempts at replication were successful. Experiments were repeated across at least 5 separate imaging sessions. No statistical method was used to predetermine sample sizes.

Reporting summary

Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article.

Author contributions

Conception and Design: T.H., L.L., and T.B. Supervision: L.L. and T.B. with input from J.M.D. Fundings: T.H., L.L., and T.B. Materials: T.H. and T.Y. Data Collection: T.H. with initial help from J.M.D. Data processing: T.H., with inputs from B.J., L.L., and T.B. Modelling: TB. Writing: T.H., L.L., and T.B.

Peer review

Peer review information

Nature Communications thanks Wallace Thoreson and the other, anonymous, reviewer(s) for their contribution to the peer review of this work. A peer review file is available.

Data availability

The processed (extracted and deconvolved) PR1 response data generated in this study are available in the GitHub repository [https://github.com/tessaherzog/A-heterogeneous-population-code-at-the-first-synapse-of-vision/tree/main/Data]. The raw two-photon data are available on request to the corresponding authors Tessa Herzog and Tom Baden (see contact details in the author list), due to their large size. Source data are available in the GitHub repository [https://github.com/tessaherzog/A-heterogeneous-population-code-at-the-first-synapse-of-vision/tree/main/SourceData]. Source Data are provided with this paper.

Code availability

Custom code for the analysis of glutamate traces is available on GitHub [https://github.com/tessaherzog/A-heterogeneous-population-code-at-the-first-synapse-of-vision/tree/main/IGORcode]” (Ref. ⁹¹).

Competing interests

The authors declare no competing interests.

Supplementary information

The online version contains supplementary material available at 10.1038/s41467-026-68757-x.