Functional, molecular and morphological heterogeneity of superficial interneurons in the larval zebrafish tectum

The superficial interneurons, SINs, of the zebrafish tectum, have been implicated in a range of visual functions, including size discrimination, directional selectivity, and looming‐evoked escape. This raises the question if SIN subpopulations, despite their morphological similarities and shared anatomical position in the retinotectal processing stream, carry out diverse, task‐specific functions in visual processing, or if they have simple tuning properties in common. Here we have further characterized the SINs through functional imaging, electrophysiological recordings, and neurotransmitter typing in two transgenic lines, the widely used Gal4s1156t and the recently reported LCRRH2‐RH2‐2:GFP. We found that about a third of the SINs strongly responded to changes in whole‐field light levels, with a strong preference for OFF over ON stimuli. Interestingly, individual SINs were selectively tuned to a diverse range of narrow luminance decrements. Overall responses to whole‐field luminance steps did not vary with the position of the SIN cell body along the depth of the tectal neuropil or with the orientation of its neurites. We ruled out the possibility that intrinsic photosensitivity of Gal4s1156t+ SINs contribute to the measured visual responses. We found that, while most SINs express GABAergic markers, a substantial minority express an excitatory neuronal marker, the vesicular glutamate transporter, expanding the possible roles of SIN function in the tectal circuitry. In conclusion, SINs represent a molecularly, morphologically, and functionally heterogeneous class of interneurons, with subpopulations that detect a range of specific visual features, to which we have now added narrow luminance decrements.

the SINs through functional imaging, electrophysiological recordings, and neurotransmitter typing in two transgenic lines, the widely used Gal4s1156t and the recently reported LCRRH2-RH2-2:GFP. We found that about a third of the SINs strongly responded to changes in whole-field light levels, with a strong preference for OFF over ON stimuli. Interestingly, individual SINs were selectively tuned to a diverse range of narrow luminance decrements. Overall responses to whole-field luminance steps did not vary with the position of the SIN cell body along the depth of the tectal neuropil or with the orientation of its neurites. We ruled out the possibility that intrinsic photosensitivity of Gal4s1156t+ SINs contribute to the measured visual responses. We found that, while most SINs express GABAergic markers, a substantial minority express an excitatory neuronal marker, the vesicular glutamate transporter, expanding the possible roles of SIN function in the tectal circuitry. In conclusion, SINs represent a molecularly, morphologically, and functionally heterogeneous class of interneurons, with subpopulations that detect a range of specific visual features, to which we have now added narrow luminance decrements.

K E Y W O R D S
cell type, interneurons, luminance detection, optic tectum, zebrafish

| INTRODUCTION
Ensuring behavioral reliability across variable environmental conditions requires our brains to extract relevant features from the visual landscape, a process which is refined over multiple levels of visual processing. Due to their robust visually mediated behavioral repertoire and their amenability to genetic and imaging tools, the larval zebrafish is a highly tractable model for interrogating the neural circuits underlying visual behavior Orger et al., 2008;Walker et al., 2013;Thiele et al., 2014;Bianco & Engert, 2015; [Correction added on 13 Jan 2021, after first online publication: ProjektDeal funding statement has been added.] Portugues & Engert, 2009;Kubo et al., 2014;Mearns et al., 2020).
In the zebrafish brain, the optic tectum is the main center for visual processing, and significant strides in mapping the circuitry of the tectum have been achieved with the identification of tectal cell types (DeMarco et al., 2020;Förster et al., 2020;Förster et al., 2018;Gabriel et al., 2012;Helmbrecht et al., 2018;Kramer et al., 2019;Robles et al., 2011;)and with the publication of a comprehensive single-cell atlas of the larval zebrafish brain (Kunst et al., 2019). Retinal ganglion cells (RGCs) send axonal projections into the tectum, where they each arborize in one of nine sublayers, the stratum opticum (SO), six laminae of the stratum fibrosum et griseum superficiale (SFGS1 through 6), the stratum griseum centrale (SGC), and the boundary region between stratum album centrale and the stratum periventriculare (SAC/SPV) (Robles et al., 2013(Robles et al., , 2014Xiao et al., 2011). In the tectum, the majority of cell bodies reside in the periventricular layer (SPV), a dense collection of periventricular projection neurons and interneurons (PVNs) below the neuropil (Nevin et al., 2010).
A third, much less abundant class of tectal neurons, named the superficial interneurons (SINs), is unique in that their cell bodies lie in the most superficial layers of the tectal neuropil, SO and SFGS1 (Del Bene et al., 2010). SINs have generally been considered to be GABAergic. Due to their spatial segregation, these cells offer a unique foothold for dissecting visual processing in the optic tectum. The sparseness of cell bodies in the neuropil has allowed for targeted electrophysiology, functional imaging, and ablation.
In the last decade, SINs have been implicated in several tectumdependent behaviors, including prey capture and looming-mediated escape (Abbas et al., 2017;Barker & Baier, 2015;Del Bene et al., 2010;Dunn et al., 2016;Hunter et al., 2013;Preuss et al., 2014;Yin et al., 2019). This broad involvement suggests that SINs either occupy a global, "general-purpose" inhibitory role in visual processing, or that functional subsets with distinct feature selectivities and specific neural connectivities exist. In support of the latter scenario, subpopulations of SINs differentially tuned to size, direction, orientation, and looming speeds have been observed (Abbas et al., 2017;Dunn et al., 2016;Förster et al., 2020;Hunter et al., 2013;Preuss et al., 2014;Yin et al., 2019). Here we contribute to this growing literature by using functional imaging and electrophysiology to investigate the tuning of SINs to whole-field luminance changes. We find that subsets of SINs encode narrow luminance steps, mostly in the OFF direction. We also characterize a new transgenic line, Tg(LCR RH2 -RH2-2:GFP) pt115-c (called LCR RH2 -RH2-2:GFP here; Fang et al., 2013) with a broader SIN labeling than the commonly used Gal4s1156t line, and investigate overlap with other molecular markers. Unexpectedly, we discovered that a substantial portion of SINs are glutamatergic. The rich diversity in feature selectivity, morphology, marker expression, and transmitter use argue in favor of the existence of multiple SIN types, each serving in parallel in feature-selective microcircuits, similar to the amacrine cells of the retina.

| Enucleation experiments
Bilateral eye removal was performed 1-2 days prior to imaging experiments. Larvae were anesthetized with 0.02% tricaine and embedded in 2% low melting point agarose (Invitrogen, Carlsbad, CA, USA). Eyes were removed surgically with a sterile 25G needle and #55 forceps (Fine Science Tools). Following the enucleation procedure larvae were removed from agarose and allowed to recover overnight. The overall health of the larvae was assessed before proceeding to imaging experiments.

| Immunostaining
Larvae (8 dpf, Gal4s1156t × UAS:EGFP) were fixed in 4% PFA in PBS overnight at 4 C followed by a 12 h incubation in 30% sucrose in PBS. Sections were washed three times with PBS (15 min per wash) and mounted in Fluoromount (Sigma-Aldrich) prior to imaging.

| Statistics
All statistical tests were performed using GraphPad Prism 6.0. When appropriate, unpaired t-tests or one-way and two-way ANOVA tests with Tukey's correction for multiple comparisons were performed. In all figures * denotes p < .05, ** denotes p < .005, and *** denotes p < .0005, n.s. denotes "not statistically significant." All error bars are standard error of the mean (SEM).

| Confocal imaging
Confocal images were acquired with Zeiss LSM700 or LSM780 microscopes with associated ZEN software. A minimum of four larvae were imaged for each condition.

| Colocalization analysis
Colocalization analysis was performed using the Coloc2 custom plug-in for ImageJ (Schneider et al., 2012). In brief, corresponding z stacks for green and red fluorescence were split. ROIs for each GFP + cell (SINs) were defined and colocalization analysis runs for both channels within the ROI in the sub-stack where the GFP+ cell body was present. A Pearson's r value coefficient for colocalization of pixel intensity within each ROI was generated. Pearson coefficients ≥0.2 were counted as colocalized in crosses of transgenic lines, and ≥0.5 for immunostainings.

| Calcium imaging and analysis
Calcium imaging was performed using either a custom-built moveable objective 2-photon microscope (MOM, Sutter Instruments, Novato, CA, USA) with ScanImage software for image acquisition (Pologruto et al., 2003) or a customized commercial 2P microscope (Femtonics 3DRC, Femtonics, Hungary;Dal Maschio et al., 2017) with associated software for image acquisition. Both microscopes used 20× objectives (Olympus, NA 1.0) and 920 nm excitation for GCaMP6s. Prior to imaging, larvae were embedded in 2.5% low melting point agarose (Invitrogen). All larvae were raised on a normal light/dark cycle and kept in a dark room prior to and during imaging (0.2 lux). Imaging F I G U R E 1 Many Gal4s1156t+ SINs are strongly tuned to changes in whole-field luminance. (a) Example SIN from a 7dpf Gal4s1156t × UAS: GCaMP6s larva shows consistent responses (shown as ΔF/F) to whole-field changes in luminance. (b) Responses to whole-field luminance are highly consistent across larvae. Responses from five SINs from four Gal4s1156t × UAS:GCaMP6s larvae. (c) Gal4s1156t+ SINs are more responsive to darkening (decreasing luminance) whole-field flashes than brightening (increasing luminance) whole-field flashes (unpaired t-test, p < .0001). (d, e) Electrophysiological recordings show increased excitatory synaptic input in response to transitions between light ON and light OFF conditions. (d) Four example traces from neurons targeted in Gal4s1156t × UAS:Kaede. EPSCs occur with greater frequency during low luminance (light OFF) conditions. (e) Total excitatory postsynaptic current is plotted for OFF and ON light conditions. For (c), n = 47 cells from 17 larvae. *** denotes p < .0005; Scale bar in (d), 20 pA, 200 ms. Error bars are SEM. Visual stimulus is shown above the corresponding example traces in (a), (b), and (d) acquisition was performed between 2.94 and 3.37 Hz (frames per second, 256 × 256 pixel dimensions). For analysis of whole-field flash and looming stimuli calcium imaging data, raw fluorescence traces were analyzed using the Time Series Analyzer V3.0 plug-in for Fiji (Pologruto et al., 2003) as previously described (Barker & Baier, 2015). In brief, ROIs were manually defined for all cell bodies and ΔF/F values calculated as follows:  Stimuli for full-screen flash and looming experiments were generated with a custom programmed graphical interface for Vision Egg (Straw, 2008). The bright loom stimulus consisted of a white expanding disc on a gray background, and the dark loom stimulus consisted of a black expanding disc on a gray background. Looming stimuli and whole-field flash stimuli were presented binocularly.
Whole-field flash stimuli were presented for 0.5 or 1.0 s and maintained for 4 or 10 s before the next luminance change (a whole-field flash) was initiated.
F I G U R E 3 A subset of Gal4s1156t+ SINs display narrow tuning to incremental luminance changes. (a) Gal4s1156t+ SINs were tested in response to a luminance step stimulus. Four progressive whole-field luminance steps from black (2.7 lux) to white (92.8 lux) were presented in both directions (darkening and brightening). Here responses from an example SIN (SIN 8 in (c)) in an 8 dpf Gal4s1156t × UAS:GCaMP6s larva are shown. Normalized luminance steps are indicated in gray, raw calcium responses are shown in black. Luminance steps of 0.5, 1.0, 1.5, and 2.0 correspond to lux changes of 8.2, 27.1, 64.5, and 90.1 lux respectively. (b) Luminance step tuning curves in 4 Gal4s1156t+ SINs demonstrate peak tuning distributed across the range of luminance steps tested (n = 14 cells, 8 larvae). The tuning curve for the SIN shown in (a) is plotted here in black. For this cell, the strongest calcium responses are observed at larger luminance steps, peaking with a luminance step of 64.5 lux and decreasing slightly at the largest step size tested (90.1 lux). (c) Luminance step responses are shown for all 14 SINs recorded in 8 larvae. Heat map shows normalized ΔF/F values for each cell. Luminance steps of 0.5, 1.0, 1.5, and 2.0 correspond to lux changes of 8.2, 27.1, 64.5, and 90.1 lux, respectively. Responses to OFF and ON steps were pooled. (d,e) Gal4s1156t+ SINs respond to bright and dark looming stimuli with similar response amplitudes and dynamics. Responses to dark (top) and bright loom (bottom) are shown for the same cell. The visual stimulus is displayed above each trace. (f) Gal4s1156t+ SINs do not show preferential tuning to bright or dark looming stimuli when compared to equivalent whole-field flashes (n = 11 cells in 5 larvae, one-way ANOVA, Tukey's correction, p = .787). n.s., not statistically significant. Error bars are SEM 2.9 | Luminance step stimuli Luminance steps were generated using custom scripts for PsychoPy (Peirce, 2008). Lux values for luminance steps were: 2.7, 10.9, 29.8, 67.2, and 92.8 lux. Each luminance step was displayed for 10 s. For analysis a custom written python script was used to select single cell ROIs, their ΔF/F values calculated and baseline changes in the traces normalized using asymmetric least square smoothing (ALS, Eilers & Boelens, 2005). For each luminance step the baseline was calculated as a mean of 2-7 frames before stimulus onset and the response as the maximum value in the first 5 frames after the luminance change (baseline subtracted). The stimulus was presented twice, and an average value from the two repetitions was calculated.

| Receptive field mapping
Receptive fields (RFs) were mapped using a checkerboard stimulus with 4 × 3 square grids of 20 or 10 × 6 square grids of 9 of the larva's visual field (total area 78 × 60 ) generated using custom scripts in PsychoPy (Peirce, 2008). Analysis was performed using a custom written python script, where single cell ROIs were selected and ΔF/F values with corrected baseline calculated. The RF was calculated using the maximum response during the presentation of each grid square and averaged across two trials. In order to calculate the size of the RF, the calculated array was interpolated by a factor of 100 for a smoothed fit, the array was thresholded to the half-maximum, and the contour diameter calculated. Ten SINs were exclusively tested with RF mapping, 6 cells with both RF mapping and luminance stimuli, and 8 cells with luminance stimuli only.

| Electrophysiology
Whole-cell patch clamp recordings were performed on 6-10 dpf  (Branchek, 1984;Branchek & Bremiller, 1984). We performed imaging experiments at the same time every day, in the early afternoon of the larva's subjective day, to account for the well-documented circadian oscillations in light sensitivity (Moore & Whitmore, 2014).
We found that any whole-field luminance step tested was sufficient to drive responses in 30% of Gal41156t+ SINs (Figure 1a,b), and this response profile was highly stereotyped across larvae ( Figure 1, n = 47 SINs out of 146 labeled cells in 17 larvae). We tested several flash durations (0.5, 1, 4, and 10 s), as shown in Figure 1a,b, and found that cells that showed luminance responses, responded consistently to all luminance transitions. However, we observed differences in the amplitude of responses when the stimulus was darkening (decreasing overall luminance) or brightening (increasing overall luminance). Significantly greater responses were observed to overall luminance decrements (Figure 1c, unpaired t-test, p < .0001), suggesting that the majority of luminance responsive Gal4s1156t+ SINs are OFF responsive cells.
3.2 | Gal4s1156t+ SINs receive increased excitatory synaptic input in response to decreases in whole-field luminance To corroborate the calcium imaging data with electrophysiology, we carried out voltage clamp recordings from SINs, targeting fluorescently labeled SINs in the Gal4s1156t × UAS:Kaede line. Across seven cells in seven individual larvae, we observed repeated and sustained membrane depolarizations following an offset of the light stimulation.
The total current from these excitatory postsynaptic currents was greater for whole-field decreases in luminance (OFF flashes) than for whole-field luminance increases (ON flashes), except for one cell where the ON and OFF currents were approximately equal (Figure 1d 3.4 | Gal4s1156t+ SIN subpopulations respond selectively to discrete OFF transitions, but not to looming We next asked how sensitively tuned the Gal4s1156t+ SINs were to discrete luminance steps. The full set of luminance values extended from 2.7 to 92.8 lux (measured at the surface of the screen). Interestingly, we observed Gal4s1156t+ SINs with selective responses to small, intermediate, and large luminance steps (Figure 3a-c).
We next tested whether Gal4s1156t+ SINs were tuned to a looming stimulus, which also produces an overall luminance change, but in a defined spatiotemporal pattern. We presented both white, expanding discs on a gray background (ON looming) and black, expanding discs on a gray background (OFF looming; Figure 3d,e). Expansion velocity was matched to that previously reported to elicit maximal escape probability (constant radial expansion of 20 /s) (Temizer et al., 2015). We found no differences in responses to bright or dark looming stimuli when compared to full-field flashes (Figure 3d

| SIN RFs sample over a large visual area
In order to assess the spatial area over which SINs are responsive, we mapped their RFs using a classic mapping strategy (e.g., Smear et al., 2007). We again used double-transgenic fish carrying Gal4s1156t and UAS:GCaMP6s. A coarse grid, composed of an array of 12 or 60 squares, each filling about 20 or 9 of the larva's visual field was presented to one eye of an embedded fish larva positioned on the stage of a two-photon microscope (Figure 4). We found that all SINs had large RFs (23 or greater, Figure 4h). Two SINs tested with a 60-square grid (with each square covering 9 of the larva's visual field) showed RFs similar in size and position to those mapped more coarsely (Figure 4a

| Gal4s1156t+ SIN responses to whole-field luminance changes require the retina
One possibility that had not been tested previously is that some SINs may possess endogenous light detection capacity. Expression of opsins has been reported in the brains of many teleost species (Cavallari et al., 2011;Fischer et al., 2013;Whitmore et al., 2000). We directly tested this possibility by surgical removal of both eyes. In the absence of retinal input, Gal4s1156t+ SINs did not show robust endogenous responses to luminance changes ( Figure 5). In only 3 of 12 SINs from enucleated larvae (n = 3 larvae) were any calcium responses observed ( Figure 5). These signals likely originated from spontaneous discharges, as they were not consistent across trials and were not synchronized to the visual stimulus. Responses to flashes in control larvae, with eyes present, often occurred within 300 ms (1-2 frames of imaging acquisition) of stimulus offset, whereas, in enucleated larvae, responses F I G U R E 6 SINs are strongly and stably labeled by the LCR RH2 -RH2-2:GFP transgene. (a) Top view of a 7dpf isl2b:Gal4 × UAS:mCherry × LCR RH2 -RH2-2:GFP larva. GFP expression is observed in putative SINs as well as in the pineal gland and a subset of photoreceptor cones. (b) Rotated confocal volume of the same larva shown in (a). Note LCR RH2 -RH2-2:GFP+ SIN cell bodies (green) line are anatomically restricted to SO (RGC axons labeled by mCherry, magenta). (c, d) Fluorescence intensity profile measurements along the yellow lines shown in (c). mCherry signal peaks correspond to SFGS, SGC, and SAC/SPV. (e) GFP expression, analyzed across multiple larvae crossed to either isl2b:Gal4 or atoh7:Gal4 x UAS-mCherry, shows that LCR RH2 -RH2-2:GFP + SINs are restricted to SO (n = 20 larvae for isl2b:Gal4, 18 larvae for atoh7:Gal4, 7 dpf). (f) Label of LCR RH2 -RH2-2:GFP+ SINs is stable over time. Number of GFP+ cells shown in 5 larvae over for each developmental age, respectively. (g-j) A single LCR RH2 -RH2-2:GFP larva at 5 dpf (g), at 6 dpf (h), at 7 dpf (i), at 8 dpf (j). Scale bars in (a, j) = 50 μm. A, anterior; L, lateral; M, medial; P, posterior. Error bars are SEM occurred, if at all, >3 s (>10 frames of imaging acquisition) after the stimulus presentation (Figure 5d). Together, these results suggest that Gal4s1156t+ SINs are not intrinsically photosensitive.
3.7 | The LCR RH2 -RH2-2:GFP transgene labels more SINs than Gal4s1156t The Gal4s1156t line has two significant limitations: it does not label all SINs, and its labeling is not restricted exclusively to the SINs; sparse labeling throughout the brain is also observed in non-SIN cells.
Thus, we sought to identify new lines with a broader label of the SIN population and less exogenous expression in non-SIN populations. To this end, we examined LCR RH2 -RH2-2:GFP, in which a regulatory element of the zebrafish RH2-2 green opsin drives GFP expression (Fang et al., 2013). This opsin promoter line had been previously reported to label only a subset of tectal neurons, along with restricted expression in green cone photoreceptors and the pineal gland (Fang et al., 2013). Intriguingly, the labeled tectal neurons were similar in morphology and position to the SINs (Figure 6a,b) with a more ubiquitous expression pattern than that observed in the Gal4s1156t line.
As an initial anatomical characterization, we crossed LCR RH2 -RH2-2:GFP transgenic fish to fish carrying both isl2b:Gal4 and UAS: mCherry. This allowed us to label RGC axons and the retinorecipient laminae in the tectum. We confirmed that the majority of GFP+ cells

| DISCUSSION
The SINs were first described in larval zebrafish by Del Bene et al. (2010), who reported that most of them were tuned to detect large (>20 ) or whole-field visual stimuli, as opposed to PVNs or RGCs, which responded as a population to stimuli of all sizes. These authors also noted that SINs expressed GABA and gad2b and concluded that they were inhibitory. Ablation studies suggested that inhibition from the SINs contributed to small-size selectivity of PVNs and thus helped tune responses to small prey items by means of a subtractive filter (Del Bene et al., 2010). In the past decade, a substantial body of work has been published, making the SINs one of the best-characterized cell classes in the zebrafish brain. Here we confirm and extend the emerging view that SINs are functionally, morphologically, and molecularly more diverse than initially considered.

| Functional heterogeneity of SINs
Previous studies demonstrated responses of SIN subpopulations to bars, dots, or gratings of varying sizes and moving in specific directions, or to looming disks (Table 1)

responses of many thousand individual tectal cells, including both
SINs and PVNs, to moving dots and gratings, looming stimuli, and sudden or gradual ON and OFF changes were recorded. One conclusion from this survey was that 85% of SINs responded selectively to some version of a dark stimulus on a bright background, that is, an OFF stimulus. The present study, using GCaMP6s imaging and wholecell patch-clamp electrophysiology, corroborates this preference for OFF stimuli by showing that Gal4s1156t+ SINs show transient responses to sudden decrements in ambient luminance. Interestingly, individual SINs were found to be selectively tuned to distinct luminance steps across all light levels tested. This observation adds to the functional diversity reported previously (Table 1). In the future, it will be important to systematically classify the SIN population by their combined feature selectivity, analogous to the seminal work done for mouse RGCs (Baden et al., 2016). Other authors have also expanded the original SIN definition. Avitan et al. (2017)

distinguished between superficial and deep SINs,
showing that the superficial population (SINs sensu stricto) showed a lower frequency of spontaneous discharges during development than the deep population (Avitan et al., 2017). Förster et al. (2020) described seven morphotypes of neurons with cell bodies in the deeper layers of the neuropil, often with bistratified or tristratified arbors. These so-called NINs, which correspond roughly to SINs with upward-oriented dendrites described here, showed graded differences in their population tuning compared to the SINs sensu stricto, although both cell classes were preferentially OFF-selective. In the present study, responses to whole-field luminance flashes did not vary measurably across neuropil layers, although the sample size may have been too small to detect a systematic difference. Interestingly, the new LCR RH2 -RH2-2:GFP line labels exclusively SINs sensu stricto and is excluded. Taken together, the available evidence points to at least ten morphological classes of SINs/NINs in the larval zebrafish tectum, three of which are classical SINs (Förster et al., 2020). and Gal4s1156t) or the incomplete labeling of glutamatergic and GABAergic populations within the vglut2a:loxP-DsRed-loxP-GFP and gad1b:loxP-DsRed-loxP-GFP lines. However, stochastic labeling cannot fully account for the molecular diversity of SINs observed. We found that none of the markers labels all SINs, and none of the markers appears to completely overlap with any of the other markers, suggesting that SINs are molecularly more heterogeneous than previously anticipated.

| Retinal origin of visual responses
Extraretinal opsins are found in the zebrafish brain (Fischer et al., 2013), and deep brain photoreceptors have been shown to influence light-mediated behavior (Fernandes et al., 2012;Fischer et al., 2013;Fontinha et al., 2020). Moreover, the specific SIN marker LCR RH2 -RH2-2:GFP was generated by linking GFP to the green cone opsin enhancer. We tested the possibility of intrinsic light responses by enucleation. In the absence of retinal input, Gal4s1156t+ SINs did not respond to whole-field ON or OFF flashes, suggesting they are not intrinsically photosensitive, at least under the experimental conditions employed here.
This result raises the question of how tuning to luminance changes emerges in the SIN population. We propose that a direct retinal synaptic input from OFF-selective RGCs is likely, although this remains hypothetical. Electrophysiological recordings can give us a clue-we see delays between visual stimulus presentation and onset of activity on the order of 100 ms. In other species, monosynaptic delays in the millisecond range have been reported between retina and tectum, and between retina and thalamus (Matsumoto & Bando, 1980;Usrey et al., 1998). However, delays between visual stimulus onset and OFF RGC spiking can be 100 ms in mouse retina (Gollisch & Meister, 2008), similar to the response latencies observed in SINs. One possibility for testing a direct RGC-SIN synapse would require electrical stimulation of RGC axons combined with electrophysiological recordings of SINs, but would likely prove difficult due to the small size of the optic tract and the short distance to the tectum in the zebrafish larva. Better still, electron microscopy reconstruction of the larval zebrafish, which has been performed for other circuits (Helmstaedter et al., 2013;Wanner et al., 2016), should provide a definitive answer regarding the presynaptic connectivity of the SINs.

| Are SINs more like the horizontal or the amacrine cells of the retina?
Based on their positions close to the retinal input layers and their monostratified morphologies, SINs have been likened to tectal horizontal cells in the older literature (Langer & Lund, 1974;Luksch & Golz, 2003;Meek & Schellart, 1978). Retinal horizontal cells allow for adaptation to overall luminance levels by acting directly on photoreceptors via feedback inhibition and on bipolar cells via feedforward inhibition (Perlman et al., 1995). In the mouse retina, targeted ablation of horizontal cells results in reductions of contrast sensitivity across all spatial frequencies tested in an optomotor assay and deficits in overall visual acuity (Sonntag et al., 2012). Similar gain control functions may be important for all levels of visual processing, and the SINs may play such a general-purpose role in the tectum, for instance, allowing it to operate under different luminance conditions.
Judging by the accumulated evidence over the past decade, however, we favor a more nuanced view of SIN function. Their tremendous molecular, functional, transmitter, and morphological diversity suggests that they act as highly specialized filters for the visual feature channels that reach the nine neuropil layers of the tectum (for a review of the functional anatomy, see Baier, 2013). A comparison to cortex or retina may be instructive, where diverse interneuron populations have been shown to operate in specialized microcircuits (Hangya et al., 2014;Jadzinsky & Baccus, 2013;Masland, 2012). In this scenario, different SIN types may be dedicated to the parallel processing of distinct visual features, more similar to the amacrine cells than the horizontal cells of the retina.