A circuit suppressing retinal drive to the optokinetic system during fast image motion

Optokinetic nystagmus (OKN) assists stabilization of the retinal image during head rotation. OKN is driven by ON direction selective retinal ganglion cells (ON DSGCs), which encode both the direction and speed of global retinal slip. The synaptic circuits responsible for the direction selectivity of ON DSGCs are well understood, but those sculpting their slow-speed preference remain enigmatic. Here, we probe this mechanism in mouse retina through patch clamp recordings, functional imaging, genetic manipulation, and electron microscopic reconstructions. We confirm earlier evidence that feedforward glycinergic inhibition is the main suppressor of ON DSGC responses to fast motion, and reveal the source for this inhibition—the VGluT3 amacrine cell, a dual neurotransmitter, excitatory/inhibitory interneuron. Together, our results identify a role for VGluT3 cells in limiting the speed range of OKN. More broadly, they suggest VGluT3 cells shape the response of many retinal cell types to fast motion, suppressing it in some while enhancing it in others.

We mined a publically available serial electron microscopic (SBEM) dataset of the adult mouse inner plexiform layer spanning >200 µm (volume 'k0725' 6 ). We reconstructed large numbers of small-field, highly branched amacrine cells known or inferred to be glycinergic 7,8 . We searched for synaptic contacts from such cells onto ON DSGCs, and focused our analysis mainly on a single previously reconstructed presumed ON DSGC (Fig. 2, main text). Numerous anatomical features confirm the identity of this cell, as an ON DSGC, as originally reported 9 (Fig. 2). Its dendrites narrowly costratify with the processes of starburst amacrine cells (SACs), with nearly the entire arbor within the ON SAC plexus (Fig. 2b-e). There, they cofasciculate with SAC processes (Fig. 2a) and receive numerous wrap-around synaptic contacts 10 from SAC varicosities (Fig. 2f, g). The presynaptic SAC processes exhibit the asymmetric connectivity that confers direction selectivity upon DSGCs [9][10][11][12] . The minor OFF branches, a common feature of ON DSGC arbors in mice 1 , receive some OFF SAC contacts. Presumptive ON-OFF DSGCs in the volume shared many of these attributes but had smaller dendritic arbors, higher branching density, and a larger fraction of their arbor in the OFF SAC plexus.
To identify possible sources of glycinergic amacrine cell inhibition to this ON DSGC, we first marked all the non-ribbon synaptic contacts we could find (n=431; Fig. 2f-i).
For each contact, we reconstructed enough of the presynaptic process to sort it into one of three groups (Fig. 2a- comprising 10% of non-ribbon inputs, with wide-field inputs nearly as common at 7% of all input (n=27). A small minority of inputs (n=23; 5%) could not be identified because too small a fraction of the arbor could be reconstructed, and were excluded from the sample.
We extensively reconstructed all of the presumptive glycinergic neurons identified by our initial screen. To our surprise, nearly all of them appeared to be VGluT3 amacrine cells, as documented below (n=38 synapses from 16 cells). Two were synapses from Type H18 amacrine cells 13,14 and one was from an unidentified medium-field type. We detected no synaptic inputs from any of the many other types of small-field amacrine cell types, even after examining a large sample of such cells reconstructed in another study 14 . We conclude that VGluT3 amacrine cells are by far the best candidate glycinergic cell type shaping the slow-speed tuning of ON DSGCs.
Multiple structural observations confirm these as VGluT3 amacrine cells ( Supplementary Fig. 3). First, their stratification is appropriate ( Supplementary Fig.   3e, f, j). They arborize most heavily in the middle of the inner plexiform layer (IPL), between the ON and OFF SAC plexuses, but extend sparse processes into and even a bit beyond those plexuses. Their cell bodies lie in the inner nuclear layer, are of appropriate size, and are distributed with spacing that suggests a regular mosaic ( Fig. 3a). Their dendritic-field diameters (106 ± 7 μm, n = 10 cells) are in line with earlier data 15,16 and overlap extensively with a coverage factor of at least 5   28,29 types, which stratify largely outside the VGluT3 plexus. VGluT3 cells also synapsed upon diverse types of amacrine cells, though only rarely onto other VGluT3 cells or SACs. We have not reconstructed most amacrine-cell targets, though WF cells stratifying between the SAC plexuses were among the most common recipients 30 . VGluT3 outputs, like their ribbon inputs, were distributed most heavily within and between the SAC plexuses ( Supplementary Fig. 3h), though they were also sparsely present in the distal IPL. Some of these are somato-dendritic synapses onto M1 ipRGCs. One unexpected feature of the VGluT3 cells in this volume was a consistent upward (roughly dorsal ) displacement of the dendritic arbor relative to the soma (Supplementary Fig. 3b-d).  In order to manipulate VGluT3 cells in a specific manner, we studied Cre-expression in the mouse line VGluT3-IRES2-cre-D that to our knowledge had not been previously used in retinal studies. To this end we crossed the VGluT3-Cre mouse with a tdTomato reporter mouse (Ai14). In retinas of these mice, nearly every VGluT3-immunopositive neuron expressed tdTomato (98.2 ± 0.6%; 9 images, 2 retinas; Supplementary Fig. 4a). VGluT3-cell dendrites formed a dense plexus concentrated between the ON and OFF SAC plexuses, as revealed by anti-ChAT immunofluorescence ( Supplementary Fig. 4e). The only other retinal cells brightly labeled with tdTomato were Müller glia, though we occasionally encountered labeled RGCs (~20/mm 2 ) and wide-field amacrine cells with somas in the GCL or INL and straight, sparsely branches processes in the IPL (~2 processes in a 200x200 µm field of view).
To gain optogenetic and chemogenetic access to VGluT3 cells, we crossed the VGluT3-Cre mouse with lines expressing channelrhodopsin (Ai32), or the hM4Di receptor (DREADD), respectively. To verify expression of the reporters in VGluT3 cells in the crossed mice, we used the YFP expressed along with ChR2 in the VGluT3 xAi32 mouse, and an antibody against HA tag was used in VGluT3 xR26 (DREADD) ( Supplementary Fig. 4b, c). VGluT3 were labeled independently using anti-VGluT3 antibodies, and SACs were labeled using anti-ChAT. In VGluT3-Ai32 mice, YFP expression was usually seen during recordings using 2-photon or single-photon illumination.
To obtain information on the morphology of the Cre expressing cells in the VGluT3-Cre line, we injected adeno-associated viruses (AAV) expressing GFP or YFP in a Cre dependent manner into the eyes of VGluT3-Cre mice. This often resulted in labeling that was sparse enough for discerning the dendrites of individual cells ( Supplementary Fig. 4d). The morphology of the labeled cells matched that of VGluT3 ACs ( 15,16,31 and SBEM data in this study). In addition, the GFP virus resulted in more non-VGluT3 cells expressing GFP than the crossed mouse line VGluT3 xAi14, possibly due to leakiness of the promoter carried by the AAV.

ON DSGC inhibition in response to optogenetic pulse trains.
While recording inhibitory currents in ON DSGCs, we delivered optogenetic pulse trains to mimic the temporal modulation of VGluT3 cells during grating motion ( Supplementary Fig. 5c). Evoked inhibitory currents persisted in postsynaptic ON DSGCs throughout a 5 s train over the full range of frequencies tested (1-8 Hz, corresponding to grating speeds of roughly 400 -3000 µm/s for experiments shown in Fig. 1). At higher stimulus frequencies, IPSCs summed temporally to produce a continuous inhibitory conductance ( Supplementary Fig. 5c, bottom traces), mirroring the conductances evoked during rapid grating motion (Fig. 1b,

Controls for optogenetic studies.
Optogenetically induced currents typically exhibited rundown, with smaller responses after repeated stimulation. In a pharmacological experiment, such rundown could masquerade as a drug effect. To assess the extent of the rundown effect, we conducted sham experiments (Supplementary Fig. 6a). Using the standard photoreceptor block but no subsequent drug application, we measured the optogenetically evoked currents at intervals similar to those in an actual pharmacological experiment. Over the time required to complete a real drug trial with two steps of drug application, the peak current dropped by 38 ± 2% (2 cells), but was not eliminated. We conclude that while rundown does contribute to the decreases of the current between steps in these experiments, the pharmacological effects we report are real.
In VGluT3 x Ai32 mice, we observed the expression of the YFP reporter in Müller glia ( Supplementary Fig. 4b). It was therefore important to demonstrate that optogenetic depolarization of Müller glia does not contribute the induced currents in ON DSGCs, especially in light of recent evidence for electrical coupling between Müller glia and amacrine cells 32 . To this end, we crossed a Müller-specific Cre mouse (GLAST-creER) 33 with the Ai32 mouse. In these animals, ON DSGCs exhibited neither excitatory nor inhibitory currents during photostimulation of the Müller cells ( Supplementary Fig. 6b).
The currents induced optogenetically in VGluT3 x Ai32 mice are due to the ChR2 transgene, and not to other light-dependent mechanisms. In control recordings of ON DSGCs in HoxD10 mice without viral or genetic expression of the channelrhodopsin, the optogenetic stimulus evoked no currents ( Supplementary   Fig. 6c). This implies that the pharmacological cocktail used in optogenetic experiments to block rod and cone influences on ON DSGCs was largely effective.
Rarely, an optogenetic stimulus appeared to trigger large current bursts; these may have resulted from incomplete blocking of rod/cone networks in some cells. Those were mostly inconsistent over trials and had a longer latency compared to persistent optogenetic responses, and were excluded from the data. A persistent ACET-resistant current was recorded in a minority of ON DSGCs, appearing as an OFF response upon termination of the LED stimulation ( Supplementary Fig. 6c).
Incomplete blockade of light responses tended to be more pronounced in the presence of strychnine.

Origin of fast-motion inhibition in ON DSGCs not blocked by DREADD.
Strychnine reduced inhibition in ON-DSGCs across speeds more than the chemogenetic suppression of VGluT3 cells did (Fig. 1d, Fig. 4d, e, Supplementary Fig.   7a, b). The chemogenetic manipulation therefore only partially suppressed glycine release onto ON DSGCs.
Overall, DREADD-mediated suppression of VGluT3 cells eliminated only about half of the inhibition induced in ON DSGCs by fast motion (Fig 5a, b). The remaining inhibitory current might be supplied by a different type of amacrine cell, but it could also stem from incomplete suppression of VGluT3 cells. Application of the DREADD ligand CNO scaled down the inhibitory currents while preserving the kinetics ( Supplementary Fig. 8a), and the scaling factor was nearly constant across different grating speeds (Supplementary Fig. 8b). This is most easily explained by incomplete suppression of VGluT3 cells. If a different amacrine type is responsible, it must exhibit temporal kinetics and ON vs. OFF sensitivity closely matched to that of VGluT3 cells.

Spots and gratings area-response functions in VGluT3 dendrites.
We compared the receptive field size and surround suppression, when measured with gratings vs. spots (Fig. 5g, h, Supplementary Fig. 10a, b). The surround stimulation using spots was a simple step in luminance whereas the grating stimulus introduced local contrast as well as motion stimuli without a change in integrated luminance. We thus tested new sets of spots and gratings, with overall contrast in the two stimuli equal and positive ( Supplementary Fig. 10c, d, see Methods). The area-response functions for spots resembled those seen previously, although the lower contrast weakened the responses. The surround suppression for masked gratings, however, was stronger than for zero mean gratings. Although the optimal size of the stimulus and the suppression at the maximal size were no longer different between spots and gratings, the shape of the response functions were still markedly different, with the response stronger for gratings at larger sizes ( Supplementary Fig. 10d). In summary, the weaker surround suppression in the case of gratings vs. spots, is both due both to a smaller change in contrast as well as to the continuous motion.