Mapping Visual Functions onto Molecular Cell Types in the Mouse Superior Colliculus

Summary

The superficial superior colliculus (sSC) carries out diverse roles in visual processing and behaviors, but how these functions are delegated among collicular neurons remains unclear. Here, using single-cell transcriptomics, we identified 28 neuron subtypes and subtype-enriched marker genes from tens of thousands of adult mouse sSC neurons. We then asked whether the sSC’s molecular subtypes are tuned to different visual stimuli. Specifically, we imaged calcium dynamics in single sSC neurons in vivo during visual stimulation and then mapped marker gene transcripts onto the same neurons ex vivo. Our results identify a molecular subtype of inhibitory neuron accounting for ~50% of the sSC’s direction selective cells, suggesting a genetic logic for the functional organization of the sSC. In addition, our studies provide a comprehensive molecular atlas of sSC neuron subtypes and a multimodal mapping method that will facilitate investigation of their respective function, connectivity, and development.

eTOC

Liu, Savier, and colleagues perform single-nucleus transcriptomics to generate a comprehensive molecular atlas of superficial superior colliculus (sSC) cell types. The authors develop a multi-modal method to unmask the molecular identity of each functionally defined sSC cell and with it reveal an inhibitory molecular subtype, Cbln4+ neurons, as direction selective.

Graphical Abstract

Introduction

Parsing the brain into cell types and neuronal subtypes has been a fundamental goal of neuroscience since the time of Ramón y Cajal. Early studies defined neurons based on morphology. With the development and application of physiological methods, neurons were then grouped into subtypes based on function. More recently, the advent of single-cell genomics has allowed classification of neuronal subtypes by their genome-wide expression profiles. These studies have generated rich information regarding neuronal diversity and identified genetic access points to specific neuronal populations for mechanistic investigations¹. However, whereas the connection between morphology and function has been firmly established, the link between functional and molecular properties is only beginning to be unraveled. This is a critical gap in understanding neuronal subtypes and their respective roles in neural computation.

Here we set out to map visual functions to molecular cell types in the superior colliculus (SC), a midbrain structure that is conserved in all vertebrates (optic tectum in nonmammalian species)^2,3. The SC is a major sensorimotor hub that integrates multimodal sensory inputs to drive reflexive behaviors⁴. Recent studies have also demonstrated that the SC plays an important role in higher cognitive functions such as visual spatial attention and decision-making^3,5,6. The superficial layers of the SC (sSC), including the stratum griseum superficiale (SGS) and stratum opticum (SO), are visual layers², which receive direct inputs from the retina and visual cortex and send outputs to multiple subcortical structures^3,7–10. The mouse SC has recently become a popular and productive model given its importance in mouse visual behaviors and the available modern neuroscience tools in this species^11–13.

In mice, more than 85% of retinal ganglion cells project to the sSC¹⁴, making it the most prominent visual center. Diverse visual responses have been reported in the mouse sSC, including compact receptive fields with overlapping On/Off regions, selectivity for stimulus direction and/or orientation, motion and speed selectivity, responses to expanding dark disks (i.e., looming), dependence on stimulus size, and suppressive responses^15–20. In particular, direction selectivity is a universal feature that has been observed in the sSC of all mammalian species studied to date¹¹. It has been demonstrated in mice that direction selectivity in the sSC is inherited from the retina²¹, and interestingly, declines with depth^18,22,23, consistent with lamina-specific projections of subtypes of retinal ganglion cells^24–27. To relate the diverse responses to neuronal subtypes, Gale and Murphy identified a few transgenic mouse lines in which Cre recombinase was expressed by populations of neurons in the sSC¹⁶. Cre-positive neurons in each of these lines appear to be morphologically distinct and have different (but overlapping) response properties and projection patterns. However, it is not clear whether each Cre-defined cell population corresponds to one or more molecular subtypes of sSC neurons. Whether molecular subtypes of SC neurons mediate different visual functions such as direction selectivity remains an open question.

In the mouse SC, single-cell genomics have been used to unravel the diverse molecular cell types^28–30, but these studies did not focus on the visual layers or relate molecular cell types to visual functions. Here, we develop a multi-modal method to link sSC neuron subtypes based on their molecular and functional features. Our results reveal molecularly distinct subtypes of neurons that differ in their responsiveness to visual stimuli and, importantly, identify one inhibitory subtype that accounts for nearly half of the sSC’s direction selective neurons. The molecular identification of this neuron subtype will facilitate future studies into its functional significance, circuitry, and development. Finally, our comprehensive cell-type atlas and multi-modal mapping method can be widely applied to reveal the molecular identities of other functional cell types in the SC and beyond.

Results

Molecular Identification of 28 neuron subtypes in the mouse sSC

We classified cell types of the adult mouse sSC by profiling their genome-wide mRNA expression with single-nuclei RNA-sequencing (snRNA-seq) (Figure 1A). After filtering out low quality cells, 41,130 cells remained, which were integrated across 3 biological replicates consisting of cells derived from different groups of mice (10 mice in total, see Methods for details of quality control and batch comparison). We then clustered these cells based on expression of highly variable genes³¹. Thirteen molecularly distinct cell clusters were identified (Figure S1A) and annotated based on their respective expression of classical cell type marker genes: neurons (Syn1+, Rbfox3+); astrocytes (Agt+); oligodendrocytes (Mbp+); oligodendrocyte progenitor cells (Pdgfra+); microglia (P2ry12+); endothelial cells (Cldn5+) and leptomeningeal cells (Lum+, Figure S1B). When hierarchically clustered according to transcriptomic similarity, neurons and non-neuronal cells separated into two clades, as expected (Figure S1B). When cells were visualized based on whether they came from male and female mice, no qualitative difference was observed in terms of cell type identities or proportions (Figure S1C).

Figure 1. Molecular identification of superficial superior colliculus (sSC) neuron subtypes.

A. Schematic of single-nuclei RNA-sequencing (snRNA-seq) workflow.

B. UMAP (uniform manifold approximation and projection) plot of 26,186 sSC neurons clustered by expression of highly variable genes and colored based on cluster identity. Each dot represents a neuron.

C. Dendrogram illustrating relatedness of 28 sSC neuron populations, followed by the cluster name, annotated based on selected marker gene(s); number of cells per cluster; number of genes detected per cluster; and plot showing average expression and percent detection of the marker gene selectively enriched for each cluster.

See also Figure S1 and S2.

To further characterize potential subtypes within the neuronal population, we re-clustered the 26,186 neurons into 28 clusters (Figure S1E) and annotated them based on their expression of specific marker genes or gene combinations (Figure 1B, C). Again, the composition or proportionality of neuronal subtypes did not differ grossly between males and females (Figure S1D). Importantly, the sample size of >26k neurons reached the saturation level of number of clusters (Figure 1A.5). Interestingly, several neuron clusters expressed genes that have been associated with cortical interneuron subtypes. For instance, one cluster (n11) specifically expressed the gene encoding vasoactive intestinal peptide (Vip) and two other neuron clusters (n12 and n13) were associated with the somatostatin gene (Sst) (Figure 1C). However, parvalbumin (Pvalb), which was previously used to dissect the sSC circuitry and linked to behavior ^32,33, was sparsely distributed across all 28 clusters, indicating that Pvalb expression does not define a specific molecular subtype in the sSC (Figure S1H). Finally, we compared our dataset to recent studies of mouse SC cell types that were based on morphology¹⁶, expression of a priori candidate molecules³⁴, or single cell transcriptomics^28–30. Notably, in many cases, single “subtypes” in previous studies split into multiple clusters when mapped to our dataset, suggesting that our analysis provided better resolution of sSC neuronal subtypes (Figure S2 and see Discussion). Our results thus reveal a greater diversity of molecularly defined neuron subtypes in the sSC (i.e., visual layers) than previously reported.

Neurotransmitter properties of sSC neuron subtypes

Next, we investigated the transmitter properties of the 28 candidate neuron subtypes based on their expression of genes for glutamatergic transmission (vGluT1/Slc17a7, vGluT2/Slc17a6 and vGluT3/Slc17a8) and GABAergic transmission (vGAT/Slc32a1, Gad1 and Gad2). We found that 10 clusters expressed vGluT2 highly but little to none of vGAT, Gad1 or Gad2, whereas the opposite was true for the other 18 clusters, indicating clear separation between excitatory and inhibitory neurons of the sSC (Figure 2A–B). These data also reveal that the sSC contains a greater diversity of inhibitory neurons than excitatory neurons.

Figure 2. Neurotransmitter properties of sSC neuron subtypes.

A. Expression of excitatory markers (vGluT1, vGluT2, and vGluT3 in magenta) and inhibitory markers (vGAT, Gad1 and Gad2 in cyan) for each cluster.

B. UMAP plot color coded based on vGluT2 and vGAT co-expression (gray, low co-expression; magenta/cyan, high co-expression).

C. Pie chart showing the percentage of excitatory and inhibitory neurons in the sSC snRNA-seq dataset.

D. Fluorescence in situ hybridization (FISH) of vGluT2 and vGAT expression in the SC. An 800 μm × 250 μm yellow rectangle is used to quantify expression and further divided into 16 sub-regions along the SC depth. Staining within the area were shown to the right: DAPI, gray; vGluT2, magenta; vGAT, cyan. Scale bars, 200 μm (left) and 100 μm (right).

E. Bar plot showing the percentage of vGluT2+ and vGAT+ neurons within the quantified area as illustrated in D. Each dot is from one image. Mean ± SEM. N = 16 images from 3 mice.

F. Plot of excitatory (magenta) and inhibitory (cyan) neuron density along SC depth.

G. Bar plot of excitatory (magenta) and inhibitory (cyan) percentage along depth.

H. FISH of Vip, vGluT2, and vGAT transcripts in SC and cortex. The areas in the white squares are shown at higher magnification to the right. Scale bar, 200 μm (left) and 50 μm (right).

I. Percentage of Vip co-localization with vGluT2 and vGAT. Mean ± SEM. N = 3 animals.

See also Figure S3.

Inhibitory neurons not only had more clusters, they also comprised the majority of neurons in the snRNA-seq dataset (59.9% of all neurons, 15,678/26,186; compared to 40.1% for excitatory neurons, 10,508/26,186) (Figure 2C). This percentage is higher than in previous reports, which ranged from ~30% in the mouse SGS³⁵ to ~50% in cat sSC³⁶. To validate our observation, we performed RNA fluorescence in situ hybridization (FISH) of vGluT2 and vGAT transcripts in the SC (Figure 2D). We found that 58.1 ± 1.5% of SC neurons were vGAT+ and 41.9 ± 1.5% vGluT2+ (Figure 2E), largely consistent with our snRNA-seq analysis. The density and percentage of inhibitory neurons (vGAT+) were even greater in the superficial layers, and then decreased with depth in the SC (Figure 2F, G). The density of vGluT2+ neurons was relatively constant along the dorsoventral axis of the SC, but their relative number increased due to the decrease of inhibitory neurons density.

Surprisingly, our snRNA-seq analysis detected expression of vGluT2 in the n11.Vip neuron cluster, suggesting that Vip+ SC neurons may be excitatory. This is contrary to what has been observed in the cortex, where Vip expression marks inhibitory interneurons³⁷, To confirm this, we used FISH to co-localize Vip transcripts with vGluT2 and vGAT in the SC and cortex (Figure 2H). The vast majority of Vip+ SC neurons expressed vGluT2 (94.2% vGluT2+) but not vGAT (1.6% vGAT+). This is in striking contrast with the neocortex, where nearly all Vip+ cortical neurons expressed vGAT (96.89%) and not vGluT2 (Figure 2I). Most of Sst+ sSC neurons expressed vGAT, similar to Sst+ cortical neurons (Figure S3 and ³⁷). Together, these results suggest that collicular and cortical neuron subtypes, even when expressing the same marker genes, may differ in fundamental features such as the expressed neurotransmitter.

Spatial organization of sSC neuron subtypes

To determine the distribution of the newly identified SC neuron subtypes, we visualized marker transcripts for a few selected clusters in the sSC using RNA FISH. The expression of many marker genes was stratified by sublayer. For example, the marker transcripts of clusters n10.Col18a1, n11.Vip, and n27.Cbln4 were only located in the uSGS, whereas n02.Npnt cluster appeared to be only in the SO sublayer (Figure 3A–B, S4A), consistent with the in situ hybridization data from Allen Mouse Brain Atlas (https://mouse.brain-map.org/)³⁸ (Figure S4C). In contrast, other clusters spanned multiple sublayers. For instance, the expression of Chrnb3 (marker of n25.Chrnb3) was distributed across the entire SGS (Figure 3C), whereas Sst was expressed in both lSGS and SO sublayers (Figure 3D). Because Sst was selectively expressed by two neuronal clusters, we assessed additional markers for these clusters (Stk33 for n12 and Dcn for n13) from Allen Brain Atlas and localized n12.Stk33/Sst neurons to lSGS but n13.Dcn/Sst neurons to the SO sublayer (Figure 3D, S4E–F). Similarly, n27.Cbln4 and n28.Epb41l4a shared a common marker gene (Cdh6) that covered the entire SGS, but expression of their unique markers was restricted to individual sublayers (n27 in uSGS and n28 in lSGS) (Figure S4E and S4G).

Figure 3. Layer-specific distribution of neuron subtypes in the sSC.

A - D. FISH showing spatial expression of marker genes that are restricted to a sublayer (A, B) or spanning multiple sublayers (C, D). Inset: UMAP plot where cells are colored based on their expression of marker genes. Corresponding clusters are highlighted with the dashed line. The laminar organization of SC is indicated below each image according to Allen Mouse Brain reference atlas. DpG: deep gray layer; InWh: intermediate white layer; InG: intermediate gray layer; SO: stratum opticum; and SGS: stratum griseum superficiale. Scale bar, 200 μm.

E. Summary table of spatial distribution of sSC neuron subtypes. Individual subtypes are restricted to upper SGS (uSGS), lower SGS (lSGS), SO, or span entire SGS (uSGS and lSGS) or entire SC.

See also Figure S4 and S5.

We further assessed the layer distribution of other clusters using data from Allen Mouse Brain Atlas. Altogether (Figure 3E and S4D), 17 of the 28 neuron clusters were localized to a specific sublayer: n10/11/27 in the uSGS, n09/12/14/17/19/24/26/28 in the lSGS, and n02/05/07/13/20/22 in the SO. Eight others resided in multiple sublayers: n01/03/04/15/16/18/25 in the entire SGS and n23 in the entire sSC. Importantly, multiple markers for the same clusters often showed the same laminar-specific pattern, supporting our assignment of neuronal clusters to specific sublayers. For instance, Vip, Nek11, and Sntb1 were markers for the n11.Vip cluster and their expression in the sSC were all restricted to uSGS (Figure S4H–I). Similarly, markers for n02.Npnt neurons, Npnt, Gpr149, and Scml2, were only expressed in the SO, and markers for the n25.Chrnb3 neurons, Chrnb3, Gabrr2, Chrna6, were found throughout SGS (Figure S4I). Together, our results reveal molecular differences between sSC neuron subtypes, which raise hypotheses about their anatomy and function. The molecular differences between sSC neuronal subtypes can be leveraged to test those hypotheses and further investigate each subtype.

Linking functional and molecular cell types

Whether molecularly defined neuron subtypes in the SC have specific visual response properties is an important but unaddressed question. One standard approach is to use marker genes through Cre/lox genetics to access genetically defined neuron subtypes and characterize their anatomy, physiology, and function. We searched for available mouse lines in which Cre recombinase is driven by marker genes identified in our snRNA-seq and RNA FISH analyses (http://www.informatics.jax.org/home/recombinase) (Figure 3E). We obtained several of these Cre driver mouse lines and examined their Cre activity in the SC. Although this approach has been successfully applied to other brain regions, here we could not establish a clear correlation between function and Cre expression driven by the candidate marker genes we examined. We ran into two major issues with this “forward” approach: a lack of correlation between the driver gene and Cre expression (e.g., in the case of Vip-Cre, Figure S5A–B), or a lack of functional specificity for the molecular cell type investigated (e.g., in the case of Sst-Cre, Figure S5C–F).

To overcome these hurdles, we developed a reverse approach: from defined functional properties to molecular identity. This was done by performing in vivo imaging of single-neuron calcium dynamics followed by RNA FISH for marker transcripts selected based on their specificity and spatial distribution. We performed landmark-based registration of images acquired in vivo and ex vivo, using sparse transgenic or viral expression of tdTomato as the landmarks (“anchor points”; see Methods for details; Figure 4A.2 and 4B.2). In vivo two-photon calcium imaging of sSC visual responses followed our previously published procedure^39,40. Briefly, a cranial window was implanted above the SC and neuronal activity was assessed by imaging the changes in fluorescence of a genetically encoded calcium indicator (AAV1.hSyn.GCaMP6f). Mice were head-fixed and live images of the sSC were acquired while a patch of sinusoidal drifting grating was presented to the mouse (Figure 4A.1). Individual neurons were manually selected offline (Figure 4A.3.), and their visual responses were analyzed, including direction selectivity. After acquiring functional responses, mice were lightly anesthetized and their sSC volumetrically imaged (Figure 4A.2.). Animals were then perfused, and the SC was horizontally sectioned and processed for RNA FISH (Figure 4B.1.). Up to four probes against the marker genes of individual clusters were consecutively stained following RNAscope HiPlex procedure. Confocal images from each round were merged after acquisition (Figure 4B.3.). Aided by the anchor points, the same neurons were identified in both in vivo and ex vivo images (Figure 4C), revealing the molecular identity of neuron subtypes that display specific visual response properties.

Figure 4. Mapping visual responses to neurons expressing identified molecular markers.

A. Measurement of visual responses in the sSC. 1. Imaging of the sSC in awake head-fixed mice using two-photon calcium imaging 2. Left panel, average imaging plane of in vivo functional characterization; right panel, sparse labelling by a red fluorescent protein. 3. Selection of regions of interest (ROIs) for the quantification of visual responses.

B. FISH of selected marker genes identified by snRNA-seq. 1. Horizontal sectioning and multiplexed FISH following in vivo imaging. 2. Acquisition of individual round of FISH for each marker gene. 3. Identification of ROIs expressing marker genes.

C. Registration of in vivo and ex vivo photomicrographs. Selected ROIs in the functional imaging are overlayed with the histological results, allowing the mapping of visual response properties onto molecularly defined cell types based on probe staining.

D. Changes in GCaMP6f fluorescence intensity obtained by 2-photon calcium imaging. Shaded areas represent the duration of stimulus presentation, with arrows representing stimulus direction.

E. Overlayed response time courses for all registered neurons, grouped by the corresponding marker genes for which they are positive. The response was at the preferred direction of each neuron, i.e., the direction that elicited the highest change from baseline.

F. Corresponding tuning curves, representing the average response for each direction aligned to the neuron’s preferred direction (vertical bar).

G. Average peak response magnitude to the preferred direction of each registered neuron as a function of marker gene identity. For non-DS neurons, the direction that caused the largest response was selected and the corresponding magnitude was plotted. For suppressed neurons, the minimum was used as the “peak response” in this plot. Mean ± SEM.

H. Distribution of gDSI of all registered neurons as a function of marker gene identity. Neurons are classified as responsive if their strongest responses were above 20% ΔF/F0 (shaded in light and dark green) and classified as direction selective (DS) if their gDSI was above 0.2 (dark green). Suppressed neurons are those that showed a decrease in response during stimulus presentation (purple shading).

I. Percentage of DS, non-DS, suppressed and non-responsive neurons in each group.

J. Venn diagram showing the overlap of neurons positive for a particular candidate marker gene and the total DS population. (total Vip: n=79; Chrnb3: n=12; Cbln4: n=63; and Itga7: n=36, see Figure S4 for animals and replicates per probe).

See also Figure S5 and S6.

Identification of a putative inhibitory DS neuron subtype

A total of 6 mice were successfully registered between in vivo functional and ex vivo molecular imaging. Each mouse was imaged at one SC location and one plane, yielding a total of 668 neurons. 344 out of the 668 recorded neurons were confidently registered with RNA FISH staining of four candidate marker genes, which were selected based on their relatively sparse distribution and location in the most dorsal part of the sSC (i.e., corresponding to the imaged plane). These included Cerebellin4 (Cbln4, cluster n27); Integrin Subunit Alpha 7 (Itga7, cluster n03); Cholinergic Receptor Nicotinic Beta 3 Subunit (Chrnb3, cluster n25); and Vasoactive Intestinal Polypeptide (Vip, cluster n11). Note that not all 4 probes were successfully stained in all mice, leading to different group sizes. Out of the registered cells in those mice, 18% were positive for Cbln4, 27% Vip+, 4% Chrnb3+, and 19% Itga7+ (Figure S6A).

We examined the response properties of each group of neurons that was positive for a particular maker gene. Representative raw traces of Ca²⁺ signals are displayed in Figure 4D, as well as overlayed responses to stimulus presentation (Figure 4E) and corresponding tuning curves (Figure 4F). Once the cells were grouped according to the expression of marker genes, a striking response specificity emerged: most Vip+ and Itga7+ neurons were barely visually responsive, whereas Cbln4+ neurons displayed highly specific responses. Interestingly, many Chrnb3+ neurons were suppressed by drifting gratings (Figure 4E), with 50% (6/12 neurons) showing a decrease in fluorescence intensity during stimulus presentation. We calculated, for each neuron, its response to a given direction of the drifting gratings (ΔF/F0, averaged over trials). The average response magnitude to the presentation of preferred drifting gratings was 0.61 for Cbln4+ neurons (0.61 ± 0.06, mean ± SEM), but only 0.19 for Vip+ (0.19 ± 0.20), 0.12 for Chrnb3+ (0.12 ± 0.09), and 0.21 for Itga7+ neuron (0.21 ± 0.025) (Figure 4G). Furthermore, we characterized neurons as visually unresponsive if their highest response overall was lower than 0.2 (ΔF/F0 < 0.2). By this criterion, the majority of Vip+ and Itga7+ neurons (n = 55/79, 69.6% and n = 25/36, 69.4% respectively) were not responsive to drifting gratings (Figure 4I). In contrast, most Cbln4+ neurons showed ΔF/F0 > 0.2 (n = 45/63, 71.4%). These results suggest that molecular subtypes of sSC neurons differ strikingly in their visual responses to gratings: Cbln4+ neurons are highly activated, many Chrnb3+ neurons are inhibited, whereas most Itga7+ and Vip+ neurons are weakly responsive.

We next assessed, among the visually responsive neurons, whether they displayed high direction selectivity (i.e., a strong preference for a particular motion direction). We aligned each tuning curve to the direction that elicited the highest response (preferred direction, Figure 4F), where a flat curve indicates little to no selectivity, while a bell-shaped curve indicates a preference for a particular direction. A clear tuning was seen for most Cbln4+ neurons, while only a few Vip+, Itga7+, and Chrnb3+ neurons displayed a high tuning among the visually responsive ones (Figure S6B). To quantify this observation, we calculated the global direction selectivity index (gDSI) for each responsive neuron and determined their direction selectivity. Most Cbln4+ neurons were direction selective (DS; n = 42/63, 66.7%, gDSI >0.2), compared to much lower percentages for Vip+ (14/79, 17.7%), Itga7+ (n = 7/36, 19.4%), and Chrnb3+ neurons (n=2/12, 16.7%) (Figure 4H–I). When considering visually responsive neurons exclusively, the vast majority of Cbln4+ neurons were DS (n = 42 vs. 3, for DS and non-DS neurons, i.e., a 14:1 ratio), while this ratio is much lower for other groups (Vip+, n=14 vs. 8, 1.75:1; Itga7+, n=7 vs. 4, 1.75:1; Chrnb3+, n = 2:1).

Out of all the DS cells, Cbln4+ neurons accounted for about half (n = 42/96, 44%; Figure 4J) and included all preferred directions (Figure S6C). This strong representation of visual responsiveness and direction selectivity is not found in the other populations that expressed one of the four selected markers or among the neurons that lacked these markers (“negative neurons,” Figure 4G–I). All these observations held true with different criteria of classifying DS cells (gDSI >0.25 or 0.15). Finally, Cbln4+ cells (n27.Cbln4) were found to express vGAT, Gad1, and Gad2 (Figure 2A and S6D), genes associated with GABAergic transmission. These findings thus indicate that Cbln4 expression marks inhibitory DS neurons in the sSC.

Discussion

Genetic coding of direction selectivity in the SC

Direction selectivity is a well-characterized feature throughout the visual system, from the retina to area MT in the cortex^41,42. We have previously shown that DS neurons in the sSC include both inhibitory and excitatory neurons^22,23,40, which inherit their selectivity from the precise convergence of retinal ganglion cells²¹. The roles that direction selectivity in the SC may play in image-forming vision remain unclear. Given that the SC is involved in looming-evoked escape and prey capture^3,11, sSC DS neurons may mediate these and other ethologically relevant behaviors. By discovering Cbln4+ neurons as an inhibitory DS cell type, our study has therefore made it possible to record and manipulate DS neurons to test these intriguing possibilities. Our discovery will also facilitate future studies to investigate the molecular mechanisms underlying how SC direction selectivity is established during development. For example, Cbln4 is known to be involved in inhibitory synapse assembly^43,44, suggesting that it may be important for local circuit development in the sSC.

Given that DS neurons include both inhibitory and excitatory ones and Cbln4+ neurons account for about half of the DS population, it is likely that we have identified most, if not all, inhibitory DS neurons. What remains unknown is the molecular identity of the excitatory DS neurons. A small population of Vip+ cells (~15%) and Itga7+ cells (~10%) were DS, though most of them were only weakly responsive. It is thus conceivable that the excitatory DS neurons may distribute across several subtypes, which would raise an intriguing question regarding the difference between inhibitory and excitatory DS neuron subtypes. On the other hand, these small populations could be due to false positive/negative errors intrinsic to all biology experiments including FISH and calcium imaging. Further studies are needed to resolve this issue, potentially using the approach we have developed here.

Cbln4+ neurons display certain heterogeneity in their visual responses, e.g., some were not responsive or responded weakly to the gratings (Fig. S6B). This is likely due to the limited spatial and temporal frequency used in this study. Similarly, only drifting gratings were used, which may explain why most Vip+ and Itga7+ groups were non-responsive. Future studies with more complex visual stimuli and behavioral context will have the potential to reveal more molecular cell types that display unique functions in the sSC. However, it is important to emphasize that, even with limited probes and visual stimuli, we were able to identify an inhibitory DS neuron type. This highlights the power of our multimodal method. In addition, compared to two recent studies that have used a similar strategy to link functions and molecular cell types in visual and somatosensory cortex^45,46, we used a commercial RNAscope HiPlex kit after narrowing down candidate genes based on anatomy. This makes it possible for practically any lab that can perform in vivo calcium imaging, to adopt our approach and determine the molecular identity of the neurons they have imaged functionally.

Neuronal subtypes in the sSC and comparison with previous studies

sSC neurons were traditionally studied and classified into subtypes based on morphology². Gale and Murphy described the visual response properties of a few morphologically defined cell types and discovered transgenic mouse lines in which Cre recombinase was expressed in these subtypes^16,47,48. These studies were significant as they enabled cell-type-specific analysis and manipulations, such as revealing the axonal projections of these cell types and their respective contributions to visual processing and prey capture behaviors^49,50. However, it was not clear whether each Cre-defined cell population corresponds to one or more molecularly defined neuronal subtypes. In fact, the Gad2-Cre line labels all GABAergic neurons in the sSC and they were classified into one group (horizontal cells). This was certainly under-classified given the unusually high percentage of inhibitory neurons in the sSC, as demonstrated by our current study (Figure 2A and Figure S2A) and previous ones^33,35. The two other mouse lines are BAC transgenic lines, where the expression of the driver gene does not necessarily recapitulate its endogenous expression, as shown by studies in the retina and other brain areas^51,52.

Byun et al. examined the expression patterns of 12 well-known molecules in the mouse SC, including transcription factors, cell adhesion molecules, neuropeptides, and calcium binding proteins³⁴. Their results revealed layer-specific expression and led to the conclusion of 10 neuronal types in the sSC. However, most of these candidate molecules are expressed by multiple neuronal subtypes identified in the current study (Figure S2B), reflecting the limitation of the candidate molecule approach.

More recent studies have instead used scRNA-seq or snRNA-seq to examine cell types in the SC^28–30. These studies have revealed a far greater number of neuronal subtypes (“clusters”) than previously reported, which allowed subtype-specific comparisons of retina inputs²⁸ and projection patterns^29,30, as well as their involvement in visual behaviors²⁹. A cluster marked by Nephronectin (Npnt) was shown to be the wide-field vertical cells which receive input from non-DS retinal ganglion cells²⁸ and project to the lateral posterior nucleus²⁹. Npnt+ is also a cluster of excitatory neurons in our data set (cluster n02) and located in the SO, consistent with that of wide-field vertical cells and the laminar expression of Npnt+ neurons reported in another study⁵³. These consistent results support the validity of scRNA-seq and associated clustering analysis. Importantly, whereas the previous atlas^29,30 included the entire SC, ours only contained the sSC. This is not a trivial difference as sSC and deeper SC layers serve distinct functions, and the deeper layers may even be better aligned with the reticular formation than the overlying visual layers⁵⁴. By focusing on the sSC, we were able to identify more of its subtypes and localize their precise laminar distribution. The fact that some clusters in the previous atlas split into multiple clusters in our dataset (Figure S2C–F), which showed a more precise layer distribution, demonstrates the finer resolution of our analysis. Interestingly, it is the inhibitory clusters that often showed the split. This, together with the high percentage of inhibitory neurons in these layers, suggest important roles of sSC inhibitory neurons, such as the direction selective Cbln4+ neurons, in visual processing.

STAR Methods

RESOURCE AVAILABILITY

Lead contact

Further information and requests for experimental data should be directed to and be fulfilled by the lead contact, Jianhua Cang (cang@virginia.edu).

Materials availability

This study did not generate any unique reagents.

Data and code availability

• The raw and processed snRNA-seq data reported in this paper are available online (GEO: GSE223155). A user-friendly interface for visualizing and exploring the snRNA-seq data is available at the Broad Institute Single Cell Portal at https://singlecell.broadinstitute.org/single_cell/study/SCP2161

• R code used in the current study is available online (Zenodo: https://doi.org/10.5281/zenodo.7630422).

• Any additional information required to reanalyze the data reported in this paper is available from the lead contact upon request.

EXPERIMENTAL MODEL AND SUBJECT DETAILS

Animals

This study used a total of 54 adult male and female mice between 2 and 9 months old from the following strains: Chat-Cre (The Jackson Laboratory, Strain #028861; RRID: IMSR_JAX: 028861), Vip-Cre (The Jackson Laboratory, Strain #010908 RRID: IMSR_JAX: 010908), SST-Cre (The Jackson Laboratory, Strain #013044 RRID: IMSR_JAX: 013044) and GAD2-Cre (The Jackson Laboratory, Strain #028867, RRID:IMSR_JAX:028867; Ai9 reporter line (The Jackson Laboratory, Strain #007909; RRID: IMSR_JAX: 007909); Chat-GFP (The Jackson Laboratory, Strain # 007902; RRID: IMSR_JAX: 007902); Tg(Ntsr1-cre)GN209Gsat (MGI: 4367043); H2b-TRAP (The Jackson Laboratory, Strain # 029789; RRID: IMSR_JAX: 029789); and C57Bl/6j (Strain #000664, RRID:IMSR_JAX: 000664). We used 10 mice for scRNA-seq of the superficial Superior Colliculus (sSC): 2 female Chat-Cre x H2b-TRAP, 3 female Chat-GFP, and 5 male Chat-GFP. Of note, these Chat reporter mice were used to obtain cholinergic neurons in a different brain region for a different project. Six C57BL/6j mice were used for fluorescent in situ hybridization (FISH) of vGluT2, vGAT, Vip, Sst and Cbln4 to reveal neurotransmitter identity. Four Vip-Cre x Ai9 and 5 Sst-Cre mice were used for the verification of Vip-Cre and Sst-Cre mouse line and the FISH of Cbln4 and Chrnb3. Two C57BL/6j and 4 Ntsr1-Cre x Ai9 mice were used for FISH of Col18a1 and Npnt. Seven Vip-Cre x Ai9, 3 Sst-Cre x Ai9 and 3 Gad2-Cre x Ai9 animals were used for optimizing procedures to register in vivo functional imaging and ex vivo RNAscope data. After the procedures were optimized, we used 5 Ai9 and 5 Vip-Cre x Ai9 mice for data collection to map visual functions to molecular markers. From these last 2 cohorts, 2 mice were excluded from each group due to poor visual responses.

All mice were kept on a 12hr light/dark cycle under standard feeding and housing conditions, with 2 to 5 animals housed per cage. All experimental procedures were approved by the University of Virginia Institutional Animal Care and Use Committee.

METHOD DETAILS

Single-nuclei RNA-sequencing

Brains were harvested after rapid decapitation to avoid stress- or anesthesia-related changes in gene expression, cooled in ice-cold Hibernate-A for 2min, placed into a chilled stainless steel brain matrix and then sectioned coronally into 1mm thick slices containing the sSC (Bregma −3.0 mm to −5.0 mm). The sSC was then visualized under a dissecting microscope, micro-dissected approximately by knife cuts, and processed into a single-nuclei suspension by using a previously published detergent-mechanical lysis protocol ⁵⁵. Briefly, we homogenized the sSC in lysis buffer and separated nuclei from cellular debris using density gradient centrifugation. After resuspending the nuclei pellet, the concentration of nuclei was counted with propidium iodide by a CellDrop automated cell counter (DeNovix, Delaware, U.S.A.). Next, we used 10X Chromium Next GEM 3’ reagent kit v3.1 according to the manufacturer’s protocol (user guide revision D) to capture each nucleus’s poly-adenylated RNA in a droplet, process it into cDNA sequencing libraries, and sequence them by Illumina Next-Seq 550 or Nova-Seq 6000 to an average of 35,998 mean reads per cell. We loaded approximately 16,500 nuclei per lane of the 10X chip. A total of 5 single-nuclei samples were processed across 3 10X chips (batches): sample 1 was batch 1; samples 2 and 3 were aliquots of the same nuclei suspension processed in parallel as batch 2; samples 4 and 5 were aliquots of the same nuclei suspension processed in parallel as batch 3. Finally, 10X Cell Ranger pipeline (version 3.0.2) was used to demultiplex and align sequencing reads to a custom “pre-mRNA” reference genome created from mouse reference genome RGCm38.98 to generate feature-barcode matrices. We note that previous studies have validated single-nuclei RNA-seq as a method of classifying neuron subtypes^56–58. Although it is possible that it may underestimate expression levels of cytoplasm-enriched transcripts, the single-nuclei approach has key advantages compared to single-cell RNA-seq, in which living cells are often enzymatically dissociated at physiological temperatures. Specifically, single-nuclei RNA-seq has comparable sensitivity for gene detection^56,58 while more accurately representing cell-type proportions⁵⁹ and avoiding transcriptional artifacts related to tissue dissociation and cell handling^60,61.

snRNA-seq data processing and analysis

snRNA-seq feature-barcode matrices were analyzed in R (version 4.0.0) with Seurat v4.0 package ³¹. Across all samples, we filtered out genes detected in fewer than three cells and cells in which we detected fewer than 200 genes, greater than 3,500 genes, or greater than 7,200 transcripts (unique molecular identifiers, UMIs). Then we applied filters to each sample based on their sample-specific distribution of quality metrics. Specifically, we filtered cells based on the number of genes detected (nGenes), the number of UMIs detected (nUMIs), and the percentage of reads mapping to mitochondrial genes (%mt): Sample 1: <450 nGenes, >3,000 nGenes, >7,200 nUMIs, >12 %mt; Sample 2: >3,500 nGenes, >7,200 nUMIs, >1.3 %mt; Sample 3: >3,500 nGenes, >5,000 nUMIs, >1.5 %mt; Sample 4: <350 nGenes, >3,200 nGenes, >7,200 nUMIs, >16 %mt; Sample 5: <350 nGenes, >3,200 nGenes, >7,200 nUMIs, >18 %mt. After that, we integrated the 5 samples to correct for technical variance including batch effects ³¹. In brief, we log-normalized the data; selected 2,000 most variable genes for each batch (“feature selection”); integrated the samples using the IntegrateData() function in Seurat; scaled each gene; performed Principal Component Analysis (PCA) to linearly reduce the dimensionality of the highly variable gene set; clustered the cells using the Louvain algorithm, based on Euclidean distance in the PCA space comprising the first 12 PCs and with a resolution value of 1.8; and performed non-linear dimensionality reduction by Uniform Manifold Approximation and Projection (UMAP)⁶² for visualization in two dimensions. We checked each cluster’s expression of cell type marker genes such as Syn1, Rbfox3, Agt, P2ry12 and Lum. Clusters potentially representing cell doublets were identified based on cluster-level and cell-level co-expression of these cell type marker genes. After removing the suspect clusters, the remaining cells were re-clustered, including the steps of feature selection, PCA, clustering with the top 8 PCs and resolution setting of 0.3. Cluster relatedness in PCA space was illustrated with dendrograms using the BuildClusterTree() function in Seurat.

Clusters in which the neuronal marker genes Syn1 and Rbfox3 were detected at a relatively high expression level were subset and re-clustered following a similar procedure as described above. Quality control clustering (to identify doublets) was done with the top 13 PCs and resolution setting of 1.5. After removing suspected cell doublet clusters based on co-expression of cell type marker genes (Syn1, Rbfox3, Agt, P2ry12 and Lum), the remaining neurons were re-clustered, including the steps of feature selection, PCA, clustering with the top 13 PCs and resolution setting of 1. Genes differentially expressed across clusters were identified using Wilcoxon Rank Sum test and filtered to only genes in a minimum fraction of 0.25 cells in either of the cell populations. Candidate marker genes were selected based on their expression specificity to the given cluster. To assess how sample size affects clustering, we randomly sampled 1k, 2k, 5k, 10k, 15k, 20k or 25k cells and independently re-clustered them using the top 13 PCs and resolution set to 1 (Figure 1A.5). The sample size of >26k neurons reached the saturation level of number of clusters.

To examine potential batch effects, we compared neuron clustering in a single batch with 2-batch and 3-batch combinations (Figure S1I). Neuronal subtypes were consistent across these iterations, with some splitting but rarely combining. The splitting of clusters is expected with more neurons, which increases statistical power and clustering resolution. Furthermore, all samples contributed similarly to the compositions of each of the 28 clusters (Figure S1F), and the percentage of reads mapping to mitochondrial genes were similar across clusters (Figure S1B &G). These data confirm the reproducibility of neuronal subtypes in the sSC across batches and samples.

We compared our results to those of previous studies in which SC cell types were defined by their expression of either a gene-driven Cre recombinase or a priori candidate molecular markers. Plotting expression of the genes identified in these previous studies showed that the majority were not specifically expressed by the neuronal subtypes identified in our snRNA-seq analysis (Figure S2A, B). We also compared our single-cell RNA-seq dataset to previously published ones by mapping each neuron from one dataset (‘query’) onto the neuronal clusters of another dataset (‘reference’) with Seurat’s MapQuery() function³¹. Compared to the previous ones, our dataset yields a finer resolution in clustering (Figure S2C, D) and so provides a more comprehensive atlas of sSC neuron subtypes (Figure S2F).

Perfusion and slicing

Animals were deeply anesthetized with a lethal dose of sodium pentobarbital and perfused with 1x PBS followed by 4% paraformaldehyde (PFA). Brains were dissected while maintaining the dorsal part of the skull and cranial window in place. After overnight in 4% PFA, brains were separated from the skull, dorsal-ventrally aligned and mounted in 5% agarose. Sections of 50μm thickness were made using a vibratome (Leica VT1000S) in PBS. Sections were mounted on a microscope slide in PBS while maintaining orientation and imaged by confocal prior to FISH (20X/0.80 Plan Apo, Zeiss LSM 800). Other brains used for RNAscope were fixed in 4% PFA for one day, cryoprotected in 30% sucrose solution for two and a half days, embedded in O.C.T. compound, frozen in −20 °C freezer and sliced coronally in 30 or 40 μm thickness by cryostat (Leica Cm1950). Slices were collected and stored in cryoprotectant in −20 °C.

Fluorescence in situ hybridization (FISH)

Sections taken from cryoprotectant were washed two times in phosphate-buffered saline (PBS; diluted from 10X stock at pH 7.2) for 5min and mounted on the slides (Fisher Cat. 12-550-15) overnight. Next day, RNAscope Multiplex v2 (Cat. 323100) or HiPlex v2 reagent kit (Cat. 324419) was used and corresponding assay was followed for the staining of coronal sections ⁶³. Briefly, slide-mounted slices were rinsed with PBS, dehydrated in 50%, 70% and 100% ethanol, circled with a hydrophobic barrier, digested by protease IV, incubated in targeted probes at 40 °C for 2h, and treated with Amp 1–3. Probes included Slc17a6 (vGluT2), Slc32a1 (vGAT), Vip, Sst, Cbln4, Npnt, Chrnb3, Col18a1, Tcf7l2, Slit3, and Itga7. For Multiplex procedure, hydrogen peroxide was added onto slide-mounted brain sections for 10 min to quench endogenous peroxidase activity; and HRP - C1 / C2 / C3, dye FITC / Cy3 / Cy5 and HRP blocker were added to bind to Amp3 and trigger the fluorescence. Whereas in HiPlex protocol, fluorophores for the corresponding channels were directly added on Amp 3 to visualize targeted gene expression. Finally, DAPI was used as a counterstain, and fluorescence images were captured by confocal microscope (20X/0.80 Plan Apo, Zeiss LSM 800).

HiPlex procedure was performed for “registration” experiments following in vivo 2-photon imaging due to its capacity for up to 12 probes hybridization in four rounds. In each round, green channels (T1/4/7/10) and red channels (T2/5/8/11) were occupied by GCaMP6f and sparsely labelled tdTomato, respectively. Far-red channels were used for Cbln4-T3, Chrnb3-T6, Itga7-T9 and Vip-T12. After each round, the same area was imaged in each brain section using the same z-stack parameters (0.45 – 0.8 μm interval between z-stacks). The slides were then soaked in 4x SSC to remove coverslip and incubated in 10% cleaving solution to cleave fluorophores, before the next round of hybridization.

In vivo two-photon calcium imaging

The details of 2-photon imaging procedures, including viral injections, imaging, visual stimulation, and data analysis, were provided in our previous publications^39,40. These procedures are only briefly described here, together with key parameters and novel technical details relevant to the registration procedure.

Surgery and calcium imaging:

To image the surface of the SC, a ~2.5mm craniotomy was performed over lambda point a few weeks before imaging. During this procedure, AAV1-Syn-GCaMP6f viral vector (RRID: Addgene_100837, pAAV.Syn.GCaMP6f.WPRE.SV40, titer 2×10¹³) was injected into the SC. The craniotomy was closed using a glass window which pushed anteriorly the transverse sinus, allowing optical access to the SC. A small titanium plate was fixed to the skull using metabond to allow head-fixation during recordings. To facilitate the registration, we used two methods to sparsely label sSC neurons, which were used as anchor points between in vivo functional imaging and ex vivo FISH histological data. In one, Vip-Cre x Ai9 mice were used. In the other, we retrogradely labeled sSC neurons by viral injection into the dLGN. Specifically, 10nL of CAV2-Cre virus (Plateforme de Virologie de Montpellier) were injected into the dLGN of Ai9 mice to label sSC neurons that project there (From Bregma: −1.8mm Anterior-Posterior, 2.2mm Medial-Lateral, 2.5mm Dorsal-Lateral). This resulted in a sparse labeling of neurons in the sSC, facilitating registration between in vivo and ex vivo images.

Five days after the surgery, mice were habituated to the imaging setup, head-fixation, and running on the cylindrical treadmill. Functional data were acquired 2 to 4 weeks after surgery using a two-photon laser scanning microscope (Ultima Investigator,s Bruker Nano Surface Division) coupled to a Ti:sapphire laser (Chameleon Discovery with or without TPC, Coherent) at 920 nm using a 16X, 0.8 NA Nikon objective. The microscope was controlled using PrairieView software (Versions 5.4) and images were acquired using resonant scanning at 30 Hz using 2X or 2.25X optical zoom at 512 × 512-pixel resolution.

Sinusoidal drifting gratings (40° patch, 100% contrast, 0.08 cpd, 2Hz, 30° steps, 12 directions) were displayed on an LCD monitor 25cm away from the mouse, aligned to the receptive field, using MATLAB Psychophysics toolbox⁶⁴. Each stimulus condition of the gratings was presented for 1 s, followed by a gray screen for 3 s, and repeated for at least 10 times in a pseudo-random manner.

Two-photon volume imaging:

Upon completion of the acquisition of functional data, animals were lightly anesthetized with isoflurane (0.5–1%) to reduce motion artifact and placed on a heating pad. A volumetric image of the entire field of view was acquired from the SC surface to ~300μm below, which is the maximal depth with sufficient optical quality for functional imaging. Images were acquired using galvo scanning and averaging two frames per depth in 1μm steps, both at 1 and 2x optical zoom. This volume was used as a reference for subsequent registration.

Registration

After performing HiPlex RNAscope followed by confocal imaging, two approaches were used to identify the same neurons in vivo and ex vivo. For the first batch of animals, where sparsely labeled neurons using retrograde tracing were used as anchor points, each round of in situ hybridization was registered to the functional images independently. For the second batch of animals, where Vip-cre x Ai9 mice were used, all rounds of FISH were first registered to generate a single z-stack using ImageJ (version 2.1.0).

After confocal acquisition for each round of FISH, z-stacks were registered along all 3 axes, allowing the matching of planes across independent acquisitions, using the green channel (GCaMP6f) as a reference. In addition, the sparsely labeled red fluorescence was used as a reference to detect the shift and rotation of each acquisition after merging. Images were generated with up to seven channels: DAPI, GCaMP6f, tdTomato, Cbln4, Chrnb3, Itga7 and Vip when all staining were successful. The resulting registered photomicrographs, containing the staining information from all probes, were then mapped as a single unit to the in vivo acquisitions. To reliably identify the same neurons across in vivo and ex vivo images, sparsely labelled tdTomato expressing neurons were first located across the z-stack acquired in vivo. Their unique relative positions were used to identify which horizontal section contained the in vivo imaged plane. Once the correct slice was identified, the registration was performed using “Control point registration” in MATLAB using the sparsely labeled cells as anchor points. A mask of the ROIs identified in the functional data was applied to the in-situ hybridization photomicrographs and transformed based on the control point registration deformation. These registered images were used to guide the identification of the same neurons in both the in vivo and ex vivo acquisitions across both the z-stacks and time series. Only neurons for which a clear GCaMP6f signal could still be found in the confocal acquisition were used to determine their RNA expression (“Registered neurons”).

Probe identification

To determine if the registered cells expressed the marker genes, registered neurons were marked in the histological data set and assigned to experimenters for RNA FISH analysis. Experimenters were blind to the neuron’s functional and molecular identities. A neuron was considered positive for a particular probe or not based on the number of puncta. This was assessed independently by two experimenters and results in disagreement were reassessed after combining the results, while the probe identification and response properties were still blind to the experimenters.

QUANTIFICATION AND STATISTICAL ANALYSIS

Analysis of anatomical distribution of excitatory and inhibitory neurons

Images of excitatory maker (vGlut2) and inhibitory marker (vGAT) expression were taken by confocal microscope (20X/0.80NA Plan Apo, Zeiss LSM800). The number of vGluT2+ and vGAT+ cells were quantified by Zen 3.4 software (blue edition) based on the puncta density. To quantify their density and distribution, an 800 μm * 250 μm rectangle was drawn from the SC surface and ~300 μm lateral of the midline of brain slice. The rectangle was further divided into 16 subregions in 50 μm bins along the SC depth. The density of vGluT2+ and vGAT+ neurons in each subregion is calculated by the following formula: density per 10,000 μm² = Cell number * 10, 000 μm² / (50 μm * 250 μm).

Analysis of two-photon calcium imaging data

We followed our published procedures to analyze the imaging data^22,23,65. Briefly, ROIs were manually drawn and ΔF/F₀ = (F − F₀)/F₀ was calculated, where F₀ was the mean of the baseline signal and F was the average fluorescence signal. The mean value of ΔF/F0 for each stimulus condition was then used for subsequent data analysis for all the responsive neurons. A neuron was considered responsive if its mean ΔF/F₀ at the preferred condition was above 0.2. To quantify the degree of direction selectivity, we calculated a global direction selectivity index (gDSI), which is the vector sum of ΔF/F₀ responses normalized by their scalar sum^{16,21–23,40}.

Key resources table.

REAGENT or RESOURCE	SOURCE	IDENTIFIER
Bacterial and virus strains
pAAV.Syn.GCaMP6f.WPRE.SV40 (AAV1)	Chen et al. 2013	Addgene, Cat #: 100837
CAV2-Cre	Hnasko et al. 2005	Plateforme de Virologie de Montpellier
Chemicals, peptides, and recombinant proteins
TSA Plus Fluorescein Reagent	Akoya Biosciences	Cat #: NEL741001KT
TSA Plus Cy3 Reagent	Akoya Biosciences	Cat #: NEL744001KT
TSA Plus Cy5 Reagent	Akoya Biosciences	Cat #: NEL745001KT
Critical commercial assays
Chromium Next GEM Single Cell 3’ GEM, Library & Gel Bead Kit v3.1, 4 rxns	10X Genomics	Cat #: 1000128
Chromium Next GEM Chip G Single Cell Kit, 16 rxns	10X Genomics	Cat #: 1000127
Single Index Kit T Set A, 96 rxns	10X Genomics	Cat #: 1000213
Invitrogen Qubit 1X dsDNA HS Assay Kit	ThermoFisher Scientific	Cat #: Q33231
RNAscope® Multiplex Fluorescent Reagent Kit v2	Advanced Cell Diagnostics	Cat #: 323100
RNAscope HiPlex12 Reagent Kit v2 (488, 550, 650)	Advanced Cell Diagnostics	Cat #: 324419
NextSeq 500/550 High Output Kit v2.5 (75 Cycles)	Illumina	Cat #: 20024906
Deposited data
Raw sequencing reads and processed gene expression counts	This paper	GEO: GSE223155
Clustered single-nuclei RNA-seq data	This paper	Broad Single Cell Portal https://singlecell.broadinstitute.org/single_cell/study/SCP2161
Experimental models: Organisms/strains
Mouse: Chat-Cre: B6J.129S6-Chattm2(cre)Lowl/MwarJ	The Jackson Laboratory, JAX	Strain #028861; RRID: IMSR_JAX: 028861
Mouse: Vip-Cre: Viptm1(cre)Zjh	The Jackson Laboratory, JAX	Strain #010908 RRID: IMSR_JAX:010908
Mouse: SST-Cre: Ssttm2.1(cre)Zjh	The Jackson Laboratory, JAX	Strain #013044 RRID: IMSR_JAX:013044
Mouse: GAD2-Cre : Gad2tm2(cre)Zjh	The Jackson Laboratory, JAX	Strain #028867, RRID:IMSR_JAX:028867
Mouse: Ai9: B6.Cg-Gt(ROSA)26Sortm9(CAG-tdTomato)Hze	The Jackson Laboratory, JAX	Strain #007909; RRID: IMSR_JAX: 007909
Mouse: Chat-GFP : B6.Cg-Tg(RP23–268L19-EGFP)2Mik	The Jackson Laboratory, JAX	Strain # 007902; RRID: IMSR_JAX:007902
Mouse: Ntsr1-cre: Tg(Ntsr1-cre)GN209Gsat	The Mutant Mouse Resource and Research Center, MMRRC	MMRRC:030780-UCD
Mouse: H2b-TRAP: B6.Cg-Gt(ROSA)26Sortm1(CAG-HIST1H2BJ/mCherry,-EGFP/Rpl10a)Evdr	The Jackson Laboratory, JAX	Strain # 029789; RRID: IMSR_JAX: 029789
Mouse: C57Bl/6j: C57BL/6J	The Jackson Laboratory, JAX	Strain #000664, RRID:IMSR_JAX: 000664
Oligonucleotides
See Table S1 for Oligonucleotides	N/A	N/A
Software and algorithms
Seurat code used for cell clustering and making figures	This paper	Zenodo: https://doi.org/10.5281/zenodo.7630422
R	R version 4.0.0	https://cran.r-project.org/bin/windows/base/old/4.0.0/
Seurat v4.0	Hao et al., 2021	https://satijalab.org/seurat/
Cell ranger 3.0.2	10X Genomics	https://www.10xgenomics.com/
Zen 3.4 software (blue edition)	ZEISS	https://www.zeiss.com/microscopy/en/products/software/zeiss-zen.html
ImageJ	ImageJ (version 2.1.0)	https://imagej.nih.gov/ij/download.html
Illustrator	Adobe	https://www.adobe.com; RRID: SCR_010279
Prism	GraphPad	https://www.graphpad.com/scientific-software/prism/
Matlab R2021a	MathWorks	“http://www.mathworks.com/products/matlab.html; RRID: SCR_001622”
Psychophysics Toolbox	Brainard, 1997	“http://psychtoolbox.org/docs/Psychtoolbox; RRID: SCR_002881”
Control point registration	Matlab	https://www.mathworks.com/help/images/control-point-registration.html

Highlights.

• snRNA-seq identifies 28 neuron subtypes and their marker genes in the mouse sSC

• The sSC is enriched with inhibitory neurons, in cell numbers and molecular subtypes

• We develop a method to connect each SC neuron’s molecular and functional identities

• We identify Cbln4+ neurons as an inhibitory subtype of direction selective neurons

Inclusion and Diversity

We support inclusive, diverse, and equitable conduct of research.