|Home | About | Journals | Submit | Contact Us | Français|
The thalamus is crucial in determining the sensory information conveyed to cortex. In the visual system, the thalamic lateral geniculate nucleus (LGN) is generally thought to encode simple center-surround receptive fields, which are combined into more sophisticated features in cortex, such as orientation and direction selectivity. However, recent evidence suggests that a more diverse set of retinal ganglion cells projects to the LGN. We therefore used multisite extracellular recordings to define the repertoire of visual features represented in the LGN of mouse, an emerging model for visual processing. In addition to center-surround cells, we discovered a substantial population with more selective coding properties including direction and orientation selectivity, as well as neurons that signal absence of contrast in a visual scene. The direction and orientation selective neurons were enriched in regions that match the termination zones of direction selective ganglion cells from the retina, suggesting a source for their tuning. Together, these data demonstrate that the mouse LGN contains a far more elaborate representation of the visual scene than current models posit. These findings should therefore have a significant impact on our understanding of the computations performed in mouse visual cortex.
The retina parses the visual scene into a set of features that are conveyed to the central visual system. At each stage, the visual scene representation can be transformed to extract new features. The textbook example of this, described by Hubel and Wiesel (Hubel and Wiesel, 1962), is the transformation from circular center-surround receptive fields in the LGN to selectivity for bars or edges of a specific orientation in primary visual cortex (V1). A hallmark of this standard model is that the only information available to V1 from subcortical relays is a set of simple ON and OFF circular receptive fields, and that other properties are computed anew in V1. Understanding the full array of visual features delivered to V1 is therefore crucial in understanding its function (Hirsch and Martinez, 2006).
Evidence has accumulated that there may be more diversity in the signals sent to LGN than generally appreciated (Field and Chichilnisky, 2007; Masland and Martin, 2007). First, a number of more complex operations than simple center-surround have been described in the retina of rodents and rabbits, including direction selectivity, local edge detectors, and sensitivity to differential motion (Gollisch and Meister, 2010). Until recently, it was thought that many of these cell types may not project to LGN, however genetic methods in mouse have shown that direction selective retinal ganglion cells (RGCs) provide monosynaptic inputs to the LGN (Huberman et al., 2009; Kay et al., 2011; Rivlin-Etzion et al., 2011), and retrograde tracing studies in primate have shown at least seven morphologically distinct RGC types that project to LGN (Dacey et al., 2003). Together, these findings raise the strong possibility that diverse visual features may arrive in the LGN.
We chose to investigate LGN response properties in the mouse, which has recently emerged as a prominent model system to study visual processing (Hubener, 2003; Huberman and Niell, 2011). A number of studies have begun to investigate the computations performed in visual cortex, (Niell and Stryker, 2008; Liu et al., 2011; Atallah et al., 2012; Lee et al., 2012; Olsen et al., 2012; Wilson et al., 2012). However, despite the importance of knowing the inputs from LGN in order to understand cortical computation (Gao et al., 2010), few studies have recorded from mouse LGN (Cang et al., 2008; Niell and Stryker, 2010; Olsen et al., 2012), and to date only one study has performed a dedicated characterization of receptive field properties (Grubb and Thompson, 2003). That study confirmed that basic LGN properties are similar in the mouse and other species, in particular the center-surround organization of the standard model. However, this survey depended on a white-noise mapping procedure, which can fail to capture many non-linear response types. A recent publication used in vivo calcium imaging to characterize direction tuning of LGN neurons, but was limited to the superficial-most 75um of LGN (Marshel et al., 2012). Thus the vast majority of the mouse LGN, both in terms of volume and response types, has remained physiologically uncharacterized.
We therefore used multisite extracellular recordings, and applied a broad set of visual stimuli, to characterize the complete repertoire of visual responses throughout the mouse LGN. Given the small size of the mouse brain it was feasible to thoroughly sample the full extent of the LGN with a moderate number of recordings, and thus avoid missing any cell types that might be localized to specific sub-regions, or that exist at relatively low overall numbers. Post-hoc reconstruction of recording sites across multiple experiments allowed us to determine the 3-dimensional organization, and to correlate this with anatomical and genetic markers. Our study, along with the recent work of Marshel et al (2012), shows that mouse LGN does not only convey the standard center-surround pathways. We demonstrate the encoding of diverse features including orientation and direction selectivity along the four cardinal axes, and cells suppressed by contrast. Furthermore, the orientation and direction selective LGN neurons displayed regional biases that were matched to incoming direction tuned retinal afferents, supporting a mechanism for their tuning based on both inheritance and pooling of retinal inputs.
Animals were maintained in the animal facility at University of Oregon or the University of California, San Diego and used in accordance with protocols approved by the UCSD and University of Oregon Institutional Animal Care and Use Committees.
Electrophysiological recordings were performed on adult C57Bl/6 mice (age 2–4 months, both male and female). Recordings were performed generally as described in (Cang et al., 2008) except for a change from urethane to isoflurane anesthesia, which eliminated the need for a tracheotomy and provides a more stable anesthetized state than an injectable agent. We performed these recordings under anesthesia, rather than a previous head-fixed awake paradigm (Niell and Stryker, 2010), in order to facilitate the presentation of a large stimulus set for complete characterization of the population. Previous studies in mouse cortex (Niell and Stryker, 2010; Andermann et al., 2011) have indicated that although firing rates can change, particularly with behavioral state, tuning properties in V1 are generally similar in awake and anesthetized recordings, and one would expect this to be further true one synapse upstream in the thalamus. However future studies focused more closely on particular response properties could investigate the effect of anesthesia and behavioral state on retinogeniculate processing.
In preparation for recording, mice were anesthetized with a surgical level of isoflurane (4% induction, 2% maintenance, in O2), and body temperature was maintained at 37.5° by a feedback-controlled heating pad. Following a scalp incision, the fascia was cleared from the surface of the skull, and a metal headplate was affixed to the skull with vetbond and dental acrylic (Niell and Stryker, 2010). The headplate provided stability for mounting the mouse, and an opening to allow access to the skull. A craniotomy of ~2mm in diameter was performed over the LGN, at 2.5mm lateral and 2.0mm anterior to the lambda suture. The exposed cortical surface was covered with 2% agarose in 0.9% saline to prevent drying and provide mechanical support.
At this point the mouse was transferred to the recording setup, and a dose of chlorprothixene (0.5 mg/kg) was administered intraperitoneally. Administration of chlorprothixene allowed us to lower isoflurane levels to 0.6%, which gave robust visual responses while maintaining an anesthetized state (Kaneko et al., 2008). Recordings were made with silicon multisite electrodes (a1×32-25-5mm-177, NeuroNexus Technologies), which were coated with a small amount of the lipophilic vital dye DiI or DiO (Invitrogen). The electrode was inserted through the craniotomy and overlying agarose using a microdrive (Siskiyou Designs), and lowered down to the LGN. At a depth of ~2500–3000µm, the LGN could be identified by rapid firing in response to either ON or OFF flashes of a small spot at a specific location in the visual field. Beyond localizing to the LGN, the electrode was placed without regard for the presence of responsive units on individual sites, and all units stably isolated over the recording period were included in analysis. Upon locating the LGN, the electrode was further embedded in agarose to increase mechanical stability, and the electrode was allowed to settle in one position for 30 minutes to obtain stable single-unit recordings. The eyes were covered with ophthalmic ointment until recording, at which time the eyes were rinsed with saline and a thin layer of silicone oil (30,000 c.s.) was applied to prevent drying while allowing clear optical transmission.
At the end of recording, the animal was sacrificed under deep anesthesia by cervical dislocation. The brain was removed immediately and fixed by immersion in 4% paraformaldehyde in phosphate buffered saline (PBS) at 4°C overnight, after which 200um coronal sections were cut with a vibratome. The sections were mounted using Vectashield with dapi (Vectorlabs) and imaged on an Olympus BX6l microscope with 2x and 20x air objectives and DP72 camera, to reconstruct electrode penetrations in the LGN.
Data acquisition was performed as described by Niell and Stryker, 2008. Signals were acquired using a System 3 workstation (Tucker-Davis Technologies) and analyzed with custom software in Matlab (MathWorks). To obtain single-unit activity, the extracellular signal was filtered from 0.7 to 7 kHz and sampled at 25 kHz. Spiking events were detected on-line by voltage threshold crossing, and a 1 ms waveform sample on 4 neighboring recording sites was acquired around the time of threshold crossing. Single-unit clustering and spike waveform analysis was performed as described previously (Niell and Stryker, 2008), with a combination of custom software in Matlab and Klusta-Kwik (Harris et al., 2000). Quality of separation was determined based on the Mahalanobis distance and L-ratio (Schmitzer-Torbert et al., 2005) and evidence of a clear refractory period. Units were also checked for stability in terms of amplitude and waveform over the course of the recording time, to ensure that they had not drifted or suffered mechanical damage. Units that were found by histology to be outside the LGN, generally above or below due to the length of the electrode, were excluded from subsequent analysis.
Visual stimuli were presented as described previously (Niell and Stryker, 2008). Briefly, stimuli were generated in Matlab using the Psychophysics Toolbox extensions (Brainard, 1997; Pelli, 1997) and displayed with gamma correction on a monitor (Planar SA2311W, 30 × 50 cm, 60 Hz refresh rate) placed 25 cm from the mouse, subtending ~60–75° of visual space. The monitor was placed either centered directly in front of the mouse or offset to 45deg, and raised or lowered up to 10cm, depending on the hand-mapped RF location of multiunit activity. In order to probe for a broad range of visual response types across an ensemble of simultaneously recorded units, we used a battery of stimuli meant to measure response parameters typically assessed in retina, LGN, and V1, as shown in Figure 1D-G.
Drifting sinusoidal gratings were presented at eight evenly spaced directions of motion, with temporal frequency of 2 and 8 Hz, and spatial frequency of 0.01, 0.02, 0.04, 0.08, 0.16, and 0.32cpd, plus full-field flicker (0cpd). Stimuli were randomly interleaved and presented for 1s duration, with 0.1s inter-stimulus interval. A gray blank condition (mean luminance) was included to estimate the spontaneous firing rate.
Contrast-modulated noise movies (Niell and Stryker, 2008) were created by first generating a random spatiotemporal frequency spectrum in the Fourier domain with defined spectral characteristics. In order to drive as many simultaneously recorded units as possible, we used a spatial frequency spectrum that dropped off as A(f) ~ 1/(f+fc), with fc = 0.05cpd, and a sharp cutoff at 0.16cpd, to roughly match the stimulus energy to the distribution of spatial frequency preferences. The temporal frequency spectrum was flat with a sharp low-pass cutoff at 10Hz. This three-dimensional (ωx, ωy, ωt) spectrum was then inverted to generate a spatiotemporal movie. To provide contrast modulation, this movie was multiplied by a sinusoidally varying contrast with 10 sec period. Each movie was 5 minutes long, and was repeated 2–3 times, for 10–15 minute total presentation.
Sparse flashing noise consisted of ON (full luminance) and OFF (minimum luminance) circular spots on a gray background, at a density such that on average 15% of the area on the screen was covered on any given frame. Spots were 2, 4, 8,16, and 32 deg in diameter, and presented such that each size made up an equal fraction of the area on the screen – e.g. number of spots was inversely proportional to their area, to ensure even coverage at each point in space by every size. In addition, twenty frames each of full-screen ON and OFF were randomly interleaved. Each movie frame was presented for 250ms followed by immediate transition to the next frame, for a total duration of 10mins.
A sparse moving noise stimulus was generated, also using ON and OFF spots, with a more limited size range (4, 8, 16 deg diameter), but each spot was assigned to move in one of eight evenly spaced directions and one of 5 speeds (10, 20, 40, 80, 160 deg/sec). Spots appeared on the appropriate edge of the screen and moved across until they disappeared on the far edge. The movie was presented for 20mins total duration. Although we hoped that this stimulus would give a second measure of direction selectivity, the direction tuning curves were not very robust, although size and speed turning were. In retrospect, this is likely due to the fact that we used circular spots, so that the edge is not always perpendicular to the direction of movement as it is for a bar or grating. Combined with the fact that stimuli were not controlled to pass directly over the center of a given neuron’s receptive field (due to parallel recording) this would lead to a range of apparent directions due to the aperture effect, whereby only the component of motion perpendicular to a contour can be detected when observed through a limited spatial window (Marr and Ullman, 1981).
Data analysis was performed using custom routines written in Matlab. Statistical significance was determined by Mann-Whitney U-test, except where otherwise stated. In figures, *** signifies p<0.001, ** p<0.01, * p<0.05. For figures representing the median of data, error bars show standard error of the median as calculated by a bootstrap. In other cases, error bars represent standard error of the mean.
Neurons in LGN can respond to a drifting grating with either an increase in mean firing rate, or by a modulation of firing up and down around the spontaneous rate, with the latter indicative of a purely linear response. We therefore calculated tuning curves for the drifting sinusoidal gratings using the mean evoked F0 (total firing rate) and F1 (modulation at grating temporal frequency) responses across trials for each stimulus condition, with the measured spontaneous rate for blank trials subtracted. We found that the F0 and F1 generally agreed for neurons with a strong F1 response, so subsequent tuning analysis was based on the F0 response. Spatial frequency tuning was computed using the mean response across all orientations, and orientation selectivity at the spatial frequency that gave the largest response. F1/F0 ratio for a unit was the ratio of the average F1 and F0 response at this spatial frequency. We presented two different temporal frequencies (TFs), 2 and 8Hz, and computed the temporal frequency preferences as (Rhi–Rlow)/(Rhi+Rlow), with Rhi and Rlow being the peak firing rates at each temporal frequency, for an index where −1 indicates only low TF response, +1 only high TF response, and 0 representing equal response to both.
Orientation and direction selectivity indices (OSI / DSI) and preference were computed from these tuning curves using a standard metric based on the circular variance, which consists of a vector sum of the responses across orientations given by
The absolute amplitude of this value gives the index, while the complex phase (or half the complex phase for OSI) gives the preferred orientation/direction. This index provides a global measure of tuning which takes into account both tuning width and depth of modulation in a single index for the clustering algorithm, and proved more robust than curve-fitting for tuning curves that aren’t densely sampled.
In order to analyze the response to contrast-modulated white noise movies, we binned the number of spikes in response to each frame of the movie. The spatiotemporal spike-triggered average (STA) of contrast-modulated movie responses was computed by the mean of the frames at a range of temporal offsets before each spike. Because we used a 1/f power spectrum for the stimulus set, the raw STA is broadened by the correlations in the stimulus set. However, because the stimulus is Gaussian and therefore only contains second-order correlations, we were able to correct the STA by normalizing its Fourier transform by the power spectrum of the stimulus set (Sharpee et al., 2004). While this can account for the under-represented high spatial frequencies, it cannot recover frequencies excluded from the stimulus (higher than 0.16cpd and 10Hz).
We used singular value decomposition (SVD) to separate the joint spatiotemporal receptive field into pairs of spatial and temporal components (Wolfe and Palmer, 1998). For all units with an evident response in the joint STA, this gave a spatial component with a clearly localized response, which was used as the spatial STA with the corresponding temporal STA. Units that did not have a spatial STA component with amplitude significantly above the noise background were left unclassified for spatial RF analysis. The spatial STA was then fit to a 2-dimensional Gaussian to determine receptive field center and amplitude. This fit cannot account for an opposing surround, as this would require a fit to a difference of two Gaussians which would be less robust. However our fits were clearly sufficient for determining RF location and polarity, and for providing a measure of RF center size. The temporal STA was used to assess how biphasic the response was, based on the ratio of the initial maximum or minimum (for ON and OFF respectively) to the subsequent minimum or maximum.
In order to analyze sparse noise movies, we computed spiked-triggered averages for ON and OFF spots separately (to avoid averaging out ON/OFF responses in non-linear units) and determined the receptive field location as the point with the largest absolute magnitude response across the two STAs. We computed peri-stimulus time histograms locked to the onset of each flashed spot that coincided with this location. Histograms were separated out based on polarity and size of the spot. The mean response during spot presentation (250ms) was used to determine response amplitude as a function of polarity and size. For the preferred size and polarity, a measure of sustained response was generated as the ratio of the mean response during spot presentation to the peak response.
For moving spots, a similar analysis was followed except that responses were locked to the time a spot first crossed the receptive field point, and were separated out by speed as well as size and polarity. The mean response throughout the time the spot was over the receptive field location was used to construct a speed tuning curve.
This analysis resulted in a set of response parameters for each unit (listed in Table 1). Each parameter was zero-centered and normalized to have unity variance across the population, and the resulting response profiles were used in a fuzzy k-means clustering algorithm (Matlab Central). We heuristically found that setting the fixed number of clusters to n=6 achieved a tradeoff between poor correlations within clusters (too few clusters) and excessive splitting (too many clusters). Setting n=5 merged the “slow” group into the DS/OS group and sON/sOFF groups, reducing the within-group correlations, and n=7 split the “slow” group into new groups whose characteristic differences were not clear.
In order to reconstruct the 3-dimensional functional organization of the LGN, we determined the location of each recorded unit and mapped it onto a common LGN template, using custom Matlab software. The outline of LGN and the electrode track from post-hoc histology were manually traced. The boundaries of the LGN are readily visible from brightfield and autofluorescence of the histology brain tissue and from dapi staining: the optic tract marks the dorsal and lateral LGN borders and a small but sharp gap in soma density marks the LGN’s medial and ventral borders; this gap represents the major site of afferent and efferent LGN fibers comprising the internal capsule and is invariant with respect to anterior-posterior position in the LGN (see dashed line depicting this boundary in Figure 8).
Using the defined geometry of the electrode, with recording sites spaced at an even 25um from the tip, we identified the spatial location of the electrode site where each isolated unit was recorded. This position was then converted into normalized co-ordinates (fraction of dorsal/ventral and medial/lateral position relative to LGN outline). The normalized position was then superimposed onto the corresponding location on the appropriate A/P section in a set of reference LGN traces. Thus, in spite of slight variations in LGN shape from experiment to experiment, the regional position within the LGN was preserved. Although histological sections were 200um thick, such that 6 sections generally span the LGN, for simplicity and to increase sampling coverage, we mapped this onto 3 evenly spaced A/P templates.
To create maps of functional properties, we performed a spatially-weighted interpolation and normalization to account for the fact that unit data consists of a set of point samples with inhomogeneous density. The location of each unit was convolved with a Gaussian weighting function (sigma = 75um). This was then used to normalize similarly weighted maps of receptive field location and functional identity from cluster analysis. Functional identity, rather than individual response properties, resulted in more distinct segregation since individual properties could often be shared across cell types.
Transgenic mice expressing GFP in direction selective RGCs (DRD4-GFP: Huberman et al., 2009; Trhr-GFP: Rivlin-Etzion et al., 2011) received bilateral intraocular injections of cholera toxin beta conjugated to Alexa 594 (CTb-594; 3µl per eye 0.5% in saline; Invitrogen). 24 hours later the animals were perfused with saline and 4% PFA, one hemisphere marked for orientation purposes and the brains cryoprotected in 30% sucrose (in 1XPBS), then serially sectioned at 40µm in the coronal plane on a freezing sledge microtome. Sections were stained free floating for anti-GFP (rabbit anti-GFP, 1:1000; Invitrogen) using previously published protocols (Huberman et al., 2008), then mounted on slides in rostral-caudal order and coverslipped using Vectashield with dapi mounting media (Vectorlabs). Patterns of dapi, CTb-594 and GFP labeling were analyzed and documented using a Zeiss A2 axioscope, 5X, 10X, and 20X air objectives and an Axiocam-HR digital camera.
Tissue sections were stained using blocking solution (10% goat or donkey serum in 1XPBS with 0.2% Triton-X detergent) for 90 minutes, followed by mouse anti-CSPG or goat FoxP2 antisera (Abcam; 1:500 diluted in blocking solution with 0.2% Triton-X detergent) for 24 hours at room temperature, then rinsed 3×60 min. in 1XPBS, and then incubated in Alexa-594 secondary antibodies (1:1000 in blocking solution; Invitrogen) for 120 min. at room temperature, and finally washed 4×90 min., then mounted onto slides and coverslipped using Vectashield with dapi mounting media. Imaging and data acquisition were as described above.
In order to determine the complete repertoire of visual responses in the mouse LGN, here we used multisite silicon electrodes to record the responses of neurons to a battery of visual stimuli, which were designed to probe for coding of both canonical and potential non-canonical features. We also used the defined geometry of the silicon electrodes to reconstruct the spatial location of different functional LGN neuron types. We combined this data with genetic and histological markers to align LGN regions with retinal inputs carrying specific qualities of visual information and with molecular markers reported to label functionally distinct LGN neurons in other species.
We performed extracellular recordings using multisite silicon electrodes in 16 adult anesthetized mice. In each recording session, we made 1–3 penetrations that targeted the LGN, for a total of 30 penetrations across experiments. Limiting the number of penetrations reduced tissue damage and facilitated subsequent identification of the fluorescently labeled electrode track. Each recording yielded an average of 9 well-isolated single units that were confirmed histologically to reside in the LGN (range = 3–16 units per recording), resulting in 257 total units recorded across all 16 mice.
To assay the spatial organization of functional cell types, we placed a fluorescent lipophilic dye on the electrode, allowing us to label, and later histologically recover, the electrode track to determine its position in the LGN. Figure 1 shows an example of one recording. A 32-channel probe (Fig 1A) spanned the depth of the LGN, leaving a clearly visible fluorescent track in a DAPI stained (nuclear label) histological section (Figure 1B). Using the measured electrode track and the boundaries of the LGN, along with the defined geometry of the recording sites on the probe, we were able to reconstruct the position of each recorded unit. Figure 1C shows example receptive fields of units recorded in this electrode penetration, with a smooth progression of receptive fields (RFs) through the visual field corresponding with position along the electrode.
We presented a battery of visual stimuli designed to probe for both canonical LGN receptive fields as well as known RGC response types. Figure 1D–G summarizes the stimulus set. A 1/f band-limited noise stimulus (Fig 1D) was used to measure spike-triggered average (STA) spatial RF and temporal response. Furthermore, by sinusoidally modulating the contrast of the noise movie (Niell and Stryker, 2008), we were able to measure how the response varies as a function of contrast: either increasing, decreasing, or non-responsive. Drifting sinusoidal gratings (Fig 1E) gave measures of orientation and direction selectivity, as well as spatial and temporal frequency response, and linearity of spatial summation in terms of the F1/F0 ratio. Sparse flashing light and dark spots of a range of diameters (Fig 1F) facilitated the identification of On, Off, or On/Off responses, as well as size selectivity and response timecourse (sustained vs. transient). Finally, a stimulus with moving spots of a range of sizes and speeds (Fig 1G) was used to probe motion selectivity.
For each unit, this stimulus set resulted in a response profile, defined by a vector of 15 computed parameters shown in Table 1. These responses were highly heterogeneous across the population, suggesting the presence of multiple functional types, which would be lost in averaged data across the population. In order to extract groups with similar response profiles, we used a clustering algorithm (fuzzy k-means) to sort the units. A small subset of neurons (18%) did not give responses to a sufficient number of stimuli to be included in the clustering algorithm; these are included below in total cell proportions, but not in the group analysis. Selecting a total of 6 clusters resulted in correlated response profiles within groups (Figure 2) that were not correlated across groups. The choice of 6 clusters was heuristic, based on a balance between averaging out heterogeneity and unwieldy continued subdivision into smaller clusters. The correlations within clusters, even when pooling across recording sessions (Figure 2C), indicates that the clustering system identifies cell types that are consistent across preparations.
Each of the groups had a distinctive characteristic, which we use to identify them. The first three groups had standard LGN center-surround responses – sustained ON (sON; n=64, 25%), sustained OFF (sOFF; n=28, 11%), and transient OFF (tOFF; n=42,16%). The next group showed a high degree of selectivity for either one direction of motion (direction selective, DS) or two opposite directions of motion (orientation selective, OS), a group which here we refer to as DS/OS (n=28, 11%). The next group showed a profound reduction in firing in response to nearly all stimuli, which we refer to as “suppressed” (n=14, 5%). The last group was more heterogeneous than the others, but shared a commonality of responding with longer latencies to slow stimuli, which we refer to as “slow” (n=38, 15%). Table 1 summarizes the response parameters used for clustering each of these groups, and we address each group in turn below.
We first discuss the three groups (sON, sOFF, tOFF) of LGN responses that resembled standard LGN responses described in previous studies (Wiesel and Hubel, 1966), to provide a reference point for the more diverse response types described below. These three groups had robust STA spatial receptive fields with a strong center response (Figure 3A). For some units an opposing surround could be observed but these were generally much weaker. The mean receptive field radius (half-width at half-maximum) was 4.9 ± 0.3° (Fig 3D). These units separated into clusters of ON or OFF, based on both the STA amplitude (Fig 3A,E) and the response to either light or dark flashing spots (Fig 3C).
The temporal response profile revealed a further distinction. The STA temporal kernel (Fig 3B) was either monophasic, which corresponded to units with a sustained firing in response to a flashed spot, or biphasic, which corresponded to units with a transient response to flashed spots. A biphasic temporal response effectively acts as a “differentiator” which detects changes in luminance. Consistent with this, we found that these units also generally fired in response to the offset of a spot of its non-preferred luminance (Figure 3C bottom). Strikingly, while we found both ON and OFF sustained responses, transient units only responded to decreases in luminance (tOFF; Fig 3F). This is consistent with previous recordings of mouse alpha ganglion cells (Pang et al., 2003; van Wyk et al., 2009).
Responses to drifting gratings were consistent with the spatial properties derived from reverse correlation, as exemplified in Fig 3G. Corresponding to a circular receptive field, this unit responded to all orientations nearly equally and showed periodic firing at the temporal frequency of the grating, indicating linearity of spatial summation as expected for units that respond to only ON or OFF. The spatial frequency tuning curve demonstrated bandpass tuning (Fig 3H) and the size tuning showed reduced response to large stimuli (Fig 3I), both of which are consistent with a center-surround organization, where the opponent surround reduces response to features larger than the center receptive field. Nearly all units in these groups had this strong size suppression, as shown in Fig 3J. Interestingly, these units showed greater response to fast-moving stimuli than slow (example in Fig 3K), which contrasts with findings for other cell-types described below.
The textbook model of vision science is that the fundamental transformation from LGN to cortex is from center-surround receptive fields to orientation selectivity. However, we found a substantial population (28/257; 11%) of neurons in the LGN that displayed a strong selectivity for either one direction of motion for sinusoidal gratings (direction selectivity- DS) or the two opposing directions of motion of a given orientation (orientation selectivity - OS). We refer to units selective to both directions of motion of a single orientation as “orientation selective” to follow recent nomenclature describing cortical tuning to moving gratings (Ringach et al., 2002; Kerlin et al., 2010; Rochefort et al., 2011). However, a more explicit description might be “bi-directional” selectivity, since they respond to two opposing directions, or “axial” selectivity since they respond to motion about a single axis.
The direction and orientation response types were classified into a single group (denoted DS/OS) by the clustering algorithm, since their other tuning properties were similar and the DS neurons alone were a small proportion of this group. Examples of three DS/OS units are shown in Figure 4. Figure 4A shows a unit tuned for anterior-directed motion, while Figure 4B shows tuning for both upward and downward motion and the unit displayed in Figure 4C shows both anterior and posterior responses. In contrast to a typical center-surround neuron (above, Fig 3G), which responds equally to all directions, these units have a dramatically decreased “null” response to the opposing or orthogonal directions of motion. Furthermore, rather than the periodic response observed in center-surround neurons, these neurons fired more continuously during the presentation of their preferred stimulus, indicating nonlinear summation.
The bottom left panels of Fig 4A–C show spatial and temporal frequency tuning for these neurons, which generally responded to high spatial frequencies and the lower of two temporal frequencies (2 vs. 8Hz) that were presented. This is consistent with the tuning for speed of moving spots (Fig 4A–C, bottom right), which had a peak at lower speeds (10–40 deg/sec).
The summary of these response properties across the population is shown in Figure 5. Direction and orientation selectivity were measured using a standard metric based on the circular variance of the tuning curve (Ringach et al., 2002), with zero representing a uniform response, and one representing a response at exactly one direction (DSI) or two opposing directions (OSI). Figure 5A,B show the distribution of OSI and DSI across the population with measurable response to drifting gratings, with the proportion assigned to the DS/OS group superimposed in red. These form a long tail off of the central peak around OSI/DSI=0.
The preferred orientation of all units in the DS/OS group is shown as a polar histogram in Figure 5C, demonstrating that nearly all the preferred orientations were along the cardinal axes of up-down and anterior-posterior. The preferred direction for the DS subgroup (DSI>0.33) is shown in 5D. We found only two directions, posterior and downwards, although given the small sample size of these neurons (n=7) we cannot rule out other directions.
As shown in Figure 5E, the contrast between periodic firing in center-surround neurons and the more continuous firing of DS/OS neurons, seen in the example neurons, was maintained across the population as measured by the F1/F0 ratio. This is a standard measure of linearity of spatial summation (Skottun et al., 1991), whereby F1/F0=0 corresponds to continuous firing (nonlinear) and F1/F0=2 corresponds to firing at only a single phase of the periodic input (linear). Furthermore, DS/OS units responded to higher spatial frequency gratings (Fig 5F) and lower speeds of moving spots (Fig 5G), both consistent with the lower speed tuning of direction selective RGCs (Weng et al., 2005). In addition, they showed a longer latency to flashed spots (Figure 5H).
Consistent with the preponderance of nonlinear responses, we were not able to map STA receptive fields for the majority of the DS/OS population (57%; 16/28). Among the units that did have a clear STA receptive field, these generally consisted of a single roughly circular region, which is quite distinct from the spatial receptive field of orientation and direction selective simple cells in cortex (McLean et al., 1994; Ringach, 2002), including in mouse (Niell and Stryker, 2008; Bonin et al., 2011). However, a small fraction of OS units had either a single elongated region (7%; 2/28) or adjacent On and Off regions (7%; 2/28), which could provide a spatial basis for their selectivity.
The group of units we term “slow” responses shared many properties with the DS/OS group (Fig 5E–H), excluding orientation selectivity. They were united by responses to lower speeds of moving spots (Fig 5G), and longer latency in response to flashing spots (Figure 5H). Furthermore, although we could map STA receptive fields for some of these units, they often responded to onset of both ON and OFF flashing spots, consistent with a lower F1/F0 ratio (Fig 5E). Qualitatively, these types of responses are often referred to as “sluggish”, a term often used for the broad category of W-like responses in other species (Casagrande, 1994).
Another theme in central visual processing is that neurons signal contrast with an increase in firing rate. This is demonstrated in Figure 6A by the typical response of a center surround LGN neuron to the contrast-modulated noise movies (Figure 1D). When the contrast is low (gray screen) the cell fires at a baseline spontaneous rate, which is then elevated as the contrast of the stimulus increases. However, the group of neurons we term “suppressed” showed the opposite response, decreasing their firing rate in response to nearly any stimulus presented. An example unit is shown in Figure 6B. This neuron has a high spontaneous firing rate when there is no contrast on the screen, and then decreases its firing until it is nearly silent during the high contrast phase of the movie.
Furthermore, this suppression of firing occurs across a large range of stimuli that drive the rest of the population. As shown in Figure 6C, during the presentation of a drifting grating of any orientation, the firing of this cell nearly stops. In addition, the spatial frequency tuning curve in 6D shows that from 0.01cpd to 0.16cpd, the neuron’s firing is strongly reduced relative to spontaneous rate. The only spatial frequencies that do not suppress firing are 0.32cpd, at the high end of the spatial acuity, and 0cpd, a full-field flicker that does not contain spatial contrast.
Figures 6E, F compare these properties to the rest of the population. This confirms that suppressed neurons as a population have a dramatically higher baseline firing rate than the rest of the population (Figure 6E), and are unique in having substantial decreases in firing rate during noise movies (Figure 6F). In fact, this decrease is approximately the same magnitude as the increase of firing rate seen in the standard population of center surround units.
Although the mouse LGN does not have the overt cellular lamination found in other species, classic studies in rat (Martin, 1986; Reese, 1988) suggested that ‘hidden’ segregation of functional pathways is indeed present in the LGN. More recently, genetic methods that label the axonal projections of specific RGC types showed that RGCs with distinct response properties project to different subregions of the LGN– for example, Off-alpha RGCs with transient light responses project to the medial portion of the LGN (Huberman et al., 2008), whereas ON/OFF direction-selective ganglion cells (DSGCs) project to a specific laminar termination zone adjacent to the optic tract (Huberman et al., 2009; Kay et al., 2011; Rivlin-Etzion et al., 2011). Furthermore, a recent morphological study of mouse LGN neurons found three categories of cells which resemble the X, Y, and W LGN neuron morphologies described in the cat; each LGN cell type had distinct, though partially overlapping, distribution in the LGN, most notably, the W-like cells were biased to reside in the region adjacent to the optic tract (Krahe et al., 2011)- the same location where the axons of On-Off direction selective RGCs terminate (Huberman et al., 2009; Rivlin-Etzion et al., 2011).
In order to assay the functional organization of the LGN, we used the location of each recorded unit (Fig 1B) to construct functional maps, by combining the sites from individual recordings onto a common LGN template. We used the normalized position within each individual reconstructed LGN, so that even if the shape of the LGN varied slightly from specimen to specimen, the relative location (anterior-posterior, dorsal-ventral, medial-lateral) would be preserved in the ensemble data set. Furthermore, to account for variations in the density of sampling, we used a spatial interpolation and normalization to calculate smooth maps of functional properties.
In mouse, the two-dimensional map of the visual field in the retina is superimposed onto a 3-dimensional structure in the LGN. This is in contrast to other species such as cats and macaque monkeys, which have distinct cytoarchitectural laminae in the LGN. In those species, each layer contains a separate two-dimensional map of retinotopic space. Given the 3-dimensional structure of the mouse LGN, it is not clear how other forms of functional organization would relate to retinotopy. Although anatomical studies have measured the retinotopic organization of RGC axons in the mouse LGN (Pfeiffenberger et al., 2005) no functional analysis of the complete representation of visual space in the mouse LGN has been carried out. By using the RF locations calculated from STA responses to band-limited and sparse noise movies, we constructed retinotopic maps for azimuth and elevation through the LGN, which we present as three A/P sections – anterior, middle, and posterior. As seen in Figure 7A, B, both azimuth and elevation are represented smoothly across the LGN, in roughly orthogonal alignment. There is a mostly complete map of visual space present in each A/P section, although the representation shifts slightly downward in visual space towards the posterior LGN. Thus, because each point in visual space is represented at multiple points throughout the LGN, there is potential for multiple representations to be spatially segregated or have regional biases, perhaps according to specific encoding properties and not unlike other species.
To test this, we mapped the location of the different functional groups described above into density maps on the LGN. The most obvious and striking organization was seen for the DS/OS group (Fig 7C), which was heavily over-represented in the posterior LGN, and throughout the LGN, DS/OS neurons were more likely to be found in the region adjacent to the overlying optic tract. Indeed, this DS/OS enriched region can be envisioned as wrapping around the dorsolateral surface of the LGN and onto the posterior pole, and thus represents a functionally segregated compartment of the mouse LGN. Other cell types showed a less pronounced bias, although the center-surround cell-types had a higher density in the anterior LGN, complementing the enrichment of DS/OS in the posterior LGN (Fig 7D). However, it is clear that there is not a strict segregation, as even at its highest the density of DS/OS cells is ~40%, and there are always intermingled cells of other types. Indeed, it is important to note that nearby units recorded at the same or at neighboring sites often had different response types.
Next we considered the distributions of LGN response properties relative to genetic and molecular markers of different types of incoming RGC afferents and LGN cell types. Our evidence for variations in DS/OS responses along the medial-lateral as well as anterior-posterior extent of the LGN (Fig 7C above) prompted us to analyze the axon termination zones arising from On-Off direction tuned RGCs as a function of their anterior-posterior location, a feature which has not been analyzed or reported previously for any DSGC type. Serial sections through the entire LGN of 8 adult mice were analyzed. The boundaries of the LGN were defined by cytoarchitecture (Fig 8A-C) and retinogeniculate terminations of all RGCs labeled by intraocular injections of the anterograde tracer cholera toxin beta conjugated to Alexa 594 (Fig 8D-E). By analyzing transgenic mice that selectively express GFP in posterior tuned direction selective RGCs we observed that axons from these RGCs terminate in the lateral LGN, close to the optic tract (Fig 8G–I) (Huberman et al., 2009; Rivlin-Etzion et al., 2011; Wei et al., 2011). Interestingly, comparison across the anterior-posterior extent of the LGN revealed a systematic expansion of DS terminations as we progressed toward the posterior pole of the LGN. Indeed, within the posterior LGN, direction selective RGCs terminated across a far greater extent of the LGN (~60%) as compared to within the anterior LGN (~20%). These observed anterior-posterior variations are nicely aligned with the electrophysiological signatures of DS/OS LGN neurons described above, and raise the possibility that the responses of DS/OS LGN neurons are the direct consequence of inputs from direction tuned retinal afferents.
Previous work addressed the question of whether there are histochemical markers of functionally distinct layers and/or cell types in the monkey and cat LGN, prompting us to ask whether the same markers delineate anatomical and/or functional subdivisions of the mouse LGN explored in our study. Chondroitin sulfate proteoglycans are reported as specific to the magnocellular layers of the monkey LGN and the “Y” (sometimes called “A”) layers of the cat LGN (Murray et al., 2008). Also, the transcription factor FoxP2 is specific to the parvocellular layers of the marmoset (Mashiko et al., 2012) and macaque LGN and X cells in carnivores (Iwai et al., 2012). We stained the mouse LGN for these markers and found that, whereas many LGN cells expressed CSPG and Foxp2 (Fig 8 J, K), neither marker displayed any obvious laminar or overt regional biases. It is interesting to note however, that not every LGN cell expressed CSPG or Foxp2 (Fig 8L-Q). Thus, if these markers represent Y/magno-like cells or X/Parvo-like cells, both cell types are scattered in “salt and pepper” fashion throughout the LGN in mouse.
Taken with our analysis of the spatial distributions of LGN neuron response properties, these molecular genetic expression studies underscore the degree to which mouse DS/OS tuned LGN neurons are strongly localized to a defined area of the LGN, as opposed to the more widespread interspersed distribution of neurons with center-surround and other response properties. To some extent this latter feature is reminiscent of mouse visual cortex, where orientation selectivity exists at the single cell level despite the absence of more macroscopic maps such as orientation pinwheels or columns (Ohki and Reid, 2007).
Our recordings and anatomical reconstructions reveal a much higher degree of functional sophistication and compartmentalization in the mouse LGN than previously suspected. Not only does the mouse LGN encode three ‘standard’ types of center-surround receptive fields, it also encodes two properties not generally thought to be present in the LGN, but that are present in the cortex: orientation selectivity and direction selectivity. The mouse LGN also encodes a suppressed-by-contrast mechanism that signals uniformity of the visual field, and a heterogeneous assembly of slow W-like cells that likely yield even more diversity. Furthermore, the mouse LGN displays a previously unknown aspect of functional organization, with direction and orientation selective units greatly enriched in the posterior and dorsolatereral LGN. Importantly, our findings provide a genetically tractable system to investigate the mechanisms of these multiple channels, including both canonical and non-canonical cell types, in cortical computations.
The use of multisite silicon electrodes had multiple advantages for our study. First, we could place them in the LGN and record any neurons present, so we were not limited to recording from cells that were readily activated. Furthermore, we could maintain recordings from an ensemble of neurons over an extended period, allowing us to present an extensive stimulus set necessary to characterize diverse cell types. The linear geometry of the probe enabled us to sample neurons throughout the volume of the LGN, and importantly, to use post-hoc histology to determine the spatial location of individual recording sites.
To seek order in the rich array of response types we found from our battery of visual stimuli, we used a clustering algorithm to generate groups with related response profiles, as has been performed previously for retinal ganglion cells (Carcieri et al., 2003; Farrow and Masland, 2011). This yielded six broad groups that could be roughly mapped onto previously described retinal or LGN cell types. This approach thereby allowed us to extract groups of neurons with similar response profiles, without relying on fixed thresholds, such as the standard 3:1 selectivity for orientation tuning. Clustering allows the entire response profile to be used, and in our case found patterns of response properties that varied together within groups. It also allows the discovery of distinctions that may not have been expected a priori. However, because the choice of number of clusters to generate was made for classification purposes, it is not to be taken as a rigorous definition of the total number of distinct channels in the LGN.
It is not clear why the diverse cell types we delineate have not been widely described in the LGN of other species, despite the fact that there are scattered examples of each unusual type described in the literature (reviewed in (Masland and Martin, 2007)). It may be that standard stimulus sets do not probe for these response types, or they may get classified as unresponsive or outliers. Furthermore, it appears they make up a higher proportion of the total LGN neuron population in mouse than other species where they have been reported. Species with high acuity may need a far greater number of center-surround units to cover the visual field; if the absolute number of non-canonical neurons needed to span the same area does not change across species, this could explain the lower density.
The presence of direction and orientation selectivity in the LGN, although not entirely unprecedented (Levick et al., 1969; Thompson et al., 1994), raises important questions about both the source of their tuning and their role in defining receptive fields properties in cortex. Direction tuned RGCs were first described in the retinas of rabbits (Barlow and Hill, 1963) and more recently were characterized in the mouse retina (Weng et al., 2005; Huberman et al., 2009; Kay et al., 2011; Rivlin-Etzion et al., 2011). A recent in vivo imaging study of mouse LGN also confirmed the existence of DS/OS (Marshel et al., 2012), although they reported primarily anterior-posterior selectivity, rather than all four cardinal directions of selectivity as observed here. The discrepancy in findings may be due to the fact that their imaging was restricted to the superficial-most 75µm of the LGN whereas the penetrating electrode recordings we describe here could survey all portions of the LGN, and thus identify a greater number and diversity of cell types.
It is interesting to note that while we found DS/OS cells strongly tuned for each cardinal axis of motion (up, down, left and right), there were few DS/OS cells with peak tuning for intermediate directions (e.g., 45 degrees). The DSGCs that project to the LGN also display this cardinal-biased tuning. Furthermore, the spatiotemporal properties of these units, including ON/OFF responses and tuning for low speeds, are consistent with DSGC’s in the retina (Weng et al., 2005). Combined with the fact that DS/OS LGN neurons lie in direct register with DSGC axon terminations along the full anterior-posterior extent of the LGN (Figs. 7 and and8),8), our data suggest that many DS/OS LGN neurons may derive their motion selectivity directly from RGC tuning.
It is particularly striking that OS responses were much more abundant than DS responses in the LGN, since a number of DS RGCs have been identified, but so far reports of OS responses in the retina are scarce. Given that mature LGN neurons receive input from three or fewer RGCs (Chen and Regehr, 2000; Jaubert-Miazza et al., 2005), our findings are consistent with the possibility that OS tuning of mouse LGN neurons arises from combined influences of 2–3 DS tuned RGCs, as in a recently proposed model based on in vivo imaging (Marshel et al., 2012). This would be a novel method of computing orientation/axial selectivity, in contrast to the classic model of OS tuning for V1 neurons arising from a linear array of center-surround receptive fields (Hubel and Wiesel, 1962). It also raises the striking developmental question of how LGN neurons could specifically restrict input to two opposed directions of motion, rather than pooling other directions as well.
Does the cortex inherit direction and/or orientation selectivity from the LGN, or is it computed anew? It appears unlikely that the orientation selectivity seen in mouse LGN is directly inherited, as mouse V1 orientation selective units are primarily simple/linear cells, whereas the DS/OS units we found in LGN were generally non-linear. Furthermore, the DS/OS population is a still a small fraction of the LGN population, relative to the population of center-surround neurons that could support OS by traditional summation models (Hubel and Wiesel, 1962; Chapman et al., 1991; Alonso et al., 2001). However, it is possible that the selectivity transmitted from the LGN could provide a scaffold for developmental processes that shape DS/OS in V1 (Rochefort et al., 2011). It is also possible that the DS/OS cells from LGN provide a dedicated channel to a subset of nonlinear neurons in V1. In either case, our results suggest the need to extend models of circuitry in the cortex of the mouse to include directionally tuned input from the LGN.
The “suppressed” group clearly corresponds to suppressed-by-contrast units, a type of rarely encountered neuron that has been described in the retina of cat (Rodieck, 1967) and rabbit (Levick, 1967), as well as the monkey LGN (Tailby et al., 2007). These responses have also been found in V1 of awake mice (Niell and Stryker, 2010), where they are strongly modulated by behavioral state: when an animal is stationary the cells are inactive, but when the animal initiates locomotion the cells greatly increase their firing rate and show the suppression to a broad range of contrast stimuli described here.
The role of suppressed-by-contrast neurons is an open question, since they run counter to most ideas of visual processing, and have only been described in a few studies. However, the results presented here and in (Niell and Stryker, 2010) demonstrate that their numbers are not insignificant in the mouse visual system. It has been proposed that they could provide a signal to control contrast gain (Troy et al., 1989), or serve to mask out regions of constant illumination (Masland and Martin, 2007). It is also possible, particularly given their “inverse” response relative to the rest of the population, that these cells are inhibitory interneurons, although their presence in the retina and cortex as well suggests a projection pathway.
The center-surround units we found resemble the standard pathways described in other species (Nassi and Callaway, 2009), parvo- and magnocellular in primates and X and Y in carnivores. Because the homology between these two pathways in primates versus carnivores is still not clear, it is difficult to map our response types directly onto the two pathways. We do find a distinction between sustained and transient responses, which often distinguishes parvo/magno and X/Y. However, similar to the results of Grubb and Thompson (Grubb and Thompson, 2003), we do not find other striking differences between the sustained and transient cells such as receptive field size or linearity. Strikingly, the three types we found do correspond closely to physiological properties of mouse alpha retinal ganglion cells, including the absence of an On-transient type (Pang et al., 2003; van Wyk et al., 2009). This suggests that alpha RGCs provide the primary input to the canonical pathways, however this remains to be tested, and indeed the lack of On-transient responses could extend to other retinal populations as well.
Furthermore, although we found some spatial correspondence between DS/OS neurons and cells morphologically described as “W” cells (Krahe et al., 2011), we did not see a clear spatial segregation of our center-surround units corresponding to Krahe et al.’s described morphological X/Y distributions, either in functional maps or anatomical markers that label these cell types in carnivores and primates. Together, these results suggest that we may not have found the appropriate distinction that correlates with these morphologies, or that different mapping methods are needed to detect their differences in distributions, if those differences exist. One imagines that the homology of mouse versus primate and cat LGN neurons may be clarified, in part, by studies of the cortical areas and cortical layers these functionally and anatomically defined LGN neuron types project to and/or receive feedback from (Briggs and Usrey, 2009; Nassi and Callaway, 2009).
It is of interest that the third classically described visual pathway through the primate LGN, termed koniocellular (or “W” in cats), generally shows less brisk visually-evoked activity, and a fairly broad range of tuning properties (Casagrande, 1994; Hendry and Reid, 2000). The “slow” group we found bears many similarities to this pathway, including long latencies and response to both ON and OFF. To date, attempts to classify the range of koniocellular/W responses have been limited, primarily due to their diversity, however the relative abundance and potential for genetic access in mice may lead to more insight. Furthermore, experiments tailored to identify other non-canonical RGC types, such as local edge detectors (Zeck et al., 2005; Russell and Werblin, 2010), may find them in this diverse group, or within the small fraction of units that were not responsive to our stimulus set.
It is worth noting that the relative proportion of center-surround and DS/OS neurons in mouse LGN is almost opposite the proportion of their counterparts in the retina. On and Off alpha RGCs represent ~5% of the total RGC population in mouse (Huberman et al., 2008; Pang et al., 2003) and yet the percentages of LGN neurons recorded with sON, sOFF, or tOFF responses was at least double that value for each type. This could be due to the fact that alpha RGCs preferentially target the LGN, relative to other RGC subtypes in the mouse and indeed, there is evidence for this (e.g., compare Huberman et al., 2008 and Osterhout et al., 2011). A non-mutually exclusive possibility is that because mouse alpha RGC axonal arbors are relatively large (Pang et al., 2003; Huberman et al., 2008; Hong et al., 2011; and A.D. Huberman, unpublished observations) they may innvervate far more LGN neurons than their absolute numbers suggest.
There is a mismatch in the opposite direction for the relative number of On-Off direction selective RGCs versus LGN neurons with DS/OS properties. Assuming approximately equal numbers of each subtype representing each of the four cardinal axes, On-Off DSGCs comprise ~40–50% of the RGCs in the mouse (Huberman et al., 2009; Rivlin-Etzion et al., 2011; Kay et al., 2011), whereas the overall number of DS/OS LGN neurons is about one-quarter that value. In contrast to alpha RGCs, On-Off DSGCs restrict their axons to a narrow region of the LGN (Figure 8), which can partly explain this relative mismatch.
Together, our data reveal the rich repertoire of visual signals contained in the mouse LGN and therefore, that are conveyed to visual cortex. Given that many of the RGC types that generate this information have been genetically identified in the mouse, it should soon be possible to causally assess the impact of specific retinal output pathways on LGN and cortical processing, using cell-type specific manipulations of activity (Zhang et al., 2007; Alexander et al., 2009). In the future, genetic identification and manipulation of specific LGN neuron types should also be immensely useful for understanding how visual information is encoded and transformed in the geniculo-cortical pathways, and their role in driving specific aspects of perception and behavior.
We thank Phong Nguyen for expert technical assistance, Dr. Sunil Ghandi, and members of the Huberman and Niell labs for thoughtful discussions and for comments on an earlier version of the manuscript. This work was supported by NINDS-NIH 5T32NS007220-30- Neurobiology Training Grant (R.N. E-D), the Searle Scholars Program (C.M.N.) and the Sloan Foundation (C.M.N.), and by grants from the Whitehall Foundation (A.D.H) and NIH R01 EY022157-01 (A.D.H).
The authors declare no competing financial interests.