Maps of cone opsin input to mouse V1 and higher visual areas

Two elements of this study are particularly novel. For one, it is the first to quantify cone inputs to mouse visual cortex; we have measured cone input in five visual areas. Next, it is the first study to identify a feature map in the mouse visual cortex that is based on well-characterized anisotropy of cones in the retina; we have identified maps of opsin selectivity in five visual areas.

Abstract

Studies in the mouse retina have characterized the spatial distribution of an anisotropic ganglion cell and photoreceptor mosaic, which provides a solid foundation to study how the cortex pools from afferent parallel color channels. In particular, the mouse’s retinal mosaic exhibits a gradient of wavelength sensitivity along its dorsoventral axis. Cones at the ventral extreme mainly express S opsin, which is sensitive to ultraviolet (UV) wavelengths. Then, moving toward the retina’s dorsal extreme, there is a transition to M-opsin dominance. Here, we tested the hypothesis that the retina’s opsin gradient is recapitulated in cortical visual areas as a functional map of wavelength sensitivity. We first identified visual areas in each mouse by mapping retinotopy with intrinsic signal imaging (ISI). Next, we measured ISI responses to stimuli along different directions of the S- and M-color plane to quantify the magnitude of S and M input to each location of the retinotopic maps in five visual cortical areas (V1, AL, LM, PM, and RL). The results illustrate a significant change in the S:M-opsin input ratio along the axis of vertical retinotopy that is consistent with the gradient along the dorsoventral axis of the retina. In particular, V1 populations encoding the upper visual field responded to S-opsin contrast with 6.1-fold greater amplitude than to M-opsin contrast. V1 neurons encoding lower fields responded with 4.6-fold greater amplitude to M- than S-opsin contrast. The maps in V1 and higher visual areas (HVAs) underscore the significance of a wavelength sensitivity gradient for guiding the mouse’s behavior.

NEW & NOTEWORTHY Two elements of this study are particularly novel. For one, it is the first to quantify cone inputs to mouse visual cortex; we have measured cone input in five visual areas. Next, it is the first study to identify a feature map in the mouse visual cortex that is based on well-characterized anisotropy of cones in the retina; we have identified maps of opsin selectivity in five visual areas.

at the front end of visual system circuitry is a mosaic of photoreceptors that tiles the retina, each containing opsins that are sensitive to a unique range of wavelengths. Most mammals are dichromats, meaning their daytime photoreceptors (i.e., cones) express variable amounts of two specific opsins. Mice are dichromats, and the pathways of their early visual system have become an important model for understanding sensory networks (Huberman and Niell 2011). In particular, many studies have characterized the spatial distribution of the mouse’s two opsins (“S” and “M”) in the photoreceptor mosaic, which exhibits considerable anisotropy (Applebury et al. 2000; Calderone and Jacobs 1995; Chang et al. 2013; Ekesten and Gouras 2005; Haverkamp et al. 2005; Lyubarsky et al. 1999; Röhlich et al. 1994; Szél et al. 1992; Wang et al. 2011). S opsin is only sensitive to ultraviolet (UV) wavelengths with a peak sensitivity at 360 nm, whereas M opsin has a peak sensitivity at 508 nm (i.e., “green”). Most cones in the mouse coexpress variable amounts of M and S opsin. This M:S-opsin ratio in each cone increases continuously from the dorsal to ventral extremes of the retina, yielding a gradient of wavelength sensitivity whereby upper and lower visual fields are mostly encoded in UV and visible wavelengths, respectively. The spatial distribution of opsin signals sent out of the retina from the retinal ganglion cells (RGCs) has been well quantified (Chang et al. 2013; Wang et al. 2011), which provides a unique opportunity to study how downstream visual areas, such as primary visual cortex (V1) and higher visual areas (HVAs) perform spatial integration from the retina.

Stimulating S opsin with UV light is necessary to study how downstream neurons compare different cone inputs for mouse color vision. Furthermore, behavioral and physiological studies emphasize the importance of stimulating S opsin to effectively drive the cortex from all parts of the visual field. For one, there is more S opsin than M opsin across the entire retina; the S:M-expression ratio is between 3 and 5 (Applebury et al. 2000; Jacobs et al. 2004; Lyubarsky et al. 1999; Wang et al. 2011). Furthermore, cone inputs to RGCs in the ventral portion of the retina appear dominated by the expression of S opsin (Wang et al. 2011). Consistently, psychophysical thresholds are significantly lower in response to UV than visible stimuli in the upper visual fields of mice lacking rod function (Naarendorp et al. 2010). This same study reported similar psychophysical thresholds of cone responses to those in primates, supporting the utility of the mouse model for studying cone signals in the cortex. Next, RGCs are able to follow much faster temporal dynamics to UV modulations than previously thought with stimuli limited to the visible spectrum (Wang et al. 2011). Finally, the first study of UV stimulation with mouse V1 recordings showed robust responses (Tan et al. 2015). However, no study of the mouse’s visual cortex has combined spatiotemporal patterns of UV and visible stimuli to quantify the contributions of S- and M-opsin inputs to the cortex. Traditionally, studies of trichromatic color vision have accomplished the task of isolating cone contributions to downstream visual areas via the method of “color-exchange,” which entails independently modulating the contrast of all three phosphors in a CRT display (Estévez and Spekreijse 1982; Gegenfurtner and Sharpe 2001). Since mice are dichromats, only two display colors are necessary to isolate cone opsin contributions. However, traditional displays do not emit power in the UV spectrum, a requirement to stimulate S opsin in the mouse’s retina. Recent studies have used custom UV and green LED displays to generate color exchange stimuli in the study of opsin inputs to mouse RGCs (Chang et al. 2013; Wang et al. 2011).

Here, we aligned UV and visible LED projectors to test for functional maps in V1 and HVAs that correspond to the dorsoventral gradient of S- and M-opsin input. We first measured retinotopic maps with intrinsic signal imaging (ISI), which served two purposes. First, it allowed us to parse visual areas (Garrett et al. 2014) and extend the study of S and M inputs in cortex beyond V1 into four HVAs (“AL,” “LM,” “PM,” and “RL”). Second, we used these retinotopic maps to register each cortical area to the cardinal axes of the retinal mosaic. After retinotopy, we measured responses to grating stimuli at multiple locations in the S- and M-color plane, which allowed us to parameterize the amount of S- and M-opsin input to each location of the cortex. The analysis produced a clear functional map in the mouse visual cortex. As predicted from the studies in the retina, there is a gradient of S:M-input ratio along the vertical retinotopy, whereby the upper visual field cortical neurons are driven mostly by S-opsin contrast, and the lower visual field neurons are driven mostly by M-opsin contrast. This illustrates that parallel pathways of wavelength sensitivity are at least partially maintained in V1 and HVAs, highlighting its importance in guiding the animal’s behavior. Also, just as ISI allows retinotopic mapping for follow-up targeting of injections or recordings to specific map locations and visual areas (Garrett et al. 2014), the opsin maps from ISI allow for the additional power of targeting functionally biased regions of the cortex.

METHODS

Animal preparation and surgery.

All experiments were approved by the University of Texas at Austin’s Institutional Animal Care and Use Committee. Twelve C57BL/6 mice (of either sex) between 2 and 4 mo of age were anesthetized with isoflurane (2% induction, 1–1.5% surgery) and implanted with a metal frame over the visual cortex for head fixation. Mice were allowed to recover for 1–2 days after head frame implantation. For intrinsic imaging experiments, mice were again anesthetized with isoflurane (2% induction) and head fixed in a custom holder. Chlorprothixene (1.25 mg/kg) was administered intramuscularly, and isoflurane was reduced to 0.25–0.8% (typically 0.5%) during visual stimulation. Anesthesia levels were titrated to maintain a lightly anesthetized state. Imaging was performed through a thinned skull or implanted window. There were no observable signal differences between the two windowing methods. The stimulated eye was first dilated with a drop of tropicamide (1%), followed by application of a thin layer of silicone oil or Stye ointment, which allows optical transmission of near UV (nUV) and visible wavelengths.

Measuring retinotopic maps and identifying borders.

A periodic drifting bar, modulated by a contrast-reversing checkerboard, was presented in the four cardinal directions. The temporal phase of the ISI response to each direction, at the drift frequency, was then used to compute the retinotopic position for each pixel (Kalatsky and Stryker 2003). The size and speed of the drifting bar at each screen location were modified to put vertical and horizontal retinotopy in units of “altitude” (in degrees), and “azimuth” (in degress), respectively (Marshel et al. 2011). Altitude and azimuth isocontours are orthogonal at all retinotopic locations. The speed of the bar was 5.1°/s and its width was 9.4°. The bar was modulated with a contrast-reversing checkerboard at a rate of 2 Hz, with 5° squares. Each drift direction was shown in 10 trials, for 120 s each. Finally, visual area borders were identified from the retinotopy using the methods described in Garrett et al. (2014) (“see Criteria 1–3”).

Intrinsic signal imaging setup.

A Dalsa Pantera 1M60 camera was attached to a Navitar Zoom 6000 lens configuration that filled the CCD sensor. The camera was connected to a Matrox Solios framegrabber via Camera Link. To trigger frame grabs at the beginning of each trial, a TTL pulse was sent from the visual stimulus computer to the Matrox board via a USB-to-TTL device (Measurement Computing 1208FS). Acquisition code for the Matrox board was written in Matlab, using the Image Acquisition Toolbox. Illumination was generated with an X-cite 110LED and a dual branch light guide. At each end of the light guide was a collimator and a long pass (590 nm) colored glass filter. Attached to the tip of the Navitar lens was a bandpass lens with a peak at 650 nm and a width of 50 nm.

Rear projection in nUV and visible spectra.

Two LED projectors (Texas Instruments Lightcrafter 4500) with spectral peaks at 405 nm (nUV) and 525 nm (Green) were used to sample S- and M-opsin contrast space. Optics in the nUV LED projector were modified with UV coatings for efficient transmission (EKB Technology). The rear projection screen consisted of a plate of borosilicate glass and rear projection film (0.8 gain Black Diamond Rear-Pro Film by Screen Innovations). There are technical advantages to using a nUV LED (405 nm), as opposed to an LED that emits power concentrated closer to the peak of the S-opsin sensitivity profile (e.g., at 360 nm). For one, calibrating the nUV LED projector can be achieved using most standard spectroradiometers. For example, the spectroradiometer that we used, a PhotoResearch PR655, captures the entire output of our nUV LED (Fig. 1E). Furthermore, most commercial rear-projection materials begin to attenuate wavelengths below ~400 nm. To account for the relatively small spectral overlap between the nUV LED output and the S-opsin sensitivity function, a strong nUV LED was used that produces ~1 W of total power at the output of the projector. The peak spectral radiance of the backprojected nUV LED projector (i.e., at 405 nm), for maximum brightness, was 0.032 W·steradian⁻¹·m⁻². For reference, the backprojected Green LED had a 0.00166 W·steradian⁻¹·m⁻² at the peak wavelength, which produced a photopic luminance of 76 candelas/m².

Fig. 1.

Overview of visual stimuli for mapping retinotopy and color in the mouse. A: 2 projectors, 1 with a near UV (nUV) LED and 1 with a Green LED, were spatially aligned onto a rear projection surface. B: periodic drifting bar was presented in 4 directions for 120 s per trial with intertrial blank of 4 s. The bars drifting in vertical (up/down) directions were transformed to encode “altitude” coordinates, as a function of temporal phase. Similarly, the horizontal (left/right) bars were transformed to encode “azimuth” coordinates, as a function of temporal phase. C: drifting sinewave gratings were shown by the 2 projectors for 2 s on each trial, with intertrial blank of 3.5 s. The gratings varied in direction and contrast; however, only contrast differed between the two projected images in a trial to stimulate different locations of S- and M-opsin space. D: projected image of the nUV and Green projectors had spectral radiance with peaks near 405 and 508 nm. The spectral sensitivity functions of the S and M opsins are also shown within the calibration range of our spectroradiometer (>380 nm). E: radiance and sensitivity functions were used to transform %Green and %nUV LED contrast to 21 points in %S and %M opsin contrast space.

Next, aligning the projectors to result in overlapping images consisted of two steps. First, the projectors were independently mounted on articulating arms and roughly centered on the rear-projection film. Next, an affine transformation was performed on the images produced by the UV projector to register them with the images produced by the Green projector. In summary, this procedure entailed first displaying a grid of points by the two projectors, where the points were defined at the same pixel locations in software. The Green and nUV points were not in register at this stage because physical alignment with the articulating arms was not possible. Next, the location of N points within the Green grid was identified relative to the nUV grid. From this series of N sample points, we used linear regression to identify the affine transformation, which then aligned the two grid patterns and our subsequent visual stimuli. The two projectors were connected to a single video port of a Macintosh with a splitter. Stimuli were coded using the Psychophysics Toolbox extensions for Matlab (Brainard 1997; Pelli 1997).

The mouse was positioned so that its left eye level was vertically centered on the projection screen, and the perpendicular bisector from the mouse's eye to the screen was 10 cm. Also, the screen was angled 30° from the mouse’s midline. The screen size was ~32 × 51 cm, horizontally and vertically. This resulted in vertical visual field coverage of approximately ± 68° and rostral to lateral coverage from 0 to 115°.

Measuring S- vs. M-opsin maps.

The two LED projectors simultaneously presented full-field drifting sinewave gratings at variable contrasts. To keep spatiotemporal frequency constant across a flat display, gratings were transformed to altitude coordinates (Marshel et al. 2011). The spatial and temporal frequencies were kept constant across trials at 0.05 and 1.5 cycle/°, which are within the bandwidth of most cells in mouse V1 (Gao et al. 2010; Niell and Stryker 2008). Drift direction varied across trials: 0, 90, 180, and 270°. The only stimulus variable that differed between the two projectors on a given trial was contrast. Sampling density of nUV and Green contrast space was chosen with the intent of robustly fitting a single parameter to account for the responsiveness along the S and M directions. The nUV projector showed three contrasts: 10, 30, and 100%. For each nUV contrast, the Green projector showed seven contrasts: −100, −50, −25, 0, 25, 50, and 100%. Negative contrast indicates that the phase of the Green sinewaves were 180° from the nUV sinewaves. Gratings were shown for 2 s and average responses were typically taken within a window starting and finishing 1 and 3.5 s after grating onset. We define this average image following stimulus onset, within each trial (i), as γ_i. A uniform screen, which was the same color and intensity as the midpoint of the sinewaves (“gray point”), was shown between trials for 3.5 s, with a luminance of 38 candelas/m². A baseline image, B_i, was computed for each trial by averaging the frames for 0.5 s immediately before the stimulus onset. Next, we computed the average of I_i = (γ_i-B_i)/B_i across all drift directions and repeats to yield a single response image for each of the 21 Green/nUV contrast combinations. Finally, each of the 21 images was smoothed with a two-dimensional (2D) Gaussian (σ = 25 μm).

Next, nUV and Green LED contrast space was converted into S- and M-opsin contrast space. The spectral radiance of the backprojected nUV and Green LEDs, R_UV(λ) and R_G(λ), were measured with a PR655 spectroradiometer. The S- and M-opsin sensitivity functions, w_S(λ) and w_M(λ), were taken from (Govardovskii et al. 2000) (Fig. 1D). The S-opsin contrast (S) and M-opsin contrast (M) were computed for each contrast of the nUV (g_UV) and Green (g_G) LED projectors:

$$S\hspace{0.17em}contrast=\frac{{{\displaystyle \sum }}_{\text{λ}}{w}_S\left(\text{λ}\right)[g_{UV}{R}_{UV}\left(\text{λ}\right)+g_G{R}_G\left(\text{λ}\right)]}{{{\displaystyle \sum }}_{\text{λ}}{w}_S\left(\text{λ}\right)[R_{UV}\left(\text{λ}\right)+R_G\left(\text{λ}\right)]}$$

$$M\hspace{0.17em}contrast=\frac{{{\displaystyle \sum }}_{\text{λ}}{w}_M\left(\text{λ}\right)[g_{UV}{R}_{UV}\left(\text{λ}\right)+g_G{R}_G\left(\text{λ}\right)]}{{{\displaystyle \sum }}_{\text{λ}}{w}_M\left(\text{λ}\right)[R_{UV}\left(\text{λ}\right)+R_G\left(\text{λ}\right)]}.$$

The mapping of g_UV and g_G to S and M contrast for each of the 21 combinations is shown in Fig. 1E. DLP projectors do not require gamma correction because their signal intensity varies linearly with the buffer values. This was nonetheless confirmed with the spectroradiometer.

Next, we computed a cortical map of the responsiveness to the S and M directions of the color space. To do this, we used linear regression to identify the response gain along the S- and M-contrast directions, at each pixel. Using linear regression and limiting the equation to two parameters allowed for rapid and robust estimates of relative opsin input at thousands of pixels. Specifically, the equation is defined as I = α|S| + β|M|, where I is the average response, above baseline, of the ISI signal to the given location in the |S| and |M| plane. Maps of α/||I|| and β/||I|| are shown in Fig. 2. ||I|| is the norm of the response across 21 color directions, for each pixel. Finally, we computed a map of 100α/(β + α) to quantify the percent contribution from S opsin (%S), at each cortical location (Fig. 3B).

Fig. 2.

Computing S- and M-opsin maps in a single animal. A: top and bottom are the timecourses from 2 pixels. Each time course is the average across 5 repeats, for a single stimulus condition. Timecourses from 3 of the 21 stimulus conditions are shown. B: average response amplitude (%dF/F) from all 21 locations in the color plane, from the same 2 pixels as in A. The bottom 2 quadrants are a reflection of the top 2 quadrants. C: model fit to the data in B, which yields a linear coefficient for the responsiveness to the %S and %M directions of the color space, which are defined as α and β. The pixel represented at top yields %S = 34%, whereas the pixel represented at bottom yields %S = 60%. D: map of α/|I| (i.e., normalized response magnitude to the S direction of color space) in 5 visual areas. The black-to-violet gradients represent an increase in responsiveness to S contrast. E: same as in D, but for the coefficient β, which is the response magnitude to the M direction of color space. Scale bar is 1 mm.

Fig. 3.

Vertical retinotopy vs. opsin maps in visual cortex. Each row is from a different mouse. A: map of vertical retinotopy. Borders were drawn automatically by combining the horizontal and vertical retinotopy gradients. Each higher visual area (HVA) was subsequently identified by observing its “sign” representation (i.e., mirror or nonmirror) and location relative to V1. Scale bar is 1 mm. B: opsin maps were computed by combining the α- and β-maps as %S = α/(α + β). Violet and green pixels represent greater responsiveness to S and M directions of color space, respectively. C: each panel gives the relationship between the opsin map (y-axis) and the vertical retinotopy (x-axis). To compute the red error bars, each pixel was placed in 1 of 7 bins, each containing the same number of pixels. The error bars represent the median ± SD. For V1 alone, we fit a sigmoidal function to the scatter of data points (gray). The value given in the inset (%/°) is the maximum slope of the sigmoid. For each HVA, we instead fit a line to the scatter of data points (not shown). The slopes are also shown for HVAs. The panel for each HVA gives V1’s sigmoidal fit to illustrate the relative balance of opsin inputs for a given retinotopic location.

Rod saturation.

For the study of relative cone-opsin inputs to cortex, it is important to ensure that we were saturating the rods to remove their contribution, especially in a rod dominated retina such as the mouse (Carter-Dawson and LaVail 1979; Jeon et al. 1998). For this, we estimated the photoisomerizations per rod per second for an intact eye that was presented with the gray point of our drifting grating stimuli. To maximize the isomerization rate and to provide experimental stability, we fully dilated the pupil of the presented eye with a drop of 1% tropicamide. In one mouse, we measured the eye diameter for several hours after application. Pupil dilation held stable at ~2 mm, as expected (Remtulla and Hallett 1985). A simple way to roughly estimate the rod photoisomerization rate in an intact eye is to multiply scotopic Trolands of our gray point by a conversion factor of 181 (Lyubarsky et al. 2004). Scotopic Trolands of our gray point was 330 Td, obtained by multiplying the scotopic luminance, 105 cd/m², by the area of the pupil, 3.14 mm². This gives an estimate of 60 K photoisomerizations·rod⁻¹·s⁻¹, which will place the retina far into a regime that produces physiological and psychophysical responses that are dominated by the cones (Naarendorp et al. 2010; Wang et al. 2011).

RESULTS

Driving the mouse retina with gratings in the S- and M-color plane.

To study how the mouse visual cortex pools cone signals from each location of the retina, we imaged responses in five cortical areas to a UV/visible display. The display consisted of two LED projectors, one nUV and one Green, aligned to produce a single rear-projected image that drives both S and M opsin in the mouse’s retina (Fig. 1A). In each mouse, we first identified retinotopic maps by recording ISI responses to periodic drifting bars in the four cardinal directions, as described previously (Kalatsky and Stryker 2003) (Fig. 1B). Contrast-reversing checkerboard modulation within the boundary of the bar stimulates both S and M opsins of the retina. Then, areas of the visual cortex were segmented using an areal boundary analysis described previously (Garrett et al. 2014). For the second stimulus set, we presented full-field drifting sinewave gratings, and independently varied the contrast of the two LED projectors (Fig. 1C). We used three nUV LED contrast values (10, 30, and 100%) and seven Green LED contrast values (−100, −50, −25, 0, 25, 50, and 100%), resulting in 21 locations in the S/M-color plane (Fig. 1E). Transformation from LED space to S- and M-opsin space was computed using the spectral radiance of the stimulus, along with spectral sensitivity functions (Govardovskii et al. 2000) (Fig. 1E). To prevent contamination of our cone measurements from rod inputs, pupils were dilated. Our calculations confirm that a dilated pupil saturates the rods for the amount of luminance produced by our adaptation point (see methods). Distance from the origin of the S- and M-opsin plane represents “total opsin contrast,” whereas the angle indicates the relative amount of S- vs. M-opsin contrast. Next, we used ISI responses to all 21 color directions to parameterize the magnitude of S and M input from each location of the retina, for each of the five visual areas.

Computing S- and M-opsin maps.

Using ISI responses to stimuli in the S/M-color plane, we quantified the strength of M- and S-opsin input to each cortical location. For the subject in Fig. 2, we identified V1 and four HVAs (posteromedial area, PM; rosterolateral area, RL; anterolateral area, AL; and lateral medial area, LM) using previously described methods (Garrett et al. 2014). At each pixel, we fit a linear equation to the ISI responses in the S/M-opsin plane: I = α|S| + β|M|, where I is the average response of the ISI signal, S is the S-opsin contrast, and M is the M-opsin contrast (Fig. 2C). The linear weights, α and β, can be used to quantify the strength of response to the S and M directions of the color plane, respectively (Fig. 2D). It should be noted that the fitted parameters only measure the absolute strength of the input along the two cardinal axes. They cannot distinguish between responsiveness in the summing (S + M; first quadrant) and opponent (S − M; second quadrant) directions. To normalize for the overall response strength at each pixel, the weights were divided by the norm of the response across 21 stimulus conditions, |I|. The maps convey the normalized weights, α/|I| and β/|I| (Fig. 2, D and E). The normalization is used to put the value at all pixels on the same unitless scale, independent of response amplitude; it is not used to quantify a direct percentage. Mouse cones mostly express M opsin in the dorsal retina and S opsin in the ventral retina, creating M- and S-expression gradients that progress in opposite directions along the dorsoventral axis. Correspondingly, our cortical maps of S and M input, α/|I| and β/|I|, exhibit gradients in the opposite direction. The opposing gradients are most obvious in V1 yet are also apparent from visual inspection of the HVAs. Next, we directly compared the maps of opsin input to the coordinates of vertical retinotopy.

Quantifying the relationship between S/M-opsin input and vertical retinotopy in 5 visual areas.

If the cone inputs from the retina are conserved topographically, the visual cortex should also exhibit a gradient of S-to-M response selectivity along the axis of vertical retinotopy. In each mouse, we characterized the relationship between vertical retinotopy and opsin input (Fig. 3, A and B). We first converted the M- and S-opsin input maps (i.e., α/|I| and β/|I|) into a single map expressed as %S = 100α/(α + β), which is the percentage of S input relative to the total M + S cone contribution (Fig. 3B). Qualitatively, it is clear in the four examples that there is a %S map in V1 and HVAs of the mouse. In particular, the %S maps in V1 progress from <50% to >50%, consistent with the retina’s profile. Furthermore, the gradient of %S can be seen to progress in the same direction as vertical retinotopy; i.e., upper visual fields have greater %S than lower visual fields. This was quantified from the joint distribution of vertical retinotopy and %S for all the pixels within each visual area (Fig. 3C). We observed strong, positive correlations between %S and vertical retinotopy in all five visual areas studied (P < 10⁻¹⁰ for all examples shown in Fig. 3C).

Next, we quantified the rate of %S change along the vertical retinotopy. First, we fit a line to the scatter plot of %S vs. vertical retinotopy in each animal and visual area. The median slopes (%S per degree of visual field), across animals, were 1.38%/° (V1; n = 12), 1.47%/° (AL; n = 10), 1.25%/° (LM; n = 12), 1.34%/° (PM; n = 8), and 1.17%/° (RL; n= 4). However, the V1 scatter plots of %S vs. retinotopy generally appeared more sigmoidal than linear, so we also fit the following function to V1: %S = A/[1 + e^{−C*(Vertical-D)}] + B. The variance of the residual error in the fit (i.e., variance of data minus fit, across pixels) was generally much higher for the linear fit; this difference ranged from 5% (“most linear”; Fig. 4, subject in fourth row) to 318% (“most sigmoidal”; Fig. 4, subject in top row). The median reduction, across 12 animals, in residual variance from the sigmoid over the line was 83%. In summary, the sigmoid was a better model than a line for the V1 data. A sigmoidal function is more consistent with the relatively sharp transition zone of S:M expression in the center of the retina (Haverkamp et al. 2005; Wang et al. 2011). The median slope of the sigmoid (at the maximum derivative), across 12 animals, was 1.8%/°.

Fig. 4.

Population summary. Each panel shows the relationship between opsin input and vertical retinotopy, for a single visual area, across multiple animals. A: for V1, the scatter plots of %S vs. vertical retinotopy were combined across 12 animals. The median (μ) of both %S and vertical retinotopy, in each animal, were subtracted before pooling them together to produce the density plot shown. To compute the red error bars, each pixel was placed in 1 of 10 bins, each containing the same number of pixels. The error bars represent the median ± SD. For V1 alone, we fit a sigmoidal function to the scatter of data points (blue). The value given in the inset (%/°) is the maximum slope of the sigmoid. B–E: just as for V1, the scatter plots of %S vs. vertical retinotopy were combined across animals. The animal yield was different for each area. For each HVA, we fit a line to the scatter of data point (not shown). The slopes are shown as an inset in A–E.

We also quantified the range of %S in V1 from the lower to upper extreme of the vertical retinotopy. We first used the sigmoidal fits for this quantification. The median value of the %S sigmoidal fits at −45 and +45° was 18 and 86%, respectively; 18 and 86% correspond to an S:M-input ratio of 18:82 and 86:14, i.e., a 28-fold change of the S:M-input ratio across the extent of the map. The median difference between %S at −45° and +45°, based on the sigmoidal fits, was 57%. Next, we quantified the minimum and maximum of %S nonparametrically, using the first and last bin in the plots of %S vs. vertical retinotopy (Fig. 3C, red, “V1”). The median %S of the first bin (lower visual field) was 23%, and the median of the last bin (upper visual field) was 63%, which corresponds to a 7.8-fold change in the S:M-input ratio. The median difference between the first and last bin was 40%. Since there are generally fewer pixels near −45 or +45°, the first and last bins are between −45 and +45° (Fig. 3C). In turn, it is expected that the sigmoidal fits yielded a greater range in %S across the V1 map.

Finally, the pixels were pooled across animals to make a single scatter plot for each of the five visual areas. Before pooling the data, the opsin map for each area was normalized by subtracting the median of %S. The combined data show that there is a highly significant trend associated with the five visual areas, whereby upper visual fields represent S opsin and lower visual fields represent M opsin (P < 10⁻¹⁰ in each). However, the sigmoidal fit of %S to the combined data has a shallower amplitude from lower to upper visual fields than is apparent in the statistics across individual animals. This is likely a consequence of trying to combine multiple scatter plots that do not exhibit the exact same shape.

In summary, the retinotopic gradient of S:M expression is clearly maintained at the level of V1 and HVAs. In V1, the ISI data shows a relatively sharp transition zone of opsin input at the center of the retinotopy, yet this sigmoidal relationship does not saturate to 0 or 100% at either extreme, which indicates significant mixing of S- and M-opsin input at all locations of V1. However, precise quantification of the signal mixing at the level of individual neurons will require single-unit measurements.

DISCUSSION

Functional organization of color in mouse V1 and HVAs.

In this study, we performed the first characterization of mouse visual cortical responses to stimuli in the S- and M-opsin contrast plane. Our study was built on previous results in the mouse retina that identify spatial variance in the cone mosaic, particularly the dorsoventral gradient of S:M-opsin expression (Applebury et al. 2000; Calderone and Jacobs 1995; Chang et al. 2013; Ekesten and Gouras 2005; Szél et al. 1992; Wang et al. 2011). Given ISI’s ability to detect retinotopic gradients in V1 and HVAs, ISI was deemed fitting to test the hypothesis that there are continuous maps of opsin input in V1 and HVAs that align with maps of vertical retinotopy. We first acquired retinotopic maps to identify visual area borders (Garrett et al. 2014) and to place our subsequent opsin input maps in the context of retinal anatomy. Next, we combined nUV and Green gratings in the classic paradigm of color exchange (Estévez and Spekreijse 1982; Gegenfurtner and Sharpe 2001) to isolate multiple points of the S- and M-opsin plane. By mapping both retinotopy and color, we were able to clearly demonstrate that the opsin gradient, which encodes different wavelengths in the retina, is maintained across multiple subsequent stages of processing. The mouse visual cortex preserves the organization of the retina’s color gradient, which yields a feature map in mouse V1 and HVAs.

Comparison between opsin maps in the retina and visual cortex.

Consistent with RGC responses (Wang et al. 2011) and cone anatomy (Haverkamp et al. 2005), we have shown a sigmoidal relationship between vertical retinotopy and the balance of S- and M-opsin input (%S) in V1. However, unlike the responses from RGCs we do not see a virtual absence of M-opsin input to the upper visual field representation of V1 (Wang et al. 2011). Rather, there appears to be a mixture of S and M input in the upper visual fields. This may be partly due to retino-cortical divergence; V1 neurons pool from a larger area of the photoreceptor mosaic than RGCs, thus integrating more M-opsin signal into the upper visual fields. Another reason may be that the ISI resolution is not adequate for completely parsing responses in the upper visual fields from the rest of the cortex; higher resolution recordings may be required to isolate signals from a “pure” S-opsin input zone lying along the dorsal edge of V1.

We found a more linear relationship between vertical retinotopy and %S in HVAs than in V1. There are a few possible reasons for this difference. First, it may simply be due to the more limited range of the visual field encoded by the HVAs (Garrett et al. 2014). In other words, a window onto a sigmoid function will look more linear. Next, there may be a physiological component leading to the difference, whereby V1-to-HVA divergence is smearing S and M signals to make the sigmoidal relationship more linear in HVAs. Divergence would also help to explain the result that slopes of %S vs. vertical retinotopy are shallower in HVAs than in V1 (Figs. 3 and 4). Lastly, an artifact from insufficient spatial resolution of ISI may inflate the linearity between %S and retinotopy in HVAs. For instance, the HVAs are smaller with steeper map gradients that are more difficult to resolve, causing an otherwise sigmoidal relationship to become linear from spatial smearing.

Quantifying maps of opsin input: suitability and limitations of the approach.

Our experimental and analytical approach stems from our goal of quantifying the S:M-input ratio across V1 and HVA retinotopy with ISI. Because ISI produces low signal-to-noise ratios, we chose to show the maximum allowable contrast along each direction of the S/M-stimulus plane: note that the parallelogram in Fig. 1E outlines the boundary of possible stimuli by our two LED projectors. For a computationally efficient, yet robust measure of the response gain along each axis of S- and M-contrast space, for thousands of pixels, we fit the ISI responses to an equation that allows 2D linear regression: I = α|S| + β|M|. However, this equation cannot distinguish between responsiveness in the summing (S + M; first quadrant) and opponent (S − M; second quadrant) directions of opsin space. We also attempted fitting the data to a different model (not shown), which allowed for parameterization of opsin opponency vs. summation. However, this additional degree of freedom gave fits that were deemed unreliable due to the asymmetry of sampling in the S/M plane. More specifically, we could not rule out that the threefold higher contrast and number of sample points in the S + M quadrant resulted in biased estimates of the preferred color direction.

The sampling density of our visual stimulus in the S/M plane was deemed suitable for fitting a linear equation (i.e., a plane), yet it is likely too sparse to fit nonlinearities that are required of a more biologically realistic model. Regardless, it seems unlikely that including additional nonlinear parameters to account for ISI responses would have significantly altered the measured balance of S- and M-opsin input. Rather, we intuited that it was more likely to add noise to the measurement. In summary, we maximized stimulus contrast and limited the parameterization of our equation to create robust estimates of absolute opsin input across a large expanse of visual cortex.

Color-luminance maps in the visual cortex.

In the primate, addition and subtraction of cone signals is thought to underlie luminance and color perception, respectively. In the mouse, along with other mammals that coexpress different opsins in each cone, an observed “summation” between opsins can arise from either coexpression or downstream addition of separate photoreceptors. On the other hand, opsin subtraction in the mouse must arise from the subtraction of different photoreceptors, as in the primate. Subtraction thus requires a photoreceptor mosaic with local inhomogeneity in opsin expression. That is, some portions of the mouse retina exhibit more inhomogeneity than others, which may be the basis for a map of color tuning. Indeed, recordings of RGCs indicate more opponency (i.e., cone subtraction) at certain locations in the retina. For one, it was shown that both direction selective and “alpha-like” RGCs can be S:M-color opponent by utilizing the steep S:M-opsin gradient near the center of the retina (Chang et al. 2013). As they mention, color opponency could similarly arise from random pooling of pure S-opsin cones and coexpressing (mostly M opsin) cones that are intermixed in the dorsal retina. A more recent study showed opponency between rods and cones by “JAMB” RGCs (Joesch and Meister 2016). More specifically, the peak wavelength sensitivity of JAMB OFF receptive field centers follow the dorsoventral S:M-opsin gradient, whereas their ON surrounds are selective for green, via rod input, across the entire retina. In turn, ventral retina JAMB cells are color opponent with UV-OFF centers and Green-ON surrounds, whereas JAMB cells in the dorsal retina are luminance preferring with Green-OFF centers and Green-ON surrounds. For the range of light intensities that activates both the center and surround of JAMB RGCs, they may also contribute to color-luminance maps in visual cortex.

Spatiotemporal frequency maps in mouse visual cortex.

In addition to the anisotropic distribution of photoreceptors, there are other examples of graded anatomy or function in the retina that may be recapitulated as a feature map in the cortex. For instance, RGC responses are able to follow much faster stimulus dynamics in the UV than visible wavelengths (Wang et al. 2011). Given our result that opsin input is conserved retinotopically in cortical areas, it seems plausible that there are also maps of temporal frequency, whereby upper visual field neurons encode faster kinetics than lower visual field neurons. Next, two lines of evidence suggest that there could be maps of spatial frequency along the axis of horizontal retinotopy. For one, it was shown that a subpopulation of Alpha-On RGCs that are thought to encode higher spatial acuity exhibit greater sampling density in the temporal retina (i.e., representing frontal visual field) (Bleckert et al. 2014). Second, it was shown that there is an increase in the cortical magnification factor (mm/°) at the frontal visual fields of V1 (Schuett et al. 2002; Wagor et al. 1980) and multiple HVAs (Garrett et al. 2014). Greater magnification factor means that more neurons are dedicated to each point of visual space, thus increasing the allowable spatial frequency bandwidth and maximum spatial frequency. In summary, there is evidence to suggest a spatial frequency map whereby frontal visual fields encode higher spatial frequencies. However, a recent study came to a different conclusion based on recordings in the upper-frontal visual field of the retinotopy using visible wavelength stimuli: there was no significant correlation between spatial frequency tuning (cutoff or preference) and horizontal retinotopy, yet there was a positive correlation between the cutoff spatial frequency and vertical retinotopy (Zhang et al. 2015).

The cortical opsin input maps we have shown are continuous and progress monotonically along an axis of the visual field, similar to retinotopy. It is worth noting that this type of map is clearly distinct from the columnar organization seen in visual cortex of larger mammals (Hubel and Wiesel 1962, 1968) and not rodents (Ohki et al. 2005). Whereas functional maps in primate V1 are locally periodic and organized to encode all features for each point of visual space (Nauhaus et al. 2016), this is clearly not the case for the mouse opsin maps observed here; the population average at each location of the retinotopic maps is tuned for a range of wavelengths that is narrower than the full range encoded across the entire population. However, single unit studies are still required to quantify the local scatter and degree of wavelength selectivity at each retinotopic location.

Opsin maps: genetic targeting of a tuned population.

Combined with the genetic advantages of using mice to study cortical circuits, the finding that opsin maps are spatially organized in mouse V1 and HVAs provides a useful experimental paradigm for targeting functional compartments. A common experimental goal of systems neuroscience is to understand the circuits leading to a neuron’s response properties by characterizing or manipulating its inputs. Rather than having a local distribution of tuning preference that is roughly uniform, as is the case with orientation (Ohki et al. 2005; Ringach et al. 2016), opsin preference, and possibly color opponency (see above), are locally biased in the cortical space. In turn, viral injections may be targeted to populations with similar chromatic tuning properties. Of course, local scatter in the maps will broaden the functional specificity of this targeting method. Future studies that measure the statistics of opsin maps at the single cell level will allow for modeling the functional specificity of expression for a given injection volume.

GRANTS

This work was supported by the University of Texas at Austin and a grant from the Whitehall Foundation.

DISCLOSURES

No conflicts of interest, financial or otherwise, are declared by the authors.

AUTHOR CONTRIBUTIONS

I.R. and I.N. conceived and designed research; I.R., G.C.-R., and H.-K.K. performed experiments; I.R. and I.N. analyzed data; I.R., G.C.-R., H.-K.K., and I.N. interpreted results of experiments; I.R. and I.N. prepared figures; I.R. and I.N. drafted manuscript; I.R., G.C.-R., H.-K.K., and I.N. edited and revised manuscript; I.R., G.C.-R., H.-K.K., and I.N. approved final version of manuscript.