Joint representations of color and form in mouse visual cortex described by random pooling from rods and cones

Abstract

Spatial transitions in color can aid any visual perception task, and its neural representation, the "integration of color and form," is thought to begin at primary visual cortex (V1). Integration of color and form is untested in mouse V1, yet studies show that the ventral retina provides the necessary substrate from green-sensitive rods and ultraviolet-sensitive cones. Here, we used two-photon imaging in V1 to measure spatial frequency (SF) tuning along four axes of rod and cone contrast space, including luminance and color. We first reveal that V1's sensitivity to color is similar to luminance, yet average SF tuning is significantly shifted lowpass for color. Next, guided by linear models, we used SF tuning along all four color axes to estimate the proportion of neurons that fall into classic models of color opponency, i.e., "single-," "double-," and "non-opponent." Few neurons (∼6%) fit the criteria for double opponency, which are uniquely tuned for chromatic borders. Most of the population can be described as a unimodal distribution ranging from strongly single-opponent to non-opponent. Consistent with recent studies of the rodent and primate retina, our V1 data are well-described by a simple model in which ON and OFF channels to V1 sample the photoreceptor mosaic randomly. Finally, an analysis comparing color opponency to preferred orientation and retinotopy further validates rods, and not cone M-opsin, as opponent with cone S-opsin in the upper visual field.NEW & NOTEWORTHY This study is the first to show that mouse V1 is highly sensitive to UV-green color contrast. Furthermore, it provides a detailed characterization of "color opponency," which is the putative neural basis for color perception. Finally, using an extremely simple yet novel random wiring model, we account for our observations.

Keywords: color; contrast; mouse; spatial frequency; visual cortex.

Graphical abstract

Figure 1.

Models of rod vs. S-cone opponency. A: illustration of the spatial distribution of opsins in the mouse retina. In the cone mosaic on the left, green and violet indicate concentration of M- and S-opsin, respectively. The green disk on the right represents the uniform rhodopsin (“R-opsin”) distribution. B: approximate spectral sensitivity functions of the two major opsins in the ventral retina [Govardovskii et al. (30)], which was the focus of this study. C: the title in each row defines the model of color opponency. The first three are built from the same ON and OFF subfields. D: each row has two linear RFs, one for R-opsin (green) and one for S-opsin (violet). E: taking the Fourier transform of each opsin’s RF in B approximates the sensitivity of the neuron to sinewave gratings with contrast that isolates either the R- or S-opsin axis of color space. F: to obtain the SF tuning to R + S (black) and R−S (red) contrast, we added and subtracted the RFs in D, respectively, and then took the Fourier transform. SF, spatial frequency.

Figure 2.

Two-photon imaging of responses to drifting gratings in the S- and R-opsin contrast plane. A: a map of vertical retinotopy in V1 of the right hemisphere, using intrinsic signal imaging. Blue-to-red is lower-to-upper fields. Subsequent two-photon imaging was targeted to the upper visual field representation. B: raw image from a two-photon imaging session, 900-µm wide. C: the local correlation image, from which 126 cell bodies were manually selected. The total across 16 ROIs using this procedure was 1,360. Bright pixels are highly correlated with the immediate neighborhood. Each small panel surrounding the perimeter shows the circled neuron’s responses to different opsin contrasts, at the neuron’s optimal drift direction and SF, averaged over five repeats. The color indicates the axis of opsin contrast, shown in E. D: upper visual field opsin sensitivities [Govardovskii et al. (30)], overlaid with the spectral power distribution of the visual stimulus. E: illustration of the stimulus set. There were five SFs, spaced 1 octave apart, ranging from 1/80 c/^o to 1/5 c/^o. In a small subset of experiments, a sixth SF was included at 1/2.5 c/^o. Above each column of gratings is a diagram of the corresponding color axis. F: each data point is a neuron that was first manually selected from one of 16 local correlation images like in C, and then shown to produce reliable parameter estimates of SF tuning and retinotopy (see methods), yielding n = 385 in this plot. The y-axis is the relative response to the S- and R-opsin isolating contrast. A value of 1 and 0 indicates that it only responded to R- and S-contrast, respectively. The x-axis is the RF’s position in vertical retinotopy, with positive values being above the midline. The overlaid dashed rectangle outlines the population used for subsequent analyses, which encloses neurons in the upper visual field, and neurons with a balanced R:S response ratio (1:2 < R:S < 2:1), totaling n = 140. This data selection allowed for the study of rod-cone color opponency with gratings that contain contrast in the R- and S-opsin color plane. ROI, region of interest; SF, spatial frequency; V1, primary visual cortex.

Figure 3.

V1’s sensitivity to color and luminance contrast, at each SF. A: population statistics of responsiveness to each contrast. For each neuron, ΔF/F at each SF and contrast was calculated as the average over all presented orientations and repeats. The plot shows statistics of ΔF/F for each contrast (S−R in red, S + R in gray), at each of the presented SFs (x-axis). The three horizontal lines of the box plots are quartiles (q₁, q₂, and q₃), and the whiskers are located at the last data point before the “outliers.” Top outliers are greater than q₃ + 1.5(q₃ − q₁). Bottom outliers are less than q₁ − 1.5(q₃ − q₁). The P value at each SF is from an unpaired t test between the two ΔF/F distributions. B: statistics of ΔF/F at the peak SF, for the two contrasts. Inset shows SF tuning from an example neuron, with its two peaks as filled-in data points. The solid line is a Gaussian fit, which was not used for the analysis in this figure. C: population statistics of relative responsiveness to S − R and S + R contrast. For each neuron, we calculated preference for S − R over S + R contrast, “C” (y-axis), at each SF (x-axis). Population statistics are shown. The P value at each SF is from a t test on C, identifying whether the mean is different from zero. D: population statistics of C_peak, where each data point is taken from the peak of each SF tuning curve, S − R and S + R. To the right of the box and whisker plot is a histogram from the same distribution. E: five examples of SF tuning and the Gaussian fit, for the S − R and S + R contrast. They are calculated by taking the mean over orientations and repeats. SF, spatial frequency.

Figure 4.

Dependencies SF tuning on opsin contrast. A–D: these small panels provide an approximate expectation of the four opponency models described in Fig. 1. The axes in each panel match the scatter plot immediately below in E and F, which compare bandpass factor (BPF) between opsin contrasts. Each circle or disk corresponds to an opponency model from Fig. 1. Open circles are color-preferring and closed circles are luminance-preferring. E: each data point shows a neuron’s bandpass factor for color (BPF_S−R) and luminance (BPF_S+R) taken from the Gaussian fits. The methods for data inclusion (n = 140) were outlined in Fig. 2. Unity line is the diagonal. On the top-left, “Δ” is the mean of BPF_S+R − BPF_S−R, and “p” is from a t test on Δ. On top-right is the Pearson correlation (r) and P value. To the right or left of each quadrant is the ratio of neurons with open (C_peak > 0) and closed (C_peak < 0) data points. F: BPF_S vs. BPF_R. Same statistics as in E are given around the perimeter. G: BPF_S vs. BPF_S+R. H: BPF_S vs. BPF_S−R. I–L: same data and layout as in A and B, but the tuning parameter is the peak location of the fitted Gaussian (cyc/°). I: peak SF for color (peak_S−R) vs. luminance (peak_S+R). Unity line is the diagonal. On the top-left, “Δ” is how much bigger peak_S−R is than peak_S+R, as a %. The P value is from a t test on log(peak_S−R) − log(peak_S+R). On the top-right is the Pearson correlation and P value comparing log(peak_S+R) and log(peak_S−R). J: same as E, but peak_S vs. peak_R. K: peak_S vs. peak_S+R. L: peak_S vs. peak_S−R. D.O., double-opponent; N.O., non-opponent; SF, spatial frequency; S.O.A., single-opponent-A; S.O.B., single-opponent-B.

Figure 5.

Dependencies SF tuning on color sensitivity. A–C: these smaller panels illustrate the approximate location of data points from the four opponency models described in Fig. 1. The axes are the same as the adjacent scatter plots in D–G. D–G: each panel is from the same population (n = 140) as in Figs. 3 and 4, now comparing SF bandpass factor (BPF) of a specified opsin contrast (y-axes) to the preference for color over luminance (C_peak) (x-axis). D: each data point is a neuron’s BPF for color (BPF_S−R) and C_peak. At top left is the Pearson correlation (r) and P value. Inset within each quadrant is the number of neurons. E: BPF for luminance (BPF_S+R) vs. C_peak. Same statistics as D are given. F: BPF_S vs. C_peak. G: BPF_R vs. C_peak. H: example SF tuning curves and Gaussian fits for S−R (red) and S + R (gray) contrast, which are numbered and circled in each scatter, D–G’. I: SF tuning curves and Gaussian fits from the same neurons as in H, but for S (violet) and R (green) contrast. D.O., double-opponent; N.O., non-opponent; SF, spatial frequency; S.O.A., single-opponent-A; S.O.B., single-opponent-B.

Figure 6.

Simulation of results in Fig. 4 with a random pooling model. A: each column shows the results of analyzing one simulated neuron that was generated by the random pooling model. The model can generate cells from three opponency classes: single-opponent-A (left 3), single-opponent-B (middle 3), and non-opponent (right 3) cells. The top shows the spatial S- an R-opsin RF, where the model’s random variables–spatial scale, and the ratio of S- to R-opsin input to each of the three Gaussian subfields–can be more explicitly visualized. The second row shows the Fourier magnitudes of S- and R-opsin RFs, to simulate SF tuning. Similarly, the third row is the Fourier magnitudes of S + R and S−R. Inset are the bandpass factors (BPFs). B: the mean SF tuning for each axis in our opsin contrast space, taken from 244 randomly drawn V1 neurons. C–F: smaller panels along the top illustrate approximate expectations of the four opponency models described in Fig. 1. The axes in each panel match the scatter plot immediately in E and F, which compare bandpass factor (BPF) of different opsin contrasts. Each circle or disk corresponds to an opponency model. Open circles are color-preferring and closed circles are luminance-preferring. G–J: analysis of 244 simulated neurons, using the same procedures that generated Fig. 4, E–H. The titles give the mean difference between BPF on the x-axis and y-axis, and P value of a paired t test. Inset within each quadrant is the ratio of color-preferring (open circles) to luminance-referring (closed circles). D.O., double-opponent; N.O., non-opponent; S.O.A., single-opponent-A; S.O.B., single-opponent-B.

Figure 7.

Simulation of results in Fig. 5 with a random pooling model. A–C: left illustrates the approximate layout of opponency models described in Fig. 1. The axes are the same as the adjacent scatter plots in D–G, where the y-axis is the bandpass factor (BPF) of a specified color contrast, and the x-axis is C_peak. Each circle or disk corresponds to an opponency model: D–G: analysis of 244 simulated neurons (same population as in Fig. 6), using the same procedures that generated Fig. 5, D–G. At top right of each panel is the Pearson correlation (r) and P value. D.O., double-opponent; N.O., non-opponent; S.O.A., single-opponent-A; S.O.B., single-opponent-B.

Figure 8.

A dependence of inseparability on orientation preference predicts marginal opponency from the cone-opsin gradient in the mesopic state. A: the colored disk is a cartoon representation of the cone-opsin mosaic, where green and violet are M- and S-opsin expressions. For the actual simulation, we used a model of this gradient based on retinal ganglion cell output [Wang et al. (39)]. Overlaid at the steepest part of the gradient is a cartoon V1 RF tuned for horizontal orientation (θ_pref = 90^°). The RF is composed of an ON and OFF Gaussian that generates a preferred SF of 0.05 cyc/^°. The green and violet traces below are the 1-D profiles of a simulated S- and M-opsin RF, calculated by multiplying the monochromatic spatial RF by the percentage of S- and M-opsin expressions at each location given by the model in Wang et al. (39). The red and black traces on bottom are the corresponding SF tuning curves for M−S and M + S contrast. This RF has lowpass tuning for M−S contrast and bandpass tuning for M + S contrast, making it inseparable for color and form. B: Same as A but the RF prefers vertical orientation and is positioned more ventral. This RF has the same bandpass tuning for M−S and M + S contrast, making it separable for color and form. C: actual data showing color-form inseparability vs. orientation preference. The y-axis values can be calculated from the scatter plot values in Fig. 4I. The x-axis is preferred orientation, relative to horizontal (90^°). Error bars show the median and SD in three bins, 0^°-to-30^°, 30^°-to-60^°, and 60^°-to-90^°.