Skip to main content
. 2024 Oct 4;13:e86860. doi: 10.7554/eLife.86860

Figure 7. Chromatic contrast tuning allows detection of ground-to-sky transitions.

(a) Distribution of green and UV contrasts of all movie inter-clip transitions (centre), separately for the four transition types, for each of which an example is shown: ground-to-sky (N=525, top left, red triangle), ground-to-ground (N=494, top right, green disk), sky-to-ground (N=480, bottom left, black downward triangle), and sky-to-sky (N=499, bottom right, purple square). Images show last and first frame of pre- and post-transition clip, respectively. Traces show mean full-field luminance of green and UV channels in last and first 1 s of pre- and post-transition clip. Black trace shows luminance averaged across colour channels. (b) Distributions as in (a), but shown as contours indicating isodensity lines of inter-clip transitions in chromatic contrast space. Density of inter-clip transitions was estimated separately for each type of transition from histograms within 10 × 10 bins that were equally spaced within the coloured boxes. Four levels of isodensity for each transition type shown, with density levels at 20% (outermost contour, strongest saturation), 40%, 60%, and 80% (innermost contour, weakest saturation) of the maximum density observed per transition: 28 sky-to-ground (black), 75 ground-to-ground (green), 42 sky-to-sky (purple), and 45 ground-to-sky (red) transitions per bin. Orange markers indicate locations of N=36 G28 maximally exciting inputs (MEIs) in chromatic contrast space (Figure 3i). (c) Tuning map of G28 retinal ganglion cells (RGCs) (N=78), created by averaging the tuning maps of the individual RGCs, overlaid with outermost contour lines from (b) (Figure 7—figure supplement 2). (d, e) Same as (c) for G21 ((g), N=97) and G5 ((h), N=33). (f) Top: Illustration of receiver operating characteristic (ROC) analysis for two RGCs, a G21 (left) and a G28 (right). For each RGC, responses to all inter-clip transitions were binned, separately for ground-to-sky (red) and all other transitions (grey). Middle: Sliding a threshold d across the response range, classifying all transitions with response >d as ground-to-sky, and registering the false-positive rate (FPR) and true-positive rate (TPR) for each threshold yields an ROC curve. Numbers in brackets indicate (FPR, TPR) at the threshold indicated by vertical line in histograms. Bottom: Performance for each cell, quantified as area under the ROC curve (AUC), plotted as distribution across AUC values for all cells (black), G21 (grey), G5 (blue), and G28 (orange); AUC mean ± SD indicated as dots and horizontal lines above histograms. (g) Boxplot of AUC distributions per cell type. Boxes extend from first quartile (Q1) to third quartile (Q3) of the data; line within a box indicates median, whiskers extend to the most extreme points still within [Q11.5×IQR, Q3+1.5×IQR], IQR = inter-quartile range. Diamonds indicate points outside this range. All plot elements (upper and lower boundaries of the box, median line, whiskers, diamonds) correspond to actual observations in the data. Numbers of RGCs for each type are indicated in the plot. (h) Illustration of stimulus with transitions with (Sky-Ground, Ground-Sky) and without (Sky-Sky, Ground-Ground) context change at different velocities (50, 150, 250, and 350 °/s) used in in silico experiments in (i, j). (i) Like (g) but for model cells and stimuli illustrated in (h) at 50/s (see (h)). (j) AUC as function of transition velocity for G28 (orange) vs. example RGC groups (‘Uncertain', G31,32; Slow-ON, G21,23,24; Fast-ON, G18,20; ON-OFF, G10; OFF, G1,5).

Figure 7.

Figure 7—figure supplement 1. Example response traces to inter-clip transitions with and without context changes.

Figure 7—figure supplement 1.

(a) Traces of example cells of different cell groups (G10, G18, G23, G24, G28) from a single recording field, responding to 33 (of 122) inter-clip transitions. Inter-clip transitions are colour-coded by transition type (red: ground-to-sky, purple: sky-to-sky, green: ground-to-ground, black: sky-to-ground). (b) The resulting tuning maps in chromatic contrast space.
Figure 7—figure supplement 2. Chromatic contrast tuning in the dorsal retina allows detection of ground-to-sky transitions.

Figure 7—figure supplement 2.

(a) Illustration of a flat-mounted retina, with recording fields in the dorsal (black circles) and ventral (white circles) retina (cross marks optic disc; d, dorsal; v, ventral; t, temporal; n, nasal). (b) Left: Distribution of green and UV contrasts of N=122 inter-clip transitions seen by a ventral group 28 (G28) RGC, coloured by transition type (red triangle, ground-to-sky; green disk, ground-to-ground; black downward triangle, sky-to-ground; purple square, sky-to-sky). Middle: Responses of example RGC in the 1 s following an inter-clip transition, averaged across transitions within the bins indicated by the grid. Right: Responses transformed into a tuning map by averaging within bins as defined by grid. Left: Responses are z-scored (μ=0, σ=1). (c) Like (b) but for a dorsal G28 RGC. (d) Tuning map of N=9 dorsal G28 RGCs, created by averaging the tuning maps of the individual RGCs. (e) Same as (d) for N=13 G21 RGCs. (f) Same as (d) for N=4 G5 RGCs. (g) Top: Illustration of receiver operating characteristic (ROC) analysis for two dorsal RGCs, a G21 (left) and a G28 (right). For each RGC, responses were binned to all inter-clip transitions, separately for ground-to-sky (red) and all other transitions (grey). Middle: Sliding a threshold d across the response range, classifying all transitions with response >d as ground-to-sky, and registering the false-positive rate (FPR) and true-positive rate (TPR) for each threshold yields an ROC curve (middle). Numbers in brackets indicate (FPR, TPR) at the threshold indicated by black vertical line in histogram plots. Bottom: We evaluated performance for each cell as the area under the ROC curve (AUC), and plotted the distribution across AUC values for all cells (black), for G5 (blue), for G21 (grey), and for G28 (orange). Among the dorsal RGCs, G28 RGCs achieved the highest AUC on average (mean±SD AUC, G28 (N=9 cells): 0.62±0.07; all other groups (N=720): 0.49±0.09, ΔAUC=0.13, bootstrapped 95% confidence interval,CI95=[0.08,0.18] Cohen’s d=1.45 two-sample permutation test G28 vs. all other groups (see Methods): p=0 with 100,000 permutations; next-best performing G24 (N=6): 0.54±0.12, ΔAUC=0.08 bootstrapped 95% confidence interval, CI95=[0.01,0.18] Cohen’s d=0.87 two-sided t-test G28 vs. G24: p=0.15 with 100,000 permutations (not significant). AUC mean ± SD indicated as dots and horizontal lines above histograms. (h) Boxplot of AUC distributions per cell type (dorsal). The box extends from the first quartile (Q1) to the third quartile (Q3) of the data; the line within a box indicates the median. The whiskers extend to the most extreme points still within [Q11.5×IQR, Q3+1.5×IQR], IQR =inter-quartile range. Diamonds indicate points outside this range. All elements of the plot (upper and lower boundaries of the box, median line, whiskers, diamonds correspond to actual observations in the data. Numbers of RGCs for each type are indicated in the plot.
Figure 7—figure supplement 3. Simulations predict tSbC cells robustly detect context changes across different speeds.

Figure 7—figure supplement 3.

(a) Illustration transition stimulus paradigm (from Figure 7h). (b) Structure of stimuli for different velocities, using a ground-to-sky transition as an example. (c) Statistics of the area under the receiver operating characteristic (ROC) curve (AUC) for the sky-ground detection task in the simulation for different velocities (G28 vs. the next-best retinal ganglion cell [RGC] group). Columns (from left): mean ± standard deviation of AUC values (top: G28; bottom: the respective best next RGC type); difference in mean AUC and corresponding bootstrapped 95% confidence intervals; Cohen’s d and p-value of a two-sample permutation test with 100,000 repeats. (d) Boxplots of AUC distributions per cell type for the different velocities (plots like in Figure 7g and j).