Awake responses suggest inefficient dense coding in the mouse retina

Abstract

The structure and function of the vertebrate retina have been extensively studied across species with an isolated, ex vivo preparation. Retinal function in vivo, however, remains elusive, especially in awake animals. Here, we performed single-unit extracellular recordings in the optic tract of head-fixed mice to compare the output of awake, anesthetized, and ex vivo retinas. While the visual response properties were overall similar across conditions, we found that awake retinal output had in general (1) faster kinetics with less variability in the response latencies; (2) a larger dynamic range; and (3) higher firing activity, by ~20 Hz on average, for both baseline and visually evoked responses. Our modeling analyses further showed that such awake response patterns convey comparable total information but less efficiently, and allow for a linear population decoder to perform significantly better than the anesthetized or ex vivo responses. These results highlight distinct retinal behavior in awake states, in particular suggesting that the retina employs dense coding in vivo, rather than sparse efficient coding as has been often assumed from ex vivo studies.

Keywords: anesthesia; awake; efficient coding; in vivo recordings; mouse; neuroscience; retinal ganglion cells.

Figure 1.. In vivo extracellular recordings from the mouse optic tract (OT).

(A, B) Schematic diagram of the experimental setup. We presented visual stimuli to a head-fixed mouse using a digital light processing (DLP) device projecting images onto a spherical screen placed laterally to the subject animal (A, front-view; B, side-view). See Methods for details and specifications. Schematic diagram (C, top-view; D, side-view) of the brain and electrode location to target the OT. (E) Histological image of a representative brain sample (coronal section, 150 µm thick) showing the electrode trace (red, DiI stain deposited on the electrode). Spike waveform (F–H; black, individual trials; white, mean) and autocorrelogram (I–K; bin size, 1 ms) for three representative units recorded from the OT of an awake mouse.

Figure 1—figure supplement 1.. Eye motion and behavioral data analysis.

(A) Example eye-tracking frame of an awake mouse eye. Cyan ellipse shows the pupil edge, while magenta dot shows its center location. Representative dynamics of simultaneously recorded data in an awake head-fixed condition (B) the presence and direction of the moving gratings stimulus; (C) firing rate heat map of simultaneously recorded seven retinal ganglion cells (RGCs); (D, E) horizontal and vertical coordinates of the pupil center location in each frame, respectively; (F) pupil size defined as a mean of the major and minor axis of the ellipse fitted to the pupil edge in each frame; and (G, running speed of the mouse). Red shades indicate eye blinks, whereas the blue vertical lines indicate saccades. Note that blinks and saccades occurred only once in a while (0.06 ± 0.03 Hz; mean ± standard deviation, 19 animals); and that the pupil size (F) is positively correlated with the running speed (G) in general (0.34 ± 0.20; 19 animals). Probability density of pupil center location (H), pupil size (I), and running speed (J) across different visual stimulation periods (from left to right, full-field sinusoidally flickering stimuli; randomly flickering full-field stimuli; and moving gratings; 19 animals).

Figure 2.. Physiological classification of retinal output responses in vivo.

(A) Representative retinal output responses to full-field contrast-inverting stimuli: top, stimulus; middle, raster graph over trials; bottom, peri-stimulus time histogram (black, mean; gray, variance; signal-to-noise ratio [SNR], Equation 1 in Methods; ON–OFF index, Equation 2). (B) SNR of the retinal output responses in different recording conditions. We set a threshold at 0.15 to identify reliably responsive cells (black) and low-quality unclassifiable cells (gray). (C) ON–OFF index distributions from the reliably responsive cells. While no apparent clusters were identified, we set a threshold at ±0.25 to categorize the response polarity into ON, ON/OFF, and OFF cells. Within the ON cells, we further identified those with an ‘OFF-suppressive’ response to the full-field flickering stimulus (Figure 3). Distribution of direction selectivity (DS)/orientation selectivity (OS) indices (Equation 3) in each recording condition (D, awake; E, isoflurane; F, fentanyl, medetomidine, and midazolam [FMM]). We set a threshold at 0.15 (with p < 0.2) to identify whether cells are OS/DS (black) or not (gray). (G–I) Fraction of identified response classes in vivo: ON (blue; OFF-suppressive in green), OFF (red), ON/OFF (orange), and the rest unclassifiable cells (‘N/A’, gray). Cells in each category were further divided based on the OS/DS properties (hued). The OFF-suppressive ON cells were prominent in the awake condition (G, 36/282 cells), but rarely observed under anesthesia (H, isoflurane, 2/325 cells; I, FMM, 1/103 cells).

Figure 2—figure supplement 1.. Retinal output was correlated with an animal’s behavioral states.

Peak cross-correlation values between retinal output and the animal’s pupil size, plotted against the p-values calculated by a shuffling analysis (1000 repetitions, see Methods; A) or the ON–OFF index (B): black, correlation with p < 0.05; gray, correlation with p > 0.05. The pupil size effects on retinal ganglion cell (RGC) firing were significantly correlated with the ON–OFF index (B; Pearson’s R = 0.38, p < 0.001). (C, D) Corresponding plots for the peak cross-correlation values between retinal output and the animal’s running speed. The locomotion effects on RGC firing were not significantly correlated with the ON–OFF index (D; Pearson’s R = 0.11, p = 0.3).

Figure 3.. Many ON retinal ganglion cells showed suppressive OFF responses in awake condition.

Mean firing rate of representative cells in response to a sinusoidally flickering stimulus with increasing contrast in the awake (A, OFF; B, OFF-suppressive ON, the same cell as in Figure 2A) or anesthetized conditions (C, ON). Overlaid with the peri-stimulus time histogram (gray) is the model fit (black, Equation 6 in Methods). The number on top indicates the fit quality (explained variance R² in Equation 5). Population data of the model parameters (D, baseline B; E, amplitude A; F, phase ϕ) across different conditions: isoflurane anesthesia (N = 147), fentanyl, medetomidine, and midazolam (FMM) anesthesia (N = 95), and awake (N = 247 in total). The sinusoidal stimulus pattern relative to the response peak is also indicated at the bottom of F. Cell types are color coded as in Figure 2 (blue, ON; green, OFF-suppressive ON; red, OFF; orange, ON/OFF). Note high baseline with negative amplitude and positive phase for the OFF-suppressive ON cells, which were predominantly found in the awake condition: ***p < 0.001; *p < 0.05; ns, nonsignificant (D, U-test; E, U-test on the absolute values; F, t-test).

Figure 4.. Retinal output showed higher temporal frequency sensitivity in awake than in anesthetized mice.

(A) Representative retinal output (gray, mean firing rate over 10 trials) in an awake condition in response to full-field sinusoidally flickering stimuli at different temporal frequencies (1.875, 3.75, 7.5, and 15 Hz), following full-field contrast inversions. Overlaid is the curve fit (Equation 4 in Methods; black). The number on top is the explained variance of the curve fit (R², Equation 5 in Methods), representing the fit quality. (B) Representative retinal output responses under isoflurane anesthesia (shown in the same format as in A). (C) Population data of the fit quality at four different stimulus frequencies in the awake (N = 248) or anesthetized conditions (isoflurane, N = 147; fentanyl, medetomidine, and midazolam [FMM], N = 95), color coded for the responses class as in Figure 2. The fit quality threshold was set to be 0.2 (black, R² ≥ 0.2; gray, R² < 0.2). (D) Fraction of the cells with the fit quality above the threshold across different conditions (awake, black line with circles; isoflurane, gray line with vertical crosses; FMM, gray line with diagonal crosses), representing the frequency tuning of the retinal output at the population level. A significantly larger fraction of cells was responsive at 15 Hz in the awake condition than in the anesthetized conditions (***p < 0.001 for both isoflurane and FMM; two-proportion z-test).

Figure 5.. Retinal ganglion cells showed faster response dynamics in awake condition than in anesthetized or ex vivo conditions.

(A) Temporal filter of a representative awake cell (gray, spike-triggered average [STA] of the full-field randomly flickering stimulus) and a difference-of-Gaussian curve fit (black) for estimating the latency of the first peak. (B) Power spectra of the example filter in A, based on the curve fit, for estimating the peak frequency. Population data of the peak latencies (C) and frequencies (D) across different conditions (light green, biphasic ON; dark green, monophasic ON; pink, biphasic OFF; violet, monophasic OFF). Here and thereafter, ***p < 0.001; **p < 0.01; *p < 0.05; ns, nonsignificant (t-test). The filter types were identified by the quadrants of the principal component analysis (PCA) biplot (see Methods for details). Population data of the temporal filters across different conditions: from top to bottom, isoflurane anesthesia (E, N = 238), fentanyl, medetomidine, and midazolam (FMM) anesthesia (F, N = 69), ex vivo (G, N = 342), and awake (H, N = 201). The four filter types are indicated on the right with corresponding colors.

Figure 6.. Retinal ganglion cells showed higher firing activity in awake condition than in anesthetized or ex vivo conditions.

Cell classes are color coded as in Figure 5 (light green, biphasic ON; dark green, monophasic ON; pink, biphasic OFF; violet, monophasic OFF). (A) Population data of the mean firing rates during the stimulus presentation period in four different recording conditions (isoflurane, fentanyl, medetomidine, and midazolam [FMM], ex vivo, and awake): ***p < 0.001; **p < 0.01; *p < 0.05; ns, nonsignificant (t-test on the logarithm of firing rates). (B) Static nonlinear gain function of a representative awake cell (the same one as in Figure 5A), estimated by the stimulus ensemble statistical techniques applied to the responses to a full-field randomly flickering stimulus (gray, Equation 7 in Methods; black, sigmoid curve fit with the midpoint at 0.71). Note a high neutral stimulus response (40 Hz) defined as the firing rate at zero filter output (i.e., in the presence of stimuli orthogonal to the cell’s spike-triggered average [STA]). (C) Population data of the neutral stimulus responses in each recording condition (in the same format as in A). Population data of the static nonlinear gain function (median for each cell type in corresponding colors; gray, interquartile range of all cells) across different conditions: isoflurane anesthesia (D, N = 238), FMM anesthesia (E, N = 69), ex vivo (F, N = 342), and awake condition (G, N = 201). (H) Population data of the midpoint of the sigmoid nonlinearity in each recording condition (Mann–Whitney U-test).

Figure 6—figure supplement 1.. Batch effects in the data sets.

Dynamic range was computed for each cell as the difference between the maximum and minimum firing rates of the static nonlinear gain function (Figure 6D–G), and shown for each recording session in columns (dark green, monophasic ON; light green, biphasic ON; pink, biphasic OFF; violet, monophasic OFF; gray line, median). Some variabilities were present across preparations in each recording condition (isoflurane, N = 44 recording sessions; fentanyl, medetomidine, and midazolam [FMM], N = 10; awake, N = 34; or ex vivo, N = 18), but the batch effects were not large enough to eliminate any particular data set.

Figure 7.. Retinal responses in vivo have a lower information rate in bits per spike but a higher rate in bits per second than those ex vivo.

Information rate conveyed by retinal output spike trains in bits per second (A) and in bits per spike (B; Equation 8 in Methods) in response to randomly flickering full-field stimuli under different recording conditions (from left to right: isoflurane, N = 231; FMM, N = 67; awake, N = 154; ex vivo, N = 328). Cell classes are color coded as in Figures 5 and 6 (dark green, monophasic ON; light green, biphasic ON; pink, biphasic OFF; violet, monophasic OFF). ***p < 0.001; **p < 0.01 (Mann–Whitney U-test).

Figure 7—figure supplement 1.. Rate coding was more preferable than latency coding for simulated awake retinal responses.

(A–D) Simulated firing rates of a representative awake cell (B) in response to a change in the stimulus contrast at 20 different levels (A, in corresponding colors). The responses were computed from the cell’s temporal filter and static nonlinear gain function (see Figures 5 and 6, respectively). The vertical dotted line in B indicates the peak response latency, from which the conditional probability distribution of the peak firing rate, p(rate|stimulus), was calculated under the assumption that the standard deviation of the Gaussian distribution follows 2.5% of the peak rates (C). The horizontal lines in B indicate ±10% from the baseline firing rate as the detection threshold to compute the conditional probability distribution of the response latency (D), p(latency|stimulus), where the standard deviation of the Gaussian was given by 2.5% of the baseline firing rate (gray shade in B). See Methods for details. (E) Information in peak rate, I_{peak rate}, in bits computed from p(rate|stimulus) as in C for each cell (from left to right: isoflurane, N = 234; FMM, N = 67; awake, N = 195; ex vivo, N = 328). Cell classes are color coded as in 5—7 (dark green, monophasic ON; light green, biphasic ON; pink, biphasic OFF; violet, monophasic OFF). ***p < 0.001; *p < 0.05 (Mann–Whitney U-test). Simulated awake cells have the highest information. (F) Information in response latency, I_latency, computed from p(latency|stimulus) as in D for each cell (shown in the same format as E). Given that both increase (red traces in D) and decrease (blue traces in D) of the firing rates are detectable, simulated awake cells have comparable information to cells ex vivo, but lower information than those under anesthesia.

Figure 8.. Linear decoding of retinal population responses worked best with awake responses.

For each recording session, we performed linear decoding of the simultaneously recorded population activity (333 ms window) in response to the full-field randomly flickering stimulus (16.7 ms bin). As a measure of the decoding performance, we then calculated Pearson’s correlation coefficient R between the presented and estimated stimuli with 10-fold cross-validation (from left to right: isoflurane, N = 44; FMM, N = 10; awake, N = 33; ex vivo, N = 18). The black line shows the exponential curve fit (Equation 10 in Methods; isoflurane, a = 0.23 ± 0.04, b = 23 ± 7, estimate with 95% confidence interval; FMM, a = 0.24 ± 0.06, b = 21 ± 6; awake, a = 0.20 ± 0.13, b = 8 ± 4; ex vivo, a = 0.20 ± 0.10, b = 72 ± 36).

Author response image 1.

Author response image 2.

Author response image 3.

Author response image 4.

Author response image 5.