Locomotion Enhances Neural Encoding of Visual Stimuli in Mouse V1

Abstract

Neurons in mouse primary visual cortex (V1) are selective for particular properties of visual stimuli. Locomotion causes a change in cortical state that leaves their selectivity unchanged but strengthens their responses. Both locomotion and the change in cortical state are thought to be initiated by projections from the mesencephalic locomotor region, the latter through a disinhibitory circuit in V1. By recording simultaneously from a large number of single neurons in alert mice viewing moving gratings, we investigated the relationship between locomotion and the information contained within the neural population. We found that locomotion improved encoding of visual stimuli in V1 by two mechanisms. First, locomotion-induced increases in firing rates enhanced the mutual information between visual stimuli and single neuron responses over a fixed window of time. Second, stimulus discriminability was improved, even for fixed population firing rates, because of a decrease in noise correlations across the population. These two mechanisms contributed differently to improvements in discriminability across cortical layers, with changes in firing rates most important in the upper layers and changes in noise correlations most important in layer V. Together, these changes resulted in a threefold to fivefold reduction in the time needed to precisely encode grating direction and orientation. These results support the hypothesis that cortical state shifts during locomotion to accommodate an increased load on the visual system when mice are moving.SIGNIFICANCE STATEMENT This paper contains three novel findings about the representation of information in neurons within the primary visual cortex of the mouse. First, we show that locomotion reduces by at least a factor of 3 the time needed for information to accumulate in the visual cortex that allows the distinction of different visual stimuli. Second, we show that the effect of locomotion is to increase information in cells of all layers of the visual cortex. Third, we show that the means by which information is enhanced by locomotion differs between the upper layers, where the major effect is the increasing of firing rates, and in layer V, where the major effect is the reduction in noise correlations.

Keywords: cortical state; decorrelation; gain; locomotion; visual cortex.

Figure 1.

Cortical state change during locomotion. a, The activity of 35–73 single neurons was recorded from the primary visual cortex of freely-moving mice. Moving gratings (6 orientations, each moving in one of two possible directions) were displayed for 1.5 s in the visual field contralateral to the recording site. Mouse movement was tracked over the course of the experiment. b, Evoked mean spike count, averaged over all visual stimuli, across behavioral conditions (N = 409 cells in 8 mice, p = 1e-47, Wilcoxon signed-rank test). Gray line indicates unity. c, CSD plot for one mouse overlaid with inferred laminar boundaries (thick gray lines). Distances at left refer to electrode location relative to the center of the array. d, Mean spike counts of cells by layer, averaged across all stimulus presentations. Layer II, N = 112; layer IV, N = 90; layer V, N = 84; layer VI, N = 123. Error bars indicate bootstrapped estimates of SE. *p < 0.05 (Wilcoxon rank-sum test). **p < 0.01 (Wilcoxon rank-sum test). ***p < 1.3e-5 (Wilcoxon rank-sum test). p values were corrected for multiple comparisons using the Holm–Bonferroni method. e, Change in mean spike count during running as a fraction of mean spike count at rest, using data from d. Numbers of samples as in d. Error bars indicate bootstrapped estimates of SE. *p < 0.05 (Wilcoxon rank-sum test). p values were corrected for multiple comparisons using the Holm–Bonferroni method.

Figure 2.

Cortical state affects single-neuron activity. a, Single-cell mutual information, I(S, R), during running and rest (N = 409, p = 8e-31, Wilcoxon signed-rank test). Gray line indicates unity. b, I(S, R) per spike (N = 409, p = 6e-12, Wilcoxon signed-rank test). Gray line indicates unity. c, Average behaviorally dependent I(S, R), within each cortical layer. Layer II, N = 112; layer IV, N = 90; layer V, N = 84; layer VI, N = 123. Error bars indicate bootstrapped estimates of SE. *p = 0.012 (Wilcoxon rank-sum test). **p = 0.002 (Wilcoxon rank-sum test). ***p = 0.0007 (Wilcoxon rank-sum test). p values were corrected for multiple comparisons using the Holm–Bonferroni method. d, Fractional change in mutual information, ΔI(S,R), within each cortical layer during running. Error bars indicate bootstrapped estimates of SE. **p = 0.002 (Wilcoxon rank-sum test). ***p < 0.0005 (Wilcoxon rank-sum test). p values were corrected for multiple comparisons using the Holm–Bonferroni method. e, Relationship between average spike count and stimulus-specific information, SSI(s). Each point indicates the SSI of a single cell to a particular grating movement direction (N = 4908). Blue line indicates fit of linear regression (R² = 0.85, p < 1e-30). f, Schematic of multiplicative (top) and additive (bottom) tuning curve shifts from rest (black) to locomotion (red). g, Sample single-cell tuning curves for evoked responses at rest (black) and during locomotion (red), with values of additive and multiplicative modulation printed above each. Bold represents significant modulation. h, Relationship between additive and multiplicative components of modulation for rapidly spiking cells (N = 409, ρ = −0.315, p = 1.2e-7). Gray points indicate cells with no significant modulation. Colors represent significant modulation for multiplicative (blue), additive (black), or both (magenta) components. Gray lines represent null hypotheses that no modulation occurs. Open circles represent data points outside of plot range. i, Average modulation across cortical layers during locomotion for cells that are significantly modulated. Error bars indicate bootstrapped estimates of SE. j, ΔI(S,R) as a function of multiplicative (left; ρ = 0.58, p = 2.4e-38) and additive (right; ρ = −0.16, p = 0.001) components of modulation (N = 409). Lines as described in h. Blue line indicates fit of linear regression.

Figure 3.

Classification of single-trial neural responses recorded during locomotion is more accurate than of those recorded at rest. a, Error in LOOCV-LDA classification of visual stimulus movement direction and orientation, as a population (All) and within particular layers. Numbers above layer labels indicate the fraction of the total population included in the decoding. Error bars indicate bootstrapped estimates of SE. b, Fractional change in decoding error with behavior (Error_run − Error_still/Error_still). More negative values indicate greater improvement during locomotion. All, All cells; SMG, significant multiplicative gain > 1; SAG, significant additive gain > 0; MG, multiplicative gain > 1; AG, additive gain > 0; ∼, entire population excluding the category specified. Numbers above layer labels indicate the fraction of the total population included. Error bars indicate bootstrapped estimates of SE. Horizontal gray line indicates no change.

Figure 4.

Noise correlations influence population representation of visual stimuli. a, The distribution of population spike counts and the sum of spikes from all neurons on a single trial overlap in the running and rest conditions. b, Classification error for grating movement direction (left) and orientation (right) as a function of population spike count. Error bars indicate bootstrapped estimates of SE. Dashed gray line indicates chance levels of performance. c, Stimulus-independent (noise) pairwise correlations shift with behavior. Error bars indicate bootstrapped estimates of SE. All, All pairs of cells; E-E, pairs of putative excitatory cells; I-I, pairs of putative inhibitory cells; E-I, pairs of one putative excitatory and inhibitory cells. Values below layer labels are number of pairs included in analysis. **Significant change during running (p < 1e-5, Wilcoxon signed-rank test). d, Noise correlations between excitatory cells by cell modulation. Error bars indicate bootstrapped SEM. All, All pairs of excitatory cells; SMG, significant multiplicative gain > 1; SAG, significant additive gain > 0; MG, multiplicative gain > 1; AG, additive gain > 0. *Significant change during running (p < 1e-3, Wilcoxon signed-rank test). **Significant change during running (p < 1e-6, Wilcoxon signed-rank test). Values below labels are number of pairs included in analysis. e, Noise correlations between excitatory cells within a single layer and across layers. Error bars indicate bootstrapped estimates of SE. Values below layer labels are number of pairs included in the analysis. For significant changes during running: *p < 0.02 (Wilcoxon signed-rank test); **p < 0.005 (Wilcoxon signed-rank test).

Figure 5.

Stimulus discriminability depends on firing rates and noise correlations. a, Schematic calculation of d′ measure. Ovals represent distribution of responses of two neurons to two visual stimuli. Black arrow between response distributions indicates difference vector upon which responses are projected, yielding distributions drawn in bottom right. Values for d′ are calculated from these overlapping distributions using Equation 6. b, Discriminability of grating movement direction, calculated on pairs of neighboring stimuli across behavioral state (p = 3e-17, Wilcoxon signed-rank test). Black points indicate d′ for a pair of stimuli (n = 248; 8 mice, 31 per mouse). Gray line indicates unity. c, Decorrelation reduces change in d′ with behavior, and increases overall d′ values. Mean improvement in d′ with correlated data (47%, p = 3e-17, Wilcoxon signed-rank test). Mean improvement in d′ with decorrelated data (31%, p = 1e-12, Wilcoxon signed-rank test). Error bars indicate bootstrapped CIs of the mean. **p < 5e-8, difference between correlated and decorrelated d′ values.

Figure 6.

Stimulus information as a function of time. a, Mutual information between cell spiking and stimulus per 10 ms time bin, averaged across all recorded cells. Error bars indicate bootstrapped estimates of SE. b, Classification error over sequential 100 ms time periods after stimulus onset. Error bars indicate bootstrapped estimates of 95% CIs. c, Classification error over various time ranges after stimulus onset. Error bars indicate bootstrapped estimates of 95% CIs. Shaded bars represent 95% CIs of the mean during the first 100 ms after neural response onset.

Figure 7.

Relationship between running speed and spike counts, population responses, and classification error. a, Residual spike counts as a function of run speed for 15 sample cells after mean visually evoked responses were subtracted. Cells were chosen randomly from the population in a single mouse; responses shown are from run trials. Blue bar represents speed of visual stimulus, 30 cm/s. b, Population spike counts as a function of the natural logarithm of mouse running speed on single running trials (black dots). Red lines indicate fit of linear regression. Panels are individual mice. c, Average LOOCV error with increasing mouse running speed for stimulus orientation (blue) and movement direction (red). Numbers of samples at each mean speed are listed at top of each panel. Error bars indicate bootstrapped estimates of SE.