Spatial clustering of tuning in mouse primary visual cortex

Abstract

The primary visual cortex of higher mammals is organized into two-dimensional maps, where the preference of cells for stimulus parameters is arranged regularly on the cortical surface. In contrast, the preference of neurons in the rodent appears to be arranged randomly, in what is termed a salt-and-pepper map. Here we revisited the spatial organization of receptive fields in mouse primary visual cortex by measuring the tuning of pyramidal neurons in the joint orientation and spatial frequency domain. We found that the similarity of tuning decreases as a function of cortical distance, revealing a weak but statistically significant spatial clustering. Clustering was also observed across different cortical depths, consistent with a columnar organization. Thus, the mouse visual cortex is not strictly a salt-and-pepper map. At least on a local scale, it resembles a degraded version of the organization seen in higher mammals, hinting at a possible common origin.

Figure 1. Experimental set-up.

(a) Activity of cells in the primary visual cortex was imaged with a two-photon scanning microscope while the mouse observed a continuous visual stimulus on a freely rotating platform. The position of the platform was monitored with an optical rotary encoder. An infrared light-reflective glass (red line) allowed a camera to image the pupil while allowing for unobstructed visual stimulation. The visual stimulus consisted of a sequence of pseudo-random sinusoidal gratings (see Methods for details). (b) Two segments illustrating the process of inferring spiking activity from imaging data. First, the calcium fluorescence corresponding to cell bodies (raw signal) and their immediate neighbourhood (neuropil) as a function of time are extracted after compensating for motion in the imaging plane. Here both the raw signal and the neuropil are normalized by the s.d. of the raw signal. The vertical scale bar corresponds to five times the s.d. of the signals. A potential contamination of the signal by the neuropil is ameliorated by projecting out a robust linear prediction of the signal based on the neuropil. Finally, the probability of spiking is inferred by non-negative deconvolution (see Methods for details). The result is a trace that is nearly identical to zero in regions devoid of spiking activity (red trace), further minimizing small contributions of the neuropil to background activity. The spike inference trace is plotted in arbitrary units.

Figure 2. Tuning similarity depends on cortical distance in mouse V1.

(a) Sample of estimates of the joint tuning of neurons in the orientation and spatial frequency plane in four cells. The origin is in the middle of the image. Kernels are individually normalized, so the maximum absolute value is one. (b) Average tuning similarity decreases as a function of cortical distance. Error bars represent ±1 s.e.m. The size of the data points denotes the significance of a rank-sum test comparing the median distribution of data at a given distance to the distribution of the rightmost bin near 200 μm. Red, dashed line shows the best exponential fit to the data. Inset: the mean tuning profile in mouse V1. The number of cell pairs in each group, in order of increasing cortical distance, are n=894, 2,538, 3,473, 3,845, 3,792, 3,399 and 2,964. (c) Demonstration of clustering within a single imaging field. Top, segmented cells. Middle, estimated kernels. Bottom, scatter plot of tuning similarity versus cortical distance. There is a statistically significant negative correlation between receptive field similarity and cortical distance. (d) Distributions of the absolute difference in preferred orientation as a function of cortical distance for cells within 50 μm of each other (left panel) and at least 150 μm away from each other (right panel) for the data in (b).

Figure 3. Tuning similarity as a function of cortical distance and depth.

(a) The tuning similarity of cell pairs within or across different cortical planes was compared with each other. The cortical distance between two cells, d, is defined as the distance between their projections on the cortical surface. Depth difference, Δ, is defined as the difference in depth between the imaging planes. (b) Average dependence of tuning similarity as a function of cortical distance and cortical depth. For each individual curve, the size of the data points denotes the significance of a rank-sum test comparing the median distribution of data at a given distance to the distribution of the rightmost bin near 200 μm.

Figure 4. Inhibitory and excitatory calcium signals have different distributions.

(a) Two-photon image of a preparation where PV cells are genetically labelled with tdTomato. (b) Sample traces for PV+ and (putative) pyramidal cells. The signals are qualitatively different. (c) The kurtosis of the signal distributions differs strongly between PV cells (shown) and that of excitatory neurons (not shown). Only 1% of the putative excitatory cells have kurtosis values lying to the left of the dashed vertical line. (d) Distribution of kurtosis for VIP and SOM cells has also low kurtosis values. (e) Clustering of tuning similarity remains largely unmodified even after removing all cells with kurtosis values less than 15.

Figure 5. Robustness of clustering.

(a) Clustering of tuning similarity remains unaffected if the analysis is restricted only to cells with sharp tuning selectivity, defined here as kernels with peak spatial frequency at values larger than 0.025 cycles per degree. (b) Clustering of tuning similarity remains unaffected after removing the ROIs with the smallest 20% areas, thereby ruling out that the occasional inclusion of dendritic processes may be responsible for the results.

Figure 6. Re-analysis of data from a seminal study shows evidence of spatial clustering.

(a) Dependence of relative preferred direction as a function of cortical distance restricted to cells that are no farther apart than 100 μm. Data are replotted from Fig. 6a of the original study. There is a statistical dependence of relative direction with cortical distance (P<0.003, r=0.22). (b) The distribution of relative directions for cells within 100 μm of each other is statistically different from that of cells distanced between 50 and 100 μm away (tailed rank-sum test, P<0.02).