Visual stimulus features that elicit activity in object-vector cells

Abstract

Object-vector (OV) cells are cells in the medial entorhinal cortex (MEC) that track an animal's distance and direction to objects in the environment. Their firing fields are defined by vectorial relationships to free-standing 3-dimensional (3D) objects of a variety of identities and shapes. However, the natural world contains a panorama of objects, ranging from discrete 3D items to flat two-dimensional (2D) surfaces, and it remains unclear what are the most fundamental features of objects that drive vectorial responses. Here we address this question by systematically changing features of experimental objects. Using an algorithm that robustly identifies OV firing fields, we show that the cells respond to a variety of 2D surfaces, with visual contrast as the most basic visual feature to elicit neural responses. The findings suggest that OV cells use plain visual features as vectorial anchoring points, allowing vector-guided navigation to proceed in environments with few free-standing landmarks.

Fig. 1. Identification of object-vector (OV) cells.

a Rate maps for Example Cell 10 from the three trials used to identify it as an OV cell. In the first trial, the environment is empty (‘Empty Box’). In the second trial, a free-standing object made of Duplo bricks is present in the environment (‘Object’). In the third trial, the same object is moved to a new location (‘Moved Object’). Rate maps show colour-coded firing rate in Hz as a function of the animal’s position. White squares mark the object location. Peak firing rate (Hz) is indicated below each map. Scale bar represents 40 cm. The example cell fires in a specific distance and direction away from the object, the defining behaviour of OV cells. b Object-centred rate maps, displaying firing rate (Hz) as a function of the animal’s distance (cm) and orientation (degrees) to the object. The object position is the centre of the map. The maps are from the ‘Empty Box’, ‘Object’ and ‘Moved Object’ trials for the cell shown in panel a. For an OV cell, we expect the maps from the ‘Object’ and ‘Moved Object’ trials to be similar. Scale bar in white, 20 cm. c Shuffling distribution of spatial information content for Example Cell 10 (same cell as in previous panels). The cell’s spike timestamps were randomly shifted along the animal’s trajectory (n = 200 permutations). For each shuffled cell, we calculated the spatial information content, which quantifies how informative the spikes are about the animal’s position. The actual spatial information content of the cell is far above the 99th percentile of the shuffling distribution. Spatial information contents were calculated using data from the ‘Object’ trial. d Shuffling distribution of OV scores for Example Cell 10. After performing shuffling (as in the previous panel) we calculated the OV score for each shuffled cell. The OV score is the Pearson correlation between the object-centred maps shown in panel b from the ‘Object’ and ‘Moved Object’ trials. The actual OV score of the cell exceeds the 99th percentile of the shuffling distribution. e Scatterplot showing the actual spatial information content (bits/spike) of OV cells, compared to the threshold value obtained from each cell’s shuffling distribution. The threshold was the 99th percentile of the shuffling distribution. All data points fall above the diagonal because, by definition, OV cells need to pass the spatial information criterion. The spatial information content was calculated on the ‘Object’ trial (a). f Scatterplot showing the actual OV score of OV cells, compared to the threshold value obtained from each cell’s shuffling distribution. The threshold was the 99th percentile of the shuffling distribution. All data points fall above the diagonal because, by definition, OV cells need to pass the OV score criterion. The OV score is the Pearson correlation between the object-centred maps in (b) from the ‘Object’ and ‘Moved Object’ trials.

Fig. 2. Object-vector cells respond to two-dimensional objects.

a Experimental design with four different trials: mice foraged in an environment with either an object absent (‘Empty Box’) or present (‘2D’, ’50%’, ‘3D’). In the three object trials, we varied the amount of the object’s volume exposed to the animal. In the 2D trial, the visible part of the object was a flat 2D surface, with the rest completely embedded into the wall. In the middle trial, the object was partially embedded into the wall so that 50% of its volume was exposed. In the 3D trial, the full volume of the object was exposed. The reference trial, used to find the cell’s vector coordinates, is shown on the left. This was the original ‘Object’ trial used to identify the cell as an OV cell (Fig. 1a, middle). b Colour-coded rate maps from example OV cell that responded as strongly to 2D surfaces as to 3D objects. Rate maps show colour-coded firing rate in Hz as a function of the animal’s position. The white square marks the object location. The dotted circle marks the ROI in which we expected the cell to fire based on its vector coordinates. The vector coordinates of the cell were found by applying the algorithm from Supplementary Fig. 3 to the ‘Reference trial’ on the left. This trial is also what we used as a template for computing the OV score. Peak firing rate (Hz) is indicated below each map. Scale bar, 40 cm. c Box-and-whisker plots of normalised firing rates (Hz) of object-vector cells (n = 30) as a function of object dimensionality. Firing rates were normalised to the maximum data point across all eight data points observed for each cell (4 experimental conditions × inside/outside ROI) so that the maximum rate is 1. The middle line (in red) on each box indicates the median, while the bottom and top lines (in blue) indicate the lower and upper quartiles, respectively. Whiskers show the range of data in each condition. Red crosses show outliers that lie more than 1.5 times outside the interquartile range. The firing rate was measured either inside (left) or outside the ROI (right). d Box-and-whisker plot of object-vector scores (ranging from −1 to 1) as a function of object dimensionality. The object-vector score is equivalent to the Pearson correlation between pairs of object-centred rate maps (see 'Method' for details). Here, the reference rate map is from the ‘Object’ trial originally used to identify the cell.

Fig. 3. Object-vector cells respond to transparent objects.

a Experimental design with three different trials: either the box was empty, had a transparent object present, or a standard object present (a Duplo tower). The reference trial, used to find the cell’s vector coordinates, is shown on the left. b Rate maps from example OV cell that responded strongly to the transparent object. Rate map conventions are as in Fig. 2. Scale bar, 40 cm. c Box-and-whisker plots of normalised firing rates (Hz) of OV cells (n = 14) as a function of discreteness of the object, inside (left) or outside the ROI (right). d OV score as a function of object type. Conventions for box-and-whisker plots are as in Fig. 2.

Fig. 4. Responses of object-vector cells to visual contrasts.

a Experimental design with four different trials: either the box was empty or had a visual contrast present. In the three trials with a visual contrast, we varied the object’s contrast from dark grey, to grey, to white (10%, 60% and 100% whiteness, respectively). The reference trial, used to find the cell’s vector coordinates, is shown on the left. b Example rate maps from OV cell clearly increasing its firing rate as a function of visual contrast. Rate map conventions are as in Fig. 2. Scale bar, 20 cm. c Box-and-whisker plots of normalised firing rates (Hz) of OV cells (n = 14) as a function of the object’s visual contrast. The firing rate was measured either inside (left) or outside the ROI (right). d Box-and-whisker plot of OV scores as a function of visual contrast. Conventions for box-and-whisker plots are as in Fig. 2.

Fig. 5. Posterior probability distributions of the average firing rates of OV cells to each object or feature.

a Visualisation of Bayesian inference. Bayes’ theorem multiplies the likelihood with a prior to give the posterior distribution summarising our state of knowledge. The likelihood shows which average firing rates (FRs) best explain the data obtained. The prior gives the probabilities of different firing rates before seeing the data. The resulting posterior shows the probability that OV cells respond with any average FR to the object. For the distributions in Fig. 5, we assigned a Gaussian likelihood and assumed a uniform prior (see 'Methods' for details). We verified that our conclusions were insensitive to these assumptions (Supplementary Figs. 7 and 8). b Posterior distributions showing the probability that OV cells respond with any average FR to each object. The plots correspond to Fig. 2 and describe the results for the 2D surface (blue curve), the partially embedded object (red curve) and the 3D object (yellow curve). The data are the firing rates of OV cells inside the region in which we expect each cell to fire based on its vector coordinates. The firing rate in the same region in the ‘Empty Box’ trial has been subtracted. This means that probability on the right (left) of 0 Hz should be interpreted as a positive (negative) response to the object. 0 Hz is marked by the stippled line. The intervals at the top represent the ‘credible region’, which is the smallest possible region containing 95% of the probability mass. The credible region can be interpreted as '‘there is a 95% probability that OV cells respond with an average FR between these bounds’. c Same as in the previous panel but probability distributions corresponding to Fig. 3 describing the results for the transparent surface (blue curve) and normal 3D tower (red curve). d Same as in the previous panels but probability distributions corresponding to Fig. 4 describing the results for the dark grey, grey and white contrast. Note that while all curves peak at positive values, the right-shift of the curves increases gradually with increasing visual contrast.