Abstract
Social recognition consists of multiple memory processes, including the detection of familiarity – the ability to rapidly distinguish familiar from novel individuals – and recollection – the effortful recall of where a social episode occurred and who was present. At present, the neural mechanisms for these different social memory processes remain unknown.
Here, we investigate the population code for novel and familiar individuals in mice using calcium imaging of neural activity in a region crucial for social memory, the dorsal CA2 area of the hippocampus. We report that familiarity changes CA2 representations of social encounters to meet the different demands of social memory. While novel individuals are represented in a low-dimensional geometry that allows for rapid generalization, familiar individuals are represented in a higher-dimensional geometry that supports high-capacity memory storage.
The magnitude of the change in dimensionality of CA2 representations for a given individual predicts the performance of that individual in a social recognition memory test, suggesting a direct relationship between the representational geometry and memory-guided behavior. Finally, we show that familiarity is encoded as an abstract variable with neural responses generalizing across different identities and spatial locations. Thus, through the tuning of the geometry of structured neural activity, CA2 is able to meet the complex demands of multiple social memory processes.
Tuned geometries of hippocampal representations meet the demands of social memory
Lara Boyle*, Lorenzo Posani*, Sarah Irfan, Steven A. Siegelbaum, Stefano Fusi
* co-first authors
Figures and captions
Figure 1. Experimental design. a) Six Amigo2-Cre subject mice were injected with a Cre-dependent GCaMP6f virus and implanted with a GRIN lens over dorsal CA2, enabling visualization of the activity of 439 CA2 pyramidal neurons. b) Implanted subjects ran through experiments in an oval arena with two stimulus mice (N1 and N2) under wire pencil cup cages, allowing interactions between the subject and stimulus mice. c) Subjects were first allowed to explore two novel individual stimulus mice in a five-minute trial. To distinguish social and spatial responses, the position of the stimulus mice were swapped in a second 5-minute trial. d) A plot of the total mean interaction time of the subjects with the stimulus mice in the two trials. No significant difference was observed for exploration of N1 or N2 in either trial: Two-way ANOVA for Partner x Trial F(1,5) = 0.0530, p=0.83. e) Example 114 simultaneously recorded deconvolved calcium traces across the two trials from a single subject mouse; f) Plot of the position of the subject along the axis defined by the center of the cups and identity of the interaction partner during the two trials. The colored lines on top and colored areas denote active interactions (sniffing) with either interaction partner. The four colors correspond to the four combinations of spatial (left versus right cup) and social (mouse N1 versus N2) variables.
Figure 2. Neural recordings of dCA2 pyramidal neurons can elucidate the geometrical relationships between represented variables. a) Scheme of the two-novels experiment. b-d) Examples of geometrical arrangements of the coding of two variables (identity and position) with different dimensionality. Each point plots the firing rates of three neurons (r1, r2, r3) measured for a given color-coded condition. Because of variability in activity, multiple instances of a given condition results in a cloud of points. b) An example one-dimensional arrangement of the four conditions (N1-left, N2- left, N1-right, N2-right) in which identity but not position can be decoded. Identity is abstract to position in this orientation. A classifier plane (yellow) can readily separate the clouds of points to decode identity. c) Example two-dimensional arrangements in which position and identity are encoded and disentangled. Left, the neurons are specialized, responding to either identity or position, resulting in a rectangular planar geometry. Two classifier planes (yellow and cyan) can readily separate the point clouds to decode identity and position. In addition, the classifiers generalize in that a classifier plane that optimally decodes identity when the animals are in the left cup will also decode identity when the animals are in the right cup (yellow planes), and vice versa. Similarly, the classifier trained to decode position when a given stimulus animal (e.g., N1) is in the two cups will also decode position when the other stimulus animal (e.g., N2) is in the two cups (cyan planes). Right, in this example, the neurons have mixed selectivity; they respond to both identity and position. However, they respond in a linear manner to the two variables, leading to a rotation of the points that maintains the rectangular-like planar 2-D geometry. This geometry also enables a linear classifier to perform with a high decoding accuracy and high generalization. Note the two geometries in c do not enable the decoding of points corresponding to the conditions represented at opposite corners of the rectangle-like shapes (N1-L & N2-R, N2-L & N1-R). This is termed the XOR dichotomy and corresponds to the conditions in the two trials. d) An example three-dimensional, tetrahedral geometric arrangement in which position, identity, and XOR can be decoded by a linear classifier. Here, position and identity are entangled. e-g) different geometries in b-d have corresponding distinct fingerprints for their ability to decode identity, position and XOR, as well as for generalized decoding of identity and position as measured by CCGP. h) Tradeoff between storage capacity and CCGP in a simulation of a Hopfield recurrent neural network. A network of N neurons was trained to store and retrieve a set of patterns with geometrical dimensionality varying from L << N to N (see Methods and Supplementary Information). In this simulation, we used L=10, N=400. To vary the dimensionality of the patterns, we added a random distortion by flipping the value of each neuron in each pattern with a given probability ranging from 0 (L-dimensional, correlated activity) to 0.5 (N-dimensional, random uncorrelated patterns) with a step size of 0.05. Curves and points represent the average over n=10 model simulations. Storage capacity was normalized to range between 0 and 1.
Figure 3. Novel and familiar identities are coded in different geometrical arrangements that support different memory requirements. a) Schemes for decoding identity, position, XOR (training and testing on data from all four conditions) and CCGP for identity and position (training on one pair of conditions and testing on the other). The numbers and colors indicate the identity of the stimulus animal under the cup (1, blue or 2, orange), the shade of the color indicates whether an animal is in the left cup (darker shades) or right cup (lighter shades) and the outline whether the data is from trial 1 (solid outline) or trial 2 (dashed outline). b-e) The scheme (b), decoding results (c, d) and proposed geometry (e) for the experiment with two novel stimulus mice. c) Open symbols show mean decoding performances compared to results of null model based on shuffled data (solid points and error bars, respectively). Novel mouse identity and position are decoded significantly above chance (id decoding = 0.76, null model = 0.50 ± 0.06; pos decoding = 0.88, null model = 0.50 ± 0.06). d) Identity CCGP and position CCGP are also significantly higher than the null model (id CCGP = 0.70, null model = 0.49 ± 0.04; pos CCGP = 0.82, null model = 0.49 ± 0.04). XOR coding does not differ from the null model (XOR decoding = 0.54, null model = 0.50 ± 0.06). e) Proposed low-dimensional geometry for social/spatial representations of two novel mice. f-i) The scheme (f), decoding results (g, h) and proposed geometry (i) for the experiment with two littermates as stimulus mice. g) Familiar mouse identity and position are decoded significantly better than chance (id decoding = 0.72, null model = 0.50 ± 0.06; pos decoding = 0.91, null model = 0.50 ± 0.06). h) Identity CCGP is not significantly greater than chance (id CCGP = 0.55, null model = 0.50 ± 0.05) while spatial CCGP and XOR decoding are significantly greater than chance (pos CCGP = 0.76, null model = 0.50 ± 0.03; XOR decoding = 0.66, null model = 0.50 ± 0.06). i) Proposed geometry for social/spatial representations of two familiar littermates. j) Difference in indicated decoding performance (Δ) in tests with two familiar stimulus mice compared to two novel stimulus mice. CCGP for identity and position were significantly greater with novel compared to familiar mice whereas XOR values were greater with familiar compared to novel mice (Δ identity CCGP = -0.14, null = 0.00 ± 0.06, p = 0.0042; Δ position CCGP = -0.06, null = 0.00 ± 0.04, p = 0.029; Δ XOR = 0.12, null = 0.00 ±0.04, p =0.0012). Values are reported as mean ± STD. Null model error bars show 2 STDs around the mean. P values are estimated from the z-score of the data values compared to the null model distributions. *p<0.05, **p<0.01, ***p<0.001.
Figure 4. Degree of familiarity is encoded in a low-dimensional format disentangled from identity and position. a) Subject mice (n=5 mice, 438 cells) were exposed to one novel stimulus mouse and one familiar stimulus mouse for a five-minute trial. The subject mice were then exposed to a second set of novel and familiar individuals in a second five-minute trial in which the position of the novel and familiar individual were swapped. b) Decoding scheme for familiarity CCGP and position CCGP. c) Familiarity and position CCGP and XOR decoding were significantly greater than the null model (familiarity CCGP = 0.74; null model = 0.51 ± 0.04; position CCGP = 0.89, null model = 0.50 ± 0.03; XOR decoding = 0.57, null model = 0.54 ± 0.02). d) Graphical representation of a geometrical model for encoding social and spatial information of novel and familiar mice by three example neurons (firing rates r1, r2, r3). Dark and light gray circles represent firing rates during specific combinations of social and spatial variables during interactions with novel and familiar animals, respectively. Increasing familiarity both shifts and distorts in neural firing space the planar, low-dimensional social-spatial representations of novel animals. e) A model based on the geometry depicted in d reproduces experimental observations (data from Fig. 3 is overlaid on predictions from the model. A best fit of the 6 parameters of the model reproduces our 10 decoding experimental data points, see Methods). Lines and shaded areas show mean ± SD calculated using 100 model simulations. Values are reported as mean ± STD. P values are estimated from the z-score of the data values compared to the null model distributions. Null model error bars show 2 STDs around the mean. * p<0.05, *** p<0.001.
Figure 5. Magnitude of geometrical distortion in familiar and novel social/spatial representations is correlated with behavior in a separate social memory task. a) Subject mice performed a social recognition memory test in which they explored an arena with cups containing one novel and one familiar stimulus mouse across two five-minute trials (with positions swapped in the second trial). b) To validate that this test assesses social recognition, a total of 12 mice underwent the test, including the 6 implanted subjects from Figure 3. The subjects spent on average significantly greater time exploring the novel compared to the familiar stimulus mouse in trial 1 but not in trial 2 (two-way ANOVA: Interaction Partner x Trial, F(1,11) = 9.208, p=0.011. Šídák’s multiple comparisons test: trial 1 p=0.0085; trial 2 p=0.75). c) Total interaction time with the novel and with the littermate across the two trials, from the subset of 6 mice that were run in two-novels and two-littermates experiments. d) XOR decoding during trials with two novel and two familiar stimulus mice for individual subject mice. Mean XOR decoding value was significantly different from chance level of 0.5 with familiar mice (0.57 ± 0.045). There was a strong trend for a greater mean XOR with familiar compared to novel stimulus mice that did not reach significance (two-novel mice: XOR mean = 0.51 ± 0.03 SEM; two familiar mice: XOR mean = 0.57 ± 0.02 SEM; n=6, Student's paired t-test, p=0.055). Points show values for the individual subjects. Vertical and horizontal lines show mean ± SEM. e) The magnitude of the difference in XOR decoding performance for individual animals was significantly correlated with behavioral preference for the novel compared to the familiar mouse in the separate social memory test of panel a (r=0.90, p=0.016). f) Interaction term from ANOVA performed for CA2 firing rates as function of mouse identity and position with two novel or two familiar stimulus mice. Mean interaction term value (position x identity) was significantly greater with two familiar mice (2.51 ± 0.12 SEM) compared to two novel mice (2.02 ± 0.18 SEM; Student's paired t-test, p=0.0084, n=6 mice). Points show values for the individual subjects. Vertical and horizontal lines show mean ± SEM. g) The interaction term was correlated strongly with behavior preference for the novel over the familiar individual in the social memory test of panel a (r=0.87, p=0.025). *p<0.05, **p<0.01.