Distinct neural dynamics in the ventral hippocampus and medial prefrontal cortex during social information processing Authors: Xinnian Wang, Zheng Ma, Yifan Luo, Yechao Han, Chuanzan Zhan, Yulin Ouyang, Yang Zhan Published: 2025 ABSTRACT SignificanceSocial interactions are fundamental to animal behavior, yet how the brain processes social information remains poorly understood. This study reveals that two key brain regions, the medial prefrontal cortex (mPFC) and ventral hippocampus (vHPC), use distinct neural coding strategies during social encounters. By combining electrophysiological recordings with advanced analytical techniques, we show that the mPFC encodes social information more robustly and with greater clarity than the vHPC. These findings advance our understanding of how distributed brain circuits process social cues and may inform future research on social dysfunction in neurological and psychiatric disorders. The ability to encode socially relevant information is crucial for social species to recognize and differentiate between individuals. This information not only helps in distinguishing one individual from another but also facilitates the formation of relationships and the exchange of important information. This process relies on specialized brain regions that are responsible for processing identity information (1–3). Ventral hippocampus (4–6) (vHPC) and medial prefrontal cortex (7–9) (mPFC) are two important brain structures in regulating social behaviors. While the majority of studies have focused on perturbing neural activities in either the vHPC or mPFC, or their connecting circuits to investigate the relationship between neuronal manipulations and behavioral changes, the mechanisms by which neural populations encode social information to distinguish one social stimulus from another relying on populational responses and oscillatory patterns remain less understood. Previous studies have shown that vHPC neurons play a critical role in social recognition memory (4, 10–13). Several output pathways of the vHPC including projections to nucleus accumbens (4) and lateral septum (14) contribute to this function. Inhibition of these pathways results in impaired social recognition behavior, suggesting that individual-specific information encoded in the vHPC may be transmitted via these projections to support social processing. Concurrently, the mPFC exhibits the ability to encode various types of social information, such as odor cues of different mice (15), social approach of different mouse targets (16–20) and interactive behavior among group members (21, 22). The mPFC sends projections to multiple brain regions, including nucleus accumbens (23), thalamus (16) and amygdala (24). The intact functionality of these projections is essential for social recognition behavior. The neuronal activity patterns observed in both the vHPC and mPFC, along with their respective downstream circuits, highlight the importance of these two brain regions in representing distinct aspects of social stimuli. However, at the population level, how these two areas differentially encode social information remains largely unknown. In this study, we recorded single-unit activities and local field potentials (LFPs) from the vHPC and mPFC during the social interaction task. We then conducted decoding analyses using neural population activity. While decoding based on single-unit activity is more commonly employed, we also extracted a range of oscillatory LFP features from multiple frequency bands recorded via parallel wired and wireless recording systems, to investigate the coding capacities and differential representations in these two brain regions. Our results revealed gamma band-specific processing in both the vHPC and mPFC, with distinct involvement of low- and high gamma activities in each brain region. Furthermore, we applied geometric analysis to population vectors and identified differences in how social information was represented across the two areas. These population-level and oscillatory patterns highlight key strategies for social information processing in the vHPC and mPFC, supporting their distinct coding mechanisms. Dysfunction or maladaptive neural coding in these brain areas may give rise to deficient social behavior. Results Differential Neuronal Responses in the vHPC and mPFC. To investigate how vHPC and mPFC neurons dynamically response to social stimuli, we performed in vivo electrophysiological recordings (SI Appendix, Fig. S1A) in freely moving male mice during the three-chamber social interaction task (25–27) (Fig. 1A), with movable tetrodes targeting either vHPC or mPFC (Fig. 1B). During the sociability phase, the mice investigated either a cup containing a juvenile social stimulus (S) or an empty cup (E), while during the social novelty phase, they chose between the familiar juvenile stimulus (S’) and a novel juvenile stimulus (N). The mice preferred S over E during the sociality phase and N over S’ during the social novelty phase as quantified by cumulative investigation time (Fig. 1 C and D and SI Appendix, Fig. S1 B and C), average duration per investigation bout (SI Appendix, Fig. S2 A and D), and average number of investigation bouts (SI Appendix, Fig. S2 B and E). Motor activities during the investigation of four stimuli were not different (SI Appendix, Fig. S2 C and F). Distinct neural responses to social stimuli in the vHPC and mPFC. (A) Schematic of the three-chamber social interaction test. The gray line represents the movement trajectory of a subject mouse. (B) Multichannel in vivo electrophysiological recordings from the vHPC or mPFC using tetrodes. (C and D) Representative heatmaps of occupancy time from three chamber social tests with tetrode arrays implanted in the vHPC (C) and mPFC (D). Discrimination index was calculated as the difference time divided by the total social interaction time, specifically comparing S vs. E during sociability phase and N vs. S’ during social novelty phase. Each dot represents a mouse. (E–H) Proportions of vHPC (E and F) or mPFC (G and H) neurons with enhanced, suppressed, and nonsignificant responses. Neuronal responses were aligned to sniffing onset within each stimulus. The gray shaded area represents the average response across individual stimuli during 0 to 2 s period. Enhanced neurons (Top, solid line), Suppressed neurons (Bottom, dotted line). vHPC: n = 306 neurons from 11 male mice. mPFC: n = 324 neurons from 10 male mice. Chi-square test, χ2 = 8.032, P = 0.0180. (I–L) Response magnitudes of individual stimuli for enhanced and suppressed neurons from the vHPC (I and J) and mPFC (K and L). Neuronal firing rates were z-score normalized using the mean firing rate during the 2 s period before sniffing onset as baseline. The Kruskal–Wallis test with Dunn’s multiple comparison correction. (I) Enhanced neurons in vHPC, n = 57. H(3) = 29.11. P < 0.0001. (J) Suppressed neurons in vHPC, n = 54. H(3) = 54.83. P <0.0001. (K) Enhanced neurons in mPFC, n = 40. H(3) = 17.28. P < 0.01. (L) Suppressed neurons in mPFC, n = 81. H(3) = 25.93. P < 0.0001. (M and N) Distribution of selectivity index in vHPC and mPFC neurons during the sociability phase (M) and the social novelty phase (N). Two-sided Kolmogorov–Smirnov test. Sociability phase, vHPC vs. mPFC: D = 0.1329, P = 0.0078. Social novelty phase, vHPC vs. mPFC: D = 0.1010, P = 0.0828. * P < 0.05, ** P < 0.01, ***P < 0.001. Data are presented as mean ± SEM. We observed both enhanced and suppressed neuronal responses to these stimuli in both brain regions. Specifically, in the vHPC, 19% of neurons exhibited enhanced responses, whereas 18% showed decreased activity (Fig. 1E). In contrast, within the mPFC, these percentages were 12% and 25% (Fig. 1G). In addition, we observed that mPFC exhibited higher firing rates during social tasks than vHPC (SI Appendix, Fig. S1E). These findings indicate that although both brain areas contain neurons that preferentially respond to social stimuli, the proportions of responsive neurons differ between the two regions. Notably, the neural response patterns varied across different stimulus categories (Fig. 1 F and H). In the vHPC, enhanced neurons exhibited stronger responses to S and N compared to E and responses to S were stronger than S’ (Fig. 1I). For the suppressed neurons, S and N elicited greater inhibition than E, with S inducing more suppression than N and S’ (Fig. 1J). In the mPFC, enhanced neurons showed stronger responses to S than to E (Fig. 1K). Suppressed neurons responded more strongly to S and N than to E, with S evoking greater inhibition than S’ (Fig. 1L). These results show that mPFC and vHPC neurons show differential responses to different social and nonsocial stimuli. We further classified neurons into putative excitatory and inhibitory types based on spike waveform characteristics and firing rates (SI Appendix, Fig. S3 A and B). In the vHPC, among 196 putative excitatory neurons, 23.97% showed enhanced responses and 12.75% showed suppressed responses. In the mPFC, the corresponding proportions were 12.94% and 26.34% among 224 putative excitatory neurons, respectively. Among 38 putative inhibitory neurons in the vHPC and 46 putative inhibitory neurons in the mPFC, only two neurons in each region exhibited enhanced responses. Meanwhile, 16 inhibitory neurons in each region showed suppressed responses. For putative excitatory neurons in the vHPC, enhanced neurons responded more strongly to N than to S or S’, whereas suppressed neurons exhibited greater inhibition in response to S compared to S’ and E (SI Appendix, Fig. S3C). In the mPFC, enhanced excitatory neurons showed stronger responses to S than to S’, and suppressed excitatory neurons were more strongly inhibited by S than by E (SI Appendix, Fig. S3D). For putative inhibitory neurons in the vHPC, suppressed neurons exhibited greater inhibition in response to S compared to E and S’ (SI Appendix, Fig. S3E). In the mPFC, although a main effect of stimulus was observed, post hoc tests did not reveal differences between individual stimuli (SI Appendix, Fig. S3F). These findings suggest that putative excitatory neurons in the vHPC and mPFC differentially encode familiar and novel social stimuli. To assess the preference of individual neurons for different stimuli, we calculated selectivity indices during both sociability and social novelty phases (18, 28). During the sociability phase, the distribution of the selectivity indices (S vs. E) differed between the two brain areas (Fig. 1M). A greater proportion of mPFC neurons exhibited selectivity indices close to zero. In contrast, vHPC neurons displayed a broader distribution, with more mid-range selectivity values. However, during the social novelty phase, no significant difference was observed in the distribution of selectivity indices (N vs. S’) between the two regions (Fig. 1N). Thus, the selectivity index revealed distinct patterns of neuronal preferences in the two brain regions. Information Representation Is Stronger in the mPFC than vHPC. To understand how neuronal populations in the vHPC and mPFC represent different social stimuli, we performed an information-theoretic analysis based on Shannon entropy (29, 30), which incorporates both the probability of each stimulus and the strength of the neuronal responses. In the three-chamber social task, the subject mouse was exposed to a series of stimuli belonging to distinct categories (E, S, N, and S’), while neural population activity was recorded during each stimulus investigation. Mutual information was used to quantify how much information the neural population responses convey about stimulus identity. Specifically, it measures the reduction in uncertainty regarding the stimulus based on single-trial neural activity, accounting for both the discriminability of mean responses across different stimuli and the trial-to-trial variability within each condition. By applying this analysis to binned population vectors, we found that social information was represented more robustly in the mPFC compared to the vHPC (Fig. 2A). Enhanced representation of social information in the mPFC compared to the vHPC. (A) Information analysis based on Shannon entropy using population vectors in the vHPC and mPFC during investigation of stimuli E, S, N, and S’. Two-way ANOVA with Bonferroni correction. vHPC: n = 12 sessions, mPFC: n = 10 sessions. Brain area: F(1, 80) = 16.53, P = 0.0001. Bin length: F(3, 80) = 11.40, P < 0.0001. (B) Decoding accuracy in classification of four stimuli (S, E, N, and E) using population vectors with varying population dimensions (1 to 10 neurons) in the vHPC and mPFC. Balanced accuracy was calculated to account for the unequal number of trials across four stimuli. The bin length was 100 ms. Two-way ANOVA with Bonferroni correction. vHPC: n = 10, 11, 11, 11, 9, 9, 9, 9, 9, 9 sessions for dimensions 1 to 10, respectively; mPFC: n = 15, 15, 15, 15, 14, 14, 13, 12, 12, 12 sessions for dimensions 1 to 10, respectively. Brain region × Dimension: F(9, 214) = 2.591, P = 0.0075. Brain area: F(1, 214) = 18.48, P < 0.0001. Dimension: F (9, 214) = 24.30, P < 0.0001. Dimension 9, P = 0.0056. Dimension 10, P = 0.0078. *P < 0.05, **P < 0.01, ***P < 0.001. Data are presented as mean ± SEM. We next investigated the differences in coding of different social stimuli between vHPC and mPFC at the population level. Using binned population vectors derived from neural recordings during stimulus investigation, we performed Bayesian decoding analysis to classify four different stimuli (E, S, N, and S’). As more neurons were included in the analysis, the decoding accuracy increased, with the mPFC showing higher accuracy at neuron dimensions of 9 and 10 (Fig. 2B). At other bin lengths, we also found that decoding accuracy was higher in the mPFC (SI Appendix, Fig. S4). We further conducted decoding analyses using population vectors composed exclusively of putative excitatory neurons. Under this condition, the mPFC maintained higher information content and decoding accuracy compared to the vHPC (SI Appendix, Fig. S5 A and B). Therefore, both information-theoretic and decoding analyses demonstrate that the mPFC exhibits a stronger representational capacity than the vHPC at the population level. Distinct Gamma Band Activities in the vHPC and mPFC. To explore how oscillatory neural activity responds to social stimuli, we analyzed LFPs recorded during the three-chamber social interaction test. To ensure the consistency and robustness of the observed LFP patterns, electrophysiological recordings were conducted using both wired (Plexon system) and wireless (Neurologger system, SI Appendix, Fig. S6) recording setups in the three-chamber paradigm. We then performed time-frequency analysis by aligning the LFP data at the onset of stimulus investigation (31–33). In the vHPC, we observed an increase in low gamma (30 to 60 Hz) power during the investigation of S, S’, and N, but not E (Fig. 3 A–C and E). A control condition (C) which included randomly selected noninvestigation periods did not show any changes in low gamma power before and after the investigation onset. We compared the magnitude of the power across different stimuli. The magnitude of S was higher than those of E and C, and the magnitude of N was higher than that of C (Fig. 3 D and F). Additionally, in data recorded using the Plexon system, we also observed that the magnitude of S’ was higher than that of C (Fig. 3D). No changes in high gamma or theta power were observed either before and after the investigation or across different stimulus conditions (SI Appendix, Fig. S7 A–H). These data demonstrate that low gamma activity in the vHPC was engaged during the investigation of social stimuli. Differential gamma oscillations in the vHPC and mPFC. (A and B) Time-frequency spectrograms of LFPs in the vHPC recorded using two different systems. (A) Plexon system. (B) Neurologger system. Data were averaged for individual stimuli across recording sessions, with time 0 aligned to the sniffing onset. Warmer colors represent increased z-scored power. (C–F) Low gamma (30 to 60 Hz) power in the vHPC. Two-sided paired t tests were used to compare prestimulus (0 to 2 s) and stimulus (−2 to 0 s) periods. (C) Plexon, n = 28 sessions. S: t(27) = 5.041, P < 0.0001; N: t(27) = 4.153, P = 0.0003; S’: t(27) = 3.377, P = 0.0022. (E) Neurologger, n = 11 sessions. S: t(10) = 4.534, P = 0.0002; N: t(10) = 9.073, P < 0.0001; S’: t(10) = 3.863, P = 0.0009. Comparisons across stimuli were performed using one-way ANOVA tests followed by post hoc Bonferroni multiple comparisons. (D) Plexon, n = 28 sessions. F (4, 135) = 8.189, P < 0.0001. S vs. E, P = 0.0075. S vs. C, P < 0.0001. N vs. C, P = 0.0129. S’ vs. C, P = 0.0245. (F) Neurologger, n = 11 sessions. F(4, 50) = 5.39, P = 0.0011. S vs. E, P = 0.0367. S vs. C, P = 0.0031. N vs. C, P = 0.0315. (G and H) Time-frequency spectrograms in the mPFC. (G) Plexon system. (H) Neurologger system. (I–L) High gamma (60 to 120 Hz) power in the mPFC. The same statistical analyses as above were used. (I) Plexon, n = 22 sessions. S: t(21) = 3.406, P = 0.0027; N: t(21) = 0.1455, P = 0.8857; S’: t(21) = 0.8896, P = 0.3838. (K) Neurologger, n = 11 sessions. S: t(10) = 3.573, P = 0.0051; N: t(10) = 3.456, P = 0.0062; S’: t(10) = 3.828, P = 0.0033. One-way ANOVA tests followed by post hoc Bonferroni multiple comparisons. (J) Plexon, n = 22 sessions, F(4,105) = 3.414, P = 0.0115. S vs. E, P = 0.0343. S vs. C, P = 0.0353. (L) Neurologger, n = 11 sessions, F(4, 50) = 3.542, P = 0.0128. S vs. E, P = 0.0247. S vs. C, P = 0.0394. *P < 0.05, **P < 0.01, ***P < 0.001. Data are presented as mean ± SEM in (C, E, I, and K). Box plots showing median, 25 to 75% percentile, 5 to 95% range and outliers in (D, F, J, and L). In the mPFC, we observed an increase in high gamma (60 to 120 Hz) during the investigation of S but not E or C (Fig. 3 G–I and K). In recordings obtained using the Neurologger system, N and S’ also elicited increased responses. The magnitude of S was greater than those of E and C (Fig. 3 J and L). Additionally, low gamma activity increased during the investigation of S before and after the investigation onset (SI Appendix, Fig. S7 I and K). The magnitude of S was greater than those of S’ and C (SI Appendix, Fig. S7 J and L). No changes were observed in theta band power either before and after investigation onset or across stimulus conditions. (SI Appendix, Fig. S7 M–P). Therefore, high and low gamma activity in the mPFC contributes to the social information processing. Coupled cross-frequency oscillatory activities are known to play roles in memory and cognitive processes (34–36). Therefore, we examined cross-frequency coupling in the two brain areas during social investigation. In the vHPC, we found stronger coupling between theta phase and both high and low gamma amplitude during social stimulus investigation compared to nonsocial stimulus investigation (Fig. 4 A–D and SI Appendix, Fig. S8A). In the mPFC, enhanced coupling was observed between theta phase and high gamma amplitude (Fig. 4 E–H and SI Appendix, Fig. S8B). Hence, we demonstrate that differential patterns of theta-gamma coupling are engaged in the vHPC and mPFC during social investigation. Theta-gamma coupling during social investigation. (A and B) Phase-amplitude coupling in the vHPC under social and nonsocial conditions, recorded using two different systems. (A) Plexon system. (B) Neurologger system. Comodulograms show the normalized modulation index between phase frequencies (2 to 20 Hz) and amplitude frequencies (30 to 150 Hz). Modulation index values were z-score normalized. Social condition includes trials from S, N, and S’. Nonsocial condition includes trials from E and C. (C and D) Normalized modulation index for low and high gamma bands in the vHPC from the Plexon (C) and the Neurologger (D) systems. Two-sided paired t tests were used to compare social vs. nonsocial conditions. (C) vHPC: n = 28 sessions, high gamma, t(27) = 5.803, P < 0.0001; low gamma, t(27) = 3.909, P = 0.0006. (D) vHPC: n = 11 mice, high gamma, t(10) = 2.3346, P = 0.0409; low gamma, t(10) = 2.946, P = 0.0146. (E and F) Average comodograms illustrating phase-amplitude coupling in the mPFC under social and nonsocial conditions from the Plexon (E) and Neurologger (F) systems. (G and H) Normalized modulation index for low and high gamma bands in the mPFC from the Plexon (G) and the Neurologger (H) systems. Two-sided paired ttests were used to compare social vs. nonsocial conditions. (G) mPFC: n = 23 sessions, high gamma, t(22) = 10.04, P < 0.0001; low gamma, t(22) = 0.6891, P = 0.4980. (H) mPFC: n = 11 mice, high gamma, t(10) = 2.532, P = 0.0298; low gamma, t(10) = 0.9922, P = 0.3445. *P < 0.05, ***P < 0.001. Data are presented as mean ± SEM in (C, D, G, and H). Taken together, using two parallel electrophysiological recording systems, we identified distinct gamma-band oscillatory patterns during social investigation. The vHPC is marked by dominant low gamma activity with theta-gamma coupling across both low and high gamma bands, whereas the mPFC shows predominant high gamma activity with theta-gamma coupling limited to the high gamma band. Spike–LFP Synergy Decoding. Distinct gamma band response patterns in the vHPC and mPFC suggest that oscillatory activities of neural data can be leveraged to decode social stimuli. To assess the decoding potential of the LFP, we extracted features across multiple frequency bands, including delta, theta, beta, low gamma, high gamma, and high frequency oscillations (HFOs), using various power metrics such as raw power, relative power, and z-scored power. We then integrated these LFP features with spike population data and applied principal component analysis (PCA) to reduce the combined dataset into a lower-dimensional representation for decoding. Our results demonstrate that combining spike and LFP data enhances decoding accuracy compared to spike data alone in both the vHPC and mPFC (Fig. 5A). Furthermore, mPFC decoding outperformed vHPC when using lower-dimensional representations derived from spike data paired with individual LFP frequency bands (SI Appendix, Fig. S9). These findings indicate that incorporating LFP information enhances decoding performance through spike–LFP synergy. Spike–LFP synergy decoding. (A) Decoding accuracy using spike only vs. combined spike–LFP population vectors in vHPC and mPFC. The Mann–Whitney test comparing spike vs. spike+LFP decoding, n = 7 sessions per group. vHPC, P = 0.0023; mPFC, P = 0.0111. (B) Spike–LFP relationships in vHPC. CCA between LFP band power and spike population latent representations. Two-way ANOVA with Bonferroni correction, n = 15 sessions. Data type × Frequency band F(5,168) = 4.097, P = 0.0015. Data type F(1,168) = 92.78, P < 0.0001. Frequency band F(5,168) = 4.232, P = 0.0012. Beta vs. high gamma: P = 0.0110; beta vs. HFOs: P = 0.0099. Actual correlations exceeded surrogates in all bands except beta. (C) Spike–LFP relationships in mPFC. Statistical analyses were identical to (B). n = 16 sessions. Data type × Frequency band F(5,180) = 2.476, P = 0.0338. Data type F(1,180) = 86.96, P < 0.0001. Frequency band F(5,180) = 2.958, P = 0.0136. Theta vs. high gamma: P = 0.0356; beta vs. high gamma: P = 0.0242; low gamma vs. high gamma: P = 0.0435. Actual correlations exceeded surrogates in all bands. * P < 0.05, ** P < 0.01. Data are presented as mean ± SEM. To further explore the relationship between spike activity and LFPs, we investigated the similarity between the latent representations derived from spike populations and LFP power (37, 38). Previous studies have shown that the interaction between spikes and LFPs varies across brain regions and frequency bands (38). Therefore, we employed canonical correlation analysis (CCA) to identify which LFP frequency bands are closely associated with the latent structure of spike activity. In the vHPC, analyses using surrogate spike data revealed significant correlations between the latent spike representation and most LFP frequency bands, except for the beta band, which showed minimal association. High gamma bands and HFO showed significant higher correlations than the beta band (Fig. 5B). In the mPFC, all LFP bands showed higher correlations with the latent spike representation compared to surrogate data. High gamma band displayed the most prominent correlation, higher than those observed in the theta, beta, and low gamma bands (Fig. 5C). These findings indicate a frequency-band-dependent relationship between spike-derived latent representations and LFP power across two brain regions. Single-Trial Pseudopopulation Decoding and Geometry Analysis. To investigate the time course of decoding performance for distinct social stimuli (S vs. N) during social interaction, we performed single-trial population decoding analyses using pseudopopulation vectors constructed from all neurons recorded across multiple sessions (15, 39). We also generated pseudofeature vectors from LFP data, dimensionally matched to the neuronal population size, to explore single-trial decoding properties of the LFPs. We observed an increase in spike-based decoding for the pseudopopulation vectors following the onset of stimulus investigation in both the vHPC and mPFC, with performance rising above chance levels (Fig. 6 A and B). Consistent with our earlier findings using binned population vectors, single-trial decoding revealed higher classification accuracy in the mPFC compared to the vHPC (Fig. 6C). Similarly, LFP-based decoding showed improved accuracy after the onset of investigation (Fig. 6 D and E), and furthermore, mPFC outperformed vHPC in classification accuracy based on LFP features as well (Fig. 6F). However, when LFP features were limited to theta, low gamma, and high gamma bands only with matched feature dimensions, social information could be decoded in the mPFC but not the vHPC (SI Appendix, Fig. S10). These results demonstrate that both spike- and LFP-based neural representations in the vHPC and mPFC carry time-dependent information about social stimuli, with the mPFC exhibiting superior decoding performance. Single-trial pseudopopulation decoding and geometry representation in the vHPC and mPFC. (A and B) Decoding accuracy for classifying S and N using spike pseudopopulations from the vHPC (A) and the mPFC (B). Colored lines represent the mean across 1,000 subsampling iterations (90 neurons per iteration). The gray lines indicate the chance level derived from 1,000 label shuffles. Mann–Whitney tests comparing 0 to 1.5 s during-stimulus accuracy (gray shaded region) to chance. For both vHPC and mPFC, P < 0.001. (C) Decoding accuracy in the mPFC and the vHPC. Mann–Whitney test, vHPC vs. mPFC, P < 0.001. (D and E) Same as (A and B) but for decoding accuracy using LFP pseudopopulations. Mann–Whitney test. vHPC: actual vs. chance, P < 0.001. mPFC: actual vs. chance, P < 0.001. (F) The same as (C), but for decoding accuracy using LFP pseudopopulations. Mann–Whitney test. vHPC vs. mPFC, P < 0.001. (G) Illustration of the calculation of geometric signal-noise angle. (H) Signal-to-noise angles in vHPC and mPFC. Mann–Whitney test. vHPC vs. mPFC, P < 0.001. *** P < 0.001. Data are presented as mean ± SEM for chance-level data in (A, B, D, and E). Violin and box plots in (C, F, and H) showing median, 25th to 75th percentiles, and whiskers extending to 1.5 times interquartile range. Recent advances in geometric approaches have established a framework for analyzing neural population dynamics during complex tasks, showing that hippocampal and prefrontal ensembles utilize specific geometric representations to encode behaviorally relevant information (40–46). To quantify potential differences in the geometric structure of neural populations between the two brain areas, we calculated the signal-noise angle in high-dimensional space (Fig. 6G). Using pseudopopulations composed of n neurons, we computed the angle between signal and noise vectors within this n-dimensional space. Our results indicated that the angles between signal and noise axes differed between the two brain regions, with the mPFC exhibiting a larger angle than the vHPC (Fig. 6H). Larger angles imply a more favorable decoding capability (47, 48), suggesting reduced information-limiting effects in the mPFC compared to the vHPC. Discussion Our findings demonstrate that mPFC neural ensembles exhibit superior encoding and decoding capabilities compared to those in the vHPC, as revealed by Shannon information analysis and decoding-based approaches. Single-trial population decoding analysis further confirms the stronger decoding performance of the mPFC. Notably, we observed that high gamma activity is predominantly involved in mPFC processing, whereas low gamma activity plays a more prominent role in the vHPC. Geometric analysis of neural population activity reveals distinct signal and noise relationships in these two regions. Collectively, our results uncover key representational characteristics shaped by both neural populations and oscillatory dynamics in the vHPC and mPFC. Gamma oscillations are known to be involved in the processing of socially relevant information (49–51). Previous studies have reported altered low gamma activity during investigation of a social stimulus in the mPFC with enhanced (49) or suppressed (50) activities. Our findings reveal increased low gamma activity in the mPFC during social stimulus investigation, aligning with earlier reports of enhanced gamma power during social interaction (49). Extending beyond low gamma activity, we also observed elevated high gamma power during social investigation, as measured by both wired and wireless recording systems. Notably, our results corroborate prior evidence of theta-high gamma phase-amplitude coupling in the mPFC during social investigation (51), underscoring the involvement of high gamma activity in this region. Compared to the role of gamma activity in the mPFC, gamma oscillations in the vHPC during social tasks remain poorly characterized. One previous study showed phase locking between neuronal spiking and high-frequency (150 to 250 Hz) ripples during rest following social recognition tasks, probably reflecting a replay or consolidation of social experience in the vHPC (12). Here, we found elevated low gamma activity in the vHPC during active social investigation. By employing dual parallel recording systems in the three-chamber social interaction paradigm, we identified distinct gamma oscillatory patterns in the vHPC and mPFC, revealing their differential engagement in social processing. The generation of gamma activity is linked to parvalbumin (PV)-expressing interneurons (52–56). Calcium imaging studies have shown that PV neurons in the mPFC signal the encounter of social stimuli (57). Optogenetic activation of the PV neurons in the mPFC can ameliorate social deficits (51, 56, 58, 59). Similarly, calcium signals from the PV neurons in the vHPC can reflect the social investigation and distinguish between two different social stimuli (10). Given the close relationship between the PV neurons and gamma oscillations and the involvement of PV neurons during social investigation, PV neurons may function as an effective modulator of the observed high and low gamma oscillations observed during social tasks. Genetic mouse models exhibiting social deficits show abnormal patterns of social information processing in the mPFC and vHPC. Specifically, Cntnap2 knockout mice exhibit reduced decoding accuracy for social odor cues in mPFC neurons alongside elevated noise levels in spontaneous mPFC activity (15). Neuroligin3 R451C knockin mice display disrupted theta-gamma coupling and impaired phase locking between neuronal spiking and both gamma activities in the mPFC (51). Similarly, Shank3 knockout mice show attenuated phase locking between vHPC neuronal spiking and high-frequency ripples (150 to 250 Hz) during post–social recognition rest periods (12). Beyond these population-level and oscillatory abnormalities, region-specific manipulations of autism-associated genes reveal distinct circuit contributions. mPFC-specific Cul3 deletion (60) and anterior cingulate cortex Shank3 deletion (61) both impair social behavior and excitatory synaptic transmission, while vHPC-specific Shank3 deletion selectively disrupts social memory (62). Importantly, targeted interventions in the mPFC in various mouse models, particularly local rescue strategies employing chemogenetic (57), genetic (20, 51, 61), or optogenetic approaches (51, 56, 59) in the mPFC, can effectively restore normal social behavior. Therefore, the mPFC and vHPC constitute critical components of the neural circuits underlying social information processing. In monkeys, prefrontal neural activities are involved in social information processing (63). In macaques, face responsive cells and population-level activity have been well documented in the prefrontal cortex (21, 64–66). Beyond face processing, ventrolateral PFC neural populations in socially interacting macaques encode dyadic social information during reciprocal grooming and threat encounters (21). Moreover, dorsomedial PFC activity in groups of macaques tracks conspecific behavioral outcomes during multiagent social tasks (22). In marmosets, neural populations in the hippocampus can encode identity information of other conspecifics (67). In humans, a neuroimaging study using fMRI during a social inference task identify hippocampus and dorsomedial PFC as two key regions can solve the task (68). Transcranial magnetic stimulation (TMS) using various protocols has been applied on dorsolateral PFC in individuals with autism spectrum disorder (69–72). In addition to the improvements of cognitive and social functions following TMS, alterations in gamma-band activity have also been reported (73). These cross-species findings highlight the roles of the prefrontal cortex and hippocampus in social cognition, as well as their potential as targets for therapeutic intervention and as sources of measurable oscillatory biomarkers with translational relevance. Anatomically, the vHPC sends direct projections to the mPFC (24, 59, 74, 75). This unidirectional pathway suggests that information in the vHPC may be relayed to the mPFC for further integration and higher-order processing. The mPFC serves as a key integration hub (76–78), receiving inputs from numerous afferent pathways, including the vHPC. Together, these inputs may contribute to the enhanced representational capacity of the mPFC. Indeed, direct manipulation of the pathway through optogenetic inhibition (59) or chronic chemogenetic activation (74) can induce social deficits. Hyperconnectivity in the pathway has been reported in Mecp2 mutant mouse model and offspring from maternal immune activation models induced by Zika virus infection (74, 79). In these models, inhibition of the pathway has been found to rescue social deficits. Future investigation could explore the synchronous oscillatory or spiking activity between the vHPC and mPFC during social behaviors. A deeper understanding of the functional interactions between these two regions and their underlying circuitry may provide critical insights into the neural substrates of social cognition and offer potential novel targets for interventions aimed at ameliorating social impairments in psychiatric disorders. Materials and Methods Animals. Adult male C57BL/6J mice (2 to 4 mo old; preoperative weight 20 to 30 g) were used as experimental subjects, while juvenile male C57BL/6 J mice (3 to 5 wk old) served as stimulus animals in the three-chamber social test. All mice were housed under a 12-h light/dark cycle (lights on 7:00 to 19:00) with ad libitum access to food and water. Environmental conditions were maintained at 20 to 26 °C and 40 to 70% humidity using an individually ventilated cage system, with a maximum of 5 mice per cage. All animal care and experimental protocols were approved by the Animal Care and Use Committee of the Shenzhen Institute of Advanced Technology (SIAT), Chinese Academy of Sciences (CAS), China. Three-Chamber Social Interaction Test. The test was conducted in a rectangular three-chamber apparatus (60 cm × 40 cm × 22 cm) following established protocols (25–27). For 3 consecutive days prior to testing, the stimulus mice were habituated to the wire cages for 15 min daily within the three-chamber apparatus. On the test day, the experimental mouse was placed 30 min prior in the behavioral testing room for environmental habituation. During the habituation phase, the subject mouse was initially placed in the center chamber and allowed 5 min for free exploration. In the sociability phase, a male stimulus mouse with no prior contact with the experimental mice was placed in one of the two small wire cages, while the other cage remained empty. The subject mouse was then permitted 10 min of free exploration. In the social novelty phase, a novel stimulus mouse was introduced to the other cage, and the subject mouse was allowed an additional 10 min of exploration. The left or right positions of stimulus mice were counterbalanced across recording sessions. Social behaviors were monitored through a digital camera positioned above the apparatus. Behavioral videos were analyzed frame-by-frame using Cineplex Editor V3 software (Plexon Inc.) and manually scored by trained observers blinded to experimental conditions. Four social stimuli were defined as E (empty cage), S (mouse A), N (novel mouse B), and S’ (familiar mouse A), when the nose of the subject mouse touched the wire cage during the electrical recordings in the three-chamber social test. During LFP analysis, an additional stimulus, C (Control), was defined as randomly selected segments at least 10 s away from any of the four stimuli (E, S, N, S’) (SI Appendix, Fig. S1D). Analyses of firing rates, LFP, and decoding performance were restricted to interaction bouts lasting longer than 2 s and separated from prior interactions by an interval of at least 2 s. In Vivo Electrophysiology. Mice were anesthetized using an isoflurane vaporizer (RWD, R580S) with gas concentration control (4% induction, 1% maintenance) and surgically immobilized in a stereotaxic frame (RWD, 68026) with ear bars. A movable multielectrode array was built, consisting of 8 tetrodes, each composed of four twisted platinum-iridium wires (17 um diameter, California Fine Wire). The microdrive assembly included a support structure with PCB boards and cap screws, polyimide tubing fixed to movable nuts for precise depth adjustment, and integrated connectors with secured electrode components. The custom-made multielectrode array was stereotaxically implanted into the mPFC (AP: 1.7 mm, ML: 0.5 mm, DV: 1.8 mm) or vHPC (AP: −3.42 mm, ML: 3.5 mm, DV: 3.5 mm). Following surgery, mice recovered for 7 d under daily veterinary monitoring and then underwent 3 d of habituation to headstage attachment. Neural signals were recorded using an Omniplex high-throughput acquisition system (Omniplex D-DHP128, Plexon Inc.), with spike activity sampled at 40 kHz (bandpass-filtered 300 to 8,000 Hz) and LFPs sampled at 1 kHz (bandpass-filtered 1 to 250 Hz). The tetrodes were incrementally advanced via microdrive screws to record from multiple depths across sessions, with data collected from 10 to 11 mice per brain region. LFP signals were also acquired at 1,600 Hz using a wireless Neurologger 2A device (80, 81). Nichrome wire electrodes (65 um, A-M system, Inc.) were individually implanted in the mPFC and vHPC. The reference and ground electrodes were positioned in the posterior and anterior region, respectively. LFP Analysis. To characterize the neural oscillatory dynamics of the LFP, time frequency analysis was conducted via the Chronux toolbox (31) in Matlab environment. The LFP signals were notch filtered to remove the 50 Hz line noise. The data were detrended and means were removed. We evaluated the behavioral trials 2 s before and after the onset of the investigation. Trials containing high-amplitude artifacts were excluded by thresholding at 1.2 SD above the mean amplitude computed at each stimulus condition. Time-frequency (31–33) power was performed using mtspecgramc function in the Chronux toolbox with a window length of 1 s and a time step of 0.01 s and the time-bandwidth product was 3, and the number of tapers was 5. LFP data were segmented into six frequency bands for statistical analysis: delta (1 to 4 Hz), theta (4 to 12 Hz), beta (15 to 25 Hz), low gamma (30 to 60 Hz), high gamma (60 to 120 Hz), and high frequency oscillation (HFOs, 120 to 150 Hz). We then calculated the relative power and the z-scored power for each frequency band. The relative power was calculated as the mean power within the band of interest divided by the total power in the power spectrum of the LFP. The z-scored power used prestimulus data as the baseline to calculate the mean and the SD. Phase-Amplitude Coupling. To measure how low-frequency phase modulates high-frequency amplitude, we computed a modulation index (MI) following established method previously described (48, 82, 83). We combined S, N, and S’ stimulus trials as social conditions and combined E and C stimulus trials as nonsocial conditions. We focus on the phase-amplitude coupling during the social investigation in the 0 to 2 s period. To mitigate the filter-induced artifacts, we used longer data period from −0.5 to 2.5 s and filtered the LFP data using a FIR filter within the respective frequency bands of interest. For low frequency of 2 to 20 Hz, we used a step of 2 Hz with bandwidth 2 Hz. For high frequency 30 to 150 Hz, we used a step of 5 Hz. The bandwidth of the amplitude signal was twice the center frequency of the phase signal, which ensured that the side peaks generated when the amplitude signal was modulated by the lower-frequency phase signal were included in the analysis (84). Hilbert transform was then performed to extract phases of low-frequency and amplitudes of high-frequency. Last, the initial and final 0.5 s of each trial were truncated, resulting in a final time window of 0 to 2 s. To evaluate the significance of the raw MI, we computed 100 surrogate indexes by randomly combining the phase and amplitude signals from shuffled trials. A normalized MI was then obtained via z-score transformation utilizing the mean and SD derived from the surrogate data. The LFPs recorded from the Neurologger system were used for the phase-amplitude coupling analysis. Spike Data Processing. Offline spike sorting was performed with Offline Sorter 4.0 (Plexon Inc.). We used well isolated units using criteria (L-ratio < 0.05, isolation distance > 20). A refractory period of 1 ms was enforced to exclude spurious spikes. A total of 304 neurons in the vHPC and 324 neurons in the mPFC were isolated. Single-unit spike data were binarized into 1 ms resolution point processes with 0/1, and firing rates were computed using 100 ms bins. Neurons with mean firing rate below 0.5 Hz were excluded. To investigate the neuronal response in an investigation trial, the behavioral trials were manually annotated frame-by-frame using Cineplex Editor software. For each investigation trial, when the animals actively interacted the cage, the onset time of the active sniffing was identified and labeled. Investigation trials with an investigation duration exceeding 2 s were used for analysis. Neuronal response was assessed by comparing the firing rates in the prestimulus period of −2 to 0 s and during the stimulus period of 0 to 2 s firing rates. A nonparametric signed-rank test was used across all stimulus trials. A P value less than 0.05 was used as the criterion for identifying responsive (enhanced or suppressed) neurons. To characterize stimulus-specific response, trial-averaged firing rate changes were normalized to baseline activity using z-score transformation. The selectivity index was calculated as (FS-FE)/(FS+FE) in sociability phase and (FN-FS’)/(FN+FS’) in social novelty phase (18). FS, FE, FN, and FS’ means the mean firing rates of the four stimuli. We classified all recorded units in both mPFC and vHPC as putative excitatory or putative inhibitory neurons based on spike waveform and firing rate characteristics. Units with a trough-to-peak (half-wave) duration shorter than 0.4 ms and average firing rates greater than 10 Hz were classified as putative inhibitory neurons; all other units were classified as putative excitatory neurons. Information Analysis. Information-theoretic measures were derived from Shannon entropy, implemented as introduced previously (30). In the three-chamber social task, the subject mouse was exposed to a series of stimuli s1,s2,…sNs, belonging to distinct categories (S, E, S’, N), while neural population activity r was recorded during each stimulus presentation. We characterized the information of neural population based on Shannon’s information theory, under which the mutual information I(S;R) between stimulus S and response R, is defined to indicate the reduction of uncertainty of the stimulus given observations about the stimulus, i.e.,IS;R=HR-H(R|S), where H(·), called the entropy, measures the uncertainty of a random variable. H(R) was defined asHR=-∑rp(r) log2 P(r). P(r) being the probability of observing r in response to any stimulus. The second entropy term H(R|S) quantified the response variability of trial-to-trial differences in the responses to the same stimulus. It was defined asHR|S=-∑s=1NsP(s)∑rPr|s log2 P(r|s), where P(r|s) was the probability of observing response r when stimulus s was investigated. P(s) was the probability of investigation of stimulus s, and was defined as Ps=N(s)/∑N(s), N(s) being the number of trials for stimulus s. To perform the information analysis, within a recording session, population vectors were constructed from simultaneously recorded neurons for each stimulus belonging to one of the four types of S, E, S’, N. Spike counts were aggregated across time bins of 75, 100, and 125 ms. Spike counts within each bin were discretized into three discrete levels as reported previously (16). To compute information metrics, population vectors comprising eight neurons were used. Recording sessions with less than eight neurons were excluded and sessions with larger populations were subsampled by randomly selecting eight neurons. To address potential estimation bias in information measures, the Panzeri-Treves bias correction method was applied (85). Population Decoding Using Spikes. Population vectors were constructed for each investigation trial by binning spike counts in 100 ms time windows. These vectors were then pooled across all trials and partitioned into training and test sets for subsequent decoding analysis. A naïve Bayesian decoder was trained under the assumption of conditional independence between neurons within a population vector. Baseline was removed for each trial by subtracting the mean prestimulus spike counts. The baseline-corrected population vectors served as features, with corresponding stimulus identities as true labels. Decoding performance was evaluated via leave-one-out cross-validation. For each iteration, one sample was held out as the test set, and a model trained on the remaining samples predicted the held-out stimulus label. This process was repeated until all samples were predicted. Stimulus-specific decoding accuracy was calculated as the percentage of correctly predicted labels. To account for unequal trial numbers across stimuli, balanced decoding accuracy was computed as the mean accuracy across all stimulus conditions; that is,Balanced Accuracy=∑i=1knii/nik, where nii denotes the number of samples of stimulus si that were correctly classified, ni is the total number of trials for stimulus si, and k is the total number of stimulus conditions. We tested the decoding performance using n-dimensional population vectors consisting of different numbers of neurons. For dimension n, sessions with fewer than n neurons were excluded from analyses. If sessions exceeded n neurons, neurons were ranked by the variance of firing rate, and the top n neurons were selected to construct the n-dimensional population vectors. Spike–LFP Synergy Decoding. To evaluate whether integrating LFPs with spike trains improves decoding performance across the four stimuli (E, S, N, and S’) compared to using spike data alone, we binned spikes and LFPs simultaneously recorded from the same tetrode using the Plexon system into 200 ms time bins. Spike and LFP features were then concatenated within the same recording session. Sessions with fewer than eight neurons were excluded. For sessions exceeding eight neurons, neurons were ranked by firing rate variance and the top eight neurons were selected to construct population vectors. LFP features included raw power, relative power, and z-scored power from all six frequency bands. After subtracting mean baseline activity (−2 to −1 s), all binned population vectors across trials were pooled together. For spike–LFP synergy decoding, PCA was applied to concatenated spike–LFP data per session to extract dominant neural covariation patterns. We retained the first eight principal components, defining an 8-dimensional neural subspace. A Naive Bayesian decoder was trained on baseline-corrected binned population vectors (−0.5 to 2 s relative to stimulus onset) using stimulus type as true labels. Decoding performance was assessed via leave-one-out cross-validation. Each time, one sample bin served as the test set while the model trained on the remaining bins predicted its stimulus label. This procedure was repeated until all bins had been predicted once. Stimulus-specific decoding accuracy was calculated as the percentage of correct predictions per condition. To mitigate bias from unequal trial counts, balanced decoding accuracy was computed as the mean accuracy across all four stimulus conditions. Spike–LFP Correlation. To investigate correlations between the power of LFP and population spiking dynamics, we implemented CCA to quantify the similarity between spikes and LFPs (37, 38). We performed this correlation analysis separately for one LFP channel in each session, generating six canonical correlation coefficients corresponding to delta, theta, beta, low gamma, high gamma, and HFOs bands. The baseline activities of spikes and LFPs from the presocial period (−2 to −1.4 s) were removed. Neural features were extracted through Gaussian kernel smoothing (σ = 50 ms) of square-root-transformed binned data (200 ms bins), computing single-neuron firing rates and LFP band power. PCA applied to spike trains defined a five-dimensional neural subspace retaining the leading components. For each frequency band, we derived canonical variates to maximize the correlation between linear projections of the five-dimensional neural spiking subspace and one-dimensional LFP power. Statistical significance of correlation coefficients was evaluated against surrogate datasets generated via tensor maximum entropy methods (86) using permutation testing (1,000 iterations) to determine whether actual correlations exceeded surrogate correlations. Single-Trial Population Decoding. To determine whether vHPC or mPFC neural activity contained sufficient information for decoding S and N stimulus identities through the time-course of social investigation, we implemented single-trial population decoding using pseudopopulations (39, 42, 43, 87) constructed from neurons or LFPs across all recording sessions. This approach leverages behavioral consistency across mice and yields results equivalent to simultaneously recorded neural ensembles (43, 88, 89). We randomly scrambled the order of trials for all neurons or LFPs included in pseudopopulations. For spike-based decoding, we aligned the spike data to the onset of social interaction and segmented into 200 ms bins. Only these sessions containing at least six trials for each social stimulus type were included into the analysis. Baseline normalization was performed by subtracting the mean presocial activity (−2 to −1 s) from all subsequent timepoints. Then we balanced the number of trials (six per stimulus) and randomly subsampled 90 neurons from the vHPC or mPFC across different mice to construct 90-dimensional pseudopopulation datasets. This subsampling was repeated 1,000 times. A Naive Bayesian decoder was trained using the baseline-corrected population vectors from −0.5 to 2 s relative to the onset of social interaction, with stimulus identities as labels. Decoding performance was evaluated via leave-one-pair-out cross-validation. Each time, one matched pair of S-N trials (each trial contains 12 bins) was set aside for testing. The model was trained on the remaining bins to predict the labels, ensuring that the training and testing datasets contained an equal number of S and N bins. This procedure was repeated until all bins had been predicted once. Stimulus-specific decoding accuracy was calculated as the percentage of correct predictions for each condition. Time-dependent decoding was performed by predicting the label at each time bin in a trial. Balanced decoding accuracy was the mean accuracy across the two stimuli. To establish a chance-level control, we generated 90-dimensional null distributions by randomly shuffling the trial labels. The shuffling procedure was also iterated 1,000 times. Statistical comparisons of decoding performance between the vHPC and mPFC was conducted during the social period (0 to 1.5 s) across different brain areas using Mann–Whitney tests. Regarding LFP-based decoding, we first aligned LFP data in a similar manner and applied the same binning and inclusion criteria. We also used six balanced trials per stimulus. There were six frequency bands, each associated with three types of power (raw power, relative power, and z-scored power) in each LFP channel. Subsequently, we randomly selected five channels from all recordings to construct a 90-dimensional LFP pseudodatasets, analogous to the spike pseudopopulation. All decoding, validation, and statistical procedures used in spike-based analyses were then applied to LFP pseudodatasets. Signal-Noise Angle. We quantified the angle (θ ∈ [0, π/2]) between signal axis and noise axis in neural population space (90). The signal axis was defined as the vector connecting mean population responses to stimuli S and N. The noise axis was computed as the first principal component of trial-to-trial variability in population activity of all social stimulus trials. Using single-trial population decoding data inputs, 1,000 signal-noise angles were computed separately for vHPC and mPFC for the statistical analysis. Statistical Analysis. Statistical analyses were performed using Graphpad Prism 9.5.0. For comparison of the neuronal responses across different stimuli, the Kruskal–Wallis test was used followed by Dunn’s multiple comparison correction. For LFP responses across different stimuli, one-way ANOVA tests followed by post hoc Bonferroni multiple comparisons were used. For comparison of distributions, the Kolmogorov–Smirnov test was used. For LFP responses between pre- and during stimulus periods, paired ttests were used. For information and decoding analysis, two-way ANOVA was used followed by Bonferroni correction. For comparisons in single-trial decoding analysis, the Mann–Whitney test was used. The details of the statistical analysis including exact statistical values and n values are provided in the figure legends. Supplementary Material Appendix 01 (PDF)