Abstract

Neural circuits supporting innate behaviors, such as feeding, exploration, and social interaction, intermingle in the lateral hypothalamus (LH). Although previous studies have shown that individual LH neurons change their firing relative to the baseline during one or more behaviors, the firing rate dynamics of LH populations within behavioral episodes and the coordination of behavior-related LH populations remain largely unknown. Here, using unsupervised graph-based clustering of LH neurons firing rate dynamics in freely behaving male mice, we identified distinct populations of cells whose activity corresponds to feeding, specific times during feeding bouts, or other innate behaviors—social interaction and novel object exploration. Feeding-related cells fired together with a higher probability during slow and fast gamma oscillations (30–60 and 60–90 Hz) than during nonrhythmic epochs. In contrast, the cofiring of neurons signaling other behaviors than feeding was overall similar between slow gamma and nonrhythmic epochs but increased during fast gamma oscillations. These results reveal a neural organization of ethological hierarchies in the LH and point to behavior-specific motivational systems, the dysfunction of which may contribute to mental disorders.

Significance Statement

The lateral hypothalamus (LH) is pivotal for regulation of innate behaviors, yet the organization of LH neuronal populations at fine temporal resolution within behavioral episodes remains unknown. Such knowledge would be crucial for understanding the contribution of LH to different innate behaviors. Here, we identified distinct groups of LH cells active at specific times within feeding bouts and, additionally, populations more continuously active during feeding, exploration, or social interaction. Cells from feeding-related populations coalesce during LH slow gamma oscillations, while fast gamma also promotes assembly across multiple behavioral populations. Our findings suggest that appetitive behaviors and phases of consummatory acts are supported by distinct LH populations. Dysfunction of their interaction and plasticity during network oscillations may contribute to eating disorders.

Introduction

In humans and other mammals, innate behaviors maintain physiological homeostasis, organize interactions with conspecifics to ensure safety and reproduction, and facilitate learning through exploratory responses to novelty. Innate behaviors are regulated by the hypothalamus, which processes various chemical, sensory, and cognitive control signals (Tinbergen, 1951; Swanson, 2000). Hypothalamic output to forebrain and midbrain regions is mainly conveyed by the anatomically complex and functionally diverse lateral hypothalamus (LH), crucial for the coordination of multiple innate behaviors (Bonnavion et al., 2016; Herrera et al., 2017; Stuber, 2023; C. Chen et al., 2024). Seminal studies showed that LH lesions dramatically reduce food intake and lead to starvation (Anand and Brobeck, 1951; Grossman et al., 1978), while the LH stimulation elicits feeding (Hoebel and Teitelbaum, 1962; Margules and Olds, 1962). These effects are brought about by functionally and neurochemically distinct LH neurons which receive appetitive and anorectic signals from within the hypothalamus, basal forebrain, and hindbrain regions (Jennings et al., 2013, 2015; Nieh et al., 2015; O'Connor et al., 2015; Carus-Cadavieco et al., 2017), respond to glucose (Oomura et al., 1974), and project to various regions within the feeding circuitry (Sternson and Eiselt, 2017). Separate populations of LH neurons are involved in food intake and reward, depending on their molecular identity and projections (Marino et al., 2020; Siemian et al., 2021b). Circuits mediating interactions with conspecifics also involve LH, where cells with partially different molecular identities are excited or inhibited during social behaviors (Nieh et al., 2016; Bai et al., 2023; Petzold et al., 2023; C. Chen et al., 2024). Interactions within and between these neuronal populations allow behaviors to be prioritized according to homeostatic needs and sensory cues (Petzold et al., 2023; Barbano et al., 2024). Studies of the other major output of the hypothalamus, the paraventricular nucleus, have highlighted the possibility that different subpopulations of cells with overlapping molecular profiles may be involved in disparate phases of food intake including its initiation, increasing satiation and meal termination (C. Li et al., 2019; M. M. Li et al., 2019). Orosensory and gastrointestinal distention signals are processed by different populations of a key hypothalamic afferent region in the brainstem, generating sequential negative feedback signals during food intake (Ly et al., 2023). It is currently unknown whether the hypothalamus features feeding phase-specific cell populations and, of particular importance for LH, whether they are also involved in other innate behaviors.

Network oscillations influence the timing of neuronal discharge and coordinate signaling within and across brain regions (Gray et al., 1989; Buzsáki and Watson, 2012) including the hypothalamus (Carus-Cadavieco et al., 2017; Samerphob et al., 2019; Tsurugizawa et al., 2019; C. Chen et al., 2024). Gamma oscillations have been implicated in feeding behavior. Food seeking in sated mice involves coordinated gamma rhythmic activity in the LH, lateral septum, and prefrontal cortex (Carus-Cadavieco et al., 2017) and is characterized by increased LH–ventral striatum gamma band coherence upon repeated contact with palatable food (Samerphob et al., 2019). Gamma local field potential (LFP) oscillations at the location of primary energy sensor neurons in the ventromedial hypothalamus were reduced in energy-deficit states in fasted mice (Tsurugizawa et al., 2019) and in mice with weight loss induced by anorexia during chronic treatment with the chemotherapeutic cisplatin (Sun et al., 2021). The impact of gamma oscillations on neuronal activity in the hypothalamus has been revealed at the resolution of single oscillation cycles (Carus-Cadavieco et al., 2017), where populations of cells preferentially fire at certain oscillation phases, thereby jointly determining efferent signaling (O'Keefe and Recce, 1993; Dejean et al., 2016). On the other hand, network synchronization during gamma oscillations involves increased activity in the population of participating neurons (Csicsvari et al., 2003; Ainsworth et al., 2012; Carus-Cadavieco et al., 2017; Sherfey et al., 2018). The joint participation of neurons in oscillations, indicated by the degree of their correlated discharge, supports their direct and polysynaptic effects on each other and on their efferent cells and influences synaptic plasticity (Csicsvari et al., 2003; Kim et al., 2016; Galuske et al., 2019). However, the organization of LH cell assemblies according to their behavioral specificity during gamma oscillations, potentially supporting the coordination of functionally similar or diverse neurons, remains unknown.

Here we used spectral clustering to analyze firing rate dynamics of LH neurons during different behaviors. Graph-based spectral clustering, which outperforms traditional methods such as k-means and hierarchical clustering (Namboodiri et al., 2019; Kariotis et al., 2021), is particularly useful for classifications of nonlinear dynamics of functionally heterogeneous neural populations (Gouwens et al., 2019; Hirokawa et al., 2019; Namboodiri et al., 2019). We found that separate populations of LH neurons recorded in a free choice model are active at particular times during feeding episodes. Other populations of LH cells were more active during social interaction and exploration of novel objects than during feeding. LH gamma oscillations were associated with either increased or decreased cofiring of neurons as compared with nonrhythmic epochs. LH cells from different groups showed distinct assembly patterns during slow and fast gamma oscillations suggesting that distinct types of neuronal synchronization in the LH specifically support feeding and facilitate the coordination of behavior-specific cells.

Material and MethodsExperimental procedures

Experimental data in this study have been used in C. Chen et al. (2024) where data collection is described in detail. Briefly, 10- to 25-week-old Vgat-ires-Cre knock-in mice (The Jackson Laboratory, used also in optogenetic stimulation experiments after recordings of spontaneous behaviors) and C57BL/6 male mice were housed under standard conditions with a 12 h light/dark cycle. The firing rate during behaviors of cells recorded in three C57BL/6 and one Vgat-ires-Cre mice (The Jackson Laboratory) was compared to ensure the absence of electrophysiological differences between the mouse lines. Prior to the experiments, mice were handled and habituated to the experimental enclosure for 3–5 d. Four-shank silicon probes (B32, NeuroNexus Technologies) were mounted on custom-made microdrives and stereotactically implanted in the LH at the following coordinates: AP, −1.58; ML, 0.8; medial shank; and DV, 5 mm. Reference and ground electrodes were screws above the cerebellum. The implant was shielded with copper mesh and secured to the skull with dental acrylic.

Data acquisition and preprocessing

The recording setup consisted of a custom enclosure with two interconnected compartments (LWH 25 × 30 × 20 cm each), with access to water, food, a conspecific, and a novel object (C. Chen et al., 2024). Mice moved ad libitum within the enclosure during recordings. Silicon probes were connected to a Neuralynx preamplifier to reduce cable movement artifacts. Signals were amplified, bandpass filtered (1 Hz–8 kHz), and acquired continuously at 32 kHz using Digital Lynx system (Neuralynx). Behavioral recordings were performed by four cameras capturing movement from different angles at 25 Hz (Motif, Loopbio). Ethograms were created using Adobe Premiere and frame-by-frame analysis of synchronized multiangle videos. Feeding was scored when the resident mouse consumed (i.e., ate) food pellets, with the minimal duration of scored feeding bouts of 4.8 s. Sniffing or tracking an intruder mouse was scored as social interaction. Novel object exploration included behaviors such as sniffing, gnawing, touching, or climbing a novel object.

Electrophysiological signals were preprocessed using the Neurophysiological Data Manager (NDManager, http://neurosuite.sourceforge.net/) and analyzed using custom-written MATLAB scripts (MathWorks). In a high-pass–filtered signal, action potentials (spikes) were identified, and spike waveforms were represented by the first three principal components and action potential amplitudes. Automatic spike sorting (https://github.com/klusta-team/klustakwik) was followed by manual cluster adjustments using auto- and cross-correlations of spike trains, Mahalanobis isolation distance between clusters (64 ± 2 across four mice), and visual comparison of waveform profiles across channels (see also Extended Data Fig. 1b in C. Chen et al., 2024). The data were obtained from an average of 82 recordings by individual probe shanks per mouse, i.e., separate spike sorting sessions, resulting in 5 ± 1 sorted units per shank/recording.

Firing rate dynamics

For individual cells, the firing rate was estimated by the convolution of spike trains with a Gaussian kernel of the size of 500 ms. To obtain the firing rate dynamics vectors for each cell during each of the innate behaviors in a temporally aligned fashion, we standardized episodes of different durations to the average duration of episodes in a recording session (Hirokawa et al., 2019). Specifically, the average duration of behavioral episodes of a given behavior in a recording session was computed, and the number of bins corresponding to the 10 ms resolution in the average episode was estimated. All behavioral episodes in a session were then partitioned using the estimated number of bins. The average firing rate for each bin was computed across different episodes of a given behavior. The firing rate vectors corresponding to three distinct behaviors—feeding, social interaction, and novel object exploration (13,914, 1,774 and 2,039 bins, respectively)—were concatenated for each cell and normalized using z-score transformation. To mitigate the impact of the behaviors duration variability and yield robust clustering (Jolliffe, 2002), we further projected the z-scored firing rate vectors onto a lower-dimensional space through principal component analysis prior to clustering. The first 128 principal components were retained, capturing 90% of the variance of the firing rate vectors.

Spectral clustering

Clustering analysis was performed on the 128-dimensional firing rate vectors of LH cells using the Scikit-learn package. Spectral clustering identifies data partitions with fewer intergroup connections and denser intragroup links and is particularly stable in higher-dimensional datasets (Von Luxburg, 2007; Hirokawa et al., 2019). The affinity matrix for clustering was computed using a k-nearest neighbor (k-NN) connectivity matrix and the cosine metric as the similarity measure for the firing rate vectors. To determine the optimal combination of the number of clusters and k-NN, which influence graph construction and thus the clustering output, we used stability analysis through bootstrapping without replacement 100 times, randomly selecting 90% of the data for each iteration (Hirokawa et al., 2019). To evaluate the stability of clustering for various combinations of cluster numbers and k-NN, we computed the adjusted Rand index (ARI; Steinley, 2004) and adjusted mutual information (AMI) score (Vinh et al., 2010). These independent metrics assess clustering stability across subsamplings as the proportion of cells with consistent clustering and the mutual dependence between the clusterings, respectively. Optimal clustering parameters (number of clusters and k-NN) were then determined as the maximal mean and mean adjusted for variance (mean/standard deviation ratio) of ARI and AMI across bootstrapped distributions. Importantly, the selected parameters enabled reliable clustering as opposed to random assignment of cells to clusters, as indicated by both the ARI and AMI score being well above zero (Steinley, 2004; Vinh et al., 2010).

While stability analysis evaluates the robustness of clustering, it does not reveal its quality, i.e., the similarity of data within clusters and their dissimilarity between clusters. For instance, low intracluster correlations may be an indication of subclusters within a cluster and the need for further segmentation. Clustering quality was estimated by Pearson’s correlations between the first five principal components of firing rate vectors for cells from the same or different clusters.

To further ensure robust clustering quality, we calculated silhouette scores (Rousseeuw, 1987), which measure the similarity of individual data samples to their assigned clusters compared with other clusters using intracluster and nearest-cluster distances, across 100 clustering iterations. Neurons consistently assigned to their respective clusters, i.e., with a silhouette score above 0 for at least 95 of the iterations, were considered for further analysis.

LFP analysis

The LFP was obtained by downsampling of the wide-band signal to 1,250 Hz using the NDManager. Slow and fast gamma oscillations were detected in the 30–60 and 61–90 Hz bandpass filtered, rectified, and smoothed signal (Carus-Cadavieco et al., 2017). Events with the amplitudes exceeding 2 SD above the noise mean for at least 25 ms were detected. The beginning and the end of oscillatory epochs were designated at times when the amplitude fell below 1 SD (Csicsvari et al., 2003). To obtain the times of nonrhythmic epochs, delta (2–4 Hz), theta (5–10 Hz), beta (15–30 Hz), and gamma (up to 120 Hz) as well as faster oscillations in the 120–200 Hz band were detected (Bender et al., 2015; Carus-Cadavieco et al., 2017; C. Chen et al., 2024). Epochs without detected oscillations exceeding 0.5 SD above the noise mean were marked as nonrhythmic. The multitaper method (NW, 3; window length of 1,024) was used to compute power spectral density.

Cross-correlations

Cross-correlograms (CCG) of cells recorded by different shanks of a silicon probe were computed using spikes during gamma or nonrhythmic epochs as trigger events and normalized by the cumulative CCG spike count. Cofiring probability was computed as the mean firing probability within ±15 ms of the trigger event. The cofiring ratio (CFR) was calculated for each cell pair as the ratio of the cofiring probability during nonrhythmic epochs to that during gamma oscillations (Fig. 6c). To robustly compare the cofiring of LH cells during gamma oscillations versus nonrhythmic epochs, we computed the CCG of the same set of cells after shifting the trigger spike time by a random variable from a uniform distribution in the [−25 25] s interval. This procedure was repeated to obtain 1,000 shuffled CCGs and control CFRs for each cell pair.

Statistical analysis

Python (https://www.python.org/) and MATLAB (MathWorks) were used for statistical analyses. Two-tailed t tests and Mann–Whitney or Kolmogorov–Smirnov (KS) tests were used for individual two-group comparisons, depending on the normality of the distributions. ANOVA, Tukey–Kramer post hoc tests, or multiple two-group tests with Bonferroni’s α correction were used for multiple group comparisons. Descriptive statistics are presented as mean ± SEM unless otherwise indicated.

ResultsUnsupervised classification of LH cells firing rate dynamics

The firing of 1,349 LH cells was recorded in four male mice using movable silicon probes in a free choice model during eating standard chow, interacting with a younger mouse of the same sex, or exploring a novel object (hereafter, exploration). These cells were recorded in parallel with at least one more cell at another silicon probe shank (see joint firing of LH cells during gamma oscillations and nonrhythmic epochs). Individual LH cells showed variable firing rate dynamics during the three behaviors. We examined the high-dimensional firing rate vectors using an unsupervised approach, spectral clustering, which identifies communities of cells connected by a similarity graph (illustrated for an example subset of cells in Fig. 1a). To estimate the optimal number of clusters and k-NN, stability analysis was performed using ARI for different combinations of parameters yielding high stability scores for the number of clusters between 2 and 11 with k-NN >20 (Fig. 1b). Further assessment of the AMI and ARI distributions indicated the most stable clustering using two followed by seven number of clusters (Fig. 1c). To compare the clustering quality between two- and seven-cluster approaches (using k-NN = 21), we computed the intracluster correlation ratio, which indicates the correlation of the data within clusters compared with correlation of the data between clusters. The dataset segmented into seven clusters showed higher intracluster correlations than with the two-cluster approach, also in relation to intercluster correlations (intracluster correlation ratio two vs seven clusters; t6 = 6.24; p = 0.0007; intracluster correlation two vs seven clusters; t(6) = 15.91; p < 0.0001; one-sample t tests; Fig. 1d–f). Thus, segmentation of the data into seven clusters was highly reproducible and properly represented the similarity of firing rate dynamics between LH neurons.

Figure 1.

Clusters of LH cells revealed by the spectral clustering of firing rate dynamics during innate behaviors. a, Scheme of the spectral clustering of firing rate vectors of 15 representative neurons during feeding (F), social interaction (S), and exploration (E). PCs, principal components. b, c, Clustering bootstrap stability analysis. b, ARI, mean of 100 bootstrap sessions, for different numbers of clusters and k-NN (n = 917 cells from 4 mice). Dashed frame, Number of cluster range with high clustering stability, estimated in c. c, Variance-adjusted ARI and AMI (mean across bootstraps/SD) for different number of clusters, mean ± SEM across k-NN from 2 to 38. Red dashed lines, Two most stable clusterizations. d, Intra- to intercluster Pearson’s correlation ratio for two- versus seven-cluster analysis. One-sample t test, seven clusters versus the mean of two clusters (dashed line); t(6) = 6.24; ***p = 0.0007. e, Intracluster correlations for two- versus seven-cluster analysis. One-sample t test, seven clusters versus the mean of two clusters (dashed line), t(6) = 15.91; ****p < 0.0001. f, Correlations of firing rate feature vectors (first 5 PCs) of 549 cells grouped into seven clusters show high intracluster correlations (squares near the diagonal) and low intercluster correlations.

Distinct LH populations signal feeding phases and further innate behaviors

To functionally characterize LH cells clustered on the basis of similar firing rate dynamics, we investigated their activity in relation to innate behaviors. We examined their firing rate patterns during feeding, exploration, and social interaction. Remarkably, even though behavior was not a predefined label in our unsupervised classification, cells in different clusters (henceforth, populations) fired closely matching these behaviors (Fig. 2a). The cumulative probability distributions of the time at which cells fired at their highest rate during their preferred behavior (i.e., the one with the highest peak firing rate across the three behaviors) were not uniform for cells from five populations active during a feeding bout (Fig. 2b; one-sample KS test; adjusted α = 0.0023; p < 0.0001). Four of these feeding-related populations clearly corresponded to temporally different phases of a feeding bout (Fig. 2a,b). These distinct feeding phase populations were sequentially active from the feeding onset (FOn), through early and late feeding (EF and LF), to feeding offset (FOff). Cells from the exploration and social interaction populations (Exp and Sol) fired homogeneously during these respective behaviors (one-sample KS test; adjusted α = 0.0023; exploration, p = 0.78; social, p = 0.22) and, together with a similar, fifth, feeding population (Fd), changed their average firing rate as a function of behavior (two-way ANOVA; population × behavior; F(4,846) = 2.01; p = 0.09; population, F(2,846) = 1.37; p = 0.25; behavior, F(2,846) = 0.30; p = 0.74). Importantly, the populations were not only evident in the combined data from multiple mice but also within individual animals (Fig. 2c). Moreover, within each animal, the proportion of cells in each identified population was similar except for one animal with a lower number of recorded neurons (Fig. 2c).

Figure 2.

Decoding of feeding phases and innate behaviors by LH cells's firing rate. a, Normalized firing rate dynamics of cells (rows) in each cluster (with margins indicated by white horizontal lines) during feeding, social interaction, and novel object exploration (columns). Data for individual cells are sorted according to the relative time of the maximal discharge in the concatenated firing rate vectors of the three behaviors. Numbers indicate average and, in brackets, peak firing rate (Hz) of cells during each behavior (mean across cells in a cluster). Clusters labeled based on the behavior when the average peak firing rate of cells was the highest: FOn, n = 64 cells from three mice; EF, n = 68 cells from four mice; LF, n = 60 cells from three mice; FOff, n = 72 cells from three mice; Fd, n = 123 cells from four mice; Sol, n = 95 cells from four mice; Exp, n = 67 cells from four mice. The color code of neuronal populations is maintained throughout the figure. Note that, unlike FOn, EF, LF, and FOff populations that are active at specific phases of feeding bouts, the instants of maximal discharge in homogeneously active Fd, Sol, and Exp reflect much less pronounced elevations of the firing rate (Fig. 4). b, Empirical cumulative distribution of cells’ peak firing times in the functional populations during behaviors of their maximal discharge, i.e., during feeding, exploration, and social interaction. One-sample KS test for the standard uniform distribution (dashed line), adjusted α = 0.0023; FOn, D = 0.51; EF, D = 0.33; LF, D = 0.33; FOff, D = 0.38; Fd, D = 0.26; p < 0.0001 for all feeding-related populations; Sol, D = 0.11; p = 0.22; Exp, D = 0.08; p = 0.78. c, Functional LH populations in different mice. Numbers in bars, n of cells. d, Average duration of behaviors during which cells were recorded, mean ± SEM across n = 549 cells from four mice. e, Empirical cumulative distribution of feeding bouts duration in recordings of cells from feeding-related clusters. Median of distributions, vertical dotted lines, FOn, 19.7 s; EF, 22.2 s; LF, 22.3 s; FOff, 25.7 s; Fd, 24.6 s; Mann–Whitney test; adjusted α = 0.005; not different between clusters; p > 0.13. X axis, log scale.

Figure 3.

Firing patterns of example LH cells from functional populations during episodes of innate behaviors. Columns, three example cells from each population. Each panel includes the following: top, raster plot of spike train; bottom, instantaneous firing rate over the duration of the behavior.

Feeding bouts were on average six times longer than exploration or social interaction episodes (Fig. 2d). Therefore, different feeding phase populations could be activated with progressively longer feeding or, alternatively, signal phases of feeding bouts, regardless of their duration. The duration of feeding episodes was consistent across different feeding-related populations (Mann–Whitney test; adjusted α = 0.005; p > 0.13; Fig. 2e). Thus, those LH populations signal phases rather than the duration of feeding.

Differences in the timing of neuronal activity during feeding and across behaviors evident in individual cells (Figs. 2a, 3) were further elucidated by examining the average firing rate dynamics. During feeding, cells from the feeding populations showed a consistently high level of activity, signaling the relative time within feeding episodes (Fig. 4a,b, Fig. 2b). In contrast, during exploration and social interaction, cells from all populations showed temporally homogenous activity, with cells from the exploration and social interaction populations firing at the highest firing rates during these behaviors (Fig. 4a,b). We next examined the activity of individual LH cells from each population during different behaviors. Activity during feeding was strongly inversely correlated with firing during exploration and social interaction for all populations (Pearson's correlation during feeding vs social interaction; −0.72 ≤ r ≤ −0.49; p < 0.0001; and vs exploration, −0.84 ≤ r ≤ −0.61; p < 0.0001; Fig. 4c). However, firing during social interaction did not typically predict the response during exploration, suggesting an independent representation of these behaviors in LH (Fig. 4c).

Figure 4.

Dynamics and relative activity of functional LH populations during innate behaviors. a, The firing rate of LH populations during feeding (F), social interaction (S), and exploration (E) (a.u., z-scored during the three innate behaviors, mean ± SEM across cells). The color code of neuronal populations is maintained throughout the figure. b, Distributions of the average firing rate (kernel density estimation) during behaviors. One-way ANOVA; mean firing rate during feeding, F(6,542) = 267.81; during social interaction, F6,542 = 207.46; during exploration, F6,542 = 137.00; p < 0.0001 for all the behaviors. Dashed line, mean of z-scored firing rate distributions. c, Correlation of the mean firing rate (a.u., z-scored) between the three behaviors for feeding phase (left plots) and other behavior-specific populations (right plots). Contours represent the probability density estimated for each population. Pearson’s correlation between F and S: FOn, r = −0.63; EF, r = −0.49; LF, r = −0.57; FOff, r = −0.72; Fd, r = −0.56; Sol, r = −0.65; Exp, r = −0.69; between F and E: FOn; r = −0.66; EF, r = −0.74; LF, r = −0.61; FOff, r = −0.78; Fd, r = −0.79; Sol, r = −0.80; Exp, r = −0.84; p < 0.0001 for all r; between S and E: LF, r = −0.31; p = 0.0155; other correlations, p > 0.05.

Next, we examined the anatomical localization of neurons from the recorded populations within the LH. For this analysis, 505 cells recorded from three mice in close anteroposterior planes (1.34–1.58 posterior to the bregma, Fig. 5a) were used. We computed two-dimensional probability density distributions of the recorded cells in mediolateral versus dorsoventral aspects, i.e., according to the shanks of a silicon probe and to their advancement, respectively. The probability density distributions were computed for the individual LH populations and for all the populations combined, the difference between the two probability densities at each location indicates the location-specific enrichment with cells from an individual population (Fig. 5b). FOn, EF, and FOff populations featured correlated anatomical distributions; they were enriched in the lateral aspects of LH without clear dorsoventral differences (Fig. 5b,c). In contrast, the localization pattern of LF cells active during LF differed from FOn, EF, Fd, and Exp (r close to zero) and moderately correlated with FOff and Sol (r = 0.22, 0.27, respectively). LF cells were enriched in the dorsomedial and ventrolateral LH (Fig. 5b). The localizations of Fd, Sol, and Exp, homogenously active during individual innate behaviors, were distinct (Fd and Sol, r = −0.25; p = 0.0122; Fd and Exp, r = 0.11; p = 0.30; Sol and Exp, r = 0.18; p = 0.07) and poorly correlated with the localization of cells active during phases of feeding bouts.

Figure 5.

Anatomical localization of functional LH populations. a, Reconstruction of recording electrodes positions in the LH; red lines indicate the path of electrodes’ advancement from the implantation until the last recording session in three mice. b, The difference of the probability density between individual LH populations and all recorded cells at each anatomical location. mt, mammillothalamic tract; 3V, third ventricle. c, Pearson's correlation between histograms shown in b. All correlations are significant, p < 0.0422 to p < 0.0001, except for EF versus Sol, Fd, LF versus FOn, EF, Fd, Exp, Fd versus FOff, Exp, Sol versus Exp, p > 0.05. The correlations were computed on 2D anatomical locations parcellated into 38.8 μm DV bins and 70 μm ML bins, 10 × 10 grid.

Feeding phase populations of LH neurons had also distinct electrophysiological properties. On the one hand, cells from feeding populations active in the second half of feeding bouts had lower average firing rates during the whole recording than the FOn and EF cells active at the beginning of a feeding bout or cells from the populations homogeneously active during behaviors (1.7 Hz, LF, 1.9 Hz, FOff cells vs 3.2 Hz, FOn, 2.6–3.5 Hz in other populations; p < 0.05 in individual comparisons by Mann–Whitney test; Table 1). On the other hand, spike width of LF cells was longer than that of EF and of individual populations of homogeneously active cells (p < 0.05 in individual comparisons by Mann–Whitney test; spike width was estimated at 25% of the spike amplitude; Csicsvari et al., 1999).

Table 1.

Physiological properties of functional LH populations

Together, the populations active in the first half and later but not before the termination of feeding bouts may be enriched in GABA and orexin cells, respectively (Mileykovskiy et al., 2005; Karnani et al., 2013), the two anatomically intermingled LH cell types (Karnani et al., 2013). In particular, the localization of LF cells at the analyzed anteroposterior level dorsally and laterally to fornix and physiological properties of LF match those of orexin cells (Mileykovskiy et al., 2005; Karnani et al., 2013; Diaz et al., 2023). Lateral LH aspects enriched with the Sol population (Fig. 5) feature melanin-concentrating hormone (MCH) neurons involved in social behaviors (Allen Brain Atlas, 2011; Yao et al., 2012; Sanathara et al., 2021; Barretto-de-Souza et al., 2023).

Joint firing of LH cells during gamma oscillations and nonrhythmic epochs

Optogenetic studies suggest that populations of hypothalamic neurons strongly influence innate behaviors when activated synchronously (Jennings et al., 2013; Nieh et al., 2016; Carus-Cadavieco et al., 2017; Kohl et al., 2018). During spontaneous behavior, such collective activity is postulated to require sufficiently strong mutual and/or afferent connectivity of behavior-specific cells to enable their temporal coordination (Hebb, 1949). The formation of cell assemblies can be facilitated by gamma oscillations, as suggested by studies in the hippocampus (K. D. Harris et al., 2003; O'Neill et al., 2008). To investigate assemblies of cells from the functional LH populations, we first characterized joint discharge in the whole population of recorded LH neurons during gamma oscillations and epochs without network oscillations, hereafter nonrhythmic epochs (Fig. 6a,b). We computed CCGs of 10,838 pairs of LH neurons recorded by different silicon probe shanks, using spikes fired during locally recorded gamma oscillations or nonrhythmic epochs as trigger events. Preferential cofiring during nonrhythmic as compared with gamma oscillation epochs was then quantified as the CFR during these epochs for each cell pair (Fig. 6c). The majority of cell pairs increased cofiring during gamma oscillations compared with that during nonrhythmic epochs as indicated by CFR < 1 (Fig. 6d–g). This agrees with the overall increase in synchronization during gamma oscillations predicted by computational models (Börgers et al., 2008) and demonstrated in the hippocampus (Csicsvari et al., 2003) and in the lateral septum, the main forebrain input of the LH (Risold and Swanson, 1996; Carus-Cadavieco et al., 2017). At the same time, a remarkably high proportion (∼40%) of LH neurons reduced their cofiring to varying degrees compared with nonrhythmic epochs (Fig. 6d–g). Slow and fast gamma oscillations were associated with different cofiring changes in individual cell pairs. Specifically, cell pairs with low CFR during slow gamma oscillations (i.e., they were more coordinated during slow gamma) showed higher CFR during fast gamma oscillations (i.e., lower coordination during fast gamma; Fig. 6h, data points below the diagonal). Conversely, many cell pairs with high CFR during slow gamma had lower CFR during fast gamma (Fig. 6h, data points above the diagonal).