Neural Coding of Appetitive Food Experiences in the Amygdala

Neurobiology of Learning and Memory 155 (2018) 261–275

Contents lists available at ScienceDirect

Neurobiology of Learning and Memory journal homepage: www.elsevier.com/locate/ynlme

Neural Coding of a Appetitive Food Experiences in the Amygdala Jun Liu

a,b,1

c,1

a

c

d

, Cheng Lyu , Meng Li , Tianming Liu , Sen Song , Joe Z. Tsien

a,⁎

T

a

Brain and Behavior Discovery Institute, Medical College of Georgia, Augusta University, Augusta, GA 30912, USA The Brain Decoding Center Key Laboratory, Banna Biomedical Research Institute, Yunnan Academy of Science and Technology, Xi-Shuang-Ban-Na Prefecture, Yunnan 666100, China c Department of Computer Science and Brain Imaging Center, the University of Georgia, Athens, GA 30602, USA d McGovern Institute for Brain Research and Center for Brain-inspired Computing Research, Department of Biomedical Engineering at Tsinghua University, Beijing, China b

A R T I C LE I N FO

A B S T R A C T

Keywords: Amygdala Food experiences Emotion Pattern-discrimination and categorization Cell assemblies

Real-life experiences involve the consumption of various foods, yet it is unclear how the brain distinguishes and categorizes such food experiences. Despite the crucial roles of the basolateral amygdala (BLA) in appetitive behavior and emotion, how BLA pyramidal cells and interneurons encode food experiences has not yet been well characterized. Here we employ large-scale tetrode recording techniques to investigate the coding properties of pyramidal neurons vs. fast-spiking interneurons in the BLA as mice freely consumed a variety of foods, such as biscuits, rice, milk and water. We found that putative pyramidal cells conformed to the power-of-two-based permutation logic, as postulated by the Theory of Connectivity, to generate specific-to-general neural cliquecoding patterns. Many pyramidal cells exhibited firing increases specific to a given food type, while some other pyramidal cells increased firings to various combinations of multiple foods. In contrast, fast-spiking interneurons can increase or decrease firings to given food types, and were more broadly tuned to various food experiences. We further show that a subset of pyramidal cells exhibited rapid desensitization to repeated eating of the same food, correlated with rapid behavioral habituation. Finally, we provide the intuitive visualization of BLA ensemble activation patterns using the dimensionality-reduction classification method to decode real-time appetitive stimulus identity on a moment-to-moment, single trial basis. Elucidation of the neural coding patterns in the BLA provides a key insight into how the brain’s emotion and memory circuits performs the computational operation of pattern discrimination and categorization of natural food experiences.

1. Introduction Eating foods is one of the most basic experiences conserved across all animal species (Baxter and Byrne, 2006; Craig, 1918; Everitt, Cardinal, Parkinson, and Robbins, 2003), and its dysregulation in neural circuits processing feeding behaviors can be detrimental to health, leading to obesity and diabetes (Morton, Cummings, Baskin, Barsh, and Schwartz, 2006; Morton et al., 2006; Nishijo, Uwano, Tamura, and Ono, 1998; Nectow et al., 2017). Eating foods can produce rich emotions and vivid memories, which often consist of multiple factors, such as visual and olfactory attractiveness, texture, food palatability and social factors (Berridge, 1996; Berthoud, 2004; GuvenOzkan and Davis, 2014; Hsu, Hahn, Konanur, Noble, Suarez, Thai, Nakamoto, and Kanoski, 2015; Kadohisa, Rolls, and Verhagen, 2005a; Peng, et al., 2015; Scott, 2005; Simon, de Araujo, Gutierrez, and Nicolelis, 2006; Zhang et al., 2003), as well as prior experiences

(Adaikkan and Rosenblum, 2015; Carballo-Marquez, Vale-Martinez, Guillazo-Blanch, and Marti-Nicolovius, 2009; Cui, Lindl, Mei, Zhang, and Tsien, 2005; Rampon, et al., 2000). Substantial progress has been made toward the better understanding of the receptors and primary sensory responses to simple odors and tastants (Fontanini, Grossman, Figueroa, and Katz, 2009; Hallem and Carlson, 2006; Scott, 2005; Simon et al., 2006; Zhang et al., 2003) as well as motivational aspects of food-seeking behaviors regulated by reward pathways (e.g., dopamine circuits) (Berridge, 2007; Dayan and Balleine, 2002; Hommel et al., 2006; Palmiter, 2008; Szczypka, et al., 1999; Wise, 2006) and energy homeostatic pathways (e.g., the lateral hypothalamus, arcuate nucleus, etc.) (Carter, Soden, Zweifel, and Palmiter, 2013; de Araujo, et al., 2006; Ko et al., 2015; Stanley, Urstadt, Charles, and Kee, 2011; Tschop, Smiley, and Heiman, 2000; Wu, Clark, and Palmiter, 2012; Nectow et al., 2017). The amygdala – such as the basolateral amygdala (BLA) – is well

⁎

Corresponding author. E-mail address: [email protected] (J.Z. Tsien). 1 Equal contribution. https://doi.org/10.1016/j.nlm.2018.08.012 Received 12 January 2017; Received in revised form 2 August 2018; Accepted 14 August 2018 Available online 17 August 2018 1074-7427/ © 2018 Published by Elsevier Inc.

Neurobiology of Learning and Memory 155 (2018) 261–275

J. Liu et al.

known for its roles in regulating emotion, especially using fear learning and memory models (Cahill, 2000; Davis, 1992; Kluver and Bucy, 1997; LeDoux, 2000; Saunders, 2000). However, using head-restrained animals or the force-feeding of simple tastants (intra-orally via a tube or pipette), researchers have reported that cells in the amygdala or other parts of the brain circuits responded to simple tastants such as NaCl, citric acid or sucrose solutions as well as to other ingestive features related to temperature, texture, palatability, etc. (Fontanini et al., 2009; Hallem and Carlson, 2006; Kadohisa, Verhagen, and Rolls, 2005b; Nishijo et al., 1998; Scott, 2005; Simon et al., 2006; Zhang et al., 2003). These experiments support the notion that the amygdala, especially the BLA, represents a key site for processing appetitive behaviors. Lesion and pharmacological inhibition experiments showed the amygdala is required to generate both specific and general behavioral responses to appetitive experiences (Murray, Gaffan, and Flint, 1996; Petrovich and Gallagher, 2003; Wise, 2006; Wu et al., 2012). Given the fact that real-life experience typically involves the consumption of various foods over the course of a given meal, it is surprising that the question of how the amygdala achieves discrimination, categorization and generalization of various food experiences has not been well investigated. In the present study, we employed large-scale in vivo tetrode recording techniques to examine how the BLA responded as the mice consumed biscuits, rice, milk and/or water during the selfinitiated, freely behaving state. Given the existence of distinct cell types in BLA, we further examined the question of whether/how distinct cell types, such as excitatory pyramidal cells vs. inhibitory interneurons, differ in their responses to natural food experiences. Moreover, we investigated whether BLA cells would respond during rapid behavioral habituation over the course of consuming the same food item. Finally, by combining large-scale recording techniques with dimensionalityreduction methods, we provided the intuitive visualization of real-time BLA neural ensemble traces for various food experiences on a momentto-moment basis.

bucket inside the chronic recording rooms a week prior to surgery. During this period, the animals were also handled daily to minimize stress from human social interaction. On the day of the surgery, the animal was given an intraperitoneal injection of 60 mg/kg ketamine (Bedford Laboratories, Bedford, OH) and 4 mg/kg Domitor (Pfizer, New York City, NY) prior to the surgery. The head of the animal was secured in a stereotaxic apparatus, and an ocular lubricant was used to cover the eyes. The hair above the surgery sites was removed, and Betadine solution was applied to the surface of the scalp. An incision was then made along the midline of the skull. Hydrogen peroxide was placed onto the surface of the skull so that bregma could be visualized. The correct positions for implantation were then measured and marked. Three mice were implanted with 128-channel tetrodes (64-channel bilaterally), two mice were implanted with 64-channel on the left side. For fixing the microdrive headstage, four holes for screws (B002SG89S4, Amazon, Seattle, WA) were drilled on the opposing side of the skull and, subsequently, the screws were placed in these holes with reference wires being secured to two of the head screws. Stereotaxic coordinates used for targeting the BLA are as follows: 1.7 mm posterior to bregma, 3.5 mm lateral, −4.0 mm ventral to the brain surface. Craniotomies for the tetrode arrays were then drilled, and the dura mater was carefully removed. Afterwards, the electrodes were inserted slightly above the BLA. The microdrive was secured to the skull with dental cement, and the reference wires from the connector-pin arrays were soldered such that there would be a continuous circuit between the ground wires from the head screws and those from the connector-pin arrays. Finally, the copper mesh was used to surround the entire headstage to aid in protection and to reduce noise during recordings. The animals were then awakened with an injection of 2.5 mg/kg Antisedan. The animal was allowed to recover post-surgery for three to five days before recordings. Then, the electrode bundles targeting the BLA were slowly advanced over several days in small daily increments.

2. Methods

2.3. Behavioral paradigm and in vivo recording

2.1. Ethics statement

Prior to the recording experiments, mice were placed on a fooddeprivation schedule to reduce their weight to 80–85% of their baseline weight. They were fed with mouse chow in their home cages in the morning each day (limited to 2 g/mouse/day). Water was available at all times in the home cages. For desensitization-based alternative food preference test, the mice were habituated to three different foods - namely, KOKUHO premium quality rice (∼20 mg), rodent diet (Harlan Laboratories, ∼20 mg per pellet) and a milk droplet (25 µl, made from instant nonfat dry powder at 25 g in 50 ml water, Stop & Shop) in the cages overnight for three days. On the preference test day, each subject mouse was fed seven rice pellets first. Then, each mouse was given the rice as well as another different food, rodent diet or milk droplets. The food preference test was quantified by computing the total consumption number of rice pellets vs. another different food in the first 14 consumptions. All data were presented as mean ± s.e.m. Differences were considered significant if p value was < 0.05. For tetrode-recording experiments, we first recorded the BLA neural activity in freely-behaving mice in the home cage for at least 30 min as a baseline. Three different types of foods (20 mg rice, 20 mg rodent biscuit pellet, and 25 µl milk droplets) were delivered to the small petri dish located in the home recording chamber. Each food (pellet or droplet) was delivered with a 15–30 s time interval after consumption within a given food session. Mice typically consumed the food within 10–20 s. The time intervals between the different food sessions were 5–10 min. The order of delivery was randomized. In a small set of recording experiments, a water bottle was also available and the number of drinking bouts by these mice was monitored. The recordings were continued for an additional 30 min after all the appetitive experiments were completed. The experiments were videotaped by a camera placed

All animal work described in the study was carried out in accordance with the guidelines established by the National Institutes of Health regarding the care and use of animals for experimental procedures and the protocols approved by the Institutional Animal Care and Use Committee at the Medical College of Georgia at Augusta University (Approval AUP number: BR10-12-392). 2.2. Construction of tetrode headstages and animal surgery We employed adjustable 128-channel or 64-channel tetrode microdrives to target the BLA bilaterally (Lin, et al., 2005; Lin, Chen, Xie, Zaia, Zhang, and Tsien, 2006b). Tetrodes and headstages were constructed using the procedures as we have previously described (Lin et al., 2006b). To construct tetrodes, four wires (Fe-Ni-Cr, Stablohm 675, 13-µm diameter or 90% platinum, 10% iridium, 17-µm diameter, California Fine Wire Company, Grover Beach, CA) were twisted together using a manual turning device and soldered with a low-intensity heat source (variable temperature heat gun 8977020, Milwaukee, Brookfield, WI) for six seconds. The impedances of the tetrodes were measured with an electrode impedance tester (Model IMP-1, Bak Electronics, Umatilla, FL) to detect any faulty connections, and our tetrodes were typically around 0.5 MΩ. The insulation was removed by moving the tips of the free ends of the tetrodes over an open flame for approximately one second. The tetrodes were then placed into appropriate polyimide tubes. Only tetrodes, but not the surrounding polyimide tubes, were inserted into the brain tissue, thereby minimizing the tissue damage. Adult male wild-type mice (5–7 months old) were housed in a large 262

Neurobiology of Learning and Memory 155 (2018) 261–275

J. Liu et al.

was longer than five bins, three bins, two bins for a given category of stimulation, respectively, then it is considered a significant neuronal response to the stimulus. These criteria are determined empirically. To facilitate the comparison between units that exhibit different increases/ decreases over baseline activities, we used the transformation Ri = (fresp, i −fpre, i )/(f0 + fresp, i ) . Here, fresp, i represents the average firing rate during the detected neuronal response after the stimuli i , fpre, i is the average firing rate during baseline before the food stimuli i , and f0 is the averaged basal firing rates of all isolated units. Note that this transformation allows for uniform quantification of the significant changes in firing patterns for units with both low- and high-baseline firing rates. For measuring the population response significance, neural activity was calculated by comparing the firing rate after stimulus onset (in 500ms bin size) with the firing rate recorded during the baseline periods (10-s before food consumption) using a Z-score transformation. Z-score values were calculated by subtracting the average baseline firing rate established over the defined duration preceding the stimulus onset from individual raw values and by dividing the difference by the baseline standard deviation.

above the recording chamber. At the end of the chronic recording experiments, the mice were anesthetized and a small amount of current (20 µA, 7 sec) was applied to the recording electrodes in order to mark the positions of the tetrode bundles. The actual electrode positions were confirmed by histological Nissl staining using 1% Cresyl Echt Violet. The stability of the in vivo recordings was judged by waveforms at the beginning, during and after the experiments. Only stable units were included for the present analysis. 2.4. Data processing and spike sorting The neuronal activity was recorded by a Plexon multi-channel acquisition processor system, and waveforms were collected using 56 points with 1400 µsec time width. The recorded spike activities were processed in the manner also previously described (Chen, Wang, and Tsien, 2009; Li, Zhao, Lee, Wang, Kuang, and Tsien, 2015; Lin et al., 2006b; Zhang, Chen, Kuang, and Tsien, 2013), and then sorted using the MClust 3.5 program (http://redishlab.neuroscience.umn.edu/ MClust/MClust.html). First, the recorded data were filed in Plexon system format (*.plx). Then, the recorded data were aligned and converted into a Neuralynx System format (*.ntt). Next, the MClust 3.5 program was used to isolate different spiking units. Only units with less than 0.1% in spike intervals within a 1-msec refractory period and clear boundaries, as judged by L-ratio < 0.5 and an Isolation-Distance calculation > 15, were included in the present analysis. Because the pyramidal cells have low firing rates and long spike widths, whereas the interneurons show fast spiking and short spike widths (Quirk, Repa, and LeDoux, 1995), here, we developed a method to characterize the waveform difference of pyramidal cells vs. interneurons by using three waveform features: trough-to-peak width, halfwidth after trough, and the differential integral of area shape ΔAafterpeak (Supplementary Figure 1B). Trough-to-peak width was defined as the duration from trough to peak. Half-width after trough was defined as the width between the points when waveform rose to or fell to half the height of the peak. ΔAafterpeak was defined as the area between the waveform and the line segment of the peak and the last point of the waveform. The feature ΔAafterpeak values could be positive or negative, depending on the shape of the waveform after peak. The k-means method was employed to achieve automated cell-type clustering, and we found that the vast majority of the units recorded from BLA were putative excitatory principal cells (∼84.1%). Besides the waveform features, the firing rate is another major criterion to separate pyramidal cells and interneurons (Herry et al., 2008; Quirk et al., 1995). Since some putative pyramidal cells were reported to have high spontaneous firing rates to be 4.2 ± 0.8 Hz (Herry et al., 2008), we accordingly used 5 Hz as a threshold to further tighten the criterion for separating putative pyramidal cells vs. interneurons. Although a good portion of the recorded units fell outside these criteria and were excluded from the present study, it allowed a more stringent selection of putative pyramidal cells vs. fast-spiking interneurons. Supplementary data associated with this article can be found, in the online version, at https://doi.org/10.1016/j.nlm.2018.08.012.

2.6. Hierarchical categorical clustering Hierarchical clustering methods were used to investigate the stimulus responses of the overall population of the simultaneously recorded units. The procedure was similar to the one previously described (Lin et al., 2005; Zhang et al., 2013). This analysis was performed on a transformed neuronal response: T = log(1 + |R|) . Here, R is a n × m matrix representing the neuronal responses of n units during m stimulus, and | | denotes the absolute value. An agglomerative hierarchical cluster tree was created from the standardized Euclidean distances. Then, a categorical sorting was applied to facilitate the visualization. That is, units were sorted by the number of stimuli to which these units responded. After the categorical sorting, the units which were non-responsive were put on the bottom of the matrix, whereas the units responding generally to all three types of foods were located at the top of the matrix. 2.7. Assessing non-randomness in response patterns In order to test if the response patterns of the recorded ensembles deviated from random distribution, we constructed a null (independent random) model, which assumed that each neuron in the given region has an independent response probability to a given pattern, and the response probabilities of this neuron to different stimuli can be different. In other words, each event would randomly activate a subset of the neurons with a different average ensemble size, and the activation patterns of different stimuli are independently chosen and do not interact. Accordingly, we normalized the observed histogram of counts for eight neuron response types to four different stimuli (example shown below) by the total number of counts to obtain the response pattern distributions denoted as {pi ({Ej})} , where i goes from 1 to 8, and j from 1 to 3, and Ej can be 0 or 1. Example histogram is shown below:

2.5. Characterization of principal unit responses Clique Clique Clique Clique Clique Clique Clique Clique

To determine whether a recorded unit was responsive to a given appetitive stimulus, we used the first video frame video showing the food consumption as time zeros to calculate a peri-event histogram using a 500-ms bin size. The neural activities 10 s prior to appetitive consumption were used as a baseline to determine confidence intervals. To assess significant changes in firing, we used the following two steps: First, the 80% confidence intervals were used as a threshold to detect if there was any significant increased positive response in firing-rate within 20 s after appetitive stimulus. Second, if the firing rate increase reached 95%, 99.9%, 99.999% confident intervals, and the duration

#1: #2: #3: #4: #5: #6: #7: #8:

282 cells (non-responsive units) 31 cells (responded to milk only) 14 cells (responded to biscuits only) 10 cells (responded to rice only) 10 cells (responded to biscuits and milk) 6 cells (responded to rice and milk) 9 cells (responded to biscuits and rice) 9 cells (responded to all three foods)

The probability of generating the combined pattern is the product of individual independent probabilities, i.e., pi (E1 = 0, E2 = 1, 263

Neurobiology of Learning and Memory 155 (2018) 261–275

J. Liu et al.

the number of recorded neurons is much higher than the number of repeated trials. In practice, the matrix SW can be rendered invertible using a regularization technique which changes each class covariance matrix based on the following formula: Ωi' = (1−λ )Ωi + λ I, where Ωi is the covariance matrix for the ith class, λ is a regularization parameter between 0 and 1, and I is the identity matrix. We determined the parameter λ automatically for each data set based on the optimization procedure we developed previously (Osan, Zhu, Shoham, and Tsien, 2007); each particular choice is determined by the particular distributions within each data set. After computation of the N −1 discriminant dimensions was computed, we projected the neural patterns during food consumption in this low-dimensional encoding subspaces. We then used the multivariate Gaussian distribution probability functions 1 exp(−(x −m)t Ω−1 (x −m)/2 ) to fit the projections for (P (x ) = N /2 1/2

E3 = 0) = p (E1=0) × p (E2 = 1) × p (E3 = 0) . The independent probabilities can be obtained by summing all pi s where a given event occurs. For example, p (E1=0) can obtained by summing all pi s with E1 = 0 , regardless of the value of the other variables. To generate the error bars, we used a resampling procedure by running 1000 Monte-Carlo simulations. In each simulation, we assigned each data point to one of those eight-clique patterns by drawing a random number weighted by the probability of occurrence of each pattern. The five percent and 95 percent values were plotted on the graphs, and the p value for the probability of a given pattern being significantly different from the actual distribution is assessed at a level of p < 0.05 after dividing the p value by 8 to reflect the Bonferroni multiple hypothesis testing correction. 2.8. Local field potential analysis

(2π )

| Ω|

The local field potential (LFP) data was processed by the FPAlign (a LFP time alignment utility provided by Plexon) to correct the filter induced phase delay. Further analyses were carried out with these aligned LFP data by custom-written MATLAB programs. For the detection of theta oscillations, the LFP data was first bandpass filtered (4–12 Hz). Then a theta (5–10 Hz)/delta (2–4 Hz) ratio was calculated in a 2-s window. Epochs with more than seven consecutive time windows in which the ratio was > 4 were identified as theta episodes (Csicsvari, Hirase, Czurko, Mamiya, and Buzsaki, 1999). To calculate the phase relationship between unit activity and theta oscillation, the trough for each theta cycle was identified and were assigned instantaneous phase of 0 rad. The phase value of each spike was obtained by linear interpolation. Then the histogram was calculated with 18degree bin size, and a Rayleigh's test was performed to determine the significance (p < 0.01) of phase modulation of theta oscillations. For the detection of gamma oscillations, the LFP data was bandpass filtered at 30–80 Hz. Gamma power (root mean square) of the filtered signal was calculated in 25-ms sliding windows (1-ms step). Two standard deviations above the mean power was used as the threshold for detecting gamma episodes (Csicsvari, Jamieson, Wise, and Buzsaki, 2003). The power spectral densities and power spectra analyses were conducted in NeuroExplorer (Version 5, Nex Technologies), then the value of averaged power spectral-density of theta and gamma oscillation was sent to MATLAB for figure plotting.

each class. We subsequently enhanced our intuition about the relationships among classes by visualizing the 2σ boundary ellipsoids for each class. We tested the robustness of our MDA statistical model by employing different partitions of the training and test data points. In general, we find that the performances for our model do not depend strongly on the particular choice of the training and test data selection. In addition, we used a sliding-window method to monitor the evolution of the population state throughout the duration of the experiment and to identify the occurrences of patterns over repeated trials (Lin et al., 2005; Osan et al., 2007). 250 msec sliding window was used. Dynamic trajectories were plotted as they moved from the rest cluster towards the corresponding food clusters. Inspection of the clusters generated by our use of MDA technique indicated that all three types of food stimulations were successfully classified. To assess the performance of the MDA-classification method, we partitioned the data set in a training data set and a test data set. Since this method can overfit the training data, especially when the dimension of the input data is high, we evaluated the generalization performances on the test data set. Using the projected training data points, we fit the lower-dimensional cluster for each class with Gaussian distributions (ellipsoids). We then evaluated the distance from the center of each cluster for all test data and assigned each point to the closest cluster to compute the percentages for correct classification for each mouse (Table 1).

2.9. Multiple discriminant analysis projection

3. Results

Multiple Discriminant Analysis (MDA) projection method was used to classify the neural responses corresponding to different food experiences into different ensemble-coding classes (Lin et al., 2005). We computed firing frequencies (f) in ten 500-ms time-bins before and after food stimuli. Baseline activities were characterized by computing the average firing rates during time intervals preceding the food stimuli. We set aside randomly chosen population activities from one of each type of food; this constitutes our test data set. The rest of the sampled population activities were then used to train our MDA statistical model. The matrix of mean responses during each category (rest and startle states) were then computed and used to compute the between-class N scatter matrix SB = ∑i = 1 ni (mi−m)(mi−m)t : Here ni is the number of elements in each class, N is the number of classes, is the mean vector for each class, m is the global mean vector and the symbol t indicate the transpose operator. To take into account the variations occurring for each class we also computed the within-class scatter matrix SW , which N N is defined as: SW = ∑i = 1 Ωi = ∑i = 1 ∑x ∈ Di (x −mi )(x −mi )t . Here Di represents the set of population responses triggered by the ith startle type. Using these two matrices, it follows that a set of at most N −1 discriminant projection vectors can be determined by computing the eigenvalue decomposition of the matrix SW −1·SB . For our data sets, the class covariance matrices SW were non-invertible, which was a direct consequence of data under-sampling, since

3.1. Specific-to-general coding patterns by pyramidal cells despite of the salt-and-pepper intermingled cellular distribution We employed 128-channel tetrode arrays to monitor activity patterns of large numbers of neurons in the BLA while mice freely consumed three different types of food items - namely, rodent biscuits, rice, and milk (Fig. 1A). To promote eating behaviors, the mice were subjected to food restrictions (limited to 2 g per day) for three to five days Table 1 Prediction power in the MDA-computed encoding subspace for different food categories. Mouse

Rest

Biscuits

Rice

Milk

A B C D E

97 89 98 93 98

99 100 90 71 84

100 70 100 84 100

100 96 100 100 100

Caption: Percent correct predictions were evaluated by using 1000 random combinations of training/test data for each mouse. For validation purpose, test data was set aside from the training set (6 out of 7 trials). The total numbers of simultaneously recorded BLA cells for the mouse A, B, C, D, and E were 127, 140, 147, 55, and 67, respectively. 264

Neurobiology of Learning and Memory 155 (2018) 261–275

J. Liu et al.

were performed while mice were exposed to the above three types of foods. To minimize potential emotional effects due to changes of environment, we conducted the recording experiments in their home buckets. A single food pellet or milk droplet was delivered to a dish

prior to the recording experiments. Tetrodes were implanted into the BLA bilaterally with 64 channels on each hemisphere (see Supplementary Fig. 1A for electrodes placements). Once large numbers of units were obtained from multiple channels, the recording experiments

(caption on next page) 265

Neurobiology of Learning and Memory 155 (2018) 261–275

J. Liu et al.

Fig. 1. The specific-to-general coding patterns by putative pyramidal cell populations in the BLA. (A) Three distinct foods (rodent biscuit, raw rice, and milk) were presented to mice. (B) An example of broadly tuned, putative pyramidal cell responded to all three types of foods. Peri-event raster (upper subgraph) and peri-event histograms (lower subgraph) of this general-response cell to rodent biscuit, rice, and milk are shown from left to right. The population response of the general neural clique (n = 9 units) is shown by the Z-score plot on the right. (C) Pyramidal cells responded to a combination of two different foods. An example putative pyramidal cell responding to rodent chow and rice, but not to milk, is shown in the top row. An example cell responding to rodent chow and milk, but not to rice, is shown in the middle row. An example putative pyramidal cell responding to rice and milk, but not to rodent chow, is shown at the bottom row. The population response of these neural cliques and their corresponding cell numbers are shown on the right column. (D) Specific pyramidal cells responding only to one type of food. An example cell responding specifically to rodent biscuit is shown on the top row. An example cell responding specifically to rice is shown in the middle row, whereas the milkspecific cell is shown at the bottom row. The population response of these neural cliques and their corresponding cell numbers are shown on the right column. (E) Hierarchical clustering shows seven distinct pyramidal cell cliques, via permutation and combination rule, to process food experiences. Y-axis of the heat map lists the number of 371 putative pyramidal cells, whereas X-axis labels correspond to the rodent biscuit, rice and milk, respectively. Color scale bars indicate the logarithm transformed responsiveness of putative pyramidal units averaged over seven trials. (F) The percentage of distribution of different response types among the recorded BLA pyramidal cell population.

high selectivity to only one type of food (Fig. 1D). To obtain an overview of BLA response patterns, we applied a hierarchical clustering method (Lin et al., 2005) to assess and visualize firing changes of these 371 putative pyramidal cells. A total of seven distinct cell subclasses was identified (Fig. 1E). Overall, we found that 24% of these recorded cells (89 of 371) increased their firings upon appetitive stimulation, whereas 76% of the population (282 out of 371 cells) were unresponsive (Fig. 1F). Among these responsive neurons, the majority (∼62%, 55 out of all 89 responsive pyramidal cells) belonged to specific cell cliques (Fig. 1F). The units which specifically reacted to milk were among the largest portion (∼34%, 31 out of 89 responsive cells) (see the bottom subpanel in Fig. 1D). This likely reflected the evolutionary selection of such a natural preference for its vital role in ensuring healthy postnatal development. Many other units showed specific firing increases to biscuit or rice (∼16% and ∼11%), respectively (see the top and middle subpanels in Fig. 1D). We noted that the duration of firing changes to these foods typically lasted 10 s or more, matching well to the amount of actual time in consuming each rice, biscuit, or milk droplet (see the population Z-score plots on the right side column of Fig. 1, as well as Supplementary Figure 4). Our analysis reveals that the units exhibiting combinatorial response patterns to two types of foods constituted about 28% of responsive pyramidal cells (25 of out 89 responsive cells, or 7% of all recorded pyramidal cells) (Fig. 1F). For example, some of these putative pyramidal cells responded only to rodent biscuit and rice – but not to milk – while other cells reacted to rodent biscuit and milk, but not to rice. In yet another set of cells, they did not respond to the biscuit, but reacted strongly to rice and milk (Fig. 1C). These cells covered various permutation and combination of two types of food. As evident from the distribution analysis, we found that the general responsive cells (∼10%, 9 out of 89 cells), which reacted to all three foods, had the smallest percentage (∼2% of all recorded pyramidal cells) (Fig. 1F). This different ratio among specific-to-general cells suggests that the BLA devotes a large percentage of cells for discrimination of different

located inside the cage and repeated seven times, with a 15–30 s timeinterval after consuming each food item. This self-initiated, voluntary behavioral protocol mimics real-world experiences in food consumption. The mice were allowed to rest for 5–10 min after the consumption of a given type of food (e.g. 7 biscuits) before exposure to a different type (i.e., 7 droplets of milk or 7 rice). The order in delivering various food items was randomized in different animals. Using this protocol, we recorded a total of 536 units from five wildtype mice. To characterize response patterns among distinct neuron types, we separated the recorded BLA units into putative pyramidal cells and fast-spiking putative interneurons based on the combination of the waveform characteristics such as the time duration from the trough to peak, waveform shape, and peak half width (Supplementary Fig. 1B). The resulting putative pyramidal cells and interneurons were further threshold by the firing rates: less than 5 Hz for pyramidal cells and greater than 5 Hz for interneurons. In addition, we assessed the stability of recorded single units by examining their spike waveforms before, during, and after appetitive stimulation (Supplementary Fig. 2A showing a representative pyramidal cell, Supplementary Fig. 2B showing a representative fast-spiking interneuron). This ensured that changes in firing patterns to food consumption were not due to signal contamination upon consummatory behavior. Of these 536 units recorded, 371 putative pyramidal units met the above criteria and were used for peri-event spike raster and histogram analysis. The time zeros used for peri-event spike rasters were identified from the recorded videotape (30 frames/second) as the moments when the mouse approached the milk, biscuit, or rice, and started to eat from the dish (Supplementary Fig. 3A). On average, mice consumed these foods in about 9–12 s (Supplementary Fig. 3B). Interestingly, we found that putative pyramidal cells exhibited a variety of food-response selectivity, including the general-responsive cells that increased their firings to all three types of food experiences (Fig. 1B) and the subgeneral cells that responded to the combination of two types of food consumption (Fig. 1C). Moreover, many BLA pyramidal cells exhibited

Fig. 1. (continued) 266

Neurobiology of Learning and Memory 155 (2018) 261–275

J. Liu et al.

Fig. 2. Specific-to-general cell-assembly distribution pattern differs significantly from the independent random-connectivity model. Summary statistics for the independent model (random connectivity model) for the BLA datasets. Clique #1: 282 non-responsive units; Clique #2: 31 cells responded to milk only; Clique #3: 14 cells responded to biscuits only; Clique #4: 10 cells responded to rice only; Clique #5: 10 cells responded to biscuits and milk; Clique #6: 6 cells responded to rice and milk; Clique #7: 9 cells responded to biscuits and rice; Clique #8: 9 cells responded to all three foods. Error bars reflect 95% confidence intervals; stars indicate significant differences from the actual distribution of empirical data as assessed by a resampling procedure. * indicates p < 0.05. N = 371 putative pyramidal cells.

or milk (Fig. 3A–C, the lower subpanels). Population Z-score plots of the selected group responses (n = 10 for each group) at the first trial vs. seventh trial also showed that distinct sub-populations of BLA units exhibited different sensitivity to repetition even within this seven-trial protocol (Fig. 3A–C, see the right column, as well as Supplementary Fig. 6). Additional analysis of these rapidly desensitized cells showed that these cells tended to distribute across specific-to-general neural cliques (Fig. 3D). Moreover, at the behavioral level, the mice which consecutively consumed the same food item (e.g., seven rice pellets) would prefer to switch to different food items (e.g., milk or rodent biscuits) as demonstrated in the food-habituation-and-choice-preference test (Fig. 3E). Therefore, the rapid desensitization of some pyramidal cells may underlie or contribute to such a behavioral habituation.

food experiences, while also engages in extracting various combinatorial features for categorization and generalization. To examine whether this specific-to-general coding pattern was also present in the individual BLA microcircuit, we performed hierarchical clustering analysis on single mouse datasets. Indeed, we found the same permutation-based combinatorial patterns in single datasets (Supplementary Fig. 5, a total of 101 putative pyramidal cells from the mouse-A was listed). Therefore, this specific-to-general, cell-assembly coding is preserved in both the single datasets as well as the pooled datasets, despite of the salt-and-pepper intermingled anatomical distribution of these responsive pyramidal cells in the BLA. 3.2. Specific-to-general coding patterns among pyramidal cell assemblies were nonrandom The next question we asked is whether such specific-to-general cell assembly logic reflects a stochastic process or pre-configured nonrandom strategy. Theoretically, if given sufficient convergence and divergence in auto-associative recurrent model (such as CA3), it is possible that a random wiring-based developmental strategy can potentially give rise to specific-to-general combinatorial connectivity. Although the BLA is known to be largely non-recurrent, feedforward circuit, we asked whether the specific-to-general cell-assembly pattern logic observed in the BLA pyramidal cell population reflects a random or non-random distribution-pattern. Accordingly, we assessed the response probability-distribution by using the independent-connectivity model. Our analysis revealed that the observed specific-to-general distribution pattern in the BLA pyramidal cell population differed significantly from the stochastic random patterns predicted by the independent-connectivity model (Fig. 2, p < 0.05). This result indicates that the specific-to-general cell-assembly pattern in the BLA was highly nonrandom, possibly reflecting the genetically pre-programmed connectivity patterns.

3.4. Gamma oscillation was diminshed during food consumption Because some neural circuits are known to exhibit dynamic network-level oscillation which can, in turn, modulate cell-firing properties (Buzsaki and Wang, 2012; Fries, 2009; Likhtik, Stujenske, Topiwala, Harris, and Gordon, 2014; Lisman and Jensen, 2013; Popescu, Popa, and Pare, 2009), we asked how these food-responsive pyramidal cells might interact with the local field potential (LFP) during food experiences. We found that in the amygdala, theta (4–12 Hz) and gamma (30–80 Hz and 90–130 Hz) rhythms were the main types of LFP oscillations before, during, and after food consumption (Supplementary Fig.7A-B). Intriguingly, we noted that gamma oscillation (30–80 Hz) tended to decrease during the actual eating of foods, while theta seemed to fluctuate (or even increase) before or during food consumption (Supplementary Fig. 7C and D). We found that theta-modulated pyramidal cells were locked to different phases of theta oscillation (Supplementary Fig. 7E). Both theta-modulated and theta-unmodulated pyramidal neurons were observed among responsive and non-responsive cell populations (Fig. 4A). Interestingly, we found that many pyramidal cells were also tightly coupled with gamma oscillations (30–80 Hz). These gamma-coupled cells exhibited the highest firing probability at the trough of gamma (Supplementary Fig. 7D). Still some BLA pyramidal cells did not show modulation by gamma oscillation (Fig. 4B). At the population level, our analysis revealed that a higher percentage of the responsive cell cliques exhibited theta coupling (with the general clique cells having the highest percentage showing theta coupling at 66.7%), whereas food-unresponsive pyramidal cells had the lowest percentage in theta-coupling (24.8%) (Fig. 4C). Likewise, foodresponsive pyramidal cells that belonged to the general and sub-general cliques were all coupled to gamma oscillation (Fig. 4D). In contrast, a

3.3. A subset of pyramidal cells exhibited rapid desensitization to repeatedly eating of the same food Because eating food is a highly dynamic process and its hedonic or emotional value may change from moment to moment (i.e., by repetitively eating the same food items), we asked whether consumption of the same food pellets over the course of seven repetitions may influence temporal firing patterns of these BLA pyramidal cells. While many responsive units seemed to show similar firing increases across all seven trials (Fig. 3A–C, the upper subpanels), interestingly, we noticed that a subset of pyramidal cells tended to reduce their firing responses by the seventh trial in comparison to the first trial of eating rodent biscuit, rice 267

Neurobiology of Learning and Memory 155 (2018) 261–275

J. Liu et al.

Fig. 3. Habituated vs. non-habituated response patterns in pyramidal cells over the course of seven trials. (A) Two distinct examples of putative pyramidal cells showed different dynamics over the seven rodent biscuit trials. The top unit showed similar response robustness between the first and seventh trials, whereas the bottom unit seemed to habituate over the course of consuming seven rodent biscuits. The top row in the peri-event raster represents spike train on the first trial, whereas the bottom row corresponds to the seventh trial. Z-score plots showed the averaged responses of such 10 representative units from these two distinct subtypes. (B) Two representative pyramidal cells showed different dynamics within the seven trials of drinking milk droplets. The top unit showed similar response robustness over seven trials, whereas the bottom unit exhibited habituation over the course of consuming seven milk droplets. Z-score plots showed the averaged responses of 10 representative units from these two distinct sub-types. (C) Two representative pyramidal cells showed different dynamics within the seven trials of eating rice. The top unit showed similar response robustness over seven trials, whereas the bottom unit exhibited habituation over the course of consuming seven rice. Z-score plots showed the averaged responses of ten representative units from these two distinct sub-types. (D) Food-responsive cells exhibiting rapid habituation were distributed across specific-to-general neural cliques. (E) After being fed seven rice pellets, the subject mice consumed more other food than rice in the first 14 trials. n = 20, ***p < 0.001, paired t test. All data are plotted as mean ± s.e.m.

exhibit LFP-related firing dynamics. We found that 75% of the FS interneurons were coupled to theta rhythm (Fig. 6A, the left pie), whereas all these interneurons were gamma-locked (30–80 Hz) (Fig. 6A, the right pie). In addition, we noted that Type-1 interneurons (n = 3 cells) had diverse phase-locking, one cell was coupled to the trough of theta (Fig. 6B, the top panel), whereas the remaining two cells seemed to be coupled to the peak of theta (Fig. 6B, the bottom panel). Yet all of them showed peak firing at the trough of gamma (Fig. 6B, the right column). Type-2 interneurons (n = 5 cells) all exhibited theta modulation with elevated firings around the peak of theta (Fig. 6C, left panel). They also showed the uniformed coupling to the trough of gamma (Fig. 6C, right panel). In contrast, Type-3 interneurons (n = 4 cells) did not show any significant modulation by theta (Fig. 6D, left panel), but they were all coupled to the trough of gamma (Fig. 6D, right panel). The detailed distribution of responsive interneurons over the theta and gamma cycles is listed in the Supplementary Table 1. For food-nonresponsive interneurons, some also showed theta coupling, whereas other did not (Supplementary Fig. 8, the left column), but all of them exhibited strong coupling effects with gamma (Supplementary Fig. 8, the right column).

greater fraction of food-nonresponsive pyramidal cells (34.4%) did not show any significant coupling with gamma (Fig. 4D). These results suggest that many food-responsive pyramidal cells were coupled with theta and low gamma, suggesting that distinct subclasses of pyramidal cells may be subjected to differential network-level modulations by local interneurons and/or inputs from other brain regions. 3.5. Fast-spiking interneurons were broadly tuned to food consumption Next, we examined how fast-spiking (FS) interneurons in the BLA responded to food experiences. Twenty putative FS interneurons met our major criteria. We found about 40% of putative FS cells (n = 8) did not change their firings, whereas the other 60% (n = 12) did. However, we noticed the major differences in the patterns of these interneurons in comparison to those of pyramidal cell populations. These FS cells typically responded to all three foods (Fig. 5). Moreover, some increased their firings during food consumption (Type-1 interneurons) (Fig. 5A and B), whereas others uniformly reduced their firings upon eating foods (Type-2 interneurons) (Fig. 5C and D). In addition, several FS cells had non-uniform responses – namely, increased firing to solid foods (such as to biscuits and rice), but suppressed firing to milk droplets (Type-3 interneurons) (Fig. 5E and F). Overall, FS interneurons lacked specific selectivity to a given type of food items, but rather exhibited a rich diversity in their temporal patterns (increase or decrease in firings), possibly reflecting their distinct roles in compartmentalized regulation of pyramidal cell firing pattern and local circuitry dynamics. We then investigated whether and how BLA interneurons might

3.6. Decoding and visualization of real-time neural ensemble traces in the BLA Finally, we took advantage of our simultaneously recorded large number of cells and examined the real-time BLA representation of food experiences on a single trial basis. Because neurons in the brain are widely known to react to external stimulations with great response 268

Neurobiology of Learning and Memory 155 (2018) 261–275

J. Liu et al.

Fig. 4. Some pyramidal cells were coupled to theta and/or gamma oscillations. (A) Both food-responsive and non-responsive pyramidal cells exhibited either thetacoupled or uncoupled firing. (B) Both food-responsive and non-responsive pyramidal cells showed either gamma-coupled or uncoupled firings. (C) The percentage of theta-coupled pyramidal cells in general, sub-general, specific and non-responsive clusters. (D) The percentage of gamma-coupled pyramidal cells in general, subgeneral, specific and non-responsive clusters.

269

Neurobiology of Learning and Memory 155 (2018) 261–275

J. Liu et al.

Fig. 5. Diverse temporal response patterns among fast-spiking interneurons. (A) An example of Type-1 FS-interneurons exhibited general increases in firings to the consumption of rodent biscuit, rice, and milk. (B) The population responses for this general increase FS units are shown by Z-score plot. (C) An example of Type-2 FSinterneurons exhibited general decreases in its firings to the consumption of rodent biscuit, rice, and milk. (D) The population responses for this general decrease FS units are shown by Z-score plot. (E) An example of the Type-3 FS interneurons, which increased firings to rice and rodent chow but decreased firing to drinking milk. (F) The population responses of units shown in (E) are shown by Z-score plot.

moment basis (Fig. 7). The sliding-window plotting gave rise to high temporal-resolution dynamical trajectories in the MDA spaces, showing that the baseline state trajectories typically stayed inside the resting ellipsoid (Fig. 7A), whereas neural traces representing each type of food experience exhibited robust, stable trajectories upon eating each item (Fig. 7B). These robust real-time encoding patterns were in stark contrast to highly variable spike trains for each individual neuron in a given trial. To examine the robustness and accuracy of this MDA-classification method, we partitioned the BLA datasets into a training dataset (randomly choosing 6 out of 7 trials as training trials) and a test data (the remaining one trial as test trial). By evaluating distance from the center of each ellipsoid cluster (using Gaussian distributions) from all test trial points, we can assess how the simultaneously recorded ensemble patterns provides prediction for the correct classification of stimulus identity. We found that overall ensemble coding of food categorization achieved good performance, ranging from 90 to 100%, reflecting the greater numbers of responsive cells recorded (Table 1). We noted that the classification performance is highest for milk, due to the fact that the number of milk specific responder cells tended to be more prevalent from the recorded datasets (Supplementary Fig. 10). Taken together, these ensemble coding patterns provided the first demonstration that BLA cell assemblies generated robust real-time dynamic patterns during consuming natural foods. Such specific-to-general cell-assembly patterns are ideal to account for both real-time pattern-discrimination and pattern-categorization in the brain’s neural circuits.

variability from trial to trial, the typical approach is to average their responses over many trials. While the peri-event spike-raster histograms provided the basic assessment of individual neurons’ response properties, over-the-trial averaging technique inherently led to loss of valuable information regarding real-time ensemble representation of stimuli as to how the brain achieves robust cognition on a moment-to-moment basis. To visualize the real-time representation and classification of food experiences in the BLA, we employed a multi-discriminant analysis (MDA) method, a supervised dimensionality reduction-based statistical pattern classification method (Chen et al., 2009; Lin et al., 2005; Osan et al., 2007; Zhang et al., 2013) to our large-scale datasets. Our MDA analysis revealed that real-time ensemble firing patterns during the baseline periods (prior to a food-consumption session) formed a tight cluster (resting state) in low-dimensional encoding subspaces (Fig. 7A), whereas real-time ensemble patterns corresponding to the consumption of rice, milk and rodent biscuits formed distinct, well-separated encoding clusters (Fig. 7B). To further monitor the real-time transient dynamics of BLA ensemble patterns, we combined the MDA method with a sliding-window technique for assessing the moment-to-moment temporal dynamics of neural processing (Chen et al., 2009; Zhang et al., 2013). By using the fixed matrix coefficients produced by the MDA or PCA method, we then computed the transient projection of neural ensemble responses throughout the recorded datasets with a sliding window. As such, the temporal evolution of the ensemble activity patterns corresponding to individual food consumption could be directly visualized as transient dynamical trajectories in these encoding sub-spaces on a moment-to270

Neurobiology of Learning and Memory 155 (2018) 261–275

J. Liu et al.

Fig. 6. Food-responsive interneurons showed different coupling to theta and/or gamma oscillations. (A) BLA interneurons (75% of them) exhibited theta coupling, in comparison to 100% of them showed gamma coupling. (B) Two foodresponsive Type-1 interneurons, one was coupled to the trough of theta (the upper left panel), another interneuron coupled with the peak of theta (the bottom left panel). Both cells showed gamma-coupling. (C) An example foodresponsive Type-2 interneurons exhibited theta- and gamma-coupling. (D) Type-3 food-responsive interneurons showed no coupling to theta, but were coupled to gamma.

4. Discussion Our present study has investigated how distinct cell subpopulations in the BLA encode various natural food experiences. Identification of both pyramidal cells and FS interneurons responding to food consumption under the naturally-occurring consummatory state provides important insights into how the amygdala engages in appetitive behavior, which has been previously investigated by pharmacological and lesion or c-fos imaging experiments (Balleine and Killcross, 2006; Carballo-Marquez et al., 2009; Everitt et al., 2003; Holland, Petrovich, and Gallagher, 2002; Killcross, Everitt, and Robbins, 1998; Parkinson, Robbins, and Everitt, 2000; Petrovich and Gallagher, 2003; Petrovich, Holland, and Gallagher, 2005; Petrovich, Setlow, Holland, and Gallagher, 2002; Robbins, 2000). Our large-scale tetrode recording techniques revealed that BLA pyramidal cells exhibited specific-to-general coding patterns despite of their salt-and-pepper intermingled cellular distribution. In contrast, interneurons lacked specific responses to a given food, but they showed broader responses to food experiences. For the pyramidal cell population, we found that the vast majority of the food responsive pyramidal cells (91.7% of all responsive pyramidal cells) exhibited selectivity to a specific food item (62.5%) and a sub-combination of two food items (29.2%), suggests that pyramidal cells in the BLA gear toward discriminative coding for a specific or sub-combination of food item(s). This is interesting when considering the vast literature that gustatory responses to simple tastants (i.e. NaCl, citric acid, sucrose solution) in gustatory cortical or subcortical neurons have been found to be broadly tuned (Fontanini et al., 2009; Hanamori, Kunitake, Kato, and Kannan, 1998a; Hanamori, Kunitake, Kato, and Kannan, 1998b; Katz, Simon, and Nicolelis, 2002; Simon et al., 2006). However, the present study is highly consistent with our recent report showing that specific-to-general encoding of milk, rodent biscuit, rice, and sugar pellets by BLA pyramidal cell populations (Xie, Fox, Liu, Lyu, et al., 2016). In that study, in response to four different foods, a total of 15 BLA pyramidal cell cliques were identified. These pyramidal cell assemblies formed specific-to-general coding patterns, conforming to the power-of-twobased permutation logic as postulated by the Theory of Connectivity to be the basic cell-assembly level motif for brain computation (Li, Liu, and Tsien, 2016; Tsien, 2015a; 2015b). It is interesting to note that such a power-of-two-permutation and combination- based specific-to-general pyramidal cell-clique organization is very similar to the cell-assembly computational logic observed in the CA1 pyramidal cells of the hippocampus in processing of various fear experiences (Lin et al., 2005; Lin et al., 2006b; Xie, Fox, Liu, Lyu, et al., 2016). This supports the notion that power-of-two-based permutation logic represents an evolutionarily conserved organizing principle to construct, cell assemblies or memory engrams (Li, et al., 2016; Tsien, 2015a; 2015b). For processing appetitive experiences, specific-food responsive pyramidal cells in the BLA encode specific but complex features about each individual food items, whereas general responsive cells are ideal to signal the generalized experience about

271

Neurobiology of Learning and Memory 155 (2018) 261–275

J. Liu et al.

Fig. 7. Classification, visualization and dynamic decoding of real-time BLA cell ensemble patterns encoding of distinct food experiences. (A) Ensemble responses visualized in MDA sub-spaces shows seven trials for each type of three food items from the mouse-1. Each dot within an ellipsoid is the statistical ensemble pattern from a single trial of a given food item. The boundaries of each ellipsoid reflect the 2σ boundaries with Gaussian distributions of seven trial results in the MDA encoding sub-spaces. Three real-time ensemble traces (green, pink and purple trajectories) during the resting state prior to each stimulation are shown by applying the sliding-window technique. (B) Robust real-time cell-assembly trajectories corresponding to each food experience were revealed by the sliding-window technique. Seven trajectories per food item were shown. They were robust and stable, in stark contrast to the high variability of individual neuronal responses.

Our study also raises a set of very interesting questions as to how these pyramidal cells achieve such selective tuning properties. It will also be valuable to systematically dissect out how sub-compositions of food items contribute to the observed features (i.e., salt, sucrose, water, texture, and/or fat content of the biscuits or milk, etc.) (Fletcher, Ogg, Lu, Ogg, and Boughter, 2017; Kadohisa et al., 2005b; Nishijo et al., 1998). Interestingly, we have noted that some BLA pyramidal cells increased their firings to all three foods. This may provide a mechanism that can give rise to abstraction of general features or knowledge (Bowers, 2009; Hromadka, Deweese, and Zador, 2008; Tsien, 2015b). While it is important to further dissect how relatively simple inputs from distinct sensory modalities are integrated to give rise to different food properties, neural coding at the level of the amygdala can be tightly regulated internally by different physiological and motivational states as well as previous memory experiences (Beshel and Zhong, 2013; Carter et al., 2013; Kadohisa et al., 2005a; LeDoux, Farb, and Romanski, 1991; Murray et al., 1996; Nishijo et al., 1998; Peng et al., 2015; Stettler and Axel, 2009; Zhan and Luo, 2010). Given the fact that the amygdala is a key site for integrating not only food taste but also its texture, smell, temperature, visual appeal, and oromotor signals (Geschwind, 1965; Isaacson, 2010; Killcross et al., 1998; Nishijo et al., 1998; Norgren, 1976; Parkinson et al., 2000; Piette, Baez-Santiago, Reid, Katz, and Moran, 2012; Turner and Herkenham, 1991) as well as prior social and visceral experiences (Carleton, Accolla, and Simon, 2010; Rampon et al., 2000), it is likely that BLA responses may have already integrated various high-order features that would be specific enough to account for categorical features of distinct food experiences, as we have indeed observed here. Another novel finding in the present study is that some of the specific-responsive pyramidal cells exhibit rapid desensitization in firing robustness when the mice repeatedly ate several pellets or milk droplets. This is consistent with rapid behavioral habituation observed in mice, that is, the mice preferred to eat other food items after consecutively consuming several rice pellets (Fig. 3E). This rapid habituation may be related to taste habituation which is different from the satiety-induced changes (Rolls, Rolls, Rowe, and Sweeney, 1981). Such cellular property may explain nicely why variety in a meal typically enhances food intake (Rolls, et al., 1981). Another interesting point is that the pyramidal cells exhibiting rapid habituation tended to distribute across specific-to-general neural cliques. Future study will be necessary to dissect out whether these cells reflect the holistic changes in food motivational or emotional values/decision-making processes,

eating foods, and sub-combinatorial cells provide categorization for various combinations (e.g., solid foods versus liquid foods, etc.). Because specific-to-general pyramidal cell organization was also observed in individual single datasets, this strongly suggests that this coding strategy was not an artifact resulting from pooling different data from multiple animals, but was indeed executed within a single circuit. To define the upper limit or the optimal of food categories in the mouse BLA, it will be interesting to increase more food types and categories. In a small set of experiments, we also recorded from the mice which consumed water from the drinking bottle in home cages in addition to the rodent biscuit, rice and milk. Indeed, we found that general pyramidal cells responded to water as well as those three foods (Supplementary Fig. 10A), whereas some units reacted to a combination of three foods (biscuit, milk and water, but not rice) or specifically only to water (Supplementary Fig. 10B and 10C). While more categorical foods and datasets will be necessary and highly desirable, however, one experimental constraint is that the number of foods a mouse can or would eat at a given session is limited. From a computational perspective, there are several strategies to construct the specific-to-general pyramidal cell circuitry. One possible mechanism is via stochastic wiring strategy, which may be possible for the completely auto-associative recurrent architecture (Marr, 1971; Rolls, 2013). However, the BLA pyramidal cell network is largely a feed-forward network and differs from the classic auto-associative network such as the CA3. Indeed, our analysis using independent random-connectivity model strongly indicates that the observed specific-to-general pyramidal cell-assembly organization in the BLA deviated significantly from the random connectivity pattern, as the Theory of Connectivity predicted (Li et al., 2016; Tsien, 2015a; 2015b). Interestingly, we found that cells responding specifically to milk constituted the largest proportion in comparison to cells encoding other foods. As a result, correct classification for real-time ensemble patterns for drinking milk was found to be the highest (Table 1). Ecologically and evolutionarily speaking, this also makes a good sense because milk plays an essential role in nurturing pups during the postnatal period (Tsien, 2015b). In the future, single-cell resolution of tracing and imaging of synaptic fibers and their wiring diagrams might be useful to inform how this cell-assembly coding logic is generated during development. Toward this end, simpler model organisms such Drosophila and its larvae may be ideal systems to functionally map the coding properties and wiring diagrams (Chiang et al., 2011; Guven-Ozkan and Davis, 2014; Hallem and Carlson, 2006; Jefferis and Hummel, 2006). 272

Neurobiology of Learning and Memory 155 (2018) 261–275

J. Liu et al.

which may modulate BLA cells’ responses to all foods in a general manner. It will be interesting in future experiments to determine whether the general food cells in the BLA are preferentially regulated by inputs from the VTA DA neurons. In summary, our large-scale recording experiments revealed that BLA pyramidal cells employed a specific-to-general coding strategy to encode various food experiences, whereas interneurons lacked specificity to a given food experience, but exhibited complex temporal dynamics. Moreover, some of the BLA pyramidal cells exhibited rapid desensitization over food repetition, consistent with behavioral habituation. In addition, ensemble activity patterns of these BLA cell assemblies were visualized as real-time population-level neural traces, allowing robust discrimination and categorization of various food experiences in any given single trial. Identification of the organizing logic of cell-assembly engram and its dynamic operation in the amygdala forms a basic conceptual framework to investigate how the brain encodes appetitive cognitions and behaviors.

via preferentially receiving inputs from the ventral tegmental area (VTA) and/or prefrontal cortex such as the anterior cingulate cortex (ACC). In contrast to the specific-to-general coding patterns within pyramidal cells, we showed that FS interneurons in the BLA lacked specificity for a given food item. Yet these interneurons exhibited three distinct temporal dynamics. While Type-1 and Type-2 interneurons showed opposite response dynamics (increased firing vs. decrease firing to all three foods), Type-3 interneurons increased firing upon eating solid foods (biscuits and rice), but reduced firing while drinking milk. The “solid-foods versus liquid-foods” phenomenon needs to be further studied in future by increasing the items in each categories. It has been reported that inhibiting cortical parvalbumin (PV) interneurons in the prefrontal cortex suppressed gamma oscillations in vivo, whereas driving these interneurons was sufficient to generate emergent gammafrequency rhythmicity (Sohal, Zhang, Yizhar, and Deisseroth, 2009). In considering that both Type-2 BLA interneurons exhibited decreased firing and low-gamma reduced its power during food consumption, Type-2 FS interneurons might belong to PV interneurons. Further characterization of interneuron identities and how they modulate pyramidal cells will be also essential (Klausberger and Somogyi, 2008), using Cre-lox-mediated optogenetic methods (Tsien, 2016). At the network level as indicated LFP oscillation, we found that theta and low gamma were rather prominent in the BLA. Interestingly, theta seemed to remain largely unchanged, and low-gamma was diminished during food consumption. We noted that many food-responsive BLA cells were not coupled with theta or gamma were diminished during food consumption. Currently, it remains unknown whether LFP participates at all in the other phases of appetitive experiences, such as replay. One possible function of pyramidal cells/ gamma-coupling may be related to the organization of BLA cell assemblies and dynamic interactions with other brain regions during postlearning consolidation or memory recall (Bauer, Paz, and Pare, 2007; Bocchio, Nabavi, and Capogna, 2017; Popescu et al., 2009; Stujenske, Likhtik, Topiwala, and Gordon, 2014). This possibility can be studied/ explored in more detail using multi-site simultaneous recording techniques (Xie, Fox, Liu, and Tsien, 2016). Finally, while measuring neuronal responses to a given set of stimuli, it is important to understand the tuning properties of neurons. The deciphering of real-time neural code requires that the operation needs to overcome trial-to-trial response variability. As evident from peri-event spike raster histogram, individual neurons responded to each trial with enormous variability, resulting in large errors in the prediction of food identity on a single trial and moment-to-moment basis. While this response variability can be averaged out over multiple trials, the brain is unlikely to use this practice to achieve real-time coding. To study the real-time neural coding of food experiences, we took advantage of a simultaneously recorded large number of cells and applied dimensionality-reduction methods which has enabled us to visualize real-time neural ensemble traces on a moment-to-moment basis. The robust classification of real-time BLA representation of distinct food experiences, in turn, supports the notion that real-time cognition is a result of cell-assembly neural coding (Li and Tsien, 2017; Lin, Osan, and Tsien, 2006a). Additional merit of this population coding strategy is that it can provide resilience to the potential sudden loss of a single neuron. Such a graceful degradation can be valuable during neurodegenerative diseases and aging. In addition to the study of how the BLA encodes real-time food experiences, one can further explore the question of how changes in animals’ appetitive, motivational or thalamic circuits may modify BLA coding (Berridge, 2007; Carter et al., 2013; Dayan and Balleine, 2002; de Araujo et al., 2006; Hommel et al., 2006; Ko et al., 2015; Palmiter, 2008; Stanley et al., 2011; Szczypka et al., 1999; Wu et al., 2012). For example, ventral tegmental area (VTA) dopamine neurons (DA), which generate positive and hedonic signals, is known to project directly to the BLA (Grace and Rosenkranz, 2002; Rosenkranz and Grace, 2002),

Author contributions JZT conceived the project and then designed the experiments with JL. JL generated the datasets and analyzed datasets together with CL, T.L. and JZT. S.S and ML analyzed independent connectivity model. JZT wrote the paper with input from all other authors. Acknowledgements This work is supported by a Georgia Research Alliance BrainDecoding Project grant, and the Brain Decoding Key Laboratory grant from the Yunnan Province Department of Science and Technology to JL and JZT, and the Chinese Natural Science Foundation grant (NSFC#91332122) to SS. We thank Dr. Hui Kuang for suggestions on phase-locking analyses, Colby Polonsky for the art illustration, and Sandra E. Jackson for proofreading. The authors declare no competing financial interests. References Adaikkan, C., & Rosenblum, K. (2015). A molecular mechanism underlying gustatory memory trace for an association in the insular cortex. Elife, 4. Balleine, B. W., & Killcross, S. (2006). Parallel incentive processing: An integrated view of amygdala function. Trends in Neurosciences, 29, 272–279. Bauer, E. P., Paz, R., & Pare, D. (2007). Gamma oscillations coordinate amygdalo-rhinal interactions during learning. Journal of Neuroscience, 27, 9369–9379. Baxter, D. A., & Byrne, J. H. (2006). Feeding behavior of Aplysia: A model system for comparing cellular mechanisms of classical and operant conditioning. Learning & Memory, 13, 669–680. Berridge, K. C. (1996). Food reward: Brain substrates of wanting and liking. Neuroscience & Biobehavioral Reviews, 20, 1–25. Berridge, K. C. (2007). The debate over dopamine's role in reward: The case for incentive salience. Psychopharmacology, 191, 391–431. Berthoud, H. R. (2004). Mind versus metabolism in the control of food intake and energy balance. Physiology & Behavior, 81, 781–793. Beshel, J., & Zhong, Y. (2013). Graded encoding of food odor value in the Drosophila brain. Journal of Neuroscience, 33, 15693–15704. Bocchio, M., Nabavi, S., & Capogna, M. (2017). Synaptic plasticity, engrams, and network oscillations in amygdala circuits for storage and retrieval of emotional memories. Neuron, 94, 731–743. Bowers, J. S. (2009). On the biological plausibility of grandmother cells: Implications for neural network theories in psychology and neuroscience. Psychological Review, 116, 220–251. Buzsaki, G., & Wang, X. J. (2012). Mechanisms of gamma oscillations. Annual Review of Neuroscience, 35(35), 203–225. Cahill, L. (2000). Modulation of long-term memory storage in humans by emotional arousal: Adrenergic activation and the amygdala. In J. P. Aggleton (Ed.). The amygdala: A functional analysis (pp. 425–445). ((2nd ed)). New: York Oxford University Press. Carballo-Marquez, A., Vale-Martinez, A., Guillazo-Blanch, G., & Marti-Nicolovius, M. (2009). Muscarinic transmission in the basolateral amygdala is necessary for the acquisition of socially transmitted food preferences in rats. Neurobiology of Learning and Memory, 91, 98–101. Carleton, A., Accolla, R., & Simon, S. A. (2010). Coding in the mammalian gustatory system. Trends in Neurosciences, 33, 326–334. Carter, M. E., Soden, M. E., Zweifel, L. S., & Palmiter, R. D. (2013). Genetic identification

273

Neurobiology of Learning and Memory 155 (2018) 261–275

J. Liu et al.

155–184. LeDoux, J. E., Farb, C. R., & Romanski, L. M. (1991). Overlapping projections to the amygdala and striatum from auditory processing areas of the thalamus and cortex. Neuroscience Letters, 134, 139–144. Li, M., Zhao, F., Lee, J., Wang, D., Kuang, H., & Tsien, J. Z. (2015). Computational Classification Approach to Profile Neuron Subtypes from Brain Activity Mapping Data. Scientific Reports, 5. Li, M., Liu, J., & Tsien, J. Z. (2016). Theory of connectivity: nature and nurture of cell assemblies and cognitive computation. Frontiers in Neural Circuits, 10, 34. Li, M., & Tsien, J. Z. (2017). Neural code-neural self-information theory on how cellassembly code rises from spike time and neuronal variability. Frontiers in Cellular Neuroscience, 11, 236. Likhtik, E., Stujenske, J. M., Topiwala, M. A., Harris, A. Z., & Gordon, J. A. (2014). Prefrontal entrainment of amygdala activity signals safety in learned fear and innate anxiety. Nature Neuroscience, 17, 106–113. Lin, L. N., Chen, G. F., Xie, K., Zaia, K. A., Zhang, S. Q., & Tsien, J. Z. (2006b). Large-scale neural ensemble recording in the brains of freely behaving mice. Journal of Neuroscience Methods, 155, 28–38. Lin, L., Osan, R., Shoham, S., Jin, W., Zuo, W., & Tsien, J. Z. (2005). Identification of network-level coding units for real-time representation of episodic experiences in the hippocampus. Proceedings of the National Academy of Sciences of the United States of America, 102, 6125–6130. Lin, L., Osan, R., & Tsien, J. Z. (2006a). Organizing principles of real-time memory encoding: Neural clique assemblies and universal neural codes. Trends in Neurosciences, 29, 48–57. Lisman, J. E., & Jensen, O. (2013). The theta-gamma neural code. Neuron, 77, 1002–1016. Marr, D. (1971). Simple memory: A theory for archicortex. Philosophical Transactions of The Royal Society London B Biological Sciences, 262, 23–81. Morton, G. J., Cummings, D. E., Baskin, D. G., Barsh, G. S., & Schwartz, M. W. (2006). Central nervous system control of food intake and body weight. Nature, 443, 289–295. Murray, E. A., Gaffan, E. A., & Flint, R. W. (1996). Anterior rhinal cortex and amygdala: Dissociation of their contributions to memory and food preference in rhesus monkeys. Behavioral neuroscience, 110, 30–42. Nectow, A. R., Schneeberger, M., Zhang, H., Field, B. C., Renier, N., Azevedo, E., ... Friedman, J. M. (2017). Identification of a Brainstem Circuit Controlling Feeding. Cell. 170(3), 429–442.e11. https://doi.org/10.1016/j.cell.2017.06.045. Nishijo, H., Uwano, T., Tamura, R., & Ono, T. (1998). Gustatory and multimodal neuronal responses in the amygdala during licking and discrimination of sensory stimuli in awake rats. Journal of Neurophysiology, 79, 21–36. Norgren, R. (1976). Taste pathways to hypothalamus and amygdala. The Journal of Comparative Neurology, 166, 17–30. Osan, R., Zhu, L. P., Shoham, S., & Tsien, J. Z. (2007). Subspace projection approaches to classification and visualization of neural network-level encoding patterns. Plos One, 2. Palmiter, R. D. (2008). Dopamine signaling in the dorsal striatum is essential for motivated behaviors – Lessons from dopamine-deficient mice. Molecular and Biophysical Mechanisms of Arousal, Alertness, and Attention, 1129, 35–46. Parkinson, J. A., Robbins, T. W., & Everitt, B. J. (2000). Dissociable roles of the central and basolateral amygdala in appetitive emotional learning. European Journal of Neuroscience, 12, 405–413. Peng, Y. Q., Gillis-Smith, S., Jin, H., Trankner, D., Ryba, N. J. P., & Zuker, C. S. (2015). Sweet and bitter taste in the brain of awake behaving animals. Nature, 527, 512-+. Petrovich, G. D., & Gallagher, M. (2003). Amygdala subsystems and control of feeding behavior by learned cues. Amygdala in Brain Function: Bacic and Clinical Approaches, 985, 251–262. Petrovich, G. D., Holland, P. C., & Gallagher, M. (2005). Amygdalar and prefrontal pathways to the lateral hypothalamus are activated by a learned cue that stimulates eating. Journal of Neuroscience, 25, 8295–8302. Petrovich, G. D., Setlow, B., Holland, P. C., & Gallagher, M. (2002). Amygdalo-hypothalamic circuit allows learned cues to override satiety and promote eating. Journal of Neuroscience, 22, 8748–8753. Piette, C. E., Baez-Santiago, M. A., Reid, E. E., Katz, D. B., & Moran, A. (2012). Inactivation of basolateral amygdala specifically eliminates palatability-related information in cortical sensory responses. Journal of Neuroscience, 32, 9981–9991. Popescu, A. T., Popa, D., & Pare, D. (2009). Coherent gamma oscillations couple the amygdala and striatum during learning. Nature Neuroscience, 12, 801–807. Quirk, G. J., Repa, C., & LeDoux, J. E. (1995). Fear conditioning enhances short-latency auditory responses of lateral amygdala neurons: Parallel recordings in the freely behaving rat. Neuron, 15, 1029–1039. Rampon, C., Tang, Y. P., Goodhouse, J., Shimizu, E., Kyin, M., & Tsien, J. Z. (2000). Enrichment induces structural changes and recovery from nonspatial memory deficits in CA1 NMDAR1-knockout mice. Nature Neuroscience, 3, 238–244. Robbins, B. J. E. R. N. C. J. H. J. A. P. T. W. (2000). Differential involvement of amygdala subsystems in appetitive conditioning and drug addiction. In J. P. Aggleton (Ed.). The amygdala: a functional analysis (pp. 353–390). ((2nd ed)). New: York Oxford University Press. Rolls, E. T. (2013). The mechanisms for pattern completion and pattern separation in the hippocampus. Frontiers in Systems Neuroscience, 7, 74. Rolls, B. J., Rolls, E. T., Rowe, E. A., & Sweeney, K. (1981). Sensory specific satiety in man. Physiology & Behavior, 27, 137–142. Rolls, B. J., Rowe, E. A., Rolls, E. T., Kingston, B., Megson, A., & Gunary, R. (1981). Variety in a meal enhances food intake in man. Physiology & Behavior, 26, 215–221. Rosenkranz, J. A., & Grace, A. A. (2002). Cellular mechanisms of infralimbic and prelimbic prefrontal cortical inhibition and dopaminergic modulation of basolateral amygdala neurons in vivo. Journal of Neuroscience, 22, 324–337. Saunders, J. P. A. R. C. (2000). The amygdala—what's happened in the last decade? In J.

of a neural circuit that suppresses appetite. Nature, 503, 111-+. Chen, G. F., Wang, L. P., & Tsien, J. Z. (2009). Neural population-level memory traces in the mouse hippocampus. Plos One, 4. Chiang, A.-S., Lin, C.-Y., Chuang, C.-C., Chang, H.-M., Hsieh, C.-H., Yeh, C.-W., ... Hwang, J.-K. (2011). Three-dimensional reconstruction of brain-wide wiring networks in Drosophila at single-cell resolution. Current Biology: CB, 21, 1–11. Craig, W. (1918). Appetites and aversions as constituents of instincts. The Biological Bulletin, 34, 91–107. Csicsvari, J., Hirase, H., Czurko, A., Mamiya, A., & Buzsaki, G. (1999). Oscillatory coupling of hippocampal pyramidal cells and interneurons in the behaving Rat. Journal of Neuroscience, 19, 274–287. Csicsvari, J., Jamieson, B., Wise, K. D., & Buzsaki, G. (2003). Mechanisms of gamma oscillations in the hippocampus of the behaving rat. Neuron, 37, 311–322. Cui, Z. Z., Lindl, K. A., Mei, B. Q., Zhang, S., & Tsien, J. Z. (2005). Requirement of NMDA receptor reactivation for consolidation and storage of nondeclarative taste memory revealed by inducible NR1 knockout. European Journal of Neuroscience, 22, 755–763. Davis, M. (1992). The role of the amygdala in conditioned fear. In J. P. Aggleton (Ed.). The amygdala: neurobiological aspects of emotion, memory, and mental dysfunction (pp. 255–306). New York: Wiley-Liss. Dayan, P., & Balleine, B. W. (2002). Reward, motivation, and reinforcement learning. Neuron, 36, 285–298. de Araujo, I. E., Gutierrez, R., Oliveira-Maia, A. J., Pereira, A., Nicolelis, M. A. L., & Simon, S. A. (2006). Neural ensemble coding of satiety states. Neuron, 51, 483–494. Everitt, B. J., Cardinal, R. N., Parkinson, J. A., & Robbins, T. W. (2003). Appetitive behavior – Impact of amygdala-dependent mechanisms of emotional learning. Amygdala in Brain Function: Basic and Clinical Approaches, 985, 233–250. Fletcher, M. L., Ogg, M. C., Lu, L., Ogg, R. J., & Boughter, J. D., Jr. (2017). Overlapping representation of primary tastes in a defined region of the gustatory cortex. Journal of Neuroscience, 37, 7595–7605. Fontanini, A., Grossman, S. E., Figueroa, J. A., & Katz, D. B. (2009). Distinct subtypes of basolateral amygdala taste neurons reflect palatability and reward. Journal of Neuroscience, 29, 2486–2495. Fries, P. (2009). Neuronal gamma-band synchronization as a fundamental process in cortical computation. Annual Review of Neuroscience, 32, 209–224. Geschwind, N. (1965). Disconnexion syndromes in animals and man I. Brain, 88, 237–294. Grace, A. A., & Rosenkranz, J. A. (2002). Regulation of conditioned responses of basolateral amygdala neurons. Physiology & Behavior, 77, 489–493. Guven-Ozkan, T., & Davis, R. L. (2014). Functional neuroanatomy of Drosophila olfactory memory formation. Learning & Memory, 21, 519–526. Hallem, E. A., & Carlson, J. R. (2006). Coding of odors by a receptor repertoire. Cell, 125, 143–160. Hanamori, T., Kunitake, T., Kato, K., & Kannan, H. (1998b). Responses of neurons in the insular cortex to gustatory, visceral, and nociceptive stimuli in rats. Journal of Neurophysiology, 79, 2535–2545. Hanamori, T., Kunitake, T., Kato, K., & Kannan, H. (1998a). Neurons in the posterior insular cortex are responsive to gustatory stimulation of the pharyngolarynx, baroreceptor and chemoreceptor stimulation, and tail pinch in rats. Brain Research, 785, 97–106. Herry, C., Ciocchi, S., Senn, V., Demmou, L., Muller, C., & Luthi, A. (2008). Switching on and off fear by distinct neuronal circuits. Nature, 454, 600–606. Holland, P. C., Petrovich, G. D., & Gallagher, M. (2002). The effects of amygdala lesions on conditioned stimulus-potentiated eating in rats. Physiology & Behavior, 76, 117–129. Hommel, J. D., Trinko, R., Sears, R. M., Georgescu, D., Liu, Z. W., Gao, X. B., ... DiLeone, R. J. (2006). Leptin receptor signaling in midbrain dopamine neurons regulates feeding. Neuron, 51, 801–810. Hromadka, T., Deweese, M. R., & Zador, A. M. (2008). Sparse representation of sounds in the unanesthetized auditory cortex. PLOS Biology, 6, e16. Hsu, T. M., Hahn, J. D., Konanur, V. R., Noble, E. E., Suarez, A. N., Thai, J., ... Kanoski, S. E. (2015). Hippocampus ghrelin signaling mediates appetite through lateral hypothalamic orexin pathways. Elife, 4. Isaacson, J. S. (2010). Odor representations in mammalian cortical circuits. Current Opinion in Neurobiology, 20, 328–331. Jefferis, G. S., & Hummel, T. (2006). Wiring specificity in the olfactory system. Seminars in Cell and Developmental Biology, 17, 50–65. Kadohisa, M., Rolls, E. T., & Verhagen, J. V. (2005a). Neuronal representations of stimuli in the mouth: The primate insular taste cortex, orbitofrontal cortex and amygdala. Chemical Senses, 30, 401–419. Kadohisa, M., Verhagen, J. V., & Rolls, E. T. (2005b). The primate amygdala: Neuronal representations of the viscosity, fat texture, temperature, grittiness and taste of foods. Neuroscience, 132, 33–48. Katz, D. B., Simon, S. A., & Nicolelis, M. A. (2002). Taste-specific neuronal ensembles in the gustatory cortex of awake rats. Journal of Neuroscience, 22, 1850–1857. Killcross, A., Everitt, B., & Robbins, T. (1998). Dissociable effects of excitotoxic lesions of amygdala sub-nuclei on appetitive conditioning. Journal of Psychopharmacology, 12, A4. Klausberger, T., & Somogyi, P. (2008). Neuronal diversity and temporal dynamics: The unity of hippocampal circuit operations. Science, 321, 53–57. Kluver, H., & Bucy, P. C. (1997). Preliminary analysis of functions of the temporal lobes in monkeys (Reprinted from Archives of Neurology and Psychiatry, vol 42, pg 979, 1939). Journal of Neuropsychiatry and Clinical Neurosciences, 9, (1997)). Ko, K. I., Root, C. M., Lindsay, S. A., Zaninovich, O. A., Shepherd, A. K., Wasserman, S. A., ... Wang, J. W. (2015). Starvation promotes concerted modulation of appetitive olfactory behavior via parallel neuromodulatory circuits. Elife, 4. LeDoux, J. E. (2000). Emotion circuits in the brain. Annual Review of Neuroscience, 23,

274

Neurobiology of Learning and Memory 155 (2018) 261–275

J. Liu et al.

38, 669–671. Tsien, J. Z. (2016). Cre-Lox neurogenetics: 20 years of versatile applications in brain research and counting. Frontiers in Genetics, 7, 19. Turner, B. H., & Herkenham, M. (1991). Thalamoamygdaloid projections in the rat – A test of the amygdalas role in sensory processing. Journal of Comparative Neurology, 313, 295–325. Wise, R. A. (2006). Role of brain dopamine in food reward and reinforcement. Philosophical Transactions of the Royal Society B-Biological Sciences, 361, 1149–1158. Wu, Q., Clark, M. S., & Palmiter, R. D. (2012). Deciphering a neuronal circuit that mediates appetite. Nature, 483 594 U112. Xie, K., Fox, G. E., Liu, J., Lyu, C., Lee, J. C., Kuang, H., ... Tsien, J. Z. (2016). Brain computation is organized via power-of-two-based permutation logic. Frontiers in Systems Neuroscience, 10, 95. Xie, K., Fox, G. E., Liu, J., & Tsien, J. Z. (2016). 512-channel and 13-region simultaneous recordings coupled with optogenetic manipulation in freely behaving mice. Frontiers in Systems Neuroscience, 10, 48. Zhan, C., & Luo, M. (2010). Diverse patterns of odor representation by neurons in the anterior piriform cortex of awake mice. Journal of Neuroscience, 30, 16662–16672. Zhang, H. M., Chen, G. F., Kuang, H., & Tsien, J. Z. (2013). Mapping and deciphering neural codes of NMDA receptor-dependent fear memory engrams in the hippocampus. Plos One, 8. Zhang, Y., Hoon, M. A., Chandrashekar, J., Mueller, K. L., Cook, B., Wu, D., ... Ryba, N. J. (2003). Coding of sweet, bitter, and umami tastes: Different receptor cells sharing similar signaling pathways. Cell, 112, 293–301.

P. Aggleton (Ed.). The amygdala: a functional analysis (pp. 1–30). ((2nd ed)). New: York Oxford University Press. Scott, K. (2005). Taste recognition: Food for thought. Neuron, 48, 455–464. Simon, S. A., de Araujo, I. E., Gutierrez, R., & Nicolelis, M. A. (2006). The neural mechanisms of gustation: A distributed processing code. Nature Reviews Neuroscience, 7, 890–901. Sohal, V. S., Zhang, F., Yizhar, O., & Deisseroth, K. (2009). Parvalbumin neurons and gamma rhythms enhance cortical circuit performance. Nature, 459, 698–702. Stanley, B. G., Urstadt, K. R., Charles, J. R., & Kee, T. (2011). Glutamate and GABA in lateral hypothalamic mechanisms controlling food intake. Physiology & Behavior, 104, 40–46. Stettler, D. D., & Axel, R. (2009). Representations of Odor in the Piriform Cortex. Neuron, 63, 854–864. Stujenske, J. M., Likhtik, E., Topiwala, M. A., & Gordon, J. A. (2014). Fear and safety engage competing patterns of theta-gamma coupling in the basolateral amygdala. Neuron, 83, 919–933. Szczypka, M. S., Rainey, M. A., Kim, D. S., Alaynick, W. A., Marck, B. T., Matsumoto, A. M., & Palmiter, R. D. (1999). Feeding behavior in dopamine-deficient mice. Proceedings of the National Academy of Sciences of the United States of America, 96, 12138–12143. Tschop, M., Smiley, D. L., & Heiman, M. L. (2000). Ghrelin induces adiposity in rodents. Nature, 407, 908–913. Tsien, J. Z. (2015b). Principles of intelligence: on evolutionary logic of the brain. Frontiers in Systems Neuroscience, 9, 186. Tsien, J. Z. (2015a). A postulate on the brain's basic wiring logic. Trends in Neurosciences,

275