|Home | About | Journals | Submit | Contact Us | Français|
Animals learn which foods to ingest and which to avoid. Despite many studies, the electrophysiological correlates underlying this behavior at the gustatory-reward circuit level remain poorly understood. For this reason, we measured the simultaneous electrical activity of neuronal ensembles in the orbitofrontal cortex, insular cortex, amygdala and nucleus accumbens while rats licked for taste-cues and learned to perform a taste-discrimination Go/No-Go task. This study revealed that rhythmic licking entrains the activity in all these brain regions, suggesting that animal's licking acts as an “internal clock signal” against which single spikes can be synchronized. That is, as animals learned a Go/No-Go task there were increases in the number of licking coherent neurons as well as synchronous spiking between neuron pairs from different brain regions. Moreover, a subpopulation of gustatory cue-selective neurons that fired in synchrony with licking exhibited a greater ability to discriminate among tastants than nonsynchronized neurons. This effect was seen in all four recorded areas and increased markedly after learning, particularly after the cue was delivered and before the animals made a movement to obtain an appetitive or aversive tastant. Overall, these results show that throughout a large segment of the taste-reward circuit, appetitive and aversive associative learning improves spike-timing precision, suggesting that proficiency in solving a taste discrimination Go/No-Go task requires licking-induced neural ensemble synchronous activity.
Animals learn to use multisensory cues to procure appetitive and avoid aversive food sources. During feeding, the animal's behavior is reinforced by a rewarding (acceptance) or aversive (rejection) stimulus. Both appetitive and aversive learning depend on interactions between components of a taste-reward circuit that includes the orbitofrontal cortex (OFC), the amygdala (AMY) (Paton et al., 2006), the nucleus accumbens (NAcc) and the insular cortex (IC) (Gottfried et al., 2003; Stalnaker et al., 2007). To date, however, relatively little is known about how this circuit as a whole mediates an animal's ability to learn to select appropriate foods.
The taste-reward circuit is comprised of a highly interconnected neural network (Cavada et al., 2000) that is involved with multiple aspects of ingestive behavior, associative learning and reward expectation (de Araujo et al., 2006; Rolls, 2007). The IC, which contains the primary gustatory cortex, is a multimodal area that processes taste, visceral, somatosensory and hedonic information (Katz et al., 2001; Accolla and Carleton, 2008). The OFC integrates information from several primary sensory systems (Cavada et al., 2000) and assays the relative reward value of sensory stimuli, including those associated with foods (Tremblay and Schultz, 1999; Rolls, 2007). Dysgranular OFC also encodes the economic value of foods (Padoa-Schioppa and Assad, 2006). Rodent OFC and AMY neurons fire selectively to sensory cues according to their predictive value of a reward (Schoenbaum et al., 1998). The NAcc is thought to translate emotional-motivational information, generated by limbic regions, into movements and actions to obtain food and to avoid punishment (Mogenson et al., 1980). Neurons from the IC, OFC and AMY project to NAcc, defining a circuit that processes information about gustatory cues, their predictive reward value and motivational significance (Pecina and Berridge, 2005).
Rodents actively sample sensory stimuli using rhythmic and stereotypic behaviors such as sniffing (Kepecs et al., 2006), whisking (Fanselow and Nicolelis, 1999) and licking (Travers et al., 1997), at roughly similar frequencies (theta band, 4–12Hz). It has been suggested that sniffing and whisking allow animals to process continuous real-world sensory stimuli into discrete chunks or cycles (Kleinfeld et al., 2006), with each cycle serving as a temporal frame to synchronize neural activity across multiple brain structures (Nicolelis et al., 1995). In contrast to other sensory systems, there is a paucity of information regarding the potential physiological role of neuronal oscillations in taste-guided behaviors. Thus, it remains unknown whether rhythmic licking may function as an internal temporal frame to synchronize the firing of populations of neurons throughout the taste-reward circuit. Equally untouched is the question of whether neuronal synchronous firing would enhance the ability of the taste-reward circuitry to discriminate among cues. Here, we addressed these two fundamental questions by simultaneously recording the activity of neural ensembles located in the IC, OFC, AMY and NAcc, while rats learned to perform a new version of a taste discrimination Go/No-Go response task. Briefly, we found that, throughout the taste-reward circuit, learning improves spike-timing precision, suggesting that widespread licking-induced neuronal synchronicity is important in the solution of taste-guided discrimination tasks.
Seven male Long-Evans rats (350–450 g) were obtained from Harlan (Harlan Laboratories, USA). Surgical procedures followed the methods described elsewhere (de Araujo et al., 2006). A movable microwire array comprised of 16 formvar-coated tungsten wires (35-μm diameter) was unilaterally implanted in each of the four brain regions, OFC, NAcc, IC and AMY. The locations of microelectrode implants are found in Figure S1. One of the seven subjects received bilateral implants only in OFC and IC. In all experiments, one electrode was used as reference for FEC software (Plexon Inc. Dallas, TX). The electrode arrays were advanced ~40 μm such that for each experimental session a new set of single units was recorded. All protocols were approved by the Duke University Institutional Animal Care and Use Committee.
Procedures were essentially the same as described previously (de Araujo et al., 2006). That is, neural electrical activity from the four implanted brain areas was simultaneous recorded using a Multichannel Acquisition Processor (Plexon Inc. Dallas, TX). Spikes were sorted using Offline Sorter software and the stability of waveform shape across a session was confirmed by using the Waveform Tracker software (Plexon Inc. Dallas, TX). Further details are found in Figure S2.
All experiments were performed in an operant box that contained two widely spaced (~17 cm) drinking compartments. Each compartment contained a photobeam lickometer (MedAssociates, VT) that was used to register each lick. The sipper tube was comprised of a bundle of 12 20-gauge stainless steel tubes cemented together in a larger steel tube (inner diameter, 7.5 mm), connected to solenoid valves (Stapleton et al., 2006). Each solution was maintained under air pressure to ensure a single delivery of ~20 μl drop of liquid (in 10 ms). Access to each compartment was restricted with a sliding door operated with a hydraulic pump. The behavioral task and details about the quantification of learning are shown in Figure 1. Briefly, prior to surgery, all rats were trained on the taste discrimination task. A trial began when the door opened allowing access to the cue-compartment. Rats were then allowed to lick the tube and received a drop of water at the fourth lick. Rats were required to continue licking the dry sipper five additional times and on the sixth lick, they received a cue consisting of a drop of either 0.1M NaCl or 0.1M monopotassium L-glutamate (MPG). These two tastants were chosen because rodents can distinguish between them independently of sodium content (Maruyama et al., 2006). In the first session, one of these tastants was randomly chosen as either the positive (C+) or negative (C−) cue. After cue delivery, access to the outcome compartment was allowed and, although no more licks were required, rats continued licking, on average, 1.2 s before moving to the outcome port. Rats had 10 s after cue delivery to move from the cue compartment to the outcome compartment, where after three empty licks they could receive the signaled outcome. If after 10 s no response was observed, both doors closed thereby terminating the trial. In the outcome compartment, the positive cue (C+) was associated with three deliveries of 0.4M sucrose and the negative cue (C−) signaled the availability of up to three deliveries of 1 mM QHCl.
Figure S6A shows a schematic representation of the behavioral strategies employed to solve the task. In our Go/No-Go task, water deprived animals initially adopt a behavioral strategy that favors responding to both cues with a “Go” response, (“always Go behavioral strategy”). Note that at the beginning of the session the predictive value of either of the cues is unknown to the rat, thus exploring the outcomes of both cues with a “Go” response will provide them with the experience needed to learn which cue predicts a reward (sucrose) and which predicts punishment (quinine). After several trials in which learning had taken place, rats switched to a discriminative behavioral response pattern, where they respond to C− cue by withholding a “Go” response (correctly avoiding quinine), whereas rats maintained making a “Go” response to C+ cue. Rats respond correctly for nearly 100% of the C+ trials throughout the entire session (Figure S6B). Therefore, since both trials types are intermingled, it is not until the rat starts to avoid quinine that it is possible to know from a behavioral viewpoint that the rat has learned that the C+ predicts sucrose. Indeed, rats that do not learn aborted the task and kept drinking for the cues but stopped making a Go-response or simply stopped working. For this reason, we used both cues in order to obtain the “learning trial” (see definition below “Behavioral analysis of learning” and Figs.S6C–D).
This initial training was followed by serial reversal training in which the response contingencies of the tastants were reversed. Once the rats learned the new contingency (usually after one or two extra-sessions covering 2 or more days), subsequent reversals were issued. Electrode implantation did not occur until at least three complete reversal cycles were learned. Therefore, animals were familiar with the taste-discrimination task and with the tastants used as cues, but after each reversal they unlearned or suppressed previous cue-outcome contingencies so that their behavior could dynamically adapt to the new task rules.
During the recordings, the animal's behavioral performance was analyzed online. If the state space algorithm (described below) determined that the subject learned, the trials were subdivided into Pre and Post-learning phases. Furthermore, if in the Post-learning phase a stable performance was observed (defined as correct performance above 80% during at least 60–100 trials), then another within-session reversal was issued. This training protocol continued as long as quality recordings were obtained. In these later sessions, however, rats unlearned the original meaning of the cues but never, in the same session, exhibited a complete reversal. The results from the within-session reversal are not presented in this manuscript. However, we note that the changes in the Post-learning epoch described in this manuscript were significantly reduced in a few trials after the within-session reversal (Gutierrez, Simon and Nicolelis, unpublished results).
Since rats in these experiments were trained in a continuous serial-reversal task, we note that our operant definition of “Pre-learning” and “Post-learning” phases refers specifically to the collection of trials before and after the “learning trial (see definition below)”. Thus, this definition does not imply occurrence of learning of a “novel” cue-pair association.
We used a state-space algorithm (Smith et al., 2004) to determine within a single session the occurrence of a “learning trial.” Briefly, this algorithm modeled learning as a dynamic process from the time series of binary responses across trials, where “0” represents an incorrect response and “1” a correct response. Smith et al., 2004 demonstrated that this method estimates the learning curve and its confidence interval more accurately and earlier than several currently accepted methods (i.e., the moving average method). In the state-space algorithm, “the learning trial” is defined as the first trial in which, from the point of view of an ideal observer (that is, one that has knowledge of the performance throughout the entire session), there is a reasonable certainty (95% confidence) that the subject is performing above chance level, and that correct performance will be maintained for the rest of the session (see Figure 1B, right). In this model, the learning curve (confidence interval) and certainty were estimated using the expectation maximization (EM) algorithm (Smith et al., 2004). However, this algorithm does not treat learning as a gradual process but rather as a statistical event. In this regard, the “learning trial” does not mean that learning occurred on that particular trial but rather it reflects the trial at which there is sufficient evidence that the subject began performing above chance level and from which its correct performance remained for the rest of the session.
The sequence of correct and incorrect behavioral responses was obtained following the standard nomenclature; a correct trial is comprised of both “Hit” and “Correct Rejection” trial types. (A “Hit” is when C+ is delivered and the subject makes a “Go” response and drinks at least one drop of sucrose. A “Correct Rejection” occurs when C− is delivered and the subject withholds a “Go” response to avoid punishment, quinine HCl). An incorrect trial comprises the pool of “False Alarms” and “Miss” trials types. A “False Alarm” occurs when C− is followed by a “Go” response, meaning the subject will drink at least one drop of quinine. A “Miss” occurs when C+ is followed by a No-Go response, and the subject will not receive the sucrose reward. By using these definitions, together with this algorithm, a quantitative estimate of the “learning trial” was obtained (see Figs. 1B–C). In this task, from the first lick in the cue-compartment the mean inter-trial-interval was 24.5 seconds ± 1.2. On average, a single session lasts for 70.9 min ± 5.4, and rats achieved the learning trial in 16.2 min ± 1.3.
All data analysis was performed using MATLAB and “R.” Unless otherwise indicated we employed the standard error of the mean.
Multi-taper spectral analysis and coherence were computed by segmenting two univariate binned point processes (licking and spike PSTHs) into chunks (Jarvis and Mitra, 2001). The coherence “C” between licking and the spike trains was computed using the formula: C(f) = Ixy/√(Ixx Iyy), where Ixx represents the spectrum of licking behavior, Iyy the spectrum of neuronal activity and Ixy is the cross-spectrum of licking and spike spectrum. Note that the coherence “C” is normalized to range between 0 and 1. Finally, f is the frequency where coherence was computed (4–10 Hz). This frequency band corresponds to that normally observed in freely licking behavior (Spector et al., 1998). The confidence interval of the coherence, C(f), and significance threshold (at alpha 0.05%) were computed with a Jackknife method and finite size corrections using the procedures developed by (Jarvis and Mitra, 2001). A neuron was classified as licking-coherent only if its lower confidence interval (95%) crossed the significance threshold. Figure S3 shows an example of coherence calculation in a single neuron. Each chunk of data corresponded to the first lick in the cue-port up to 2.5 s of activity. This time essentially represents the entire period that rats licked the sipper tube in the cue-compartment (Fig. 2). The maximum coherence value and frequency were computed using the Chronux 1.50 software package (www.chronux.org). Multi-taper coherence was also computed cycle-by-cycle in windows of 170 ms aligned to each lick in the cue-compartment (Figs. 4D, E; Fig. S4). This window is large enough to observe theta rhythm. Tapers 1 and 2 were used for this analysis, as were frequencies between 0 to 60 Hz (see Fig. S4A). In Figure 9C, we determined whether After Cue (AC) cue-selective cells were licking-coherent, thus we employed a single window from 0.2 to 1 s, after cue delivery. In the Post-learning phase, and both C+ and C− trials were combined.
In this analysis, a neuron from one region was compared with a neuron simultaneously recorded in a different brain area. For instance, if two licking-coherent neurons were recorded in OFC, three in AMY and two in IC then there were six OFCAMY pairs, four OFC-IC pairs and six AMY-IC pairs. The variability in each neuron's response to repeated presentations of the same stimulus was removed by subtracting the shift predictor from the raw cross-correlation. The shift predictor was constructed by shuffling the trials of one neuron (reference cell), this procedure was repeated 100 times, and the average cross-correlation values were subtracted from the original raw cross-correlations. Then, the absolute peak value of this cross-correlation was identified (Grossman et al., 2008).
Neuronal firing modulations shown in Figure 2 were identified using a Wilcoxon rank sum test at an alpha level <0.05.
Phasic neuronal firing modulations were identified using a Wilcoxon rank sum test at an alpha level <0.05. Briefly, the spike counts over the first 150 ms after cue-delivery were compared against the neuronal firing produced during the 150 ms following a prior empty lick. In cases where the inter-lick interval of the previous lick overlapped with cue delivery, the next to last dry lick before the cue delivery was employed as the baseline. This approach determined if cue delivery evoked a significant neuronal response relative to that produced by an empty lick, thus providing a rationale to separate oromotor-driven firing modulations from genuine taste cue-evoked neuronal activity (Fig. 2).
To determine whether the neurons responsive to the delivery of the cues were also responsive to other gustatory stimuli (e.g. water, sucrose, quinine), we followed the same procedure outlined above for the cues, namely by calculating the differences in activity between the wet and dry licks. This criterion allowed us to test whether neurons responded to other gustatory stimuli, but not whether these evoked responses were significantly different among gustatory stimuli (Table S1). It is important to emphasize that in the present study we did not explore the contribution of a temporal code (in 150 ms), which has been shown to also carry taste information (Di Lorenzo and Victor, 2003; Stapleton et al., 2006).
Neurons with excitatory activity before cue delivery were identified using a Wilcoxon rank sum test at an alpha level < 0.05. Neurons with this firing pattern had to meet two selection criteria: (1) They needed to show a significant inhibition during the first lick after cue delivery. Thus, we compared the spike counts over the first 150 ms after cue-delivery against the firing rate produced during 150 ms following a prior empty lick. (2) The spike counts over the first 500 ms before cue-delivery had to be significantly greater than the firing rate in the subsequent 500 ms after cue-delivery.
Neurons that fired after cue delivery were also identified using a Wilcoxon rank sum test at an alpha level < 0.05. These neurons showed a significant increase in firing rate during the first second after cue delivery, in comparison with the second before cue-delivery. Neurons that showed phasic activity (Upon Cue responses) were not included in this category. In Figure 2, both AC and BC neurons were ranked according to their root mean square of their firing rate.
The onset and peak latency were computed from peri-event stimulus histograms (PSTHs, bin width 1 ms) using a 200 ms window after cue delivery. The PSTHs were smoothed through convolution with a Gaussian function with a standard deviation of 10 ms. The onset latency was defined as the first bin exceeding 3 standard deviations (s.d.) above background activity. The onset latency and average peak time are displayed as a function of cue type, learning phase and brain area in Table S2.
It is generally accepted that neuronal firing rates follow a Poisson distribution. Therefore to determine whether spiking activity discriminated between cues in each learning phase, the spike counts were modeled using a Poisson generalized linear model (Stapleton et al., 2006). The GLM is more appropriate to use than a two-way ANOVA test, which assumes the spike trains are normally distributed. Details of the Poisson GLM model can be found in Supplementary Methods –Data Analyses, GLM.
The choice probability represents the accuracy with which an ideal observer, given previous knowledge of the firing rate distributions, could predict from the response of the neuron the subject's anticipated decision (Go/No-Go). The choice probability was computed with a standard receiver operating characteristic (ROC) curve, since this is the test that is classically used to detect choice probability (Britten et al., 1996). The ROC analysis takes on values between 0 and 1, where 0 or 1 represent perfect predictive power and values near 0.5 reflect a random association between neuronal response and behavioral response and indicate that the neuronal response provides no information for predicting the animal's choice. We analyzed activity only from C− trials in the Post-learning phase, since in those trials rats performed both Go and No-Go responses, in accordance with the inclusion criteria described by Setlow et al. (2003). For Upon Cue (UC) cue-selective neurons, we integrated the activity over the first 150 ms and for the After Cue (AC) cue-selective neurons over the interval 0.2–1 seconds. Significance of the ROC curve was computed using a permutation test, at an alpha of 0.05%. In this test, the labels of Go/No-Go responses were shuffled and the ROC value was calculated from 1000 permuted samples.
The ability of single neurons to discriminate the cues as a function of task performance was quantified using a non-parametric method for statistical pattern recognition called optimized learning vector quantization (see Supplementary Methods –Data Analyses, OLVQ). For all OLVQ analyses, spike trains were binned in 10 ms bins and both firing rates and spike timing information were always used for cue-discrimination. A “moving window” analysis was used to assess the time-course of information about the cues. OLVQ was applied sequentially to a 300 ms window of single unit activity that “moved” in 50 ms steps through 0.5 s before and 1.0 s after the time of cue delivery at 0 sec (Figs. 7D, 9A, 9B). This analysis reveals a continuous quantitative readout (50 ms steps) of the recorded population's ability to distinguish between the C− and C+ cues.
Adult rats were trained in a taste discrimination Go/No-Go response task (Fig. 1A). To solve this Go/No-Go task, rats needed to determine which taste cue predicted the reward (C+) and which predicted the aversive outcome (C−). Subjects initially responded to each cue by making a “Go” response, and received the corresponding outcome until the cues acquired a predictive value. At this point, rats learned to avoid quinine (correct rejection, No-Go response), while continuing to respond (Go response) following the positive cue (Hit) to obtain sucrose (also see Fig. S6A).
In an individual session, learning was quantified using a state-space algorithm in which trials were categorized as either Pre-learning or Post-learning (Fig. 1B–C, also see Figure S5 for all learning curves). We found that for C− trials the “learning trial” was at 95.4±6.8 trials (Fig. S6C), that is it occurred significantly later than the learning trial obtained by combining both cues (56.8±4.9 trials; t-test(88)=−4.5,p-value<0.0001). Thus, the “learning trial” was estimated by combining both cues together (Fig. S6D). Parenthetically, the state-space definition of the onset of learning occurs earlier than that found using the often applied 90% trials correct criterion (73.1±5.9 trials; t-test(88)=−2.1,p-value<0.037; (Schoenbaum et al., 1998)). In addition, during the Pre-learning phase, rats performed at chance level 52.8±0.005%, but in the Post-learning phase, subjects performed correctly in 88.4±0.008% of the trials (mean ± s.e.m.; n=45 sessions). Thus showing that they can extract taste information from one lick of a tastant and use this information to guide their behavior (Halpern and Tapper, 1971).
During the delivery of the two hedonically positive tastants used as cues (0.1M NaCl; 0.1M MPG), animals licked the tube in a stereotypic rhythmic manner. For these two cues (1 s after cue delivery- in the Delay epoch, see below), the mean lick frequency was not significantly different C− = 5.9±0.1 Hz, and C+ = 6.0 ±0.09 (F(1,178)=0.99, p=0.32; n=45). In addition, during the Delay epochs of both the Pre- and Post- learning phases, rats did not discriminate among the cues by changing their licking patterns (see Fig. S7).
As expected, during both Pre- and Post-learning phases, the rats licked more times after the delivery of sucrose (Hit) than quinine (False Alarm). By calculating the probability that a trial contained only a single lick in the first 1.0 s after the first delivery of sucrose or quinine, we found that throughout this period rats continued licking for sucrose. Indeed, in a small fraction of the Hit trials, rats received only one drop of sucrose (Pre-learning 0.016±0.007 and Post-learning 0.008±0.004 fraction trials; Wilcoxon test p= 0.93 n.s.). In contrast, the fraction of False Alarm trials in which they licked only once after quinine delivery significantly increased, from 0.30±0.045 (pre-learning) to 0.63±0.042 (Post-learning; Wilcoxon test p<0.001). These results demonstrate that rats can learn to detect (and reject) quinine in a single lick (~150 ms).
We recorded the activity of 1,110 single neurons distributed across four taste-reward structures: OFC, IC, Nacc and AMY (see Fig. S1). The distribution of neurons in these areas was: OFC (n=449), NAcc (n=306), IC (n=243) and AMY (n=112).
Figure 2 shows a stack of 526 (of the 1,110) normalized Z-score PSTHs for all the OFC, IC, AMY and NAcc neurons that exhibited a significant modulation in firing rate from −1.5 to 1 s centered around cue delivery (0s). During this 2.5 s interval, the PSTHs included three main epochs of the Go/No-Go task (Anticipation, Cue [C+ left panel, C− right panel], and Delay). The Anticipation epoch consisted of three dry licks (L) followed by water (W) delivery, followed by five more dry licks. The Cue epoch is delineated by the vertical white lines (representing a single lick ~150ms) and contains the response to the delivery of the cue at time 0 s. This was followed by a Delay epoch that consisted of several more dry licks (L).
The observed patterns of neuronal firing modulation in these components of the taste-reward circuit were classified into two broad, but not mutually exclusive, categories: licking coherent (neurons 1–278) and event related (neurons 279–526). A total of 414 neurons showed a significant coherence with licking (see below), and of these, 278 were plotted in the licking-coherent category (Fig. 2). The remaining 136 licking coherent neurons were plotted in the event related category since they fired during a particular epoch in relation to cue delivery. Therefore, 112 neurons with no licking-coherent activity were classified as event-related (see Figure S8). The event related neurons (136+112 = 248 in total) were further characterized by their responses in relation to the cue delivery. These included neurons that fired before cue delivery (BC), upon cue delivery (UC), and after cue delivery (AC). Below each category is described in detail. Inspection of the color-coded population of PSTHs indicates that throughout the four brain areas, the neuronal responses to cues may be entrained to the licking cycle.
Next, we analyzed the level in which the firing of each of the 1,110 neurons was coherent with rhythmic licking. In this context, coherence is defined as a measure of the interdependence of licking and neural activity in the relevant frequency domain (see Fig. S3). A coherence of 0 or 1 means that two signals are completely uncorrelated, or completely correlated in frequency and phase, respectively. The phase of the coherence ranges between ± π radians and indicates the extent to which a neural discharge follows the lick (set at 0 radians corresponding to the tongue contacting the sipper; Fig. 3A). Positive phase values (>0 to π radians) indicate that the action potentials of a given neuron tend to occur during the early phases of the lick cycle, meaning that such neurons would rapidly fire after the lick. For neurons that fire with a negative phase (−π to <0 radians), their discharges tend to peak later in the lick cycle (Figs. 3A, D). The distribution of the phase of coherence (average phase across all licks in the cue-epoch) for each brain region is shown in Figure 3C.
Neurons with licking-coherent activity were present in the four brain areas and in all subjects (Table S3). However, the proportion of licking-coherent neurons differed among brain areas (Fig. 3B; χ2(3)= 1.45, p < 0.0001); their distribution was IC (71%, 173/243), AMY (60%, 68/112), OFC (34%, 153/449), and NAcc (6.5%, 20/306). A comparison of the number of licking-coherent neurons observed when the animals were in the cue and outcome compartments are presented as Table S4 (see Supplementary Results -Coherence). The IC and the AMY exhibited the greatest percentage of these cells and both regions contained more such neurons than the OFC (p<0.0001; Fig. 3B). In turn, all three areas, IC, AMY and OFC, were significantly different from the NAcc (p<0.0001; Fig. 3B). The small proportion of coherent neurons recorded in NAcc is most likely due to the prominent inhibition observed in this region during consummatory behavior (Nicola et al., 2004; Roitman et al., 2008).
The average coherences in the different brain areas were: IC (0.25±0.009) and AMY (0.24±0.013) followed by the OFC (0.21±0.008) and NAcc (0.17±0.011). The IC and the AMY exhibited the greatest coherence values and both regions showed higher coherence than the OFC. In turn, all three areas, IC, AMY and OFC, were significantly different from the NAcc (F(3,410)=5.3, p=0.0012).
Figures 4A–C depicts representative examples from different brain areas that illustrate typical licking-coherent firing patterns. Figure 4A depicts a neuron from the AMY that fired throughout the behavioral task in phase with the animal's rhythmic licking. This neuron exhibited a maximum coherence with licking of 0.64 at 6.8Hz, and relative to the lick cycle, its action potentials were phase-locked at 2.6 radians. Figure 4B shows a response from the IC that, prior to water delivery, had virtually no firing. Nevertheless, upon water and the cue delivery, it produced a strong transient response that decayed in an oscillatory manner during the subsequent dry licks. Thus, in contrast to the neuron shown in Figure 4A which fired in phase throughout the trial, this neuron fired in phase with licking only after a stimulus was delivered.
Figure 4C shows an example of two simultaneously recorded neurons, one in AMY and the other in OFC, that commenced firing two licks after receiving the cue and whose action potentials were phase-locked to specific lick cycles (see vertical dashed lines). This type of response shows that the resulting synchronous activity was not a consequence of somatosensory input as might be the case for the example shown in Figure 4A. Instead, despite being located in two distinct brain areas these neurons fired in synchrony in a manner that covaried with licking. Thus, neurons across multiple brain structures may fire synchronously at specific lick cycles during relevant epochs of the behavioral task (see also Fig. S4).
To determine the proportion of neurons that fired in synchrony during each lick cycle at theta frequencies, we employed the multi-taper coherence analysis to establish that the proportion of neurons with significant coherence was larger after learning in the OFC, AMY and IC, but not in NAcc (Fig. 4D). This analysis also showed that the proportion of coherent neurons was greatest when the animals received either water or a tastant cue (Fig. 4E). These results suggest that licking may act to synchronize the activity of neurons from different brain regions in a dynamic and learning-dependent manner.
Next, we determined whether licking-induced oscillations enhance coincident spiking between neuron pairs from distinct brain areas. To address this issue, we analyzed the cross-correlations between pairs of licking-coherent neurons before and after learning (Grossman et al., 2008), during the first second after cue-delivery. Since no pair of licking-coherent cells was obtained in the NAcc, this area was not included in this analysis. Overall, during the Delay Epoch there were 168 neurons with significant coherent activity. From these, we recorded 37 pairs of OFC-AMY neurons, 43 OFC-IC pairs and 57 AMY-IC neuron pairs (see Fig. 5D). By measuring the maximum peak of the cross-correlation, we found that coincident spiking between licking coherent neurons increased significantly after learning (Pair t-test, p < 0.0001; Fig. 5). Such an increase was also observed in pairs of licking-coherent neurons from the OFC-AMY, OFC-IC and AMY-IC (Fig. 5D, “Licking-coherent”). As a control, we performed the same analysis for all pairs of neurons in which firing in the Delay Epoch was not coherent with licking (Fig. 5D, “Non-coherent”). In this subpopulation, there were 891 OFC-AMY, 1129 OFC-IC, and 272 AMY-IC neuron pairs. We observed that, after learning, coincident spiking between non-coherent OFC-AMY, OFC-IC and AMY-IC neuronal pairs also showed larger cross-correlation peaks (Fig. 5E). These correlation peaks, however, were significantly smaller than for licking-coherent cells (p-value<0.01; Fig. 5E). This result shows that, after learning, licking-induced oscillations increased the probability of coincident spiking between disparate brain regions, such as the OFC-AMY-IC. This effect, however, was much more pronounced among licking-coherent neurons.
In addition to licking-coherent neurons, the neural populations also contained cells that fired in relation to a salient event such as the delivery of a cue. As shown in Figure 2, we further classified the activity of this population according to its timing relative to the delivery of the cue (Before (BC), Upon (UC) and After (AC) cue delivery).
The BC population contained 22 neurons. Their firing modulations were characterized by a buildup of activity prior to cue delivery, followed by a rapid return to baseline after cue delivery (~150 ms; Fig. 2). In this regard, even though the rats continued licking upon and after cue delivery, the decrease in neuronal activity was maintained, indicating that this neuronal population did not encode purely oromotor information. Interestingly, BC neurons were found only in the OFC and IC. In the OFC, only 4% (18/449) of the neurons exhibited such firing behavior. Overall, 88% (16/18) of these OFC neurons displayed significant phase-locked activity with respect to licking (mean coherence 0.226±0.027 s.e.m.; mean frequency = 6.82±0.16 Hz). In the IC, only 1.6% (4/243) of the neurons showed the same kind of anticipatory activity (see BC; Fig. 2). All four IC neurons showed significant coherence with licking.
Figure S9 illustrates the response of an OFC neuron in the post-learning phase in which the firing rate increased or ramped up prior to cue delivery. The activity increased in a sigmoidal manner from the time the subject entered the cue compartment, licked the dry sipper, received water, licked the dry sipper again and received either of the cues (C+ raster is shown). This firing pattern, which was not significantly different for either cue (see PSTH's below raster), transiently decreased in the first 100 ms post-cue and reached baseline levels 50 ms later. Figure S9B depicts the population PSTH of the above noted 18 OFC neurons. For this population, neuronal activity increased as the animal entered the cue compartment until water was delivered, whereupon it slightly decreased. The activity then increased again until a cue was delivered, whereupon the response rapidly decreased and returned to baseline in 137 ms, before the onset of the next lick. Overall, this population of OFC neurons may reflect the expectation of the delivery of salient liquids but not their identity.
We also determined whether the cue-anticipatory activity seen in the Post-learning phase was also present in the Pre-learning phase. In this regard, we compared the average population PSTH (from −3 to 0 sec before cue delivery, bin width 25 ms) of the 18 OFC neurons in the pre-learning phase (data not shown) against the population response over the same period in the post-learning phase (Fig. S9B). We found no significant differences in anticipatory activity between Pre- and Post-learning phases (Repeated Measure ANOVA, F (1,68) =0.18, p=0.67 n.s.). We also did not find any difference between cues (RM ANOVA, F (1,68) =0.0037, p=0.95 n.s.). Therefore, in both learning phases, this OFC population seems to anticipate the delivery of the cues (but not their identity). Likewise, the four IC neurons with cue anticipatory activity acted in a similar manner to the OFC neurons (data not shown).
Another neuronal subpopulation in our classification scheme, named UC, exhibited firing rates that transiently changed upon cue delivery (see UC; Fig. 2). These neurons constituted 15% (168/1110) of the neural population and were present in similar proportions in all four recorded brain areas (χ2(3)= 2.74, p = 0.9). About 50% (84/168) of all UC neurons displayed phase-locked activity with respect to licking (mean coherence 0.24±0.013 s.e.m.; mean frequency = 6.8±0.09 Hz). To characterize the UC population (n=168), we determined the onset latency of phasic firing modulation by finding the first 1 ms bin in the PSTH that was at least 3 s.d. above baseline. This analysis revealed no significant difference among brain regions (F(3,248)=0.21, p=0.88). The mean onset latencies observed were OFC 47 ms ± 3, NAcc 44 ms ± 4, IC 49 ms ± 5 and AMY 42 ms ± 3, suggesting that there was a parallel activation of these four areas upon cue delivery (Table S2).
To obtain a better estimate of the dynamics of the UC neurons, we also analyzed their peak latencies. For each of the recorded brain areas this analysis revealed the existence of a characteristic temporal activation pattern (see Fig. S10). The time to maximum firing rate (peak) was significantly different among the four areas (F(3,248)=8.2, p=0.0001). Specifically, the peak firing activity of AMY neurons occurred earlier than peak responses from the OFC (p<0.0001), NAcc (p=0.0007) and IC (p=0.002). These data show that as a population, neurons in the AMY had a faster time to peak than OFC and NAcc neurons (Table S2). In the IC, the peak distribution exhibited a very broad activation pattern (see Fig. S10).
Interestingly, because the firing rate modulation of UC neurons was broadly tuned, 83% (140/168) of them could not be used to discriminate between the two cues (“non-cue selective UC”) (Table S1; Supplementary Results -Cue Epoch). Thus, even though most individual UC neurons could not discriminate the identity of these tastants, their increase in firing rate could at least mediate cue detection.
How then do rats discriminate among the cues to perform the taste guided behavioral task? In this context, we found that during learning, 17% (28/168) of the UC population developed cue-selectivity in the Post-learning phase (“cue-selective UC”). That is, although the firing rate of these neurons did not discriminate among cues in the Pre-learning phase, they did after learning occurred. Figure 6A, B show the development of cue-selectivity of an OFC neuron throughout the session. This neuron, which was not activated by licking the sipper per se, responded similarly to both cues in the Pre-learning phase, but developed a stronger phasic response for C+ in the Post–learning phase. Figure 6 also shows the average population of cue-selective UC neurons sorted by their preferred cue (Fig. 6C, prefer C+; Fig. 6D, prefer C−). Although most cue-selective neurons responded to both cues, their responses were often of different magnitudes. Moreover, the development of cue-selectivity was specific to the cues and was not observed during water delivery (see Fig. S11). We found no difference in the distribution of cue-selective UC neurons among brain areas: 29% (5/17) in the AMY, 17% (13/74) in the OFC, 16% (6/38) in the NAcc, and 10% (4/39) in the IC (χ2(3)=3.2, p=0.36). Overall, these data indicate that, upon cue delivery (~150 ms), 17% of the UC neurons contained detailed information about the associative significance of the cues and are dynamically reorganized upon learning.
A significant percentage of non-cue selective UC neurons also changed their firing modulation properties after learning (Fig. S12A). For instance, 26% (36/140) of UC neurons showed significant evoked responses to cue delivery in the Pre-learning phase, which became smaller and/or non-significant in the Post-learning phase (Fig. S12A –referred as Only-Pre). About 53% (74/140) of them developed significant cue-related firing activity only in the Post-learning phase (Only-Post) and 21% (30/140) displayed significant evoked responses in both learning phases (Pre-Post). The number of neurons from each category and brain region is shown in Figure S12B.
The Delay epoch encompasses a period in which animals had already sampled the cue, but continued to lick until they decided to make a Go/No-Go response. A small subpopulation of AC neurons (n=58) displayed a tonic firing pattern during the Delay epoch (see Fig. 2; neurons labeled “AC”). We observed that 8.5% (38/449) of the total number of OFC neurons exhibited this type of tonic firing pattern. In the other areas, the following percentages were found: NAcc 2.6% (8/306), IC 3.2% (8/243), and AMY 3.5% (4/112), indicating that the percentages of such responses were more frequently found in the OFC (χ2(3)=16.17, p=0.001). Overall, 55% (32/58) of these neurons displayed significant phase-locked activity with respect to licking (mean coherence 0.27±0.02 s.e.m.; mean frequency 6.4±0.12 Hz).
In addition to the AC neurons described above, a much larger population of AC neurons (n = 196) changed their firing pattern upon learning. To identify these, we searched for neurons that during the Delay epoch, or 200–1000 ms after cue-delivery, discriminated between the cues. We named such neurons AC cue-selective. The example in Figure 7 shows that no activity was evoked during the initial trials in the pre-learning phase, but as the animal learned the task, the neuron's firing activity increased during the Delay epoch. In particular, in the Post-learning phase, the neuron's activity for C+ trials was markedly enhanced when compared to the C− trials (Fig. 7B). In general, we also found that cue-selective AC neurons discriminated among cues in both Pre- and Post-learning phases, although the vast majority (77%, 152/196) was selective only in the Post-learning phase (Table S5).
The OFC contained the largest percentage of these type of neurons at 24.7% (111/449), followed by the AMY at 17.8% (20/112), NAcc at 15% (46/306) and the IC at 7.8% (19/243). These proportions were different among brain regions (χ2(3)= 33, p <0.0001, n=196; Fig. 7C.
The C+ or C− distribution of cue-selective neurons was found by pooling the results from all four-brain regions. Under these conditions, we found a slight, but significant, increase in the proportion of C− (116) relative to C+ (80) neurons (χ2(1)= 6.25, p <0.012, null probability 0.5). However, the proportion of C+ and C− neurons was not significantly different across individual brain regions. In the OFC, 66 cue-selective neurons fired more for C−, whereas 45 preferred the C+ cue (χ2(1)= 3.6, p <0.057). For NAcc neurons, 29 fired more for C− and 17 for C+ (χ2(1)= 2.63, p <0.105) whereas for AMY neurons, 14 had firing greater for C− and 6 for C+ (χ2(1)= 2.45, p <0.1175). Finally, the IC had 7 firing more for C− and 12 for C+ (χ2(1)= 0.84, p <0.3588). Therefore, the four brain regions tested were responsive to both appetitive and aversive cues, with a slight bias for aversive cues.
As mentioned in the behavioral results, to solve this Go/No-Go task, rats needed to determine which taste cue predicted the reward (C+) and which predicted the aversive outcome (C−). Although it is possible that rats can learn the meaning of both cues independently, our data suggest that both learning processes (C+ and C−) can influence each other. In this regard, we found neuronal activity that suggests a possible interaction between cues, as shown in the example in Fig. S6E, and even in the absence of any apparent behavioral change for C+ trials, this neuron began developing cue selectivity for C+ trials, a few trials after the first correct rejection of quinine (see Figure S6E). We suggest that this type of neuronal activity not only reflects an interaction between C− and C+ cues, but also that in this task rats probably first had to learn that the C− cue predicted quinine before changes in neuronal C+ cue selectivity were observed.
To quantify the influence of learning on this AC neuronal subpopulation (n = 196), we used the activity of individual neurons to classify C+ and C− trials (Methods –Data analysis OLVQ). The temporal evolution of the percentage of trials that were correctly classified in each brain area is shown in Figure 7D. In the Pre-learning phase, only OFC neurons were found to discriminate the cues above chance level (RM ANOVA; F(110,30)=4.2, p<0.0001). However, in the Post-learning phase, a significant increase in correct cue classification was found in all four brain areas (RM ANOVA; for OFC, F(1,220)=24.27, p<0.0001; AMY, F(1,38)=6.74, p=0.013; NAcc, F(1,90)=24.2, p<0.0001; IC, F(1,36)=14.7, p<0.0005). Therefore, throughout these components of the taste reward circuit, the ability of neurons to discriminate between cues improved with learning.
We next asked whether these 196 cue-selective AC neurons could encode the acquired cue's predictive value of the outcomes (reward or punishment) and/or whether they encoded the instrumental response to be performed (Go/No-Go). To answer these questions, we performed a receiver operating characteristic (ROC) analysis to measure the subjects' choice probability (Britten et al., 1996). In this analysis, a significant choice probability suggests that the neuronal activity provides information for predicting the subsequent behavioral response that subjects will perform (Go or No-Go response). We found that 19% (22/111; 7 preferred C+ and 15 C−) OFC, 20% (4/20; 4 preferred C−) AMY, 16% (3/19; 1 preferred C+ and 2 C−) IC and 9% (4/46; 2 preferred C+ and 2 C−) NAcc neurons also encoded the rat's anticipated motor decision. Thus, a relatively small population of these neurons also contained information regarding the anticipated instrumental choice of the subjects.
This result was verified when we only analyzed the subpopulation of coherent, cue-selective AC neurons. In total, we found that, in the Post-learning phase in the delay epoch, 46 neurons (26-OFC; 8-AMY; 10-IC; 2-NAcc) were both coherent and cue-selective. In the OFC only 30% (8 /26 neurons) of AC cue-selective coherent cells also encoded the rat's anticipated motor decision (Go/No-Go). In the AMY 12.5% (1/ 8) exhibited these characteristics, whereas in the IC (0 /10) and NAcc (0 / 2) none of these cells showed a significant choice probability. This suggests that a relatively small population of licking-coherent AC cue-selective neurons also contains information regarding the anticipated instrumental choice of the subjects. Thus, the large majority of AC cue-selective coherent cells encoded the predictive value of the cues independently of the anticipated motor decision.
Similar results were found for 3 out of 28 cue-selective UC neurons (3/28 = 10.7 %) which displayed a significant “choice probability” (3 neurons from OFC and none in NAcc, IC and AMY).
We also determined whether the 196 cue-selective AC neurons could utilize the animal's licking pattern as a reference signal and/or whether licking-induced synchronous firing could improve their ability to discriminate among the cues. In the Post-learning phase, we first analyzed whether each of these neurons showed a significant coherence during the Delay epoch (Fig. 8). We found that in total, 23% (46 /196) of the neurons fired coherently with licking, with the IC having the largest proportion (52%, 10/19), followed by the AMY (40%, 8/20), OFC (23%, 26/111) and NAcc (4%, 2/46). A similar number of these neurons responded preferentially for C+ (n = 25) and C− (n = 21).
Interestingly, we found that licking coherent neurons did not use the entire lick cycle as a reference (Fig. S13). As found in the olfactory system during respiratory cycles (Bathellier et al., 2008), different parts of the licking cycle are not similarly informative. In this regard, we found that the subpopulation (n=46) of coherent AC-cue selective neurons do not use the entire lick cycle as an internal reference.
We also found that coincident spiking between pairs of licking coherent cue-selective AC cells was enhanced by learning. This conclusion emerged from the analysis of 46 coherent AC cue-selective cells. For this analysis, we measured the cross-correlations in the neuronal pairs simultaneously recorded from different brain regions during the Delay period. A total of 13 pairs were analyzed: 7 OFC-AMY pairs, 2 OFC-IC and 4 AMY-IC. After learning, the cross-correlation peaks of the 13 pairs of such neurons increased. Thus, while the mean peak values in the pre-learning was 13.1 coincidences/sec ± 2.6, after learning it increased to 25.6 coincidences/sec ±5 (Paired t-test(12) =−3.4, p-value 0.0046;n=13). Likewise, we found that after learning the cross-correlation peaks between cue-selective and non-cue-selective coherent neuron pairs also increased (see Figs. 5A, B for examples of individual neuron pairs; the mean cross-correlation peak of 61 neuronal pairs was 12.5±1.2 in the Pre-learning; it increased to 28.7±2.5 coincidences/sec after learning; Paired t-test (60) =−7.4, p-value < 0.0001). These results suggest that a subpopulation of cue-selective AC neurons synchronizes their activity with licking, and may transfer information about the acquired predictive value of the cues to other licking coherent neurons located in different brain structures of the taste-reward circuit (Gregoriou et al., 2009).
Using AC cue-selective neurons, we tested the hypothesis that neurons that exhibited synchrony with licking contained more information about the cues than those that do not (Fig. 9A). These analyses revealed that, during the Delay epoch, coherent neuronal populations were significantly better at identifying the cue than non-coherent neurons, but only in the Post-learning phase (Pre-learning; RM ANOVA, F(1,194)=0.03, p=0.86 n.s.); Post-learning; RM ANOVA, F(1,194)=18.1, p<0.0001). Nevertheless it is important to note that after learning, both coherent and non-coherent neurons improved their cue classification (coherent neurons; RM ANOVA, F(1,90)=21.2, p <0.0001; non-coherent neurons, F(1,298)=41.8, p < 0.0001). We then analyzed each brain region independently and found that coherent neurons displayed greater cue discrimination than non-coherent neurons in all four brain regions (Fig. 9B). Although, the cue-classification of licking-coherent neurons was similar across different brain regions (RM ANOVA; F(3,1302)=1.5, p=0.2; see four blue lines in Figure 9B), not all brain regions showed the same enhancement. For example, during almost all the Delay epoch (from 0.45 to 1.0 s), the OFC was the only region exhibiting a consistently higher performance for coherent over non-coherent cells. The other three regions showed enhanced performance during a limited window of the Delay epoch. Since only two NAcc neurons were coherent and cue-selective, no conclusions were drawn for this particular area. Using the OLVQ as a classifier, we found that the level of coherence during the Delay epoch correlated significantly with cue-discrimination. That is, in the Post-learning phase, the larger the coherence, the greater the cue discrimination ability of a neuron (Fig. 9C and Fig. S14). A good correlation between the level of coherence and cue-discrimination was found in the OFC (Pearson-Pairwise correlation; r = 0.44, p-value = 0.02 -when one outlier was removed). Larger correlations were found in the AMY (r = 0.74, p = 0.03) and IC (r = 0.79, p = 0.006). Overall, these results indicate that both coherent and non-coherent neurons contribute to the decoding of the cue identity, but neurons that fired in synchrony with licking clearly performed significantly better in this task.
To better distinguish among sensory cues, we explored whether licking coherent cells might rely on additional information contained in either their firing rates or their spike timing. To identify which of these mechanisms better explained our findings, we first determined the amount of discrimination obtained by using only mean firing rates. This analysis revealed no significant difference between licking-coherent and non-coherent cue-selective AC cells of the IC, AMY and the OFC (Fig. S15), indicating that firing rates alone did not convey the extra cue-information of licking-coherent cue-selective cells.
To test the hypothesis that additional cue-information derived from the neuron's spike timing, we shuffled the spikes of all 46 licking coherent cue AC selective cells. This procedure removed all potential cue-information from the neuron's spike timing, without compromising information contained in the cell's average firing rate. When the spike timing information was eliminated from these neuronal responses, their ability to discriminate among the cues became nearly identical to that of non-coherent neurons (Fig. 10). This demonstrated that in the Delay epoch spike timing likely conveyed the extra cue-information of cue-selective and licking coherent AC neurons in comparison with non-coherent neurons.
Therefore, it follows that the spike timing of coherent cue selective AC neurons should be more precise than of non-coherent cue-selective cells. Indeed, we found this to be the case. In general, one can measure how precisely and reliably a neuron will fire during a lick cycle (spike-timing precision) by calculating the standard deviation (s.d.) of the neuron's firing phase: the smaller the s.d. of the firing phase (for a given phase of the lick cycle), the greater the spike-timing precision (Fig. 11A). By plotting the s.d. of the neuronal firing phase for coherent and non-coherent cue-selective cells at each lick cycle (Fig. 11B), we observed that coherent cells tended to fire with a significantly smaller s.d. phase than non-coherent cells (Wilcoxon Test, p-value < 0.05). Therefore, licking-induced synchronization likely improves taste discrimination because AC cue-selective cells, in addition to being able to discriminate cues by modulating their firing rates, also fired with higher spike-timing precision. In fact, in contrast to non-coherent cells, licking coherent cells significantly improved the precision of their spike timing during the Delay epoch (smaller s.d. firing phase) in the Post-learning phase (Wilcoxon Test p-value < 0.001; Fig. 11C). This suggests that one of the major effects of learning the task was the reduction of spike timing variability, which consequently allowed cue selective licking-coherent neurons to improve cue-discrimination.
In this paper we addressed how a taste stimulus gains control of behavior through associative learning and how active sampling of gustatory cues, through rhythmic licking, influences neural activity across multiple brain regions. We observed that as rats learned to perform a Go/No-Go taste discrimination task, neuronal responses to initially non-predictive taste-cues became more distinct and predictable in different brain areas, showing that learning induces a significant functional reorganization of neural activity throughout major components of the taste-reward circuit. We also observed that neurons that fired in synchrony with licking exhibited greater cue-discrimination than non-synchronized neurons and that this effect increased with learning. We attribute this to an enhancement in spike-timing precision of licking-coherent neurons.
We found that rats can learn to discriminate and identify two hedonically positive tastants (the cues), as well as a negatively hedonic tastant (quinine), using only a single lick comprising about 150 ms (Halpern and Tapper, 1971) (Figs. 1B–C, also see Results: Behavior). In contrast to this rapid response, tastants delivered via intraoral cannulae (IOC) require a much longer time to be detected (Katz et al., 2001). One possible reason is that actively sampling tastants, by freely licking, may induce an “active brain state” (Poulet and Petersen, 2008) that integrates volition and expectation (Gutierrez et al., 2006), whereas delivering tastants via IOCs, produces a passive stimulation. In agreement with our results, previous findings in the somatosensory (Krupa et al., 2004), auditory (Eliades and Wang, 2008), olfactory (Fuentes et al., 2008) and also in the gustatory system (Wilkins and Bernstein, 2006) have clearly shown that sensory responses vary significantly according to whether animals obtain a stimulus passively or actively.
One aspect of the subjects' behavior in the cue compartment is that, after receiving the cue, rats continued to lick the empty sipper for an average of 1.2 s (or 5–7 additional dry licks). This suggests that before emitting a behavioral response, the subjects used this self-imposed time to continue accumulating information. In this regard, we found that the Delay epoch contains information about the acquired predictive value of cues (Fig. 7), which may explain why rats favor accuracy over speed in a Go/No-Go discrimination task (Abraham et al., 2004; Friedrich, 2006).
Based on the fact that licking produces a quasi-periodic input into the CNS (4–9 Hz)(Halpern, 1983), we hypothesized that, like whisking or sniffing (Nicolelis et al., 1995; Kepecs et al., 2006; Kleinfeld et al., 2006), licking may act to fragment the flow of sensory information into discrete chunks during each rhythmic cycle. In the present study, we found neurons with activity that was phase-locked with respect to the licking rhythm (Figs. 2–4). This is in agreement with previous results where licking-related activity was reported in the IC, AMY and OFC (Yamamoto et al., 1988; Nishijo et al., 1998; Gutierrez et al., 2006). Some of these neurons may encode oromotor information, as they fire in synchrony with licking, independently of whether water or a cue was delivered (Fig. 4A). However, for the majority of neurons, phase-locked activity did not arise from pure oromotor input since the neuronal activity was commonly found to be in synchrony with licking only during specific epochs of the trial. Moreover, the proportion of neurons with licking coherent activity increased after learning. These results showed that licking behavior has a more dynamic function than previously expected. Under this new scheme, neurons in the taste-reward circuit may use a lick cycle as an internal temporal frame, and thus licking may serve to coordinate (“phase-lock”) the activity of neurons distributed across multiple brain structures.
During the task presented here, the animal is licking the entire time, however particular ensembles of neurons, distributed over the four recorded areas, track various states of the task (Fig. 2). It follows that the distributed neural activity generates a continuous flow of information that reflects all sensorimotor transformations required to solve the task, from cue expectation, to cue detection and identification, to just prior to the initiation and execution of a Go/No-Go response. In summary, together with the anticipatory activity (BC) and phasic neurons (UC), AC neurons can completely characterize each salient epoch of the trial. We propose that, in a well-rehearsed behavioral sequence, where rats lick continuously, neuronal ensembles throughout the taste–reward circuit can represent the state-to-state transitions of the behavioral task (Sutton and Barto, 2000; Jones et al., 2007; Belova et al., 2008).
The presence of neurons with cue-expectant activity (BC) in the OFC indicates that this area has a role in monitoring relevant task events (Feierstein et al., 2006). UC neurons displayed both rapid and transient (<150 ms) responses. Although most of them were uninformative about the identity of the cue, they may be involved in cue detection. Moreover, the responses of the non-cue selective UC neurons were modulated by learning. The functional reorganization of phasic responses observed during learning may reflect different levels of attentional processing to the cues (Ifuku et al., 2003). Taken together, these results indicate that phasic responses in the taste–reward network are dynamically reorganized by learning.
We found that cue-selectivity could arise even after brief taste stimuli presentation and continue for about 1 second (in the subsequent empty licks). Therefore, our results further extend previous studies (Schoenbaum et al., 1998; Setlow et al., 2003) by showing that the predictive reward value of the cues can be rapidly encoded in one single lick cycle. That is, in less than 150 ms, the firing of a subpopulation of phasic neurons contains sufficient information to permit subjects to obtain a reward and avoid punishment.
With respect to the neural changes that occurred with learning, the cue-selective AC responses were the most intriguing. This is because after the animal had obtained the tastants (cues), their activity contained information that may be used to discriminate among the cues. Yet, animals continued to lick without getting any fluid before making the decision of whether or not to go. The OFC and the AMY contained more neurons with cue-selectivity in the Delay epoch followed by NAcc. In contrast, the IC, which contains the primary gustatory cortex, had the smallest proportion of cue-selective AC neurons, perhaps as a consequence of its role in gustatory processing (Ifuku et al., 2003). Thus, despite the small proportion, IC neurons also represented information about the predictive value of the cues (Fig. 7D) (Stapleton et al., 2007; Lara et al., 2009). Therefore, the predictive reward value of gustatory cues appears to be widely distributed in the OFC-AMY-NAcc-IC circuit. However, different degrees of specialization occur in this network.
We identified two types of cue-selective AC neurons: one that synchronizes with some phase of the licking cycle and a second that does not (Fig. 8). In general, cue-selective AC neurons that synchronize their activity with licking were significantly better at decoding the cue identity than non-coherent neurons (Fig. 9A). Subsequent analysis indicated that coherent cells fire with higher spike timing precision than non-coherent neurons (Fig. 11B) and it is this greater degree of spike timing that conveys the extra cue-information of licking coherent and cue-selective AC neurons (Fig. 10). In fact, after learning, the spike-timing precision of coherent, cue selective AC-cells increased (Fig. 11C), suggesting that learning reduces spike timing variability, allowing licking-induced oscillations to enhance cue-discrimination. In this regard, it has been shown that sniffing-induced neuronal oscillations (at theta rhythm) enhance stimulus discrimination by ensuring action potential precision in the olfactory system (Schaefer et al., 2006). Therefore, rhythmic behaviors associated with the two chemical senses seem to share several neuronal coding mechanisms.
The rat orbitofrontal cortex is composed exclusively of agranular cortical areas, similar to the caudal regions of monkey OFC (Ongur and Price, 2000). However, rats lack a clear homolog of the medial and rostral regions of monkeys OFC, which contain a dysgranular and granular layer IV, respectively. Based on this cytoarchitectonic evidence, it has been suggested recently that the rat OFC may be a model to study the function of the agranular most caudal region of monkey OFC (Wise, 2008), which receives inputs from olfactory and gustatory cortices (Cavada et al., 2000). In this regard, in rats, we found that licking rhythmically induced phase-locked activity across several brain regions, including the OFC. At present, it is not clear whether licking induces neuronal synchrony in non-human primates. However, we note that monkeys masticate rhythmically (<3Hz), and they have cortical neurons, whose firing is phase-locked to mastication and to chewing rhythms (Yao et al., 2002).
In summary, we characterized the neurophysiological correlates of a large component of the taste-reward circuit during learning of a taste-guided behavioral task. Our results suggest that to compute the acquired predictive value throughout this circuit, appetitive and aversive associative learning appears to employ synchronous activity distributed among neuronal ensembles. As such, we propose that to understand the meaning of gustatory stimuli and to make behavioral decisions based upon them, like sniffing and whisking, licking constitutes another type of active exploratory rhythmic behavior that rats employ to synchronize neuronal activity across multiple brain structures.
The authors wish to thank Jim Meloy for invaluable technical support and Susan Halkiotis and Eric Thomson for their comments on the manuscript. The project described was supported by Award Number R01DC001065 from the National Institutes on Deafness and Other Communication Disorders (NIDCD). The content is solely the responsibility of the authors and does not necessarily represent the official views of the NIDCD or NIH.