Introduction The coordinated, rhythmic activity of neuronal populations gives rise to oscillations in local field potentials (LFP) and electroencephalograms at a broad range of frequencies . Throughout the brain, these oscillations potentially constitute clocking mechanisms against which to reference and coordinate the timing of neural firing. Synchronization of these rhythmic activities is likely to reflect or underlie functional interactions between neurons within a defined brain structure, or between disparate populations in distinct structures [2,3]. Equally, abnormal synchronization may impair functional interactions and contribute to complex cognitive disorders such as schizophrenia [4,5] and attention deficit/hyperactivity disorder [6–8]. At the level of single neurons, synchrony is evident in the consistent temporal relationships between the firing patterns of interconnected cells. These are most commonly quantified using cross-correlation techniques [9–11]. For example, the correlated firing of cortical neurons is implicated in visual processing , attention , and motor learning . However, the majority of these studies are based on recordings from single brain regions, and are unable to address the nature of correlated activities in networks spanning multiple structures. Do correlations also underlie coordination between anatomically and functionally related brain regions? If so, are they also mediated or reflected by oscillatory population activities at the LFP level? Theta rhythms are 4- to 12-Hz oscillations consistently associated with complex behaviors presumed to require mnemonic processing and/or decision-making, for example spatial exploration in rodents , working memory in primates , and navigation and working memory in humans [17,18]. Dynamic, behavioral modulation of theta rhythms may therefore indicate or mediate cross-neuronal and/or cross-structural interactions during these behaviors. Theta rhythms are found in many mammalian brain structures, but are most prominent in the rodent hippocampus . Here, the firing of individual “place cells”—hippocampal principal excitatory neurons with spatial receptive fields —is coordinated (“phase-locked”) with respect to the local theta rhythm. Thus the action potentials of a given neuron tend to occur during a preferred phase of the theta cycle. This phase-locking of hippocampal spike-timing to ongoing LFP oscillations is an important example of temporal coding in the brain  and—in concert with the related phenomenon of phase precession—has been proposed to allow higher-order coding of spatial information than that imparted by the firing-rate pattern alone . It has also been proposed that hippocampal theta rhythms may coordinate neural activity during sensorimotor integration  or information encoding [24,25]. Neuronal firing phase-locked to the hippocampal theta rhythm has also been described in cingulate cortex , amygdala , entorhinal cortex , striatum , and, most recently, the rat prefrontal cortex [30,31]. As in the hippocampus, phase-locking in the prefrontal cortex is also accompanied by phase precession . This raises the possibility that phase-locking may play a broader role in defining the temporal relationships between cross-structural activities. However, it remains to be established how these phase relationships influence the firing of connected neurons, and how—or whether—they relate to behavior or hippocampal function. For example, while Siapas et al.  included radial arm maze and T-maze tasks among the conditions during which they collected data, analyses made no attempt to relate medial prefrontal cortex (mPFC) phase-locking to ongoing behaviors. If mPFC phase-locking remains constant while behavioral demands vary, it is unlikely to reflect functional interactions between coactive structures. Similarly, Hyman et al.  recorded during running on a linear track and random foraging, and were therefore unable to explicitly relate variations in phase-locking with variations in behavioral demands. Siapas et al.  suggested that mPFC phase-locking may play some role in the formation of long-term memories that require that transfer of information from the hippocampus to the neocortex. Consistent with this role, monosynaptic projections from hippocampus to the deep layers of mPFC [33,34] do exhibit activity-induced plasticity , and LFP oscillations in the two structures are correlated during slow-wave sleep . However, lesions or disruptions of either mPFC or the hippocampus impair spatial working memory [37–39], providing functional evidence that the two structures also interact on-line during behaviors that require short-term mnemonic processing . To investigate the nature of functional interactions that relate to decision-making, we made simultaneous tetrode recordings of extracellular action potentials and LFP from CA1 and mPFC to examine coordination of activity in these two structures during a spatial working-memory task designed to dissociate neural activity related to performance (e.g., running and orienting behaviors) from activity related to mnemonic or decision-making processes. We find that spike timing and theta-rhythmic activities in CA1 and mPFC become more coordinated during epochs of the task associated with peak mnemonic and decision-making load. Results Behavior Data are presented from eight recording sessions from six rats. Rats ran 15–25 trials per 20- to 30-min recording session. Each trial of the task comprised a “forced-turn” (sample) and “choice” (test) epoch (Figure 1A), and was subdivided into a number of stages corresponding to different sections of the maze (Figure 1B). Rats were trained to asymptotic performance prior to electrode implantation (Figure 1C), and performed the task at 83% ± 5.0% (mean ± standard error of the mean) correct during the eight sessions presented here. Rats began at a reward point (F1 or F2), then ran towards the central arm of the maze (stage 1 in Figure 1B). A moveable barrier directed them down the central arm towards the choice point (stage 2), when they were required to choose a left turn if the trial started from F1, or a right turn if the trial started from F2. If the correct turn was made (stage 3), chocolate reward was delivered remotely to C1 or C2; incorrect turns were not reinforced. Rats then returned to the choice point (stage 4) and were directed by another moveable barrier back down the central arm towards the forced-turn end of the maze (stage 5). Here, the barrier directed them to reward at F1 or F2 (stage 6), with left or right turns selected at random from trial to trial. The location of the forced-turn end of the maze was varied between animals. Stages 1 and 2 (choice epoch) presumably invoked spatial working-memory processes: rats were required to “hold in mind” the location of the starting reward—or the direction of the turn between stages 1 and 2—in order to choose between C1 and C2 at the opposite end of the maze. By stage 3, their decision had already been made, and any working-memory requirement was negated. Stages 4–6 (forced-turn epoch) never required active working memory, since routes back to F1 or F2 were always predetermined by the barriers. A primary concern during subsequent analyses was to dissociate overt behavioral differences (such as running speed) from differences in neural activity. For example, stages 1 and 6 both occurred on the same sections of the maze and yet made contrasting demands upon spatial working memory. However, rats ran more than twice as fast during stage 6 (44 ± 1.0 cm/s) than they did during stage 1 (18 ± 5.8 cm/s), a behavioral difference that may confound interpretation of differences between neural activities during the two stages. Similarly, rats ran at 17 ± 2.1 cm/s during stage 3 and at 31 ± 1.8 cm/s during stage 4. Most analyses therefore focused on the central three-quarters section of the central arm, where mean running speeds were similar during both choice and forced-turn epochs (38 ± 1.9 and 40 ± 2.0 cm/s, respectively). Restricting analyses to the central arm also excluded sections of divergent running trajectories at the very ends of the arm (Figure S1). As rats crossed the central arm during forced-turn epochs, their route was predetermined by the moveable barrier. In contrast, during choice-direction epochs, rats were required to choose between left or right reward arms, employing spatial working memory to guide their decision. Since overt behavior was similar during both epochs, differences between neural activity during runs across the central arm in the forced-turn and choice directions therefore reflect mnemonic and/or decision-making processes, rather than the simple behavioral demands of the task. Behavioral Correlates of CA1 and mPFC Firing One hundred and sixty-five mPFC and 149 CA1 neurons (ten neurons from ventral CA1) active on the maze were classified as putative pyramidal neurons or fast-spiking interneurons (4% of recorded mPFC neurons, 5% of CA1 neurons) on the basis of spike width, firing rate, and burst-firing characteristics. The classification was based on similar schemes derived from intracellular recordings [41,42] (Table 1), although only pyramidal neurons were used for subsequent analyses. The data of Siapas et al.  and Hyman et al.  set a precedent for examining the coordination of dorsal CA1 and mPFC activities. However, the most prominent hippocampal projections to mPFC arise from ventral CA1/subiculum ; initial experiments were therefore designed to compare the properties and interactions of dorsal CA1 neurons with ventral CA1/subiculum neurons. The low yield of well-isolated units from ventral regions precluded a systematic comparison between dorsal and ventral subregions in this study, although the basic properties of ventral CA1 pyramidal cells were comparable with those of dorsal neurons (see Table 1). The behavioral correlates of mPFC neuronal activity tended to be more spatially distributed than for CA1 neurons (Figure 2), with no population bias towards one maze region (mean mPFC firing rate 4.7 ± 0.4 Hz on reward arms [including reward points] and 5.2 ± 0.5 Hz on the central arm). Neither was mPFC population activity biased towards one task epoch. For example, overall central-arm firing rates were comparable in forced-turn and choice directions (5.3 ± 0.5 Hz and 5.2 ± 0.5 Hz, respectively). Despite these similar population firing rates during different task epochs, the central-arm firing rates of individual mPFC neurons did tend to distinguish between runs in the two directions and between different routes in the choice direction (Figure 3). A “directional index” for each neuron was defined as the magnitude of the difference between mean firing rates during forced-turn and choice directions, divided by the overall mean firing rate on the central arm. “Preference index” was defined as the magnitude of the difference in mean firing rates during F1 → C1 and F2 → C2 trials divided by the mean choice-direction firing rate. Thus both indices ranged from zero (firing rate identical in both epochs/routes) to one (fired only in one epoch). The mean directional index of mPFC neurons was 0.27 ± 0.02, and the mean preference index was 0.34 ± 0.04. This biased activity was also evident in CA1 (mean directional index 0.60 ± 0.05, preference index 0.62 ± 0.16) and was reminiscent of the CA1 activity previously described on similar linear tracks [44,45]. Slight differences in running trajectory or head direction may contribute to these firing-rate effects. However, these behavioral parameters did not vary consistently with trial type (Figure S1) and are unlikely to explain the firing-rate biases in their entirety. mPFC firing tended to be more sustained than that of CA1, with the average mPFC neuron firing at more than 10% of its maximum rate across 52% ± 3% of the central-arm area, whereas average CA1 place-cell firing covered only 22% ± 2% of the central arm. Similar “delay firing” activity in rat mPFC has been reported by Baeg et al. . However, these firing-rate data do not address the nature of interactions between CA1 and mPFC. Behavioral Modulation of Cross-Correlations between CA1 and mPFC Spike Times We first investigated coordination of hippocampal and prefrontal activities by quantifying the temporal alignment of neuronal firing using cross-correlation of spike times from CA1–mPFC unit pairs coactive during different behavioral epochs. Peak cross-correlation coefficients (bin size 100 ms, maximum lag ± 200 ms) were normalized by spike counts, and bias-corrected by subtracting values obtained when trials were shuffled with respect to one another. Spike trains from 50 CA1–mPFC pairs coactive during stages 1 and 6 (see Figure 1B) showed significantly higher cross-correlation coefficients during stage 1 than during stage 6 (0.029 ± 0.002 versus 0.017 ± 0.002; p 180°) phase of theta will have mutually increased firing probabilities during repeated ∼50-ms windows. The net effect will be increased correlated activity amongst these groups of neurons. This proved to be the case during choice epochs, when improved cross-correlations between CA1–mPFC unit pairs paralleled enhanced phase-locking and coherence. Similarly, the value of each neuron's mean preferred phase during a given behavioral epoch will consistently dictate the order in which that neuron fires relative to neurons with different mean preferred phases. Theta rhythms can therefore mediate the consistent timing relationships needed to establish functional connectivity between two structures, in this case serving to dynamically incorporate currently relevant spatial information into decision-making processes. How is CA1–mPFC synchrony enhanced during specific behavioral epochs? One possibility is simply through simultaneously increased theta-modulated activity in these two structures. However, the enhanced correlations and phase-locking were independent of changes in theta power, since overall theta power was similar during forced-turn and choice epochs. In addition, the coherence measures are normalized by the power spectra of CA1 and mPFC LFPs and are consequently insensitive to power changes. Together, these data therefore imply selective alignment of CA1 and mPFC theta-rhythmic activity during choice-direction runs across the central arm of the maze. Since the instantaneous frequency of theta rhythms can vary rapidly over time, this alignment requires active synchronization of CA1 and mPFC rhythms in order to maintain their coordination with respect to each other. Similarly, maintenance of enhanced phase-locking of mPFC neurons to hippocampal theta cannot arise simply as a consequence of firing rates and LFP being modulated similar frequencies, but rather necessitates precise temporal control of spiking relative to theta rhythms . This level of control may be exerted by monosynaptic projections from the hippocampus, which have direct influence on mPFC interneurons . Since the synchronization of oscillatory networks is often attributed to the activity of local inhibitory interneuronal networks , these projections may provide the anatomical and physiological foundations for coherence between these two structures. Interestingly, the most dorsal part of CA1 does not contribute direct projections to mPFC , raising the possibility that dorsal CA1 interacts with mPFC via ventral CA1/subiculum. Whilst we did not observe qualitative differences between ventral and dorsal CA1 neurons in terms of their correlations with coactive mPFC neurons or their phase-locking properties, further experiments are required to establish the functional consequences of the known anatomical connections in this system. There is no evidence for direct reciprocal projections from mPFC back to the hippocampal formation , making it tempting to speculate that the hippocampus drives mPFC firing rather than vice versa. Although our data do not unequivocally address the directionality of hippocampal–prefrontal interactions, this is supported by Siapas et al. , who suggest that mPFC neurons phase-lock to CA1 theta that occurs ∼50 ms in advance of their spikes; this also proved to be the case during the working-memory task employed here. However, whether behavioral-dependent enhancement of these coordinated activities is achieved entirely through hippocampal–prefrontal connectivity or via some third party that influences both CA1 and mPFC remains to be established. Whilst enhanced theta-frequency coordination coincided with peak working-memory load, it is possible that theta-frequency interactions between CA1 and mPFC do not pertain solely to working memory or decision-making. For example, attention or reward expectancy may also vary between choice and forced-turn behavioral epochs. It should also be noted that some degree of phase-locking and coherence remained evident during runs across the central arm in the forced-turn direction. These residual interactions may reflect some working-memory-related aspect of spatial behavior common to both task epochs, such as updating route or task-rule information. CA1–mPFC synchrony—neuron-pair correlations, phase-locking, and LFP coherence—also fell to these control levels during runs towards the choice point on error trials. We cannot determine the stage of the task at which the errors originated, or whether they were due to failures in mnemonic, attentional, or decision-making components. However, the fact that the degree of CA1–mPFC synchrony can be used to predict behavioral outcome strongly suggests that these electrophysiological phenomena are indeed signatures of cross-structural interactions. In summary, our data reveal correlations between behavioral demands and cross-structural neural synchrony: theta-frequency coordination between CA1 and mPFC peaks during behavioral epochs presumed to require effective communication between these two structures. It follows that disruption of such complex cross-structural communication is likely to generate behavioral impairments. For example, schizophrenia is associated with altered GABAergic function in hippocampal and prefrontal interneurons , and is widely presumed to involve disrupted functional connectivity of the prefrontal cortex [4,58,59]. Interestingly, schizophrenic patients do show spatial working-memory impairments . The theta-rhythm-mediated coordination of hippocampal–prefrontal activity that we describe here may reflect the nature of cross-structural coordination at network and neuronal levels, and may contribute to both the clinical diagnosis of the impaired interactions likely to underlie cognitive disorders and to characterizing animal models of these diseases. Materials and Methods All procedures were performed in accordance with the Massachusetts Institute of Technology Committee on Animal Care and the National Institutes of Health guidelines. Six male Long-Evans rats (2–6 mo) were mildly food-deprived (to 85% of free-feeding body weight) and trained to run a continuous spatial-alternation task (see Figure 1A). Each trial comprised distinct sample and test epochs. The contingency was set such that, for example, a rat forced to turn to his right during the forced-turn epoch had to choose a left-hand turn to win reward during the subsequent choice epoch. Forced-turn direction was varied randomly, with no more than three consecutive trials in one direction. The relative location of the forced-turn end of the maze was varied between animals. Every effort was made to constrain running to overlapping linear trajectories by using a narrow track (6 cm). Furthermore, analysis epochs excluded reward points and turning points. Each rat was trained to asymptotic performance (two consecutive days of at least 80% choice-correct) over a period of 12–14 d before surgery, then implanted with arrays of adjustable tetrode recording electrodes targeted to the mPFC (+3.2 mm, +0.6 mm from bregma) and ipsilateral dorsal CA1 (−3.6 mm, +2.2 mm). In two rats, tetrodes were also targeted to ventral CA1 (−6.3 mm, +6.2 mm). Differential recordings of extracellular action potentials (sampled at 31.25 kHz per channel, filtered between 600 Hz and 6 kHz) and continuous LFP (sampled at 3.125 kHz per channel, filtered between 1 and 475 Hz) were made using Keithley Instruments acquisition boards (DAS-1802HC [http://www.keithley.com]). Local reference electrodes were placed in overlying (for dorsal CA1 recordings) or adjacent (for ventral CA1 recordings) white matter, or in a proximal cortical region without spiking activity (2.4–2.7 mm below the pial surface for mPFC recordings). Positioning in white matter was achieved on the basis of characteristically flat LFP recordings (with no hippocampal sharp waves or ripples) and an absence of action-potential activity in this region. Only hippocampal LFP data taken from dorsal CA1 are presented here. Electrolytic lesions established tetrode tip positions at the end of each experiment (Figure 1A). Data presented here are taken from eight recording sessions from the six rats. Action potentials were assigned to individual neurons by off-line, manual clustering using Xclust software (M. A. Wilson). Subsequent analyses employed a combination of in-house software (M. A. Wilson) and custom Matlab code (MathWorks, Natick, Massachusetts, United States). Firing with inter-spike intervals of between 2 and 15 ms was defined as bursting (minimum inter-burst interval 150 ms). The Complex Spike Index (Table 1) combined a measure of bursting with a measure of the likelihood that spikes later in bursts were smaller in amplitude than spikes earlier in bursts . Spatial information was calculated according to Skaggs et al. . Most analyses compared firing on the central three-quarters section of the central arm in the two running directions, and were restricted to neurons that fired at least 50 spikes in both choice and forced-turn epochs (in order to allow reliable circular statistics). LFPs were down-sampled (to 600 Hz) and band pass filtered between 4 and 12 Hz, then maxima and minima detected and thresholds established to extract theta peak and trough times. Only peaks or troughs greater than one standard deviation from the mean amplitude of the filtered LFP were included (63% ± 6.6% of all maxima during choice epochs, 61% ± 9.5% during forced-turn epochs). Each spike was assigned a theta phase between 0 and 360° by linear interpolation of the spike time relative to the enveloping pair of peak (phase 180°) and trough (phase 0 or 360°) times. The Rayleigh test of uniformity was used to assess the resulting phase distributions for deviations from the circular uniform distribution. Circular statistics were calculated according to Fisher . Multi-taper spectral analysis  was used to calculate power spectra and coherence for LFP data. This technique takes advantage of short-time-window Fourier analysis to reduce artifacts caused by non-stationary elements in the data (since data can be assumed to be stationary within the short sliding time-windows). The significance of trial-by-trial magnitude of the coherence during central-arm crossings was calculated according to Jarvis and Mitra (with coherence values greater than 2/√[(number of trials) × (number of tapers)] considered significant at p = 0.05) . Data in the text and figures are given as mean ± standard error of the mean. Statistical comparisons between forced-turn, choice-correct, and choice-error conditions were performed on the two groups of animal means using Wilcoxon rank sum tests. Supporting Information Figure S1 Details of Movement Trajectories and Running Speed (A) Raw data showing positional samples (taken from LEDs mounted on the rat's head, 30-Hz sampling rate) during a single run-session. Blue box marks the central three-quarters section of the central arm used throughout analyses comparing choice and forced-turn directions. Scale bar 6 cm. (B) Mean central-arm trajectories averaged across all rats and trials for the different behavioral epochs (± 1 standard deviation marked by the width of the shaded area). Upper panel compares choice-correct runs (solid red line and dark red shading) with choice-error runs (dashed red line and lighter shading). Lower panel compares choice-correct with forced-turn runs (grey line and shading). Trajectories show considerable overlap on this section of the maze; systematic variations in trajectory are therefore unlikely to explain the enhanced coordination seen during choice-direction runs. (C) Mean running speeds across the central arm (same color scheme as in [B]). (2.8 MB PDF). Click here for additional data file. Figure S2 Raster Plots Illustrating Trial-By-Trial Firing on the Central Arm of the Maze for the Three Neurons Shown in Figure 2 A CA1 pyramidal cell is shown in (A), while (B) and (C) show mPFC pyramidal cells; the two neurons in A and B were recorded simultaneously. Spikes (shown by the tick marks on the rasters) were parsed into choice-direction (middle column) and forced-turn-direction (right column) runs. Corresponding linearized mean firing rates (Gaussian-smoothed, with a kernel width of 8 cm) are shown below each raster. (2.3 MB PDF). Click here for additional data file.