Interindividual differences in memory system LFP activity predict behavioral strategy on a dual-solution T-maze

Abstract

Individuals can use diverse behavioral strategies to navigate their environment including hippocampal-dependent place strategies reliant upon cognitive maps and striatal-dependent response strategies reliant upon egocentric body turns. The existence of multiple memory systems appears to facilitate successful navigation across a wide range of environmental and physiological conditions. The mechanisms by which these systems interact to ultimately generate a unitary behavioral response, however, remain unclear. We trained twenty male, Sprague-Dawley rats on a dual-solution T-maze while simultaneously recording local field potentials that were targeted to the dorsolateral striatum and dorsal hippocampus. Eight rats spontaneously exhibited a place strategy while the remaining twelve rats exhibited a response strategy. Interindividual differences in behavioral strategy were associated with distinct patterns of LFP activity between the dorsolateral striatum and dorsal hippocampus. Specifically, striatal-hippocampal theta activity was in-phase in response rats and out-of-phase in place rats and response rats exhibited elevated striatal-hippocampal coherence across a wide range of frequency bands. These contrasting striatal-hippocampal activity regimes were 1) present during both maze-learning and a 30min pre-maze habituation period and 2) could be used to train support vector machines (SVM) to reliably predict behavioral strategy. Distinct patterns of neuronal activity across multiple memory systems, therefore, appear to bias behavioral strategy selection and thereby contribute to interindividual differences in behavior.

Introduction

The interplay between external stimuli and internal brain state gives rise to distinct patterns of neuronal activity and consequent behavioral output (Baldassarre et al., 2012; Hesselmann et al., 2008; Monto et al., 2008). Behavioral choice can therefore arise in response to changes in external conditions and/or the brain’s internal milieu. In the dual-solution T-maze (Packard & McGaugh, 1996; Tolman, 1948) individual rats can successfully locate an appetitive food reward using one of two distinct behavioral strategies. This behavioral assay, therefore, provides a powerful approach to investigate the neuronal underpinnings of interindividual differences in behavior.

Behavioral strategy on the dual-solution T-maze can be experimentally biased by varying either external maze conditions or the internal state of test subjects (Chersi & Burgess, 2015; Ferbinteanu, 2019; Goodroe et al., 2018; Packard & Goodman, 2013; Retailleau et al., 2012). In the presence of heterogeneous spatial cues, rats and humans typically navigate this maze using a spatial map (“place strategy”). By contrast, with homogeneous spatial cues (or the absence of salient cues), individuals typically use egocentric body turns to navigate the maze (“response strategy”). Place and response strategies have primarily been associated with neuronal activity in the hippocampus and striatum, respectively. Individuals exhibit greater hippocampal activation when using a place strategy and striatal/caudate activity increases when using a response strategy (Bohbot et al., 2007; Iaria et al., 2003). Hippocampal inactivation, through either local lidocaine injection or excitotoxic lesions, impairs place strategy expression, while similar striatal inactivation produces a corresponding deficit in response strategy (Lee et al., 2008; Packard & McGaugh, 1996). Thus, distinct navigational strategies appear to 1) provide behavioral flexibility in response to external conditions and 2) arise in part from the activity of distinct memory systems.

Like the effects of spatial cue availability, a diverse set of experimental manipulations that alter neuronal signaling also affect the expressed behavioral strategy. Levels of circulating sex hormones (Korol et al., 2004; Spritzer et al., 2013; Zurkovsky et al., 2007), acute stressors (Sadowski et al., 2009; Schwabe et al., 2009), sleep deprivation (Hagewoud et al., 2010), and metabolite availability within the hippocampus and/or striatum (Canal et al., 2005; Carlson et al., 2010; Gold et al., 2013) can each alter the relative expression of place and response strategies. Similarly, interindividual differences in brain structure, including hippocampal grey matter volume (Bohbot et al., 2007) and fractional anisotropy (Iaria et al., 2008), have been shown to correlate with navigational performance. Collectively, these results demonstrate that internal brain state can dynamically alter the expression of navigational strategy. The ability to switch between place and response strategies as a function of internal brain state may provide a means by which individuals can adapt their navigational strategy to their current physiological needs (Shelton et al., 2013).

Despite the wealth of research characterizing physiological factors that produce differences in the expression of place and response strategies, it remains unclear how internal brain state affects neuronal activity within and/or between hippocampal and striatal networks to produce behavioral variability. In the present study, we trained 20 Sprague-Dawley rats on a dual-solution T-maze under constant experimental conditions in which place and response strategies were equally likely to be expressed. By simultaneously recording local field potentials (LFP) targeted to the dorsolateral striatum and dorsal hippocampus both before and during maze learning, we sought to characterize how intrinsic patterns of neuronal activity affect the expression of distinct behavioral strategies. We observed that place and response rats exhibit spontaneous differences in striatal-hippocampal coherence and phase-locking that were present during pre-maze habituation and persist throughout maze learning. Using support vector machines (SVM) trained with LFP data from either habituation or maze learning, we were able to predict the behavioral strategy of individual rats with high sensitivity and specificity. Interindividual differences in spontaneous striatal-hippocampal activity, therefore, are associated with behavioral strategy.

Methods

Surgery

3-4-month-old, male Sprague-Dawley rats (n=20, Charles River; Wilmington, MA) were housed under standard laboratory conditions (12hr light/dark cycle, access to food and water ad libitum). Prior to surgery, rats were given a subcutaneous analgesic (Meloxicam; 2mg/kg; MWI, Boise, ID) and an intramuscular antibiotic (Penicillin; 100,000 units/kg). Under isoflurane anesthesia (3.5% induction, 2-3% maintenance), stereotactic surgery was performed to implant local field potential (LFP) electrodes (Teflon-coated stainless steel wire, 0.008in diameter; A-M systems, Sequim, WA) targeted to the dorsal hippocampus (anterior-posterior (AP): −3.3mm, medial-lateral (ML): +2.4mm, dorsal-ventral (DV): −3.2mm) and dorsolateral striatum (AP: −0.9mm, ML: +3.4mm, DV: −4.5mm). LFP leads were referenced to a screw affixed to the skull above the cerebellum. An additional screw above the cerebellum served as a ground. LFPs and ground/reference screws were connected to a commercially available headmount (8201-SS, Pinnacle Technologies, Lawrence, KS). All wires, screws, and the headmount were housed within a plastic cap (to facilitate subsequent wireless recordings) and affixed in place with dental acrylic (Lang Dental; Wheeling, IL). 12 – 24 hours post-surgery, all rats received a post-operative analgesic (Meloxicam; 2mg/kg). Rats had a minimum of seven complete days to recover following surgery. These methods and those below were performed in accordance with the National Institutes of Health Guide for the Care and Use of Laboratory Animals and were approved by Middlebury College’s Institutional Animal Care and Use Committee.

Behavioral testing on a dual-solution T-maze

After recovery from surgery, acute food restriction was undertaken to increase appetitive motivation during behavioral training/testing. Rats were provided limited access to food each day with the intent of reducing body weight to 80-85% of expected free-feeding weight over the course of 7-10 days. During this food restriction period, each rat was handled daily for 5-10 minutes and given pieces of Fruit Loops (the appetitive reward used during behavioral testing; Kellogg’s, Battle Creek, MI).

Upon reaching <85% of expected free-feeding weight, rats were trained and tested on a dual solution T-maze while LFP activity was wirelessly recorded from the dorsal hippocampus and dorsal striatum. Rats were briefly anesthetized (~5-10min) with isoflurane to connect the implanted headmount to a wireless amplifier (100x amplification, 0.5Hz high pass filter)/Bluetooth transmitter (8200-K6-SL, Pinnacle Technologies) that sat within the plastic cap atop the head. LFP data were continuously collected (200Hz sampling rate; Sirenia Acquisition, Pinnacle Technologies) until the completion of behavioral testing. Upon recovery from anesthesia, rats were moved to a dimly lit room adjacent to the behavioral testing room for pre-maze habituation (habituation times ranged from 30 to 90min, although all data analyses from this period only encompassed the last 30min of the habituation period). During the habituation period, rats remained in their home cage with access to water, ad libitum. Throughout habituation, rats were undisturbed by the experimenters with most rats engaging in an assortment of behaviors including quiet wake, active exploration, and sleep.

Following this habituation period, rats were moved into a dimly lit testing room with sparse visual cues on the walls. Rats were then trained on a dual-solution T-maze (opaque acrylic, arm length/width: 16 x 5 inches, wall height: 5.25 inches; Figure 1A) that can be successfully solved using either an allocentric “Place” strategy (i.e. external visual cues to guide navigation) or an egocentric “Response” strategy (i.e. specific body turn direction to guide navigation). During each training trial, rats were permitted up to 90s to make an arm choice and locate the food reward. Only one arm choice was permitted each trial. At the completion of each trial, rats were returned to their cage while the maze was cleaned to remove potential odor cues. Training continued until the rat reached a predetermined learning criterion (8/10 trials with correct arm choices). Upon successful acquisition of the task, rats were returned to their cage for five minutes. Following this five minute period, three probe trials were administered (one min inter-trial interval) in which rats were placed into a novel starting arm located directly across from their original starting arm (Figure 1A). For each probe trial, if the rat selects the choice arm in which the food reward was previously located, the trail is indicative of a “Place” learner (i.e. they appeared to navigate to the same visual cues). By contrast, selecting the opposing choice arm was indicative of a “Response” learner (i.e. they appeared to maintain the same direction of body turn). During probe trials, food reward was not present in either choice arm to limit the formation of novel positive associations during probe trials. Of note, either behavioral strategy can be successfully employed during training, but the specific strategy a rat is using is only characterized by their probe trial choice. Ultimately, the probe trials were used to classify whether an individual rat was employing a “Place” or “Response” strategy. Following previous literature (Canal et al., 2005; Chang & Gold, 2003; Korol et al., 2004; Packard, 1999; Spritzer et al., 2013), we opted to classify learning strategy based on performance in the first probe trial (to limit potential confounds of new learning occurring during probe trials). However, qualitatively similar results were obtained across our analyses if we instead classified behavior as the strategy chosen on the majority of the three probe trials (with only one rat’s classification changing as a result of this alternative classification strategy). All training and probe trials were recorded to facilitate subsequent behavioral analyses.

Figure 1.

Behavioral responses on a dual-solution T-maze. A) Schematic of the dual-solution T-maze during training and probe trials. Degrees reflect head direction vector angles used in C/D/E with −90 and 90 degrees indicating a left and right turn, respectively. B) Rats readily learn the location of food reward in a single training session. Trial block reflects a moving window of 5 trials. C) Instantaneous head direction observed between the approach and choice points is depicted across three example trials in which the rat ultimately turned left. D) Smoothed (six-sample moving window linear average) head direction for the same three trials depicted in C. By taking a linear average, head direction movements that result in sustained midline maze crossings produce a measure of head direction that crosses 0 degrees (dotted-lines above). Zero degree crossings within each trial were counted as a measure of vicarious trial and error (VTE). E) Average VTE counts during the first (Early) and last (Late) half of training trials for rats that exhibited a place strategy (dark bars) or a response strategy (light bars). * depicts VTE averages for each individual rat. Significant effects of response strategy, trials (early/late), and their interaction on VTE were observed (all p’s <0.05).

Experimental Design and Statistical Analyses

All electrophysiological data were processed and analyzed using custom scripts in Mathworks MATLAB (Natick, MA). Circular statistics were calculated using the MATLAB toolbox, CircStat (Berens, 2009). Repeated measures ANOVAs were performed with SPSS (IBM, Armonk, NY). All data are presented as mean ± the standard error of the mean. The data that support the findings of this study are available from the corresponding author upon reasonable request.

To analyze behavior during T-maze training, video recordings were manually scored for each trial to identify when each rat was placed into the maze, reached the decision point, and made an arm choice (Figure 1A). Total trial duration was also recorded. To more directly analyze rat behavior at the choice point, we used a behavioral tracking software (EthoVision XT with Multiple Body Points Module; Noldus, Leesburg, VA) to continuously assess the rat’s head direction. Previous reports indicate that at choice points, rats often engage in vicarious trial and error (VTE; Muenzinger & Gentry, 1931; Redish, 2016; E. C. Tolman, 1938), a commonly observed behavior in which rats pause and appear to survey possible choices. Choice-point, head direction vectors from EthoVision were exported to Matlab and a measure of VTE was calculated by counting the number of times these vectors crossed the maze midline.

Multiple analytical approaches (e.g. band-limited power (BLP), coherence, phase-locking) were used to characterize LFP activity in the hippocampus and striatum during maze training and pre-maze habituation. For both maze-training and habituation data, BLP and coherence were calculated in 1s epochs while values for instantaneous phase were obtained for every sampling point (200Hz). Analyses of LFP data during maze training were conducted separately within each trail (i.e. from maze-entry to maze-exit). Habituation data were analyzed separately within pre-determined 1min episodes, as there were no trials during this undisturbed period.

To characterize activity within each of these brain regions, we calculated band-limited power (BLP) across a range of frequencies (delta: 2-4Hz; theta: 5-9Hz; alpha: 10-14Hz; beta: 18-30 Hz; low-gamma: 40-70Hz; and high-gamma: 80-99Hz). Herein, LFP power spectra were first calculated via Welch’s method in 1s epochs (Hamming window). Resultant power spectra were averaged within each frequency bin and averaged across trials. To facilitate comparisons across all rats, trial-averaged power spectra for each rat were normalized by the average broadband power observed across all of that rat’s trials.

Two approaches were undertaken to assess relationships in LFP activity between the dorsal hippocampus and dorsolateral striatum. First, similar to our power spectra analyses above, we calculated magnitude-squared coherence across 1s epochs and calculated average, band limited coherence values across all training trials or during habituation. Second, we assessed striatal-hippocampal theta phase locking. Herein, raw LFP hippocampal and striatal signals for each trial (i.e. from the time rats are placed in the maze to the time they were removed) were filtered in the theta range (zero-phase Chebyshev Type II filter, band pass: 5.0 – 9.0Hz). Using these filtered signals, instantaneous phase was calculated from the angle of the complex vector obtained from the Hilbert transform (Dash et al., 2018; Vanhatalo et al., 2004). Instantaneous striatal-hippocampal phase differences (IPD) were calculated: IPD = e ^ (i * (P₁-P₂)), in which P₁ and P₂ are the instantaneous phase of the striatum and hippocampus, respectively. The average of IPD is a complex number in which the imaginary component is the average phase difference for that trial and the real part is the phase locking factor (PLF). The PLF ranges from 0-1 with values near 0 representing a uniform phase distribution and values near 1 indicating perfect phase synchrony (Dash et al., 2018; Vanhatalo et al., 2004). To assess the statistical significance of our observed PLFs, we compared these values to a distribution of randomized PLFs derived following the same approach after a Monte Carlo resampling of striatal and hippocampal trial data.

Lastly, we attempted to determine whether LFP activity could be used to train a support vector machine (SVM) to predict behavioral strategy. Two SVM models (linear kernels) were trained using seven features each (striatal-hippocampal theta phase differences and band-limited coherence from either maze training or habituation). K-Fold 10, cross-validation of SVM models was performed to evaluate model performance. Herein, data from rats are split into ten groups of two rats each. One group of rats is withheld as a testing set while the model is trained on the remaining nine groups. This processes is repeated (withholding a different group each time to serve as the testing set). Posterior probabilities (the predicted likelihood that a given set of features is derived from a rat that expressed a particular behavioral strategy) were calculated from these cross-validated SVM models. Using these posterior probabilities, we constructed receiver operating characteristic (ROC) curves and calculated the area under the curve (AUC) to measure SVM model performance. To determine whether the performance of these SVM models differed from that expected from chance alone, we compared observed model performance to distributions comprised of SVM models trained using randomized feature data. For each random SVM model, we first shuffled scores within each feature across the twenty rats. We then followed our original procedures to train the SVM model, calculate posterior probabilities, and construct ROC curves. Distributions depicting chance SVM performance were constructed from 500 random SVM models.

Results

Dual-solution T-maze Behavior

We trained twenty rats to find a food reward within a dual-solution T-maze that can be solved with either an egocentric “response” strategy or an allocentric “place” strategy (Figure 1A). Despite each individual rat being trained under similar conditions, our experimental approach did not appear to bias rats to develop a specific behavioral strategy; the number of rats that exhibited a place strategy (N=8) or a response strategy (N=12) during post-learning probe trials did not significantly differ from that expected by chance alone (binomial test, p=0.25). Moreover, average probe trial duration did not significantly differ (t (18) = −0.09, p=0.93) between place (7.44 ± 1.90s) and response (7.73 ± 2.39s) rats.

Behavioral strategy did not significantly affect the number of trials required to learn the location of the food reward (t (18) = 1.59, p=0.13) with rats requiring an average of 18.5 ± 1.63 trials to reach the predetermined behavioral learning threshold (8 correct choices within the past 10 trials with the last 2 trials also having to be correct; Figure 1B). However, rats exhibiting a response strategy had significantly shorter average trial durations than those exhibiting a place strategy (10.35 ± 1.39 seconds and 22.79 ± 5.55 seconds, respectively; t (18) = 2.44, p=0.03). This difference in average trial duration could reflect the differential demands of cognitive processes necessary for a deliberative allocentric strategy as compared to a habitual egocentric strategy (Redish, 2016; Schmidt et al., 2013). Place and response rats did not significantly differ in their behavior during the approach to the choice point: average approach times did not differ across behavioral strategies (t (18) = 1.38, p=0.18) nor did exploratory rearing behavior (t (18) = 1.01, p=0.33). Instead of differences in running speed during approach altering trail durations, we observed that differences in trial duration largely arose at the choice point; after completing their approach to the choice point, place animals took on average 13.30 ± 3.79 seconds to enter a choice arm while response animals took only 4.14 ± 0.66 seconds.

Differences in vicarious trial and error (VTE, Muenzinger & Gentry, 1931; Redish, 2016; E. C. Tolman, 1938)) could account for these observed differences in average trial duration. VTE has been reported to increase when animals exhibit a place strategy as opposed to a response strategy and when place/response strategies are in apparent competition (Schmidt et al., 2013). As a measure of this exploratory behavior, we counted alternations in head direction between the choice arms while the rat was at the choice point and compared this measure of VTE across behavioral strategies and between early/late training trials (Figure 1C/D). Place rats exhibited significantly more VTE than response rats (F (1, 18) = 7.37, p=0.01). Moreover, as compared to the first half of trials, significantly less VTE was observed in the second half of trials (F (1,18) = 5.09, p=0.04), with this effect appearing to be driven by a reduction in place animal VTE (F (1,18) = 5.36, p=0.03; see Figure 1D). Thus, VTE appears elevated when rats employ a deliberative allocentric approach and early in the learning process when uncertainty is high. Consistent with these results, we also observed a trend for more “failed” trials (i.e. when a rat does not make an arm choice in 90s) in place rats than in response rats (place: 4.25 ± 1.52 failed trials, response: 1.75 ± 0.46; t (18) = 1.85, p = 0.08).

Dorsolateral Striatal and Dorsal Hippocampal Phase Differences

Neuronal activity within the hippocampus and striatum has previously been associated with place and response strategies, respectively (Bohbot et al., 2007; Iaria et al., 2003; Lee et al., 2008; Packard & McGaugh, 1996). Differences in behavioral strategy employed by rats during the ambiguous T-maze, therefore, may arise from changes in neuronal activity within these structures and/or as a result of distinct patterns of signaling between these brain regions. To address these possibilities, we first explored phase differences between band-pass filtered dorsal hippocampal and dorsolateral striatal LFP activity within rats undergoing training on the T-maze. As evident in the individual examples presented in Figure 2A/B, place and response strategies were associated with distinct patterns of signaling within the theta band (5-9Hz); hippocampal-striatal theta activity was consistently in phase for response learners and out of phase for place learners. Similar phase differences were apparent across all training trials and all rats (Figure 2C). Phase differences were not uniformly distributed in place or response learners (Rayleigh test (Berens, 2009): z = 38.54, p=6.27 x 10⁻¹⁹; z = 92.06, p=4.49 x 10⁻⁴⁸, respectively). Mean phase differences for place learners (109.01 ± 18.23°) significantly differed from those of response learners (11.27 ± 10.43°; Watson-Williams test (Berens, 2009): F (1,18) = 15.94, p=8.54 x 10⁻⁴; Figure 2D). Additionally, theta phase differences in place, but not response, rats significantly differed from a mean phase difference of 0° (circular mean test, Berens, 2009), p<0.05). To determine whether these apparent phase differences reflect consistent phase locking, we calculated the phase-locking factor (PLF) for our observed data and compared it to a distribution of PLFs derived from shuffled hippocampal-striatal LFP pairings. As evident in Figure 2E, both place and response rats exhibited significant phase locking (Monte Carlo resampling, p<0.01). Thus, while theta activity in both place and response rats is significantly phase locked between the dorsal hippocampus and dorsolateral striatum, theta is synchronous in response rats and out-of-phase in place rats.

Figure 2.

Hippocampal and striatal phase differences during maze learning. A) Filtered theta activity in the hippocampus and striatum during across a representative training trial in a response (top) and place (bottom) rat (maze locations: a – approach, c – choice). B) Polar histograms of instantaneous phase differences derived from the representative data in A depict the proportion of time each phase difference was observed. C) Polar histogram for place and response rats across all rats and all trials. D) Average phase differences for place (grey) and response (black) rats differed significantly (p<0.001). E) Phase locking factors (PLF) for observed data (solitary bars) and shuffled data (histograms). Significant phase locking (p<0.01) was observed for both place and response rats.

Dorsolateral Striatal and Dorsal Hippocampal LFP Power and Coherence

To further describe patterns of neuronal activity associated with behavioral strategy, we characterized LFP band-limited power (BLP) in place and response rats during maze learning. Within each rat, power spectra during each trial were first normalized to the mean broadband power across all trials (see Figures 3A/C for individual examples of BLP within the dorsal hippocampus and dorsolateral striatum across a single trial for a place rat and a single trial for a response rat). We then calculated BLP across six different frequency bands and averaged BLP values across all trials and all rats as a function of behavioral strategy (see methods and Figure 3B/D). As expected by the typical 1/f nature of the power spectrum (Bédard et al., 2006), we observed a significant effect of frequency band on power within both the hippocampus (F (5,90) = 83.20, p=3.19 x 10⁻³²) and the striatum (F (5,90) = 92.99 p=5.14 x 10⁻³⁴). We did not, however, observe a significant interaction between frequency band and behavioral strategy in either brain region (hippocampus: F (5,90) = 0.80, p=0.55; striatum: F (5,90) = 0.22, p=0.96). Similarly, no individual t-tests to assess potential differences in BLP within a single frequency band between place and response learners were statistically significant (range of p values; hippocampus: 0.21 – 0.44, striatum: 0.09 – 0.95). Thus, differences in BLP in the two brain regions of interest during maze learning are not associated with behavioral strategy.

Figure 3.

Hippocampal and striatal LFP power and coherence during maze learning. A/C) Hippocampal (A) and striatal (C) LFP power across a representative trial for a place and representative trial for a response rat. B/D) Average band-limited power within the hippocampus (B) and striatum (D) across all trials for all place and response rats. Band-limited power (BLP) data were normalized within each rat as a function of that rat’s mean broadband power. E) Magnitude-squared coherence between hippocampus and striatum during a representative place and representative response trial. F) Average coherence across all trials for all place and response rats. Response rats exhibit significantly higher hippocampal-striatal coherence in the theta, alpha, and beta bands (* p<0.05).

Despite a lack of overall differences in maze BLP, place and response learners could differ in the coordinated expression of hippocampal and striatal activity within distinct frequency bands. To test this possibility, we calculated magnitude-squared coherence between dorsal hippocampal and dorsolateral striatal LFPs while rats performed on the maze (see Figure 3E for individual examples). Across all trials and all rats (Figure 3F), we observed a significant main effect of frequency on coherence (F (5,90) = 15.41, p =6.30 x 10⁻¹¹), a significant interaction between frequency and learning strategy (F (5,90) = 3.06, p =0.01), and a trend for a main effect of strategy (F (1,18) = 3.80, p=0.07). Post-hoc t-tests revealed that response learners exhibited significantly higher coherence across theta, alpha, and beta bands than place learners (p = 0.03, 0.02, and 0.04, respectively). Different patterns of activity across the dorsal hippocampus and dorsolateral striatum during maze learning, therefore, appear associated with each behavioral strategy.

Behavioral Strategy Associated Differences in Dorsolateral Striatal and Dorsal Hippocampal LFP Power and Coherence are present prior to maze exposure

As all rats were trained under similar experimental conditions, differences in striatal-hippocampal activity associated with behavioral strategy were unlikely to have arisen as consequence of experimental approach. Instead, they may reflect interindividual differences present prior to maze learning. We therefore analyzed dorsal hippocampal and dorsolateral striatal LFP activity during a 30min habituation period prior to maze exposure. Similar to that observed during maze learning, striatal-hippocampal theta coupling differed between place and response learners during habituation (Figure 4A). During habituation, phase differences were not uniformly distributed in place or response learners (Rayleigh test, Berens, 2009): z = 89.47, p=5.61 x 10⁻⁴⁴; z = 263.19, p=7.32 x 10⁻¹⁵¹, respectively; Figure 4B). Dorsal hippocampal and dorsolateral striatal theta oscillations in place rats were consistently out of phase (106.69 ± 17.74°), but were typically in phase in response rats (22.10 ± 7.08°). These phase differences between place and response learners were significantly different (Watson-Williams test, (Berens, 2009): F (1,18) = 16.92, p=6.53 x 10⁻⁴; Figure 4C). Overall, average theta phase differences observed in rats across habituation were highly correlated with average theta phase differences across maze learning (circular correlation (Berens, 2009): ρ_cc (18) = 0.85, p=0.002). Interindividual differences in striatal-hippocampal theta activity, therefore, appear present prior to introduction to the maze, continue during maze training, and ultimately are associated with distinct behavioral strategies.

Figure 4.

Differences in theta phase and hippocampal-striatal coherence between place and response learners are present prior to maze exposure. A) Instantaneous theta phase differences during a 30min habituation for each rat. B) Polar histogram of instantaneous phase differences during habituation (A) depict the proportion of time each phase difference was observed. C) Average phase differences for place (grey) and response (black) rats during habituation differed significantly (p<0.001). D) Average coherence during habituation for all place and response rats. Response rats exhibit significantly higher hippocampal-striatal coherence in the delta, theta, alpha, and beta, and low gamma bands (* p<0.05).

Like theta phase, differences in striatal-hippocampal coherence associated with behavioral strategy also appear to be present prior to maze learning (Figure 4D). During habituation, we observed significant main effects of both frequency (F (5,90) = 7.05, p=2.21 x 10⁻⁵) and behavioral strategy (F (1,18) = 4.43, p=0.05) on striatal-hippocampal coherence. No significant interaction between frequency and strategy was observed (F (5,90) = 0.88, p = 0.50). Post-hoc t-tests revealed that response learners exhibited significantly higher coherence across delta, theta, alpha, beta, and low gamma bands than place learners (all p’s <0.05; Figure 4D). Thus, elevated striatal-hippocampal coherence is present in response learners prior to maze learning. Moreover, these coherence differences were strongly maintained during maze learning; significant correlations were observed between habituation coherence and maze learning coherence across all six frequency bins (range of r values: 0.72 – 0.93, all p’s <0.001).

Interindividual Differences in in Dorsolateral Striatal and Dorsal Hippocampal LFP Activity predict behavioral strategy

Interindividual differences in hippocampal-striatal activity present during habituation (and expressed during maze learning) may have biased rats to adopt a particular behavioral strategy on the dual-solution T-maze. Using data from either habituation or maze learning, we trained support vector machines (SVM) to classify behavioral strategy from a feature set comprised of striatal-hippocampal phase differences and coherence. Once trained, these SVM classifiers were used to calculate posterior probabilities (i.e. the predicted likelihood that a rat will exhibit a particular behavioral strategy). SVM classifiers reliably segregated place and response rats (Figure 5A/B) when using either habituation (Cohen’s d = 1.87) or maze learning data (d = 1.73). By contrast, when we trained these SVM classifiers using shuffled feature data (see methods), place and response rats where poorly classified (d = 0.11 and 0.24, shuffled habituation and maze learning, respectively). Indeed, posterior probabilities derived from these shuffled SVM models were clustered around 0.6 for both place and response rats (Figure 5A/B), as expected from a random classifier given the observed 60/40 proportion of response/place rats observed in the present study.

Figure 5.

Support vector machine (SVM) classification of place and response learners. A/B) Histograms of the posterior probability of being a response learner for each rat as derived from SVM classifiers trained on either habituation (A) or maze (B) data. Data are plotted separately for rats that were behaviorally identified as place (light) or response (dark) learners. Shuffled distributions were generated by randomly shuffling data between rats within each feature prior SVM classifier training. C/D) Receiver operating characteristic curves derived from posterior probabilities presented in A/B. Note that SVM models trained on observed data (dark) have high sensitivity and specificity while the average SVM model trained on shuffled data (light) performs near chance (dotted line). Insets depict the area under the curve (AUC) for observed and shuffled data.

To further analyze the output of our SVM classifiers, receiver operating characteristic (ROC; Figure 5C/D) curves were calculated from the posterior probabilities described above. ROCs derived from shuffled habituation (area under the curve (AUC) = 0.47 ± 0.02) or maze data (AUC = 0.49 ± 0.02) reflect classifiers that operate near chance levels. By contrast, ROCs derived from observed habituation (AUC = 0.86) or maze data (AUC = 0.88) reflect classifiers that reliably predict behavioral strategy and perform significantly better as compared to shuffled classifiers (Monte Carlo resampling, both p’s <0.01). Indeed, ROCs for both habituation and maze data reflect classifiers that can exhibit both high sensitivity and selectivity (i.e. high true positive rates and low false positive rates). Features of striatal-hippocampal activity, therefore, can serve as reliable predictors of behavioral strategy on the dual-solution T-maze.

Discussion

We successfully trained 20 rats on the dual-solution T-maze. Consistent with previous reports (Devan & White, 1999; McDonald & White, 1994), a similar proportion of rats (N=8) expressed a place strategy as those that expressed a response strategy (N=12), despite all rats having been trained under the same experimental conditions. Behavioral strategy did not alter the number of trials required to learn the location of the food reward, although place rats spent significantly more time at the choice point and exhibited significantly more vicarious trial and error (VTE). Interindividual differences in behavioral strategy were associated with contrasting patterns of neuronal activity between memory systems previously associated with response and place strategies: 1) striatal-hippocampal theta activity was in-phase in response rats and out-of-phase in place rats and 2) response rats exhibited elevated striatal-hippocampal coherence across a wide range of frequency bands. These differences were already present during pre-maze habituation and were consistently maintained across habituation (e.g. Figure 4A) and throughout each trial during maze learning (e.g. Figures 2B and 3E). Instead of reflecting specific behavioral components of maze performance (e.g. running speed, learning, etc.), these measures therefore appear to reflect spontaneous differences in brain activity that may bias individual rats to develop a particular behavioral strategy when exposed to a maze environment that did not have external cues strongly promoting either behavioral strategy. Consistent with this idea, we were able to train support vector machines (SVM) to predict the behavioral strategy of individual rats using either pre-maze or maze-learning LFP data. Distinct dorsolateral striatal and dorsal hippocampal activity regimes, therefore, appear to promote the adoption of distinct behavioral strategies.

Multiple memory systems appear to support navigation of the environment including a hippocampal-dependent “place” system reliant upon external spatial cues and a striatal-dependent “response” system reliant upon egocentric body turns (Ferbinteanu, 2019; Gold et al., 2013; Johnson et al., 2007; Mizumori et al., 2004; Packard & Goodman, 2013; Tolman, 1949). In the present study, place and response strategies were equally effective for learning the location of an appetitive reward (i.e. number of trials to criterion did not significantly differ by behavioral strategy). Place rats, however, spent significantly more time at the choice point during which time they expressed significantly more VTE. VTE has been proposed to reflect deliberative decision making (Papale et al., 2012; Redish, 2016) and has previously been shown to be higher in rats forced to use a place strategy as opposed to a response strategy (Schmidt et al., 2013). Our results in which individual rats spontaneously express one of these behavioral strategies extend these previous observations and further suggest that differences in VTE reflect strategy-dependent differences in the need for deliberation independent of external environmental conditions. Consistent with this idea, we also observed that VTE decreased in association with learning. Thus, while distinct behavioral strategies contribute to behavioral flexibility across diverse environmental conditions (Ferbinteanu, 2019; Goodroe et al., 2018; O’Doherty et al., 2017; Packard & Goodman, 2013), a deliberative place strategy may impose additional cognitive demands than a habitual response strategy. Mechanisms that enable an animal to select an appropriate navigational strategy in response to external conditions and/or internal needs are therefore critical for generating optimal behavior (Shelton et al., 2013).

While the presence of multiple memory systems appears to enhance behavioral flexibility, the mechanisms by which these disparate systems can produce a unitary behavioral output remains unclear. Pharmacological inactivation of core structures of distinct memory systems selectively impairs behavioral strategy (Packard & McGaugh, 1996) thereby suggesting that these systems can function independently. Subsequent research however, has challenged the independence of these systems by demonstrating that they may interact with one another in apparent collaboration (Brown et al., 2012; Ferbinteanu, 2016), competition (An et al., 2018; Bohbot et al., 2007; Lee et al., 2008), or both depending upon task familiarity (Jacobson et al., 2012). Consequently, it has been proposed that memory systems process information in parallel and that interactions between these parallel systems ultimately contribute to behavioral output (McDonald et al., 2004; Mizumori et al., 2004; White & McDonald, 2002). Mechanistically, dynamic reconfiguration of communication between these networks to enable the formation of transient memory meta-systems may enable the classically defined multiple memory systems to generate the appropriate behavioral output in response to internal and/or external conditions (Ferbinteanu, 2019).

In the present study, we observed differences in dorsolateral striatal/dorsal hippocampal LFP activity that appear consistent with such a dynamic network model. Spontaneous adoption of a place or response strategy was not associated with changes in neuronal activity within discrete memory systems. Specifically, we did not observe any significant difference in dorsal hippocampal or dorsolateral striatal band-limited power (BLP) across any frequency as a function of behavioral strategy. This similarity may reflect extensive parallel processing within hippocampal and striatal networks; during navigation using either behavioral strategy these structures exhibit extensive functional redundancy as individual neurons within each region are responsive to spatial location, cue changes, movement, directional heading, and/or reward location (Mizumori et al., 2004; Yeshenko et al., 2004). By contrast, place and response strategies may be characterized by differences in communication between memory systems. We observed significant differences in striatal-hippocampal theta phase coupling associated with behavioral strategy. Moreover, as compared to response rats, place rats exhibited significantly lower striatal-hippocampal coherence across a wide range of frequencies. While these phase coupling and coherence data do not provide direct, causal evidence for communication between these structures (e.g. they may instead arise from inputs from another shared regions), they may reflect reconfigurations of communication between these memory systems that result in the selective expression of behavioral strategy. Indeed, using only features reflective of striatal-hippocampal communication (i.e. theta phase differences and band-limited coherence), we were able to train SVMs to reliably predict behavioral strategy.

The specific patterns of dorsolateral striatal and dorsal hippocampal LFP activity observed in our study may reflect the different functional contributions of these brain regions to behavioral strategy selection and/or expression. In addition to functional differentiation of hippocampal-dependent place strategies and striatal-dependent response strategies, subregions of the striatum appear to make distinct contributions to behavior. Broadly speaking, activity within the dorsolateral striatum has been associated with stimulus-response and habitual behaviors, the dorsomedial striatum with choice and behavioral flexibility, and the ventral striatum with reward salience and valuation (Devan & White, 1999; Ferbinteanu, 2016; Ragozzino et al., 2009; Thorn et al., 2010; van der Meer et al., 2010). As compared to place rats, we observed that response rats exhibited significantly greater LFP coherence between dorsolateral striatum and dorsal hippocampus across many frequency bands. Similarly, it has been previously reported that dorsolateral striatum/hippocampal theta coherence 1) exhibits a wide range across individual rats and 2) reliably increases in rats that successfully learn a cue-based associative T-maze (DeCoteau et al., 2007b). Strikingly, cue-dependent tasks have also been associated with coherent theta activity between dorsolateral and dorsomedial striatum (DeCoteau et al., 2007a) and ventral striatum and the hippocampus (Lansink et al., 2009). Thus, elevated coherence across striatal subregions and the hippocampus appears to be associated with the expression of both cue-dependent and spontaneous response strategies.

The in-phase dorsolateral striatal/dorsal hippocampal theta oscillations that we observed in response rats may provide a mechanism to generate such coherent activity. During cued navigation on the T-maze, phase-amplitude coupling between theta oscillations and higher frequency power has been described within and between the dorsolateral striatum and hippocampus (Tort et al., 2008). Interestingly, this coupling appears to be asymmetric, with striatal theta having a more pronounced impact on hippocampal activity than in reverse (Tort et al., 2008). Given the putative role for the dorsolateral striatum in producing habitual behavior, this asymmetry may explain why we observed coherent activity between the dorsolateral striatum and hippocampus associated with spontaneous adoption of a response strategy. Synchronous theta oscillations between the dorsolateral striatum and hippocampus may therefore: 1) promote information transfer between neuronal populations (Vidaurre et al., 2018; Wilson et al., 2015), 2) provide a mechanism by which higher frequency activity can be reconfigured across multiple memory systems (Tort et al., 2008), and 3) contribute to a communication regime that facilitates response behavior. By contrast, when we observed out-of-phase dorsolateral striatal/hippocampal theta, LFP coherence was diminished across a wide range of frequencies and rats typically adopted a place strategy.

A limitation of the present study is that we are unable to address the root cause(s) of interindividual differences in neuronal activity associated with place or response strategies. One possibility is that systemic variation in electrode placements between place and response rats was responsible for differences in LFP activity observed across behavioral strategy. For example, theta phase differences in the present study could simply reflect the gradual progression in theta phase that occurs in association recording depth within the hippocampus (Bragin et al., 1995). Although we do not have histological verification of electrode placements to directly preclude this possibility, we find this explanation unlikely as 1) we have no reason to expect that electrode location would systematically vary in association with behavioral strategy and 2) place/response rats did not exhibit significant differences in theta or gamma power despite these electrophysiological measures also systematically varying as a function of hippocampal depth (Bragin et al., 1995).

Related concerns also limit our ability to fully characterize the neuronal origins of our striatal LFPs. While some studies support a local origin for striatal LFP activity (e.g. DeCoteau et al., 2007a; Tanaka & Nakamura, 2019; Tingley et al., 2018), the non-laminar organization of the striatum may cause currents generated within the striatum to cancel and thereby render the striatal LFP particularly sensitive to volume-conducted signals generated within nearby structures including the hippocampus (Carmichael et al., 2017; Lalla et al., 2017). Should the striatal theta activity observed in the present study primarily have a non-striatal origin, theta coherence and phase-differences detailed herein may reflect behavioral strategy dependent alterations in neuronal activity within hippocampal networks as opposed to across multiple memory systems as detailed above. Here, interactions between the medial septum and hippocampus may contribute to the distinct patterns of theta coordination associated with behavioral strategy. Distinct interneuron populations within the medial septum fire at the peak and trough of hippocampal theta (Borhegyi, 2004) and septal/hippocampal coupling exhibits state-dependent modulation (Dragoi et al., 1999). Behavior and brain state further modulate theta phase coupling within hippocampal networks including across the multiple hippocampal dipoles (Montgomery et al., 2009) and interneuron populations (Klausberger et al., 2003) that produce theta oscillatory activity. Thus, extensive opportunity for reconfiguration of theta coupling appears to be characteristic of septal/hippocampal circuits and may consequently contribute to behavioral variability. Future studies that directly characterize the source(s) of the dorsolateral striatal and dorsal hippocampal LFP activity described herein, however, appear critical for understanding the neuronal underpinnings of behavioral strategy selection.

As external testing conditions were identical for each rat however, our observed behavioral differences appear to have arisen from interindividual differences in internal brain state. Such differences could be structural; previous reports indicate that greater hippocampal grey matter volume is associated with the spontaneous adoption of a place strategy (Bohbot et al., 2007), and increased fractional anisotropy of the right hippocampus is related to more efficient use and formation of a cognitive map (Iaria et al., 2008). Alternatively, differences in behavioral strategy may reflect more transient alterations in internal brain state. Changes in neurotransmission (e.g. glutamate (Packard, 1999), acetylcholine (Chang & Gold, 2003; McIntyre et al., 2003; Pych et al., 2005), or dopamine (Braun et al., 2015; Lex et al., 2011)) or metabolite availability (e.g. glucose (Canal et al., 2005) or lactate (Newman et al., 2017)) within the hippocampus and/or striatum have all been associated with the expression of a particular behavioral strategy. Each of these neurotransmitters can likewise modulate theta oscillatory activity (Buzsáki, 2002; Costa et al., 2006; Lemaire et al., 2012; Zhang et al., 2010) and may thereby contribute to the distinct striatal/hippocampal activity regimes that we observed associated with interindividual differences in behavioral strategy. It remains to be seen whether the multitude of external and internal factors that shape the expression of a particular behavioral strategy during navigation (Chersi & Burgess, 2015; Ferbinteanu, 2019; Goodroe et al., 2018; Packard & Goodman, 2013; Retailleau et al., 2012) do so through the formation of similar neuronal activity regimes. Dynamic reorganization of activity across multiple memory systems, however, appears to be an effective mechanism for the selection of behavior. Further study is merited to determine whether these systems reconfigure in a dichotomous manner thereby producing the expression of either a fully place or fully response strategy, or whether an individual rat’s behavioral strategy selection exists along a continuum (with reorganization instead shifting the probability of implementing a given strategy).