Abstract
Communication between the cerebellum and forebrain structures is necessary for motor learning and has been implicated in a variety of cognitive functions. The exact nature of cerebellar–forebrain interactions supporting behavior and cognition is not known. We examined how local and network activity support learning by simultaneously recording neural activity in the cerebellum, amygdala, and anterior cingulate cortex while male and female rats were trained in trace eyeblink conditioning. Initially, the cerebellum and forebrain signal the contingency between external stimuli through increases in theta power and synchrony. Neuronal activity driving expression of the learned response was observed in the cerebellum and became evident in the anterior cingulate and amygdala as learning progressed. Aligning neural activity to the training stimuli or learned response provided a way to differentiate between learning-related activity driven by different mechanisms. Stimulus and response-related increases in theta power and coherence were observed across all three areas throughout learning. However, increases in slow gamma power and coherence were only observed when oscillations were aligned to the cerebellum-driven learned response. Percentage of learned responses, learning-related local activity, and slow gamma communication from cerebellum to forebrain all progressively increased during training. The relatively fast frequency of slow gamma provides an ideal mechanism for the cerebellum to communicate learned temporal information to the forebrain. This cerebellar response-aligned slow gamma then provides enrichment of behavior-specific temporal information to local neuronal activity in the forebrain. These dynamic network interactions likely support a wide range of behaviors and cognitive tasks that require coordination between the forebrain and cerebellum.
SIGNIFICANCE STATEMENT This study presents new evidence for how dynamic learning-related changes in single neurons and neural oscillations in a cerebellar–forebrain network support associative motor learning. The current results provide an integrated mechanism for how bidirectional communication between the cerebellum and forebrain represents important external events and internal neural drive. This bidirectional communication between the cerebellum and forebrain likely supports a wide range of behaviors and cognitive tasks that require temporal precision.
Introduction
The cerebellum (CB) has traditionally been viewed as a motor structure, but emerging evidence suggests that network communication between the cerebellum and forebrain systems plays key roles in various cognitive functions (Schmahmann and Sherman, 1998; Buckner, 2013; Koziol et al., 2014; Parker et al., 2017; Sokolov et al., 2017; Liu et al., 2022). Despite the importance of cerebellar contributions to cognition, the mechanisms underlying cerebellum–forebrain communication remain elusive. The cerebellum may communicate with the forebrain through network oscillations that are driven by temporal patterns of neuronal activity in cerebellar projection neurons (Parker et al., 2017; McAfee et al., 2019).
Measurement of communication among brain areas during learning can be accomplished by targeted electrophysiological recording of the local neuronal firing and network oscillations. Network oscillations in both theta (4–12 Hz) and slow gamma (25–50 Hz) ranges show increases in power to the training stimuli used in various learning paradigms (Fell et al., 2001; Kahana et al., 2001; Hasselmo et al., 2002; Bastiaansen and Hagoort, 2003; Jensen and Lisman, 2005; Osipova et al., 2006; Rizzuto et al., 2006; Colgin et al., 2009; Nyhus and Curran, 2010; Colgin, 2013, 2016; Lisman and Jensen, 2013). Moreover, theta and gamma oscillations show synchronization across brain areas after learning has taken place (Osipova et al., 2006; Benchenane et al., 2010; Nyhus and Curran, 2010; Colgin, 2013, 2016; Voloh et al., 2015). Neural processing supporting learning and retrieval has been proposed to depend on coordination between local and network activity both within and across brain areas (Hyman et al., 2005; Benchenane et al., 2010; Dzirasa et al., 2009; Jones and Wilson, 2005; Kim et al., 2011; Sigurdsson et al., 2010).
Local activity in the cerebellum generates adaptively timed behavioral responses, and this learning-related cerebellar activity is particularly well documented for associative learning in eyeblink conditioning (Choi and Moore, 2003; Freeman et al., 2005; Halverson et al., 2010, 2015, 2018, 2022; McCormick and Thompson, 1984b; ten Brinke et al., 2017). It is also well established that theta power in the cerebellum is correlated with both learning rate and behavioral expression in eyeblink conditioning (Hoffmann and Berry, 2009). Cerebellar output not only generates the temporally specific neural drive for behavioral timing but also communicates that temporal information with other brain areas (Clark et al., 1984; Halverson et al., 2010; Pisano et al., 2021; Sears and Steinmetz, 1990; Siegel and Mauk, 2013). Eyeblink conditioning is therefore an ideal behavioral context for examining the nature of cerebellum–forebrain communication.
The CB receives input from the central amygdala (AM) and anterior cingulate cortex (ACC) via the pontine nucleus, which is crucial for eyeblink conditioning (Farley et al., 2016; Kalmbach et al., 2009; Siegel et al., 2012; Weible et al., 2007). Lesions or inactivation of either the AM or ACC prevent eyeblink conditioning when a temporal gap (500 ms) occurs between the training stimuli (trace conditioning; Siegel et al., 2015; Weible et al., 2000). These studies provide evidence that the CB and forebrain structures communicate to support learning and expression of properly timed cerebellum-driven motor responses. This communication has been proposed to be unidirectional from the forebrain to the cerebellum; however, the role of bidirectional communication has not received much consideration. In the current experiment, we simultaneously recorded local activity and network communication among the CB, AM, and ACC during trace eyeblink conditioning. Our experimental question was to determine the specific nature of dynamic cerebellar interactions with the ACC and AM during progressive learning stages. We focused on cerebellar bidirectional interactions with the AM and ACC because these areas provide relatively direct input to CB, and each area is essential for initial learning of trace conditioning (Farley et al., 2016; Kalmbach et al., 2009; Kronforst-Collins and Disterhoft, 1998; Siegel et al., 2015; Takehara et al., 2003).
Materials and Methods
Subjects
The subjects were seven male and two female Long–Evans rats (250–400 g). The rats were individually housed in the animal colony in Bowen Science Building at the University of Iowa. All rats were maintained on a 12 h light/dark cycle and given ad libitum access to food and water. All procedures were approved by the Institutional Animal Care and Use Committee at the University of Iowa.
Surgery
One week before the start of recordings and behavioral training, rats were removed from their home cages and anesthetized with isoflurane. After the onset of anesthesia, electromyograph (EMG) electrodes (stainless steel) were implanted into the upper left orbicularis oculi muscle. The EMG electrode leads terminated in gold pins in a plastic connector. A bipolar stimulating electrode for delivering the shock [unconditioned stimulus (US)] was implanted subdermally, caudal to the left eye. During surgery, a custom-made hyperdrive array fitted with a 16-tetrode electronic interface board (EIB; EIB-36-16TT or EIB-72-QC-Small, Neuralynx) was implanted over the left IPN, right AM and right ACC. Stereotaxic coordinates taken from bregma (+ for posterior), midline (+ for lateral), and the skull surface (+ for ventral) targeting three ROIs were as follow: +11.4 mm, +3.0 mm, and +5.6 mm for the CB; +2.2 mm, +4.5 mm, and +7.5 mm for the AM; 0 mm, +0.5 mm, and +0.5 mm for the ACC (Fig. 1A). For each placement, the skull surface was marked for each tetrode bundle and drilled out with a square blunt drill bit. Skull fragments were carefully removed along with the dura matter over each placement under visual guidance. Eight holes were also drilled for eight skull screws. A silver wire attached to one skull screw was used for the EMG ground, and a silver wire attached to another skull screw was used for the EIB ground. Bundles of tetrodes were lowered to the surface of the brain and sealed with low-viscosity silicon (Kwik-Sil, World Precision Instruments). Seven tetrodes (six recording, one reference) in the CB bundle and six tetrodes (five recording, one reference) in both AM and ACC bundles were lowered to the brain surface. Tetrodes were slowly advanced so that each was ∼0.5 mm above the targets. Each independently movable tetrode was composed of four nichrome wires (12 µm diameter, Kanthal Palm Coast), twisted and partially melted together to form a tetrode. Tetrodes targeting CB were gold plated to reduce final impedance to 500–1000 kΩ, and CeA and ACC tetrodes were plated to 250–500 kΩ measured at 1 kHz (Nano Z, Neuralynx). The EMG, bipolar electrode, skull screws, and hyperdrive were secured to the skull with bone cement (Biomet Bone Cement R, Zimmer Biomet). A 3D printed screw top cone was also cemented around the hyperdrive to protect the tetrodes between recording sessions.
Conditioning procedure
Rats were initially habituated to the conditioning chamber for at least 2 d before training. The tone conditioned stimulus (CS) used in training was a 250 ms 2 kHz pure tone (85 dB). Each rat initially received an unpaired session consisting of 100 presentations of the 250 ms tone and 25 ms shock to obtain a baseline of behavioral and single neuronal responses to the training stimuli. Rats then received daily paired sessions of trace eyeblink conditioning with the same 250 ms tone CS, a 500 ms stimulus free trace interval followed by the 25 ms periorbital shock US (0.8–2.0 mA; Fig. 1C). Training sessions consisted of 100 trials per day with a pseudorandom distribution of intertrial intervals between 18 and 42 s that averaged 30 s. Trace conditioning sessions consisted of 10 blocks of 9 paired CS-US presentations and a CS-alone trial. Before the start of each session the EMG, bipolar electrode, and EIB were connected to lightweight tethers that were connected to a commutator that allowed the rat to freely move inside the conditioning chamber. Rats were allowed to acclimate to the conditioning chamber for at least 5 min before the start of each session. These sessions continued until rats reached ∼80% conditioned responses (CRs) for two consecutive sessions.
Simultaneous neural recording
Neural data acquisition has been described in detail in our earlier work (Halverson et al., 2010). Briefly, tetrode bundles over the CB, AM, and ACC were advanced to within 0.5 mm of the targets during surgery using 0–80 screws (McMaster-Carr) to precisely control tetrode depth. During daily turning sessions (∼1 h), tetrodes were advanced in 25 µm steps, and changes in neuronal activity were closely monitored for each target area. Once stable single neurons were encountered in all three target structures, neuronal recordings and behavioral training commenced. Neuronal signals were first preamplified at unity gain. The signals were then amplified to fit into windows ranging from 250 to 1500 µV. The amplification was the same for each channel of a given tetrode. Neural signals that exceeded a channel amplitude threshold (typically 40–55 µA) were digitized and stored at 32 kHz (Cheetah software, Neuralynx). Each target area had a unique local reference tetrode throughout the data acquisition to prevent local activity of one area from contaminating the others. Having a unique local reference also helps to minimize the possibility of systematic artifacts, such as volume conduction, from contaminating the single neuron and field potential recordings.
Neuron isolation and criteria
Waveform characteristics of the single neurons were plotted as a scatterplot of one of the electrodes versus another in the peak, energy, and valley planes. The peak presentation is the maximum height (positive amplitude) of the waveform after crossing threshold. The energy presentation represents the square root of the sum of the squared points for the entire waveform. The valley presentation is the maximum depth (negative amplitude) of the waveform. All isolated single neurons were represented by clusters of points on these different presentations and the initial boundary for a given cluster (single neuron) was established using peak (Fig. 1B). Peak was used as the primary feature for isolating single neurons as this feature was used for thresholding. Any potential cluster that was not completely above the threshold in the peak presentation for a given tetrode was not included in the analysis. Energy and valley presentations were used to further differentiate clusters initially isolated in peak. Cluster boundaries were always drawn in the same presentation space (e.g., peak to peak), and no cross-presentations were used for isolation (e.g., peak to energy). On the initial separation, each cluster was further assessed with its physiological parameters such as interspike interval, autocorrelation, firing rate, number of spikes, stability over time, likelihood ratio (l-ratio), and isolation distance, in addition to the histologic verification of the recording location. All single neurons were isolated through this manual cluster cutting process, the MClust toolbox by David Redish in MATLAB (MathWorks). All identified single neurons used for analysis were stable throughout the recording session (∼1 h).
Histology
After completion of the experiment, small marking lesions were made through selected tetrodes that yielded isolated single neurons. Marking lesions were made by passing 10 µA of anodal DC current through the EIB for 10 s. Rats were then perfused, tetrodes were retracted from the brain, and the brain was removed. Brains were cryoprotected in 30% sucrose for 3 d and sectioned at 40 µm on a cryostat. Brain slices were mounted on slides and stained with thionin. Recording locations in each structure throughout the experiment were verified using a light microscope to identify final tetrode depth and by comparing the final depth with tetrode turning records kept during turning sessions (Halverson et al., 2010).
Statistical testing on behavioral and neural data
All statistical tests were performed using custom-written codes in MATLAB (MathWorks) or IBM SPSS Statistics. The statistical significance level was set at 0.05 for all analyses, unless otherwise stated. When repeated-measures ANOVA was conducted, the least significant difference procedure was performed as a post hoc pairwise comparison. Group data are presented as mean ± SE. When appropriate, iterative trimming procedures were applied to dependent measures to prevent outliers from distorting representative values. Samples containing values beyond mean ± 2 SDs were iteratively excluded before taking the means for the subsequent analytic tests.
Analytical design
Trace conditioning sessions were sorted and binned into three different learning phases. Initial training (IT) represented the initial two paired sessions. The learning phase (LRN) represented a midpoint of training for each rat when CR percentage showed a significant increase relative to the beginning of training (see below, Learning during trace eyeblink conditioning). The retention phase (RET) represented the final two sessions of training when rats showed the highest CR percentage. Both single neurons and local field potentials (LFPs) during the learning phases were compared with each other and with the baseline activity acquired during the unpaired session (UP). Neural responses during the 250 ms CS period, two 250 ms periods of the trace interval, and two 250 ms periods of the post-US across the learning phases were quantified and compared. Data from each period were further grouped by trials when rats generated a CR or not (CR vs non-CR trials). In addition to aligning neural activity by the external stimuli (tone CS), we also aligned data to the onset of the cerebellar-driven behavior (CR onset). The neural signals around the CR onset from three target structures across the learning phases were assessed and compared. The same quantification procedures (e.g., window size, binning, or smoothing methods, z-normalization, etc.) were applied to both CS-aligned and CR-onset-aligned versions of analyses.
Single-neuron data analysis
Peristimulus time histograms (PSTHs, 10 ms bins) of single-neuron activity for CB, AM, and ACC were created to investigate neuronal responses during conditioning. Analysis of CS-aligned single-neuron activity was compared with an equal pre-CS baseline period using the Wilcoxon signed-rank matched-pairs test (Halverson et al., 2010; p < 0.05). Each single neuron was classified as CS/trace/US-responsive based on significant differences from baseline activity. For each neuronal response category, z-normalized PSTHs (normalized to pre-CS baseline activity) of responsive single neurons were constructed aligned to both CS and CR onset for each target structure for the UP and during each phase of training. Based on these PSTHs, we calculated the average neuronal response of each target area during different learning phases. Average neuronal responses for each area were compared across the different learning phases. Also, for the CS-aligned PSTHs, the results were further divided into CR and non-CR trials to investigate any learning-related changes in neuronal activity. The same analyses were done with different bin sizes (20 and 50 ms) to verify any observed effects were not because of binning artifacts. Results were consistent across bin sizes, 10 ms bins were used for all analyses. Putative glutamatergic (fast firing, >10 Hz baseline rate) and putative GABAergic (slow firing, <10 Hz baseline rate) CB single neurons were identified with established firing rate criteria to investigate possible differences in response properties during trace eyeblink conditioning (Uusisaari et al., 2007). Average CS- and CR-aligned responses for both types were similar, so both types of putative CB neurons were pooled for analysis. For the ACC, a total of eight putative interneurons (baseline rate >30 Hz) were eliminated from analysis to focus on cortical output neurons.
Spectral analyses
All LFP-related data analyses were performed with customized MATLAB codes as described previously (Del Rio-Bermudez et al., 2017, 2020). Briefly, one LFP channel from each structure per subject was chosen for subsequent analysis. The channel selection was based on (1) tetrode placement within the target loci, (2) signal-to-noise ratio, (3) and the numbers of single neurons isolated on a given tetrode (Bragin et al., 1995). Power spectra were generated using fast Fourier transform after filtering out low-frequency noise (DC ∼2 Hz). For time–frequency power analyses, the signals were convolved using a complex Morlet wavelet. The Morlet wavelet was built as follows: The frequency band of interest was divided into 30 bins, with each bin accommodating the lowest rhythms containing at least two to three cycles per second. Bandpass filtering was done similarly to extract specific frequency bands. Theta rhythms were defined as 4–12 Hz oscillations (Kim et al., 2016). Slow gamma rhythms were defined as 25–50 Hz oscillations (Fig. 1C). Beta and fast gamma rhythms were defined as 15–25 Hz and 65–140 Hz, respectively. Phase information of oscillations was extracted using the Hilbert transformation (Le Van Quyen et al., 2001). Phase-related statistical tests (p < 0.05) were performed using a circular statistic toolbox for MATLAB (Berens, 2009). For statistical tests, all measures were normalized using an equal duration pre-CS period to z transform CS- and CR-aligned LFPs.
LFP–LFP spectral coherence
After convolving the raw signals using a complex Morlet wavelet for each frequency, the spectral coherence was computed as follows:
Granger connectivity
Wiener–Granger connectivity was computed with BSMART toolbox (Cohen, 2014; Cui et al., 2008), as described in the earlier study (Del Rio-Bermudez et al., 2017). Briefly, the prediction errors were computed in a given time, at a specific frequency range (e.g., theta), in each pair of target structures. The computation resulted in two directional values (i.e., A to B and B to A) and were then averaged to simplify interpretation. A matching pretrial baseline period was used to normalize the values to allow for z-transformed Granger scores for statistical testing.
Phase-amplitude coupling
Phase-amplitude coupling between theta (4–12 Hz) and slow gamma (25–50 Hz) oscillations were computed following techniques established from previous studies (Canolty et al., 2006; Jirsa and Müller, 2013; Voytek et al., 2013). Either theta rhythms from each local area were used for cross-frequency coupling. After the LFPs around the target event periods were convolved by complex Morlet wavelet, the theta phase was extracted by Hilbert transformation (Le Van Quyen et al., 2001), and the time window matching slow gamma power was computed. Slow gamma was sorted along the theta phase changes around the events of interest. Results were Z-transformed from equal baseline time bins, and the resulting values were used for statistical tests of population data.
Phase-locking analyses
Raw LFPs from target structures were bandpass filtered to either theta or slow gamma frequency then used for Hilbert transformation (Le Van Quyen et al., 2001). The spike train of a single neuron against simultaneously recorded theta or slow gamma phase was used to generate a spike-phase distribution. If neuronal firing frequently occurred at a specific phase, the resulting distribution would be significantly different from a uniform distribution (Rayleigh's test, p < 0.05). The proportion of the single neurons with a spike-phase distribution that was statistically different from a normal distribution was used as dependent measures. To investigate the extent that neural oscillations encoded external events during both CR and non-CR trials, we computed theta phase-locking values (PLVs) at CS onset. We used 500 ms around CS onset (250 ms prior and posterior) to compute PLVs with a 10 ms sliding window. The amount of PLVs around CS onset was quantified as the slope in each session for each rat. The slopes in both CR and non-CR trials were statistically compared with chance level (0, in which slopes were zero) using one-sample t tests (alpha < 0.05).
Results
Learning during trace eyeblink conditioning
During trace eyeblink conditioning, rats learn to associate a conditioned stimulus (CS, i.e., tone) with an unconditioned stimulus (US, i.e., periorbital shock). After repeated paired presentations of the CS and US, rats acquire an adaptively timed eyelid closure CR before onset of the US. The eyelid CR is part of a synergy of defensive CRs (Heiney et al., 2021). Rats showed a progressive increase in CR percentage with successive training sessions (Fig. 1E). Behavioral data were grouped by the initial two sessions (IT), sessions when rats reached a learning criterion of >15% increase in CRs compared with IT (LRN), and the final two paired sessions (RET). For the grouped data, a repeated-measures ANOVA revealed a main effect of session (F(2,16) = 38.30, p < 0.001, η2 = 0.83), and pairwise post hoc comparisons indicated that the CR percentage increased from IT to LRN and from LRN to RET (p values < 0.004).
Simultaneous recordings in the ACC, AM, and CB and associative learning. A, Recording locations in the right ACC and AM and left CB. Coronal sections showing representative tetrode placements (left) and all accurate placements (right, color coded arrows and dots) in each area. B, Isolating single neurons from multiunit activity in each area. Boundaries are drawn around clusters of points representing the peaks of waveforms in two dimensions (see above, Neuron isolation and criteria). The color-coded and numbered waveforms for each cluster are shown. C, Trace eyeblink conditioning schematic showing the duration of each event indicated by different colored backgrounds; the tone CS (250 ms, gray), free trace interval (500 ms, blue), and shock US (25 ms, pink). Below the conditioning schematic are LFPs from the ACC (green), AM (red), and CB (blue). For each brain area the top trace is the raw LFP for a single conditioning trial. Middle and bottom traces show theta (4–12 Hz) and gamma oscillations (25–50 Hz), respectively. D, Representative regular coherograms (left) and the matching imaginary coherograms (right) aligned to CS onset. To ensure that the coherence measures were not because of artifact effects, volume conduction insensitive imaginary coherence was computed for cross-validation. Very similar patterns of coherence values were observed in both time and frequency domains. E, Top, Overlaid (gray) and average (black) eyelid EMG responses from IT, LRN, and RET phases. Bottom, Average (±SEM) CR percentage for each training session (black); learning curves for individual rats are plotted in gray.
Neuronal firing rate and local field potentials
Learning during trace conditioning requires communication among the AM, ACC, and CB (Hattori et al., 2014; Plakke et al., 2009; Siegel et al., 2015; Weible et al., 2007). Because of this distributed neural network, trace conditioning provides an opportunity to investigate both local information processing and network oscillatory activity. Local information processing represents nodes in a circuit where information about the training stimuli is relayed from the AM and ACC to the cerebellum. Theta and gamma frequency oscillations in the forebrain and cerebellum represent the network state. We investigated how LFPs change across phases of learning. Analysis of both single-neuron activity and LFPs in specific frequency ranges (e.g., theta, 4–12 Hz, and slow gamma, 25–50 Hz) were aligned to CS onset and CR onset. Alignment of LFPs to external stimuli (CS) or to internally driven learned responses (i.e., CRs) allowed the investigation of how neural oscillations that are potentially driven by different mechanisms change across learning phases.
Stimulus- and behavior-aligned theta power increased with learning
Early increase in contingency-related theta power
We observed a substantial increase in theta power during IT (Fig. 2A,B) when rats transitioned from unpaired presentations of the CS and US to paired training (F(3,54) = 7.84, p < 0.001, η2 = 0.30; UP vs IT, p < 0.01). The increased theta power slightly decreased during the LRN phase (IT vs LRN, p < 0.05), but remained elevated above UP levels through RET (p values < 0.05) in all three structures. Importantly, this observed increase in theta power was specific to CR trials. We observed no interaction and no difference in theta power between unpaired sessions and each phase of paired training during non-CR trials (F(3,51) = 1.96, p = 0.13, η2 = 0.10; Fig. 2B). We tested for possible differences in theta power between CR and non-CR trials across learning phases. There was a main effect of CR condition (F(1,16) = 17.44, p < 0.001, η2 = 0.52) and an interaction between the CR condition and learning phase (F(3,48) = 5.46, p < 0.01, η2 = 0.25). Post hoc comparisons indicated there was higher theta power in CR trials relative to non-CR trials during each phase of training (p values < 0.05; Fig. 2B).
Theta and slow gamma power. A, Spectrograms from the ACC, AM, and CB aligned to CS onset showing average power during UP, IT, LRN, and RET. B, Bar graphs (±SEM) representing CS-aligned average z-normalized theta power across training phases for ACC (green), AM (red), and CB (blue). Filled and open bars during paired sessions represent CR trials and non-CR trials, respectively. C, Theta phase resetting during each learning phase. From UP through the end of training (RET), theta phase of all three areas showed significant phase resetting to CS onset. We observed similar CS-related phase resetting during CR (blue) and non-CR (red) trials. This analysis indicates that phase resetting is not learning-related and that rats show similar phase resetting responses to the CS even during unpaired presentations of the tone when no learning has occurred (black/gray). D, Averaged power spectrograms for each area aligned to CR onset. White dotted lines indicate CR onset. E, Bar graphs (±SEM) representing CR-aligned average normalized theta (left) and slow gamma (right) power across phases of training for ACC, AM, and CB. The asterisk indicates significant differences from UP (above plot) or significant difference between CR and non-CR (above bars). Individual data points are plotted in gray.
The theta power results suggest two possibilities; (1) higher theta power levels increase the likelihood of CRs relative to situations where theta power is similar to unpaired levels, or (2) CR production by the cerebellum causes an increase in theta power. To address these two possibilities, we aligned LFPs by CR onset to investigate how the internal drive of the cerebellum might influence theta oscillations. Similar to the results under the CS alignment, we found a significant increase in theta power with learning under the CR alignment (F(3,54) = 20.28, p < 0.001, η2 = 0.53). This change in theta power was because of the increase during IT when the CR percentage was lowest (IT vs all other learning phases, p values < 0.001; Fig. 2D,E). Similar to the CS-aligned results, the increased theta power observed during IT remained above the UP levels throughout each phase of training (p values < 0.05). If the neural drive of the CB to produce CRs was also driving an increase in theta power, we would expect to see a progressive increase in power as the CR percentage increased. However, theta power was not elevated further across training in either the CS- or CR-aligned results (Fig. 2B,E). Additional analysis comparing the CS-aligned and the CR-onset-aligned theta power across the learning phases confirmed this idea. There were significant main effects of learning (F(3,54) = 22.86, p < 0.001, η2 = 0.56) and an interaction (F(3,54) = 6.95, p < 0.001, η2 = 0.28) between CS- and CR-aligned theta power and learning phase. The sharp increase from UP to IT was responsible for most of the effects observed (p values < 0.001). After the initial increase, theta power tended to decrease across training. If CB activity driving CR expression also generates theta activity, the pattern would be the opposite because more CRs were produced in later training sessions. During IT, theta power was higher when CR onset aligned than when CS aligned (p < 0.001), but theta power was not different between the two theta alignments in the other phases of learning (Fig. 2E). Overall, our results suggest that theta power quickly increases after rats switch from unpaired presentations of stimuli to associative learning. Because of the rapid nature of this increase in power and the fact that theta power does not parallel an increase in CR percentage, we hypothesize that this early increase in theta power represents a learning-driven attentional state shift to facilitate or initiate the upcoming learning process. Similar analysis of beta oscillations (data not shown) were inconclusive.
Theta phase resetting did not change as a result of learning
We examined whether neural oscillations showed phase resetting to CS onset during each phase of training. Phase resetting has been used as a reliable tool to examine how ongoing neural oscillations might encode information triggered by an external event (Canavier, 2015; Wilson and Cowan, 1972). Phase resetting is an indication of neural processing of an external stimulus and might not be modulated by learning but rather serves as a neural correlate for attention, which is a basic building block of learning. With this analysis we wanted to verify that any learning-related changes in LFPs or single neurons were not just because of inattentiveness to the CS. This is especially important during non-CR trials, as lack of attention to the CS would be a simple explanation for failure to produce a CR on a given trial. From UP through the end of the training (RET), the theta phase of all three areas showed significant phase resetting to CS onset (p values < 0.05; Fig. 2C). These results indicate that any differences found in power and coherence analyses were because of learning and not because of the inattentiveness of the rat toward the CS.
Learning-related changes in single-neuron responses during the training stimuli and trace interval
We recorded a total of 409 single neurons from the AM, 88 neurons from the CB, and 338 neurons from the ACC from 9 rats (7 male, 2 female). The activity of single neurons was isolated from the multineuron recording using a manual cluster cutting procedure (see above, Materials and Methods). Data from misplaced tetrodes were not used. Figure 1 shows examples of the recording locations and tetrode placement in each area. Neuronal recordings were taken during daily training sessions. Average baseline firing rate for all AM neurons was 13.4 Hz, 45.1 Hz for CB neurons, and 7.4 Hz for ACC neurons. Tables 1 and 2 show summaries of the basic changes in firing rate of single neurons during training and event responsiveness, respectively (Wilcoxon signed-rank tests, alpha = 0.05).
Single-neuron firing properties
Proportion of responsive neurons across learning
Figure 3A shows individual examples of single-neuron activity and averaged activity (Fig. 3B–E) in all three areas with significant increases in spiking during the CS and trace intervals in the different phases of training. Single neurons with a significant increase in firing rate during the 250 ms tone were classified as CS responsive (Fig. 3B–C). The initial response profile for CS-responsive neurons was a phasic increase in firing to CS onset and a lower sustained response during the CS that faded back to baseline in the trace interval. As training progressed from LRN to RET, CS-responsive neurons in the CB also showed a substantial increase in activity during the trace interval. Tone-responsive neurons in the AM and ACC showed a progressively larger response to tone onset and more sustained activity into the trace interval as training progressed to RET. During RET, the sustained response in the trace interval observed in the AM was larger in magnitude than during IT and LRN. In the RET phase, the trace response observed in the AM was also larger than the response observed in the ACC (Fig. 3B). Single neurons with a significant increase in firing rate during either the first or second 250 ms epoch of the trace interval were classified as trace responsive (Fig. 3D,E). The initial response profile for trace-responsive neurons was also a phasic increase in firing to tone onset and a lower sustained response throughout the tone that persisted into the trace interval. The percentage of single neurons showing this persistent response during the trace interval increased across training in all three areas. For trace-responsive neurons, the CB showed the largest learning-related percentage increase in the trace interval across the training phases. Neurons classified as trace responsive in the ACC showed an increase in response magnitude during the trace interval in the LRN and RET phases relative to IT. Figure 3, F and G, shows heat maps of the normalized firing rate of each neuron. These results suggest that trace conditioning leads to more persistent activity in the AM and to a lesser degree in the ACC as training progresses (Fig. 3F).
CS- and trace-responsive neurons. A, Peristimulus histograms of individual CS-responsive single-neuron activity from the ACC (green), AM (red), and CB (blue) during each phase of training, UP, IT, LRN, and RET. B, D, Population average of CS- (B) and trace-responsive (D) single neurons during CR trials (solid lines) and non-CR trials (dashed lines) aligned to CS onset. C, E, Population average of CS- (C) and trace-responsive (E) single neurons aligned to CR onset. F, Heat maps showing the normalized activity from ACC, AM, and CB for each CS- and trace-responsive single neuron during each phase of training (vertical axis). CR trials (top), non-CR trials (bottom). Bottom, Duration of the CS (gray) and trace (blue) intervals. G, Heat maps showing the normalized activity from ACC, AM, and CB for each CS- and trace-responsive single neuron during each phase of training (vertical axis) aligned to CR onset (white dotted line).
We then examined the normalized activity of all CS- and trace-responsive single neurons separately for CR and non-CR trials to assess learning-specific activity (Fig. 3B,D). Throughout training, large differences in firing rates between CR and non-CR trials in the CB was an expected result as neurons in the interpositus nucleus drive expression of the learned behavior (Choi and Moore, 2003; Freeman et al., 2005; Green and Arenos, 2007; Halverson et al., 2010, 2015, 2018, 2022; Heiney et al., 2021; McCormick and Thompson, 1984b; ten Brinke et al., 2017; Fig. 3B,D). As rats produced a larger percentage of CRs, the neuronal activity in CB driving the learned behavior was also present on a larger percentage of trials relative to IT. Larger increases in the average firing rate were also observed during CR trials relative to non-CR trials in trace-responsive single-neuron activity during the LRN phase in the ACC and AM (Fig. 3B,D). This CR-related increase in firing was most pronounced in the AM during RET, suggesting that the large increase in persistent activity observed in the AM is specific to expression of learning. Increased firing rates during CR trials in all three areas in both CS- and trace-responsive single neurons was most evident when rats were performing the most CRs during RET. The clear difference in CR versus non-CR firing was a result of learning as evidenced by different activity profiles throughout the learning process, which progressively developed as then CR percentage increased. Average activity for both CS- and trace-responsive single neurons in all three areas during non-CR trials at each phase of learning was approximately equal at each phase of training and similar to responses observed during UP training when no learning had occurred (Fig. 3B,D). During training, increases in ACC and AM activity related to CRs were slightly delayed compared with those in the CB. Learning-related activity was specific to the expression of CRs in all three areas. In the CB, this CR-related increase drove the CR, whereas in the ACC and AM, the differential firing may have increased the likelihood of a CR on a given trial by providing greater input to the cerebellum related to the CS and trace intervals.
We also examined the activity profiles of CS- and trace-responsive neurons in all three areas across training when the activity was aligned to CR onset (Fig. 3C,E). Figures 3, F and G, show heat maps of the normalized firing rate of CS- and trace-responsive neurons aligned to CR onset. In these neurons we observed increases in CB activity before CR onset earlier in training relative to both the ACC and AM, which is consistent with the cerebellum driving the learned behavior. This learning-related activity in CB is also possibly driving the CR-related increases observed in the ACC and AM. CS- and trace-responsive neurons in the AM, and ACC to a lesser extent, showed a progressive increase before CR onset from LRN to RET. AM trace-responsive neurons matched the timing and magnitude of the CB pre-CR activity during RET (Fig. 3C,E). In contrast to the CB and AM, the learning-related increase in the ACC peaked after CR onset. These results suggest that as the CB drives expression of the learned behavior, forebrain areas start to progressively increase their representation of the CS and trace intervals.
Slow gamma power aligned to the learned response increased with learning
CR-aligned increase in slow gamma power
Specific roles for the involvement of forebrain slow gamma in learning and memory-related paradigms have been proposed in previous studies (Bieri et al., 2014; Bragin et al., 1995; Colgin and Moser, 2010; Fries, 2009; Harris et al., 2003). It is unknown how slow gamma might be involved in communicating between the cerebellum and forebrain during learning. As slow gamma has been shown to be critical for prospective temporal coding of upcoming events (Bieri et al., 2014), we hypothesized that output from the cerebellum to produce the CR might also communicate the precise temporal aspects of the CR to forebrain structures through an increase in slow gamma power.
CS-aligned slow gamma differed from theta in that there were no significant main effects of learning phase and no interaction of CR versus non-CR trials. These results suggest that slow gamma power is not relevant to the CS-US contingency or learning. However, event-related increases in slow gamma power are by nature more difficult to uncover relative to theta because of the faster oscillation frequency of slow gamma (Gliske et al., 2016; López-Cuevas et al., 2013). If slow gamma is related to CR production it would be difficult to detect the short temporal component of slow gamma when LFPs are aligned to CS onset, especially considering that CR onset occurs at different times (relative to CS onset) across trials. Therefore, if slow gamma is more related to output of the CB, which drives the temporal kinematics of the CR, than processing of external stimuli, we would be more likely to detect learning-related increases in slow gamma power when LFPs were aligned to CR onset. Under CR onset alignment (Fig. 2D,E), there was a significant main effect of learning phase (F(3,45) = 4.64, p < 0.01, η2 = 0.24). Slow gamma power during paired training was significantly increased from UP levels (p values < 0.01) and marginally increased in RET (p = 0.06). These results suggest that slow gamma in the ACC and AM is driven by temporally specific cerebellar output, which also drives expression of the learned response (Fig. 2E).
Theta and slow gamma coherence increased with learning
CS-aligned theta coherence increased with the onset of paired training (IT) and was greater on CR trials relative to non-CR trials
The simultaneous nature of our recordings allowed the investigation of how critical brain regions for learning become synchronized across training. We computed LFP–LFP coherence for each learning phase. Spectral power analysis revealed that theta power sharply increased at the onset of paired training, especially during CR trials. Guided by this finding, coherence analyses focused on examining how CS-aligned theta synchronized between brain areas during CR trials across training (Fig. 4). We found main effects of learning (F(3,57) = 4.57, p < 0.01, η2 = 0.19) and regional pair (F(2,19) = 4.12, p < 0.05, η2 = 0.30). Post hoc comparisons indicated that theta coherence was higher than UP levels throughout paired training (UP vs each learning phase, p values < 0.05; Fig. 4C). Similar to theta power, theta coherence did not progressively increase across learning phases. Theta coherence was higher between ACC-AM and CB-ACC than between CB-AM (Fig. 4C; p values < 0.05). The two ACC-related pairs were statistically comparable. These results suggest that via the theta rhythm the ACC might play a larger role in coordinating with the AM and CB during learning. The ACC likely accomplishes this through theta-mediated top-down control (Cavanagh and Frank, 2014; Colgin, 2013; Fries, 2015). During non-CR trials there was a main effect of learning (F(3,60) = 4.04, p < 0.05, η2 = 0.17; Fig. 4A,C). The significant main effect of learning was mainly because of the sharp increase in theta coherence during the transition from UP to IT (p < 0.05). This elevated theta coherence showed divergence by CR condition. Although theta coherence during CR trials remained elevated throughout each phase of training, theta coherence between each pair during non-CR trials decreased to UP levels (Fig. 4C). There were main effects of learning phase (F(3,57) = 5.48, p < 0.01, η2 = 0.22) and an interaction between learning phase and CR condition (F(3,57) = 2.88, p < 0.05, η2 = 0.13). The main effect of CR condition was marginally significant (F(1,19) = 4.23, p = 0.05, η2 = 0.18). Post hoc comparisons showed that theta coherence during each phase of paired training was higher than UP in both CR conditions (p values < 0.05), and coherence during CR trials was higher than non-CR trials in the RET phase (p < 0.05). These results suggest that initially paired training produces large changes in theta power and coherence in cerebellum and forebrain. As training progresses, CS-aligned theta power and coherence remains high during CR trials, whereas both power and coherence are low during non-CR trials. Theta acts as a gate for expression of cerebellar-driven responses that is independent of the amount of learning.
Theta and slow gamma coherence. A, Heat maps showing coherence between each pair of oscillations (4–50 Hz range) from the ACC, AM, and CB aligned to CS onset for individual session examples during UP, IT, LRN, and RET. Dotted white line represents CS duration (start of trace interval). B, Bar graphs (± SEM) representing CS-aligned average z-normalized theta coherence for ACC-AM (light blue), CB-ACC (dark-green), and CB-AM (purple). Filled bars during paired sessions represent CR trials, white bars with colored outline represent non-CR trials. The asterisk indicates significant differences from UP. C, Heat maps of regular coherograms and imaginary coherograms confirm the coherent results were not due to artifacts such as volume conduction. D, Individual session heat maps showing coherence for each pair of oscillations for each area aligned to CR onset (white dotted line). E, Bar graphs (± SEM) representing CR-aligned average normalized theta (left) and slow gamma (right) coherence across phases of training for ACC-AM, CB-ACC and CB-AM. The asterisk indicates significant differences from UP levels. Individual data points are plotted in gray. sGamma, Slow gamma.
In addition to our observation that theta power increases rapidly during the transition from UP to IT, the finding that theta coherence also shows the same pattern suggests this synchrony is part of a state shift. The network for associative learning might be using theta frequency synchrony to functionally connect distant brain regions (Wingerden et al., 2010). Previous studies using various cognitive tasks in humans suggest that theta reflects general attention for the task rather than contributing toward a specific cognitive function (Caplan et al., 2001; Cavanagh and Frank, 2014; Kahana et al., 2001; Raghavachari et al., 2001). To test this idea, we aligned LFPs to CR onset and investigated theta coherence between areas. If the increased theta synchrony between brain areas during paired training was related to the neural drive to generate CRs, the CR-aligned theta coherence should be greater than the CS aligned. Main effects of learning phase, regional pairs, and the interaction were not statistically significant (Fig. 4B,D). The absence of a learning effect in CR-aligned results indicates that the differential pattern of theta coherence was not because of the increase in cerebellar output to drive CR expression. Similar analyses of coherence for beta oscillations for each area were not significant (data not shown).
CR-aligned slow gamma coherence increased with the onset of paired training (IT)
Spectral analysis of slow gamma aligned to CR onset revealed a significant increase in power after the transition from UP to IT (Fig. 2D). This result is consistent with previous findings that gamma oscillations are modulated by direct decisions made during a task rather than general task difficulty (Caplan et al., 2001; Kahana et al., 2001; Raghavachari et al., 2001). Slow gamma may be playing a similar role during associative learning, with fast neural oscillations communicating precise timing of the behavior (CR) rather than the parameters of the task. Multiregional neural synchrony of slow gamma was examined to test this hypothesis. When LFPs were aligned to the CS, none of the factors in the CR and non-CR conditions was significant. However, when LFPs were aligned to CR onset there was a significant main effect of learning phase (F(3,66) = 6.07, p < 0.01, η2 = 0.22; Fig. 4B,D). Post hoc tests revealed that CR-aligned slow gamma coherence was different between UP and all other learning phases (p values < 0.05). These results suggest that presentations of the CS, even during CR trials, are not sufficient to drive synchrony in slow gamma activity. However, the neural output of the cerebellum to drive CR expression is associated with an increase in slow gamma synchrony between brain areas. The findings that presentations of an external stimulus do not drive increases in slow gamma power or coherence, but temporal alignment of a cerebellar-driven behavior is consistently related to large increases in slow gamma power and coherence, provide converging evidence indicating that the cerebellum is simultaneously driving both behavior and increases in slow gamma.
Network switch from encoding to retrieval indicated by a shift in cross-frequency phase-amplitude coupling
The current results suggest that theta and slow gamma are important neural signals during learning. Previous studies proposed that local activity in forebrain systems might promote effective downstream information processing through generation of the relatively slow theta rhythm. Specifically, theta oscillations would maintain long synaptic delays to enable distributed brain regions to be more coordinated to an external event (Colgin, 2013; Harris et al., 2003; Jones and Wilson, 2005; Paz et al., 2008; Stujenske et al., 2014). Slow gamma, which is a relatively faster oscillation than theta, would then ride on theta to provide additional information for downstream processing by distant cell assemblies (Buzsáki and Wang, 2012; Harris et al., 2003). Gamma oscillations could also contribute to forming transient links among cell assemblies (Colgin et al., 2009). Given the relationship between theta and gamma, the phases of theta and slow gamma power should show cross-frequency synchrony across learning, especially when oscillations are aligned to CR onset. With CR alignment, we investigated how slow gamma power was distributed on valleys and peaks of theta.
CR-aligned local theta phase to slow gamma power coupling switched from encoding to retrieval across learning
LFPs were aligned to CR onset, and slow gamma power was then sorted relative to the phase of theta (Fig. 5A,B). First, we examined how slow gamma power from each area was distributed according to each area's own local theta phase. During UP, slow gamma power was uniformly distributed across theta, forming a flat distribution. However, once paired training began (IT), slow gamma clustered on the valley of local theta in each area. During LRN, both the AM and CB maintained a peak of slow gamma power on the valley of local theta. However, the magnitude of phase-locking was lower, and the peak in each area was shifted to the earlier local theta phase. In the ACC, the distribution during LRN was flat, indicating no phase-locking. Once the rats entered into the RET phase, the slow gamma power distribution with local theta became bimodal in all three areas. We quantified these observations by computing the slow gamma power ratio by the valley (170–190°) and peak (350–10°) of theta. If slow gamma power was equally distributed between valleys and peaks, this ratio would be 0.5. However, greater slow gamma concentration around theta valley would result in a ratio >0.5, but more concentration around theta peak would be <0.5. Results showed that slow gamma power distributions related to the local theta of each region were different across training phases. There was a significant main effect of learning phase (F(3,72) = 2.90, p < 0.05, η2 = 0.11). Post hoc comparisons showed the ratio values during RET were significantly lower than both IT and LRN (p values <0.05). The increase from UP to IT was marginally significant (p = 0.08). During UP, the ratio values were ∼0.5, indicating that slow gamma power was similarly distributed along local theta phases. However, the ratio values increased after the transition from UP to IT and then decreased below 0.5 during RET. Previous studies suggest that encoding and retrieval occur via different theta phases (Buzsáki and Wang, 2012; Harris et al., 2003; Osipova et al., 2006; Wingerden et al., 2010). The current results are consistent with earlier studies by showing differential slow gamma power distributions on valleys early in learning, suggesting encoding of information and a transition of slow gamma onto the peaks of theta, suggesting retrieval of information after learning has occurred.
A–D, Local (A–B) and CB (C–D) theta phase–slow gamma power coupling. A, C, Population averages (±SEM) for the magnitude of slow gamma power from ACC (green), AM (red), and CB (blue) coupled to theta phase during each phase of training. No slow gamma local or cerebellar theta phase coupling was observed during UP. During the transition from UP to IT slow gamma was more locked to the valley of theta (180°). As training progressed, slow gamma became more phase-locked to the theta peak (360°). B, D, Theta valley-to-peak slow gamma power ratio for three areas for each phase of training. The asterisk indicates significant differences. Individual data points are plotted in gray.
CR-aligned cerebellar theta phase to slow gamma power coupling switched from encoding to retrieval across learning
Previous studies proposed different functional roles of CB output with other brain structures. Two different but not mutually exclusive models, inverse and forward models, have proposed specific functions for cerebellar output to either drive motor commands (inverse model) or provide a prediction about kinematic properties of the upcoming movement (forward model; Ebner et al., 2011; Green and Angelaki, 2010; Imamizu and Kawato, 2009; Lisberger, 2009; Medina, 2011). To efficiently learn the current association, the cerebellar activity drives not only CR expression (inverse model) but also provides response information to the forebrain to precisely adjust its timing (forward model). The CR-aligned increase in slow gamma power and synchrony is a likely mechanism for the CB to communicate the properties of the CR to forebrain areas. As cerebellar computation heavily relies on theta oscillations for learning and behavioral expression of CRs (Casula et al., 2016; Dave et al., 2020; Hoffmann and Berry, 2009), we hypothesized that CB is communicating CR timing to the forebrain through temporally specific oscillations. We would expect slow gamma in each region to be concentrated on the valley of CB theta during IT and shift to the peak of CB theta during RET. This would be consistent with the forward model of CB communication, in that early in training (IT) the CB is communicating specific temporal information with the forebrain. After sufficient repetitions (LRN, RET) forebrain communication would be enriched with this temporal information learned from the CB. Our results show that local activity of the AM, CB, and ACC tended to phase-lock on cerebellar theta valleys more than peaks during IT. As the training progressed, slow gamma power became more concentrated around the peaks rather than valleys, especially during RET (Fig. 5C,D). Statistical tests revealed a significant main effect of learning phase (F(3,72) = 6.42, p < 0.001, η2 = 0.21). Post hoc comparisons revealed significant pairwise differences for all three areas of interest; slow gamma power to CB theta phase showed an equal distribution during UP that shifted toward the CB theta valley during IT (p < 0.05). The valley-concentrated slow gamma power further shifted toward CB theta peaks at the end of training (RET vs all other learning phases, p values < 0.05). These results support the hypothesis that cerebellar network activity provides feedback to the ACC and AM. This feedback is focused on encoding temporal information during the IT and then shifts to communicating temporal information learned from cerebellum during RET. This early training encoding phase that transitions to retrieval could be a networkwide phenomenon as it was observed in a similar way with respect to both local and cerebellar theta.
Progressive increase in CR-aligned cerebellar slow gamma Granger directionality to forebrain across learning
Cerebellar activity drives CR expression and increases in slow gamma power and coherence. Slow gamma from AM and ACC were collapsed to investigate the directionality of outbound/inbound slow gamma communication with cerebellum (Fig. 6). There was a significant main effect of learning phase (F(3,24) = 3.400, p < 0.034, η2 = 0.298) for CB inbound CS-aligned Granger prediction (Fig. 6B). This was because of a large increase in CB inbound Granger prediction during the transition from UP to IT, which remained elevated throughout training. When oscillations were aligned to CR onset, we observed a large increase in both CB inbound and outbound Granger prediction. For CR-aligned CB inbound slow gamma Granger prediction there were main effects of alignment (CS vs CR aligned; F(1,8) = 6.160, p < 0.038, η2 = 0.435) and learning phase (F(3,24) = 3.438, p < 0.033, η2 = 0.301). For CR-aligned CB outbound slow gamma Granger prediction there were also main effects of alignment (CS vs CR aligned; F(1,8) = 6.398, p < 0.035, η2 = 0.444) and learning phase (F(3,24) = 4.263, p < 0.015, η2 = 0.348). For CR-aligned CB inbound slow gamma Granger prediction the results were because of an increase during LRN, which decreased in RET (Fig. 6B). In contrast, for CR-aligned CB outbound slow gamma Granger prediction we observed a linear increase in CB communication to forebrain as training progressed similar to the increase in CR percentage (Fig. 6A). We did not observe this linear increase in any other LFP analyses. This result is also similar to the time course of single-neuron CR-specific activity across training phases. These results suggest that the increase in CR-aligned slow gamma power and coherence are being communicated from the CB to the forebrain progressively as learning occurs.
Granger directionality. A, B, CS- and CR-aligned (green and blue, respectively) Granger prediction for outbound (A) and inbound (B) slow gamma between cerebellum and forebrain across training.
We also investigated CS-aligned theta and slow gamma Granger prediction between each area across the different learning phases. We found no significant Granger prediction in CS-aligned theta; however, there was a main effect of learning phase for slow gamma (F(1,8) = 6.340, p < 0.050, η2 = 0.444), and the interaction between learning phase and directionality was marginally significant (F(1,8) = 4.960, p < 0.07, η2 = 0.380). These results suggest that CS-aligned slow gamma communication from ACC to AM increases more than the opposite direction after the switch from UP to IT. This slow gamma communication shows a progressive increase across learning phases and could be facilitated by the increase in CR-aligned slow gamma communication from CB to both forebrain structures.
Increases in phase-locking indicate interactions between local and network activity at the onset of learning
Both coherence and cross-frequency coupling results indicate that network activity from the cerebellum and forebrain became more synchronized in a way that suggests encoding during IT that progressively switches to retrieval during RET. Our next question was how local neuronal activity in each area synchronized with slow and fast oscillations across training. Among the isolated single neurons (N = 835), the CS- and trace-responsive neurons (N = 297) were used to compute the PLV (Nokia et al., 2012; Nokia and Wikgren, 2014) against theta and slow gamma (Fig. 7).
Theta and gamma spike phase-locking (PVL). A, Individual trial examples of how single-neuron spikes (black filled squares) correspond to theta (left) and slow gamma (right) phase. sGamma, Slow gamma. TR marks the onset of the trace interval. B, PLV values (±SEM) for ACC, AM, and CB theta and AM slow gamma for CR (left) and non-CR (right) trials for each phase of training. Asterisk indicates significant differences.
Theta phase-locking with neuronal activity increased with the onset of paired training (IT)
Theta from each area was used to compute PLVs with CS- and CR-aligned local single-neuron spiking during CR and non-CR trials (Fig. 7). For ACC, AM, and CB theta, PLVs with all regional spikes increased after the transition from UP to IT in both CR and non-CR trials. There was a significant main effect of learning phase (F values > 5.72, p values < 0.01, η2 values > 0.42). Post hoc tests indicated that PLVs from paired sessions were significantly higher than UP levels in general, especially during the transition from UP to IT (p values < 0.05). PLVs during non-CR trials in the LRN phase for CB theta and PLVs during RET with AM theta showed significant differences compared with UP values (p values < 0.05). There were no significant PLV main effects or interactions when theta network activity and local single-neuron activity were aligned to CR onset. Spike-theta PLV analyses showed that local spike activity in the AM, CB, and ACC became more synchronized with network activity after the transition from UP to IT (Fig. 7). The increases were similar between CR and non-CR trials suggesting that spike-theta synchronization may reflect general attention toward the CS-US contingency shift rather than specific learning-related changes (Caplan et al., 2001; Cavanagh and Frank, 2014; Kahana et al., 2001; Raghavachari et al., 2001).
Slow gamma phase-locking with neuronal activity increased with the onset of paired training (IT)
The same analyses and statistical tests were conducted with slow gamma oscillations from the ACC, AM, and CB regional spike activity (Fig. 7). Slow gamma PLVs from both the ACC and CB, regardless of CR or non-CR trials, were stationary across learning phases. The AM slow gamma PLV, however, showed an interaction between learning phase and region during CR trials (F(6,24) = 2.93, p < 0.05, η2 = 0.42). A.M. slow gamma PLVs with AM local neuronal activity showed an increase during the transition from UP to IT. This elevated PLV value decreased throughout LRN and RET (post hoc comparisons, UP vs IT and LRN, p values < 0.05; UP vs RET, p = 0.17; IT and LRN vs RET, p values < 0.05). A.M. slow gamma PLVs with AM activity were significantly higher than AM slow gamma PLVs with CB spikes during the LRN phase (p < 0.05). These results indicate that fast oscillations in AM become more synchronized with local AM spike activity only after the contingency between CS and US is established in IT. Overall, PLV results suggest that local single-neuronal spike activity becomes more synchronized with subpopulation network activity, and the single-neuron subpopulation level synchronization was driven by the CS rather than by the internal signal driving the CR. Learning-related changes in PLVs related to theta were more prevalent relative to slow gamma PLVs. As the increase in PLVs during theta were not specific to CR or non-CR trials, this likely indicates that neurons in general are more likely to fire when theta is present after the CS-US contingency is established.
Discussion
In the current study we used trace eyeblink conditioning to investigate how local (single neurons) and network (oscillations) processing in the forebrain (AM and ACC) and cerebellum interact during the process of transforming stimulus information into an adaptively timed motor response. Network activity in the CB, ACC, and AM during trace eyeblink conditioning suggests how brain states might dictate expression of learning and how the specific engagement of different oscillation frequencies depends on what the animal is experiencing versus what the animal is doing. During the transition from UP to IT we observed a large increase in CS-aligned theta power in the AM and ACC selectively during CR trials, which persisted throughout training (Fig. 2). This increase in CR-related theta power was subsequently observed in the CB during the LRN phase. In general, there was greater theta power during CR trials relative to non-CR trials in each phase of training. In contrast, we did not observe significant differences in CS-aligned slow gamma power between CR and non-CR trials. The rapid changes in network activity contrast with local processing in the CB, which showed CR-specific activity before the forebrain, whereas network processing showed changes in forebrain theta power before changes in CB theta (Figs. 2B, 3B,D). These results suggest that the transition from UP to IT produces an immediate state change to signal the CS-US contingency and to facilitate the development of CR-specific activity in the cerebellum. The coherence and phase synchrony results suggest that the ACC, AM, and CB communicate CS-US contingency information via theta oscillations at both network (Fig. 4C) and local levels (Fig. 7B). The CR-trial selective increase in forebrain theta power at CS onset (what the animal is experiencing; Fig. 2B) indicates that all three areas engage together in processing and signaling the CS-US contingency, which, in turn, makes expression of CRs more likely.
The increase in in theta and slow gamma power of all three structures after CR onset (what the animal is doing; Fig. 2E) suggests that the internal cerebellar drive to produce the learned behavior may train itself and forebrain areas after processing the CS-US contingency. It is the first evidence, to our knowledge, that the CB is the source of the dramatic increase in slow gamma power/communication observed in multiple brain areas during expression of a cerebellar-dependent behavior. This evidence comes from the difference between analyses aligning oscillations to what the animal is experiencing (tone CS) and what the animal is doing (CR; adaptively timed eyelid closure). We only observed an increase in slow gamma power and synchrony when oscillations were aligned to expression of a learned behavior that is driven by output from CB (Figs. 2E, 4D). Importantly, we did not observe an increase in slow gamma power when oscillations were aligned to external stimuli. Future studies examining slow gamma while optogenetically manipulating cerebellar output during expression of a learned movement that requires precise timing (e.g., eyeblink conditioning) are needed to confirm that the CB is the source of the slow gamma signal.
Local processing of individual neurons in the CB and forebrain structures showed dynamic learning-related changes at different times across training. The CR-related increases in cerebellar firing occurred before onset of the behavioral response, which is consistent with the CB providing the motor drive for the adaptive timing (Choi and Moore, 2003; Halverson et al., 2010, 2015, 2018; Heiney et al., 2021; McCormick and Thompson, 1984a). The CR-related increases in firing observed in the AM and ACC occurred later in training relative to the CB, with a large difference in activity on CR and non-CR trials during RET. Importantly, activity during non-CR trials in all three structures, during each phase of training, did not differ from firing rates during UP, when no learning had occurred. Changes in local activity were similar to changes in network activity (theta power/theta synchrony); during CR trials we observed large increases in stimulus-aligned theta power and synchrony but not during non-CR trials. As we only observed increases in local firing and network theta power/synchrony on CR trials the implication is that theta either gates and/or facilitates learning-related changes in local activity at each stage of training. As the CR percentage increased there were further learning-related changes in local activity in AM and ACC. These learning-related changes in local activity during the trace interval in the forebrain were also similar to the progressive increase in slow gamma communication from cerebellum to forebrain across training (Figs. 3B,D, 6A). This result suggests that repeated slow gamma communication from the cerebellum to the forebrain about cerebellar-driven behavior could gate/facilitate local learning-related changes in forebrain activity, providing enrichment of the temporal aspects of behavior in forebrain neurons.
The effect of the internal drive to produce CRs, and potentially facilitate network communication, is especially evident in the dramatic increase in CR-aligned slow gamma power observed in all three structures in each phase of training. We found CS-aligned increases in theta power that were further facilitated when LFPs were aligned to CR onset (Fig. 2). We also found an immediate increase in theta coherence between AM and ACC and between CB and both forebrain structures during the transition from UP to IT. As rats showed more CRs during the learning and retention phases, theta coherence progressively differentiated, showing elevated synchrony during CR trials (Fig. 4). Aligning the LFPs to CR onset made theta coherence fall to UP levels and eliminated the dynamic pattern of CS-aligned theta coherence across phases of training. The results may indicate the theta-based functional synchrony among areas are prerequisites, not determinants, for the CRs once CS-US contingency changes develop (Fig. 4C). Afterward, theta coherence among pairs is modulated by learning and becomes more closely tied to later adaptively timed CRs. When slow gamma is temporally aligned with the internally driven CR onset we found a dramatic increase in slow gamma power in each phase of training. The absence of CS-aligned slow gamma power/coherence changes across training phases indicates that different mechanisms are controlling communication in the slow gamma frequency compared with theta (Figs. 2, 4). These results suggest that forebrain structures are the primary driver of increases in theta power to an external stimulus when animals transition from unpaired to paired training, and accompanying internal cerebellar output for CRs can further facilitate theta. The opposite is true for slow gamma; external stimuli are not sufficient to drive increases in slow gamma power, whereas the internal cerebellar drive reveals a dramatic increase in slow gamma power in the CB and forebrain. Other cross-regional analyses support the idea that the forebrain and CB have bidirectional communication via theta and slow gamma. Specifically, cross-frequency coupling results suggest a possible mechanism of how communication occurs from learning to retrieval. The gradual shifts of slow gamma power concentration on the theta valley toward the theta peaks across training indicate how encoding and retrieval of the CS-US contingency occur in the cerebellar–forebrain network (Fig. 5).
It is well established that theta has a strong influence on both general behavioral states (attention and arousal) and neuronal plasticity (e.g., high theta facilitates LTP; Albers et al., 2013; Bazelot et al., 2015; Hamilton et al., 2020; Kahana et al., 2001; Orr et al., 2001; Raghavachari et al., 2001; Rizzuto et al., 2006; Voloh et al., 2015). Gamma oscillations have been associated with theta phase and specific goal-oriented behaviors and are also thought to facilitate neuronal plasticity (high gamma might gate spike-timing-dependent plasticity; Lee et al., 2009). Moreover, recent evidence suggests that the CB is important for coherence in different frequency oscillations among forebrain structures (McAfee et al., 2019; Liu et al., 2022). Our results provide the basis for an integrated hypothesis for how external cues have an impact on local single-neuron processing and theta network oscillations early during initial training to facilitate the induction of learning-related changes in local activity and expression of learning (Fig. 8). We hypothesize that the initial increases observed in network theta power, driven by the forebrain to signal the contingency between the CS and US, facilitates the induction of synaptic plasticity (and/or excitability) in both the forebrain and CB. Once local processing in the CB starts to consistently provide the adaptively timed neural drive for CR expression, this temporal information is repeatedly communicated via network transmission to the forebrain through internally driven slow gamma oscillations. Because of the relatively faster frequency relative to theta, gamma oscillations provide an ideal mechanism for the cerebellum to communicate precise temporal information with the forebrain. We further hypothesize that the increase in local CR-specific activity in the forebrain during LRN and RET is the result of repeated adaptively timed slow gamma signaling from the cerebellum. This local (single neuron) increase in firing is selective for CR trials, which strongly suggests the CB is the source of the CR-aligned slow gamma signal to forebrain and resulting local single-neuron plasticity later in training in the AM and ACC. We propose that this cerebellar–forebrain network functions in much the same way during any behavioral or cognitive task that requires computation of external stimuli to produce properly timed behavior.
Hypothesized mechanisms underlying local and network plasticity in the forebrain (ACC, green; AM, red) and CB (blue). Brain schematics show the direction of primary communication, and numbers indicate learning-phase-specific sites of plasticity. Communication between CB and AM/ACC is multisynaptic. For each learning phase the activity of an ideal single neuron from each area (color coded) along with LFPs and eyelid EMG (black) are shown. During IT (left column) we observed an increase in CS-aligned theta power and coherence (gray dotted line) in ACC and AM, first, that serves as an attentional state shift to signal the contingency between CS and US and promotes the induction of CB plasticity. In the LRN phase (middle column) CS-aligned theta power and coherence remain elevated to promote induction of plasticity in CB neurons to drive adaptive CR timing and temporally specific slow gamma back to forebrain (dotted blue line in brain) Second, when CRs are present (gray dotted line during the Trace interval in LRN and RET) there is a large increase in CR-aligned slow gamma power (LFP, bottom, during LRN and RET) and coherence. During RET (right column), we observed similar levels of CS-aligned theta power and coherence and an increase in CR-aligned slow gamma output and coherence from CB to forebrain. This temporally specific slow gamma acts to enrich single neurons in forebrain with temporal information from the CB. Third, this enriched temporal information is then communicated back to the CB to facilitate computations that make CRs more likely on subsequent trials. Non-CR trials are associated with a lack of these local and network dynamics.
Footnotes
This work was supported by National Institutes of Health–National Institute of Neurological Disorders and Stroke Grant NS088567 (J.H.F.) and Kyungpook National University Research Fund, 2022 (J.K.). We thank Drs. Isabel Muzzio and Marc Normandin for feedback regarding the data analyses and Dr. Yong-Seok Lee for comments on the manuscript.
The authors declare no competing financial interests.
- Correspondence should be addressed to Hunter E. Halverson at hunter-halverson{at}uiowa.edu
