Delta Rhythm Orchestrates the Neural Activity Underlying the Resting State BOLD Signal via Phase–amplitude Coupling

Abstract

Spontaneous ongoing neuronal activity is a prominent feature of the mammalian brain. Temporal and spatial patterns of such ongoing activity have been exploited to examine large-scale brain network organization and function. However, the neurophysiological basis of this spontaneous brain activity as detected by resting-state functional Magnetic Resonance Imaging (fMRI) remains poorly understood. To this end, multi-site local field potentials (LFP) and blood oxygenation level-dependent (BOLD) fMRI were simultaneously recorded in the rat striatum along with local pharmacological manipulation of striatal activity. Results demonstrate that delta (δ) band LFP power negatively, while beta (β) and gamma (γ) band LFPs positively correlated with BOLD fluctuation. Furthermore, there was strong cross-frequency phase–amplitude coupling (PAC), with the phase of δ LFPs significantly modulating the amplitude of the high frequency signal. Enhancing dopaminergic neuronal activity significantly reduced ventral striatal functional connectivity, δ LFP–BOLD correlation, and the PAC effect. These data suggest that different frequency bands of the LFP contribute distinctively to BOLD spontaneous fluctuation and that PAC is the organizing mechanism through which low frequency LFPs orchestrate neural activity that underlies resting state functional connectivity.

Introduction

Resting state functional Magnetic Resonance Imaging (fMRI) has evolved into a powerful tool to study systems-level brain functions. Dysregulation in resting state brain networks has been implicated in a number of neuropsychiatric and neurodegenerative disorders. Furthermore, with noninvasive brain stimulation (e.g., transcranial magnetic stimulation, TMS) emerging as a potential tool to treat a wide range of psychiatric and neurological diseases, functional connectivity has been proposed to serve as a biomarker to classify patients and to predict TMS outcomes in depression (Drysdale et al. 2017). However, such clinical potential may be limited by the still poor understanding of the physiological basis of resting state functional connectivity (rsFC), underscoring a critical need for preclinical models to investigate this phenomenon.

A notable, but poorly understood feature in rsFC networks, is their relatively widespread spatial localization, which does not center on specific, fine anatomical structures as often observed in the evoked fMRI response (Lu et al. 2004; Fukuda et al. 2006). For example, the human default mode network (DMN) encompasses a number of cortical association and limbic regions, including medial prefrontal, posterior cingulate, inferior parietal and temporal regions (Raichle et al. 2001). Furthermore, although most rsFC networks appear to be grounded by known anatomical connections, some prominent networks, such as the visual network covering bilateral primary visual cortex, are known to have very weak monosynaptic connections (Van Essen et al. 1982; Vincent et al. 2007). These observations raise a fundamental question about the physiological basis of rsFC: What are the neural mechanisms that mediate the long-range and spatially disparate rsFC brain networks, especially in those absent of direct monosynaptic anatomical connectivity?

Several studies have investigated the neurophysiological correlates of spontaneous BOLD fluctuations in humans (Laufs et al. 2003; He et al. 2008; Nir et al. 2008; Chang et al. 2013; Hiltunen et al. 2014), non-human primates (Shmuel and Leopold 2008; Schölvinck et al. 2010; Magri et al. 2012; Hutchison et al. 2015) and rodents (Lu et al. 2007;2016; Pan et al. 2011, 2013). In two independent studies where electrophysiology and fMRI were recorded in separate sessions (Lu et al. 2007; He et al. 2008), we and others have reported that in the somatosensory cortex, where strong rsFC is detected, strong synchrony in the low frequency LFP signal was also consistently detected, while synchrony in the high frequency (γ) LFP was only present in certain brain states (He et al. 2008). However, given the nature of “spontaneous” and “ongoing”, the resting state fMRI signal, by definition, never repeats itself in space and time. The above two studies, although shedding light on the neurophysiological basis of resting state fMRI signal, are intrinsically limited.

To that end, it would be ideal to simultaneously record the fMRI and LFP signal. In one such study with awake monkeys, Schölvinck et al. (2010) reported that the spontaneous fluctuations in broadband LFPs measured from a single cortical site exhibited widespread correlations with fMRI signals over nearly the entire cerebral cortex, suggesting that the global component of BOLD fluctuations is coupled with neuronal activity. This study, however, cannot explain the regional specificity as observed in rsFC brain networks (Biswal et al. 2010; Smith et al. 2013). In a subsequent study involving concurrent fMRI and LFP recordings in anesthetized rats, Pan and colleagues (Pan et al. 2013) found strong correlation between ultraslow LFP (<0.5 Hz) and resting state BOLD signal. Human studies combining fMRI with scalp electroencephalography (EEG), or magnetoencephalography (MEG) reported correlation in fluctuations between BOLD and broadband LFP signals (Laufs 2008; Brookes et al. 2011). However, EEG and MEG typically suffer from poor spatial resolution and relatively imprecise localization of the electromagnetic field patterns associated with neural current flow, which is particularly problematic in resting state MRI, since the detected ongoing activity is weak, and precludes signal averaging (Deligianni et al. 2014).

Taken together, these studies suggest a complex picture regarding the neural basis of spontaneous BOLD fluctuations. Critically, none of the studies cited above employed an experimental procedure to parametrically modulate brain activity—an important approach to minimize spurious correlations between electrophysiological recording and fMRI, given that these two are dramatically different readouts of brain function. Furthermore, no unified neurophysiological theory that explains the observations described above have been proposed.

Brain oscillatory electrical activity does not occur in isolation, instead the amplitude of high frequency oscillations is often coupled to the phase of the low frequency (δ,θ) oscillations, forming the so-called cross-frequency phase–amplitude coupling (PAC) (Buzsáki and Draguhn 2004; Schroeder and Lakatos 2009). Low frequency oscillations, in particular, the δ oscillations, have been traditionally considered to index sleep stages (Steriade and Amzica 1998). Recent data demonstrate that such low frequency oscillations reflect rhythmic shift in the excitability of local neuronal ensembles. Indeed, hierarchical control of neuronal excitability by cross-frequency PAC has been suggested (Lakatos et al. 2005), with the phase of δ oscillations modulating the amplitude of θ oscillation, and θ phase modulating γ amplitude, leading to the proposition that γ is the slave to, not the master of, lower frequency activity (Schroeder and Lakatos 2009). In another study involving simultaneous surface EEG, intracortical LFP and multiunit activity (MUA) measurement (Whittingstall and Logothetis 2009), EEG power in single trials within the γ band (30–100 Hz) and the δ band (2–4 Hz) phase are significant predictors of the MUA response. MUA response was strongest only when increases in EEG γ power occurred during the negative-going phase of the delta wave. These data together suggest that the cross-frequency PAC effect may be a fundamental mechanism of brain activity. Furthermore, in a recent study employing simultaneous intracranial EEG–fMRI in human subjects performing a finger tapping task (Murta et al. 2017), the PAC effect between β and γ signals explained a significant amount of variance in the BOLD signal in addition to the α, β, and γ band power, emphasizing the role of PAC in neurovascular coupling.

From a cellular physiology perspective, different frequency components in the LFP reflect different aspects of local neuronal activity, with the γ band LFP predominately generated by synchronous firing and hyperpolarization of different neuronal ensembles (Belluscio et al. 2012; Buzsáki et al. 2012), while the low frequency bands reflect subthreshold synaptic processes including receptor-mediated excitatory and inhibitory currents and the maintenance of transmembrane electrochemical gradients. Since such processes are metabolically expensive (Attwell and Laughlin 2001), we hypothesize that different LFP frequency components and the cross-frequency PAC effect that orchestrates these frequency components should have hemodynamic fingerprints as reflected by spontaneous fluctuations in the fMRI signal. In the present study, we attempt to address this question by employing concurrent intracortical local field potential recordings and fMRI in a preclinical model that takes advantage of well-known striatal physiology.

The striatum is a key structure involved in adaptive learning and in linking motivation to behavior. It is the major projection target of mesencephalic dopamine neurons and also receives key glutamatergic afferents from several brain regions, including the medial prefrontal cortex, hippocampus, and amygdala (see Fig. 1) (Voorn et al. 2004). Previous studies have demonstrated that medium spiny neurons (MSNs), which constitute over 90% of striatal neurons (Gerfen 1988), exhibit prominent UP-DOWN two-state fluctuations in membrane potential (Leung and Yim 1993; Stern et al. 1998; Goto and O’Donnell 2001). Striatal MSNs fire preferentially in the UP state, and voltage-dependent potassium channels appear critical in regulating these two-state fluctuations (Wilson and Kawaguchi 1996). By modulating hippocampal glutamatergic input to the ventral striatum via microinjection of lidocaine (a voltage-gated Na⁺ channel blocker), O’Donnell and Grace were able to modulate MSN cell excitability, resting membrane potential and LFP frequency components (O’Donnell and Grace 1995). Thus although LFP of different frequency components are coupled through cross-frequency PAC effect, one could disentangle this relationship in well-defined anatomical pathways through pharmacological approaches. Motivated by this study, we developed a concurrent fMRI-electrophysiological recording technique to perform chronic repetitive recordings with microelectrode arrays covering the entire striatum from its dorsal lateral to the ventral medial domains. In addition, by agonizing the AMPA (α-amino-3-hydroxy-5-methyl-4-isoxazolepropionic acid) receptors within the ventral tegmental area (VTA) micro-circuitry, we systematically modulated dopamine release and the neuronal activity in the striatum (the major projection target of VTA dopamine neurons). We conclude that LFPs of different frequency bands contribute distinctively and differentially to the observed BOLD fluctuations, and that cross-frequency phase–amplitude coupling (PAC) is the organizing mechanism through which low frequency LFPs orchestrate neural activity that underlies the BOLD rsFC.

Figure 1.

Overview of experimental design and hardware setup. Simultaneous fMRI and LFP recording was performed in rats under both baseline condition (pre-AMPA) and following pharmacological modulation (post-AMPA). (A and B) MRI-compatible linear electrode array (16 contacts) was implanted in rat striatum covering the dorsolateral and ventrolateral domains. AMPA was injected into the ventral tegmental area (VTA) via the guide cannula to enhance the activity of VTA dopaminergic neurons, which project primarily to the ventral striatum, modulating striatal LFP signal. (C) Hardware setup for microinjection and electrophysiological recording inside a 9.4T MRI scanner. (D) Experimental timeline. Animals were implanted with microelectrode array and guide cannula. Longitudinal imaging studies were performed after at least 7 days of surgical recovery. AMPA, α-amino-3-hydroxy-5-methyl-4-isoxazolepropionic acid. Note, the amplifier for the 15th channel was consistently dead across all recordings in our recording system; data from this channel was ignored.

Materials and Methods

Overview of Experimental Design

Figure 1 illustrates the overall experimental design and the hardware/software setup for simultaneous fMRI and intracranial multichannel electrophysiological recording. Rats were implanted with an MRI compatible linear microelectrode array that covered the entire dorsolateral to ventromedial striatum (Fig. 1A,B). A guide cannula was implanted into the VTA (Fig. 1A,B). Following 7 days of surgical recovery, concurrent MRI and electrophysiological recordings were performed. After baseline “resting state” scans, AMPA was infused into the VTA via the guide cannula, which is known to enhance dopamine release in the nucleus accumbens (NAc), the major target of VTA dopamine neurons (Karreman et al. 1996). As a control procedure, on separate scan days, the same animals also received a VTA saline injection (see Supplementary Fig. 1). Special care was taken to minimize electrical artifacts in LFP recordings induced by the MRI scanner and to ensure that only 1 μl AMPA solution was infused into the VTA since the microinjection pump was located outside the magnet (Fig. 1C). Methodological details are given below.

Subjects

A total of 29 male Sprague-Dawley rats (275–450 g; Charles River) were included in the study. Nineteen rats were used to establish baseline striatal rsFC; 10 rats were implanted with microelectrode arrays and guide cannulas for pharmacological manipulations. Rats were single-housed in a colony room with a 12-h light/dark cycle with access to water and food ad lib. All experiments were performed during the light phase. All procedures were approved by the NIDA-IRP Animal Care and Use Committee.

Survival Surgery: Intracranial Microelectrode Array and Guide Cannula Implantation

Using aseptic procedures, a silicon-based MRI compatible 16-channel linear electrode array was implanted into rat striatum such that LFP signals form most of the dorsal lateral to the ventral medial striatal domain were recorded. The electrodes were 15 μm thick, 10 mm in length, with 16 30 μm gold electrode contacts centered 200 μm apart (Neuronexus, Ann Arbor, MI). Animals were anesthetized under 2% isoflurane; core body temperature, respiration and arterial oxygenation levels were continuously monitored and maintained within their normal physiological range throughout the surgery. A 1 mm² window was made above the nucleus accumbens (NAc) (AP: +1.2 mm, ML: 3.9 mm, DV: −7.0 mm). The dura was carefully removed using a 27-gauge needle and a #10 scalpel blade. The electrode array was lowered into brain using a motorized micro-driver at a 20° angle (medial-lateral). Dental cement was used to fix the electrode in place following a 5 minute rest to allow the tissue to reset. A 22-gauge plastic guide cannula (7 mm below pedestal, Plastics One, Roanoke, VA) was implanted into the VTA (AP: −5.0 mm, ML: 0.6–0.8 mm). A cannula dummy was inserted into the guide cannula to prevent foreign matter from entering the brain. Finally, a ground/reference screw was placed over the cerebellum just above the surface of the brain. The screw was made from brass. It was soldered to the ground and reference wires from the probe using silver solder. Rats were allowed to recover for a minimum of 7 days before any fMRI/electrophysiological recording experiments were conducted. Gentamicin and Buprenorphine were administered (i.p.) to reduce infection and pain during the recovery period.

fMRI Data Acquisition

We followed the same imaging protocol as described previously (Lu et al. 2012). Rats were initially anesthetized with 2% isoflurane until the hind limb reflex was absent, followed by a loading dose (0.01 mg/kg, i.p.) of dexmedetomidine (Dexdomitor), an α2-adrenergic receptor agonist. A continuous subcutaneous infusion of Dexdomitor (0.01 mg/kg/h) was then initiated. The rat was transferred to an MRI compatible holder and the head stabilized using a customized bite bar and ear holders. The rat’s body temperature was monitored using a rectal temperature probe and maintained at 37 ± 0.5 °C with the help of a water-heating pad. Respiration rate was continuously monitored (SA Instrument Inc., NY, USA). Oxygen saturation was monitored via a noninvasive MRI-compatible pulse oximeter (Starr life sciences corp., PA, USA) connected to a hind paw, O₂ saturation was maintained at >96% by adjusting oxygen concentration of the inhaled air mixture. Isoflurane was subsequently lowered to 0.5–0.75% for the remainder of the study.

MRI data were acquired on a Bruker Biospin 9.4 T scanner equipped with an actively shielded gradient coil running ParaVision 6.0.1 (Bruker Medizintechnik, Karlsruhe, Germany). An 86 mm volume transmitter coil was used for RF excitation, and a 2-cm OD surface coil was placed directly over the skull encompassing the area of the electrode and cannula placement for MR signal reception. Tuning and matching of the coils were performed on every subject. T2-weighted anatomical images were acquired using a rapid acquisition with relaxation enhancement (RARE) sequence. The scan parameters were as follows: TR = 2500 ms, TE = 40 ms, RARE factor = 8, field of view (FOV) = 3.5 × 3.5 cm², matrix size = 256 × 256. Functional images were acquired using a gradient echo echo-planar imaging (EPI) sequence. Parameters: FOV = 3 × 3 cm², matrix size = 64 × 64, TR = 1500 ms, TE = 13 ms, four 1 mm thick slices. Three of the slices were located within the striatal region with the fourth slice covering the VTA. The data acquisition bandwidth was 250 kHz. The midsagittal view of the anterior commissure decussation (−0.36 mm from bregma) was used as the fiducial landmark to standardize slice localization within and between subjects. Prior to acquiring concurrent electrophysiology-fMRI data (n = 10), we acquired data from 19 animals without electrode implantation, which were used to identify and determine the consistency of resting state striatal and other brain networks. For this group of rats, the number of slices was set to 11. fMRI data were initiated 90 min after the loading dose of dexmedetomidine. A series of studies from our lab has demonstrated stable animal physiology and robust BOLD signal beginning at this time latency (Lu et al. 2012; Brynildsen et al. 2016; Hsu et al. 2016). Each resting state fMRI scan lasted 7 min and 34 s and comprised 300 repetitions and 4 dummy scans. In studies that involved VTA AMPA microinjections, pre- and post- AMPA scans were conducted sequentially, up to 50 min post-AMPA injection.

Multiple-channel LFP Signal Recording Concurrent with MRI Scan

We developed a method to record LFPs inside the 9.4 T MRI scanner. Once the animal was secured in the MRI holder, a customized jumper cable (1.5 feet in length; Plexon Inc., Dallas, TX) was attached directly to the probe implant (Neuronexus, Ann Arbor, MI). This jumper cable allowed for the headstage to be placed outside of the gradient and RF coils to reduce the MRI-induced artifacts. An MRI compatible headstage (gain = 1) was attached to the jumper cable and the amplifier. LFPs were recorded using a multichannel data acquisition system (Plexon Inc., Dallas, TX) and amplified at a gain of 1000, notch filtered at 60 Hz, and digitally sampled at either 5 or 10 kHz. In additional to the 16 channels from the electrode array, the respiratory waveform and MRI trigger signal that indexed the beginning of each imaging slice were recorded on 2 separate channels. The trigger signal was used for subsequent removal of the MRI artifacts in the LFP traces. The amplifier for the 15th channel was consistently dead across all recordings in our system; noise data from this channel was not included for analysis.

Pharmacological Manipulation

The striatum is the major projection target of midbrain dopamine neurons, with the ventral striatum (nucleus accumbens, NAc) receiving projections mainly from the VTA, while the dorsal striatum receives mainly substantia nigra, pars compacta afferents (Ikemoto 2007). More recent studies have demonstrated, using opto-fMRI, that the dorsal striatum becomes activated in response to optogenetic stimulation of VTA dopaminergic neurons (Beier et al. 2015; Ferenczi et al. 2016; Decot et al. 2017; Lohani et al. 2017). A previous study (Karreman et al. 1996) demonstrated that microinjection of AMPA (RS-alpha-amino-3-hydroxy-5-methyl-4-isoxazolepropionic acid), an ionotropic glutamate receptor agonist, into the VTA enhanced dopamine release in the ventral striatum. We took a similar approach to modulate ventral striatal neuronal activity via AMPA injections (Bristol, UK, 1 μl, 100 μM in saline vehicle) into the VTA during LFP and fMRI data were acquired. Specifically, a fused silica needle (Plastics One, Roanoke, VA) was lowered into the implanted guide cannula and connected to a micro-infusion pump via 1.5 m long polyethelene tubing (PE 10). The pump was located on a plane level to the animal inside the scanner to minimize solution backflow or leakage. Efforts were taken to avoid air bubbles inside the infusion line.

Resting State MRI Data Analysis

Imaging analyses were performed using the AFNI software package (Cox and Hyde 1997). Spatial independent component analysis (ICA, Melodic package from FSL software, Oxford University, UK) was performed on each scan session to identify and subsequently remove artifacts associated with respiration and cardiac pulsation. Images from each session were co-registered onto a common 3D space anatomical image (Lu et al. 2010). The registered data were subjected to a processing pipeline consisting of slice-time correction, linear and quadratic trend removal, spatial smoothing with a Gaussian kernel [full width at half maximum (FWHM) = 0.6 mm]. Data were high-pass filtered at 0.01 Hz for subsequent group ICA analysis and bandpass filtered with a cutoff frequency of 0.01–0.1 Hz for the seed-based correlation analysis. Three ICA components covering the dorsal, middle, and ventral striatum were consistently identified.

Three seed regions were defined as the center voxels within the 3 striatal ICA component maps, which corresponded to ventral (NAc), middle, and dorsal domains of the striatum (Voorn et al. 2004). Voxel time courses from each seed were averaged and served as the reference function in a whole brain correlation analysis (Biswal et al. 1995). The resulting cross-correlation coefficients were transformed to Z scores. Data from pre-AMPA baseline scan, first, second, third and fourth post-AMPA injection scans were grouped and subjected to linear mixed-effects modeling analysis (3dLME in AFNI) to determine the effect and time course of AMPA on BOLD rsFC. An uncorrected P < 0.02 with a cluster size of 17 voxels yielded a P_corrected < 0.05 significance level, accounting for multiple comparisons (AFNI program 3dclustsim).

LFP data analysis

LFP data were exported to EEGLAB (Delorme and Makeig 2004), which provides a convenient way for visually examining data consistency from multiple recording channels. LFP data were analyzed in Matlab (MathWorks, MA, USA) with function calls to EEGLAB routines. LFP data analyses included MRI artifacts removal, phase–amplitude coupling (PAC) quantification, band-limited power calculation, and LFP spectral analysis as detailed below.

Removal of MRI-induced Artifacts

MR imaging induces well-known artifacts in LFP traces when both signals are obtained simultaneously. Our MRI artifact removal algorithm included the following steps: first, the starting time of each imaging slice was identified based on concurrently acquired trigger signal from the scanner. Each 60 ms LFP segment immediately following the trigger signal had high-amplitude (up to ±5 mV) fast changing signals. These segments were replaced by linear interpolation. The resulting signal was then low-passed filtered to 400 Hz using an eighth order Butterworth filter. The Matlab filter function filtfilt() was applied to filter the data, which avoids phase nonlinearity and group delays. The filtered data were down-sampled to 1 kHz. The linear interpolation minimized ringing effects in the filtering process resulting from high frequency artifact spikes. The linearly interpolated segments were then replaced by cubic-spline interpolation of the data from 35 ms before and 35 ms after the artifact segments. Finally, LFP data were low-passed filtered to 100 Hz and down-sampled at 250 Hz using a Butterworth filter and the Matlab filtfilt() function.

To determine the potential distortion of the LFP signal from this artifacts correction method, clean LFPs were acquired without concurrent MRI scans and analyzed using identical steps described above using the trigger signal from another scan session. As a comparison, the same data were then down-sampled to 250 Hz (low-passed at 100 Hz) without the artifact removal steps. The waveforms with and without the data interpolation steps essentially overlapped, and the power spectra of the signals following these two processing methods were essentially identical, suggesting minimal distortion of the LFP signal resulting from the MRI artifact removal method.

Quantification of Cross-Frequency Phase–Amplitude Coupling (PAC)

The coupling between the phase of a low frequency signal and the amplitude of a high frequency signal was quantified using an “Envelope-to-Signal” coupling (ESC) method (Bruns and Eckhorn 2004; Onslow et al. 2011). The ESC measure calculates the correlation between the amplitude envelope of the filtered high frequency signal, A_famp, and the filtered low frequency signal, Y_fph.

$$ES{C}_{famp,fph}=r\left(A_{famp},Y_{fph}\right)$$ (1)

Signal filtering was achieved via convolution of the raw LFP signal with complex Morlet wavelets w(t, f₀) that have a Gaussian shape both in the time (${\sigma }_t$) and in the frequency domain (${\sigma }_f$) around its central frequency f₀ (Tallon-Baudry et al. 1997):

$$w\left(t,f_0\right)=A\cdot \exp \left(-\frac{t^2}{2{{\sigma }_t}^2}\right)\cdot \exp \left(2i\pi f_{0}t\right)$$ (2)

with ${\sigma }_f=1/\left(2\pi {\sigma }_t\right).$ Wavelets are normalized so that their total energy is 1, with the normalization factor A being equal to ${\left({\sigma }_t\cdot \sqrt{\pi }\right)}^{1/2}$. A wavelet family is characterized by a constant ratio $\left(f_0/{\sigma }_f\right)$ , and is set to 7, which has been shown to provide a good trade-off between temporal and spectral resolution (Torrence and Compo 1998). The instantaneous amplitude A_famp is calculated as the absolute value of the analytic signal, and Y_fph is equal to the real part of the signal after Morlet wavelet transformation.

LFP–BOLD Correlation Analysis

A previous study demonstrated that the evoked BOLD response correlated with LFP power in the γ band, reflecting input to and local processing within the imaging region (Logothetis et al. 2001). It is unknown whether the same relationship extends to spontaneous BOLD fluctuations. As such, we calculated LFP band-limited power, and investigated its relationship with BOLD signal recorded concurrently (herein noted as LFP–BOLD).

LFP band-limited power was computed as follows: The LFP signal was filtered into 6 classical frequency bands (delta (δ): 1–4; theta (θ): 5–8; alpha (α): 9–14, beta (β): 15–30; gamma (γ): 30–50; and high gamma (hγ): >70 Hz) (Leopold et al. 2003). Signal between 50 and 70 Hz was discarded to avoid potential contamination from 60 Hz power noise. A bandpass filter was realized by sequentially filtering the data using a high-pass Butterworth filter followed by a low pass Butterworth filter. The filters were designed to have sharp band transition in Matlab (with 8 or 12 orders). Data were filtered using Matlab filtfilt() function. LFP signal in individual bands were full-wave rectified by taking their absolute value to derive the band-limited power time course (Leopold et al. 2003), which was then down-sampled to MRI repetition time (TR) by averaging power within each TR (1.5 s). The slice-trigger signal from the scanner was used to align the LFP signal with the EPI time course. After down-sampling, band-limited power time courses had the same temporal resolution as the resting fMRI data.

The calculation of LFP–BOLD correlation was similar to seed-based functional connectivity correlation analysis except that the reference function was the band-limited power time course. There were 6 band-limited power time courses from each of the 15 recording channels. Each power time course was cross-correlated with the BOLD time course from the whole brain, resulting in 6 LFP–BOLD correlation maps, which were transformed to Z scores. Within each LFP frequency band, the LFP–BOLD correlation maps were similar across channels. We thus averaged the correlation maps across channels to improve the signal-to-noise ratio. If artifacts were seen in some channels, the LFP–BOLD correlation from these channels were discarded from averaging. For each frequency band, group pre-AMPA baseline LFP–BOLD correlation maps were derived by pooling data across all animals and performing a t-tested against zero. The AMPA effect on LFP–BOLD correlation was analyzed in a similar fashion. Specifically, pre-AMPA baseline LFP–BOLD, and each post-AMPA time point from all rats and all scan sessions were analyzed using 3D linear mixed effect modeling with the AFNI 3dLME package (Chen et al. 2013). This calculation was repeated for LFP–BOLD correlation at all 6 frequency bands. An uncorrected P < 0.02 with a cluster size of 17 voxels yielded a P_corrected < 0.05 significance level, accounting for multiple comparisons (AFNI program 3dclustsim).

Results

Three Distinct Functional Connectivity Components define the Rat Striatum

A previous rsfMRI study using the same anesthetic protocol identified a default mode network along with sensory and motor networks in the rat (Lu et al. 2012). Consistent with that study and in addition to those networks, we now also identified a robust rsFC network localized to the bilateral striatum. Figure 2A shows striatal functional connectivity derived from a group ICA (n = 19 rats, total of 58 scan sessions). The 3 components roughly overlap the electrode placement (Fig. 2B,C) and with the 3 striatal functional domains previously reported (Voorn et al. 2004). They are also very similar to that previously identified in awake marmosets by our lab (Belcher et al. 2013), suggesting that these functional connectivity patterns are conserved across species and are not significantly affected by the current anesthesia regimen. While strong functional connectivity in rat striatum was observed previously (Lu et al. 2007; Zhao et al. 2008; Hutchison et al. 2010; Pan et al. 2013), none of these studies identified 3 distinct striatal components as reported herein. Intriguingly, there are only very weak (or no) monosynaptic connections between the bilateral striatum (McKenzie 2013). The possible neural mechanism that underlies such strong bilateral functional connectivity is discussed below.

Figure 2.

Resting state functional connectivity identified by independent component analysis (ICA). Three component maps cover the dorsal, middle and ventral domains of the rat striatum (A, top to bottom). A linear silicon-based microelectrode array was implanted such that the 16 contacts of the array covered most of the striatal domains as in the BOLD functional connectivity. The electrode location can be readily identified using traditional spin echo (SE) and gradient echo (GE) sequences. Small but nevertheless appreciable signal loss is seen in images acquired with a GE echo-planar imaging (EPI) sequence (B, bottom). Panel C shows histological reconstruction illustrating the electrode in the striatum (top) and guide cannula above ventral tegmental area (VTA, bottom). Numbers bellow figures indicate coordinates relative to bregma in millimeter.

Unilateral VTA AMPA microinjections modulate striatal rsFC

We applied a seed-based correlation analysis approach to analyze the effects of VTA AMPA injection on striatal rsFC using AFNI 3dLME statistical package (Chen et al. 2013) (Fig. 3A), with time window as the independent variable. The seed voxels were placed at the center of mass from the striatal ICA component maps shown in Figure 2A. The main effect of AMPA on striatal rsFC is shown in Figure 3B. The “main effect” map shows only those voxels that had significant changes in the correlation values across time windows. Thus AMPA injections modulated FC only in the NAc, the major target of VTA dopamine neurons (Fig. 3B). All maps were threshold at P < 0.05 after corrections for multiple comparisons. Figure 3C depicts the striatal subdivision based on Paxino’s rat brain atlas (top) and Voorn et al. 2004 (bottom; with permission from Elsevier).

Figure 3.

Effects of VTA AMPA microinjection on BOLD functional connectivity. (A) Pre-AMPA BOLD functional connectivity was analyzed using a seed-based correlation approach with the seed defined based on the ICA maps in Figure 2. VTA AMPA microinjection significantly modulated BOLD functional connectivity in the NAc (B) . Note only those voxels that had significant changes in the correlation values across time windows would be significant in the main effect map. All maps were thresholded at P < 0.05 after correction for multiple comparisons. Panel (C) delineates NAc core (Acbc) and shell (AcbSh) based on Paxinos’ rat brain atlas. The dorsal, middle, and ventral subdivision of the rat striatum is also shown in panel C (bottom, reprinted from Voorn et al. Trends Neurosci. 2004;27(8)468–74, with permission from Elsevier). The dorsal striatum consists of caudate-putamen complex, the middle and ventral domain consist of NAc and the striatal elements of the olfactory tubercle (OT). Note that VTA AMPA microinjection modulated BOLD functional connectivity in Acbc and AcbSh, the major targets of VTA dopaminergic neurons. Numbers below figures indicate coordinates relative to bregma in millimeter. AMPA, α-amino-3-hydroxy-5-methyl-4-isoxazolepropionic acid; ac, anterior commissure.

Relationship between LFP–BOLD Correlations Pre- and Post- VTA AMPA Microinjection

Having demonstrated that VTA AMPA microinjections modulated BOLD functional connectivity in the NAc (Fig. 3B), we then investigated the neural mechanism underlying this observation. We reasoned that the LFP signal that drives BOLD functional connectivity should be modulated by VTA AMPA injections in a similar manner, both spatially and temporally. The LFP data acquired simultaneously with the BOLD fMRI data offers a unique opportunity to address this assumption.

LFP Signal Modulated by VTA AMPA Microinjection

After the removal of MRI-induced artifacts (see Methods and Supplementary Fig. 2), one can appreciate that the striatal LFP signal was dominated by low frequency oscillations in the delta (δ) range (Fig. 4), consistent with previous observations (Leung and Yim 1993; Wilson and Kawaguchi 1996; Goto and O’Donnell 2001). Figure 4B illustrates power spectra from a representative electrode contact located in dorsal, middle, and ventral striatum, corresponding to the 3 ICA identified striatal components (Fig. 2A). Both the amplitude and frequency of the oscillations were modulated by VTA AMPA microinjections (Fig. 4B, red trace). Somewhat surprising, LFP signal in the dorsal striatum was also modulated by VTA AMPA injection. However, a series of recent studies employing optogenetic stimulation of VTA dopaminergic neurons reported robust fMRI responses in the dorsal striatum, with the lateral VTA shown to project to dorsal striatum (Ferenczi et al. 2016; Decot et al. 2017; Lohani et al. 2017).

Figure 4.

Effects of VTA AMPA microinjection on striatal LFP signal. (A) Microinjection of AMPA into the VTA modulated both the amplitude and frequency of the spontaneous LFP signal (the pre- and post-AMPA waveforms were derived from the blue and red boxes, respectively). The 3 columns in (B) are spectral analysis from 3 electrode channels recorded from the dorsal, middle, and ventral striatum (NAc), respectively.

Pre-AMPA Baseline Frequency Band-dependent LFP–BOLD Correlations

We next explored the relationship between LFP fluctuations and BOLD fluctuations within the striatum. Our data analysis strategy was similar to the above seed-based BOLD FC analysis, except that the reference function was replaced with the LFP band-limited power time course filtered LFP signal into 6 conventional frequency bands

For a given neuronal event, there is a well-known hemodynamic delay in the evoked BOLD response of about 4–5 s in human motor cortex, which varies across brain regions (Handwerker et al. 2012). In contrast, the delay is not well characterized in the rat brain, although a recent study using intrinsic optical signal reported a hemodynamic response lag of ~2 s behind the neuronal calcium signal (Matsui et al. 2016). As such, we computed LFP–BOLD correlations with different hemodynamic delays by shifting the LFP power time courses backward and forward up to 6 TRs (−9 s to +9 s). As shown in Supplementary Figure 3, forward shifting LFP signal by 3 s (2 TRs) gave the strongest LFP–BOLD correlation; shifting the LFP in both the γ band and high γ band generated similar time shift-dependent LFP–BOLD correlation patterns. We thus applied a hemodynamic delay of 3.0 s in the LFP–BOLD correlation analysis.

As shown in Figure 5A, we found a significant negative correlation between the δ LFP and the BOLD signal, while positive correlations were seen between the β, γ and hγ LFPs and the BOLD signal. Critically, irrespective of the directionality in LFP–BOLD correlation across these frequency bands, the correlation maps all have similar spatial patterns, covering ventral, and medioventral domains of the striatum. LFP–BOLD correlations in the α and θ bands were marginally significant and were not specifically localized to the striatum (see Supplementary Fig. 4).

Figure 5.

Effects of VTA AMPA microinjection on LFP–BOLD correlation. (A) Pre-AMPA LFP–BOLD correlation maps in the δ(1–4 Hz), β(15–30 Hz), γ(30–50 Hz) and high γ(>70 Hz) frequency range. LFP–BOLD correlation in the θ (5–8 Hz) and α(9–14 Hz) bands was marginal and is shown in Supplementary Figure 4. Although all bands display similar spatial correlation patterns, the δ LFP–BOLD correlation was negative, while a positive correlation was seen between the other 3 frequency bands. AMPA modulated the LFP–BOLD correlation only in the δ band (B). Note only those voxels that had significant changes in the LFP–BOLD correlation across time windows would be significant in the main effect map. All maps were thresholded at P < 0.05 after correction for multiple comparisons.

Unilateral AMPA Microinjection into VTA Modulates Striatal LFP–BOLD Correlation

We next investigated how VTA AMPA microinjections modulate LFP–BOLD correlations. LFP–BOLD correlation data from the pre-AMPA baseline and up to 37.5 min post-AMPA were subjected to voxel-wise 3dLME statistical analysis (Chen et al. 2013). The “main effect” map shows only those voxels that had significant changes in LFP–BOLD correlation across time windows. As shown in Figure 5B, we found significant modulation in the LFP–BOLD correlation only in the δ band, but not in any other frequency band. Voxels with significant AMPA modulation were clustered within the NAc, ventral lateral striatum, and insular cortex.

NAc Cross-frequency PAC Effect is Modulated by VTA AMPA Microinjections

As shown in Figure 5A, basal LFP–BOLD correlations exhibit striking frequency band-dependence, with the δ band LFP–BOLD correlation being negative while a positive correlation was seen between the β and γ LFP–BOLD. This observation raises an intriguing question: what is the neuronal mechanism that underlies such distinction? Since the cross-frequency PAC effect has been proposed to orchestrate the LFP signal at different bands, we next investigated the PAC effect before and after AMPA modulation.

Figure 6A shows raw LFP traces from a representative electrode channel. In this case, the amplitude of the signal between 9 and 30 Hz was consistently enhanced at the troughs of the low frequency signal (δ, Fig. 6A top). We quantified the PAC effect across frequencies up to 100 Hz by calculating the correlations between the waveform of the low frequency and the envelope of the high frequency signal derived from a wavelet transformation (Onslow et al. 2011). Thus, a negative PAC value indicates enhanced amplitude in the high frequency LFP that occurs at the trough of the low frequency signal. Similar PAC effects were observed across all 16 electrode channels. As a result, PAC values from dorsal, middle, and ventral striatal electrode locations were averaged across animals. The LFP signal phase in the 1–4 Hz (δ band) frequency range strongly modulated the amplitude of higher frequency (8–50 Hz) LFPs. Figure 6B shows F-statistical maps comparing PAC effects pre-and post-AMPA microinjection across data acquisition time windows, each lasting for 7.5 min across the 3 regions of the striatum. Regions in red indicate significant AMPA modulations across time; bottom row shows post-hoc t-statistics comparing the PAC effect between pre-AMPA baseline and 7.5-min post-AMPA microinjections (both after Bonferroni correction). Thus, when we applied PAC analysis to the electrophysiological data we found that the phase of the δ band modulates the power of the higher frequency band LFP signals (Fig. 6B). In addition, AMPA administration induced a significant reduction in PAC values in the 8–50 Hz (α, β, θ, and γ band) and above 70 Hz (high γ band) frequency ranges.

Figure 6.

(A) Illustration of high-amplitude LFP signal with the frequency in the range of 9–30 Hz occurring consistently at the trough of the low frequency signal. Phase–amplitude coupling (PAC) was quantified based on the raw waveform of the low frequency signal and amplitude waveform of the high frequency signal derived from wavelet transformation (right). (B) (Top row): F-statistical maps comparing PAC effects pre- and post-AMPA microinjection across data acquisition time windows. Regions in red indicate significant AMPA modulation effect across time; bottom row shows t-statistics comparing PAC effect between pre-AMAPA baseline and 7.5 min post-AMPA microinjection (both after Bonferroni correction). AMPA injection induced significant reduction in PAC values in frequency ranges shown in blue.

Figure 7 illustrates a summary of the quantitative temporal profiles of BOLD FC (A), LFP–BOLD correlation (B) and cross-frequency PAC (C,D) pre- and post-AMPA modulation. Thus VTA AMPA microinjections induced a reduction in BOLD functional connectivity in the NAc, which was accompanied by a reduction in negative correlation strength of LFP–BOLD coupling in the δ band and a reduction in the ventral striatum PAC effect in the same time windows.

Figure 7.

Time course plot of BOLD functional connectivity (A), LFP–BOLD correlation (B) and cross-frequency PAC (C) in the ventral striatum pre- and post-AMPA microinjection into the ventral tegmental area. Compared with pre-AMPA baseline (BL), AMPA microinjection caused a reduction in BOLD functional connectivity, along with a reduction in negative correlation strength in LFP–BOLD as well as in the PAC effect. F-statistical map comparing PAC pre- and post-AMPA microinjection (D).

Together, these results suggest that the striatal δ LFP oscillations regulate the higher frequencies imbedded within these oscillations which can be identified using PAC and modulated via intra-VTA injection of AMPA.

Discussion

We demonstrated robust BOLD functional connectivity within the bilateral striatum, which was modulated by AMPA microinjections into the VTA. Simultaneously recorded LFP and BOLD data demonstrated that the δ band LFP power was negatively correlated with BOLD functional connectivity, while β and γ band LFP power was positively correlated with BOLD functional connectivity. The α and θ band significant correlations with the BOLD signal were not specifically localized within the striatum. The δ, β, γ and hγ band LFP–BOLD spatial correlations had a pattern similar to that of the BOLD functional connectivity. VTA AMPA microinjections modulated both δ LFP–BOLD correlations and BOLD functional connectivity in the NAc, the major target of VTA dopaminergic projections. Notably, BOLD–LFP correlations in other frequency bands were not significantly modulated by AMPA microinjections, despite the fact that they shared similar pre-AMPA LFP–BOLD correlation patterns (Fig. 5). Temporally, a reduction in BOLD functional connectivity was accompanied by a reduction in negative correlation strength between δ band LFP and the striatal BOLD signal. Further, there were strong cross-frequency PAC effects in the LFP signal. The LFP waveform in the δ band was negatively correlated with the LFP signal amplitude at higher frequencies, and the PAC effect was significantly reduced by VTA AMPA microinjections (Figs 4–6). Taken together, our data suggest a mechanistic role of the δ band LFP signal upon rsFC BOLD via a cross-frequency PAC mechanism.

Implications for Functional Connectivity Networks

Previous studies in both awake and anesthetized animals have established prominent two-state spontaneous δ LFP oscillations in the rat striatum, which can be detected both intra- and extracellularly. Dendritic voltage-dependent potassium channels in medium spiny neurons appear to play a major role in generating striatal δ oscillations (Wilson and Kawaguchi 1996). Notably the extracellular LFP signal is always in opposite phase of the intracellular membrane potential (Leung and Yim 1993; Wickens and Wilson 1998; Goto and O’Donnell 2001). We extended the above findings and found a strong PAC effect, demonstrating that the slow LFP signal not only modulates cell excitability, it also modulates the amplitude of high frequency activity. It is conceivable that neuronal populations contribute to a given frequency oscillation will be more or less likely to be excited depending upon the states of the population oscillation. This synchronization can create temporal windows for segregating cortical populations, which can separate information intake and transfer processes. Neuronal populations in the same state will be more likely to interact, exchange information, and modulate synaptic plasticity (Fries 2005; Cavanagh and Frank 2014). Furthermore, rhythmic changes in excitability has been proposed to instantiate transient functional networks between spatially distal sites (von Stein and Sarnthein 2000; Fries 2005). Thus with the pivotal role of slow oscillations in BOLD functional connectivity as shown here, the modulation of the high frequency signal by the phase of the low frequency signal may play a fundamental role in organizing large-scale network activity (e.g., the default mode network) as identified by resting state MRI.

Establishing the key role of δ LFP oscillations in rsFC could also help explain some puzzling observations in BOLD functional connectivity studies. For example, as stated above, there are no or very few monosynaptic connections between homotopic regions of striatum, yet, we and others have observed prominent bilateral striatal functional connectivity. Similar observations have been made in the visual cortex, where bilateral connections are sparse (Raichle 2011). What is known (Leung and Yim 1993) is that bilateral striatal δ oscillations are highly synchronized. Similarly, the rat and human DMN (Raichle et al. 2001; Lu et al. 2012) are composed of a number of association and limbic cortices that are spatially segregated, with minimal known monosynaptic connections between regions, for example between posterior parietal cortex and hippocampus (Reep et al. 1994), and between posterior cingulate cortex and medial prefrontal cortex (Carmichael and Price 1996). A testable hypothesis is that these structures organize into a network by synchronous low frequency fluctuations, which in turn modulate neuronal excitability among them.

Comparison with Previous Studies

In a previous study where epidural EEG signal and fMRI were recorded in separate sessions(Lu et al. 2007), we reported region-specific, anesthetic dose-dependent fMRI resting-state functional connectivity in bilateral primary somatosensory cortex (S1FL) in the rat brain. Furthermore, the synchrony in δ EEG power exhibited similar region-specificity and dose-dependency, i.e., a reduction in S1FL functional connectivity was accompanied by a reduction in δ EEG power synchrony between bilateral S1FL electrode pairs. However, we were not able to interrogate the direct relationship, in particular, the polarity of the correlations between these two types of signal since they were not collected simultaneously. In the present study, with the development of concurrent multichannel electrophysiological recording and fMRI, and with the aid of pharmacological circuit modulation, we further interrogated this relationship and found that a reduction in BOLD functional connectivity was accompanied by a reduction in negative correlation strength between the δ band LFP and BOLD signal. The findings from the current study support the proposition that spontaneous δ oscillations plays a pivotal role in network synchrony as observed in rsFC.

There are two contrasting hypotheses regarding the neural mechanism of rsFC: one emphasizes higher frequency (γ) LFP signals (Nir et al. 2008; Shmuel and Leopold 2008), while contrasting data indicate a role of low frequency LFP (Lu et al. 2007; He et al. 2008; Pan et al. 2013). In spite of technical limitations in each of the studies mentioned above, there is no unified transduction coupling theory that explains the experimental data under different conditions. The present study unambiguously demonstrates that, at least in rat striatum, although β, γ and high γ LFP frequencies all contribute to the observed resting state fMRI signal, they are organized via a cross-frequency PAC mechanism in which the phase of the δ LFP modulates the amplitude of LFPs at higher frequencies, which in turn modulated the BOLD signal. Importantly, a reduction in BOLD functional connectivity was accompanied by a reduction in negative correlation strength of LFP–BOLD only in the δ band, but not other bands. Our finding that, via a cross-frequency PAC mechanism, δ band LFP orchestrates neural activity offers a unified transduction coupling theory to explain the neural mechanism of the resting state fMRI signal.

Positive/negative Correlation of LFP Band-limited Power, BOLD Signal and PAC Effect

Elucidating the neurophysiological basis of the BOLD signal has been a subject of intensive research for almost two decades. It is now generally accepted that in most cases the BOLD response correlates with LFP and MUA, although scenarios exist when BOLD signal and MUA diassociate (Lauritzen 2001). The study by Shih and colleagues is particularly relevent to the current study (Shih et al. 2009): unilateral noxious electrical stimulation induced robust reduction in fMRI signal accompanied by heightened (instead of decreased) neuronal firing in rat dorsal striatum. The LFP signal, an electrophysiological readout reflecting synaptic inputs and local neural processing, appears to better correlate with the BOLD signal (Logothetis 2003). Additional studies along this line reported a complex relationship between LFP power of individual bands and the BOLD signal, with the directionality of the correlations could be positive or negative (Niessing et al. 2005; Magri et al. 2012). A recent human study reported that the cross-frequency PAC effect accounted for a significant amount of variance in the BOLD response to finger tapping that cannot be explained by the LFP power changes alone (Murta et al. 2017). However, it remains unknown why LFP–BOLD coupling could be positive or negative, how PAC mediates the LFP–BOLD correlation, and whether the PAC effect could predict the directionality of LFP–BOLD correlations.

In the absence of abrupt change in LFP power time-locked to an external task, resting state fMRI has revealed a complex relationship between LFP band-limited power and BOLD signal has been reported in anesthetized monkeys and rats (Thompson et al. 2014; Hutchison et al. 2015). PAC in the LFP signal was also observed in the rat somatosneory cortex (Sotero et al. 2015). A recent monkey study by Hutchison et al. (2015) reported opposite polarity in correlation between LFP band-limited power and spontaneous BOLD signal. Specifically, delta band LFP power and BOLD was negative in sign, while LFP power in beta-and gamma band and BOLD was positive. Our data are in line with this report. A critical question is how to interpret the directionality of the correlations. Further work is needed to address this observation.

Technical Considerations and Limitations

We implanted a single linear electrode array to the rat striatum in an attempt to cover the 3 ICA component maps as shown in Figure 2. However, our electrode placement covers mostly the dorsal-medial and the ventral-medial part of the striatum (NAc), but to a less extent, the lateral part of rat striatum (middle component), since it is well-established that VTA dopaminergic neurons project mostly to NAc, while substantia nigra dopaminergic neurons project mostly to the dorsal striatum (Ikemoto 2007); our AMPA modulation site is VTA, and the target site is NAc. To perfectly cover LFP signal from middle ICA component would require another electrode array, which adds in another layer of complexity, and was not explored.

We found that the cross-correlation between δ LFP and the fMRI signal peaked at about −0.3, but the average was about −0.1; the cross-correlation between γ LFP and the fMRI signal was similar, but with an opposite sign. These data imply that fluctuations in LFP explain a small portion of variance in the fMRI signal, consistent with a previous report (Scholvinck et al. 2010). A number of factors should be taken into account when interpreting this correlation result. Our LFP–BOLD correlation analysis depends, to a certain degree, on the assumption of linearity in neurovascular coupling. In evoked fMRI some nonlinearity is known to exist (Birn et al. 2001). In the case of resting state fMRI, to the best of our knowledge, this assumption has not been tested. Second, BOLD signal relies on a mismatch in oxygen supply and oxygen consumption, which is the result of a complex interplay between blood flow, blood volume and oxygenation. Blood flow regulation ultimately relies on the release of vasoactive substances from both neuronal and glial cells (Attwell and Iadecola 2002; Iadecola and Nedergaard 2007). Our measurement of extracellular potential is at best a surrogate of neuronal processes. Third, our LFP recording was hardware limited to frequencies >0.3 Hz, thus the contribution from ultraslow component of the LFP signal was ignored. Such contribution has been reported to be significant (Pan et al. 2013). On the other hand, even under optimal conditions where the fMRI and neural signals are driven by a strong visual stimulus using a block design, only about 50% of the BOLD variance can be explained by the LFP signal (Logothetis et al. 2001).

Of the pre-AMPA LFP–BOLD correlation maps shown in Figure 5A, there was minimal significant correlation in the dorsal lateral domain. One potential reason is that in general, we found the correlation between LFP band-limited power and BOLD was generally low, as discussed above. Compromise in image quality after electrode implantation likely has contributed to this, since the width of the electrode array is larger at the top than at the bottom. We found that the LFP–BOLD correlation maps across channels were similar, and thus we averaged the data to improve the SNR. Similar LFP–BOLD correlation map across channels were somewhat surprising, but this might reflect the fact that LFP signal is not as local as one has anticipated. Previous studies from different labs on different brain areas have drawn different conclusions with regard to “how local the local field potential is”. The spatial spread of LFP depends on the type of electrodes, geometry, local circuits, cell types (Buzsáki et al. 2012), from within approximately 200–400 μm of the recording electrode (Katzner et al. 2009; Xing et al. 2009) to 600–1000 μm (Berens et al. 2008), to 1.5–2 mm (Hoffman et al. 2007; Nauhaus et al. 2009). Nevertheless, by modulating the activity of VTA dopamine neurons via AMPA microinjection, we found statistical changes in BOLD functional connectivity and LFP–BOLD correlation in the nucleus accumbens, the major target of VTA dopamine neurons.

Another limiting factor is the use of anesthesia in our experiments. Many anesthetics are known to compromise neurovascular coupling (Aksenov et al. 2015), and induce low frequency oscillation synchrony of membrane potentials. Such slow oscillations are often used to index sleep stages (Steriade and Amzica 1998). These factors together raise the question whether the observed phenomenon is derived from or biased by the use of anesthesia or a reflection of sleep-related brain states. Although anesthesia may have influenced our results, several lines of evidence argue against the proposition that the current findings are an epiphenomenon of anesthesia or sleep. First, striatal δ oscillations have been observed not only in anesthetized rats but also in rats in a quiet, yet awake state (Leung and Yim 1993). Second, δ activity is not only seen in sleep, but it also participates in general information processing. For example, step-like shifts in membrane potential in visual cortical neurons are modified by presentation of visual stimuli (Douglas et al. 1991; Anderson et al. 2000). In both humans and rodents, frontal δ and θ activity is associated with cognitive control and organizing goal-directed activity (Parker et al. 2015; Emmons et al. 2016). Third, alterations in δ activity are linked to various brain diseases. For example, Babiloni et al. (Babiloni et al. 2006) reported that interhemispheric and fronto-parietal δ synchronization is reduced in patients with mild Alzheimer’s disease and vascular dementia. Indeed, it has been suggested that the UP and DOWN membrane potential fluctuations are a characteristic of an idling system during slow-wave sleep or quiet awake state and that they reflect a widely synchronized network activity (Destexhe et al. 1999; Goto and O’Donnell 2001).

In summary, our data demonstrate that while spontaneous striatal BOLD and LFP power in the δ, β, γ, and hγ band had similar spatial correlation patterns, modulation of striatal activity following AMPA microinjections into the VTA modulated striatal BOLD–LFP correlation only in the δ band, which also corresponded to BOLD functional connectivity changes. The phase of δ LFP significantly modulated the amplitude of the high frequency LFP, while the PAC effect was significantly reduced by AMPA microinjection. These data together strongly suggest that the δ LFP plays a key role in orchestrating the neural activity that drives spontaneous fMRI fluctuations.

Funding

This work was supported by the Intramural Research Program of the National Institute on Drug Abuse, NIH and UT Health San Antonio.