Systems-level neurophysiological state characteristics for drug evaluation in an animal model of levodopa-induced dyskinesia

Abstract

Disorders affecting the central nervous system have proven particularly hard to treat, and disappointingly few novel therapies have reached the clinics in recent decades. A better understanding of the physiological processes in the brain underlying various symptoms could therefore greatly improve the rate of progress in this field. We here show how systems-level descriptions of different brain states reliably can be obtained through a newly developed method based on large-scale recordings in distributed neural networks encompassing several different brain structures. Using this technology, we characterize the neurophysiological states associated with parkinsonism and levodopa-induced dyskinesia in a rodent model of Parkinson's disease together with pharmacological interventions aimed at reducing dyskinetic symptoms. Our results show that the obtained electrophysiological data add significant information to conventional behavioral evaluations and hereby elucidate the underlying effects of treatments in greater detail. Taken together, these results potentially open up for studies of neurophysiological mechanisms underlying symptoms in a wide range of neurological and psychiatric conditions that until now have been very hard to investigate in animal models of disease.

diseases affecting the central nervous system (CNS) are a rapidly growing concern that puts a great economic burden on society (Olesen et al. 2012) and cause major suffering for afflicted individuals and their families. Unfortunately, these diseases have also proven particularly hard to treat. Despite impressive advances in the field of molecular biology in recent decades, few novel therapeutic options have reached clinics, notwithstanding corporate and regulatory efforts to break the trend (Graul 2008). A major challenge in the development of new CNS therapies is the limited understanding of the basic processes governing normal brain functions as well as the pathophysiological changes that ultimately cause the symptoms. For these reasons, the methodological approaches in drug discovery and development have often been limited to rather simplistic experimental readouts. In preclinical studies, the evaluation of novel compounds typically involve characterization of changes in animal behavior in combination with postmortem tissue analyses with little information about the ongoing CNS changes causing the actual symptoms or the underlying physiological effects of an intervention. To make matters worse, several neurological and psychiatric conditions are not directly associated with overt changes in behavior, which makes them even more challenging to model in experimental animals. To gain an insight into such internal CNS processes, chronic electrophysiological recordings are a particularly promising approach that can give real-time access to neurophysiological activity patterns associated with physiological processes during natural conditions (Gervasoni et al. 2004; Lehew and Nicolelis 2008). Building on this technology, large-scale sampling of neurophysiological signals from diverse brain regions could allow for the characterization of brain states that explains the difference between healthy and diseased states as well as how these states are altered by drugs aimed at treating the disease. Although clearly a great experimental challenge, such detailed information on neurophysiological states obtained in valid animal models of CNS disease could significantly help to increase the rate of progress in research aimed toward new treatments for CNS disorders.

In fact, even recordings performed in single locations of the brain, such as those that have been obtained in Parkinson's disease (PD) patients implanted with electrodes in the subthalamic nucleus (STN) and the internal part of the globus pallidus (GP) for the purpose of therapeutic deep-brain stimulation, have provided novel insights into pathological processes potentially underlying symptoms in this disease (Brown et al. 2001; Brücke et al. 2012; Lalo et al. 2008). In similar experiments on animals implanted with multiple electrodes, additional neurophysiological features have been identified that are thought to be associated with motor symptoms on and off medication. In particular, the parkinsonian hypokinetic state has been linked to an increased cell firing rate in the STN (Albin et al. 1989; Bergman et al. 1994; Levy et al. 2000), synchronized cell firing in cortex (Goldberg et al. 2002), striatum (Goldberg et al. 2004), GP (Nini et al. 1995), and STN (Bergman et al. 1994) as well as abnormally strong local field potential (LFP) oscillations in the beta band (approximately 10–35 Hz) present across the entire corticobasal ganglia network (Costa et al. 2006; Fuentes et al. 2010; Hammond et al. 2007; Stein and Bar-Gad 2013). Dopamine replacement therapy alleviating parkinsonian symptoms has been shown to suppress these aberrant activity patterns concomitantly (Costa et al. 2006; Gilmour et al. 2011; Kreiss et al. 1997; Santana et al. 2014). Unfortunately, following long-term dopamine replacement therapy, the therapeutic window frequently narrows to such an extent that treated subjects rapidly transition from parkinsonism to dyskinesia as the drug plasma concentration rises. In this situation, oscillatory activity in other parts of the LFP frequency spectrum has been reported to be markedly altered following levodopa administration in patients suffering from involuntary dyskinetic movements as a medication side effect. Low-frequency oscillations in the theta range (4–10 Hz), for example, have attracted particular attention over the years (Alam et al. 2013; Alegre et al. 2012; Alonso-Frech et al. 2006), and, more recently, characteristic gamma oscillations (at ∼80 Hz) in a rat model of PD were found to be strongly associated with levodopa-induced dyskinesia (Dupre et al. 2013; Halje et al. 2012). Equivalent high-frequency oscillations have also been reported in STN recordings in Parkinson patients, sometimes referred to as finely tuned high gamma, but have in these studies primarily been thought to reflect the prokinetic state associated with the therapeutic effect of the medication (Brown 2003; Cagnan et al. 2014; Cassidy et al. 2002; Sharott et al. 2005).

To clarify the association between different aberrant neuronal activity patterns and the expression of motor symptoms and to obtain a more comprehensive description of the neurophysiological state on a systems level, we have here made use of a technology developed in our laboratory that lets us perform large-scale multistructure recordings in awake behaving rats (Fig. 1, A and B; Ivica et al. 2014) in the most commonly used model of PD [the 6-hydroxydopamine (6-OHDA)-lesioned rat; Nadjar et al. 2009]. Applying this technology, we have investigated: 1) the neurophysiological state of the corticobasal ganglia-thalamic circuit that is associated with parkinsonism; 2) the neurophysiological state that is associated with levodopa-induced dyskinesia; and 3) the behavioral and electrophysiological effects of experimental and clinically used drug interventions aimed at alleviating dyskinetic symptoms.

Fig. 1.

Parallel recordings in 8 different structures of the corticobasal ganglia-thalamic loop in each hemisphere made possible with high-density recording arrays. A: microelectrode recording wires (n = 128) are distributed to target relevant brain structures (circles mark positions of single 30-μm tungsten wires; 250 μm center-to-center spacing within groups). B: the relative arrangement of wire groups is guided by a custom-made 2-dimensional (2D) array and a 3D aligner. Wires are electrically linked to a connector via a printed circuit board (PCB). RFA, rostral forelimb area; M1, primary motor cortex; DLS, dorsolateral striatum; DMS, dorsomedial striatum; GP, globus pallidus; Th, thalamus; STN, subthalamic nucleus; SNr, substantia nigra pars reticulata.

METHODS

Animals.

Four adult female Sprague-Dawley rats (230–250 g) were used in the study. The animals were kept on a 12:12-h light-dark cycle and received food and water ad libitum. All experiments were approved in advance by the Malmö/Lund ethical committee of animal experiments.

6-OHDA lesions and levodopa priming.

Rats were anesthetized with fentanyl-medetomidine (0.3 mg/0.3 kg ip injection) and fixed in a stereotaxic frame. The animals received two injections of 6-OHDA hydrochloride (3.0 μg/μl free base dissolved in 0.02% ascorbate saline) into the medial forebrain bundle of the right hemisphere at the following coordinates from bregma and cortical surface (Lundblad et al. 2002): injection site (I), 2.5 μl: tooth bar (TB): −2.3; anteroposterior (AP): −4.4; mediolateral (ML): −1.2; and dorsoventral (DV): −7.8; injection site (II), 2.0 μl: TB: +3.4; AP: −4.0; ML: −0.8; DV: −8.0. Moderate motor impairments including asymmetric posture and gait and reduced contralateral forelimb dexterity were generally apparent 2 wk after lesioning. One week after lesioning, animals were given daily doses of levodopa (6 mg/kg) for 2 wk. After 2 wk of treatment, the animals that showed moderate to high levels of dyskinetic symptoms after having been challenged with 12 mg/kg levodopa were implanted and included in the study.

Implantation surgery.

Implantations were performed under fentanyl-medetomidine anesthesia (0.3 mg/0.3 kg ip) at least 3 wk after 6-OHDA lesions. Microwire electrodes were implanted in both hemispheres. The eight regions targeted in each hemisphere were: rostral forelimb area (RFA; a rodent supplementary motor area), primary motor cortex (M1), dorsolateral striatum (DLS), dorsomedial striatum (DMS), GP, ventrolateral/ventroanterior nuclei of the thalamus (VL/VA; projecting to motor cortex), STN, and substantia nigra pars reticulata (SNr). Center coordinates in relation to bregma and the cortical surface were in the following structures: RFA: AP: +3.75, ML: ±2.0, DV: −1 (Neafsey and Sievert 1982); M1: AP: +1.5, ML: ±2.8, DV: −1.0 (Gioanni and Lamarche 1985); the DLS: AP: +0.2, ML: ±3.8, DV: −4; DMS: AP: +0.2, ML: ±2.8, DV: −4 (West et al. 1990); GP: AP: −1.0, ML: ±3, DV: −5.5 to −7.2 (Chen et al. 2011); VL/VA: AP: −2.6, ML: ±1.75, DV: −6.5 (Paxinos and Watson 2007); STN: AP: −3.5, ML: ±2.3, DV: −7.5 to −8.2 (Tai et al. 2003); and SNr: AP: −5.4, ML: ±2.4, DV: −7.8 to −8.8 (Wang et al. 2010). The implant was fixated with dental acrylic, which was attached to screws in the skull. After surgery, the anesthesia was reversed by atipamezole hydrochloride (5 mg/kg ip), and buprenorphine (0.5 mg/kg sc injection) was administered as postoperative analgesic. The animals were allowed to recover for 1 wk after surgery before testing commenced.

Experimental procedure.

During recording sessions, animals were placed in a transparent cylinder (250 mm in diameter), and their behavior was recorded with digital video in parallel with the electrophysiological recordings (synchronized via an external pulse generator; Master-8; A.M.P.I.). The same paradigm was used in each experiment. First, the rat was recorded for ∼30 min to establish baseline conditions. Second, the rat received an intraperitoneal injection with 12 mg/kg levodopa (levodopa methyl ester hydrochloride) and 12 mg/kg benserazide [serine 2-(2,3,4-trihydroxybenzyl)hydrazide hydrochloride]. The time point of this injection marks the beginning of the experimental timeline, i.e., t = 0 min. Dyskinesia developed 10–20 min after levodopa injection and reached its peak severity ∼60 min after levodopa injection. In experiments not involving further pharmacological intervention, the recordings continued until the dyskinesia diminished spontaneously (approximately 2–3 h after dyskinesia onset). Experiments involving additional drug treatment are described below.

Pharmacology.

Following levodopa injection and the manifestation of dyskinesia, a number of pharmacological substances were evaluated with respect to their antidyskinetic effects. Injection time points were chosen such that each drug would exhibit its therapeutic effect during the time of peak dyskinesia. Once the pharmacological effect of the serotonin 5-HT_1A receptor agonist 8-OH-DPAT (1 and 0.4 mg/kg ip, t = ∼60 min) had been established, the specificity of the intervention was verified by injection of the 5-HT_1A antagonist WAY-100,635 (0.5 and 0.4 mg/kg ip, t = ∼100 min), which effectively reversed the effect of 8-OH-DPAT. The neurophysiological and behavioral effects of the clinically used antidyskinetic drugs amantadine hydrochloride (an NMDA receptor antagonist, 50 and 50 mg/kg ip, t = ∼60 min), diazepam (a positive allosteric modulator at GABA_A receptors, 5 mg/kg ip, t = ∼60 min), and levetiracetam (a presynaptic calcium channel inhibitor, 80 and 120 mg/kg ip, t = ∼30 min) were also evaluated. All drugs used in this study were obtained from Sigma-Aldrich, Sweden, and doses were chosen according to previously published studies (Coppola et al. 2010; Kannari et al. 2001; Peixoto et al. 2005; Tronci et al. 2014).

Assessment of dyskinesia severity.

The scoring of dyskinesia was performed offline using an adapted version of the scoring methods of abnormal involuntary movements (AIMs) described by Lundblad and colleagues (2002). In summary, three different types of AIMs (orolingual, forelimb, and axial dyskinesia) were scored with respect to their severity for monitoring periods of 1 min with 5-min intervals. In addition, contraversive rotations with respect to the lesioned side were also quantified, as rotational behavior is correlated with general dyskinetic symptoms in this model (Breger et al. 2013). Forelimb and axial AIMs and rotations were rated on a scale ranging from 0 to 3 where 0 equals no dyskinesia and 3 equals continuous dyskinesia. Orolingual dyskinesia was less clearly detectable in the videos and was therefore scored as 1 when detected and 0 otherwise. The measures for all AIMs and rotations were normalized per category (0, 1) and then added together to produce a total AIM value (0, 4) for each assessed 1-min period. This combined value was taken to indicate the overall severity of the dyskinesia at any given time.

Recording electrodes.

For details on electrode design, see Ivica et al. (2014). In brief, formvar-insulated tungsten wires (33 μm) were arranged into 16 groups of arrays (8 per hemisphere; Fig. 1A) with 250-μm wire spacing in each horizontal dimension and fixed to the length corresponding to the implantation site for each group. Each array consisted of a minimum of 5 recording channels and 1 reference channel. All wires were connected to a custom-made printed circuit board and linked via connectors/adaptors to the preamplifier of the acquisition system. A 200-μm thick silver wire was attached to the skull screws and used as a ground connection from the animal to the recording system.

Signal acquisition.

Neuronal activity was recorded with the Neuralynx multichannel recording system using a unity gain preamplifier (HS-36; Neuralynx). LFP signals were filtered between 0.1 and 300 Hz and digitized at 1,017 Hz. Unit activities were filtered between 600 and 9,000 Hz and digitized at 32 kHz. Thresholds for storage of spiking events in each channel were automatically set to 3 SDs of the unfiltered signal.

Spike sorting.

Action potentials were sorted manually into unit clusters using Offline Sorter (Plexon). Waveform features used for separating the units were, e.g., valley and peak amplitude or the 1st 3 principal components (PCs) of all of the 32-element vectors defining the sampled waveforms for a given data set. A cluster was classified as single unit when <0.1% of the spikes in a defined cluster occurred within the refractory period (set to 1.6 ms) and as multiunit otherwise (Harris et al. 2000).

Frequency analysis of LFPs.

To emphasize local sources of the measured electrical potential (and to minimize effects of the choice of amplifier reference), bipolar LFP time series were computed offline from all unique pairs of electrodes from the same structure. For each of these time series, time-frequency spectrograms were calculated over the entire frequency range with a multitaper method (Pesaran 2008; 50% overlapping 8-s windows, time-bandwidth product 4, 7 tapers) implemented in Chronux 2.0 (Mitra and Bokil 2008). Power line noise (50 ± 2 Hz and 1st harmonic at 100 ± 2 Hz) was removed from the power spectral densities. To identify oscillations in a certain part of the frequency spectrum better, each individual power spectrum was normalized to the pink-noise background. That is, the noise background was estimated once for each 8-s window and for each bipolar channel separately. Because of the complexity of the data, it was not possible to pick enough frequency bands with pure pink noise in all structures and conditions manually to get unbiased estimates of the noise background. Instead, we divided the whole frequency axis (from 1 to 200 Hz) into 20 logarithmically spaced bands (1–1.3, 1.3–1.7, …, 151.3–200) and used the median power of each band for the fitting of the pink-noise power curve, S(f) = b/f^a. The pink-noise normalization allowed us to describe deviations from the pink-noise floor in terms of the unit dB_pink, defined as:

$$S_{\text{dB}\left(\text{pink}\right)}\left(f\right)=10{\log }_{10}\frac{S\left(f\right)}{S_{\text{pink}}\left(f\right)},$$

where S(f) and S_pink(f) have the dimension power per frequency (i.e., V²/Hz) and S_dB(pink) is expressed in the dimensionless unit dB_pink.

As a final step, an average spectrogram was calculated for each structure based on the pink-noise-corrected spectrograms for each individual local bipolar LFP time series.

In Figs. 2–4 and 7, the obtained spectrograms were further averaged over time for each behaviorally classified state to obtain average spectra for the different states.

Fig. 2.

Changes in neurophysiological activity patterns in the STN and M1 with the onset of dyskinesia. A, top: examples of local field potential (LFP) spectrograms from recordings in the STN and M1 in the lesioned hemisphere during a 90-min period including a time period before and following the onset of dyskinesia (dashed line; t = 0 min corresponds to time point of levodopa injection). Bottom: close-up of the low-frequency range of the spectrograms shown in the top row [power is expressed in decibels relative to the estimated pink-noise floor (dB_pink)]. B: time-averaged spectra from 9 recordings (≥20 min per state and recording) for the parkinsonian period (gray, individual recordings; black, average) and the dyskinetic period (pink, individual recordings; red, average).

Fig. 4.

Spectral state differences per structure divided by animal. The average LFP spectral difference vectors in the recorded structures for: top, control vs. PD; middle, control vs. dyskinesia; and bottom, [control + levodopa (LDA) vs. dyskinesia] over all recordings averaged per animal. Note that the spectral difference (control vs. dyskinesia) is shown rather than (control vs. dyskinesia)_ortho (to illustrate the true spectral difference without orthogonality constraints). Colored dots indicate significant differences between the compared states for the corresponding frequency bin and structure (Wilcoxon rank sum, P < 0.05, Bonferroni-corrected for multiple tests).

Fig. 7.

Systemic treatment with a 5-HT_1A receptor agonist alleviates dyskinesia and alters the neurophysiological state. A: severity of dyskinesia scored during 1-min periods once every 5 min (marked by crosses). Dashed lines indicate times of drug injections (levodopa was administered twice in this experiment to reach the dyskinetic state, represented by the 1st 2 lines). AIM, abnormal involuntary movement; au, arbitrary units. B: spectrogram from all recorded structures in the lesioned hemisphere showing the relative change in LFP spectral contents throughout an example experiment where a dyskinetic rat was treated with 8-OH-DPAT (0.4 mg/kg ip at t = 123 min) to reduce dyskinesia. This drug effect was subsequently reversed by treatment with WAY-100,635 (0.4 mg/kg ip at t = 163 min). C: cellular activity showed clear differences between states (color code represents deviation from the mean firing rate across all 4 conditions for each unit; units are ordered in rows according to the mean firing rate during the nontreated parkinsonian state, and the colored boxes to the left of each unit indicates structure with same color codes as in D). D: the mean differences in LFP spectral contents between the control condition and the nondyskinetic 8-OH-DPAT-treated state shown in B, summarized for each structure separately.

Systems-level neurophysiological states.

To visualize and identify systems-level neurophysiological states, we relied on the average, pink-noise-normalized spectrograms that were calculated for each structure in each recording session, as described above. Each such spectrogram consists of a series of individual spectra reflecting the frequency content between 2 and 120 Hz with 0.5-Hz resolution in that structure during an 8-s window. The electrophysiological samples (made up of 8-s recording segments) included from each state were selected from within a time interval during a steady-state as defined by behavioral criteria (dyskinesia score; see Table 1 for a summary of the number of samples obtained in each animal and state). Samples were defined such that one sample contained the concatenated spectra from all structures for one such 8-s window. Thus, for one recording session, the number of samples (n) becomes equal to the number of 8-s windows, and the number of variables (p) becomes equal to the number of frequency bins (2× the frequency range) multiplied by the number of structures. Pooling all recording sessions in one animal results in a number of samples (n_pooled) equal to the number of 8-s windows in all of these recording sessions, whereas p stays the same.

Table 1.

Data size and classification performance

	Animal 1	Animal 2	Animal 3	Animal 4
Number of Samples Used
Control	4,281	1,661	509	434
Control + levodopa	4,311	1,346	217	194
Parkinson's disease	4,281	1,661	509	434
Dyskinesia	4,311	1,346	217	194
Levetiracetam	449
Amantadine	1,723
8-OH-DPAT	973
WAY-100,635	298
Diazepam	749
Comparison of Classification Performance Between Individual Structures and All Structures
RFA	0.7242	0.5793	0.5970	0.9865
M1	0.6650	0.8387	0.6233	0.9570
DMS	0.5725	0.6875	0.6306	0.8631
DLS	0.5710	0.7220	0.5763	0.9100
GP	0.5743	0.5148	0.8957
Thalamus	0.6182			0.9005
STN	0.5152	0.3755	0.6366	0.8997
SNr	0.4305	0.4487	0.7746	0.9514
All	0.9910	0.9782	1	1

A summary of the total number of 8-s samples simultaneously collected in all structures per state (rows) and animal (columns) is shown above. Classification performance for 4 states (control, control + levodopa, Parkinson's disease, and dyskinesia) was evaluated for each and all structures (rows) in all animals (columns) using the 30 1st principal components and is presented below as the fraction of correctly classified states. Note that the best performance was always reached when all structures were used. 8-OH-DPAT, 5-HT_1A receptor agonist; WAY-100,635, 5-HT_1A antagonist; RFA, rostral forelimb area (a rodent supplementary motor area); M1, primary motor cortex; DMS, dorsomedial striatum; DLS, dorsolateral striatum; GP, globus pallidus; STN, subthalamic nucleus; SNr, substantia nigra pars reticulata.

A first aim was to obtain a two-dimensional visualization of the samples describing the spectral differences to the control state along the axes control vs. PD and control vs. dyskinesia. The following steps have been taken. The data are normalized such that the mean and SD over each variable become 0 and 1, respectively. The samples normalized in this way are denoted by s_i, i = 1, …, n. Next, the origin of the coordinate system is shifted to become equal to the cluster center of the control state, i.e., the mean over all samples belonging to the control state (s̄_control) is subtracted from each sample: s̃_i = s_i − s̄_control. By this, each shifted sample (s̃_i) describes the spectral differences to the mean control state. To obtain a two-dimensional representation of the data, an x- and y-axis are defined to point from the cluster center of the control state, i.e., the origin of the shifted coordinate system, to the PD and dyskinesia cluster center, ${\overline{\tilde{\mathit{s}}}}_{\text{PD}}$ and ${\overline{\tilde{\mathit{s}}}}_{\text{dys}}$, respectively. However, for the y-axis, its projection on an axis orthogonal to the x-axis will be shown. The projection onto the x- and y-axis is furthermore normalized such that the PD and dyskinesia cluster centers will have an x- and y-value equal to 0 and 1, respectively. Mathematically, the value on the x-axis for s̃_i can be obtained from:

$$x_i={\tilde{\mathit{s}}}_i\cdot \frac{{\overline{\tilde{\mathit{s}}}}_{\text{PD}}}{{\Vert {\overline{\tilde{\mathit{s}}}}_{\text{PD}}\Vert }_2^2},$$

and the value on the orthogonal y-axis can be obtained from:

$$y_i^{\text{ortho}}={\tilde{\mathit{s}}}_i\cdot \frac{{\overline{\tilde{\mathbf{s}}}}_{\text{dys}}^{\text{ortho}}}{{\Vert {\overline{\tilde{\mathbf{s}}}}_{\text{dys}}^{\text{ortho}}\Vert }_2^2}\hspace{0.17em}\text{with}$$

$${\overline{\tilde{\mathit{s}}}}_{\text{dys}}^{\text{ortho}}={\overline{\tilde{\mathit{s}}}}_{\text{dys}}-\frac{{\overline{\tilde{\mathit{s}}}}_{\text{PD}}\cdot {\overline{\tilde{\mathit{s}}}}_{\text{dys}}}{{\Vert {\overline{\tilde{\mathit{s}}}}_{\text{PD}}\Vert }_2^2}{\overline{\tilde{\mathit{s}}}}_{\text{PD}},$$

where · denotes the dot product, and || ||₂ denotes the l₂-norm. Figures 3A and 9D show the results of this visualization. Furthermore, in Fig. 3B, the vectors ${\overline{\tilde{\mathit{s}}}}_{\text{PD}}$ and ${\overline{\tilde{\mathit{s}}}}_{\text{dys}}^{\text{ortho}}$ are illustrated, whereas Fig. 7D illustrates the eight structure components that make up the vector ${\overline{\tilde{\mathit{s}}}}_{\text{8-OH-DPAT}}$ (i.e., the cluster center of the 8-OH-DPAT treatment state in the shifted coordinate system). Finally, for Fig. 3, C and D, and Fig. 5, the above analysis has been performed for each structure separately (i.e., without concatenating the spectra from all structures), and the distributions of x_i for the control and the PD cluster are shown in Fig. 3C, whereas Fig. 3D shows the distributions of y for the control and the dyskinetic cluster. Note that we chose not to use y_i^ortho in Fig. 3D but rather the more intuitive distribution defined by:

$$y_i={\tilde{\mathit{s}}}_i\cdot \frac{{\overline{\tilde{\mathit{s}}}}_{\text{dys}}}{{\Vert {\overline{\tilde{\mathit{s}}}}_{\text{dys}}\Vert }_2^2},$$

which exclusively depends on the difference between the control and the dyskinetic states.

Fig. 3.

Systems-level neurophysiological states associated with parkinsonism and dyskinesia. A: systems-level state descriptions in 4 rats based on LFP recordings in the corticobasal ganglia-thalamic loop [dark blue, control; black, Parkinson's disease (PD); red, dyskinesia; light blue, control with levodopa]. The x-axis denotes the direction in LFP spectral space where the difference between the control condition and the parkinsonian state is the largest, and the y-axis represents the largest difference between the control and dyskinetic state orthogonal to the x-axis [(control vs. dyskinesia)_ortho]. Note the close clustering of data points from each state (each small dot represents the state coordinate during an 8-s period, and shaded clouds denote dot densities) and the great similarity of the states in separate recordings (filled triangles indicate cluster centers for the states in each recording; animal 1: n = 9; animal 2: n = 4; animal 3: n = 1; animal 4: n = 1; classification performances were for the 4 animals: 0.9910, 0.9782, 1, and 1; all pairwise comparisons of cluster medians were significant, P < 0.001, Wilcoxon rank sum). B: the average spectral differences in the 8 structures for control vs. PD and (control vs. dyskinesia)_ortho over all 9 recordings in animal 1. C and D: histograms illustrating the state separability in each structure with data from all 9 recordings. C: control vs. PD. D: control vs. dyskinesia. The distributions were obtained by projecting the data onto the 1 dimension represented by the spectral difference vector.

Fig. 9.

Systems-level characterizations of pharmacological interventions alleviating dyskinesia. A: reduction in normalized dyskinesia scores following systemic treatment in the same rat with 4 different drugs in 7 separate recordings. Wilcoxon signed-rank tests for significant effects on individual AIM scores between pre- and posttreatment showed significant reductions (P < 0.05, after Bonferroni corrections with n = 16 comparisons) for orolingual dyskinesia (OL): amantadine 2, 8-OH-DPAT 2; forelimb dyskinesia (FL): amantadine 2, 8-OH-DPAT 1, diazepam; and axial dyskinesia (Ax): amantadine 2, 8-OH-DPAT 1 and 2, diazepam. B: overview of the corresponding systems-level neurophysiological states induced by the different pharmacological interventions based on the spectral contents of recorded LFPs. Note that each drug clusters in a separate region of the illustrated space spanned by the 1st 3 principal components (PCs; classification performance with 3 PCs was 0.82; cf. Fig. 10; all pairwise comparisons of cluster medians were significant, P < 0.001, Wilcoxon rank sum). C: cluster classification performance shown as a function of number of brain structures included in the electrophysiological measurement [red, average value for all possible combinations of x-structures; blue, best combination of x-structures (the composition of the best combinations are listed for 1–6 structures); classification performance when all 8 structures were used reached 99.94% for the n = 5,421 samples with 8 states; this performance was significantly higher than what was obtained using fewer structures except for n = 7 structures, P < 0.05, Wilcoxon signed-rank test with Bonferroni correction for multiple comparisons]. D: representation of the systems-level state induced by each of the drugs in 2D space with axes defined by the main spectral differences control vs. PD and (control vs. dyskinesia)_ortho. r² Between the individual AIM scores shown in A and mean coordinate value in the (control vs. dyskinesia)_ortho dimension of the states shown in D were: OL = 0.729, FL = 0.777, Ax = 0.621, rotation (Rot) = 0.566, total = 0.724.

Fig. 5.

Histograms illustrating the state separability of all recordings shown per animal. Top: control vs. PD. Bottom: control vs. dyskinesia. The distributions were obtained by projecting the data onto the 1 dimension represented by the spectral difference vector (for example, the vector pointing from the center of the control cluster to the PD cluster). Three of the animals were used for evaluation of electrical microstimulation in a separate set of experiments and are consequently lacking recording electrodes in that structure, Th (n = 2) and GP (n = 1). Notably, this missing information was largely compensated for by the parallel recordings in the other structures as indicated by the histograms in the rightmost column.

Quantification of state separability.

To quantify the separation between states in terms of classification performance, it was necessary first to reduce the dimensionality of the data using PC analysis (PCA). We used the singular value decomposition PCA algorithm with variance weighting (MATLAB). Generally speaking, given a data set with n samples and p variables, all samples can be represented in a p-dimensional coordinate system. PCA can be thought of performing a high-dimensional rotation of this coordinate system according to:

$$\mathit{T}=\mathit{S}\mathit{W},$$

where S and T are n × p matrices representing the samples in the original and the rotated coordinate system, respectively, and W is the p × p rotation or weight matrix. In PCA, the weight matrix W is constructed such that the p variables in the new coordinate system are uncorrelated over the data set. Furthermore, the first variable in this coordinate system, i.e., the first PC, will, by definition, capture the most variance of the data set, the second PC will capture the second most variance in a perpendicular dimension to the first, and so forth. Thus, would one only keep the first two PCs, one would automatically obtain a representation of the data set in the two-dimensional plane in which the data are most spread out, allowing a convenient visualization of the high-dimensional data and, e.g., the identification of clusters. For example, in Fig. 9B, the first three PCs, obtained from applying PCA to the pooled data in one animal, are shown. Such a visualization can complement visualizations based on the method described in the previous section where differences between selected states are emphasized by projection onto the state difference vector.

After dimensionality reduction with PCA, a Gaussian mixture model was fitted to the data (MATLAB fitgmdist function). The number of Gaussian components in the model was set to be equal to the number of experimental conditions (e.g., control, control and levodopa, PD, and dyskinesia), and the starting conditions for the optimization (means, covariances, and mixing proportions) were calculated by assigning samples from the same experimental condition to one Gaussian component. The performance of the model was then estimated by assigning each sample to the Gaussian component with the largest posterior probability (weighted by the component probability) and calculating the average number of correct classifications. Generally, the classification performance improved as more PCs were added until a plateau was reached (cf. Fig. 10). Chance level of correctly assigning a data point to 1 of n states corresponds to p = 1/n. As a compromise between the risk of overcompressing the data and the cost of performing heavy calculations, we settled on consistently using 30 PCs for all quantifications of classification performance.

Fig. 10.

Classification performance as a function of the number of PCs used. The classification performance for the 8 states shown in Fig. 9B is plotted as a function of the number of PCs used to represent the full space. The black line shows the performance when all 8 structures are used together. The colored lines show the performance when only data from a single structure are used. The dashed line represents chance level of correctly assigning a data point to 1 of the 8 states. In this comparison, each structure was represented by the average LFP spectral contents of all electrode pairs in the structure. It can be noted that despite that the number of electrodes differed (range: 5–9), classification performance was similar using the different individual structures. As expected, classification performance was higher when combining the information in all structures. It was also confirmed that the number of PCs used (n = 30) to compress the data before numerical comparisons of state separability (e.g., in Fig. 9C) was sufficiently high to avoid significant information loss (the performance curves appear to have plateaued much earlier).

To complement classification performance as a measure of state separability, classic frequentist hypothesis tests were performed to test for significant differences between states. For each possible state pair, the data were projected orthogonally onto a line going through the means of the 2 distributions, e.g., the line defined by the vector s̄_PD − s̄_dys. The distributions (now 1-dimensional) were then tested with a standard Wilcoxon rank sum test corrected for multiple comparisons (Bonferroni).

Tissue preparation and immunostaining.

Animals were anesthetized with a lethal dose of sodium pentobarbital (100 mg/kg), and heads were fixated in 4% paraformaldehyde. Brains were removed, postfixed in paraformaldehyde overnight, and then transferred to 30% sucrose PBS solution at 4°C overnight for cryoprotection. Using a cryostat, tissue was sectioned in 50-μm-thick coronal slices and mounted on charged slides. The placement of electrodes was verified in coronal brain sections stained with cresyl violet (CV) in two animals. The extent of the lesions was confirmed by tyrosine hydroxylase (TH) immunohistochemistry.

CV staining.

Sections were stained with 0.1% CV powder in distilled water (dH₂O) and 0.3% glacial acetic acid solution for 5 min. Sections were then rinsed for 1 min in dH₂O and dehydrated with 70, 95, and 99.5% EtOH for 1 min each and then immersed in 100% xylene for 5 min (×2) before mounted with DPX mounting media.

TH staining and quantification.

Brain sections were washed in 0.01 M PBS (5 min), 0.3% hydrogen peroxide diluted in methanol (20 min), and 0.05% PBS/Tween (5 min) and then were incubated in 10% normal goat serum for 30 min followed by incubation with primary antibody rabbit anti-TH (1:500; Chemicon) overnight at room temperature. On the following day, sections were rinsed in 0.05% PBS/Tween (5 min) and incubated with biotinylated goat anti-rabbit (1:200; Vector Laboratories) for 2 h. After that, all sections were rinsed in PBS/Tween 0.05%, incubated in Avidin/Biotin Complex (ABC Kit; Vector Laboratories) for 1 h, and stained with 3,3′-diaminobenzidine and H₂O₂.

TH striatal optical densitometry was assessed using ImageJ software (National Institutes of Health) as described previously (Fuentes et al. 2009) in areas adjacent to the striatal recording sites. The optical density of the ipsilateral corpus callosum was used as staining background and was subtracted from striatal values before comparison.

Statistical methods.

All statistical tests used to assess significant group difference are specified in results and in the respective figure legends.

Materials, data, MATLAB code, and protocols used in this publication are readily available from the authors on request.

RESULTS

Experiments performed.

To clarify which neurophysiological activity patterns are associated with parkinsonism and dyskinetic states in PD, we performed parallel multistructure neuronal recordings in 8 different parts of the corticobasal ganglia-thalamic loop using the described novel methodology. In total, 15 separate recording sessions were performed in 4 unilaterally 6-OHDA-lesioned dyskinetic rats (repeatability was evaluated by performing 9 separate recordings in the same subject in experiments performed several weeks apart and reproducibility by performing similar recordings in 4 different subjects). Postmortem TH staining adjacent to the recording electrodes showed a complete loss (100%) of dopaminergic terminals in posterior parts of the striatum ipsilateral to the lesion with some remaining terminals in anterior areas (average striatal denervation ∼74%). In 7 of these experiments, additional pharmacological interventions aimed at reducing dyskinetic symptoms were also investigated as a proof of principle for the use of the developed technology in characterization of experimental treatment of disease.

Recordings in STN/M1 confirm previously reported changes in neuronal activity patterns.

From the obtained recordings, we could confirm the presence of narrow-band high-frequency gamma oscillations in M1 (as previously documented in rodents) and theta oscillations in STN (as previously documented in humans) after the transition from the parkinsonian to the dyskinetic condition following levodopa treatment (Fig. 2, A and B; Alonso-Frech et al. 2006; Halje et al. 2012). A notable difference between these two phenomena was, however, that narrow-band gamma oscillations at no instance were observed in either the intact hemisphere during dyskinesia or the lesioned hemisphere of non-levodopa-treated animals as opposed to theta oscillations that were more abundantly present (in particular, during periods of increased motor activity). From the spectrograms presented in Fig. 2A, it is clear that the spectral contents in the parkinsonian condition varies over time [examinations of the video recordings revealed that these changes were associated with changes in behavioral state of the animal, in agreement with previous studies (Avila et al. 2010; Brazhnik et al. 2014; Delaville et al. 2014)]. In contrast, following a transient frequency-tuning at the onset of dyskinesia, the spectral characteristics in the dyskinetic state were relatively stable throughout the dyskinetic period.

Within an individual, the theta/gamma power changes in STN/M1 were consistent across recordings (average power spectra from 9 example recordings from the animal shown in Fig. 2A are presented in Fig. 2B). On the other hand, between rats, peak frequencies within the different bands were found to vary somewhat. On average over all of the recordings, there was an increase in LFP power for the theta band (3–9 Hz) when comparing the dyskinetic state to the baseline before levodopa administration [Wilcoxon signed-rank tests revealed that the 1.9-dB increase in STN was significant (P = 0.0004), whereas the 0.3-dB increase in M1 was weakly significant (P = 0.05) and did not survive correction for multiple comparisons]. For the gamma band (65–100 Hz), a significant increase was only found in M1 (Wilcoxon signed-rank tests revealed +0.1 dB, P = 0.4, for STN and +3.0 dB, P = 0.0004, for M1). In this context, it is also interesting to note that a comparison to the levodopa-treated control side revealed that the theta increase following levodopa administration may partly be related to the induced increase in motor activity in contrast to the changes in gamma, which are more specific to the dyskinetic state. Wilcoxon signed-rank test for differences in the increase of power in the theta band between the STN in the two hemispheres before and after treatment showed that the side difference was not quite significant [P = 0.054 after Bonferroni correction for multiple comparisons (n = 4)]. For the gamma band, on the other hand, the corresponding power increase in M1 after treatment was significantly higher in the lesioned hemisphere compared with control (P = 0.0032).

Multistructure recordings reveal systems-level brain states.

Based on these confirmatory findings in M1 and STN, it is expected that abnormal activity patterns should arise under similar conditions also in other parts of the highly interconnected circuits making up the corticobasal ganglia-thalamic loop. Moreover, although these specific frequency bands that have been highlighted in earlier studies indeed showed clear changes in relation to the transition from parkinsonian to dyskinetic state, it is evident that other parts of the frequency spectrum also displayed changes (which appeared to differ between M1 and STN; Fig. 2, A and B). In the subsequent analyses, we therefore included all simultaneously recorded structures from both hemispheres and did not restrict LFP analyses to delimited frequency bands. Recordings from four different conditions were analyzed: 1) from the intact hemisphere OFF-levodopa, representing the control condition; 2) from the lesioned hemisphere OFF-levodopa, representing the parkinsonian state; 3) from the lesioned hemisphere ON-levodopa during periods with dyskinesia, representing the dyskinetic state; and 4) from the intact hemisphere ON-levodopa, representing a second control condition in the drug-treated state. Recordings from these different conditions were divided into separate data sets and analyzed individually for each rat. LFPs and the firing rates of individual neurons were both examined. For LFPs, frequency power spectra (based on the spectral contents between 2 and 120 Hz with a 0.5-Hz resolution) were calculated during 8-s sample periods for all brain structures. To describe the neurophysiological state of an animal at different time points during the experiment, the power spectra from the different structures were combined into a single vector, thereby essentially creating a unique coordinate in this high-dimensional space for each 8-s time bin. Similarly, for firing rates, the neurophysiological state was also described by a unique coordinate for each 8-s period created from the vector comprising the average firing rate of all recorded neurons during each sample period.

In all recorded experiments, behavioral observations confirmed that animals quickly transitioned into a stable severely dyskinetic state following levodopa treatment and remained in this condition with uninterrupted dyskinesia for an average time period of 160 ± 22 min (corresponding to the reported period of elevated levodopa concentrations; Carta et al. 2006) unless other pharmacological interventions were carried out. This was expected given that medial forebrain bundle lesions are known to cause a severe model of PD where practically no therapeutic window for dopamine replacement therapy remains following a brief period of levodopa treatment (Winkler et al. 2002). When pooling LFP data from parkinsonian and dyskinetic animals from multiple recordings and plotting the coordinates in a two-dimensional space chosen to facilitate the comparison of the two pathological states (i.e., where the x-axis represents the difference vector between the parkinsonian and control state and the y-axis the difference vector between the dyskinetic and control state in the direction orthogonal to the x-axis), it became obvious that data sampled from time periods belonging to each specific state clustered in separate parts of the plane (Fig. 3A). Moreover, this LFP-based state description proved to be very robust across experiments performed in each animal (see Table 1 for details on state classification performance).

To get a better understanding of the underlying physiological differences separating the states, the spectral content in each structure was analyzed in further detail. In Fig. 3B, the mean of the spectral differences that chiefly separate the control from the parkinsonian state and the dyskinetic from the control state in Fig. 3A (i.e., the axes spanning the plane) is plotted for all eight brain structures in animal 1, which has the largest number of recordings. Note the increase in relative LFP power in the beta band in several structures in the parkinsonian state as well as the theta and gamma peaks in the dyskinetic state (Fig. 3B, top and bottom, respectively). However, certain variability between subjects in terms of the exact difference spectra that separate the states was also observed (Fig. 4). These interindividual differences could be expected given inherent variability between individuals relating to brain circuit anatomy, the exact locations of the recording electrodes, signal-to-noise levels, etc. On the other hand, the great similarities in the state representations (Fig. 3A) show that comparisons of similar states across subjects can be made even though the absolute differences between states in terms of LFP spectral contents may vary between individuals.

To investigate the relative contribution from the eight different brain structures for state separation, we also analyzed state classifications based on the LFPs recorded in each single structure (for details on calculations, see methods). In Fig. 3C, the state separations control vs. PD and control vs. dyskinesia obtained for each structure are shown separately. These analyses show that the LFP spectral contents in, for example, cortex and STN constitute relatively reliable biomarkers for these three states [see also Halje et al. (2012)]. Nonetheless, state separation for any individual structure was clearly not as robust as the multistructure data: for example, whereas the average classification performance for all recordings was 98.6% using data from all structures, it was reduced to 85.6% when using data from M1 and STN only (cf. Fig. 2B), which corresponds to a >10-fold higher error rate than when all 8 structures are included (histograms for all animals are included in Fig. 5 and classification performance in Table 1).

We next analyzed changes in neuronal activity. Here, the requirement of sampling unit activity from the same neurons across states limits comparisons to changes observed within each structure across different experimental conditions. Hence, in the unilateral 6-OHDA PD model, direct comparisons between the control and the parkinsonian/dyskinetic state cannot be obtained with single-cell resolution. Even so, when analyzing unit activity of cells located in the lesioned hemisphere, we found that several neurons clearly altered their firing rates during dyskinesia compared to the parkinsonian state (increased: RFA 2/6, DMS 7/10, DLS 6/11, GP 0/9, thalamus 9/9, STN 2/9; decreased: RFA 3/6, DMS 3/10, DLS 3/11, GP 9/9, thalamus 0/9, STN 5/9; P < 0.05, Wilcoxon rank sum test). Consequently, these 2 states could be reliably separated in a manner similar to the LFP-based clustering (in the corresponding multivariate analysis across the 2 states, i.e., ON-/OFF-levodopa). See Fig. 6 for example state plots based on unit data (average classification performance for PD vs. dyskinesia was in this case 99.3%).

Fig. 6.

State plots based on changes in neuronal firing rates. Left: heat plots of all individual unit activities from the lesioned hemisphere during different states. Each row on the y-axis represents the activity of a unit throughout an experiment, normalized to its respective maximal firing rate [max(r); color codes denoting recording structures as in Fig. 3B]. Vertical white lines indicate times of drug injections during the recording and onset of dyskinesia (Dysk. onset; based on manual behavioral scoring). Ldopa, levodopa; 8OHDPAT, 5-HT_1A receptor agonist; Way100635, 5-HT_1A antagonist. Right: systems-level state descriptions based on unit activity in the corticobasal ganglia-thalamic loop in the lesioned hemisphere. The x-axis denotes the direction in unit activity space where the difference between the parkinsonian and dyskinetic state is the largest, and the y-axis represents the largest difference between the parkinsonian and drug-induced state orthogonal to the x-axis. The firing rate difference between PD and dyskinesia for units in the respective structures were, expressed in z-scores [median/interquartile range (IQR)]: RFA: 0.56/0.63, DMS: 0.63/1.19, DLS: 0.58/1.07, GP: 0.79/1.93, Th: 1.33/0.93, STN: 0.68/0.78. Classification performance of the 3 states in this 2D projection were for the 4 panels: 0.9708, 0.9645, 0.8556, and 0.7641. All pairwise comparisons of cluster medians were significant, P < 0.001, Wilcoxon rank sum.

Ameliorating dyskinetic symptoms using a serotonin agonist.

For the vast majority of Parkinson patients, dopamine replacement therapy effectively improves a range of symptoms and remains the therapeutic approach of choice (PD Med Collaborative Group et al. 2014). The possibility to prolong the levodopa treatment period before complications arise by reducing drug-induced dyskinetic symptoms has therefore attracted a lot of interest in recent years (Crosby et al. 2003; Huot et al. 2013; Olanow et al. 2000). One such approach is the use of 5-HT agonists aiming to control the efflux of dopamine from serotonergic terminals of dorsal raphe neurons by stimulation of 5-HT autoreceptors (Carta et al. 2007; Svenningsson et al. 2015). The rationale for this method stems from the notion that dyskinesia is partly caused by a dysregulation in dopaminergic signaling and that serotonergic terminals synthesizing dopamine via aromatic l-amino acid decarboxylation (AADC) release it in an uncontrolled manner (the AADC enzyme is in serotonergic neurons responsible for the synthesis of 5-HT but can also convert levodopa into dopamine). Accordingly, a pharmacological intervention targeting presynaptic 5-HT receptors on these neurons could potentially harness the uncontrolled synaptic release of dopamine. To evaluate the potential of this approach from a systems-level point of view, we first administered the 5-HT_1A agonist 8-OH-DPAT systemically during peak dyskinesia and subsequently reversed the effect of the drug by treatment with the 5-HT_1A antagonist WAY-100,635 ∼40 min later. Dyskinetic symptoms were quantified during different phases of the experiment with respect to prevalence of the AIMs observed (Lundblad et al. 2002).

The 5-HT_1A agonist was found to ameliorate dyskinesia effectively, and this effect was fully reversible by the antagonist [Fig. 7A; mean normalized scores (0, 1): dyskinesia = 0.72, 8-OH-DPAT = 0.01, WAY-100,635 = 0.75; Kruskal-Wallis: P < 0.001; Dunn posttest for group differences: dyskinesia vs. 8-OH-DPAT, P < 0.01; 8-OH-DPAT vs. WAY-100,635, P < 0.001; dyskinesia vs. WAY-100,635, not significant based on dyskinesia scores >5 min after each injection (2nd injection for levodopa)]. It was noted, however, that whereas the dyskinesia was practically eliminated, other behavioral abnormalities appeared to be present in the 8-OH-DPAT-treated state (i.e., an abnormally flat body posture and recurring forepaw movements, resembling previous observations connected to excessive serotonergic stimulation; Jacobs 1974). The recorded neurophysiological signals in the 8 different brain regions revealed a clear shift away from the dyskinetic state in both LFPs [Fig. 7B; MANOVA (ANOVA with frequency bands as dependent variables), P < 0.001; mean distance in 1st canonical dimension were: dyskinesia vs. 8-OH-DPAT = 24.3, 8-OH-DPAT vs. WAY-100,635 = 17.1, dyskinesia vs. WAY-100,635 = 7.1; and all groups were significantly different from each other, P < 0.001, Wilcoxon rank sum; cf. Dupre et al. (2013)] and unit activity (Fig. 7C; MANOVA, P < 0.001; mean distance in 1st canonical dimension were: dyskinesia vs. 8-OH-DPAT = 10.9, 8-OH-DPAT vs. WAY-100,635 = 9.5, dyskinesia vs. WAY-100,635 = 1.4; and all groups were significantly different from each other, P < 0.001, Wilcoxon rank sum). However, a closer comparison to the control state revealed that certain differences remained between the 8-OH-DPAT-treated state and the control condition. In particular, low-frequency oscillations (delta/theta and beta) showed a deviant pattern (Fig. 7D; see Fig. 8 for corresponding spectrograms from all structures for the intact hemisphere). These remaining differences between the 8-OH-DPAT-treated state and control conditions, together with the observation that normal behavior was not fully reinstated with this drug, suggest that aspects of the motor behavior other than those captured with the dyskinesia score are relevant for the interpretation of the electrophysiological state in this case.

Fig. 8.

LFP spectrograms from the contralateral, nonlesioned hemisphere recorded in parallel with data shown in Fig. 7B with 8-OH-DPAT administered as a dyskinesia-reducing agent. Vertical lines indicate times of drug injections during the recording. Thal, thalamus.

Investigating drug effects in a systems-level neurophysiological state space.

Because pharmacological manipulations targeting 5-HT_1A receptors clearly have the potential to reduce dyskinesia but nevertheless induce neurophysiological activity patterns that in some parts of the brain differ considerably from the control state, it would be of relevance to characterize the drug-induced state at a systems level. More generally, condensed systems-level descriptions could conceivably offer a more straightforward way to compare complex brain states following interventions that involve diverse changes in different parts of the CNS and in different neurotransmitter systems. Thus, to test the potential of the developed technology for the experimental evaluation of drug candidates and other novel therapeutic interventions, we next characterized brain states, based on LFPs recorded in the eight structures in the same animal, following treatment with 8-OH-DPAT and three other drugs with putative antidyskinetic effects: amantadine, levetiracetam, and diazepam (Pahwa et al. 2006; Pourcher et al. 1989; Stathis et al. 2011). In parallel, behavioral assessments of the reduction of dyskinesia was quantified for all four drugs. The drugs were administered systemically to reach maximum effect at the time point of peak dyskinesia (where the animals displayed severe dyskinesia corresponding to 79 ± 4% of the maximum compound dyskinesia score). The alleviation of dyskinetic symptoms differed between the drugs [P < 0.001, Kruskal-Wallis; ranging from no detectable effect for levetiracetam (Wolz et al. 2010) to clear alleviation of symptoms, e.g., 8-OH-DPAT], and in some cases the effect also varied either during or across experiments. In specific, following diazepam treatment, intermittent periods of AIMs were present even though dyskinesia was otherwise almost completely abolished, and in the case of amantadine the alleviation of dyskinesia was relatively weaker in 1 of the 2 experiments performed (Fig. 9A).

To get an overview of the entire data set, LFP data from 6 recordings (1 experiment was excluded due to poor recording quality) were 1st represented in a common PC space spanned by the 1st 3 PCs. Remarkably, unique and clearly separable clusters were found for each of the drugs even in this low-dimensional representation (Fig. 9B; for details on calculations, see methods). The state separation was quantified with a classifier with 8 states that achieved near-perfect classification performance (>99.9%) using 30 PCs (Fig. 10). This tight clustering of neurophysiological states induced by the same treatment in separate experiments performed weeks apart clearly indicates that the drug-induced systems-level states were specific and robust.

To clarify further to what extent activity patterns in different brain structures contributed to the combined state description, we next analyzed how well the different drug-induced states could be separated using only a subset of the recorded structures. Hence, the classification performance of the 8 states shown in Fig. 9B was calculated for all 255 possible combinations of structures (Fig. 9C). As expected, a higher number of brain structures generally improved classification performance. It was also noted that although motor cortex and STN together turned out to be the most informative pair, different combinations of structures resulted in the most accurate state classification depending on the total number included because the fraction of shared (redundant) information in a given structure will depend on which other structures are included.

To compare the contribution from specific frequency bands, we analyzed how well the states could be separated using only the theta band (3–9 Hz), only the beta band (10–35 Hz), only the gamma band (65–100 Hz), or any combination with 2 or 3 of these bands. To make the comparison fair, we used PCA to reduce the dimensionality to 8 in all test cases before classifying. The classification performances were: theta = 24%, beta = 20%, gamma = 25%, theta and beta = 22%, theta and gamma = 35%, beta and gamma = 29%, and theta, beta, and gamma = 32%. This should be compared with the classification performance of 96% obtained when using the full spectrum (when also reduced to 8 PCs).

Having confirmed that each drug-induced state could be reliably identified based on multistructure LFP data, we wanted to plot the different states using the same two-dimensional space as in Fig. 3A to facilitate the comparison to the 2 pathological states that the pharmacological interventions were aimed to alleviate (i.e., PD and dyskinesia). To enable us to pool data from different subjects despite potential interindividual differences, the parkinsonian and dyskinetic states were used as reference states defining the subspace onto which other states were then projected (the robustness of this approach was initially verified in a control experiment by training a PD/dyskinesia/amantadine classifier in 1 rat and cross-validating it in another rat with or without calibration to the PD and dyskinesia reference states; see Fig. 11).

Fig. 11.

Robustness of state space calibration across subjects shown by cross-validation of the amantadine-treated state. A classifier with 3 states (a Gaussian mixture model for parkinsonian, dyskinetic, and dyskinesia treated with amantadine) was trained in the subspace spanned by the parkinsonian and dyskinetic state in 1 animal and tested in the analogous subspace in a 2nd animal. The concentric circles represent the Gaussians corresponding to the 3 states (black, PD; red, dyskinesia; green, amantadine) that were fitted using data from the 1st animal only. The green crosses show the positions of the samples from the amantadine-treated state from the 2nd animal after calibration using the 2 reference states. The purple crosses show the positions of the same samples but without calibration. With calibration, the amantadine samples from the 2nd animal were correctly identified 85% of the time (i.e., the true-positive rate), which is only a slight decrease from the 89% achieved with the samples from the 1st animal, i.e., the data on which the classifier was trained. The corresponding false-positive rates were 7 and 3%, respectively. As a comparison, the true-positive rate without calibration was 26%.

In this subspace, spanned by the basis vectors control vs. PD and (control vs. dyskinesia)_ortho, the state induced by each drug was plotted separately (Fig. 9D). In agreement with the results from the 8-OH-DPAT experiments, it is clear that whereas several of the drugs produced reductions of dyskinetic symptoms (as shown in Fig. 9A), the neurophysiological state was nevertheless not fully normalized and in many cases partly reverted toward the PD state. In this context, it deserves mentioning that although dyskinetic symptoms were clearly reduced by some of the drugs, other aspects of the motor behavior appeared somewhat abnormal (amantadine: poor hindlimb-to-forelimb coordination, arching of back, and postural deficits; levetiracetam: very minor reduction in dyskinesia and flat body position; diazepam: mostly immobile but dyskinetic in association with movements).

DISCUSSION

Experimental treatment of CNS disease is conventionally evaluated in animals by documentation of changes in behavior. This approach has, however, a number of drawbacks. First, assessments of motor behavior only give indirect information on the underlying brain states that the therapy aims to treat, making it almost impossible to deduce the specific pharmacological/neurophysiological effects of the treatment. Second, the sensitivity and robustness in assessments of animal behavioral are usually not sufficient to allow for differentiation between several related CNS states. Third, unbiased measures based on automated procedures are still rare, making the testing procedures highly dependent on proper training of skilled observers, and reduce reproducibility between laboratories.

Ever since the first electrophysiological measurements were carried out in awake subjects, it has been known that the electrical activity of neurons frequently tend to synchronize into rhythmic patterns that vary depending on the state of the brain (Berger 1929). The results presented in the current study confirm previous findings suggesting an association between LFP oscillations within certain frequency intervals in specific regions of the corticobasal ganglia-thalamic circuit and various motor symptoms in PD (de Hemptinne et al. 2015; Hammond et al. 2007). More importantly, however, through the use of the developed techniques, previous findings can now be complemented with significantly more elaborate state characterizations based on large-scale multistructure recordings. The added value of these large-scale multistructure recordings was tested quantitatively by comparisons against the same recordings where state classifications were based on information obtained from fewer structures, showing a higher classification performance with higher number of structures (Fig. 9C). Similarly, we show that using the entire spectral contents in LFP recordings rather than the power in a few preselected frequency bands greatly improves state classification. It should also be noted that, by aligning the assessed systems-level states in each individual to a number of reference states, interindividual differences in activity patterns associated with each state are compensated for, which makes it possible to pool data across subjects without ad hoc realignment of data [this is a well-known problem in, for example, comparisons of spectral LFP contents between parkinsonian subjects (Kühn et al. 2009)]. In particular, in practical applications where, for example, therapeutic effects of a drug are evaluated and disease mechanisms are not of primary concern, this approach can be beneficial.

Using the developed method, we have here shown that robust and detailed representations of the pathophysiological conditions associated with motor symptoms in a rodent model of PD can be obtained. In addition, the complex and diverse effects of a number of different pharmacological interventions aimed at treating motor symptoms could also be characterized on a systems level. It may be worth noting in this context that although a representation based on the systems-level electrophysiological differences among the parkinsonian, dyskinetic, and control states is a natural starting point for the investigation of antidyskinetic interventions, adding other reference states/conditions to the analyses (e.g., information about the behavioral state, the effects of other drugs, etc.) will further help elucidate additional features of each state. Also, for pairwise comparisons of states such as direct comparison of the effects of two different drugs, difference spectra is a natural starting point for further analyses. In any case, the very rich data sets obtained with the described method potentially open up for a much more exploratory/data-driven approach, which can be very beneficial in this field of research due to the extreme complexity of the systems studied (Finkbeiner et al. 2015).

With regard to the animal model used in these experiments, it should be cautioned that the unilateral 6-OHDA medial forebrain bundle lesion model of PD has certain limitations. First, and most importantly, because the nonlesioned hemisphere is used as control, some comparisons cannot be made in a straightforward manner between pathological and nonpathological states (for example, changes in the firing rate of individual neurons), and a certain degree of variability is inevitably inherent to the model due to differences of the exact recoding locations in different hemispheres. Second, it cannot be assumed that the physiology of the intact hemisphere in a hemilesioned rat is entirely comparable to that of a nonlesioned animal due to potential biological adaptations that have occurred to compensate for the lesion-induced contralateral deficits. A few examples of such physiological changes have in fact been reported (Breit et al. 2008; González-Hernández et al. 2004; Kish et al. 1999). In an attempt to estimate how large these differences are, we quantitatively compared differences between intact hemispheres of lesioned and nonlesioned rats using multistructure recordings in different animals. Although not reaching significance, group differences were nevertheless confirmed (the average difference in median Euclidean distance to the nonlesioned reference condition was for intact hemispheres in hemilesioned rats 140% higher than that of contralateral hemispheres in nonlesioned animals; i.e., the median distance to group 1 for group 5 compared with the mean of groups 2 and 4 in Fig. 12). Third, although the severe lesions used in the model are beneficial for the study of dyskinesia, the limited therapeutic window for levodopa treatment precludes detailed analyses of the therapeutic effects of this drug. A strength of the unilateral model is, on the other hand, that certain factors affecting the general neurophysiological state are easier to control for in bilateral recordings with an internal control condition such as the degree of drowsiness/alertness, periods of immobility/locomotion, etc.

Fig. 12.

Control experiment with LFP spectra of the intact hemisphere in hemilesioned rats vs. LFP spectra in nonlesioned animals. Two experiments each were conducted in four nonlesioned rats, A–D, in the following referred to as RecA₁–D₁ and RecA₂–D₂. From each of these recordings, 10 min were chosen for further analysis. The power spectral densities (PSDs) in dB_pink during each 10-min period were then computed for each structure as described in methods, i.e., based on 8-s windows with 50% overlap. Samples containing the concatenated spectra from all structures in 1 hemisphere were constructed for each such 8-s window, resulting in 149 samples each for the left and right hemisphere during the analyzed 10-min period. The same was done for a 10-min period during the OFF- and ON-levodopa period, respectively, in 1 recording each of the hemilesioned animals 1-4. In summary, this resulted in the following data sets, each having a size of 149 samples × 4 animals: group: 1, left hemispheres from RecA₁–D₁; 2, right hemispheres from RecA₁–D₁; 3, left hemispheres from RecA₂–D₂; 4, right hemispheres from RecA₂–D₂; 5, control hemispheres from hemilesioned animals 1-4 OFF-levodopa; 6, control hemispheres from hemilesioned animals 1-4 ON-levodopa; 7, lesioned hemispheres from hemilesioned animals 1-4 OFF-levodopa (i.e., PD state); and 8, lesioned hemispheres from hemilesioned animals 1-4 ON-levodopa (i.e., dyskinetic state). Displayed is the similarity of the samples in each data set to the mean over all samples in group 1 with the similarity being measured as the Euclidean distance of each sample to this mean. Box represents 25th to 75th percentile, i.e., the IQR, and red line marks median value. Whiskers mark the range for values 1.5× IQR above or below the 75th or 25th percentile, respectively; data points outside of this range are marked as outliers. Blue asterisks denote median values for individual hemispheres in each group. Significant differences (black asterisks) between these median values on a group level were found between groups 7 and 8 and the control group [1; P < 0.05, t-test with Bonferroni correction for multiple comparisons (n = 7)].

In relation to previous publications using 6-OHDA-lesioned rats, it is worth pointing out that certain differences have been observed between recordings in anesthetized preparations compared to awake behaving animals. In particular, urethane-anesthetized 6-OHDA-lesioned rats have been reported to display beta oscillations with a somewhat lower oscillation frequency than awake animals (Brazhnik et al. 2014). Instead, awake rats typically display two types of beta oscillations that are dependent on the behavioral state (Avila et al. 2010; Brazhnik et al. 2014; Delaville et al. 2014). These oscillations (<15 and 20–35 Hz, respectively) were indeed present also in the current study (e.g., Fig. 2A, M1 before levodopa).

In addition to the presented measures, changes in functional connectivity between different structures, reflected in increased LFP coherence and correlated spiking activity of cells in anatomically connected structures, have also been implicated in the pathophysiology of PD (Fuentes et al. 2010; Hammond et al. 2007; Santana et al. 2014). Such measures have not been included in the state analyses to this point, but it is probable that the addition of pairwise coherence/correlation measures of neuronal activity within and between structures would help to improve further the performance of state classifications and would be a natural complement given the multistructure recording design.

This methodology could also be combined with several of the recently developed techniques for genetic manipulations of neuronal subpopulations that are, to date, primarily performed in mice. The presented findings indicate that several brain structures should preferably be targeted. Thus, to adapt the method to a smaller brain, it would be recommendable to scale down the number of electrodes used to target each brain structure in such experiments rather than reducing the number of structures.

Because motor symptoms are cardinal features of PD, neurophysiological states in parkinsonian and dyskinetic rats could here be directly matched to quantitative behavioral assessments of the displayed symptoms, essentially providing a validation of the neurophysiological readouts for the studied conditions. Following antidyskinetic treatment, an apparent mismatch was sometimes observed between the reduction in dyskinesia score and the corresponding changes in systems-level brain state (although the coordinate values in the dimension control vs. dyskinesia indeed correlated well with dyskinesia scores; see Fig. 9 legend). The discrepancy observed can, however, largely be explained by the fact that a behavioral characterization solely based on dyskinesia score does not capture a whole range of other motors symptoms that were here only described qualitatively. If more detailed behavioral assessments had been carried out with quantitative assessment scales that were adapted to include a wider range of motor symptoms, it is probable that the behavioral state descriptions would be better correlated to the neurophysiological activity states recorded in these motor circuits. Notably, however, such behavioral assessments are technically very challenging to carry out and will likely require more advanced automated procedures (e.g., Palmér et al. 2012; Santana et al. 2015). In addition, it is well-known that PD also includes nonmotor symptoms, and in several other disorders few overt signs, if any, may be associated with a specific pathological condition. In this situation, CNS state characterizations on a more holistic level could help opening up a new window into otherwise hidden internal processes in conditions such as persistent pain states, psychosis, depression, etc. We therefore envision that this technology could have an important use in the development of future treatments for a range of neurological and psychiatric conditions. More fundamentally, however, the knowledge gained from improved descriptions of how different brain structures interact to create mental states and complex behaviors in health and disease using a technology that bridges all the way from the scale of single-cell activity to systems-level states has potentially wide-reaching implications for neuroscientific research in general.

GRANTS

The study was supported by grants from the Bergvall, Crafoord, Kockska, Michael J. Fox, Olle Engkvist, Parkinson, Parkinson Research, Segerfalk, Åhlén, and Åke Wiberg Foundations, MultiPark, Hjärnfonden, Svenska Sällskapet för Medicinsk Forskning (SSMF), BABEL (Erasmus Mundus), and Vetenskapsrådet (VR) Grant 325-2011-6441.

DISCLAIMERS

The sponsors had no role in the study design, the collection, analysis, and interpretation of the data, the writing of the report, or the decision to submit the article for publication.

DISCLOSURES

No conflicts of interest, financial or otherwise, are declared by the author(s).

AUTHOR CONTRIBUTIONS

P.H. and P.P. conception and design of research; M.T., I.B., and N.I. performed experiments; M.T., I.B., U.R., P.H., and P.P. analyzed data; M.T., I.B., U.R., P.H., and P.P. interpreted results of experiments; M.T., U.R., and P.H. prepared figures; M.T., U.R., P.H., and P.P. drafted manuscript; I.B. and P.P. edited and revised manuscript; M.T., I.B., U.R., P.H., and P.P. approved final version of manuscript.