Differential dopaminergic modulation of spontaneous cortico–subthalamic activity in Parkinson’s disease

  1. Abhinav Sharma  Is a corresponding author
  2. Diego Vidaurre
  3. Jan Vesper
  4. Alfons Schnitzler
  5. Esther Florin  Is a corresponding author
  1. Institute of Clinical Neuroscience and Medical Psychology, Medical Faculty, Heinrich-Heine University Düsseldorf, Germany
  2. Department of Psychiatry, University of Oxford, United Kingdom
  3. Department of Clinical Health, Aarhus University, Denmark
  4. Department of Neurosurgery, University Hospital Düsseldorf, Germany
  5. Department of Neurology, Center for Movement Disorders and Neuromodulation, Medical Faculty, Heinrich-Heine University Düsseldorf, Germany

Abstract

Pathological oscillations including elevated beta activity in the subthalamic nucleus (STN) and between STN and cortical areas are a hallmark of neural activity in Parkinson’s disease (PD). Oscillations also play an important role in normal physiological processes and serve distinct functional roles at different points in time. We characterised the effect of dopaminergic medication on oscillatory whole-brain networks in PD in a time-resolved manner by employing a hidden Markov model on combined STN local field potentials and magnetoencephalography (MEG) recordings from 17 PD patients. Dopaminergic medication led to coherence within the medial and orbitofrontal cortex in the delta/theta frequency range. This is in line with known side effects of dopamine treatment such as deteriorated executive functions in PD. In addition, dopamine caused the beta band activity to switch from an STN-mediated motor network to a frontoparietal-mediated one. In contrast, dopamine did not modify local STN–STN coherence in PD. STN–STN synchrony emerged both on and off medication. By providing electrophysiological evidence for the differential effects of dopaminergic medication on the discovered networks, our findings open further avenues for electrical and pharmacological interventions in PD.

Introduction

Oscillatory activity serves crucial cognitive roles in the brain (Akam and Kullmann, 2010; Akam and Kullmann, 2014), and alterations of oscillatory activity have been linked to neurological and psychiatric diseases (Schnitzler and Gross, 2005). Different large-scale brain networks operate with their own oscillatory fingerprint and carry out specific functions (Keitel and Gross, 2016; Mellem et al., 2017; Vidaurre et al., 2018b). Given the dynamics of cognition, different brain networks need to be recruited and deployed flexibly. Hence, the duration for which a network is active, its overall temporal presence, and even the interval between the different activations of a specific network might provide a unique window to understanding brain functions. Crucially, alterations of these temporal properties or networks might be related to neurological disorders.

In Parkinson’s disease (PD), beta oscillations within the subthalamic nucleus (STN) and motor cortex (13–30 Hz) correlate with the motor symptoms of PD (Marreiros et al., 2013; van Wijk et al., 2016; West et al., 2018). Beta oscillations also play a critical role in communication in a healthy brain (Engel and Fries, 2010). (For the purposes of our paper, we refer to oscillatory activity or oscillations as recurrent but transient frequency-specific patterns of network activity, even though the underlying patterns can be composed of either sustained rhythmic activity, neural bursting, or both [Quinn et al., 2019]. Disambiguating the exact nature of these patterns is, however, beyond the scope of this work.) At the cellular level, loss of nigral dopamine neurons in PD leads to widespread changes in brain networks, to varying degrees across different patients. Dopamine loss is managed in patients via dopaminergic medication. Dopamine is a widespread neuromodulator in the brain (Gershman and Uchida, 2019), raising the question of whether each medication-induced change restores physiological oscillatory networks. In particular, dopaminergic medication is known to produce cognitive side effects in PD patients (Voon et al., 2009). According to the dopamine overdose hypothesis, a reason for these effects is the presence of excess dopamine in brain regions not affected in PD (MacDonald et al., 2011; MacDonald and Monchi, 2011). Previous task-based and neuroimaging studies in PD demonstrated frontal cognitive impairment due to dopaminergic medication (Cools et al., 2002; Ray and Strafella, 2010; MacDonald et al., 2011).

Using resting-state whole-brain MEG analysis, network changes related to both motor and non-motor symptoms of PD have been described (Olde Dubbelink et al., 2013a; Olde Dubbelink et al., 2013b). However, these studies could not account for simultaneous STN–STN or cortico–STN activity affecting these networks, which would require combined MEG/electroencephalogram (EEG)–LFP recordings (Litvak et al., 2021). Such recordings are possible during the implantation of deep brain stimulation (DBS) electrodes, an accepted treatment in the later stages of PD (Volkmann et al., 2004; Deuschl et al., 2006; Kleiner-Fisman et al., 2006). Combined MEG–LFP studies in PD involving dopaminergic intervention report changes in beta and alpha band connectivity between specific cortical regions and the STN (Litvak et al., 2011; Hirschmann et al., 2013; Oswal et al., 2016). Decreased cortico–STN coherence under dopaminergic medication (ON) correlates with improved motor functions in PD (George et al., 2013). STN–STN intra-hemispheric oscillations positively correlate to motor symptom severity in PD without dopaminergic medication (OFF), whereas dopamine-dependent nonlinear phase relationships exist between inter-hemispheric STN–STN activity (West et al., 2016). Crucially, previous studies could not rule out the influence of cortico–STN connectivity on these inter-hemispheric STN–STN interactions.

To further characterise the differential effects of dopaminergic medication and delineate pathological versus physiological-relevant spectral connectivity in PD, we study PD brain activity via a hidden Markov model (HMM), a data-driven learning algorithm (Vidaurre et al., 2016; Vidaurre et al., 2018b). Due to the importance of cortico–subcortical interactions in PD, we investigated these interactions with combined spontaneous whole-brain magnetoencephalography (MEG) and STN local field potentials (LFPs) recordings from PD patients. We study whole-brain connectivity including the STN using spectral coherence as a proxy for communication based on the communication through coherence hypothesis (Fries, 2005; Fries, 2015). This will allow us to delineate differences in communication OFF and ON medication. Furthermore, we extended previous work that was limited to investigating communication between specific pairs of brain areas (Litvak et al., 2011; George et al., 2013; Hirschmann et al., 2013). Moreover, we identified the temporal properties of the networks both ON and OFF medication. The temporal properties provide an encompassing view of network alterations in PD and the effect of dopamine on these networks.

We found that cortico–cortical, cortico–STN, and STN–STN networks were differentially modulated by dopaminergic medication. For the cortico–cortical network, medication led to additional connections that can be linked to the side effects of dopamine. At the same time, dopamine changed the cortico–STN network towards a pattern more closely resembling physiological connectivity as reported in the PD literature. Within the third network, dopamine only had an influence on local STN–STN coherence. These results provide novel information on the oscillatory network connectivity occurring in PD and the differential changes caused by dopaminergic intervention. These whole-brain networks, along with their electrophysiological signatures, open up new potential targets for both electric and pharmacological interventions in PD.

Results

Under resting-state conditions in PD patients, we simultaneously recorded whole-brain MEG activity with LFPs from the STN using directional electrodes implanted for DBS. Using an HMM, we identified recurrent patterns of transient network connectivity between the cortex and the STN, which we henceforth refer to as an ‘HMM state’. In comparison to classic sliding window analysis, an HMM solution can be thought of as a data-driven estimation of time windows of variable length (within which a particular HMM state was active): once we know the time windows when a particular state is active, we compute coherence between different pairs of regions for each of these recurrent states. Each HMM state itself is a multidimensional, time-delay embedded (TDE) covariance matrix across the whole brain, containing information about cross-regional coherence and power in the frequency domain. Additionally, the temporal evolution of the HMM states was determined. The PD data were acquired under medication (L-DOPA) OFF and ON conditions, which allowed us to delineate the physiological versus pathological spatio-spectral and temporal changes observed in PD. To allow the system to dynamically evolve, we use time delay embedding. Theoretically, delay embedding can reveal the state space of the underlying dynamical system (Packard et al., 1980). Thus, by delay-embedding PD time series OFF and ON medication, we uncover the differential effects of a neurotransmitter such as dopamine on underlying whole-brain connectivity. OFF medication, patients had on average a Unified Parkinson’s Disease Rating Scale (UPDRS) part III of 29.24 ± 10.74. This was reduced by L-DOPA (176.5 ± 56.2 mg) to 19.47 ± 8.52, indicating an improvement in motor symptoms.

Spontaneous brain activity in PD can be resolved into distinct states

Using an HMM, we delineated cortico–subthalamic spectral changes from both global source-level cortical interactions as well as local STN–STN interactions. Three of the six HMM states could be attributed to physiologically interpretable connectivity patterns. We could not interpret the other three states within the current physiological frameworks both OFF and ON medication and they are therefore not considered in the following (see Figure 2—figure supplement 1). The connectivity between different brain regions for each state was visualised for the frequency modes shown in Figure 1. Figures 24 show the connectivity patterns for the three physiologically meaningful states in both the OFF (top row) and ON medication condition (bottom row). We refer to the state obtained in Figure 2 as the cortico–cortical state (Ctx–Ctx). This state was characterised mostly by local coherence within segregated networks OFF medication in the alpha and beta band. In contrast, there was a widespread increase in coherence across the brain from OFF to ON medication. Therefore, ON medication, the connectivity strength in the alpha and beta band was not significantly different from the mean noise level. Figure 3 displays the second state. A large proportion of spectral connections in this state enable cortico–STN communication via spectral coherence (Lalo et al., 2008; Litvak et al., 2011; Hirschmann et al., 2013; Oswal et al., 2013; van Wijk et al., 2016) and thus we labelled this as the cortico–STN state (Ctx–STN). This state was characterised by connectivity between multiple cortical regions and the STN OFF medication, but increased specificity of cortical–STN connectivity ON medication. Finally, Figure 4 shows the third state. Within this state, highly synchronous STN–STN spectral connectivity emerged, both OFF and ON medication and therefore we named it the STN–STN state (STN–STN). The spectral characteristics of this state largely remain unaffected under the influence of dopaminergic medication. In the following sections, we describe these three states in detail.

Data-driven frequency modes.

Each plotted curve shows a different spectral band. The x-axis represents frequency in Hz and the y-axis represents the weights obtained from the non-negative matrix factorisation (NNMF) in arbitrary units. The NNMF weights are like regression coefficients. The frequency resolution of the modes is 0.5 Hz. Panels A and B show the OFF and ON medication frequency modes, respectively. Source data are provided as Figure 1—source data 12.

Figure 2 with 1 supplement see all
Cortico–cortical state.

The cortico–cortical state was characterised by a significant increase in coherence ON compared to OFF medication (see panel B). Due to this, no connections within the alpha and beta band ON medication were significantly higher than the mean (panel C). However, in the delta band, ON medication medial prefrontal–orbitofrontal connectivity emerged. (A and C) Each node in the circular graph represents a brain region based on the Mindboggle atlas. The regions from the atlas are listed in Table 1 along with their corresponding numbers that are used in the circular graph. The colour code in the circular graph represents a group of regions clustered according to the atlas (starting from node number 1) STN contacts (contacts 1, 2, 3 = right STN and contacts 4, 5, 6 = left STN), frontal, medial frontal, temporal, sensorimotor, parietal, and visual cortices. In the circular graph, only the significant connections (p<0.05; corrected for multiple comparisons, IntraMed analysis) are displayed as black curves connecting the nodes. The circles from left to right represent the delta/theta, alpha, and beta bands. Panel A shows results for OFF medication data and panel C for the ON medication condition. For every circular graph, we also show a corresponding top view of the brain with the connectivity represented by yellow lines and the red dot represents the anatomical seed vertex of the brain region. Only the cortical connections are shown. Panel B shows the result for inter-medication analysis (InterMed) for the cortico–cortical state. In each symmetric matrix, every row and column corresponds to a specific atlas cluster denoted by the dot colour on the side of the matrix. Each matrix entry is the result of the InterMed analysis where OFF medication connectivity between ith row and jth column was compared to the ON medication connectivity between the same connections. A cell is white if the comparison mentioned on top of the matrix (either ON >OFF or OFF >ON) was significant at a threshold of p<0.05. The connectivity maps of states 4–6 are provided in Figure 2—figure supplement 1. Source data are provided as Figure 2—source data 13.

Cortico–STN state.

For the general description, see the note to Figure 2. The cortico–STN state was characterised by preservation of spectrally selective cortico–STN connectivity ON medication. Also, ON medication, a sensorimotor–frontoparietal network emerged. Source data are provided as Figure 3—source data 13.

STN–STN state.

For the general description, see the note to Figure 2. The STN–STN state was characterised by preservation of STN–STN coherence in the alpha and beta band OFF versus ON medication. STN–STN theta/delta coherence was no longer significant ON medication. Source data are provided as Figure 4—source data 13.

Table 1
Regions of the Mindboggle atlas used.

STN, subthalamic nucleus; Vis, visual; Par, parietal; Smtr, sensory motor; Tmp, temporal; Mpf, medial prefrontal; Frnt, frontal; Ctx, cortex. The colour code is for the ring figures presented as part of the results.

STN1Contact one rightSmtr-Ctx12Postcentral
2Contact two right13Precentral
3Contact three rightTmp-Ctx14Middle temporal
1Contact four left15Superior temporal
2Contact five leftMpf-Ctx16Caudal middle frontal
3Contact six left17Medial orbitofrontal
Vis-Ctx4CuneusFrnt-Ctx18Insula
5Lateral occipital19Lateral orbitofrontal
6Lingual20Pars opercularis
Par-Ctx7Inferior parietal21Pars orbitalis
8Para central22Pars triangularis
9Precuneus23Rostral middlefrontal
10Superior parietal24Superior frontal
11Supramarginal

Ctx–Ctx state is characterised by increased frontal coherence due to elevated dopamine levels

Supporting the dopamine overdose hypothesis in PD (Kelly et al., 2009; MacDonald and Monchi, 2011), we identified a delta/theta oscillatory network involving intra-hemispheric connections between the lateral and medial orbitofrontal cortex as well as the pars orbitalis. The delta/theta network emerged between the lateral and medial orbitofrontal as well as left and right pars orbitalis cortex ON medication (p<0.05, Figure 2C delta). On the contrary, OFF medication no significant connectivity was detected in the delta/theta band. In the alpha and beta band OFF medication there was significant connectivity within the frontal regions, STN, and to a limited extent in the posterior parietal regions (p<0.05, Figure 2A).

Another effect of excess dopamine was significantly increased connectivity of frontal cortex and temporal cortex both with the STN and multiple cortical regions across all frequency modes (p<0.01, Figure 2 delta, alpha, and beta). The change in sensorimotor–STN connectivity primarily took place in the alpha band with an increased ON medication. Sensorimotor–cortical connectivity was increased ON medication across multiple cortical regions in both the alpha and beta band (p<0.01, Figure 2 alpha and beta). However, STN–STN coherence remained unchanged OFF versus ON medication across all frequency modes.

Viewed together, the Ctx–Ctx state captured increased coherence across the cortex ON medication within the alpha and beta band. This, however, implies that ON medication, no connectivity strength was significantly higher than the mean noise level within the alpha or beta band. ON medication, significant coherence emerged in the delta/theta band primarily between different regions of the orbitofrontal cortex.

Dopaminergic medication selectively reduced connectivity in the Ctx–STN state

Our analysis revealed that the Ctx–STN state ON medication was characterised by selective cortico–STN spectral connectivity and an overall shift in cortex-wide activity towards physiologically relevant network connectivity. In particular, ON medication, connectivity between STN and cortex became more selective in the alpha and beta band. OFF medication, STN–pre-motor (sensory), STN–frontal, and STN–parietal connectivity was present (p<0.05, Figure 3A alpha and beta). Importantly, coherence OFF medication was significantly larger than ON medication between STN and sensorimotor, STN and temporal, and STN and frontal cortices (p<0.05 for all connections, Figure 3B alpha and beta). Furthermore, ON medication, in the alpha band only the connectivity between temporal, parietal, and medial orbitofrontal cortical regions and the STN was preserved (p<0.05, Figure 3C alpha). Finally, ON medication, a sensorimotor–frontoparietal network emerged (p<0.05, Figure 3C beta), where sensorimotor, medial prefrontal, frontal, and parietal regions were no longer connected to the STN, but instead directly communicated with each other in the beta band. Hence, there was a transition from STN-mediated sensorimotor connectivity to the cortex OFF medication to a more direct cortico–cortical connectivity ON medication.

Simultaneously to STN–cortico and cortico–cortical, STN–STN connectivity changed. In the ON condition, STN–STN connectivity was significantly different from the mean noise level across all three frequency modes (p<0.05, Figure 3C). But on the other hand, there was no significant change in the STN–STN connectivity OFF versus ON medication (p=0.21 delta/theta; p=0.25 alpha; p=0.10 beta; Figure 3B).

To summarise, coherence decreased ON medication across a wide range of cortical regions both at the cortico–cortical and cortico–STN level. Still, significant connectivity was selectively preserved in a spectrally specific manner ON medication both at the cortico–cortical (sensorimotor–frontoparietal network) and the cortico–STN levels. The most surprising aspect of this state was the emergence of bilateral STN–STN coherence ON medication across all frequency modes.

Dopamine selectively modifies delta/theta oscillations within the STN–STN state

In this STN–STN state, dopaminergic intervention had only a limited effect on STN–STN connectivity. OFF medication, STN–STN coherence was present across all three frequency modes (p<0.05, Figure 4A), while ON medication, significant STN–STN coherence emerged only in the alpha and beta band (p<0.05, Figure 4C alpha and beta). ON medication, STN–STN delta/theta connectivity strength was not significantly different from the mean noise level (p<0.05, Figure 4C delta).

OFF compared to ON medication, coherence was reduced across the entire cortex both at the inter-cortical and the STN–cortex level across all frequency modes. The most affected areas were similar to the ones in the Ctx–STN state, in other words, the sensorimotor, frontal, and temporal regions. Their coherence with the STN was also significantly reduced, ON compared to OFF medication (STN–sensorimotor, p<0.01 delta/theta, beta; p<0.05 alpha; STN–temporal, p<0.01 delta/theta, alpha, beta; and STN–frontal, p<0.01 delta/theta, alpha and beta; Figure 4B).

In summary, STN–STN connectivity was not significantly altered OFF to ON medication. At the same time, coherence decreased from OFF to ON medication at both the cortico–cortical and the cortico–STN level. Therefore, only significant STN–STN connectivity existed both OFF and ON medication, while cortico–STN or cortico–cortical connectivity changes remained at the mean noise level.

States with a generic coherence decrease have longer lifetimes

Using the temporal properties of the identified networks, we investigated whether states showing a shift towards physiological connectivity patterns lasted longer ON medication. A state that is physiological should exhibit increased lifetime and/or should occur more often ON medication. An example of the state time courses is shown in Figure 5.

Example of a probability time course for the six hidden Markov model (HMM) states OFF medication.

Note that within the main text of the paper, we are only discussing the first three states. The connectivity maps of states 4–6 are provided in Figure 2—figure supplement 1. Source data are provided as Figure 5—source data 12.

Figure 5—source data 1

Probability time course first half in relation to Figure 5.

https://cdn.elifesciences.org/articles/66057/elife-66057-fig5-data1-v1.mat
Figure 5—source data 2

Probability time course second half in relation to Figure 5.

https://cdn.elifesciences.org/articles/66057/elife-66057-fig5-data2-v1.mat

Figure 6A-C shows the temporal properties for the three states for both the OFF and ON medication conditions. Two-way repeated measures ANOVA on the temporal properties of the HMM states revealed an effect of HMM states on the fractional occupancy (FO) (F(2,96)=10.49, p<0.01), interval of visits (F(2,221513)=9783.13, p<0.01), and lifetime (F(2,214818)=50.36, p<0.01). There was no effect of medication (L-DOPA) on FO (F(1,96)=2.00, p=0.16) and lifetime (F(1,214818)=0.15, p=0.7026). Medication had a significant effect on the interval of visits (F(1,221513)=4202.96, p<0.01). Finally, we found an interaction between the HMM states and medication on the interval of visits (F(2,221513)=1949.98, p<0.01) and lifetime (F(2,214818)=172.25, p<0.01). But there was no interaction between HMM states and medication on FO (F(2,96)=0.54, p=0.5855).

Temporal properties of states.

Panel A shows the fractional occupancy for the three states for the cortico–cortical (Ctx–Ctx), cortico–STN (Ctx–STN), and the STN–STN (STN–STN). Each point represents the mean for a state and the error bar represents standard error. Orange denotes ON medication data and blue OFF medication data. Panel B shows the mean interval of visits (in milliseconds) of the three states ON and OFF medication. Panel C shows the lifetime (in milliseconds) for the three states. Figure insets are used for clarity in case error bars are not clearly visible. The y-axis of each figure inset has the same units as the main figure. Source data are provided as Figure 6—source data 16.

We performed post hoc testing on the ANOVA results. OFF medication, the STN–STN state was the one with the longest lifetime (STN–STN >Ctx Ctx, p<0.01; STN–STN >Ctx-STN, p<0.01). The Ctx–STN state OFF medication had the shortest lifetime among all three states (Ctx–STN <Ctx-Ctx, p<0.01; Ctx–STN <STN-STN, p<0.01) and the shortest interval between visits (interval of visit Ctx–STN <Ctx-Ctx, p<0.01; Ctx–STN <STN-STN, p<0.01). The largest interval between visits was for the Ctx–Ctx state OFF medication (Ctx–Ctx >STN-STN, p<0.01; Ctx–Ctx >Ctx-STN, p<0.01). The FO for the STN–STN and Ctx–STN states was similar, but significantly higher than for the Ctx–Ctx state (STN–STN >Ctx-Ctx, p<0.01; STN–STN ≈ Ctx–STN, p=0.82; Ctx–STN >Ctx-Ctx, p<0.01). ON medication, the comparison between temporal properties of all three states retained the same significance levels as OFF medication, except for the lifetime of the Ctx–STN state, which was no longer significantly different from that of the Ctx–Ctx state (p=0.98). Within each medication condition, the states retained their temporal characteristics relative to each other.

Across medication conditions, significant changes were present in the temporal properties of the states. The lifetimes for both the STN–STN and Ctx–STN state were significantly increased by medication (ON >OFF: STN–STN, p<0.01; Ctx–STN, p≤0.01) but the lifetime for the Ctx–Ctx state was not significantly influenced by medication. The Ctx–Ctx state was visited even less often ON medication (interval: ON >OFF Ctx–Ctx, p<0.01). The interval between visits remained unchanged for the STN–STN and Ctx–STN states. The FO for all three states was not significantly changed from OFF to ON medication. In summary, the cortico–cortical state was visited least often compared to the other two states both OFF and ON medication. The cortico–STN and STN–STN states showing physiologically relevant spectral connectivity lasted significantly longer ON medication.

Discussion

In this study, we parsed simultaneously recorded MEG–STN LFP signals into discrete time-resolved states to reveal distinct spectral communication patterns. We identified three states exhibiting distinct coherence patterns ON and OFF medication: a cortico–cortical, a cortico–STN, and a STN–STN state. Our results indicate a tendency of neural activity to engage in connectivity patterns in which coherence decreases under the effect of dopaminergic medication and which maintain selective cortico–STN connectivity (Ctx–STN and STN–STN states). Only within the Ctx–Ctx state did coherence increase under dopaminergic medication. These results are in line with the multiple effects of dopaminergic medication reported in resting and task-based PD studies (Jubault et al., 2009; West et al., 2016; Tinkhauser et al., 2017).

The differential effect of dopamine allowed us to delineate pathological and physiological spectral connectivity. The Ctx–Ctx state provided electrophysiological evidence in the delta/theta band for the overdose effect of dopaminergic medication in PD. Prior to the electrophysiological evidence in our study, there was only evidence through task-based or functional magnetic resonance imaging (fMRI) studies (Cools et al., 2002; Ray and Strafella, 2010; MacDonald et al., 2011). The Ctx–STN state revealed that simultaneous cortico–cortical and STN–STN interactions emerge ON medication, with spectrally and spatially specific cortico–STN interactions. In addition, ON medication, a frontoparietal motor network was present, indicating a shift from STN-mediated motor connectivity to a cortical one. These findings have not been reported in previous studies. The STN–STN state exhibited the limited ability of dopaminergic medication to modify local STN–STN delta oscillations. Our analysis also revealed significant changes in the temporal properties of the connectivity profiles, including lifetime and FO, under the effect of dopaminergic medication. This insight might in the future prove important for modifying medication as well as DBS-based strategies for therapeutic purposes.

Increased tonic dopamine causes excessive frontal cortical activity

The Ctx–Ctx state showed significant coherent connectivity between the orbitofrontal cortical regions in the delta/theta band ON medication. According to the dopamine overdose hypothesis in PD (Cools, 2001; Kelly et al., 2009; MacDonald and Monchi, 2011; Vaillancourt et al., 2013), the commonly used doses of dopaminergic medication to mitigate the motor symptoms cause the ventral frontostriatal cortical circuits to experience an excessive increase in tonic dopamine levels. This medication-induced increase is due to excessive compensation of dopamine in the ventral striatal circuitry, which experiences a lower loss of dopamine than its dorsal counterpart. The reason is that in PD dopaminergic neurons in the substantia nigra are primarily lost and therefore the dopamine depletion within the dorsal circuitry is higher than within the ventral one (Kelly et al., 2009; MacDonald and Monchi, 2011). Frontal regions involved with the ventral striatal circuitry include the orbitofrontal cortex, anterior cingulate, and the inferior temporal cortex (Cools, 2006; MacDonald and Monchi, 2011). Increased frontal cortex connectivity potentially explains the cognitive deficits observed in PD (Shohamy et al., 2005; George et al., 2013). Our detected emergence of frontal cortico–cortical coherence (between orbitofrontal and medial orbitofrontal regions) specifically in the delta/theta band could explain the cognitive deficits observed in PD due to dopaminergic medication, given the role of frontal delta/theta oscillations in cognition (Harmony, 2013; Zavala et al., 2014).

A comparison of temporal properties of the Ctx–Ctx state OFF versus ON medication revealed that the interval between visits was significantly increased ON medication, while the FO of this state was significantly reduced. In fact, the FO of the Ctx–Ctx state was the lowest among the three states. The temporal results indicate that the Ctx–Ctx state is least visited. Neural activity ON medication is not likely to visit this state, but whenever it does, its visits are of the same duration as OFF medication. Hence, the Ctx–Ctx state’s presence could explain the cognitive side effects observed ON medication in PD.

Selective spectral connectivity remains preserved with increased dopamine levels

An interesting feature of the Ctx–STN state was the emergence of local STN–STN coherence in all three frequency modes. Bilateral STN–STN coherence in the alpha and beta band did not change in the Ctx–STN state ON versus OFF medication (InterMed analysis). However, STN–STN coherence was significantly higher than the mean level ON medication (IntraMed analysis). Since synchrony limits information transfer (Cruz et al., 2009; Cagnan et al., 2015; Holt et al., 2019), the high coherence within the STN ON medication could prevent communication with the cortex. A different explanation would be that a loss of cortical afferents leads to increased local STN coherence. The causal nature of the cortico–basal ganglia interaction is an endeavour for future research.

Previous studies have reported STN–sensorimotor (Hirschmann et al., 2011; Litvak et al., 2011), STN–parietal, and STN–frontal (Litvak et al., 2011) coherence in the beta band OFF medication. Consistent with previous studies STN–sensorimotor, STN–parietal (inferior parietal), STN–frontal (insular cortex, pars orbitalis, pars opercularis, and lateral orbitofrontal), and STN–medial prefrontal (medial orbitofrontal) coherence emerged in the Ctx–STN state. In contrast, ON medication sensorimotor regions were coherent with parietal (para central) and frontal (superior frontal)/medial prefrontal (caudal middle frontal) regions in the beta frequency range. Previous research has not reported the emergence of such a coherent frontoparietal motor network ON medication. But consistent with previous research (Hirschmann et al., 2013), sensorimotor–STN coherence was reduced ON compared to OFF medication.

In addition, critical processing regarding sensorimotor decision-making involves frontoparietal regions (Gertz and Fiehler, 2015; Siegel et al., 2015; Gallivan et al., 2018; Martínez-Vázquez and Gail, 2018). Hence, the emergence of frontoparietal connectivity with motor regions points towards the physiological relevance of the Ctx–STN state. Moreover, neural activity ON medication remained longer in the Ctx–STN state as the lifetime of this state significantly increased compared to OFF medication. The finding is in line with our hypothesis that a state showing physiologically relevant spectral connectivity lasts longer ON medication.

Tonic dopamine has a limited effect on local STN–STN interactions

In the Ctx–STN state, STN–STN coherence accompanied network changes affecting cortico–STN communication ON medication, thereby likely having a functional role. In contrast, in the STN–STN state, STN–STN coherence emerged without the presence of any significant cortico–STN coherence either OFF or ON medication. This may indicate that the observed STN–STN activity in the STN–STN state emerged due to local basal ganglia circuitry. No changes were observed in the alpha and beta band in the STN–STN state ON versus OFF medication, which may indicate the inability of tonic dopamine to modify basal ganglia circuit activity. These results provide more evidence that the changes in STN–STN coherence observed in previous studies (Little et al., 2013; Oswal et al., 2013; Shimamoto et al., 2013) reflect cortical interaction affecting STN activity. Future studies should analyse changes occurring within the STN. To the best of our knowledge, we are the first to uncover modulation of STN–STN delta/theta oscillations by dopaminergic medication. Studies have shown that local basal ganglia delta oscillations, which do not require input from the motor cortex, are robust biomarkers of dopamine depletion (Whalen et al., 2020). Hence, selective elimination of delta/theta oscillations under dopaminergic medication in the STN–STN state points towards restoration of physiologically relevant network activity.

Limitations of the study

In the present study, we employed a data-driven approach based on an HMM. In order to find the appropriate model, we had to specify the number of states a priori. We selected the number of HMM states based on a compromise between spectral quality of results and their redundancy. The number of states could also be determined by selecting the one with the highest negative free energy. However, model selection based on free energy often does not yield concrete results (Baker et al., 2014). Another limitation is the use of multivariate Gaussian distributions to characterise the state covariance matrices. Although it improves the tractability of the HMM estimation process, it is by construction unable to capture higher-order statistics of the data beyond the first two moments. For example, burst activity might also be a relevant property of brain networks (Florin et al., 2015). Lastly, we would like to note that the HMM was used as a data-driven, descriptive approach without explicitly assuming any a priori relationship with pathological or physiological states. The relation between biology and the HMM states, thus, purely emerged from the data; that is, is empirical. What we claim in this work is simply that the features captured by the HMM hold some relation with the physiology even though the estimation of the HMM was completely unsupervised (i.e., blind to the studied conditions).

Besides these limitations inherent in the analysis approach, there are also some related to the experimental design. As this is a study containing invasive LFP recordings, we can never have a healthy control group. In addition, we only recorded four female patients because during the study period fewer female patients underwent a DBS surgery at our centre. To the best of our knowledge, there is no previous literature reporting a sex difference in MEG markers or the prescribed dopaminergic medication (Umeh et al., 2014). The medication led to a marked motor improvement in these patients based on the UPDRS, but the patients still have impairments. Both motor impairment and motor improvement can cause movement during the resting state in PD. While such movement is a deviation from a resting state in healthy subjects, such movements are part of the disease and occur unwillingly. Therefore, such movements can arguably be considered part of the resting state of PD. None of the patients in our cohort experienced hyperkinesia during the recording. All patients except for two were of the akinetic-rigid subtype. We verified that tremor movement is not driving our results. Recalculating the HMM states without these two subjects, even though it slightly changed some particular aspects of the HMM solution, did not materially affect the conclusions. A further potential influencing factor might be the disease duration and the amount of dopamine patients are receiving. Both factors were not significantly related to the temporal properties of the states.

To differentiate pathological and physiological network activity, we had to rely on the temporal properties of the networks. A further limitation was that all our recordings were made under resting conditions, preventing us from discerning the functional role of oscillations within the discovered networks. We opted for the current design because resting-state data allows the study of networks independent of a task and because using a specific task bears the risk that the patients are not able to properly perform it. Nevertheless, future studies should analyse the behaviour of specific networks using tasks to probe them.

Lastly, we recorded LFPs from within the STN – an established recording procedure during the implantation of DBS electrodes in various neurological and psychiatric diseases. Although for Parkinson patients results on beta and tremor activity within the STN have been reproduced by different groups (Reck et al., 2010; Litvak et al., 2011; Florin et al., 2013; Hirschmann et al., 2013; Neumann et al., 2016), it is still not fully clear whether these LFP signals are contaminated by volume-conducted cortical activity. However, while volume conduction seems to be a larger problem in rodents even after re-referencing the LFP signal (Lalla et al., 2017), the same was not found in humans (Marmor et al., 2017). Moreover, we used directional contacts, which have a smaller surface area than the classical ring contacts. Based on the available literature, our sampling rate was high enough to resolve oscillatory activity in the STN (Telkes et al., 2020; Nguyen et al., 2020).

Conclusion

Using a data-driven machine learning approach, we identified three distinct networks (states) that captured differential effects of dopaminergic medication on spectral connectivity in PD. Our findings uncovered a Ctx–Ctx state that captured the potentially adverse effects of increased dopamine levels due to dopaminergic medication. Furthermore, a Ctx–STN state was identified that maintained spatio-spectrally selective cortico–STN connectivity ON medication. We also found an STN–STN coherent state, pointing towards the limited effect of dopaminergic medication to modify local basal ganglia activity. Our findings bring forth a dynamical systems perspective for differentiating pathological versus physiologically relevant spectral connectivity in PD. Furthermore, we were able to uncover differential changes induced by altered levels of a neuromodulator such as dopamine in a completely data-driven manner without providing detailed information about large-scale dopaminergic networks to the HMM. This shows another advantage of our dynamical systems-level approach. Furthermore, our whole-brain STN approach provides novel electrophysiological evidence of distributed changes due to dopaminergic medication in brain connectivity, extending previous pairwise connectivity results reported in PD.

Materials and methods

Subjects

In total, 17 (4 female) right-handed PD patients (age: 55.2 ± 9.3 years) undergoing surgery for therapeutic STN DBS were recruited for this study. Patients had been selected for DBS treatment according to the guidelines of the German Society for Neurology. The experimental procedure was explained to all participants and they gave written consent. The study was approved by the local ethics committee (study number 5608R) and conducted in accordance with the Declaration of Helsinki. Bilateral DBS electrodes were implanted in the dorsal part of the STN at the Department of Functional Neurosurgery and Stereotaxy in Düsseldorf. The implanted DBS electrodes used were the St. Jude Medical directional lead 6172 (Abbott Laboratories, Lake Bluff, IL) and in one case the Boston Scientific Vercise segmented lead (Boston Scientific Corporation, Marlborough, MA). These electrodes have four contact heights and the two middle heights are segmented into three equally spaced contacts.

The DBS leads were externalised and we measured the patients after 1–3 days. To simultaneously acquire MEG and LFP signals, we connected the externalised leads to an EEG amplifier integrated with the MEG system. We used a whole-head MEG system with 306 channels (Elekta Vectorview, Elekta Neuromag, Finland) housed within a magnetically shielded chamber. All patients were requested to sit still and awake during data acquisition. To ensure that patients did not fall asleep, we tracked patients’ pupil diameter with an eye tracker. To remove eye blink and cardiac artefacts, electrooculography and electrocardiography were recorded along with the LFP and MEG signals. In order to co-register the MEG recording with the individual MRI, four head position indicator coils were placed on the patient’s head. Their position as well as additional head points were digitised using the Polhemus Isotrack system (Polhemus, Colchester, CT). The data were recorded with a sampling rate of 2400 Hz and a low-pass filter of 800 Hz was applied. An electrode was placed at the mastoid and all LFP signals were referenced to it.

For the clinical OFF medication state, oral PD medication was withdrawn overnight for at least 12 hr. If a patient had an apomorphine pump, this pump was stopped at least 1 hr before the measurement. First, we recorded resting-state activity in the medication OFF condition. The patients were then given their morning dose of L-DOPA in the form of fast-acting levodopa. Data were acquired in three runs of 10 min, for a total of 30 min for each medication condition. We started the ON medication measurement at least half an hour after the administration of the dose and after clinical improvement was seen. The same procedure as for the OFF medication state was followed for the ON medication measurement.

Pre-processing

Request a detailed protocol

All data processing and analyses were performed using Matlab (version R 2016b; Math Works, Natick, MA). Custom-written Matlab scripts (https://github.com/saltwater-tensor/HMM_pipeline (copy archived at swh:1:rev:277a6a0ff21ff6885815c934255f953a97e16e98)); Sharma et al., 2021a Sharma et al., 2021b and the Brainstorm toolbox (http://neuroimage.usc.edu/brainstorm/Introduction) were used (Tadel et al., 2011). To ensure artefact-free data, two people independently inspected the data visually, cleaned artefacts, and compared the cleaning output. The final cleaned data included changes agreed upon by both the people involved in cleaning. The Neuromag system provides signal-space projection (SSP) vectors for the cleaning of external artefacts from the MEG channels, which were applied. The line noise was removed from all channels with a notch filter at 50, 100, 150, …, 550, and 600 Hz with a 3 dB bandwidth of 1 Hz. The LFP recordings from the DBS electrode were re-referenced against the mean of all LFP channels. Very noisy and flat MEG/LFP channels were excluded from further analysis. Time segments containing artefacts were removed from the time series. However, if artefacts regularly occurred only in one single channel, this whole channel was removed instead. Frequently arising artefacts following the same basic pattern, such as eye blinks or cardiac artefacts, were removed via SSP. All data were high-pass filtered with 1 Hz to remove movement-related low-frequency artefacts. Finally, the data were down-sampled to 1000 Hz.

Source estimation was performed on these recordings at an individual level using each individual’s anatomy. Therefore, using Freesurfer (https://surfer.nmr.mgh.harvard.edu/, v.5.3.0), the individual cortical surfaces were extracted from the individual T1-weighted MRI scans (3T scanner and 1 mm³ voxel size). We used the overlapping spheres method with 306 spheres for the forward model. As the inverse model, we used a linearly constrained minimum variance (LCMV) beamformer. The data covariance matrix for the LCMV beamformer was computed directly from each 10 min recording. The data covariance was regularised using the median eigenvalue of the data covariance matrix. The noise covariance was obtained from an empty room recording on the same day as the actual measurement.

For each subject, the invasive entry point of the STN was identified based on intraoperative microelectrode recordings (Gross et al., 2006; Moran et al., 2006). Subsequently, the first recording height after the entry into the STN was selected to obtain the three directional LFP recordings from the respective hemisphere. In addition, we visualised the location of all electrodes using lead-DBS (Horn et al., 2019). All electrodes were properly placed within the STN – except for one (see Figure 7). To exclude that our results were driven by outlier, we reanalysed our data without this patient. No qualitative change in the overall connectivity pattern was observed.

Deep brain stimulation (DBS) electrode location for all subjects.

Lead-DBS reconstruction with all subjects. The red leads are the ones of a subject with one of the outside the STN. The red directional contacts are the ones from which the data was used for analysis.

The source-reconstructed MEG data were projected to the default cortical anatomy (MNI 152 with 15,002 vertices) and then down-sampled temporally to 250 Hz for each medication condition for every subject. We used the Mindboggle atlas to spatially reduce the data dimensions. For each of the 42 cortical regions in the atlas, a multidimensional time series consisting of the vertices within that anatomical region was extracted. To reduce the multivariate times series for each region to a single one, we employed the first principal component explaining the highest variance share in each region. The first principal component row vectors from all 42 anatomical regions were stacked into a MEG cortical time series matrix. To correct for volume conduction in the signal, symmetric orthogonalisation (Colclough et al., 2015) was applied to each subject’s resulting MEG cortical time series matrix. The row vectors of this orthogonalised matrix and the six LFPs (three each for left and right STN) were z-scored. Subsequently, they were stacked into one multidimensional time series (N by T) matrix. Here, N = 48 is the total number of nodes/regions (42 regions from the cortex and 6 LFP electrode contacts) and T denotes the length of the time dimension. This 48 by T data matrix obtained from each subject was concatenated along the temporal dimension across all subjects for each specific medication condition. Finally, to resolve sign ambiguity inherent in source-reconstructed MEG data as well as resolve polarity of LFP channels across subjects, a sign-flip correction (Vidaurre et al., 2016) procedure was applied to this final 48 by (T by number of subjects) dataset within a medication condition. The pre-processing steps were performed for OFF and ON medication separately.

HMM analysis

Request a detailed protocol

The HMM is a data-driven probabilistic algorithm which finds recurrent network patterns in multivariate time series (Vidaurre et al., 2016; Vidaurre et al., 2018a). Each network pattern is referred to as a ‘state’ in the HMM framework, such that these networks can activate or deactivate at various points in time. Here onwards, ‘state’ or ‘network’ is used interchangeably. We used a specific variety of the HMM, the TDE-HMM, where whole-brain networks are defined in terms of both spectral power and phase coupling (Vidaurre et al., 2018b). Hence, for every time point, the HMM algorithm provided the probability that a network is active. Here onwards, a contiguous block of time for which the probability of a particular network being active remained higher than all the other networks is referred to as a ‘state visit’. Hence, the HMM produced temporally resolved spatial networks for the underlying time series. In our approach, we also performed spectral analyses of these state visits, leading to a complete spatio-spectral connectivity profile across the cortex and the STN. By applying the HMM analysis to the combined MEG–LFP dataset, we were able to temporally, spatially, and spectrally separate cortico–cortical, cortico–STN, and STN–STN networks.

Estimation of the HMM

Request a detailed protocol

Since we were interested in recovering phase-related networks, the TDE-HMM was fit directly on the time series obtained after pre-processing steps described previously, as opposed to its power envelope. This preserved the cross-covariance within and across the underlying raw time series of the cortical regions and the STN. The model estimation finds recurrent patterns of covariance between regions (42 cortical regions and 6 STN contacts) and segregates them into ‘states’ or ‘networks’. Based on these covariance patterns, for each state, the power spectra of each cortical region and the coherence amongst regions can be extracted.

We opted for six different states as a reasonable trade-off between the spectral quality of the results and their redundancy. The HMM-MAR toolbox (Vidaurre et al., 2016) was used for fitting the TDE-HMM. We employed the TDE version of the HMM where the embedding took place in a 60 ms window (i.e., a 15 time point window for a sampling frequency of 250 Hz). Since time embedding would increase the number of rows of the data from 48 to 48 times the window length (also referred to as number of lags), an additional PCA (principal component analysis) (reduction across 48 by number of lags) step was performed after time embedding. The number of components retained was 96 (48 × 2). This approach follows Vidaurre et al., 2018b. To characterise each state, a full covariance matrix with an inverse Wishart prior was used. The diagonal of the prior for the transition probability matrix was set as 10. To ensure that the mean of the time series did not take part in driving the states, the ‘zero mean’ option in HMM toolbox was set to 1. To speed up the process of fitting, we used the stochastic version of variational inference for the HMM. In order to start the optimisation process, the ‘HMM-MAR’-type initialisation was used (for details, see Vidaurre et al., 2016). The HMM was fit separately OFF and ON medication.

Statistical analysis of the states

Request a detailed protocol

After the six states were obtained for HMM OFF and HMM ON medication, these states were statistically compared within each medication condition as well as between medication conditions. In addition, the temporal properties of these states were compared.

Intra-medication analysis

Request a detailed protocol

We investigated the spectral connectivity patterns across the different states within a medication condition (intra-medication or IntraMed). The objective was to uncover significant coherent connectivity standing out from the background within each frequency band (delta/theta [1–8 Hz], alpha [8–12 Hz], and beta [13–30 Hz]) in the respective states. The HMM output included the state time courses (i.e., when the states activated) for the entire concatenated data time series. The state time courses allowed the extraction of state- and subject-specific data for further state- and subject-level analysis. For each HMM state, we filtered the state-specific data for all the subjects between 1 and 45 Hz. (For state-wise data extraction, please refer the HMM toolbox wiki [https://github.com/OHBA-analysis/HMM-MAR/wiki/User-Guide].) Then we calculated the Fourier transform of the data using a multitaper approach to extract the frequency components from the short segments of each state visit. (See Vidaurre et al., 2018b for discussion on multitaper for short time data segments.) Seven Slepian tapers with a time–bandwidth product of 4 were used, resulting in a frequency resolution of 0.5 Hz and therefore binned frequency domain values. Subsequently, we calculated the coherence and power spectral density of this binned (frequency bins obtained during the multitaper step) data for every subject and every state. The coherence and the power spectral density obtained were three-dimensional matrices of size f (number of frequency bins) by N (42 cortical locations + 6 STN contacts) by N.

Based on the coherence matrices, we performed a frequency band-specific analysis. Canonical definitions of frequency bands assign equal weight to each frequency bin within a band for every subject. This might not be suitable when considering analyses of brain signals across a large dataset. For example, the beta peak varies between individual subjects. Assigning the same weight to each bin in the beta range might reduce the beta effect at the group level. To allow for inter-subject variability in each frequency bin’s contribution to a frequency band, we determined the frequency modes in a data-driven manner (Vidaurre et al., 2018b). Because we focused on interactions that are important to establish the STN–cortex communication, the identification of the relevant frequency modes was restricted to the cross-coherence between the STN–LFPs and cortical signals; in other words, the block matrix consisting of rows 1–6 (STN) and columns 7–48 (cortex). For each subject, this extracted submatrix was then vectorised across columns. This gave us a (number of frequency bins by 252 [6 STN locations by 42 cortical locations]) matrix for each state. For every subject, this matrix was concatenated along the spatial dimension across all states producing a (number of frequency bins by [252 by 6 (number of states)]) matrix. We called this the subject-level coherence matrix. We averaged these matrices across all subjects along the spectral dimension (number of frequency bins) to yield a (number of frequency bins by [252 by 6]) group-level coherence matrix. We factorised the group-level coherence matrix into four frequency modes using a non-negative matrix factorisation (NNMF) (Lee and Seung, 2001). Each of the resulting four frequency modes obtained was of size (one by number of frequency bins). The values of frequency modes are the actual NNMF weights obtained from the NNMF estimation (which, just like a regression coefficient, are unit-less, because coherence is unit-less). Three of them resembled the canonical delta/theta (delta and theta frequencies were combined into one band), alpha, and beta bands whereas the last one represented noise. Since NNMF does not guarantee a unique solution, we performed multiple instances of the factorisation. In practice we could obtain frequency modes, which showed correspondence to the classical frequency bands, within four iterations of the algorithm. At each instance, we visualised the output to ensure frequency specificity of the frequency modes. The stability of the output was ensured by using ‘robust NNMF’, which is a variant of the NNMF algorithm (Vidaurre et al., 2018b). While these frequency modes were derived in fact from coherence measures (as detailed in Vidaurre et al., 2018a), they can be applied to power measures or any other frequency-specific measure. We then computed the inner product between the subject- and group-level coherence matrix and the frequency modes obtained above. We called these the subject-level and group-level projection results, respectively.

To separate background noise from the strongest coherent connections, a Gaussian mixture model (GMM) approach was used (Vidaurre et al., 2018b). For the group-level projection results, we normalised the activity in each state for each spectral band by subtracting the mean coherence within each frequency mode across all states. As a prior for the mixture model, we used two single-dimensional Gaussian distributions with unit variance: one mixture component to capture noise and the other to capture the significant connections. This GMM with two mixtures was applied to the coherence values (absolute value) of each state. Connections were considered significant if their p-value after correction for multiple comparisons was smaller than 0.05.

Inter-medication analysis

Request a detailed protocol

To test for differences in coherence across medication conditions (inter-medication or InterMed), the first step was to objectively establish a comparison between the states found in the two HMMs fit separately for each condition. There is no a priori reason for the states detected in each condition to resemble each other. To find OFF and ON medication states that may resemble each other, we calculated the Riemannian distance (Förstner and Moonen, 2003) between the state covariance matrices of the OFF and ON HMM. This yielded an OFF states by ON states (6 × 6) distance matrix. Subsequently, finding the appropriately matched OFF and ON states reduced to a standard linear assignment problem. We found an ON state counterpart to each OFF state by minimising the total sum of distances using the Munkres linear assignment algorithm (Vidaurre et al., 2018a). This approach yielded a one-to-one pairing of OFF and ON medication states, and all further analysis was conducted on these pairs. For ease of reading, we gave each pair its own label. For example, when we refer to a ‘Ctx–STN’ state in the following sections, then such a state was discovered OFF medication and its corresponding state ON medication is its distance-matched partner. In subsequent sections, all mentions of ON or OFF medication refer to these state pairs unless mentioned otherwise.

We used the subject-level projection results obtained during IntraMed analysis to perform InterMed analyses. We performed two-sided independent sample t-tests between the matched states to compare the coherence, which was calculated between different regions of interest (see Dataset preparation). We grouped individual atlas regions into canonical cortical regions like frontal, sensorimotor, parietal, visual, medial PFC (prefrontal cortex), and STN contacts. For example, in the beta band, STN (contacts)–sensorimotor coherence in the OFF condition was compared to the STN (contacts)–sensorimotor coherence in the ON condition. The p-values obtained were corrected for multiple comparisons for a total number of possible combinations.

Temporal properties of HMM states

Request a detailed protocol

To test for changes in the temporal properties OFF versus ON medication, we compared the lifetimes, interval between visits, and FO for each state both within and across HMMs using two-way repeated measures ANOVA followed by post hoc tests. Lifetime/dwell time of a state refers to the time spent by the neural activity in that state. Interval of visit was defined as the time between successive visits of the same state. Finally, the FO of a state was defined as the fraction of time spent in each state. Extremely short state visits might not reflect neural processes, hence we only used values that were greater than 100 ms for lifetime comparisons.

Data availability

We have made the code to produce the results and generate the figures available on Github: https://github.com/saltwater-tensor/HMM_pipeline (copy archived athttps://archive.softwareheritage.org/swh:1:rev:277a6a0ff21ff6885815c934255f953a97e16e98). However, the raw data cannot be made publicly available due to the European and German data privacy laws. When signing the informed consent forms, our patients consented to using their data for research purposes, but they did not sign a form stating that their data can be shared publicly, even in anonymized form. In addition, the MRIs of their heads and brains might make them identifiable. Hence, posting them within a public repository is not possible. The raw data can be requested from the corresponding author for replication of the current results and will then be shared in an anonymized way. We are providing the intermediate Matlab data underlying the figures with our submission.

References

  1. Book
    1. Förstner W
    2. Moonen B
    (2003) A metric for covariance matrices
    In: Grafarend E. W, editors. Geodesy-the Challenge of the 3rd Millennium. Berlin, Germany: Springer. pp. 299–309.
    https://doi.org/10.1007/978-3-662-05296-9
  2. Conference
    1. Lee DD
    2. Seung HS
    (2001)
    Algorithms for non-negative matrix factorization
    Advances in Neural Information Processing Systems.

Decision letter

  1. Andreas Horn
    Reviewing Editor; Charité - Universitätsmedizin Berlin, Germany
  2. Michael J Frank
    Senior Editor; Brown University, United States
  3. Kelly Bijanki
    Reviewer
  4. Muthuraman Muthuraman
    Reviewer; University Medical Center of the Johannes Gutenberg University Mainz, Germany

Our editorial process produces two outputs: i) public reviews designed to be posted alongside the preprint for the benefit of readers; ii) feedback on the manuscript for the authors, including requests for revisions, shown below. We also include an acceptance summary that explains what the editors found interesting or important about the work.

Acceptance summary:

In this manuscript, Sharma et al., investigated the effect of dopamine administration on oscillatory whole-brain networks by means of simultaneous local field potential recordings from the subthalamic nucleus and whole-brain magnetoencephalography recordings in seventeen patients with Parkinson's disease. A key feature is the combined use of invasive and noninvasive recordings and to investigate network changes by employing a hidden markov model. They identified three physiologically interpretable spectral connectivity patterns and found that cortico-cortical, cortico-STN, and STN-STN networks were differentially modulated by dopaminergic medication. These findings provide new insights regarding the mechanisms by which dopamine and medications alter cortico-basal ganglia dynamics, and open up new directions for studying their functions.

Decision letter after peer review:

Thank you for submitting your article "Differential dopaminergic modulation of spontaneous cortico-subthalamic activity in Parkinson's disease" for consideration by eLife. Your article has been reviewed by 3 peer reviewers, and the evaluation has been overseen by a Reviewing Editor and Michael Frank as the Senior Editor. The following individuals involved in review of your submission have agreed to reveal their identity: Kelly Bijanki (Reviewer #1); Muthuraman Muthuraman (Reviewer #2).

The reviewers have discussed their reviews with one another, and the Reviewing Editor has drafted this to help you prepare a revised submission.

Essential Revisions:

1. I believe one strong concern is the first one raised by reviewer #1 in that it is important to differentiate neural activity differences in the ON medication state in whether they may originate from the effects of dopamine on the brain or the fact that the patients are moving more. Is there any possibility to address this issue? E.g. were motion parameters analysed to some degree? Does archival data exist that could be used to differentiate states at rest vs. states with movement (e.g. finger tapping)? If not this limitation should be prominently discussed.

2. While I personally think the concern of rev. 3 with contamination of STN signals by EEG is not of major concern (and the method has been established with reproducible results e.g. based on the Litvak / Hirschmann / Neumann works). Still, electrode localization could help to at least assure all leads were properly placed (or if not results would remain stable if the ones outside STN would be discarded)

3. I agree with rev. #2 that presentation can be improved. While circular connection plots are informative, it could be helpful to see to results in Figures 2-4 mapped to a brain (or sensor space). In general, I think the Results section could be stratified and presentation optimized.

Reviewer #1 (Recommendations for the authors):

Below is a list of concerns upon reviewing the manuscript, with suggestions for improvement.

1. The style of writing in the introduction is a little light on examination of the prior literature and motivation for the study. Starting at line 79, it transitions into a preview of the methods section listing of the analyses the researchers undertook and less like a true introduction setting the scene for the experiments.

2. This reviewer was surprised by the use of the word "communication", communication state, comms, etc. Perhaps this is a term of art in MEG (not this reviewer's area of expertise) but defining an MEG feature as indicative of communication seems presumptuous and merits at least brief discussion in the manuscript. Please add more background on MEG analysis and interpretation, especially in PD, in the introduction section.

3. This reviewer was unclear on why increased "communication" in the medial OFC in δ and theta was interpreted as a pathological state indicating deteriorated frontal executive function. Given that the authors provide no evidence of poor executive function in the patients studied, the authors must at least provide evidence from other studies linking this feature with impaired executive function.

4. Authors further reported that increased DA (L-DOPA administration) caused β activity to switch from STN-mediated motor network to a frontoparietal mediated one. The authors provide somewhat impoverished anatomical detail about the differences between the observed STN-mediated motor network and the known pathological β activity in STN in PD.

5. Last, authors report that DA didn't modify locally-originating STN oscillations in PD, but the detail on how they define locally-originating is unclear.

6. On line 86 the authors identify prior research limited by its investigation of specific connectivity pairs, whereas on line 104 the authors report on connectivity pairs in the current study.

7. The authors use the term "oscillations" without properly defining it in terms of neural activity and they have not addressed the analytical approach they use to confirm the activity measured on LFP or MEG reflects oscillatory activity vs. bursting activity or other neural activity types.

8. Authors need to acknowledge the role of DBS in PD therapy earlier in the study and identify that as the means by which they have access to simultaneous LFP recording. They don't even define the acronym "DBS" at its first use.

9. Authors could be more clear in lines 90-97 to define what they mean by "disruptive, "physiologically restorative", and "limited".

10. Authors repeatedly state their method allows them to delineate between pathological and physiological connectivity, but they don't explain how dynamical systems and discrete-state stochasticity support that goal. Lines 106 and 111.

11. Authors must address differences in neural activity other than DA-mediated changes in the on and off medication state. For example, when patients are off med, their UPDRS scores are elevated – do they not have movements that would pick up extra activations in MEG during "rest"? Is it possible to do a true "resting state" in active PD? At minimum, this concern must be discussed in the manuscript.

12. Figure 1: this reviewer would like to see the Y axis indicating amplitude or power, rather than arbitrary units from NNMF output. This figure also begs the question why the NMMF categorized 10-20hz oscillations as δ/theta (bump in blue line in "ON medication" and above 30Hz as α (red line, same plot)). To this reviewer, it appears as though the NMMF factorization worked effectively in the "OFF medication" data set but failed in the "ON medication" data set.

13. Patient sample appears to under-represent female patients. Is there a relationship between sex and DA metabolism/uptake/MEG markers? Perhaps simply acknowledge the imbalance in the manuscript.

14. LFP recordings were sampled from externalized DBS leads using the St. Jude directional 6172 lead except in one patient who had the Boston Scientific leads (which I think is fine to include in analysis together provided surface areas of contacts are identical). The geometry of the segmented leads needs to be addressed relative to sensitivity for various frequencies of oscillatory activity. Segmented leads have very low surface area and may require ultra-high sampling rates (30 kHz and higher) to resolve oscillatory activity.

Reviewer #2 (Recommendations for the authors):

General remarks:

1. A visualization of the state time series would be a nice addition to see the dynamics of the network. Including the probabilities would further give an impression of "how clear the states are", i.e. were they exclusive or were there some intervals where 2 states were rather active at the same time.

2. The color coding in table 1 (4x orange) does not correspond to the ring figures (2x orange/hemisphere). Both areas 17 and 18 are termed "medial orbitofrontal".

Results

As stated in the public review, there are several points concerning the presentation of the results:

3. Line 180 Supporting the dopamine overdose hypothesis in PD, we identified a δ/theta oscillatory network involving lateral and medial orbitofrontal cortical regions.

– Figure 2 also shows the inter-hemispheric connectivity of the pars orbitalis. What is the reason to not mention this in the main text?

4. Line 259: Furthermore, ON medication in the α band only the connectivity between temporal and parietal cortical regions and the STN was preserved (Figure 3C α; p < 0.05), consistent with previous findings (Litvak et al. 2011). In contrast, in the β band only STN-medial orbitofrontal connectivity remained intact (p < 0.05, Figure 3C β).

– Similar to the comment on line 180, the figure also shows α connectivity between the STN and medial orbitofrontal cortex.

5. Line 269: Finally, ON medication, a sensorimotor-frontoparietal network emerged (p < 0.05, Figure 3C β) where sensorimotor, frontal, and parietal regions were no longer connected to the STN, but instead directly communicated with each other in the β band.

– Similar to the comment on line 180: is there a reason to exclude the connection from the somatomotor cortex to the caudal middle frontal (e.g. figure 3c, β: 13r-16r)?

6. 256: Importantly, coherence OFF medication was significantly larger than ON medication between STN and sensorimotor, STN and temporal and STN and frontal cortices (p < 0.05 for all connections, Figure 3B α and β).

– The figure does not these results for STN and sensorimotor cortex

7. 263: Most previous PD studies report a decrease in the motor-STN coherence ON medication in the β band (Hammond, Bergman, and Brown 2007; Litvak et al. 2011; Hirschmann et al. 2013; Little et al. 2013; Marinelli et al. 2017) but do not indicate any changes that the sensorimotor regions might experience at the whole-brain level. In the communication state, OFF medication, STN-pre-motor (sensory), STN-frontal, and STN-parietal connectivity was present (p < 0.05, Figure 3A α and β). STN-cortical coherence was then significantly reduced ON compared to OFF medication (p < 0.05, Figure 3B α and β).

– The first sentence introduces cortico-cortical connections, which is then, however, followed by STN-cortical results. This is then followed by results on the sensorimotor-frontoparietal network, which indeed is cortico-cortical. However, later in the manuscript, the following statement suggests that the cortico-cortical network differs from the sensorimotor-frontoparietal network: "Still, significant connectivity was selectively preserved in a spectrally-specific manner ON medication both at the corticocortical level and the cortical-STN level. Furthermore, a sensorimotor-frontoparietal network emerged ON medication" (line282). It is not clear what exactly resembles the cortico-cortical network, and where the results are presented.

On the results of temporal properties:

8. Line 324: Previous research has shown that ON medication, spectrally-specific cortico-STN connectivity remains preserved in PD compared to OFF medication (Litvak et al. 2011; Hirschmann et al. 2013). This indicates the existence of functionally relevant cortico-STN loops. A decrease in coherence between the cortex and the STN has also been observed ON medication (George et al. 2013), which was correlated with improved motor functions in PD. All the connectivity effects were observed in our results for the communication state. Furthermore, in the communication state, we showed the existence of a frontoparietal sensorimotor 331 network in the β band ON medication. Recent evidence indicates that with the loss of dopamine and the start of PD symptoms, δ/theta oscillations emerge within the basal ganglia (Whalen et al. 2020). In line with these findings, STN-STN δ/theta oscillations in the local state were reduced ON medication.

– I have the impression that this paragraph is not in the right position. It repeats and even discusses the previously mentioned results and jumps from "frontoparietal sensorimotor networks" to "STN-STN δ/theta" findings.

Further, it is not clear if "All the connectivity effects were observed in our results for the communication state" relates to the citations above, especially since other states are shown to have effects on connectivity as well.

9. Additionally, the main text should provide t-scores and degrees of freedom, not only p-values. The comprehension of the results could be improved by displaying significantly different comparisons in a clear way in Figure 5 (e.g. provide * for p<0.05). However, what was the rationale for testing the medication condition and states with different t-tests? Would a 2way- repeated measures ANOVA not be more appropriate?

10. Line 353: The lifetimes for both the local and communication state were significantly increased by medication (ON > OFF: local, p < 0.01; comms, p {less than or equal to} 0.01).

and line 360: Both the local and communication state tended to last longer ON medication.

– This seems to be a repetition. Also, it is not clear what "tended to last longer" means? Is there a significant difference or a trend?

Methods

11. Line 676: Since NNMF does not guarantee a unique solution, we performed multiple instances of the factorization.

– The authors should state the exact number of repetitions. How was this number decided?Reviewer #3 (Recommendations for the authors):

1. The Authors could examine the effect of the dopaminergic medication (ON and OFF) on the recurrent oscillatory patterns of transient network connectivity within and between the cortex and the STN, as a function of the duration of the disease and/or the DRT history.

2. Re the contamination of the STN signal by volume conducted signals from the cortex. Page 20 (lines 584-586) "To correct for volume conduction in the signal, symmetric orthogonalisation (Colclough et al., 2015) was applied to each subject's resulting cortical time series matrix".

Have the authors applied this correction to the STN LFPs? If the authors cannot rule out the possibility of a contamination of STN LFPs by volume conducted signals from the cortex and cannot guarantee that the signal recorded in the STN only reflected local STN activity they should at least interpret their results with caution and discuss this limitation in their discussion.

3. The authors could make the paper more reader-friendly. In particular, could classical spectral and coherence analyses be performed to visualize the characteristics of the oscillatory activity and neural synchronization of the EMG and LFP signals in the recurrent oscillatory patterns of transient network connectivity both ON and OFF medication?

4. In this study, the authors could reconstruct the trajectories of the DBS leads within the STN (using for example the open source LeadDBS program?).

https://doi.org/10.7554/eLife.66057.sa1

Author response

Essential Revisions:

1. I believe one strong concern is the first one raised by reviewer #1 in that it is important to differentiate neural activity differences in the ON medication state in whether they may originate from the effects of dopamine on the brain or the fact that the patients are moving more. Is there any possibility to address this issue? E.g. were motion parameters analysed to some degree? Does archival data exist that could be used to differentiate states at rest vs. states with movement (e.g. finger tapping)? If not this limitation should be prominently discussed.

Regarding the higher movement ON medication during the resting condition, it would mainly be due to hyperkinesia. None of our patients did experience hyperkinesia during the recording. Unfortunately, we do not have actual movement data for our patients during the MEG recording.

Instead, we investigated the UPDRS 3 sub-scores for tremor and akinesia. According to this analysis, one patient was tremor-dominant and one of mixed type. As tremor causes additional movement in the OFF medication condition, we excluded these two patients and recalculated the HMM. ON medication results for all HMM states remained the same. OFF medication results for the Ctx-Ctx and STN-STN state remained the same as well. The Ctx-STN state OFF medication was split into two states: Sensorimotor-STN connectivity was captured in one state and all other types of Ctx-STN connections were captured in another state (see Author response image 1 and Author response image 2). The important point in the context of our study is that these different solutions preserve the biological conclusions: both with and without the two subjects a stable covariance matrix entailing sensorimotor-STN connectivity was determined, which is the main finding for the Ctx-STN state OFF medication.

Author response image 1
States obtained after removing one tremor dominant and one mixed type patient from analysis.
Author response image 2
States obtained after removing one tremor dominant and one mixed type patient from analysis.

Panel C shows the split OFF medication cortico-STN state. Most of the cortico-STN connectivity is captured by the state shown in the top row (Figure 1 C OFF). Only the motor-STN connectivity in the α and β band (along with a medial frontal-STN connection in the α band) is captured separately by the states labeled “OFF Split” (Figure 1 C OFF SPLIT).

We therefore discuss this issue now within the limitation section (page 13):

“Both motor impairment and motor improvement can cause movement during the resting state in PD. While such movement is a deviation from a resting state in healthy subjects, such movements are part of the disease and occur unwillingly. Therefore, such movements can arguably be considered part of the resting state of Parkinson’s disease. None of the patients in our cohort experienced hyperkinesia during the recording. All patients except for two were of the akinetic-rigid subtype. We verified that tremor movement is not driving our results. Recalculating the HMM states without these 2 subjects, even though it slightly changed some particular aspects of the HMM solution, did not materially affect the conclusions.”

2) While I personally think the concern of rev. 3 with contamination of STN signals by EEG is not of major concern (and the method has been established with reproducible results e.g. based on the Litvak / Hirschmann / Neumann works). Still, electrode localization could help to at least assure all leads were properly placed (or if not results would remain stable if the ones outside STN would be discarded)

We thank the review editor for this suggestion. We have reconstructed the electrode location with Lead-DBS (Horn et al., 2019). All electrodes were properly placed within the STN – except for one (see Author response image 3 and Figure 7 in the manuscript). To exclude that our results were driven by outlier, we reanalysed our data without this patient. No apparent change in the overall connectivity pattern was observed (Author response image 4 and Author response image 5). In addition, we now discuss the potential of volume conduction of signals from the cortex to the STN as a potential limitation (page 13-14): “Lastly, we recorded LFPs from within the STN –an established recording procedure during the implantation of DBS electrodes in various neurological and psychiatric diseases. Although for Parkinson patients results on β and tremor activity within the STN have been reproduced by different groups (Reck et al., 2010, Litvak et al., 2011, Florin et al., 2013, Hirschmann et al., 2013, Neumann et al., 2016), it is still not fully clear whether these LFP signals are contaminated by volume-conducted cortical activity. However, while volume conduction seems to be a larger problem in rodents even after re-referencing the LFP signal (Lalla et al., 2017), the same was not found in humans (Marmor et al., 2017).”

Author response image 3
Lead DBS reconstruction of the location of electrodes in the STN for different subjects.

The red electrodes have not been placed properly in the STN. The contacts marked in red represent the directional contacts from which the data was used for analysis.

Author response image 4
HMM states obtained after running the analysis without the subject with the electrode outside the STN.
Author response image 5
HMM states obtained after running the analysis without the subject with the electrode outside the STN.

3) I agree with rev. #2 that presentation can be improved. While circular connection plots are informative, it could be helpful to see to results in Figures 2-4 mapped to a brain (or sensor space). In general, I think the Results section could be stratified and presentation optimized.

We have stratified the Results sections according to the suggestion of reviewer 2. In addition, we are now providing glass-brain views of the connections in figures 2-4.

Reviewer #1 (Recommendations for the authors):

Below is a list of concerns upon reviewing the manuscript, with suggestions for improvement.

1. The style of writing in the introduction is a little light on examination of the prior literature and motivation for the study. Starting at line 79, it transitions into a preview of the methods section listing of the analyses the researchers undertook and less like a true introduction setting the scene for the experiments.

We have changed the introduction and added more background literature (page 2-3):

“Dopamine is a widespread neuromodulator in the brain (Gershman and Uchida 2019), raising the question of whether each medication-induced change restores physiological oscillatory networks. In particular, dopaminergic medication is known to produce cognitive side effects in PD patients (Voon et al., 2009). According to the dopamine overdose hypothesis, a reason for these effects is the presence of excess dopamine in brain regions not affected in PD (MacDonald et al., 2011; MacDonald and Monchi, 2011). Previous task-based and neuroimaging studies in PD demonstrated frontal cognitive impairment due to dopaminergic medication (Cools et al., 2002; Ray and Strafella, 2010; MacDonald et al., 2011).

Using resting-state whole-brain MEG analysis, network changes related to both motor and non-motor symptoms of PD have been described (Olde Dubbelink et al., 2013a, b). However, these studies could not account for simultaneous STN-STN or cortico–STN activity affecting these networks, which would require combined MEG/EEG–LFP recordings (Litvak et al., 2021). Such recordings are possible during the implantation of deep brain stimulation (DBS) electrodes, an accepted treatment in the later stages of PD (Volkmann et al., 2004; Deuschl et al., 2006, Kleiner-Fisman et al., 2006). Combined MEG-LFP studies in PD involving dopaminergic intervention report changes in β and α band connectivity between specific cortical regions and the STN (Litvak et al., 2011; Hirschmann et al., 2013; Oswal et al., 2016). Decreased cortex-STN coherence under dopaminergic medication (ON) correlates with improved motor functions in PD (George et al., 2013). STN–STN intra-hemispheric oscillations positively correlate to motor symptom severity in PD without dopaminergic medication (OFF), whereas dopamine-dependent nonlinear phase relationships exist between inter-hemispheric STN–STN activity (West et al., 2016). Crucially, previous studies could not rule out the influence of cortico–STN connectivity on these inter-hemispheric STN–STN interactions.”

2. This reviewer was surprised by the use of the word "communication", communication state, comms, etc. Perhaps this is a term of art in MEG (not this reviewer's area of expertise), but defining an MEG feature as indicative of communication seems presumptuous and merits at least brief discussion in the manuscript. Please add more background on MEG analysis and interpretation, especially in PD, in the introduction section.

We apologize for any confusion the loaded terminology may have caused. Following this and reviewer 2’s similar comment we renamed the states. Now the communication state is termed Cortico-STN state. In the previous version of the manuscript, we used the word “communication” to denote connectivity between regions based on spectral coherence. The terminology was based on the communication through coherence (CTC) hypothesis (Fries, 2005, 2015), but we think the new label is more descriptive. More background on MEG analysis in PD has been added to the manuscript (see previous reply). In addition, we added the following explanation to the introduction (page 3):

“We study whole-brain connectivity including the STN using spectral coherence as a proxy for communication based on the communication through coherence hypothesis (Fries, 2005, 2015). This will allow us to delineate differences in communication OFF and ON medication”

3. This reviewer was unclear on why increased "communication" in the medial OFC in δ and theta was interpreted as a pathological state indicating deteriorated frontal executive function. Given that the authors provide no evidence of poor executive function in the patients studied, the authors must at least provide evidence from other studies linking this feature with impaired executive function.

If we understand the comment correctly it refers to the statement in the abstract “Dopaminergic medication led to communication within the medial and orbitofrontal cortex in the delta/theta frequency range. This is in line with deteriorated frontal executive functioning as a side effect of dopamine treatment in Parkinson’s disease”

This statement is based on the dopamine overdose hypothesis reported in the Parkinson’s disease (PD) literature (Cools, 2001; Kelly et al., 2009; MacDonald and Monchi, 2011; Vaillancourt et al., 2013). We have elaborated upon the dopamine overdose hypothesis in the discussion on page 10. In short, dopaminergic neurons are primarily lost from the substantia nigra in PD, which causes a higher dopamine depletion in the dorsal striatal circuitry than within the ventral striatal circuits (Kelly et al., 2009; MacDonald and Monchi, 2011). Thus, dopaminergic medication to treat the PD motor symptoms leads to increased dopamine levels in the ventral striatal circuits including frontal cortical activity, which can potentially explain the cognitive deficits observed in PD (Shohamy et al., 2005; George et al., 2013). We adjusted the abstract to read:

“Dopaminergic medication led to coherence within the medial and orbitofrontal cortex in the delta/theta frequency range. This is in line with known side effects of dopamine treatment such as deteriorated executive functions in Parkinson’s disease.”

4. Authors further reported that increased DA (L-DOPA administration) caused β activity to switch from STN-mediated motor network to a frontoparietal mediated one. The authors provide somewhat impoverished anatomical detail about the differences between the observed STN-mediated motor network and the known pathological β activity in STN in PD.

Prior research has indicated that β activity within the dorsolateral STN is related to the pathology of PD (Marreiros et al., 2013; van Wijk et al., 2016; West et al., 2018). To provide more anatomical detail on the electrode location we reconstructed the electrode location using Lead-DBS (Horn et al., 2019). Except for one electrode, all electrode contacts used in the analysis were located within the dorsolateral STN. We have added a figure with the electrode locations to the manuscript (see manuscript figure 7 and Author response image 3).

Consistent with previous studies in the β frequency band OFF medication, we found STN-sensorimotor (Hirschmann et al. 2011), STN-parietal (inferior parietal) (Litvak et al., 2011), STN-frontal (insular cortex, pars orbitalis, pars opercularis and lateral orbitofrontal) (Litvak et al., 2011), and STN-medial prefrontal (medial orbitofrontal) (Litvak et al., 2011) coherence in our Ctx-STN state results. ON medication sensorimotor regions were coherent with parietal (para central) and frontal (superior frontal)/ medial prefrontal (caudal middle frontal) regions in the β band. Previous research has not reported the emergence of such a coherent network ON medication. But consistent with previous research, we found that sensorimotor-STN coherence was reduced ON medication (Hirschmann et al., 2013).

We provide details in the discussion on page 11:

“Previous studies have reported STN-sensorimotor (Hirschmann et al., 2011, Litvak et. al., 2011), STN-parietal, and STN-frontal (Litvak et al., 2011) coherence in the β band OFF medication. Consistent with previous studies STN-sensorimotor, STN-parietal (inferior parietal), STN-frontal (Insular cortex, pars orbitalis, pars opercularis and lateral orbitofrontal), and STN-medial prefrontal (medial orbitofrontal) coherence emerged in the Ctx-STN state. In contrast, ON medication sensorimotor regions were coherent with parietal (para central) and frontal (superior frontal)/ medial prefrontal (caudal middle frontal) regions in the β frequency range. Previous research has not reported the emergence of such a coherent fronto-parietal-motor network ON medication. But consistent with previous research (Hirschmann et al., 2013), sensorimotor-STN coherence was reduced ON compared to OFF medication.”

5. Last, authors report that DA didn't modify locally-originating STN oscillations in PD, but the detail on how they define locally-originating is unclear.

We apologize for not more clearly defining this term. Because only STN-STN coherence/connectivity is altered under medication in the “local state”, we termed this state “local state”. But we agree the term is confusing and changed it to “STN-STN state”. Unfortunately, identifying whether the STN oscillations are indeed originating locally within basal ganglia circuits would require a higher coverage of invasive recording locations as well as intracellular recordings. We changed the abstract to:

“In contrast, dopamine did not modify local STN-STN coherence in PD.”

6. On line 86 the authors identify prior research limited by its investigation of specific connectivity pairs, whereas on line 104 the authors report on connectivity pairs in the current study.

We are sorry for this unclear description. The line stating “Each HMM state was characterised by coherence calculated between different pair of regions” was supposed to intuitively explain what an HMM state is: a multidimensional, time-delay embedded covariance matrix capturing frequency-specific connectivity between different brain regions. Each HMM state can subsequently be analysed in a spectrally-specific manner. To eliminate this seemingly contradictory statement we changed this sentence to (page 4): “Each HMM state itself is a multidimensional, time-delay embedded covariance matrix across the whole brain, containing information about cross-regional coherence and power in the frequency domain.”

Analysing whole brain connectivity instead of investigating pre-specified bilateral connectivity pairs is an important difference to previous studies. Thus, we investigated coherence between multiple regions and report only those connectivity pairs that significantly stand out from all the possible functional connections.

7. The authors use the term "oscillations" without properly defining it in terms of neural activity and they have not addressed the analytical approach they use to confirm the activity measured on LFP or MEG reflects oscillatory activity vs. bursting activity or other neural activity types.

For the purpose of our paper, oscillatory activity or oscillations refers to recurrent but transient frequency–specific patterns of network activity. Because our analysis pipeline relies on an HMM, the oscillatory connectivity under consideration is neither exclusively made up of sustained rhythmic activity nor neural bursting, and it can be a mix of both (van Ede et al., 2018; Quinn et al., 2019). We have added this information as a footnote to the paper on page 2.

“For the purposes of our paper we refer to oscillatory activity or oscillations as recurrent but transient frequency–specific patterns of network activity, even though the underlying patterns can be composed of either sustained rhythmic activity, neural bursting, or both (Quinn et al., 2019). Disambiguating the exact nature of these patterns is, however, beyond the scope of the present work.”

8. Authors need to acknowledge the role of DBS in PD therapy earlier in the study and identify that as the means by which they have access to simultaneous LFP recording. They don't even define the acronym "DBS" at its first use.

Thank you for pointing this out. We now define DBS at its first use and added the information on DBS surgery as a further treatment option to the introduction. We added on page 2:

“However, these studies could not account for simultaneous STN-STN or cortico–STN activity affecting these networks, which would require combined MEG/EEG–LFP recordings (Litvak et al. 2021). Such recordings are possible during the implantation of deep brain stimulation (DBS) electrodes, an accepted treatment in the later stages of PD (Volkmann et al., 2004; Deuschl et al., 2006, Kleiner-Fisman et al., 2006)”

9. Authors could be more clear in lines 90-97 to define what they mean by "disruptive, "physiologically restorative", and "limited".

We have changed those lines to make the reported findings more accessible. We used “disruptive” to denote that connectivity changes due to dopamine were related to the cognitive side effects reported in literature. “Physiologically restorative” was supposed to refer to connectivity changes on medication that brought connectivity closer to normal physiological connectivity as previously reported in PD. Finally, “limited” referred to the fact that there were only changes in STN-STN coherence in that state.

The relevant passage now reads (page 3):

“For the cortico–cortical network medication led to additional connections that can be linked to the side effects of dopamine. At the same time, dopamine changed the cortico-STN network towards a pattern more closely resembling physiological connectivity as reported in the PD literature. Within the third network, dopamine only had an influence on local STN-STN coherence. “

10. Authors repeatedly state their method allows them to delineate between pathological and physiological connectivity, but they don't explain how dynamical systems and discrete-state stochasticity support that goal. Lines 106 and 111.

To recapitulate, the HMM divides a continuous time series into discrete states. Each state is a time-delay embedded covariance matrix reflecting the underlying connectivity between brain regions as well as the specific temporal dynamics in the data when such state is active. See Packard et al., (1980) for details about how a time-delay embedding characterises a linear dynamical system.

Please note that the HMM was used as a data-driven, descriptive approach without explicitly assuming any a-priori relationship with pathological or physiological states. The relation between biology and the HMM states, thus, purely emerged from the data; i.e. is empirical. What we claim in this work is simply that the features captured by the HMM hold some relation with the physiology even though the estimation of the HMM was completely unsupervised (i.e. blind to the studied conditions). We have added this point also to the limitations of the study on page 18 and the following to the introduction to guide the reader more intuitively (page 4):

“To allow the system to dynamically evolve, we use time delay embedding. Theoretically, delay embedding can reveal the state space of the underlying dynamical system (Packard et al., 1980). Thus, by delay-embedding PD time series OFF and ON medication we uncover the differential effects of a neurotransmitter such as dopamine on underlying whole brain connectivity.”

11. Authors must address differences in neural activity other than DA-mediated changes in the on and off medication state. For example, when patients are off med, their UPDRS scores are elevated – do they not have movements that would pick up extra activations in MEG during "rest"? Is it possible to do a true "resting state" in active PD? At minimum this concern must be discussed in the manuscript.

We agree that Parkinson’s disease can lead to unwanted movements such as tremor as well as hyperkinesias. This would of course be a deviation from a resting state in healthy subjects. However, such movements are part of the disease and occur unwillingly. The main tremor in Parkinson’s disease is a rest tremor and – as the name already suggests – it occurs while not doing anything. Therefore, such movements can arguably be considered part of the resting state of Parkinson’s disease. Resting state activity with and without medication is therefore still representative for changes in brain activity in Parkinson’s patients and indicative of alterations due to medication.

To further investigate the effect of movement in our patients, we subdivided the UPDRS part 3 score into tremor and non-tremor subscores. For the tremor subscore we took the mean of item 15 and 17 of the UPDRS, whereas for the non-tremor subscore items 1, 2, 3, 9, 10, 12, 13, and 14 were averaged. Following Spiegel et al., 2007, we classified patients as akinetic-rigid (non-tremor score at least twice the tremor score), tremor-dominant (tremor score at least twice as large as the non-tremor score), and mixed type (for the remaining scores). Of the 17 patients, 1 was tremor dominant and 1 was classified as mixed type (his/her non-tremor score was greater than tremor score). None of our patients exhibited hyperkinesias during the recording. To exclude that our results are driven by tremor-related movement, we re-ran the HMM without the tremor-dominant and the mixed-type patient (see Author response image 1 and Author response image 2).

ON medication results for all HMM states remained the same. OFF medication results for the Ctx-Ctx and STN-STN state remained the same as well. The Ctx-STN state OFF medication was split into two states: Sensorimotor-STN connectivity was captured in one state and all other types of Ctx-STN connections were captured in another state (see Figure 1 response letter). The important point is that the biological conclusions stand across these solutions. Regardless, both with and without the two subjects a stable covariance matrix entailing sensorimotor-STN connectivity was determined, which is the main finding for the Ctx-STN state OFF medication.

We therefore discuss this issue now within the limitation section (page 13):

“Both motor impairment and motor improvement can cause movement during the resting state in PD. While such movement is a deviation from a resting state in healthy subjects, such movements are part of the disease and occur unwillingly. Therefore, such movements can arguably be considered part of the resting state of Parkinson’s disease. None of the patients in our cohort experienced hyperkinesia during the recording. All patients except for two were of the akinetic-rigid subtype. We verified that tremor movement is not driving our results. Recalculating the HMM states without these 2 subjects, even though it slightly changed some particular aspects of the HMM solution did not materially affect the conclusions.”

12. Figure 1: this reviewer would like to see the Y axis indicating amplitude or power, rather than arbitrary units from NNMF output. This figure also begs the question why the NMMF categorized 10-20hz oscillations as δ/theta (bump in blue line in "ON medication" and above 30Hz as α (red line, same plot). To this reviewer, it appears as though the NMMF factorization worked effectively in the "OFF medication" data set but failed in the "ON medication" data set.

We apologise for the misunderstanding. The graphs in manuscript figure 1 do not reflect power, but the spectral templates derived from the NNMF method; that is, these values are the actual NNMF weights obtained from the NNMF estimation. These templates were in fact derived from coherence measures (as detailed in Vidaurre et al., 2018), which is a unit less measure. Therefore, these NNMF weights also don’t have a unit. We have clarified this in methods section titled Intra-medication analysis on page 20.

Please also note that NNMF is a data-driven technique and it is therefore natural that the spectral templates do not exactly match the canonical frequency bands. In particular, it is common that residual “bumps” emerge besides the main characteristic frequency (for example, as seen in the “α” band ON medication). However, given that all the results are statistically thresholded at a p-value of 0.05, any connections purely arising from the frequency bumps at 10-20 /20-40Hz Hz for the δ / α mode would be discarded as background noise, because the weights for the bumps observed in these modes are lower than the peak values in the respective mode. The analysis and selection of the frequency modes follows the analysis pipeline described and validated by Vidaurre et al., (2018).

13. Patient sample appears to under-represent female patients. Is there a relationship between sex and DA metabolism/uptake/MEG markers? Perhaps simply acknowledge the imbalance in the manuscript.

We now acknowledge the sex imbalance in the discussion, caused by more male patients having been admitted for DBS surgery. To the best of our knowledge, there is no previous literature reporting a sex difference in MEG markers. Previous studies have found the density of dopamine transporters to be higher in females than in males, which is in line with findings from animal studies (Lavalaye et al., 2000). On the other hand, previous work has not found any significant differences between men and women in the type and dose of dopaminergic medication used in PD (Umeh et al., 2014).

We have added the following paragraph to the limitations of study on page 13:

“In addition, we only recorded 4 female patients because during the study period fewer female patients underwent DBS surgery at our centre. To the best of our knowledge, there is no previous literature reporting a sex difference in MEG markers or the prescribed dopaminergic medication (Umeh et al., 2014).”

14. LFP recordings were sampled from externalized DBS leads using the St. Jude directional 6172 lead except in one patient who had the Boston Scientific leads (which I think is fine to include in analysis together provided surface areas of contacts are identical). The geometry of the segmented leads needs to be addressed relative to sensitivity for various frequencies of oscillatory activity. Segmented leads have very low surface area and may require ultra-high sampling rates (30 kHz and higher) to resolve oscillatory activity.

The surface area of the directional contacts is almost the same for the two systems: In both cases, the contact is 1.5mm long, but for the St. Jude electrode the diameter is 1.29mm, while for the Boston scientific lead it is 1.3mm. Our data was acquired at 2.4 KHz. Other studies using directional electrodes for LFP recordings employed 1Khz (Telkes et al., 2020) and 24 KHz down-sampled to 375 Hz for analysis (Nguyen et al., 2020) and were able to resolve oscillatory activity in the STN. Thus we are confident the sampling rate is sufficient. We now discuss this issue in limitations of the manuscript on page 14.

“Moreover, we used directional contacts, which have a smaller surface area than the classical ring contacts. Based on the available literature, our sampling rate was high enough to resolve oscillatory activity in the STN (Telkes et al., 2020; Nguyen et al., 2020).”

Reviewer #2 (Recommendations for the authors):

General remarks:

1. A visualization of the state time series would be a nice addition to see the dynamics of the network. Including the probabilities would further give an impression of "how clear the states are", i.e. were they exclusive or were there some intervals where 2 states were rather active at the same time.

We now provide a snapshot (out of the 6,425,750 time points) of the state probability time course as figure 5 in the manuscript. In addition, we provide the state time series for anyone to check and visualise.

2. The color coding in table 1 (4x orange) does not correspond to the ring figures (2x orange/hemisphere). Both areas 17 and 18 are termed "medial orbitofrontal".

Thanks – we have corrected the table.

Results

As stated in the public review, there are several points concerning the presentation of the results:

3. Line 180 Supporting the dopamine overdose hypothesis in PD, we identified a δ/theta oscillatory network involving lateral and medial orbitofrontal cortical regions.

– Figure 2 also shows the inter-hemispheric connectivity of the pars orbitalis. What is the reason to not mention this in the main text?

We thank the reviewer for pointing this out and we now also mention the missing inter-hemispheric connection in the text.

4. Line 259: Furthermore, ON medication in the α band only the connectivity between temporal and parietal cortical regions and the STN was preserved (Figure 3C α; p < 0.05), consistent with previous findings (Litvak et al. 2011). In contrast, in the β band only STN-medial orbitofrontal connectivity remained intact (p < 0.05, Figure 3C β).

– Similar to the comment on line 180, the figure also shows α connectivity between the STN and medial orbitofrontal cortex.

We have now also added this connection to the text.

5. Line 269: Finally, ON medication, a sensorimotor-frontoparietal network emerged (p < 0.05, Figure 3C β) where sensorimotor, frontal, and parietal regions were no longer connected to the STN, but instead directly communicated with each other in the β band.

– Similar to the comment on line 180: is there a reason to exclude the connection from the somatomotor cortex to the caudal middle frontal (e.g. figure 3c, β: 13r-16r)?

We added this information to the text.

6. 256: Importantly, coherence OFF medication was significantly larger than ON medication between STN and sensorimotor, STN and temporal and STN and frontal cortices (p < 0.05 for all connections, Figure 3B α and β).

– The figure does not these results for STN and sensorimotor cortex

We apologize for the confusion. Panel B of Figures 2-4 is thresholded for a p-value of 0.01, whereas the p-value for the STN-sensorimotor comparison was 0.01398 (β) and 0.02826 (α). Hence, these results were not visible in Figure 3B. We have changed panel B of figures 2-4 to a consistent threshold of 0.05.

7. 263: Most previous PD studies report a decrease in the motor-STN coherence ON medication in the β band (Hammond, Bergman, and Brown 2007; Litvak et al. 2011; Hirschmann et al. 2013; Little et al. 2013; Marinelli et al. 2017) but do not indicate any changes that the sensorimotor regions might experience at the whole-brain level. In the communication state, OFF medication, STN-pre-motor (sensory), STN-frontal, and STN-parietal connectivity was present (p < 0.05, Figure 3A α and β). STN-cortical coherence was then significantly reduced ON compared to OFF medication (p < 0.05, Figure 3B α and β).

– The first sentence introduces cortico-cortical connections, which is then, however, followed by STN-cortical results. This is then followed by results on the sensorimotor-frontoparietal network, which indeed is cortico-cortical. However, later in the manuscript, the following statement suggests that the cortico-cortical network differs from the sensorimotor-frontoparietal network: "Still, significant connectivity was selectively preserved in a spectrally-specific manner ON medication both at the corticocortical level and the cortical-STN level. Furthermore, a sensorimotor-frontoparietal network emerged ON medication" (line282). It is not clear what exactly resembles the cortico-cortical network, and where the results are presented.

We like to thank the reviewer for pointing this out. We have added that “cortico-cortical network” refers to the sensorimotor-frontoparietal network. The results are presented on page 6:

“Finally, ON medication, a sensorimotor–frontoparietal network emerged (p < 0.05, Figure 3C β), where sensorimotor, medial prefrontal, frontal, and parietal regions were no longer connected to the STN, but instead directly communicated with each other in the β band. Hence, there was a transition from STN-mediated sensorimotor connectivity to the cortex OFF medication to a more direct cortico–cortical connectivity ON medication.”

On the results of temporal properties:

8. Line 324: Previous research has shown that ON medication, spectrally-specific cortico-STN connectivity remains preserved in PD compared to OFF medication (Litvak et al. 2011; Hirschmann et al. 2013). This indicates the existence of functionally relevant cortico-STN loops. A decrease in coherence between the cortex and the STN has also been observed ON medication (George et al. 2013), which was correlated with improved motor functions in PD. All the connectivity effects were observed in our results for the communication state. Furthermore, in the communication state, we showed the existence of a frontoparietal sensorimotor 331 network in the β band ON medication. Recent evidence indicates that with the loss of dopamine and the start of PD symptoms, δ/theta oscillations emerge within the basal ganglia (Whalen et al. 2020). In line with these findings, STN-STN δ/theta oscillations in the local state were reduced ON medication.

– I have the impression that this paragraph is not in the right position. It repeats and even discusses the previously mentioned results and jumps from "frontoparietal sensorimotor networks" to "STN-STN δ/theta" findings.

Further, it is not clear if "All the connectivity effects were observed in our results for the communication state" relates to the citations above, especially since other states are shown to have effects on connectivity as well.

We have removed the mentioned paragraph and now only introduce the reason for investigating the temporal properties (page 8).

9. Additionally, the main text should provide t-scores and degrees of freedom, not only p-values. The comprehension of the results could be improved by displaying significantly different comparisons in a clear way in Figure 5 (e.g. provide * for p<0.05). However, what was the rationale for testing the medication condition and states with different t-tests? Would a 2way- repeated measures ANOVA not be more appropriate?

We thank the reviewer for pointing this out. Indeed, a two-way repeated measures ANOVA is more appropriate. We now report the results of the ANOVA and of the post-hoc tests, including the relevant degrees of freedom. Manuscript figure 6 and text (page 8) was updated accordingly.

“Two-way repeated measures ANOVA on the temporal properties of the HMM states revealed an effect of HMM states on the fractional occupancy (F (2,96) = 10.49, p< 0.01), interval of visits (F (2, 221513) = 9783.13, p < 0.01), and lifetime (F (2, 214818) = 50.36, p< 0.01). There was no effect of medication (L-DOPA) on fractional occupancy (F (1, 96) = 2.00, p = 0.16) and lifetime (F (1,214818) = 0.15, p = 0.7026). Medication had a significant effect on the interval of visits (F (1,221513) = 4202.96, p < 0.01). Finally, we found an interaction between the HMM states and medication on the interval of visits (F (2,221513) = 1949.98, p< 0.01) and lifetime (F (2,214818) = 172.25, p< 0.01). But there was no interaction between HMM states and medication on fractional occupancy (F (2,96) = 0.54, p = 0.5855). “

10. Line 353: The lifetimes for both the local and communication state were significantly increased by medication (ON > OFF: local, p < 0.01; comms, p {less than or equal to} 0.01).

and line 360: Both the local and communication state tended to last longer ON medication.

– This seems to be a repetition. Also, it is not clear what "tended to last longer" means? Is there a significant difference or a trend?

We have removed those repetitions. “Tended to last longer” was meant to indicate that the lifetime of the communication and local states was significantly increased on medication. This part now reads (page 9):

“In summary, the cortico-cortical state was visited least often compared to the other two states both OFF and ON medication. The cortico-STN and STN-STN states showing physiologically-relevant spectral connectivity on the other hand lasted significantly longer ON medication. “

Methods

11. Line 676: Since NNMF does not guarantee a unique solution, we performed multiple instances of the factorization.

– The authors should state the exact number of repetitions. How was this number decided?

Every iteration of the NNMF algorithm produces a new solution from scratch. We repeated the NNMF algorithm until the obtained frequency modes converged to an approximation of the classical frequency bands (consider this to be a Bayesian prior). This approach follows Vidaurre et al., 2018. In practice, this required 4 iterations of the NNMF algorithm. We added this information to the manuscript, on page 20.

Reviewer #3 (Recommendations for the authors):

1. The Authors could examine the effect of the dopaminergic medication (ON and OFF) on the recurrent oscillatory patterns of transient network connectivity within and between the cortex and the STN, as a function of the duration of the disease and/or the DRT history.

We would like to thank the reviewer for pointing this out. We regressed duration of disease (year of measurement – year of onset) on the temporal properties of the HMM states. We found no relationship between any of the temporal properties and disease duration. Similarly, we regressed levodopa equivalent dosage for each subject on the temporal properties and found no relationship. We now discuss this point in the manuscript (page 13):

“A further potential influencing factor might be the disease duration and the amount of dopamine patients are receiving. Both factors were not significantly related to the temporal properties of the states.”

2. Re the contamination of the STN signal by volume conducted signals from the cortex. Page 20 (lines 584-586) "To correct for volume conduction in the signal, symmetric orthogonalisation (Colclough et al. 2015) was applied to each subject's resulting cortical time series matrix".

Have the authors applied this correction to the STN LFPs? If the authors cannot rule out the possibility of a contamination of STN LFPs by volume conducted signals from the cortex and cannot guarantee that the signal recorded in the STN only reflected local STN activity they should at least interpret their results with caution and discuss this limitation in their discussion.

We apologize for not stating this more clearly. We did not apply symmetric orthogonalisation to the STN signals. We have changed the relevant sentence to (page 17):

“To correct for volume conduction in the signal, symmetric orthogonalisation (Colclough et al. 2015) was applied to each subject’s resulting MEG cortical time series matrix.”

With respect to volume conduction, we appreciate this concern and thank the reviewer for bringing it up. Marmor et al., (2017) investigated this on humans and is therefore most closely related to our research. They find that re-referenced STN recordings are not contaminated by cortical signals. Furthermore, the data in Lalla et al., (2017) is based on recordings in rats, making a direct transfer to human STN recordings problematic due to the different brain sizes. Since we re-referenced our LFP signals as recommended in the Marmor paper, we think that contamination due to cortical signals is relatively minor; see Litvak et al., (2011), Hirschmann et al. (2013), and Neumann et al., (2016) for additional references supporting this. That being said, we now discuss this potential issue in the paper on page 13-14.

“Lastly, we recorded LFPs from within the STN –an established recording procedure during the implantation of DBS electrodes in various neurological and psychiatric diseases. Although for Parkinson patients results on β and tremor activity within the STN have been reproduced by different groups (Reck et al., 2010, Litvak et al., 2011, Florin et al., 2013, Hirschmann et al., 2013, Neumann et al., 2016), it is still not fully clear whether these LFP signals are contaminated by volume-conducted cortical activity. However, while volume conduction seems to be a larger problem in rodents even after re-referencing the LFP signal (Lalla et al., 2017), the same was not found in humans (Marmor et al., 2017).”

3. The authors could make the paper more reader-friendly. In particular, could classical spectral and coherence analyses be performed to visualize the characteristics of the oscillatory activity and neural synchronization of the EMG and LFP signals in the recurrent oscillatory patterns of transient network connectivity both ON and OFF medication?

We now explain our analysis more intuitively within the manuscript (see page 4). To aid intuition on how to interpret the result in light of the methods used, one can compare the analysis pipeline to a windowing approach. In a more standard approach, windows of different time length can be defined for different epochs within the time series and for each window coherence and connectivity can be determined. The difference in our approach is that we used an unsupervised learning algorithm to select windows of varying length based on recurring patterns of whole brain network activity. Within those defined windows we then determine the oscillatory properties via coherence and power – which is the same as one would do in a classical analysis. We have added an explanation of the concept of “oscillatory activity” within our framework to the introduction (page 2 footnote):

“For the purpose of our paper, we refer to oscillatory activity or oscillations as recurrent, but transient frequency–specific patterns of network activity, even though the underlying patterns can be composed of either sustained rhythmic activity, neural bursting, or both (Quinn et al. 2019).”

Moreover, we provide a more intuitive explanation of the analysis within the first section of the results (page 4):

“Using an HMM, we identified recurrent patterns of transient network connectivity between the cortex and the STN, which we henceforth refer to as an ‘HMM state’. In comparison to classic sliding-window analysis, an HMM solution can be thought of as a data-driven estimation of time windows of variable length (within which a particular HMM state was active): once we know the time windows when a particular state is active, we compute coherence between different pairs of regions for each of these recurrent states.”

4. In this study, the authors could reconstruct the trajectories of the DBS leads within the STN (using for example the open source LeadDBS program?).

We selected the electrode contacts based on intraoperative microelectrode recordings (for details, see page 16). The first directional recording height after the entry into the STN was selected to obtain the three directional LFP recordings from the respective hemisphere. This practice has been proven to improve target location (Kochanski et al., 2019; Krauss et al., 2021). The common target area for DBS surgery is the dorsolateral STN. To confirm that the electrodes were actually located within this part of the STN, we now reconstructed the DBS location with Lead-DBS (Horn et al., 2019). All electrodes – except for one – were located within the dorsolateral STN (see figure 7 of the manuscript). To exclude that our results were driven by outlier, we reanalysed our data without this patient. No change in the overall connectivity pattern was observed (see Author response image 4 and Author response image 5).

https://doi.org/10.7554/eLife.66057.sa2

Article and author information

Author details

  1. Abhinav Sharma

    Institute of Clinical Neuroscience and Medical Psychology, Medical Faculty, Heinrich-Heine University Düsseldorf, Düsseldorf, Germany
    Contribution
    Conceptualization, Software, Formal analysis, Investigation, Visualization, Methodology, Writing - original draft
    For correspondence
    Abhinav.Sharma@med.uni-duesseldorf.de
    Competing interests
    No competing interests declared
    ORCID icon "This ORCID iD identifies the author of this article:" 0000-0002-9296-3386
  2. Diego Vidaurre

    1. Department of Psychiatry, University of Oxford, Oxford, United Kingdom
    2. Department of Clinical Health, Aarhus University, Aarhus, Denmark
    Contribution
    Software, Validation, Methodology, Writing - review and editing
    Competing interests
    No competing interests declared
    ORCID icon "This ORCID iD identifies the author of this article:" 0000-0002-9650-2229
  3. Jan Vesper

    Department of Neurosurgery, University Hospital Düsseldorf, Düsseldorf, Germany
    Contribution
    Resources, Investigation, Writing - review and editing
    Competing interests
    No competing interests declared
  4. Alfons Schnitzler

    1. Institute of Clinical Neuroscience and Medical Psychology, Medical Faculty, Heinrich-Heine University Düsseldorf, Düsseldorf, Germany
    2. Department of Neurology, Center for Movement Disorders and Neuromodulation, Medical Faculty, Heinrich-Heine University Düsseldorf, Düsseldorf, Germany
    Contribution
    Resources, Investigation, Writing - review and editing
    Competing interests
    has been serving as a consultant for Medtronic Inc, Boston Scientific, St. Jude Medical, Grünenthal, and has received lecture fees from Abbvie, Boston Scientific, St. Jude Medical, Medtronic Inc, UCB.
  5. Esther Florin

    Institute of Clinical Neuroscience and Medical Psychology, Medical Faculty, Heinrich-Heine University Düsseldorf, Düsseldorf, Germany
    Contribution
    Conceptualization, Resources, Supervision, Funding acquisition, Validation, Investigation, Methodology, Project administration, Writing - review and editing
    For correspondence
    esther.florin@hhu.de
    Competing interests
    No competing interests declared
    ORCID icon "This ORCID iD identifies the author of this article:" 0000-0001-8276-2508

Funding

Volkswagen Foundation (89387)

  • Esther Florin

The funders had no role in study design, data collection and interpretation, or the decision to submit the work for publication.

Acknowledgements

EF gratefully acknowledges support by the Volkswagen Foundation (Lichtenberg program 89387). Computational support and infrastructure was provided by the ‘Centre for Information and Media Technology’ (ZIM) at the University of Düsseldorf (Germany). We would like to thank Johannes Pfeifer for his valuable feedback on the manuscript.

Ethics

Human subjects: The study was approved by the local ethics committee (study number 5608R) and conducted in accordance with the Declaration of Helsinki. Informed consent and consent to publish the results was obtained.

Senior Editor

  1. Michael J Frank, Brown University, United States

Reviewing Editor

  1. Andreas Horn, Charité - Universitätsmedizin Berlin, Germany

Reviewers

  1. Kelly Bijanki
  2. Muthuraman Muthuraman, University Medical Center of the Johannes Gutenberg University Mainz, Germany

Publication history

  1. Received: December 23, 2020
  2. Accepted: May 12, 2021
  3. Version of Record published: June 4, 2021 (version 1)

Copyright

© 2021, Sharma et al.

This article is distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use and redistribution provided that the original author and source are credited.

Metrics

  • 863
    Page views
  • 103
    Downloads
  • 5
    Citations

Article citation count generated by polling the highest count across the following sources: Crossref, PubMed Central, Scopus.

Download links

A two-part list of links to download the article, or parts of the article, in various formats.

Downloads (link to download the article as PDF)

Open citations (links to open the citations from this article in various online reference manager services)

Cite this article (links to download the citations from this article in formats compatible with various reference manager tools)

  1. Abhinav Sharma
  2. Diego Vidaurre
  3. Jan Vesper
  4. Alfons Schnitzler
  5. Esther Florin
(2021)
Differential dopaminergic modulation of spontaneous cortico–subthalamic activity in Parkinson’s disease
eLife 10:e66057.
https://doi.org/10.7554/eLife.66057

Further reading

    1. Neuroscience
    Benjamin D Pedigo, Mike Powell ... Joshua T Vogelstein
    Research Article

    Comparing connectomes can help explain how neural connectivity is related to genetics, disease, development, learning, and behavior. However, making statistical inferences about the significance and nature of differences between two networks is an open problem, and such analysis has not been extensively applied to nanoscale connectomes. Here, we investigate this problem via a case study on the bilateral symmetry of a larval Drosophila brain connectome. We translate notions of'bilateral symmetry' to generative models of the network structure of the left and right hemispheres, allowing us to test and refine our understanding of symmetry. We find significant differences in connection probabilities both across the entire left and right networks and between specific cell types. By rescaling connection probabilities or removing certain edges based on weight, we also present adjusted definitions of bilateral symmetry exhibited by this connectome. This work shows how statistical inferences from networks can inform the study of connectomes, facilitating future comparisons of neural structures.

    1. Neuroscience
    Sinisa Pajevic, Dietmar Plenz ... R Douglas Fields
    Research Article

    Temporal synchrony of signals arriving from different neurons or brain regions is essential for proper neural processing. Nevertheless, it is not well understood how such synchrony is achieved and maintained in a complex network of time-delayed neural interactions. Myelin plasticity, accomplished by oligodendrocytes (OLs), has been suggested as an efficient mechanism for controlling timing in brain communications through adaptive changes of axonal conduction velocity and consequently conduction time delays, or latencies; however, local rules and feedback mechanisms that OLs use to achieve synchronization are not known. We propose a mathematical model of oligodendrocyte-mediated myelin plasticity (OMP) in which OLs play an active role in providing such feedback. This is achieved without using arrival times at the synapse or modulatory signaling from astrocytes; instead, it relies on the presence of global and transient OL responses to local action potentials in the axons they myelinate. While inspired by OL morphology, we provide the theoretical underpinnings that motivated the model and explore its performance for a wide range of its parameters. Our results indicate that when the characteristic time of OL’s transient intracellular responses to neural spikes is between 10 and 40 ms and the firing rates in individual axons are relatively low (⪅ 10 Hz), the OMP model efficiently synchronizes correlated and time-locked signals while latencies in axons carrying independent signals are unaffected. This suggests a novel form of selective synchronization in the CNS in which oligodendrocytes play an active role by modulating the conduction delays of correlated spike trains as they traverse to their targets.