On the Complexity of Resting State Spiking Activity in Monkey Motor Cortex

Abstract Resting state has been established as a classical paradigm of brain activity studies, mostly based on large-scale measurements such as functional magnetic resonance imaging or magneto- and electroencephalography. This term typically refers to a behavioral state characterized by the absence of any task or stimuli. The corresponding neuronal activity is often called idle or ongoing. Numerous modeling studies on spiking neural networks claim to mimic such idle states, but compare their results with task- or stimulus-driven experiments, or to results from experiments with anesthetized subjects. Both approaches might lead to misleading conclusions. To provide a proper basis for comparing physiological and simulated network dynamics, we characterize simultaneously recorded single neurons’ spiking activity in monkey motor cortex at rest and show the differences from spontaneous and task- or stimulus-induced movement conditions. We also distinguish between rest with open eyes and sleepy rest with eyes closed. The resting state with open eyes shows a significantly higher dimensionality, reduced firing rates, and less balance between population level excitation and inhibition than behavior-related states.


24
The resting state in behavioral studies is defined operationally as an experimental condition without 25 imposed stimuli or other behaviorally salient events (Raichle, 2009 , 2014). A major conclusion of these studies is that spontaneous brain activity in 29 human and monkey can be characterized as a sequence of re-occuring spatio-temporal patterns of 30 activation or deactivation resembling task-evoked activity, but present during rest (Vincent et al., 31 Manuscript submitted to eLife More specifically, we also ask if certain features of the neuronal firing during pure resting periods 91 allow for a differentiation from spontaneous and task-induced movements, preparatory periods, 92 or sleepiness. Contrary to this expectation, the motor system may show invariants, i.e., statistical 93 properties of the neuronal spiking that do not change with respect to different behavioral epochs. 94 While such comparisons have been performed on the level of local field potential (LFP) recordings, 95 e.g., the investigation of behavior-related frequency modulations (Engel and Fries, 2010; Kilavik 96 et al., 2013), to our knowledge this is the first study to perform such comparison on the level of 97 spiking activity. 98 In the following, we first detail how we performed the segmentation of REST recordings according 99 to behavior, and then explored the activity of single neurons in different behavioral states. To 100 investigate if there are comparable neuronal activity states in task-related data, we performed 101 similar analyses for the R2G data. Apart from the SU dynamics, we also focused on network 102 properties of the neuronal activities: We evaluated pairwise covariances, dimensionality of rate 103 activities, and excitatory-inhibitory balance in the different behavioral states of both REST and R2G. 104 The comparison to the R2G data enabled us to identify systematic network state changes which are 105 less pronounced in REST. 106 Results 107 We aim to determine in what regards spiking activity during rest is distinct from other behavioral 108 states like spontaneous movements, sleepiness, movement preparation, or task-induced grasping. 109 To do so, we first describe the behavioral segmentation of the REST data based on videos of the 110 monkeys during the experiment, resulting in a sequence of defined behavioral states in REST. The 111 segmentation of R2G was chosen as in previous studies on these data (Riehle et al., 2018). Then,112 we show that the behavioral segmentation is meaningful in terms of neuronal activity on two 113 different scales: On the mesoscopic scale, which incorporates the collective behavior of neurons, 114 we show that the LFP spectra differ across states. On the microscopic scale, we show that SU firing 115 is correlated to the monkeys' behavior, and examine the relation between behavior, spiking activity 116 dimensionality and excitatory-inhibitory balance. 117 Behavioral segmentation 118 Based on video recordings, each REST session (two per monkey) was segmented according to 119 the monkey's behavior, (cf. Materials and Methods: Behavioral Segmentation). Three states were 120 considered: resting state (RS)-no movements and eyes open; sleepy resting state (RSS)-no 121 movements and eyes (half-)closed; spontaneous movements (M)-movements of the whole body 122 and/or limbs (Fig. 1A). For R2G recordings, two behavioral states were defined with respect to trial 123 events. For these states, interval lengths of 500 ms were used: the first part of the preparatory 124 period (PP)-500 ms after the first cue, when the monkey waits immobile for the GO; and a task-125 related movement period (TM)-an interval containing movement onset and grasping (Fig. 1B). 126 The visual segmentation is substantiated by comparison of the LFP spectra in the above defined 127 states (Fig. 1C). The relationship between LFP and behavior has been shown in several studies, e.g.  (Pfurtscheller and Aranibar, 1979). 132 In our data, RS and PP show peaks in the range from ≈10 to ≈30 Hz (alpha/beta range), the 133 peak in PP occurs for a higher frequency than in RS. In both monkeys, M and TM contain more 134 power compared to other states in frequencies above ≈50 Hz (gamma), while beta power is reduced. 135 However, the spectrum during RSS differs between monkeys. In monkey E, RSS seems to be a 136 distinct physiological state: it shows strong slow oscillations, as to be expected (Gervasoni et  SR indicates the switch-release event-beginning of the hand movement. PP was defined as [CUE-OFF, CUE-OFF+500 ms], and TM as [SR, SR+500 ms] for monkey E, and [SR-150 ms, SR+350 ms] for monkey N (different for the two monkeys due to differences in performance speed). (C and D) Power spectral density of LFP in different behavioral states. Panels in C pertain to REST, panels in D to R2G, left for monkey E and right for monkey N, respectively. States are defined in A and B. The peak at 50 Hz in the R2G spectra is an artifact (line frequency) and was not considered.

149
A prerequisite for the following analyses is to formalize a relationship between neuronal spiking 150 activity and the behavioral states of a monkey. Therefore, we quantified the correlation between 151 SU firing and behavior. This is by no means to be taken as a decoding approach, but rather as a 152 substantiation for the approach taken above to differentiate between behavioral states in REST.
153 Figure 2A shows     primarily between REST and R2G recordings, thus between spontaneous and task-related behaviors.   Network firing properties 263 We now turn towards the analysis of coordinated firing as opposed to single unit dynamics. Co-  dimensionality). The higher the PR, the more eigenvectors (principle components) are needed to 287 capture most of the variance of the dynamics. We performed an analysis for the REST and also for 288 the R2G states (0.5 s slices were concatenated to 3 s slices). To make the PR of different experiments 289 and recordings comparable, we normalized to the total number of SU obtained in each session.
290 Figure 4B shows that the dimensionality varies over time (shown for monkey E during REST 291 experiment). It changes with relation to behavior consistently across monkeys (as shown in Fig. 4C). 292 This is true for all sessions (see also Tab. 6). The PR is highest during RS and PP and lowest during 293 TM. The RSS state in both monkeys is clearly distinct from RS, its PR being more similar to the one 294 obtained for M, as seen in the covariance distributions. The spread of the values is notably higher 295 in REST than in R2G states, especially in monkey N. We performed a quantitative analysis of how the balance between bs and ns SUs relates to the 337 behavioral states, on the time scale of 100 ms, for our REST and R2G data. 338 Firstly, we asked if there was a state-specific prevalence of ns or bs activity.  Spearman rank correlation between (bs, ns) and PR

423
(2018), we find a (slightly) lower spike count variability during task related movements (TM) than 424 during movement preparation (PP) and vice versa for the spike time irregularity 1 . 425 A new finding of our study is a pronounced difference in variability between REST and R2G 426 states, i.e., between spontaneous and task-related behavior. All REST states show a significantly 427 higher spike count variability and a higher firing irregularity than the R2G states. These differences 428 are probably due to the behavioral constraints present in the R2G but not in the REST experiments. 429 During R2G task, the monkey received visual input to control periods of waiting or arm movements, 430 resulting in well-defined behavioral states and partially constrained mental states with a more 431 regular and reliable firing. In contrast, during REST experiments, the monkey itself decided what to 432 do (e.g. movement preparation or onset), resulting in a less well-defined behavior and its timing.  Finally, reliable covariance estimation necessitates very long data slices (Cohen and Kohn, 2011). 505 To satisfy this requirement, in R2G data we had to concatenate slices from 6 consecutive trials 506 into 3 s slices for the analysis of covariance and participation ratio. Thus, a single PR value results 507 from averaging over six independent recording periods in contrast to the continuous REST data. 508 However, this approach can be justified by our observation of a low inter-trial variability obtained 509 for 0.5 s slices of the R2G data.  Another typical claim of network simulations is the assumption of a balanced state, see above. 531 The modeling literature discusses different types of balance (Deneve and Machens, 2016). Many 532 studies assume a cancellation of excitation and inhibition in the input to neurons based on a balance 533 between the strength and number of excitatory and inhibitory afferent connections (Poil et al., 2012). 534 Perfect balance in this context corresponds to a critical point, where network dynamics exhibits 535 avalanche-like behavior (Beggs and Plenz, 2003 (Renart et al., 2010). The latter cancellation is caused by excess inhibitory feedback (Tetzlaff et al.,   540   2012) and, in excitatory-inhibitory networks, is accompanied by correlations between excitatory 541 and inhibitory spiking (Renart et al., 2010). These correlations can be quantified on the level of 542 neuronal output. Therefore, we here study balance based on the correlation between population 543 activities.  (Murphy et al., 1985), may boost the modulation of firing 561 rates on the population level. This could lead to higher pairwise covariances and subsequently lower 562 dimensionalities than expected in artificial networks with a well controlled input structure. We find 563 that such a decrease in dimensionality, is, for example, particularly pronounced during task-induced 564 movements. This again points out the necessity to separate between rest and movements in order 565 to avoid potential unrealistic mismatch between input and output statistics of spiking models.  assuming that a superposition of many spatially embedded networks yields an enlarged spatial 644 extent than a single such network (cf. Fig. 1 in Appendix 2). Likewise, the high dimensionality 645 observed during RS agrees well with the hypothesis of a superposition of several sub-networks.
646 647 Yet another question concerns the definition of "rest" in general: how to define it in other cortical 648 areas than motor cortex, e.g., in sensory systems? For the auditory system one would intuitively 649 assume that silence or white noise as auditory input represents the resting condition. Similarly, 650 for the visual system one could use a uniform or noise background as visual input. The choice of 651 "eyes-closed" as rest condition would, however, represent a different behavioral state compared to 652 our assumption of sleepy rest being a qualitatively different condition. 653 Given all the issues concerning the definition of "rest" and the behavioral segmentation, together 654 with the superposition of RSNs on the scale of brain areas, one could claim that it is futile to attempt 655 to characterize the spiking activity during an assumed resting state. However, our results clearly 656 demonstrate a set of significant differences between the spiking activity in motor cortex during 657 "rest" as compared to other behavioral conditions. 658

659
We demonstrate that spiking activity in monkey motor cortex during rest differs significantly from 660 other spontaneous and task-related behavioral states, for example sleepiness and movements. 661 The main characteristics of the resting state activity are low average firing rates combined with a 662 high variability of single-unit spiking statistics, and a pronounced complexity as indicated by a less 663 coordinated firing which results in a higher dimensionality of the network activity. We show that 664 and explain why neuronal network models should be validated against resting state data, aiming to 665 enhance the trend towards realistic network models that account for the heterogeneity in neuronal 666 data. We hope that our study is just the beginning of the characterization of "rest" on the level of 667 spiking neurons. More specific analysis is needed to quantify transient activities, their relation to 668 the balance between exitatory and inhibitory population activities, and to provide an automated 669 algorithm for the behavioral segmentation of REST recordings.

671
We first describe the two types of experimental recordings analyzed in this paper: resting state 672 (REST) and reach-to-grasp (R2G) data, the latter obtained during a behavioral task. Then, we explain During a resting state session, the monkey was seated (but not fixated) in a primate chair. The 685 chair was positioned so as to prohibit the animal from reaching any objects. There was neither a 686 particular stimulus nor any task, the monkey was free to look around and move spontaneously. 687 In addition to the registration of brain activity, the monkey's behavior was video recorded and  REST recording was performed subsequent to an R2G recording session. Only the E2 session was 701 recorded directly before an R2G session which is probably the reason for the missing RSS intervals. 702 The monkey was rather twitchy, impatiently waiting for the R2G tasks, because R2G experiments 703 include a reward while there was no reward during REST recordings.   The COV between spike trains and was defined as: of the neuronal dynamics: (2) The PR thus quantifies how many eigenvectors are necessary to explain a significant part of variance 852 in dynamics described by , see Fig. 8A. 853 The PR is low if most of the variability is captured by the first few eigenvectors. A large PR 854 indicates that many eigenvectors are necessary to capture the dynamics-a sign of high complexity. 855 In order to test the robustness of our results, we performed our analysis with different bin sizes. 856 The result is shown in (Fig. 8B). Here, all bin sizes revealed the same PR-dependent ordering of 857 behaviors. This suggests that our results are robust to the choice of bin size. The multiscale balance between putative excitatory (bs SUs) and inhibitory (ns SUs) population 868 firing was examined similarly to the procedure proposed in Dehghani et al. (2016). We considered 869 timescales from 1 ms to 10 s. For a given timescale, pooled spikes from bs and ns units were binned 870 and z-scored separately, resulting in bs and ns population activities (whole recording, no separation 871 into behavioral states). Then, the putative inhibitory population activity was subtracted from the 872 putative excitatory activity (see Fig. 5A). If this difference was close to zero, i.e., if pooled ns and bs 873 spike counts were nearly identical, the network activity was called balanced. If this was the case for 874 multiple time scales (i.e., bin sizes), it was called multiscale balance. 875 Since we observed some deviations from balance for bin sizes larger than 30 ms, we quantified 876 these deviations in a state-resolved manner. For each REST and R2G session of a given monkey, 877 we binned the 3 s time slices (concatenated from six consecutive trials of 0.5 s for R2G data) into 878 100 ms bins. Next, we applied two methods to quantify the balance between population activities. 879 Firstly, the same as for the multiscale balance, we z-scored the population activities, using 880 the respective mean and standard deviation of the whole recording (not state-specific). Then, we 881 calculated, separately for each state, the difference between the z-scored bs and ns population 882 activity of each 100 ms bin in each time slice: A negative value indicated a domination of ns activity 883 while a positive value meant that the bs activity was higher. Fig. 5C shows the corresponding 884 state-resolved histograms. 885 Secondly, we calculated the Spearman rank correlation between raw bs and ns population 886 activities for each time slice: The higher the correlation (bs, ns), the more strict the instantaneous 887 balancing between the ns and bs populations (cf. (Renart et al., 2010; Tetzlaff et al., 2012)). The 888 state-resolved results are presented in box plots (Fig. 6A). 889 To investigate the relationship between balance and dimensionality, we calculated the Spearman 890 rank correlation between (bs, ns) and PR ′ for each monkey, pooled over all REST and R2G sessions, 891 respectively (Fig. 6C, D and Tab. 6). 892 Biswal B, Zerrin YF, Haughton VM, Hyde JS. Functional connectivity in the motor cortex of resting human brain Transient activities 1100 To analyze the instantaneous balance, we correlate putative excitatory and inhibitory population activities in different behavioral states (sleepy rest RSS, rest RS, and movements M). We find a significantly reduced correlation (i.e., balance) during RS compared to M for monkey E, and during RS and M compared to RSS in monkey N, significantly only between M and RSS (cf. 6A and B). We also observe numerous transient increases in the population spike counts (Fig. 5B). Such simultaneous peaks contribute to higher correlation values between the two neuronal populations. The prevalence of this deviations differs between behavioral states. Fig. 1 and Tab. 1 below show that the distributions of population activities during M (monkey N, ns population) or both RSS and M (monkey E, both populations) are characterized by higher standard deviations than expected from higher mean values. RSS of monkey N shows lower means and slightly higher standard deviations than RS in ns population, pointing to the same conclusion. Both relations serve as footprints of an increased number of narrow peaks in population spiking during non-resting states. Given the transient peaks in the population spike counts during M and RSS, we suspect the following relationship between balance and transients in the population activity: Whenever one of the population activities transiently increases, the other one is forced to do the same due to the recurrent coupling between putative excitatory and inhibitory neurons, yielding higher correlation value and thus more balance. The participation ratio quantifies the dimensionality in the network activity space. One could ask how this relates to the distribution of neuronal activity in physical space. In analogy to large-scale resting state studies which find widely distributed networks of brain areas that are particularly active during rest (Biswal et al., 1995; Raichle, 2009; Deco et al., 2011), we estimated the spatial spread of active SUs in the different behavioral states of the REST recordings. To this end, we calculated the average spatial distance from each active SU to the center of mass of the spiking activity during sleepy rest (RSS), rest (RS), and movements (M). An active SU emitted at least one spike during the respective 3 s slice; the center of mass is given by the average coordinates of all active SUs. We thus characterized the mean spatial spread of the activity around the center of mass in each behavioral state. (2018). The differences in the spatial confinement of active SUs were small but consistent across sessions and monkeys. For monkey E, we find that the activity during RS exhibits a higher spatial spread than during M, even if the difference is only weakly significant ( < 0.01 in session E1 and < 0.05 in session E2). In monkey N, only the second session shows a significant difference, namely a larger spatial spread in RS compared to M ( < 0.01). In summary, we show a tendency of the SU activity during rest to be distributed over a larger spatial region than during movement, which may relate to higher dimensionality quantified by PR.