Modulation of visual gamma oscillation by excitatory drive and the excitation/inhibition balance in the visual cortex

Cortical gamma oscillations are generated in circuits that include excitatory (E) and inhibitory (I) neurons. Prominent MEG/EEG gamma oscillations in visual cortex can be induced by static or moving high-contrast edges stimuli. In a previous study in children, we observed that increasing velocity of visual motion substantially accelerated gamma oscillations, and led to the suppression of gamma response magnitude. These velocity-related modulations might reflect the balance between neural excitation induced by increasing excitatory drive, and efficacy of inhibition. Here, we searched for functional correlates of visual gamma modulations and assessed their development in 75 typically developing individuals aged 7-40 years. Gamma oscillations were measured with MEG in response to high-contrast annular gratings drifting at 1.2, 3.6, or 6.0°/s. In adults, we also recorded pupillary constriction as an indirect measure of excitatory drive. Pupil constriction increased with increasing velocity, thus suggesting increased excitatory drive to the cortex. Despite drastic developmental changes in gamma frequency and response strength, the magnitude of the velocity-related gamma modulations – a shift to higher frequency and amplitude suppression – remained remarkably stable. Based on the previous simulation studies, we hypothesized that gamma suppression might result from excessively strong excitatory drive caused by increasing motion velocity and reflects a tradeoff between overexcited excitatory and inhibitory circuitry. In children, the stronger gamma suppression correlated with higher IQ, suggesting importance of an optimal E/I balance for cognitive functioning. The velocity-related changes in gamma response may appear useful to assess E/I balance in the visual cortex. Highlights Speeding visual motion increases frequency and suppresses power of gamma. Gamma frequency decreases from 7 years on, faster in children than in adults. Magnitude of velocity-related modulation remains stable across 7-40 years. Gamma suppression correlates with gamma frequency and IQ in children. Gamma suppression may reflect efficiency of excitation/inhibition balance.

Despite the urgent need for biomarkers of the E/I balance in the human brain (Ecker, et al., 2013;Levin and Nelson, 2015), valid non-invasive measurements are still lacking.
Recently, the stimulus-induced high-frequency gamma oscillations in brain activity measured with EEG/MEG have attracted considerable attention as the putative noninvasive indicators of an altered E/I balance in cortical networks (Levin and Nelson, 2015;Nelson and Valakh, 2015). Gamma oscillations (30-100 Hz) are generated by populations of interconnected excitatory and inhibitory neurons in response to sufficiently strong excitatory drive (Buhl, et al., 1998;Munk, et al., 1996). These oscillations critically depend on fast-spiking (FS) parvalbumin-sensitive (PV+) interneurons that are capable of quickly synchronizing activity of the principle cells by inhibiting them through a negative feedback loop (Takada, et al., 2014). Because gamma synchronization depends on the excitatory state of both E and I circuitry, the strength of the gamma response to external input is intimately related to the E/I balance in neural networks. Increasing intensity of the excitatory input usually induces an increasingly strong gamma response, and the gamma oscillations are particularly powerful when excited E circuitry is not properly balanced by inhibition (i.e. in case of high E/I ratio) (see e.g. (Yizhar, et al., 2011).
In contrast to previously reported results, our recent MEG study in children revealed a strong suppressive effect of increasing stimulus velocity on the gamma response power that was paralleled by a prominent increased in gamma peak frequency . These unexpected findings imply that highly powerful visual input might cause gamma power suppression instead of enhancement.
Although unexpected, the velocity-related gamma suppression phenomenon resembles that which has been observed in some invasive studies in primates that used visual stimuli of high luminance contrast (Hadjipapas, et al., 2015;Jia, et al., 2013;Roberts, et al., 2013). These studies demonstrate a non-linear bell-shaped dependency of visual gamma power on luminance contrast.
The transition from gamma power enhancement to suppression as a function of input drive can be explained by a computational model of PING gamma oscillations suggested by Borgers and Kopell (Borgers and Kopell, 2005). The transition from low to modest levels of excitatory input to the Inhibitory-cells (I-cells) increases the number of recruited neurons, thereby facilitating gamma oscillations at the populational level, and eliciting an increase in gamma power. However, a too strong excitatory drive may disrupt synchronization between I cells and/or E-I synchronization and lead to a reduction in the power of gamma oscillations. A bell-shaped dependence of gamma power as a function of the strength of excitatory drive is thus expected: an initial increase up to some critical intensity is followed by suppression at yet higher intensities.
In the context of the Borgers and Kopell's model, our findings of gamma suppression suggest that stimuli moving with a sufficiently high speed may result in a failure of the excessively activated I-cells to synchronize network activity, and consequently correspond to a descending branch of the bell-shaped curve. Therefore, the gamma suppression driven by an increase in stimulus velocity seems to reflect a distinct neural phenomenon, which highlights the nonlinear relationships between external drive, E/I ratio, and gamma power modulation in the human brain. The model predicts that the stronger and more rapid suppression of gamma oscillation power at the descending branch of the bell-shaped curve occurs when excitation of the E-cells is effectively reduced by inhibitory synaptic currents, i.e. at a relatively lower E/I ratio. On the other hand, weaker and slowly developing gamma suppression is expected in case of higher E/I ratio. Thus, modulation of gamma oscillations by an increase in the stimulus velocity could pave the way to characterization of E/I balance in normal and disturbed visual networks and, if confirmed in systematic studies of typical individuals and individuals with neurodevelopmental conditions may have important clinical implications.
In this study we aimed to characterize the incidence, properties and individual variability of gamma suppression in a large population of neuro-typical children and adults through a broad age range (6 -40 years), thereby clarifying whether the gamma suppression is age-specific or a more general phenomenon. In order to modulate the excitatory drive to the visual cortex, we used the 'visual motion' paradigm that revealed the paradoxical suppression of gamma power in our previous studies in children Stroganova, et al., 2015). We expected that maturation of GABA-, and glutamate-mediated neurotransmission would change the frequency and power of gamma response and affect their modulations as a function of excitatory drive to the cortex.
To confirm that an increase in stimulus velocity elicits higher excitatory input to the visual cortex, we measured pupillary reaction in a subsample of our participants. A transient constriction of the pupil can be provoked by visual motion without increment in light flux, reflecting an increase in the neuronal activity within the visual cortical areas (Barbur, et al., 1998;Sahraie and Barbur, 1997;Wilhelm, et al., 2002). Therefore, if faster motion of the annular grating does intensify excitatory drive to the visual cortex, it is likely to be accompanied by greater constriction of the pupil.
An important prediction derived from Borgers and Kopell's model (Borgers and Kopell, 2005;Cannon, et al., 2014) is a link between a magnitude of gamma suppression and the efficacy of neural inhibition in the visual cortex. Here, we addressed this prediction in two ways. First, because relatively high excitability of inhibitory neurons should facilitate gamma suppression and accelerate gamma oscillations (Ferando and Mody, 2015;Mann and Mody, 2010), we expected that the magnitude of gamma suppression would correlate with individual variations in gamma frequency. Second, because more efficient inhibition in the visual cortex was shown to be related to better visual perceptual skills and higher IQ (Cook, et al., 2016;Melnick, et al., 2013), we expected that magnitude of suppression of gamma oscillations in response to increasing velocity of visual motion will correlate positively with IQ.

2.1.1.Children
Fifty typically developing boys between 6 to 15 years of age were recruited for the study from local schools in Moscow. The inclusion criterion was absence of neurological or psychiatric diagnoses. In all but one child (his parents refused the additional visit for IQ assessment) the IQ was assessed with the Kaufman Assessment Battery for Children KABC II (Kaufman and Kaufman, 2004).
Two control children were later excluded from the MEG data analysis. One of the two had excessive myogenic artifacts during MEG recording and another one failed to follow the instruction to respond with a button press. Of the 48 boys left, two were later excluded from the group analysis due to the reasons described in the Results section. This resulted in a final sample size of 46 children (age range 6.7 -15.5 years, mean 11.0 years, SD = 2.2).
The investigation was approved by the Ethical Committee of the Moscow University of Psychology and Education. All children provided their verbal assent to participate in the study and were informed about their right to withdraw from the study at any time during the testing. Written informed consent was also obtained from a parent/guardian of each child.

Adults
The initial adult sample included twenty seven adult subjects. Twenty two adult participants were recruited in Gothenburg. The inclusion criterion was an absence of reported neurological or psychiatric conditions. One male subject (20 years-old) was later excluded from the MEG group analysis because of a post-hoc discovery of an unreported epilepsy. His results are described separately. The age of the other 26 adult participants (3 females) ranged from 19.2 to 40.1 years (mean 27.9 years, SD=6.3). In 20 of 26 the participants (all males) the IQ has been assessed using Wechsler Adult Intelligence Scale, 4 th Edition (WAIS-IV) (Wechsler, 2010).
The study was approved by the Gothenburg Regional Ethical Review Board (Regionala etikprövningsnämnden i Göteborg). All participants gave written informed consent after the experimental procedures had been fully explained.

Experimental task
We applied the same experimental paradigm as that used in our previous studies in children Stroganova, et al., 2015). The subjects sat in a dim magnetically shielded room, with their head resting against the back of and positioned as closely as possible to the top of, the helmet-shaped surface of the helium dewar. For children, we used a PT-D7700E-K DLP projector that presented images with 1280 х 1024 screen resolution and 60 Hz refresh rate. For adults, the stimuli were presented using FL35 LED DPL projector with 1920 х 1080 screen resolution and 120 Hz refresh rate. The stimuli consisted of black and white annular gratings with a spatial frequency of 1.66 circles per degree of visual angle and outer diameter of 18 degrees of visual angle (Fig. 1A, insert). They were generated using Presentation software (Neurobehavioral Systems Inc., USA). The grating appeared in the center of the screen over black background and drifted to the central point at one of the three velocities: 1.2°/s; 3.6°/s; 6.0°/s, referred to below as 'slow', 'medium' and 'fast'. The frequencies of black to white transitions at a given position in the visual field produced by these velocities were 4, 12, and 20 Hz, respectively.
Each trial began with the presentation of a white fixation cross in the centrum of the display over a black background for 1200 ms, that was followed by a moving grating. After moving for a random interval varying between 1200 and 3000 ms, the stimulus stopped and remained on the screen for the length of the 'response period'. The response period was individually adjusted in children (see  for details) and was fixed at 1000 ms in adults. To keep participants alert, we instructed them to detect termination of the motion and to respond by pressing a button as soon as the motion stopped. If no response occurred within the response period the grating was substituted by a discouraging message "too late!" (in Russian/Swedish for participants in Moscow/Gothenburg) that remained on the screen for 2000 ms, after which a new trial begun. The participants were instructed to constantly maintain their gaze at the fixation cross in the intervals between the stimuli.
Individual omission and commission (responses that occurred during motion or earlier than 150 ms following termination of motion) errors were measured in all participants except two children for whom the behavioral responses were not recorded due to equipment error. The error trials were excluded from the MEG analysis.
The experiment included three blocks of trials; stimuli of different velocities were presented in all three blocks in a random order. The participants responded with either the right or left hand (50% trials each) in a sequence that was counterbalanced between blocks and participants. Each stimulus velocity was presented 30 times within each experimental block resulting in 90 trials per stimulus velocity. In order to minimize visual and mental fatigue, short (3-6 s) animated cartoon characters were presented between every 2-5 stimuli. The subjects' head position was continuously monitored and adjusted to a common position using MaxFilter™ (v2.2).

Pupillometry
Pupil size was measured in 19 of 27 adult participants using an EyeLink 1000 binocular tracker. We calculated the amplitude of the maximal pupil constriction (i.e., minimal pupil size) during the 0 to 1200 ms stimulus interval as a percentage of the average pupil size during the-100 to 0 ms pre-stimulus interval: % relative constriction = (A pre-mean -A post-min )/A pre-mean *100 These values were averaged separately for slow, medium and fast conditions. Trials contaminated by blinks in the window of interest (-100 to 1200 ms), those preceded by cartoons and error trials were excluded from analysis.

MEG recording
In children, neuro-magnetic brain activity was recorded at the Moscow MEG Centre (the Moscow State University of Psychology and Education) using a 306-channel detector array (Vectorview; Neuromag, Helsinki, Finland) positioned in a magnetically shielded room (AK3b; Vacuumschmelze, Hanau, Germany). In adults, MEG was recorded at the NatMEG Centre (The Swedish National Facility for Magnetoencephalography, KarolinskaInstitutet, Stockholm) using a similar 306-channel system (ElektaNeuromag TRIUX) located in 2-layer magnetically shielded room (Vacuumschmelze GmbH). Both systems comprise 102 identical triple sensor elements. Each sensor element consists of three superconducting quantum interference devices, two with orthogonal planar gradiometer pickup coils and one with a magnetometer pickup coil configuration. Four electrooculogram (EOG) electrodes were used to record horizontal and vertical eye movements. EOG electrodes were placed at the outer canti of the eyes and above and below the left eye. To monitor the heartbeats in children one ECG electrode was placed at the manubrium sterni and the other at the mid-axillary line (V6 ECG lead). In adults the ECG electrodes were placed at left and right sides of the chest under the collarbone. MEG, EOG, and ECG signals were recorded with a band-pass filter of 0.03-330 Hz in Moscow and 0.1-330 Hz in Stockholm, digitized at 1000 Hz, and stored for off-line analysis.

MEG data preprocessing
The data was first de-noised using the Temporal Signal-Space Separation (tSSS) method (Taulu and Hari, 2009) implemented in MaxFilter™ (v2.2) with parameters: '-st'=4 and 'corr'=0.90. For all three experimental blocks, the head origin position was adjusted to the initial head origin position in the block #2. For further pre-processing we used the MNEpython toolbox (Gramfort, et al., 2013) and custom Python and Matlab scripts.
The de-noised data was filtered between 1 and 145 Hz. To remove biological artifacts (blinks, heart beats, and in some cases myogenic activity), we then applied independent component analysis (ICA). The MEG periods with too high (4000e-13 fT/cm for gradiometers and 4e-12 fT for magnetometers in adults, 4000e-13 fT/cm for gradiometers and 8e-12 fT for magnetometers in children ) or too flat ( 0.1e-12 fT/cm for gradiometers and 1e-13 fT for magnetometers) amplitudes were excluded from the analysis. The number of independent components was set at dimensionality of the raw 'SSS'ed' data (usually around 70). We further used an automated MNE-python procedure to detect artificial EOG and ECG components, which we complemented with visual inspection of the ICA results. The number of rejected artifact components was normally 1 for vertical eye movements, 0-3 for cardiac and 0-6 for myogenic artifacts.
The artifact-corrected data was epoched from -1.0 to 1.2 s relative to the stimulus onset. The epochs containing strong muscle artifacts were excluded by thresholding the mean absolute value of the high-frequency (>70 Hz) signal. The threshold was set at 3 standard deviations from the power of the 70-Hz high-passed signal averaged across channels. The epochs left were visually inspected for the presence of undetected highamplitude bursts of myogenic activity and the epochs contaminated by such artifacts were manually marked and excluded from the analysis. The number of artifact-free epochs with correct button-press responses for the 'slow', 'medium' and 'fast' conditions was on average 70.5, 70.8 and 70.5 in children and 79.7, 79.3 and 78 in adults.

Time-frequency analysis at the sensor level
The time-frequency analysis at the sensor level has been performed in all participants. In this study only planar gradiometers were analyzed. For spectral estimation we used multitaper analysis with discrete prolate spheroidal sequences (DPSS). We estimated the spectra in overlapping windows of 400 ms (shifted by 50 ms) at frequencies from 2.5 to 120 Hz. The frequency step was 2.5 Hz, the number of cycles was set at frequency/2.5 and the bandwidth was 4, which resulted in spectral smoothing ±5 Hz around each center frequency. To decrease the contribution of phase-locked activity related to the appearance of the stimulus on the screen, as well as photic driving that could be induced by the temporal frequency of the stimulation and screen refresh rate (or their interaction), we subtracted averaged evoked responses from each single data epoch using 'subtract_evoked' MNE function. To assess individual gamma peak frequency we first defined the frequency bin with the maximal power of the gamma response. Because the spectra of the gamma responses were not always perfectly Gaussian, we adjusted the peak frequency by calculating a weighted average over whose frequency bins (subject/condition-specific band) where the post-stimulus/baseline ratio exceeded 2/3 of the absolute maximum. The weighted gamma peak power was calculated in dB, as 10*log10(post-stimulus/baseline), where poststimulus/baseline is an average ratio over these frequency bins.
To assess the reliability of stimulus-related increase in gamma power in each subject and for each velocity condition we compared the single epochs' average pre-and poststimulus power at the frequency bin of the maximal gamma response using Mann-Whitney U test.

Time-frequency analysis at the source level
The structural MRIs (1 mm 3 T1-weighted) were available for 23 of 28 adult participants. For these participants we performed source analysis using FieldTrip open source software (http://www.ru.nl/neuroimaging/fieldtrip/). We sought to check for the similarity between gamma response parameters measured at the source and the sensor levels, as well as to assess position of the most significant gamma increase ('maximally induced voxel') in each of the velocity conditions. Prior to analysis each subject's brain has been morphed to the MNI template brain using linear normalization and 0.6 mm grid. To perform source localization we then adapted the two-step approach. As the first step time-frequency analysis was performed on the artifact-free epochs (-0.9 to -0.1 pre-stimulus and 0.4 to 1.2 post-stimulus) using multitaper method. The analysis window was centered at the sensor-defined subject/condition specific frequency ±25 Hz. The DICS inverse-solution algorithm (Gross, et al., 2001) was used to derive the common source spatial filters. Subsequently, bootstrap resampling source statistics was performed (with 10000 Monte Carlo repetitions) to verify for each participant and condition presence of a significant (p<0.05) brain cluster of poststimulus increase of gamma power in the visual cortex (L/R cuneus, lingual, occipital superior, middle occipital, inferior occipital, or calcarine areas according to the AAL atlas (Tzourio-Mazoyer, et al., 2002)).
At the second step the signal was filtered in 30-120 Hz range, the 'virtual sensors' time series were extracted for each participant and velocity condition from each of 6855 brain voxels, and the time-frequency analysis (multitaper method, ±5 Hz smoothing, ~1Hz frequency resolution) has been performed. The voxel with the most significant poststimulus increase of 45-90 Hz power in visual cortex was found. The gamma response parameters were defined for the average of this and the 25 closest voxels. To assess the source-derived gamma response frequency and amplitude we calculated post-/pre-stimulus ratio for each frequency in the broad gamma range (30-120 Hz) and then assessed weighted subject/condition specific frequency and amplitude using the approach applied in the sensor space.

Statistical analysis
Statistical analysis was performed using STATISTICA-13 software (Dell Inc, 2015). To test for the effect of the velocity condition (hereafter Velocity) on MEG parameters and pupil contraction amplitude, we used the repeated measures analysis of variance (ANOVA). The Greenhouse-Geisser correction for violation of sphericity has been applied when appropriate. The original degrees of freedom and corrected p-values are reported together with epsilon (ε). Partial eta-squared (η 2 ) was calculated to estimate effect sizes. ANOVA was also used to assess the effect of age group. Spearman coefficients were used to calculate correlations between variables. Details about the statistical analysis are given in the Results section.

Figure 1.
Pupillary constriction response to the annular gratings drifting towards the center of the display at three different velocities ('Slow' -1.2 °/s, 'Medium' -3.6 °/s; and 'Fast' -6.0 °/s). Vertical bars denote 0.95 confidence intervals. The insert shows the stimulus display; the arrows indicate motion direction.

Incidence of gamma responses to visual motion in the sensor space in children and
adults.  Under the slow and medium velocity conditions, gamma response peaks were highly reliable (response vs. baseline: p<0.0001) in all the adults and in the majority of the children.
For the third velocity, the highly reliable peaks were detected in the majority of the adults and in about half of the children.

Comparison of the source-and sensor-derived gamma power and amplitude.
The source analysis was performed in 23 of 26 adults for whom the structural MRI data were available. The DICS beamformer analysis showed presence of a significant cluster of gamma increase in visual cortex in response to the 'slow' and 'medium' Velocity in all these participants. In response to the 'fast' Velocity, a significant cluster was present in 19 of 23 participants. In all three conditions, the most frequent location of the 'maximally induced voxel' was in the calcarine sulcus (in 17/23 subjects in the 'slow' condition, in 16/23 subjects in the 'medium' condition and in 11/18 subjects in the 'fast' condition, see also the Supplementary table 1 for the MNI coordinates of gamma response for each subject). Figure   3 illustrates source distribution of the gamma response obtained using DICS beamforming technique. No significant shift of the 'maximally induced voxel' in either x, y or z coordinate has been observed between the three velocity condition (rmANOVA: all p's>0.4). The mean intra-individual distance between 'maximally induced voxels' was 5.4 mm (sd=6.2) for the 'slow' vs 'medium' conditions, 10 mm (sd=7.6) for the 'medium' vs 'fast' conditions, and 9.   * Only subjects with reliable gamma increase at both sensor and source levels were included into the frequency correlation analysis.
The high correlations between source-and sensor-derived gamma response parameters (Tab. 2) suggest close correspondence between sensor-and source-derived results. Moreover, source-space analysis showed that in all velocity conditions the major source of gamma oscillations was situated in the overlapping regions of the primary visual cortex. Considering that the structural MRI data were not available for all of the participants we further restricted analysis to the sensor space data.

Power
Here and in all further statistical analyses that examined the effect of velocity on the gamma response power at the group level, we included only those participants who demonstrated the reliable (p<0.0001) gamma peaks under the 'slow' condition. This strict criterion was applied because the magnitude of the velocity-related power suppression can be reliably estimated only when a reliable gamma response under the slow velocity condition is present. This criterion did not affect the size of the adult sample, but reduced the number of children in the analysis (adults: 26, children: 37). The children excluded from the gamma power modulation analysis (n=9; mean age=9.4 years; range: 6.7 to 12.6 years) were significantly younger (M-W test: z=2.33; p=0.02) than the rest of the children in the sample (n=37; mean age=11.4; range: 7.7 to 15.5 years).
The majority of participants, regardless of age, demonstrated gradual attenuation of gamma response with increasing stimulus velocity from 'slow' to 'medium' and further to 'fast' (Fig. 4A). Two examples of typical gamma spectra are shown in Figure 5A. In a few subjects, the transition from 'slow' to 'medium' condition elicited an increase (or no change) in gamma power, succeeded by its decrease under the 'fast' condition (Fig. 5B). One adult and two children showed virtually no change in gamma response power, even at the highest velocity (Fig. 5C,   In adults the repeated measures ANOVA also showed highly reliable reduction of gamma response power with changing velocity from 'slow' to 'medium' and further to 'fast' condition (F (2,50) =83.1, ε=0.83, p<10e -6 ; η 2 =0.77; Fig. 6A).

Frequency
A viable estimate of the gamma peak frequency is possible only when the peak is reliable.
We estimated the frequency of an individual gamma response peak only if the stimulusinduced power enhancement was significant at the p<0.0001 level. For the analysis of the velocity effect on gamma frequency ('frequency ANOVA'), we included those participants who had the reliable peaks under all three experimental conditions (21 children, 19 adults).

Power
The magnitude of the stimulus-induced gamma response significantly increased with age in children between 6 and 15 years, but did not change in adults between 19 and 40 years (Tab. 3). In figure 7A we present the age-related changes in power of gamma response for the 'slow' condition.  To analyze possible differences in velocity-related changes of gamma response magnitude in children and adults, we performed an ANOVA with the main factor Age-Group (adults, children) and repeated measures factor Velocity (3 conditions). This revealed a highly significant effect of Velocity (F (2, 122) =172.8, ε=0.87, p<1e -6 ; η 2 =0.74), but no effect for the Age-Group or the Age-Group*Velocity interaction (p's>0.6). In both the child and the adult samples, the gamma peak frequency decreased with age (Tab 3). In figure 7B we present the age dependent changes in the peak frequency of gamma response induced by 'slowly' drifting gratings. Notably, the drop in gamma frequency was more rapid during childhood (1.712 Hz per year) than during adulthood (0.644 Hz per year).

Gamma power suppression and gamma frequency
Based on the animal data and computational model reviewed in the introduction, we anticipated that subjects with relatively higher gamma frequency and/or greater velocityrelated increase of gamma frequency (i.e. those who might have greater excitation of I neurons or greater velocity-related increase in their excitation) would demonstrate 1) greater overall magnitude of gamma power suppression with transition from 'slow' to 'fast' velocity and/or 2) more 'rapid' gamma power suppression that would occur already at a milder level of excitatory drive (with transition from 'slow' to 'medium' velocity).
To assess gamma power suppression, we introduced two parameters. The first one - and characterizes how fast with increasing velocity the gamma suppression occurs.
Considering that the denominator was positive in all the participants included in the analysis, the gSS was positive if the suppression took place already at the 'medium' velocity.
It was negative in the case of initial power increase in the 'medium' velocity condition with the following decrease in the 'fast' condition (e.g. Fig. 5B).
The POWslow, POWmedium and POWfast were calculates as an absolute difference between mean post-stimulus and baseline power in the respective condition and for the subject/condition-specific band (see Methods). The higher gSM value indicates stronger gamma suppression with increasing excitatory drive. The higher gSS value signifies that the relatively greater part of the gamma suppression occurs already with the moderate excitatory drive (i.e. at the 'medium' velocity).
When all subjects with the reliable gamma response in the 'slow' condition (p<0.0001) were included into analysis, in children gSS correlated positively with gamma frequency in the 'fast' (R=0.6, p<0.01) and 'medium' (R-0.36, p=0.05) conditions (Tab. 5.) These correlations could not be accounted for by age, because gSS did not significantly change with age (R(29)=-0.3, p>0.1). This means that children who had higher frequency of gamma oscillations at high and medium velocity, suppressed gamma response more rapidly, i.e. already at the transition from the 'slow' to the 'medium' velocity. Unlike that in children, no correlations with gamma response frequency were found in adults (Tab.5) .
We further tested whether the velocity-related changes in gamma response amplitude correlated with changes in gamma frequency. In adults, the slow-to-medium # Note that the number of subjects included in each analysis depends on the presence of a reliable gamma peak in the corresponding velocity condition.

Gamma power suppression and IQ
In

Discussion
The current study examined the paradoxical suppression of visual gamma oscillations caused by increasing velocity of drifting annular gratings. In a large sample of children, we replicated our previous findings  indicating hypothesized that the suppression of gamma power with increasingly strong excitatory drive reflects the capacity of inhibitory neurons to counteract excitation in the visual networks (i.e., the E/I ratio). Since the E/I ratio depends, amongst other factors, on the tonic excitability of inhibitory neurons, which in turn strongly affects the frequency of gamma oscillations, we expected that the suppression of gamma power would correlate with gamma frequency parameters. Following the prediction from the model of (Borgers and Kopell, 2005), we anticipated that the greater suppression of gamma power would reflect lower E/I ratio that might be beneficial for an individual subject's level of cognitive functioning (Cook, et al., 2016;Melnick, et al., 2013). These predictions were confirmed in children, but only partially in adults, suggesting that E/I balance may be based on different mechanisms during childhood and adulthood.

Gamma parameters in the source and sensor space.
Comparison of the source and sensor analysis results shows close correspondence of the gamma parameters measured in the source and sensor space. This result agrees well with the findings of Tan et al (Tan, et al., 2016) who reported high reliability of both source-and sensor-derived parameters of visual gamma response and their close correspondence.
Importantly, in all three velocity conditions, the individual maxima of gamma responses were most frequently located in the primary visual cortex (calcarine gyri, see Supplementary Materials) and we did not find any evidence for a systematic shift of the maximum with changes in motion velocity. These findings suggest that generation of gamma activity in all three velocity conditions was based on activity of the same neural circuits.

Developmental changes in strength and frequency of visual gamma response
Visual stimulation with moving annular gratings resulted in prominent increase of induced gamma power under conditions of 'slow', 'medium' and 'fast' stimulus velocities in both child and adult samples (Fig. 2). In children, the strength of the gamma response increased between 6 and 15 years of age, while in adults it remained stable between 19 and 40 years of age (Tab.2, Fig. 7A). This developmental growth in the power of induced gamma oscillations was accompanied by a drop in their frequency that continues during adulthood but at a much slower pace (Tab.4 , Fig. 7B).
A developmental increase in the strength of visual gamma response is compatible with previous reports of a maturational rise in auditory gamma oscillations evidenced by increase in amplitude and/or inter-trial coherency of 40-Hz auditory steady-state responses (ASSR) (Cho, et al., 2015;Edgar, et al., 2016). Cho et al (Cho, et al., 2015) observed that the amplitude of the ASSRs increased in children between 8 and 16 years of age, but then dropped between 15-16 and 20-22 years. Although we had very few subjects within the 15 -22 year age range, (which precludes statistical analysis of the latter effect), the presence of very high gamma response values in the older children in our sample and their absence in adults (Fig. 7A) suggests that the reduction of gamma amplitude between adolescence and early adulthood may be a common feature of both the auditory and visual cortices.
Several factors may underlie developmental changes in gamma response power. The main contributor might be the maturation of inhibitory GABAergic neurotransmission that in primates is protracted into early adulthood (Le Magueresse and Monyer, 2013). Changes in subunit composition of GABAR (GABA-Receptor) (Hashimoto, et al., 2009), as well as increasing 'on-demand' GABA release (Pinto, et al., 2010) produce a developmental shifts toward stronger, more precise inhibitory currents, thereby resulting in highly coherent activity of principal cells in the upper gamma frequency band (Doischer, et al., 2008).
Developmental slowing of the visual gamma oscillations observed in our study is in line with results of Gaetz and colleagues (Gaetz, et al., 2012) who reported a decrease in the frequency of gamma oscillations induced by static visual gratings in subjects aged 8-45 years.
However, the small number of children (12 subjects) and much larger number of adults (47 subjects) included in their sample did not allow Gaetz et al. to perform a separate analysis of developmental trajectories in the maturing and the aging brain. Our study shows that the rate-and most probably the mechanism -of the developmental decrease in gamma frequency differ in children and adults.
The frequency of the induced gamma oscillations predominantly depends on the excitatory state of FS PV+ inhibitory neurons with higher frequency related to their higher tonic excitability (Anver, et al., 2011;Ferando and Mody, 2015;Mann and Mody, 2010).
There is convincing evidence of a developmental shift towards lower NMDA-mediated tonic excitation of the FS PV+ cells, as well as of a developmental reduction of the number of FS PV+ cells exhibiting excitatory NMDA-mediated currents (Wang and Gao, 2009). On the other hand, tonic inhibitory regulation of the FS PV+ interneurons gradually increases during postnatal development (Le Magueresse and Monyer, 2013). These maturational transformations leading to subtler tonic excitability of inhibitory neurons may well explain the highly pronounced and rapid decrease in the frequency of visual gamma oscillations that we observed between childhood and adolescence. Of note, the expression of the NMDAR units (NR1, NR2a, NR2b) proceeds to decrease even in the mature brain -from young to middle-aged adulthood (Adams, et al., 2008). These changes in NMDA transmission might explain the more gradual decrease of visual gamma frequency between 19 and 40 years of age in our study.

Modulation of gamma oscillations by excitatory drive.
Increasing velocity of visual motion from 1.2 to 6.0°/s (which also shifts the temporal frequency of the stimulation from 4 to 20 Hz) was associated with significantly stronger relative pupillary constriction (Fig. 1). It has been previously found that pupil constriction can be induced by changing attributes of a visual stimulation (e.g. spatial structure, color or motion coherence) without any changes in light flux (Barbur, et al., 1998;Sahraie and Barbur, 1997;Wilhelm, et al., 2002). These pupillary reactions are thought to reflect the activation of cortical areas involved in processing of respective features of the visual stimuli (Barbur, et al., 1998). In our paradigm, the increase in pupil constriction from slow to fast motion velocity is likely to reflect increasing activation of the motion-processing cortical areas (including V1, V5 and beyond) that in turn causes transient weakening of the central sympathetic inhibition of the parasympathetic Edinger-Westphal nucleus of the midbrain (Wilhelm, et al., 2002). Thus, the pupillometry findings support our interpretation of the effect of the motion velocity on gamma response as being a direct consequence of the rising excitatory drive to the visual cortex. The latter is further supported by the observation that the gratings moving at fast velocity were often perceived by the participants as most intense and even unpleasant.
Our results suggest that the frequency and power of induced gamma oscillations strongly depend on the strength of the excitatory drive. The increasing excitatory drive with increasing motion velocity from 'slow' to 'fast' lead to acceleration of gamma oscillations in all participants with reliably detectable gamma peaks (Fig. 4B). The mean increase in frequency was almost the same in children (15.3 Hz) and adults (14.6 Hz).
Experimental studies in animals (Mann and Mody, 2010), as well as simulation studies (Moca, et al., 2014), suggest that excitation of inhibitory FS interneurons is the main factor affecting the frequency of gamma oscillations. The recent DCM modeling study by Shaw and colleagues (Shaw, et al., 2017) has shown that reduction of gamma frequency by GABA reuptake inhibitor tiagabine is partly mediated by its effect on the time constant of the inhibitory interneurons. Therefore, the increase in gamma frequency with speeding the visual motion might be at least partly mediated by increasing excitation of FS interneurons in response to increasing excitatory drive.
The gamma response power decreased with increasing motion velocity in all participants who had reliable gamma responses under the 'slow' condition, with the exception of three subjects (one adult and two children; Fig. 4A, dashed lines), all of whom had neurological conditions revealed post-hoc (epilepsy, Arnold-Chiari malformation, minor motor ticks and frequent nightmares). In the majority of the participants, regardless of their age, the gamma response decreased already in the transition from 'slow' to 'medium' velocity (see two examples in Fig. 5A), although in a few subjects the 'medium' velocity augmented the gamma response, which started to attenuate only under the 'fast' velocity condition (Fig. 5B).
The robustness of the gamma suppression with increasing velocity of visual motion as well as its overall similarity in children and adults suggest that this phenomenon is not age-specific and reflects a very basic mechanism underlying the generation of gamma oscillations in the visual cortex. Concurrently, as we mentioned in the introduction, the gamma suppression phenomenon is at odds with the results of other MEG studies that consistently reported increases in gamma response power with increasing excitatory drive to the visual cortex (Hadjipapas, et al., 2015;Hall, et al., 2005;Muthukumaraswamy and Singh, 2013;Peiker, et al., 2015;Perry, et al., 2015;Siegel, et al., 2007;Swettenham, et al., 2009).
The opposite behavior of gamma oscillations in our and other studies fits well the predictions from the Borgers and Kopell model (Borgers and Kopell, 2005) and may be explained by an inverted U-shaped relationship between strength of excitatory drive and power of gamma oscillations (Fig. 9). The reliable enhancement of gamma response observed in MEG studies that manipulated excitatory drive with luminance contrast (Hadjipapas, et al., 2015;Hall, et al., 2005;Perry, et al., 2015), eccentricity (van Pelt and Fries, 2013), motion coherence (Peiker, et al., 2015;Siegel, et al., 2007), or compared stationary and slowly moving stimuli (Muthukumaraswamy and Singh, 2013;Swettenham, et al., 2009) might be explained by a relatively weak excitation produced by such manipulations at the ascending portion of the bell-shaped curve (Fig. 9A). The previous reports on MEG gamma enhancement with increasing luminance contrast, motion coherence, decreasing eccentricity, or transition from stationary to slowly drifting gratings are compatible with the effects of modest stimulus intensity at the upward branch of the bell-shaped curve. In our studies suppression of the gamma response with increasing stimulus velocity has been induced by strong excitatory drive whose intensity was close to (slow velocity) or well above (fast velocity) the suppression transition point. B. The schematic representation of the effects of optimal (blue) and elevated (orange, red) E/I ratios on the gamma power modulation. The increased E/I ratio shifts the transition point towards higher excitatory drive and results in reduced or absent gamma power suppression.
At a certain high level of the external excitatory drive (the 'suppression transition point'), when the FS cells are sufficiently excited, they lose their capacity to support gamma oscillations, since they either walk out of phase with pyramidal cells or desynchronize (Borgers and Kopell, 2005;Borgers and Walker, 2013;Cannon, et al., 2014). It is conceivable that in the majority of our subjects (with the exception of a few cases pictured in Fig. 5B and 5C), the excitatory drive produced by the 'slow' motion was close to or above the 'suppression transition point' and that the further increase of velocity up to the medium and high values put the visual networks in the high excitatory state corresponding to the descending portion of the bell-shaped curve, thus resulting in relative suppression of gamma response power.
Remarkably, the model by Borgers and Kopell explains while the PING gamma oscillations are not supported beyond the 30-90 Hz frequency range (Kopell, et al., 2010).
Therefore, the gamma amplitude reduction in the 'fast' velocity condition is well in line with

Gamma suppression and its relation to the E/I balance in visual networks
The Borgers and Kopell's model states that the gamma suppression with increasingly strong excitatory drive occurs more easily when excitatory postsynaptic currents in pyramidal cells are well compensated by inhibitory ones (Borgers and Kopell, 2005). Therefore, one may expect that at the same level of a sufficiently high external excitatory drive, the gamma power suppression will be relatively stronger when there is stronger excitation of inhibitory neurons and/or weaker excitation of excitatory pyramidal cells (Fig. 9B). Since a greater tonic excitation of the inhibitory neurons also accelerates gamma oscillations (Mann and Mody, 2010), one can expect that individual variability in gamma suppression parameters and gamma frequency are partially controlled by the same neural mechanism. Therefore, individuals with higher gamma peak frequency should be likely to display greater velocityrelated gamma power suppression, which also has a steeper slope, i.e. gamma power suppression peaks at a relatively lower excitatory drive. By the same token, both the magnitude and the slope of gamma suppression should be reduced in case of an elevated E/I ratio driven by either excessive excitation of excitatory circuitry or by relatively weak inhibition. The predicted link between gamma suppression parameters, gamma frequency and E/I balance was supported by our data in several ways.
First, we found the predicted positive correlations between gamma suppression parameters and peak gamma frequency (in children) or changes in frequency (in adults) ( Table 4). The dissimilar pattern of the correlations may reflect low power of the analysis, but may also reflect different mechanisms of the E/I balance in the developing and mature brain.
Secondly, consistent with results of the previous behavioral (Melnick, et al., 2013) and proton magnetic resonance spectroscopy (H1-MRS) (Cook, et al., 2016) studies linking the strength of neural inhibition in the visual cortex with cognitive functions, we found that the stronger and faster gamma response suppression correlated with higher IQ in children (Tab. 6, Fig. 8). This correlation suggests that lower E/I ratio in the visual cortex in associated with better cognitive functioning in neurotypical children. It is however unclear if the E/I balance in the visual cortex is important by itself or as a reflection of the widespread E/I level in the brain. Absence of significant correlations between general IQ and gamma suppression in adults suggests that an optimal E/I balance in the visual cortex might be less important for the general cognitive functioning of the mature brain.
Thirdly, the deviations from the typical gamma suppression values seemed to be associated with abnormally elevated E/I ratio in the visual cortex. Three out of 66 (39 children and 27 adults) participants with highly reliable gamma peaks in the 'slow' velocity condition failed to suppress gamma oscillations even at the highest velocity/excitatory drive (see Fig. 4C for the spectra of two of these subjects), and their post-hoc examination revealed neurological conditions that might be associated with cortical hyper-excitability (epilepsy, Arnold-Chiari malformation and a case with minor motor tics and frequent nightmares). Although more systematic investigation is warranted, these outliers' data are generally in line with the prediction that the low or absent gamma suppression at the descending branch of the bell-shaped curve is indicative of global imbalance in excitation within the visual cortical networks.
Of note, the unbalanced excitation in visual cortex of subjects with neuropsychiatric disorders might be reflected not only in an absent/reduced gamma power suppression at the descending branch of the bell-shaped curve, but also as an abnormal gamma power gain at its ascending branch. Indeed, Peiker et al (Peiker, et al., 2015) have found abnormally strong modulation of gamma power by gradually increasing coherence of visual motion in people with autism spectrum disorders. Investigation of gamma response modulation as a function of excitatory drive at both the ascending and descending branches of the bell-curve would add important knowledge about E/I balance in normal and atypical brain.

Gamma modulations and age
Highly reliable suppression of gamma power and increase of gamma frequency with increasing velocity of visual motion has been observed in both children and adults (Fig. 6).
The developmental stability of the gamma modulations suggests that the global E/I balance under increasing excitatory drive is efficiently regulated in the healthy brain throughout the broad age range of subjects included in the study. Moreover, given the substantially higher gamma frequency in children as compared to adults (Tab. 1, Fig. 7B), the adjustment of the global E/I balance in the immature brain seems to be achieved at a higher level of excitability of both E and I neurons. The developmental stability of the gamma response modulations is in line with an in vitro observation that the number of inhibitory synapses on the neuronal branches scale together with the number of excitatory synapses during development, resulting in a relatively stable E/I ratio across several developmental stages (Liu, 2004).
However, a few words of caution are warranted regarding the lack of a reliable developmental trend in gamma modulation parameters. According to animal studies, the maturation of neural processes underlying the E/I balance in auditory and visual cortices takes place during early sensitive periods of brain development (Froemke, 2015). All of our participants were older than 6 years of age, and it is not unlikely that infants and toddlers would present gamma modulation parameters that are still age-dependent due to a continuing developmental fine-turning of the E/I balance in the visual cortex. This could also be true for the youngest children from our sample. These children were more frequently excluded from the gamma suppression analysis due to a low signal to noise ratio of the gamma response. While only including cases with highly reliable gamma responses in terms of signal-to-noise ratio allowed us to reduce the impact of noise on the gamma suppression parameters, it also lowered the proportion of youngest participants in our child sample.

Conclusions
Although a link between gamma oscillations and the E/I balance in neural networks is hardly disputable, the hope that the frequency or power of human gamma response may serve as a single non-invasive indicator of the E/I balance is perhaps overly simplistic. Here, we investigated the phenomenon of gamma power suppression caused by gradual increase in stimulus velocity in children and adults and provided evidence for its potential role as a putative biomarker of the E/I balance in the visual cortex. Our study resulted in two main novel findings.
First, we found a substantial developmental slowing of visual gamma oscillations from already 7 years of age, and that therefore cannot be explained by 'aging' (Gaetz, et al., 2012). This finding challenges the prevailing view, which predicts that maturational changes in neurotransmission should improve precision with which populational neural activity is synchronized and, as a result, increase both gamma synchronization and gamma frequency (Uhlhaas, et al., 2010). Given the animal data (Mann and Mody, 2010;Wang and Gao, 2009), the likely mechanism for developmental slowing of gamma oscillations is the developmental decrease in tonic excitability of the PV+ FS inhibitory neurons in the visual cortex.
Functionally, given the role of inhibitory transmission in neural plasticity, faster juvenile visual gamma oscillations may be beneficial for perceptual learning, which proceeds with higher intensity in younger as compared to older brains.
The second novel finding is the developmental stability of the suppressive effect of an increasing velocity of visual motion on the strength of gamma response coupled with an acceleration of gamma oscillations. Although the suppressive effect of excessively strong excitatory drive on MEG gamma power is well in line with the theoretical predictions (Borgers and Kopell, 2005;Cannon, et al., 2014), to our knowledge it has not been demonstrated before, with the exception of our earlier studies in children Stroganova, et al., 2015). Because gamma oscillations facilitate information transmission (Vinck, et al., 2013), their effective suppression with a 'too strong' excitatory drive may have a protective function via reducing the number of cortical assembles engaged in information flow and preventing cortical hyper-excitation in response to sensory stimuli.
Given that the slope of gamma suppression correlated with the frequency of the gamma response in the developing brain, we suggest that these parameters are sensitive to partly inter-related neural processes. However, while the frequency of gamma oscillations predominantly depends on the tonic excitability of inhibitory neurons (Mann and Mody, 2010), efficient gamma suppression may reflect a global balance between mutually counteracting excitatory and inhibitory circuits. The link between the magnitude of gamma suppression and IQ values in children points to the importance of an optimal E/I balance for cognitive abilities. A remarkable developmental stability of gamma suppression in healthy brain and its atypical absence in a small proportion of our participants with neurological abnormalities allow us to speculate that a gamma suppression deficit may characterize a disturbed E/I balance in the hyper-excitable visual cortex in individuals over a wide age range. Due to the lack of direct neurophysiological evidence in the context of the same 'visual motion' experimental paradigm, this interpretation should be treated with caution. However, it allows us to make several testable predictions for future research regarding functional consequences of individual variability in gamma suppression.
We expect that a gamma suppression deficit at the high levels of excitatory drive will characterize: patients with seizure onset zones located in the occipital cortex, neuro-typical individuals with visual hypersensitivity, and patients from clinical groups characterized by sensory modulation problems. Furthermore, a deficit in gamma suppression is likely to correlate with other non-invasive indexes of unbalanced excitation/inhibition in the visual cortex provided by, e.g., the H-MRS and visual perceptual tasks.
In conclusion, we suggest that velocity-related changes in gamma oscillations characterize individual variation in E/I balance and efficacy of neural inhibition in the visual cortex and may provide a valuable inhibition-based biomarker for patients with neuropsychiatric disorders.

Funding
This work is financed by Torsten Söderberg Foundation (M240/13to CG) and the Russian