Chronic stimulation of cultured neuronal networks boosts low-frequency oscillatory activity at

theta and gamma with spikes phase-locked to gamma frequencies

This article has been downloaded from IOPscience. Please scroll down to see the full text article.

2012 J. Neural Eng. 9 026015

(http://iopscience.iop.org/1741-2552/9/2/026015)

Download details:

IP Address: 130.251.200.3

The article was downloaded on 13/03/2012 at 07:25

Please note that terms and conditions apply.

View the table of contents for this issue, or go to the journal homepage for more

Home Search Collections Journals About Contact us My IOPscience

IOP PUBLISHING JOURNAL OF NEURAL ENGINEERING

J. Neural Eng. 9 (2012) 026015 (12pp) doi:10.1088/1741-2560/9/2/026015

Chronic stimulation of cultured neuronalnetworks boosts low-frequency oscillatoryactivity at theta and gamma with spikesphase-locked to gamma frequenciesStathis S Leondopulos1, Michael D Boehler1, Bruce C Wheeler2

and Gregory J Brewer1,3

1 Department of Medical Microbiology, Immunology and Cell Biology, Southern Illinois University,School of Medicine, Springfield, IL 62794-9626, USA2 J. Crayton Pruitt Family Department of Biomedical Engineering, University of Florida, Gainesville,FL 32611-6131, USA3 Department of Neurology, Southern Illinois University, School of Medicine, Springfield,IL 62794-9626, USA

E-mail: gbrewer@siumed.edu

Received 13 September 2011Accepted for publication 24 January 2012Published 23 February 2012Online at stacks.iop.org/JNE/9/026015

AbstractSlow wave oscillations in the brain are essential for coordinated network activity but have notbeen shown to self-organize in vitro. Here, the development of dissociated hippocampalneurons into an active network with oscillations on multi-electrode arrays was evaluated in theabsence and presence of chronic external stimulation. Significant changes in signal powerwere observed in the range of 1–400 Hz with an increase in amplitude during bursts.Stimulation increased oscillatory activity primarily in the theta (4–11 Hz) and slow gamma(30–55 Hz) bands. Spikes were most prominently phase-locked to the slow gamma waves.Notably, the dissociated network self-organized to exhibit sustained delta, theta, beta andgamma oscillations without input from cortex, thalamus or organized pyramidal cell layers.

S Online supplementary data available from stacks.iop.org/JNE/9/026015/mmedia

1. Introduction

The hippocampus has long been considered the centerof memory storage and recall. It contains a recurrentneural network in the CA3 region that has beenshown by experimentation and modeling studies to storeepisodic memory—consecutive memory states or ‘items’ thattogether comprise a complete memory (Rolls 2010). Thehealthy mammalian hippocampal network develops from itsembryonic state to maturity with intermittent inputs fromthe cortex, brainstem, thalamus and other areas. Dynamicmemory occurs as these inputs modify synaptic connectionsthrough long-term potentiation, depression (LTP and LTD)

(Rolls 2010) and spike-time dependent plasticity (Colgin andMoser 2010), both Hebbian and non-Hebbian mechanisms.

Although episodic memory formation cannot be directlyobserved, there are methods that gauge the time-varyingsignatures of electrical activity in the brain, includingelectro-encephalograms (EEG) that offer an estimate of theglobal rhythms (in the 1–140 Hz range) from post-synapticpotentials over the entire brain or a large portion thereof(Brown et al 2010). Also, micro-electrode arrays (MEAs)can simultaneously probe local field potentials (LFP) andsingle units from multiple locations in the brain at once.Moreover, the LFP is defined over a similar frequency bandto the EEG allowing precise associations to be drawn betweenspike times and the phase of particular EEG bands. In this

1741-2560/12/026015+12$33.00 1 © 2012 IOP Publishing Ltd Printed in the UK & the USA

J. Neural Eng. 9 (2012) 026015 S S Leondopulos et al

context, although MEAs have been used extensively to recordfrom slice preparations as early as the 1980s (Jobling et al1981, Wheeler and Novak 1986), early studies of dissociatedneurons on MEAs (Gross 1979, Pine 1980, Thomas et al 1972)have focused attention on action potentials or spikes, largelyignoring physiologic frequencies in the range of 1–140 Hz andtheir association to spike timing.

Recent studies of the hippocampus in the intact ratbrain have revealed that memory storage and retrieval areclosely associated with oscillations in the hippocampal LFP(Colgin and Moser 2010). These oscillations tend to appearin four significant forms: delta or sharp waves (1–4 Hz),theta (4–11 Hz), beta (11–30 Hz) and gamma (30–140 Hz).Determining the origins of these oscillations is an importantstep toward understanding the brain’s capacity for signalprocessing, routing of information, attention and control.Studies of hippocampal slices in isolation from the brain revealan emergence of theta oscillations through both excitatory andinhibitory localized feedback loops within the CA1 region(Goutagny et al 2009). Cholinergic activation as well asGABAA-mediated inhibitory post-synaptic potentials (IPSPs)have been associated with oscillations of the LFP in the gammarange (Fisahn et al 1998). In the hippocampus, slow gamma(20–60 Hz) has been observed to originate in the CA3 region,while fast gamma (60–140 Hz) comes from the dentate gyrus(Colgin et al 2009, Fisahn et al 1998). Here we inquiredwhether these low-frequency oscillations could self-organizein a network of isolated hippocampal neurons and whetherchronic stimulation affects the activity.

Cultured neural networks are grown in vitro after theindividual neurons have been extracted from their naturalenvironment in the fetal brain. They are predominantlypyramidal neurons with less numerous inhibitory neurons andastroglia (Banker and Cowan 1977, Brewer et al 1993). Theydevelop synaptic connections amongst themselves followedby the appearance of spontaneous action potentials. However,the isolation of these networks deprives them of the externalstimulation they would receive from other brain regions suchas the entorhinal cortex. Indeed, in the ocular cortex ofrats in vivo, synaptic development is impeded by sensorydeprivation, which indicates that external inputs are neededfor full network development (Maffei et al 2006). In contrast,studies of hippocampal networks in vitro show training effectsor synaptic modification in neural networks that have beenprobe-stimulated for 40 min prior to recording (Eytan et al2003, Stegenga et al 2009). Also, gamma oscillations andsynchronous spiking have been observed in hippocampal slicesimmediately following brief (150–250 ms) tetanic stimulation(Whittington et al 2001, 1997). In addition, long-term(chronic) stimulation over a period of days resulted in highersynapse density (Brewer et al 2009) together with increasedspontaneous spike rates and evoked responses observed duringdevelopment of cultured hippocampal networks (Ide et al2010).

To our knowledge, no previous study has detected thetaor slow gamma waves in cultured hippocampal neuronalnetworks, much less the correlation of action potential eventsto these waves (spike phase synchrony) in culture. Here, we

find that low-frequency oscillations develop spontaneouslyin these networks along with phase-locked spiking, and thatchronic stimulation induces a sustained increase in oscillatoryactivity observed primarily in the delta (1–4 Hz), theta(4–11 Hz) and slow gamma (30–55 Hz) bands.

2. Methods

2.1. Neuron culture and stimulation

Networks were formed from cultures of embryonic day 18Sprague–Dawley rat hippocampal neurons (E18HIPP) at adensity of 50 000 cells per cm2 in the serum-free NbActiv4medium (Brewer et al 2008) on Multichannel System micro-electrode arrays (MEA60 with 30 μm diameter electrodesspaced at 200 μm; Reutlingen, Germany). Cultures wereseparated into three groups (five arrays per group).

Two experimental groups received daily paired-pulsestimulation lasting 1 h and 3 h, respectively (Brewer et al2009, Ide et al 2010), while the control group received noexternal stimulation. Chronic stimulation was applied duringculture development over a period of 12 days starting on day 8in vitro. The stimulation consisted of five pairs of pulses (50 msinter-pulse interval, biphasic, 100 μs/phase duration positivefirst with a wait of 5 s between pairs) performed on each arrayelectrode chosen in a pseudorandom sequence and repeatedfive times. The mode and level of this paradigm are consistentwith reports in the literature (e.g. Eytan and Marom (2006)).The paired pulses with a delay of 50 ms were used to maximizepre-synaptic activation and constituted a minimum number ofstimuli that could avoid short-term plasticity while enhancingthe probability of post-synaptic potentials (Soleng et al 2004).The wait time of 5 s between pulse pairs was empiricallychosen as a minimum that avoided desensitization.

2.2. Recording and signal processing

The data samples were obtained at day 20 ( ± 2 days) as3 min recordings (one recording/array, five arrays/condition)sampled at 40 kHz and amplified 1100 × using a MultichannelSystem amplifier with a bandwidth of 8 Hz–3 kHz. The signalfrom a reference electrode was subtracted from each recording,and data samples were digitally processed using a modificationof SpyCodeV2.0 software (Bologna et al 2010) includingcustom MATLAB scripts. Customizations and scripts wereprogrammed using MATLAB version 7.10.0.499 (R2010a)with MathWorks Signal Processing and Statistics Toolboxes.

2.3. Spike and burst detection criteria

In preparation for spike detection, recordings were digitallyfiltered through a 400-tap Hamming windowed high-pass FIRfilter with a 300 Hz cut-off. Background noise levels werecomputed as the minimum standard deviation of 1 s windowsover the entire recording. To minimize false positives withmaximum sensitivity, spikes were detected as peak-to-peakdeviations of 11 × the standard deviation of the noise (Ideet al 2010) in 2 ms sliding windows with a post-spike dead timeof 1 ms. To automatically reject shorted or open electrodes,

2

J. Neural Eng. 9 (2012) 026015 S S Leondopulos et al

active channels were selected with a minimum backgroundnoise level of 1 μV and a maximum of 10 μV, respectively.To eliminate false spike detections, real spikes were identifiedas having a peak-to-peak height in the range of 24–400 μVand ‘half-height’ widths in the range of 150–900 μs, wherethe height is considered the distance of the spike peak from 0.Active electrodes were selected as having spike rates between0.01 and 100 Hz, while bursts were defined as a series of atleast five spikes separated by no more than a 100 ms inter-spikeinterval.

2.4. Spectrogram processing

In preparation for spectrogram processing, recordings weresubjected to successive decimation stages followed by Fouriertransforms of 1 s windows from each 180 s recording.Windows were chosen small enough to preserve resolutionin time, but large enough to detect delta oscillations as lowas 1 Hz. The decimation stages consist of a 200-tap, low-pass Hamming windowed FIR filter with a 4900 Hz cut-off and a down-sampling to 10 kHz followed by a 200-tap, low-pass Hamming windowed FIR filter with a 950 Hzcut-off and down-sampling to 2 kHz (Geckinli and Yavuz1978). A Fourier transform was then obtained for every 1 swindow for each channel-electrode, resulting in a spectrogramof 1 Hz resolution. Next, to equalize the signal distortionscaused by the analog amplifier characteristic, a point-by-point correction weight was applied to the spectrum usingthe inverse of the amplifier response (figure S1, available atstacks.iop.org/JNE/9/026015/mmedia).

To ensure that single unit spikes were not causing impulseresponse artifacts (or ringing) of the digital filter, the unitimpulse response of the gamma band FIR filter was inspectedand visually compared to gamma oscillations. The frequencycharacteristics and impulse response of the filter are shown infigures S2A and S2B, respectively. Further, the magnitude ofthe ringing in the time domain (figure S2B) is 3000-fold lowerthan the unit impulse, while the peak-to-peak amplitude ofgamma oscillations observed in figures 2(A) and (B) displayonly a two-fold decrease compared to the raw signal in thesame figure. Also, in figure S3 we applied the band-pass filterfor the gamma band to both a low-pass filtered and high-pass filtered version of the raw signal separately. We thusdetermined that the ringing artifact (figure S3F) accountedfor less than 0.05% of the network gamma waveform(figure S3B).

2.5. Frequency bands and spike phase processing

To obtain the phase at which each spike occurred within theslow wave, first the slow wave was enhanced using FIR filters.In particular, the down-sampled signal (at 10 kHz samplingrate) was further processed using a Hamming windowed FIRband-pass filter of length 20 000 (for delta), 10 000 (for betaand theta) and 5000 (for gamma) to enhance the componentin the frequency band of interest, that is delta (1–4 Hz), theta(4–11 Hz), beta (11–30 Hz) or gamma (30–55 Hz). All filteringoperations were time-compensated to account for the inherent

linear-phase lag of the Hamming window and the phase–frequency characteristics of the analog amplifier (figure S1B).The time stamps of detected spikes were then compared withthe band-filtered wave to determine the phase of the spike. Forexample, given spike time Ts is between wave peaks at T1 andT2, then spike phase (ϕ) is computed as in equation (2.1):

ϕ = 2 · π · Ts − T1

T2 − T1. (2.1)

Thus, ϕ = 0◦ (or ϕ = 360◦) corresponds to the peak of thewave, and ϕ = 180◦ the trough.

The tendency of spikes to occur in phase with a particularrhythm (modulation) was measured as a deviation from auniform distribution of spike phases. Thus, a modulation index(MI) was adopted based on the Kullback–Leibler distancebetween the observed phase distribution (P) of N bins andan equivalent uniform distribution Q (Tort et al 2010, Onslowet al 2011):

MI = 1

log(N)·

N∑n=1

Pn · log

(Pn

Qn

). (2.2)

2.6. Glutamate and GAD immunostaining

MEA cultures were immunostained to detect excitatory andinhibitory puncta and somata. Cultures were fixed with a 4%paraformaldehyde, 4% sucrose solution in phosphate bufferedsaline (PBS) for 20 min. Non-specific sites were blockedand cells were permeabilized for 5 min in 5% normal goatserum (NGS) and 0.5% Triton X-100 (TX-100) in PBS.Primary and secondary antibodies were diluted in 5% NGS,and 0.05% TX-100 in PBS. To label excitatory connectionswe used primary rabbit anti-glutamate (1:200; ChemiconAB133) with secondary AlexaFluor-488 (1:1000; MolecularProbes A11034) anti-rabbit antibody. To label inhibitoryconnections, we used mouse anti-glutamic acid decarboxylase(GAD) 65 (1:250; Sigma G1166) with secondary AlexaFluor-568 (1:2000; Molecular Probes A11031) labeled anti-mouseantibody. Cells were incubated overnight with primaryantibodies at 4 ◦C and for 1 h at room temperaturein secondary antibodies. Bisbenzamide was used to labelsomata (final concentration 1 μg ml−1 in PBS, 2 min roomtemperature). Immunostaining densities were collected atconstant exposure through a 20 × objective and analyzed withImage-pro+ (Media Cybernetics) at constant threshold.

2.7. Statistics

ANOVA was used to compare effects of 0, 1 and 3 hchronic stimulation with significance set to p < 0.001.The reported n is degrees of freedom for the combinationsof all measures. Means and standard errors are reportedunless otherwise specified. Long-tailed distributions of thespectra were treated as log-normal models (a naturallogarithmic transformation defined in table S1). Choosing alog-normal statistical model over a straight normal and root-normal model was based on a number of criteria includingvisual inspection (figure S4), proximity of the mean tothe median (figure S5) and kurtosis/skewness measures(figure S6) computed using the formulae in table S2. Total

3

J. Neural Eng. 9 (2012) 026015 S S Leondopulos et al

(A) (B)

(C) (D) (E)

Figure 1. (A) Bursts in non-stimulated networks have a longer duration and are less frequent than in (B) 3 h chronically stimulated networksdisplaying short bursts with higher spike frequency. (C) Spike rate within bursts increases with chronic stimulation. (D) Burst durationdecreases with chronic stimulation. (E) Bursts are separated by shorter intervals in chronically stimulated networks.

(or integrated) power within each band also had a long-taileddistribution modeled here as log-normal (figure S7, availableat stacks.iop.org/JNE/9/026015/mmedia) as were the reportedburst measures (figure S8).

3. Results

Chronically stimulated and non-stimulated networks werecompared in terms of burst measures, spectrograms and spikephase distributions for each slow wave frequency band.

3.1. Burst analysis

We confirmed and extended our previous findings that,compared to non-stimulated networks (figure 1(A)), chronicstimulation increases the tendency of spikes to group togetherinto shorter, more intense and more frequent bursts of activity(figure 1(B)). Averages over five arrays per condition showthat stimulated networks display roughly a 20% increase inmean intra-burst spike rate (C) and a 25% reduction in burstduration (D), complemented by a 30% decrease of inter-burstinterval (E).

3.2. Band waves in dissociated neuron networks

We hypothesized that the network would self-organize toproduce slow wave oscillations in the resting membranepotential and drive the spiking activity as observed in vivo(Bragin et al 1995, Lisman 2005, Lisman et al 2005,Montgomery et al 2008) and in slices (Traub et al 2004,Whittington et al 1997). Low-frequency band filtering of theraw signals from three-week-old networks revealedoscillations in the delta band (1–4 Hz), theta band (4–11 Hz),beta band (11–30 Hz) and gamma band (30–55 Hz) occurringprimarily during bursts (figures 2(A) and (B)). Figures 2(C)

and (D) show the five most active channels superimposed overthe beta and gamma waves. In many instances, the beta andgamma oscillations were visibly mirrored in the raw signal, buttheir phases were sometimes different on adjacent recordingsites.

3.3. Spectrograms and duration of stimulation

Spectrograms of MEA recordings were computed separatelyfor 1 s segments containing bursts and 1 s segmentscontaining no spikes. Bursty segments constituted 5%, 9%and 8% of the total number of segments for 0, 1 and3 h chronically stimulated networks, respectively, while non-spiking segments constituted 80%, 72% and 77% of thetotal number of segments (figures 3(A)–(D)). Spectrogramsof segments containing spikes but not bursts (12%, 14% and12% of total segments) were very similar to spectrogramsof segments with no spikes, except for a decrease in deltaband energy (figures S9 and S10). The prominence ofgamma activity (30–55 Hz) in bursts is seen for both 1 and3 h day–1 stimulated networks (figures 3(B) and (C)), butis absent from the non-stimulated networks (figure 3(A)).Moreover, the 1 h day–1 (figure 3(B)) stimulated networksdisplay a preference for the delta band (1–4 Hz), while the3 h day–1 stimulated networks (figure 3(C)) show moretheta band (4–11 Hz) activity. Not surprisingly, the non-spiking windows have a spectrogram that is very similar toa control spectrogram without neurons obtained from an arraycontaining only saline solution (figure 3(D)).

With respect to the duration of chronic stimulation,spectrograms of non-spiking segments show a monotonicincrease in band power in all slow wave bands except delta(1–4 Hz) (figure 3(E)). In contrast, spectrograms of burstysegments (figure 3(F)) show a monotonic increase only in

4

J. Neural Eng. 9 (2012) 026015 S S Leondopulos et al

(A)

(B)

(C)

(D)

Figure 2. (A) Signals filtered to specific frequency bands show increases in wave amplitude during burst events. (B) Details of specificfrequency bands show slow wave progression within a single burst event. (C) The five most active channels in one recording aresuperimposed with corresponding beta oscillation during a burst event. Some portions of the beta wave display synchronicity betweenelectrodes (latter half of the segment) while other portions do not (earlier half) in this 550 ms sample. (D) Gamma oscillations follow thebaseline of the raw signal. Gamma also increases in amplitude during a burst event and is occasionally synchronous between electrodes inthis 550 ms sample.

5

J. Neural Eng. 9 (2012) 026015 S S Leondopulos et al

(A) (B)

(C) (D)

(E) (F)

Figure 3. (A) For non-stimulated networks, the spectrogram of bursty signal segments has significantly greater power than the spectrogramof non-spiking segments. (B) Power in the delta (�) and slow gamma (�s) bands is particularly prominent for 1 h chronically stimulatednetworks. (C) Power in the theta (�) and slow gamma (�s) bands is particularly prominent for 3 h chronically stimulated networkscompared to beta (B), mid gamma (�m), fast gamma (�f) and ripple (R) bands. (D) Spectrogram of a recording from an MEA containingsaline solution (control without neurons) shows similarity to spectrograms of non-spiking segments. (E) Comparison of spectrograms ofnon-spiking segments shows an increase in low-frequency (1–200 Hz) power for networks that have undergone 3 h daily stimulation.(F) Comparison of spectrograms of segments containing bursts shows prominence of delta (�) in the 1 h chronically stimulated networksand a shift into the theta (�) range for the 3 h chronically stimulated networks. A monotonic increase in gamma with chronic stimulationduration is evident. 60 Hz harmonics ± 3 Hz have been removed from all the plots.

the theta and slow gamma bands. For non-spiking segments,integrated power increased by 15% (from non-stimulated to3 h stimulated networks) in the delta band and 22–24% in theta,beta and slow gamma bands (figure 4(A)). For bursty segments(figure 4(B)), integrated power increased by 23% in the deltaband, 38% in theta and 20–21% in beta and slow gamma bands.Overall, the power in bursty segments was three- to five-foldhigher than in non-spiking segments.

3.4. Relationship of spike time to wave oscillation phase

In observing network organization, we tested the hypothesisthat networks develop coordination of spike activation atcertain phases of slow wave oscillations (a) in a self-organizedmanner or (b) in response to chronic stimulation. We examinedthe timing relationship or phase-locking of single unit activity(each spike) to each band of slow wave oscillations up toslow gamma. We found that phase-locking was strongest inthe slow gamma wave as shown in the example burst offigure 5(A). However, isolated spikes (far from bursts) did notshow strong correlations with gamma phase and amplitude(figure 5(B)). For gamma waves, the distribution of all spikephases in five non-stimulated arrays produced a peak at

the upstroke of the cycle near 225◦ (figure 6(A)). Chronicstimulation did not affect the phase of the synchrony asobserved in the stimulated arrays (figure 6(B)). In the case ofbeta waves, spikes were phase-locked near 270◦ (figures 6(C)and (D)), though less prominently than in gamma, and withlittle effect of chronic stimulation. In theta, the distributionof all spike phases of each experimental condition showedlittle preference or focus (figures 6(E) and (F)). Delta phasemodulation of spikes was broadly distributed near the peakof the wave (0◦) for all experimental conditions (figures 6(G)and (H)). Overall, spike phase becomes more focused in thedelta, beta, and particularly the slow gamma bands, but is notsignificantly influenced by chronic stimulation.

Looking at spike timing and phase on an electrode-by-electrode basis revealed a strong preference of spikesto occur at particular phases in each of the slow wavebands, hence a high MI (figure 7, upper trace). However,inasmuch as this phase varied among electrodes, the meandistribution of the corresponding array acquired a weaker MI(figure 7, lower trace). Chronic stimulation had little effect onthe MI of spike phases; hence all experimental groups werecombined in the results of figure 7. Overall, for both local (byelectrode) and aggregate (by array) phase distributions, the

6

J. Neural Eng. 9 (2012) 026015 S S Leondopulos et al

(A)

(B)

Figure 4. (A) For non-spiking segments, total power per frequency band displays a monotonic increase with stimulation for all bands exceptfor the delta band. (B) For segments containing bursts, total power per frequency band is higher with stimulation for all bands but the rippleband, displaying a monotonic increase with stimulation duration for the theta (4–11 Hz) and slow gamma (30–55 Hz) bands. Error bars are± standard error. N is the number of 1 s segments analyzed for each band. Per cent increase from minimum to maximum is indicated.p values indicate overall effect of duration of chronic stimulation at each band.

theta band displayed the smallest deviations from uniformity,while gamma displayed the strongest MI.

3.5. Chronic stimulation alters the ratio of excitatory toinhibitory neurons

We evaluated the effect of chronic stimulation onthe composition of the network as excitatory neuronsimmunostained for glutamate and inhibitory neurons stainedfor GAD. Figure 8(A) shows nearly a 50% reduction ininhibitory neuron (GAD) density associated with chronicstimulation. The density of excitatory neurons (figure 8(B))was unaffected by chronic stimulation and exceeded theinhibitory staining by four to eightfold. Thus, chronic

stimulation evoked a marked decrease in the relative numberof inhibitory to excitatory connections that could explain thehigher intra-burst spike rates in figure 1(C).

4. Discussion

Networks in this study that were chronically stimulatedproduced spikes in shorter bursts, with higher spike ratesand higher burst rates (lower inter-burst interval) associatedwith a decrease in the inhibitory neuron composition of thenetwork. Spectral analysis revealed that most of the oscillatoryactivity in our networks occurred during bursts with peaks inthe delta (1–4 Hz), theta (4–11 Hz) and slow gamma (30–55 Hz) bands. Furthermore, paired-pulse chronic stimulation

7

J. Neural Eng. 9 (2012) 026015 S S Leondopulos et al

(A)

(B)

Figure 5. (A) Spikes in the gamma wave (30–55 Hz) tend to occur near 225◦ (−135◦) from the peak in the wave. (B) Isolated spikes do notinduce significant filter ringing.

caused a monotonic increase in power at theta and gammafrequency bands. Other stimulation frequencies and trainswere not explored. Notably, these effects occurred withoutthe complex inputs of extra-hippocampal brain regions andwithout the ordered tri-synaptic network in layers of pyramidalneurons. Phase-locking of spikes was strongest in the gammawave 225◦ (−135◦) or upstroke of the cycle (the peak beingat 0◦) for all arrays, and was also strong in the remainingbands albeit with varied phase relationships across electrodes.Since extracellular recordings are associated with the time-differential of the trans-membrane voltage (Claverol-Tintureand Pine 2002), this suggests spike timing at a particularphase of the intracellular IPSP. Although spike timing wassignificantly correlated with all slow wave bands, chronicstimulation did not cause a significant shift in phase or increasein modulation strength. These results have implications for thecoding of information in bursts, time and frequency.

The tendency of spikes to occur in bursts is a phenomenonassociated with network maturation (Wagenaar et al 2006)while the overall increase in spike and burst rates isconsistent with the presence of more excitatory connectionsover inhibitory connections. Although theta waves have beenassociated with both inhibitory and excitatory connections(Kendrick et al 2011), our observed increased ratio ofexcitatory to inhibitory composition with chronic stimulationsuggests that less inhibitory drive contributes to the increasedtheta power.

Spikes consistently follow the beta and gamma waves(at ∼225◦) even after chronic stimulation, suggesting a closerelation of the observed oscillation to the corresponding singleunit activity. If these oscillations correspond to sub-thresholdcurrent events such as excitatory or inhibitory post-synaptic

potentials mediated by specific receptors, then associationsmay be captured between slow waves observed on any oneelectrode and single unit activity of other electrodes in thearray. This is partially confirmed by the tendency of burstevents to be a network-wide phenomenon, observed to occursimultaneously on multiple array channels. Thus, as slowwaves increase in amplitude during these events, they alsocoincide with increased single unit activities on other channels.Further analysis using cross-correlations will elucidate thisrelationship more clearly.

Theta and gamma oscillations are proposed to servea number of purposes related to the encoding of episodicmemory, higher order associations and learning (Buzsaki 2005,Colgin and Moser 2010, Lisman and Buzsaki 2008). Thepresence of these bands in cultured networks and their closeassociation with single unit spikes suggests that networksof hippocampal neurons may form an inherent mechanismof spike phase synchrony that is only modulated throughfeedback connectivity with the cortex.

4.1. Slow gamma band power and phase-locking

The spectrograms in figure 3(F) and bar charts of figure 4(B)indicate that chronic stimulation has a more prominent effecton the slow gamma band rather than the fast gamma band.In the intact brain, a large proportion of neurons in the CA3are synchronized to CA1 at the slow gamma cycle, while themedial entorhinal cortex (MEC) region is largely in tune withthe CA1 fast gamma (Colgin et al 2009). In the context of ourcultured networks, the prominence of slow gamma over thefast gamma range may be due to the absence of MEC neuronsin our preparations.

8

J. Neural Eng. 9 (2012) 026015 S S Leondopulos et al

(A) (B)

(C) (D)

(E) (F)

(G) (H)

Figure 6. Polarization (in phase) of all spike times for (A) gamma wave—no stimulation, (B) gamma wave—3 h chronic stimulation,(C) beta wave—no stimulation, (D) beta wave—3 h chronic stimulation, (E) theta wave—no stimulation, (F) theta wave—3 h chronicstimulation, (G) delta wave—no stimulation, and (H) delta wave—3 h chronic stimulation.

Patch-clamp studies of whole cells within slices indicatesynchronous firing within the gamma band associated withactivation of metabotropic glutamate receptors after brief (15to 20 pulses, 50 μs pulse width, 80–100 Hz) tetanic stimulation(Whittington et al 2001, 1997). In the present study, it may bethat our stimulation protocol induces a similar metabotropicreaction in the short term. However, after repeated (chronic)application of the protocol over a period of 20 days, thenetwork response may become sustained or learned throughLTP.

Other mechanisms reported to facilitate gammaoscillations in slice preparations and in vivo studies includeexternal pacemakers, gap junctions and feedback loops withinthe hippocampus, or a combination of these (Mann and Paulsen2005). In isolated networks of dissociated hippocampal cells,

external pacemakers are ruled out. However, there is evidencesuggesting that oscillations are enabled by both inhibitoryand excitatory paths acting in concert (Mann and Paulsen2007). Our finding that the ratio of excitatory-to-inhibitoryconnections increases with chronic stimulation is furtherevidence to this.

Slow gamma waves in the intact hippocampus appear toresult from IPSPs that connect CA3 neurons to CA1 for thepurpose of memory recall (Colgin et al 2009, Fisahn et al1998). The results of our study show a strong correlationbetween spike times and the upstroke (225◦) phase of thegamma wave. Further studies in which specific receptorsare pharmacologically inhibited are needed to confirm thatGABAA receptors (Vreugdenhil et al 2005) are responsiblefor this oscillation in culture.

9

J. Neural Eng. 9 (2012) 026015 S S Leondopulos et al

Figure 7. MI is a normalized form of the Kullback–Leibler distancebetween the distribution of phases and a uniform probabilitydistribution (Tort et al 2010, Onslow et al 2011). MI reflectsdeviation from uniformity (0 is purely uniform). MI is greater ineach slow wave band for individual electrode distributions (allexperimental and control conditions included) than for thecorresponding array-wide distributions. In array-wide distributions,spikes are strongly modulated with the beta and slow gamma bands,but not significantly with other bands. p values indicate significanceof deviation from zero (single-tailed t-test).

(A) (B)

Figure 8. (A) Networks without chronic stimulation have nearlytwice as many inhibitory GAD+ immunofluorescent soma (openbar) compared to networks with 3 h day–1 chronic stimulation(hatched bar). (B) Excitatory glutamate (GLU) positive somaimmuno-fluorescence is similar for unstimulated (open bar) orstimulated (closed bar) networks (n = 2 MEAs per condition, eightfields per MEA).

Spike timing was also correlated with the beta cycle near270◦, more so on individual neurons than in the network asa whole, although the modulation strength was less than thatof the slow gamma waves. It is interesting that the beta bandencompasses our paired-pulse frequency of 20 Hz (50 ms pulseinterval). This raises the question of whether shorter or longerpulse intervals would produce stronger effects in gamma ortheta compared to beta. In intact hippocampal slices, powerin the beta region increases following the decay of gamma

activity induced by tetanic stimulation (Bracci et al 1999,Vreugdenhil et al 2005).

4.2. Theta band power increases with chronic stimulation indissociated networks

Studies of the hippocampus in isolation reveal an emergenceof theta oscillations through both excitatory and inhibitorylocalized feedback loops within the CA1 region (Goutagnyet al 2009). Our studies suggest that dissociated neuronsfrom the embryonic hippocampus have the capacity to self-organize these feedback loops, independent of external inputs.In the intact brain, the medial septum-diagonal band of Broca,possibly in conjunction with feedback connections to andfrom the hippocampus, plays a central role as a pacemakerin theta rhythm generation (Buzsaki et al 2002). In awakeand behaving rats, studies of the hippocampus reveal a roleof theta in modulating the gamma wave (Bragin et al 1995).In studies of the intact hippocampus, theta is implicated as aswitch in the memory encoding and recall cycle. In particular,slow gamma and recall are associated with spikes appearingon the falling edge (90◦) of theta (Colgin et al 2009). Ourfinding of lower levels of GAD-positive inhibitory neuronstaining with stimulation suggests that the increase in theta andgamma waves was driven by a shift toward more excitatoryover inhibitory connections.

4.3. Sharp waves, the delta band and bursts

We find delta or sharp waves (1–4 Hz) in association withsynchronous bursts in our networks that increase with chronicstimulation. These findings suggest that sharp waves and thetaoscillations are also inherent to dissociated neural networksand do not exclusively emerge in the dense connectivitypatterns of the intact hippocampus. Short-term stimulationdecreases bursts (Stegenga et al 2010, Wagenaar et al 2005).However, we find that chronic stimulation increases burst rates.

Sharp waves (0.3–4 Hz) are associated with synchronouspopulation bursts from recurrent connections in the CA3(Buzsaki 1986, Csicsvari et al 2000). Both in vivo andin vitro studies of oscillations in the sharp wave range reveal anintrinsic capability of the hippocampus to generate waveformsin this frequency range without external input (Csicsvari et al1999, Kubota et al 2003). Sharp waves in the delta band arebelieved to be responsible for associating (or consolidating)large numbers of cell assemblies (or memory items) duringslow wave sleep, awake immobility, and idle tasks (Buzsaki1986, Wagner et al 2004). Delta waves are also associatedwith drowsiness and sleep (Molle et al 2006). With a shiftfrom a wakeful to a drowsy state in rats, the hippocampalEEG spectrum shifts abruptly from theta to the delta (Tani andIshihara 1988). Although this shift was not evident betweenour non-stimulated and 1 h stimulated networks, there was arespective increase in theta band power (figures 4(A) and (B))and a marked ‘theta-shift’ between the spectrograms of our1 h and 3 h stimulated networks (see figure 3(F)). This raisesthe question of whether the absence of external stimulationimposes a ‘sleep’ state on the network. Conversely, some form

10

J. Neural Eng. 9 (2012) 026015 S S Leondopulos et al

of external stimulation may be required to activate an ‘awake’state, as in the intact hippocampus.

In summary, the presence of low-frequency oscillationsand their association with spikes and bursts in networks fromdissociated hippocampal cells suggests that cultured networksof neurons tend to form connections that self-assemble theseoscillations. Chronic stimulation enhances these waveforms,suggesting a possible emulation of in vivo conditions.Spikes were most prominently phase-locked to the upstroke(225◦–270◦) of the slow gamma and beta waves regardlessof chronic stimulation, and also phase-locked to local deltaand theta oscillations suggesting an inherent mechanism ofspike phase synchrony in networks of hippocampal neurons.Thus, hippocampal neurons self-assemble to emulate in vivoconditions even when they are dissociated from each other inculture at the embryonic stage of development. This findingfurther complements the existing evidence of a highly adaptiveand modular hippocampus. Also, this enables unprecedentedpharmacologic access in the investigation of the origins andpropagation of slow wave oscillations in cultured hippocampalnetworks.

Acknowledgments

This work was supported in part by NIH NS052233 (BCW,PI). The authors would like to thank Professor S Martinoia,Dr Michela Chiappalone and all the authors of SpyCodeV2.0MEA analysis software. GJB invented the NbActiv4 culturemedium used in these studies and obtained from his company,BrainBits LLC. Other authors have no competing financialinterests to declare. SSL analyzed nearly all the data and wrotethe paper. GJB and BCW designed the research and editedthe paper. MDB isolated neurons, prepared and stimulatedcultures, collected signals, analyzed data and wrote portionsof the paper.

References

Banker G A and Cowan W M 1977 Rat hippocampal neurons indispersed cell culture Brain Res. 126 397–425

Bologna L L, Pasquale V, Garofalo M, Gandolfo M, Baljon P L,Maccione A, Martinoia S and Chiappalone M 2010Investigating neuronal activity by SPYCODE multi-channeldata analyzer Neural Netw. 23 685–97

Bracci E, Vreugdenhil M, Hack S P and Jefferys J G R 1999 On thesynchronizing mechanisms of tetanically induced hippocampaloscillations J. Neurosci. 19 8104–13

Bragin A, Jando G, Nadasdy Z, Hetke J, Wise K and Buzsaki G1995b Gamma (40–100 Hz) oscillation in the hippocampus ofthe behaving rat J. Neurosci. 15 47–60

Brewer G J, Boehler M D, Ide A N and Wheeler B C 2009 Chronicelectrical stimulation of cultured hippocampal networksincreases spontaneous spike rates J. Neurosci. Methods184 104–9

Brewer G J, Boehler M D, Jones T T and Wheeler B C 2008NbActiv4 medium improvement to Neurobasal/B27 increasesneuron synapse densities and network spike rates onmultielectrode arrays J. Neurosci. Methods 170 181–7

Brewer G J, Torricelli J R, Evege E K and Price P J 1993 Optimizedsurvival of hippocampal neurons in B27-supplementedNeurobasal, a new serum-free medium combinationJ. Neurosci. Res. 35 567–76

Brown E N, Lydic R and Schiff N D 2010 General anesthesia, sleep,and coma N. Engl. J. Med. 363 2638–50

Buzsaki G 1986 Hippocampal sharp waves: their origin andsignificance Brain Res. 398 242–52

Buzsaki G 2005 Theta rhythm of navigation: link between pathintegration and landmark navigation, episodic and semanticmemory Hippocampus 15 827–40

Buzsaki G, Csicsvari J, Dragoi G, Harris K, Henze D and Hirase H2002 Homeostatic maintenance of neuronal excitability byburst discharges in vivo Cereb. Cortex 12 893–9

Claverol-Tinture E and Pine J 2002 Extracellular potentials inlow-density dissociated neuronal cultures J. Neurosci. Methods117 13–21

Colgin L L, Denninger T, Fyhn M, Hafting T, Bonnevie T,Jensen O, Moser M B and Moser E I 2009 Frequency ofgamma oscillations routes flow of information in thehippocampus Nature 462 353–7

Colgin L L and Moser E I 2010 Gamma oscillations in thehippocampus Physiology 25 319–29

Csicsvari J, Hirase H, Czurko A, Mamiya A and Buzsaki G 1999Fast network oscillations in the hippocampal CA1 region of thebehaving rat J. Neurosci. 19 RC20

Csicsvari J, Hirase H, Mamiya A and Buzsaki G 2000 Ensemblepatterns of hippocampal CA3-CA1 neurons during sharpwave-associated population events Neuron 28 585–94

Eytan D, Brenner N and Marom S 2003 Selective adaptation innetworks of cortical neurons J. Neurosci. 23 9349–56

Eytan D and Marom S 2006 Dynamics and effective topologyunderlying synchronization in networks of cortical neuronsJ. Neurosci. 26 8465–76

Fisahn A, Pike F G, Buhl E H and Paulsen O 1998 Cholinergicinduction of network oscillations at 40 Hz in the hippocampusin vitro Nature 394 186–9

Geckinli N C and Yavuz D 1978 Some novel windows and a concisetutorial comparison of window families IEEE Trans. Acoust.Speech Signal Process. 26 501–7

Goutagny R, Jackson J and Williams S 2009 Self-generated thetaoscillations in the hippocampus Nat. Neurosci.12 1491–3

Gross G W 1979 Simultaneous single unit recording in vitro with aphotoetched laser deinsulated gold multimicroelectrode surfaceIEEE Trans. Biomed. Eng. 26 273–9

Ide A N, Andruska A, Boehler M, Wheeler B C and Brewer G J2010 Chronic network stimulation enhances evoked actionpotentials J. Neural Eng. 7 16008

Jobling D T, Smith J G and Wheal H V 1981 Active microelectrodearray to record from the mammalian central nervous systemin vitro Med. Biol. Eng. Comput. 19 553–60

Kendrick K M, Zhan Y, Fischer H, Nicol A U, Zhang X and Feng J2011 Learning alters theta amplitude, theta-gamma couplingand neuronal synchronization in inferotemporal cortex BMCNeurosci. 12 55

Kubota D, Colgin L L, Casale M, Brucher F A and Lynch G 2003Endogenous waves in hippocampal slices J. Neurophysiol.89 81–9

Lisman J 2005 The theta/gamma discrete phase code occurringduring the hippocampal phase precession may be a moregeneral brain coding scheme Hippocampus 15 913–22

Lisman J and Buzsaki G 2008 A neural coding scheme formed bythe combined function of gamma and theta oscillationsSchizophr. Bull. 34 974–80

Lisman J E, Talamini L M and Raffone A 2005 Recall of memorysequences by interaction of the dentate and CA3: a revisedmodel of the phase precession Neural Netw. 18 1191–201

Maffei A, Nataraj K, Nelson S B and Turrigiano G G 2006Potentiation of cortical inhibition by visual deprivation Nature443 81–4

Mann E O and Paulsen O 2005 Mechanisms underlying gamma(‘40 Hz’) network oscillations in the hippocampus—amini-review Prog. Biophys. Mol. Biol. 87 67–76

11

J. Neural Eng. 9 (2012) 026015 S S Leondopulos et al

Mann E O and Paulsen O 2007 Role of GABAergic inhibition inhippocampal network oscillations Trends Neurosci. 30 343–9

Molle M, Yeshenko O, Marshall L, Sara S J and Born J 2006Hippocampal sharp wave-ripples linked to slow oscillations inrat slow-wave sleep J. Neurophysiol. 96 62–70

Montgomery S M, Sirota A and Buzsaki G 2008 Theta and gammacoordination of hippocampal networks during waking andrapid eye movement sleep J. Neurosci. 28 6731–41

Onslow A C, Bogacz R and Jones M W 2011 Quantifyingphase-amplitude coupling in neuronal network oscillationsProg. Biophys. Mol. Biol. 105 49–57

Pine J 1980 Recording action potentials from cultured neurons withextracellular microcircuit electrodes J. Neurosci. Methods2 19–31

Rolls E T 2010 A computational theory of episodic memoryformation in the hippocampus Behav. Brain Res. 215 180–96

Soleng A F, Baginskas A, Andersen P and Raastad M 2004Activity-dependent excitability changes in hippocampal CA3cell Schaffer axons J. Physiol. 560 491–503

Stegenga J, le Feber J, Marani E and Rutten W 2009 The effect oflearning on bursting IEEE Trans. Biomed. Eng. 56 1220–7

Stegenga J, le Feber J, Marani E and Rutten W L 2010Phase-dependent effects of stimuli locked to oscillatory activityin cultured cortical networks Biophys. J. 98 2452–8

Tani Y and Ishihara T 1988 Changes in EEG associated withsleep–awake behavior in young adult versus aged adultFischer-344 rats Physiol. Behav. 44 389–92

Thomas C A Jr, Springer P A, Loeb G E, Berwald-Netter Yand Okun L M 1972 A miniature microelectrode array tomonitor the bioelectric activity of cultured cells Exp. Cell Res.74 61–6

Tort A B, Komorowski R, Eichenbaum H and Kopell N 2010Measuring phase-amplitude coupling between neuronaloscillations of different frequencies J. Neurophysiol.104 1195–210

Traub R D, Bibbig A, LeBeau F E, Buhl E H and Whittington M A2004 Cellular mechanisms of neuronal populationoscillations in the hippocampus in vitro Annu. Rev. Neurosci.27 247–78

Vreugdenhil M, Bracci E and Jefferys J G 2005 Layer-specificpyramidal cell oscillations evoked by tetanic stimulation in therat hippocampal area CA1 in vitro and in vivo J. Physiol.562 149–64

Wagenaar D A, Madhavan R, Pine J and Potter S M 2005Controlling bursting in cortical cultures with closed-loopmulti-electrode stimulation J. Neurosci. 25 680–8

Wagenaar D A, Pine J and Potter S M 2006 An extremely richrepertoire of bursting patterns during the development ofcortical cultures BMC Neurosci. 7 11–7

Wagner U, Gais S, Haider H, Verleger R and Born J 2004 Sleepinspires insight Nature 427 352–5

Wheeler B C and Novak J L 1986 Current source density estimationusing microelectrode array data from the hippocampal slicepreparation IEEE Trans. Biomed. Eng. 33 1204–12

Whittington M A, Doheny H C, Traub R D, LeBeau F E Nand Buhl E H 2001 Differential expression of synaptic andnonsynaptic mechanisms underlying stimulus-induced gammaoscillations in vitro J. Neurosci. 21 1727–38

Whittington M A, Stanford I M, Colling S B, Jefferys J Gand Traub R D 1997 Spatiotemporal patterns of gammafrequency oscillations tetanically induced in the rathippocampal slice J. Physiol. 502 (Pt 3) 591–607

12