Spatiotemporal characteristics and pharmacological modulation of multiple gamma oscillations in the CA1 region of the hippocampus

Multiple components of “γ-oscillations” between 30–170 Hz in the CA1 region of the hippocampus have been described, based on their coherence with oscillations in other brain regions and on their cross-frequency coupling with local θ-oscillations. However, it remains unclear whether the different sub-bands are generated by a single broadband oscillator coupled to multiple external inputs, or by separate oscillators that incorporate distinct circuit elements. To distinguish between these possibilities, we used high-density linear array recording electrodes in awake behaving mice to examine the spatiotemporal characteristics of γ-oscillations and their responses to midazolam and atropine. We characterized oscillations using current source density (CSD) analysis, and measured θ-γ phase-amplitude coupling by cross frequency coupling (CFC) analysis. Prominent peaks were present in the CSD signal in the mid- and distal apical dendritic layers at all frequencies, and at stratum pyramidale for γslow (30–45 Hz) and γmid (50–90 Hz), but not γfast (90–170 Hz) oscillations. Differences in the strength and timing of θ-γslow and θ-γmid cross frequency coupling, and a lack of coupling at the soma and mid-apical region for γfast oscillations, indicated that separate circuit components generate the three sub-bands. Midazolam altered CSD amplitudes and cross-frequency coupling in a lamina- and frequency specific manner, providing further evidence for separate generator circuits. Atropine altered CSD amplitudes and θ-γ CFC uniformly at all locations. Simulations using a detailed compartmental model were consistent with γslow and γmid oscillations driven primarily by inputs at the mid-apical dendrites, and γfast at the distal apical dendrite. Our results indicate that multiple distinct local circuits generate γ-oscillations in the CA1 region of the hippocampus, and provide detailed information about their spatiotemporal characteristics.


INTRODUCTION
Oscillations in the brain span a wide range of frequencies and play a variety of roles in different brain structures. Different frequency oscillations are thought to be generated by distinct cellular and network mechanisms (Buzsáki, 2006). In general, higher frequency oscillations reflect the synchronous activity of locally connected cell assemblies, whereas lower frequency oscillations support longer-range coordination and communication (Singer, 1993;Singer and Gray, 1995;Fries, 2005).
Their broad frequency range, lack of distinct peaks, and local expression driven by external structures, has complicated the analysis of underlying cellular mechanisms of γ-oscillations. Locally within the CA1 region, both perisomatic inhibition and apical dendritic feed-forward inhibition are thought to contribute to the generation and expression of γ-oscillations, with periodic suppression by θ-frequency inhibitory inputs resulting in θ-γ coupling (White et al., 2000;Freund, 2003;Buzsáki and Wang, 2012;Campanac et al., 2013;Zemankovics et al., 2013;Lasztóczi and Klausberger, 2014). However, it remains unclear whether the different sub-bands are generated by a single broadband oscillator that is coupled to multiple external inputs, or by separate oscillators that incorporate distinct circuit elements.
To distinguish between these two possibilities, we examined the spatiotemporal characteristics of γ-oscillations and their responses to midazolam and atropine in awake behaving mice. Midazolam is a benzodiazepine that modulates γ subunitcontaining γ-aminobutyric acid type A receptors (GABA A Rs). These are the major inhibitory receptors found at inhibitory synapses in the hippocampus and throughout the forebrain. (Dundee et al., 1980;Sigel and Buhr, 1997;Rudolph et al., 1999;Rudolph and Möhler, 2004). Atropine is a competitive muscarinic acetylcholine receptor (mAChR) antagonist that reduces excitability of pyramidal cells and interneurons and alters transmitter release at a subset of inhibitory nerve endings (Levey et al., 1995;Hájos et al., 1997;Qian and Saggau, 1997;Rouse et al., 2000;Cea-Del Rio et al., 2010). Atropine has been shown to alter γ-oscillations in the CA1 region in vivo and in vitro (Fellous and Sejnowski, 2000;Traub et al., 2000;Whittington et al., 2000;Mann et al., 2005). We characterized the patterns of the underlying currents using current source density (CSD) analysis, and quantified θ-γ phase-amplitude coupling using CFC analysis. We found that γ slow (30-45 Hz) and γ mid (50-90 Hz) oscillations had similar CSD and CFC spatial profiles under drug-free conditions, but that midazolam altered these two sub-bands differently in stratum pyramidale. The γ fast (90-170 Hz) oscillation was distinct from both γ slow and γ mid oscillations in its pattern of CSD amplitudes across different laminae, as well as in its CFC profile. Atropine uniformly increased CSD amplitudes across all bands. In computer simulations employing a CA1 pyramidal neuron compartmental model, the observed physiological activity patterns were best reproduced by γ slow and γ mid oscillatory inputs at the level of the mid-apical dendrite and γ fast oscillations at the distal apical dendrite. Our results thus indicate that distinct local circuits generate the different γ-oscillators, and that the different sub-bands are differentially modulated by midazolam but not atropine.

IN VIVO
The details of surgical methods were described previously. The results presented here belong to the same recorded data set from the same animals used for theta frequency band (4-12 Hz) oscillation analysis (Balakrishnan and Pearce, 2014). The experimental protocol was approved by the University of Wisconsin Institutional Animal Care and Use and complied with National Institutes of Health guidelines.
In brief, seven adult male homozygous mice derived from heterozygous breeding pairs carrying the GABA A R α5-H105R mutation were implanted with 15 μm thick, 3 mm long, 16channel linear array recording probes (NeuroNexus Technologies "A" type probe, CM-series package, 50 μm spacing). These "pseudo wild type" animals do carry addition genetic material introduced as part of the gene-targeting strategy, and although no changes in gene expression are expected or have been described, it is possible that some exist. Surgery was performed under isoflurane anesthesia using sterile technique. The skull was secured in a stereotaxic apparatus (KOPF® Instruments, Model 900) with a heated platform maintained by a circulating water bath (Stoelting®). Respiratory rate, temperature, and movement were monitored throughout surgery. The target recording area was the CA1 region of the dorsal hippocampus, with recording sites spanning the layers between the alveus and the hippocampal fissure. Post-surgical analgesia was provided by injecting mice with buprenorphine (0.1 mg/kg s.c.). Animals were monitored during recovery, and the first recording session was performed 7-10 days after electrode implantation.
Following the completion of all experiments, electrode location was verified by histological analysis (Figure 1A). In addition to estimating the laminar positions of recording sites based on histology, the location of the pyramidal cell layer was established as the recording site that showed the maximal amplitude of the ripple oscillation (170-250 Hz) during immobility.
Local field potentials (LFPs) were recorded using a Tucker Davis Technology (TDT®) recording system (System 3). Signals were band pass filtered from 2-6000 Hz using a zero-phase distortion digital Butterworth filter (filtfilt), acquired at a 12 KHz sampling frequency, and then extracted to a MATLAB® v2008-readable format by lowpass filtering at 500 Hz and downsampling to 1 KHz. Data were analyzed using custom-written MATLAB® v2012b routines. Time-stamped behavioral scoring was performed manually with a TDT® scoring box (BBOX) connected to the recording system. To characterize the LFP, CSD, and drug effects on γ slow (30-45 Hz), γ mid (50-90 Hz), and γ fast (90-170 Hz), contiguous data segments greater than 1 s, obtained during exploratory behavior, were filtered into individual frequency bands using band pass Butterworth filters (filtfilt).
Two drug treatments were used: (1) Midazolam 1.25 mg/kg subcutaneous (s.c.), which in other experiments was established as the dose that impairs freezing to context in mice by 50% (V. Rau and E.I. Eger II, unpublished data); and (2) Atropine sulfate 50 mg/kg intra-peritoneal (i.p.) which is saturating with respect to behavioral and EEG effects (Buzsáki et al., 1983;Hentschke et al., 2007). Sterile 0.9% saline was used as the vehicle/diluent and was also used for control injections.
For an individual recording session, a mouse was randomly administered drug or saline-control. Only one recording session per day was performed for each mouse. Recording sessions were comprised of three blocks ( Figure 1B). (1) A mouse was placed in a rectangular glass aquarium (15 × 30 cm) with an open top, and a 10-min pre-injection block of baseline EEG/LFP activity was recorded. (2) The mouse was removed from the glass aquarium, either drug or saline was injected, and the mouse was returned to the aquarium. (3) 40 min after midazolam, or 20 min after atropine administration (or these same durations for the respective saline-controls), the mouse was moved to a transparent circular plastic enclosure of diameter 25.5 cm and a post-treatment block of EEG/LFP activity was recorded for 10 min. For midazolam, only 3-min data blocks were analyzed, as this period of time corresponded to the time during which animals explored their new environment in fear conditioning studies.
We used current source density (CSD) analysis to localize the individual gamma sub-band filtered signals and estimate their strength. The CSD was derived from band pass filtered LFP signals using the cubic spline iCSD method introduced by Pettersen et al. (2006) (MATLAB® Toolbox CSDplotter). The amplitudes of the sinks/sources in the CSD signal for individual frequency bands were quantified as the root mean square of the CSD signal (rmsCSD) for each contiguous 1-s segment of filtered LFP data. Three distinct peaks were evident in the spatial distribution of the CSD. The peak observed at the somatic recording site is referred to as S pole , at mid-apical site as M-AD pole and at the distal apical region as D-AD pole . For statistical evaluation, rmsCSD values for a treatment block were normalized to the median value of the pre-injection block. For each frequency band, the normalized rmsCSD values for each peak and for each drug/saline-control treatment block from all animals were pooled and plotted. To illustrate the variability between different data segments, we plotted cumulative frequency or cumulative probability distributions. One-Way ANOVA with Bonferroni posttest (GraphPad Prism® v5) was used to test significance between drug and saline-control for each of the CSD poles, with each data point the average rmsCSD amplitude of 1-s data segments normalized to the median value of the pre-injection recording block.
The strength of cross frequency coupling (CFC) between the phase of the θ-oscillations and the amplitude of the γ slow , γ mid , and γ fast oscillations (i.e., phase-amplitude coupling) were quantified by the modulation index (MI), as described by Tort et al. (2010). The phase-frequency was divided into 18 bins of 20 • . The MI was computed in steps of 2 Hz for phase-frequency (i.e., the frequencies at which the θ-oscillation phase was used for CFC) and 2 Hz for amplitude-frequency (i.e., the frequencies at which the amplitude of the γ-oscillation for each phase-frequency was computed). Since we found that the peak frequency of θ and the maximum θ-γ coupling were both within the 6-10 Hz band, further analysis concentrated on MI values within this range. As the modulation index is a measure of deviation from uniform distribution, the values from the drug treatment blocks were directly compared to the saline-control blocks for each of the frequency bands (θ-γ slow /θ-γ mid /θ-γ fast ) using Two-Way ANOVA with Bonferroni posttest. A minimum of 30 s of data was used to compute CFC. For the 3 min block of midazolam and its salinecontrol, the analyzed segments averaged 82 ± 37 and 129 ± 27 s, and for the 10 min block for atropine and its saline-control, 534 ± 51 and 270 ± 61 s respectively.

IN SILICO
Local field potentials recorded in vivo are subject to low-pass filtering by the brain tissue. Higher frequencies are attenuated more than lower frequencies, as the tissue acts like a "system of coupled RC circuits" (Bédard et al., 2006). To evaluate how this effect, together with the directional voltage attenuation seen in the CA1-PC (Carnevale et al., 1995), would affect the relationship between the input strength and the CSD profile, we simulated oscillatory inputs at three different frequencies, at three different sites corresponding to proposed physiological inputs. We then compared the spatial distribution profiles for derived CSD profiles with our in vivo results.
Simulations were run in the NEURON v7.1 simulation platform, using a CA1 pyramidal cell model embedded in an extracellular matrix. In brief, an "extracellular stimulation and recording" program (Carnevale, 2005) was integrated with a CA1 pyramidal cell model (Poirazi et al., 2003;Carnevale and Hines, 2009) (Senselab-ModelDB accession number 20212). We simulated oscillatory LFPs at 16 recording sites separated by 50 μm along a linear track parallel to the long-axis of the pyramidal neuron. Using an approach that we and others have followed previously (Kopell et al., 2010b;Balakrishnan and Pearce, 2014), inputs were designed as point processes, with time-varying excitatory and inhibitory synaptic inputs simulated as oscillatory conductances, the reversal potentials of which were set to 0 mV or −75 mV respectively ( Figure 1D). Values for conductances were chosen such that they were subthreshold for action potential generation (somatic inhibitory conductance = 0.003 mho/cm 2 ; dendritic excitatory conductance = 0.0004 mho/cm 2 ; dendritic inhibitory conductance = 0.01 mho/cm 2 . Three input locations were selected for independent simulations: (1) Soma input (inhibitory only; compartment = soma[1]); (2) Mid-apical dendritic input (inhibitory or excitatory; compartment = apical_dendrite[68]); and (3) Distal apical dendritic input (inhibitory or excitatory; compartment = apical_dendrite[92]) ( Figure 1D). Three frequencies of sinusoidally varying conductances-30 Hz (∼γ slow ), 59 Hz (∼γ mid ) and 111 Hz (∼γ fast )-were independently simulated at each of the locations. Except for these characteristics of oscillatory inputs, model parameters were the same as those we used previously for our simulations of θ-band oscillations (Balakrishnan and Pearce, 2014). CSD profiles were derived from the 16-channel LFPs, and quantified using root mean square over the last 1750 ms of the 2000 ms simulation (the first 250 ms of data discarded to eliminate fluctuations due to model settling time). The CSD profiles were computed in same fashion as for in vivo data.

CSD SPATIAL PROFILES
In vivo recordings of LFPs obtained during open field exploration showed that higher frequency γ-oscillations were present together with lower frequency θ-oscillations ( Figure 1C, Supplementary Figure 1), as other investigators have reported previously (Tort et al., 2008;Colgin et al., 2009;Belluscio et al., 2012). We examined the spatial pattern of the current sinks and sources that produce each of the three higher frequency bands using CSD analysis.  Figure 2C) activity. Distinct peaks, corresponding to individual poles of oscillating dipoles, were especially evident at sites positioned within the mid-apical dendrite (M-AD, recording sites 7-9) and distal apical dendrite layers (D-AD, recording sites 12-15), but also near the soma (S, recording sites 2-3) in this example.
To visualize the pattern and variation in the CSD poles over the full analysis period, we calculated the root mean square of the CSD signal (rmsCSD) for each 1 s data segment, and plotted the combined data for each of the frequency bands (Figures 2D-F). Again, it was evident that the D-AD pole and M-AD pole amplitudes were substantially larger than the S pole at all frequencies.
In addition, compared to the D-AD pole and M-AD pole , which were consistently present in all recordings in all animals, the S pole showed substantial variability between animals; it was not discernible in the γ slow band in two mice, or in the γ mid band in four mice, and it was completely absent in the γ fast oscillations in all mice.
This analysis thus showed that γ-oscillations are driven by current sources and sinks concentrated in three distinct regions-the soma, the mid-apical dendrite, and the distal apical dendritebut that the two dendritic inputs predominate for all three sub-bands.

EFFECT OF MIDAZOLAM ON rmsCSD AMPLITUDES
Previous studies have implicated GABAergic synaptic transmission in the generation of γ-oscillations expressed by hippocampal an exploring mouse, for γ slow , γ mid , and γ fast oscillations respectively. In all bands the amplitude of the CSD was highest at the fissure (electrodes 13-14), with a large peak also present in the mid-apical region. The relative amplitude of the somatic peak decreased as the frequency increased, and was absent for γ fast .
slices (Traub et al., , 2003. To test the effect of the GABA A receptor modulator midazolam on γ-oscillations in vivo, we administered a sub-hypnotic but amnestic dose and assessed effects on the different sub-bands by comparing the cumulative probability distributions of rmsCSD peaks for γ slow , γ mid and γ fast oscillations (Figure 3). For γ slow , midazolam increased the amplitude of the S pole ( Figure 3A: midazolam/saline-control, n = 402/367 data points from 5/4 mice p < 0.001 One-Way ANOVA) and the D-AD pole ( Figure 3C: midazolam/salinecontrol, n = 458/503 data points from 7/6 mice p < 0.001 One-Way ANOVA), but had no effect on the M-AD pole ( Figure 3B: midazolam/saline-control, n = 458/503 data points from 7/6 mice). In contrast, for γ mid and γ fast oscillations, midazolam significantly decreased the amplitude of rmsCSD at all locations (γ mid -Figures 3D-F: midazolam/saline-control, S pole : n = 266/234 data points from 3/3 mice p < 0.001; M-AD and D-AD pole : n = 508/595 data points from 7/6 mice p < 0.001, γ fast -Figures 3G,H: midazolam/salinecontrol, n = 535(M-AD pole ), 531(D-AD pole )/648(M-AD pole ), 647(D-AD pole ) data points from 7/6 mice p < 0.001 One-Way ANOVA). This pattern of effects-an increase in the slowest component primarily at the soma, and decreases in the faster two components at all locations-is consistent with the participation of GABAergic inhibition in the generation or expression of all three sub-bands. The location-and frequency-dependent pattern of changes suggests that oscillations in the different frequency bands arise through distinct cellular mechanisms, with midazolam differentially modulating their strength.

ATROPINE INCREASED rmsCSD AMPLITUDES FOR ALL FREQUENCY BANDS AT ALL LOCATIONS
Like midazolam, atropine impairs hippocampal function. However, it does so through a distinct molecular mechanism-by antagonizing mAChRs. To further explore the pharmacologic sensitivities of the different sub-bands, we administered atropine and measured its effect γ slow , γ mid , and γ fast oscillations. Figure 4 summarizes the effects of atropine (orange) compared to salinecontrols (gray) on the rmsCSD peaks of γ slow , γ mid , and γ fast oscillation activity in the CA1 region. Unlike midazolam, which had different effects on different peaks, atropine increased the amplitudes of γ slow , γ mid , and γ fast oscillations at all locations.

CROSS-FREQUENCY COUPLING PROFILES
Previous studies have found that the different sub-bands differ in cross-frequency coupling with the θ-oscillation (Tort et al., 2008Colgin et al., 2009;Belluscio et al., 2012;Scheffer-Teixeira et al., 2012). We explored layer-specific characteristics of CFC by measuring the amplitude of γ-oscillations as a function of the phase of the local θ-oscillation. Figure 5 shows an example from one animal during the baselinerecording period, prior to injecting a drug or saline. Here, the amplitude of the γ-oscillation is plotted as a function of the phase of the θ-oscillation, for all three sub-bands at each of three recording sites (Figures 5A,D for γ mid than for γ slow or γ fast oscillations at all locations. This pattern was observed consistently across animals; on average, the amplitude modulation of γ mid was ∼6 times that of γ slow and ∼4 times that of the γ fast oscillation, as assessed by comparing the modulation index (MI) for each of the sub-bands at D-AD. Although the rmsCSD amplitudes of γ slow and γ mid were higher in the M-AD than at the soma (Figures 2D,E), the phase-amplitude coupling was lowest in the M-AD, confirming that MI values are independent of CSD amplitude (Tort et al., 2010). A second difference between the CFC of the three sub-bands was evident in the timing of their phase-amplitude coupling with the local θ-oscillation. In all cases, the phase at which amplitude modulation was the greatest was shifted by 180 • across layers, with the peak at ∼240 • at the soma coinciding with a trough at the D-AD (Figures 5D,G,J,F,I,L). This pattern matches the 180 • phase reversal of the θ-filtered LFP in the CA1 region (Green et al., 1960;Winson, 1974;Bland et al., 1975;Balakrishnan and Pearce, 2014). However, the timing of the CFC in the midapical region differed: for γ mid the phase of maximum amplitude modulation matched that of the soma, whereas the phase of maximum amplitude modulation of γ slow matched the timing of the D-AD.
Taken together, our findings that the different sub-bands differed in the strength of their cross frequency coupling, and in the timing of the coupling in the mid-apical region, support a model in which separate oscillatory circuits generate γ slow , γ mid , and γ fast oscillations; a single broadband oscillator would be expected to be more uniformly coupled in strength and timing to the phase of the modulating θ-oscillation.

EFFECT OF MIDAZOLAM AND ATROPINE ON CFC
In addition to characterizing their CFCs under drug-free conditions, we examined the effects of midazolam and atropine on the three sub-bands. An example of the effect of midazolam at the somatic recording site in an individual animal is shown in Figures 6A-F, comparing oscillations following administration of midazolam (Figures 6B,D,F) vs. saline (Figures 6A,C,E). Average effects on MI in all animals are shown in Figures 6G,H for midazolam and its saline-control, and Figure 7 for atropine and its saline-control. Following administration of midazolam, both the peak frequency of the modulating θ-oscillation and the peak frequency of the modulated γ-oscillation shifted to lower values (Figures 6A,B), so that the maximum modulation occurred within the γ slow frequency band rather than γ mid , as seen during the pre-injection baseline ( Figure 5) and following saline injection (Figures 6A,C,E).
Comparing the modulation index values for all animals at the somatic recording site, we found a significant increase in the MI for θ -γ slow CFC ( Figure 6G: midazolam/saline-control, n = 5/5 p < 0.01, Two-Way ANOVA) but no significant change at the M-AD and D-AD recording sites ( Figure 6G: midazolam/salinecontrol, M-AD: n = 3/5, D-AD: n = 6/6 p > 0.05, Two-Way ANOVA). In contrast, midazolam significantly decreased the MI of θ-γ mid CFC at the somatic recording site (Figure 6H: midazolam/saline-control, n = 5/5 p < 0.01, Two-Way ANOVA) and again had no significant effects at the M-AD and D-AD recording sites ( Figure 6H: midazolam/saline-control, M-AD: n = 3/5, D-AD: n = 6/6 p > 0.05, Two-Way ANOVA). Atropine had no significant effects on the CFCs for either θγ slow (Figure 7A) or θ -γ mid (Figure 7B) at any of the recording sites ( Figures 7A,B: atropine/saline-control, n = 4/5 p > 0.05, Two-Way ANOVA). For the θ-γ fast CFC, only modulation of the D-AD was large enough to analyze, and there were no effects on MI for either drug (midazolam/saline-control, n = 6/6 p > 0.05, Two-Way ANOVA; atropine/saline-control, n = 4/5 p > 0.05, Two-Way ANOVA) (data not shown). Neither midazolam nor atropine altered the phase of θ at which the amplitude peaks appeared for any of the sub-bands. The lack of effect of atropine on CFC indicates that although muscarinic receptors modulate the circuits that generate γ-oscillations, they do not modulate the components that underlie phase-amplitude coupling within the CA1 network. By contrast, the opposite effects of midazolam on θ-γ slow and θ-γ mid CFC is consistent with coupling between inhibitory circuits as the mechanism for θ-γ coupling, and it provides further evidence that the underlying circuitry differs for these two oscillations.

COMPARTMENTAL MODELING
The in vivo rmsCSD spatial profiles showed that the M-AD pole and D-AD pole were larger than the D-AD pole , for all three subbands (Figure 2). These CSD peaks could be produced by local active synaptic inputs, or they could reflect the locations of passive returns from spatially distant inputs, as occurs for θfrequency oscillations (Balakrishnan and Pearce, 2014). We used computer simulations to determine how single inputs at different frequencies (30, 59, 111 Hz) and locations (soma, midapical, distal-apical dendrite) would be expected to influence the rmsCSD profiles in the presence of frequency-dependent attenuation (Figure 8). Excitatory and inhibitory inputs were simulated at each location as oscillatory conductance changes. The CSD was then derived from the extracellular LFP, and its spatial spread and amplitude pattern were evaluated. The patterns produced by oscillatory mid-apical dendrite inputs at 30 and 59 Hz matched the CSD spatial profiles that we observed in vivo for γ slow and γ mid (Figures 8D,E), with large amplitude oscillations present at both M-AD and D-AD sites and a smaller amplitude oscillation at the soma. The pattern produced by an oscillatory input at the D-AD matched the in vivo profile of the γ fast oscillation, which lacked a peak at the soma (c. f. Figures 2, 8). However, simulating an oscillatory input at the soma produced a large-amplitude oscillation at the soma, a smaller peak in the M-AD region, and no peak at the D-AD-a pattern that did not match any profile recorded in vivo.
Since changes in membrane shunting have been invoked for many drug effects, including midazolam, we also considered how a midazolam-induced increase in membrane conductance might influence local CSD amplitude. Contrary to our physiological results, an increase in membrane leak conductance led to a decrease in CSD amplitudes at the soma, both for oblique excitatory as well as somatic inhibitory input (data not shown). Therefore, we conclude that midazolam effects cannot be simply a change in passive propagation due to membrane shunting effects.
These modeling results thus showed that the spatiotemporal characteristics we observed for γ slow and γ mid oscillations matched the expected patterns produced by M-AD inputs, and that the profile for γ fast oscillations matched the D-AD input pattern. We expect that combinations of oscillatory drivers would generate profiles with mixed characteristics, but we did not explore CSD profiles produced by multiple inputs in any detail.

DISCUSSION
In this study we sought evidence for local differences in the expression or generation of γ slow (30-45 Hz), γ mid (50-90 Hz), and γ fast (90-170 Hz) oscillations in the CA1 region of the hippocampus, based on their spatiotemporal profiles, their cross frequency coupling with the local θ-oscillation, and their pharmacological modulation. CSD analysis revealed prominent peaks in the mid-and distal-apical dendrites for all three subbands, and a smaller peak at the soma that was variably present for the γ slow and γ mid , but absent for γ fast oscillations. Differences in the strength and timing of θ-γ slow and θ-γ mid cross frequency coupling, and a lack of coupling at the soma and mid-apical region for γ fast oscillations, suggested that separate biophysical processes generate the three sub-bands. The lamina-and frequency specific modulation by midazolam of both CSD amplitude and cross-frequency coupling provided further evidence for separate underlying generator circuits. Based on these results, we conclude that distinct local circuits generate γ slow , γ mid , and γ fast oscillations in the CA1 region of the hippocampus.

MULTIPLE γ-OSCILLATIONS IN THE HIPPOCAMPUS
Investigations conducted in recent years by a number of laboratories have revealed that "γ-oscillations" can be divided into several distinct components. As new information has emerged, the nomenclature and definition of gamma sub-bands has varied. Based upon their coherence and phase locking with cells in other brain regions, "slow" (25-50 Hz) and "fast" (50-150 Hz) γ-oscillations in the CA1 region were described initially, with "slow gamma" driven or entrained by input from the CA3 region, and "fast gamma" by input from the ECtx (Bragin et al., 1995;Charpak et al., 1995;Middleton et al., 2008;Colgin et al., 2009). Detailed characterization of θ-γ cross frequency coupling revealed distinct "high frequency oscillations" (HFOs-also referred to as ε-oscillations) in the 110-160 Hz range, and indicated that γ-oscillations in the lower frequency range can be separated into "γ-slow" (30-60 Hz) and "γ-mid" (60-100 Hz) oscillations Tort et al., 2012). Studies of θ-γ phase-phase coupling, another form of CFC, revealed additional frequency-dependent differences (Belluscio et al., 2012). Our present results further support the separation of "γ-oscillations" into three distinct sub-bands generated by different biophysical processes. In keeping with the nomenclature of Belluscio et al. (2012), we refer to them here as γ slow , γ mid , and γ fast oscillations.

MECHANISM OF γ-OSCILLATION GENERATION-CONTRIBUTION OF EXTERNAL INPUTS
Oscillations in the CA1 region at frequencies below ∼70 Hz were proposed to be entrained or driven by γ-oscillations in the CA3 region (Middleton et al., 2008;Colgin et al., 2009). Our present results support this hypothesis, as patterns of CSD spatial profiles derived from physiological recordings in vivo (Figure 2) matched computer simulations with inputs in the mid-apical region (Figures 8D-F). Thus, active inputs in the mid-apical region appear to contribute to γ slow and γ mid oscillations. Interestingly, this situation is unlike that for θ-oscillations, in which case a passive mid-apical peak is created by the overlapping return currents from active inputs located at the soma and distal apical dendrites (Balakrishnan and Pearce, 2014).
By contrast, the maximal current sources and sinks driving γ fast oscillations were located in stratum lacunosum-moleculare. θ-γ fast CFC was also maximal in this layer. In keeping with prior suggestions that input from the ECtx drives or entrains γ-oscillations at higher frequencies, we found a correspondence between physiological results showing that the S pole is absent in vivo for γ fast (Figure 2F), as it is in computer simulations with distal apical dendritic input (Figures 8G-I). However, in some prior studies, γ fast oscillations were reported to be concentrated in the superficial layers, i.e., stratum oriens and stratum pyramidale (Belluscio et al., 2012;Scheffer-Teixeira et al., 2012;Tort et al., 2012). Our present finding that γ fast oscillations are maximal in stratum lacunosum-moleculare, in register with afferent input from ECtx, thus supports the association between γ fast in CA1 and ECtx (Lasztóczi and Klausberger, 2014).

MODULATION OF γ-OSCILLATIONS BY MIDAZOLAM
Since we used an amplitude-independent measure of cross frequency coupling (Tort et al., 2010), the differences between sub-bands in CFC characteristics and drug responses (Figures 5, 6G,H) provided additional evidence of different intrinsic and network mechanisms. Were there a single broadband oscillator that was simply shifted to lower frequencies by midazolam, power might have appeared to go down in the higher frequency band, and up in the lower frequency band, but these changes would not have altered the characteristics of CFCs within those bands. We found the most prominent differences in stratum pyramidale, where midazolam increased γ slow CSD amplitude and θ-γ slow CFC, but decreased γ mid CSD amplitude and θ-γ mid CFC (Figures 3A,D, 6). The γ mid and γ fast oscillations responded similarly to midazolam as well as atropine; however their baseline CSD and CFC characteristics differed, as described above.
One of the models proposed for generation of gamma oscillations involves an interaction between excitatory neurons and inhibitory interneurons Kopell et al., 2010a). Previous research has shown that PV-BCs resonate at γ-frequency (Pike et al., 2000) and impose a strong inhibition at the soma of CA1-PCs (Freund and Buzsáki, 1996;Buzsáki, 2002). Also, PV-BCs fire at ∼35 Hz (Campanac et al., 2013), and PV-BCs regulate perisomatic γ-oscillations without any effect on γ-oscillations in dendritic regions (Lasztóczi and Klausberger, 2014). Taken together, these findings suggest that the contribution of PV-BCs is restricted to the γ slow frequency band. These BC's in turn are known to receive inhibitory input from O-LM interneurons (Elfant et al., 2008). Hence the effects of midazolam on the CSD amplitude of γ slow , and on θ-γ slow CFC at stratum pyramidale, likely arise from its modulation of inhibitory synapses embedded in a network involving interactions between CA1-PCs, basket cells, and O-LM interneurons, and not exclusively on CA1-PC GABA A Rs. Moreover, the differential distribution of GABA A Rs composed of different subunits on CA1-PCs and interneurons likely influences midazolam's effects as well (Nusser et al., 1996;Somogyi et al., 1996;Brünig et al., 2002;Serwanski et al., 2006;Salesse et al., 2011).
In contrast to γ slow , midazolam decreased the CSD amplitude of γ mid at all locations, and decreased the θ-γ mid CFC at the soma. Since midazolam enhances GABAergic inhibition (Sigel and Buhr, 1997;Rudolph et al., 1999;Rudolph and Möhler, 2004), and inhibition is thought to contribute critically to the generation of γ-oscillations (Traub et al., , 2003Jackson et al., 2011;Chen et al., 2014), this finding is therefore somewhat counterintuitive. We suggest two possible explanations. First, γ mid might be generated by temporally and spatially overlapping excitatory and inhibitory dendritic inputs. In this case, enhanced inhibition may offset excitation and thereby reduce oscillation amplitude, as suggested for θ-oscillations based on computer simulations (Balakrishnan and Pearce, 2014). Indeed, a recent study described interneurons in the stratum lacunosum moleculare coupled to 60-100 Hz oscillations of urethane-anesthetized rats, and slightly higher frequencies in awake head-fixed mice, recorded exclusively from the apical dendrite region (Lasztóczi and Klausberger, 2014). Midazolam could increase the strength of GABAergic input onto pyramidal cells from such interneurons, offsetting excitatory components of γ mid and/or γ fast oscillations in the CA1 region. Alternatively, midazolam may globally suppress oscillations in the inhibitory networks that generate γ mid and γ fast oscillations, so that even if inhibitory input onto pyramidal neurons is enhanced, the net effect is a reduced inhibitory current in CA1-PCs. It is also possible that midazolam differentially affects the CA1 local circuitry by prolonging the GABA IPSC decay constant such that it is not able to resonate at the higher frequency inputs. Our present results do not allow us to distinguish between these possible explanations.
The midazolam-induced decrease in θ-γ mid CFC matches the change produced by deletion of the γ2 subunit in fast-spiking interneurons (Wulff et al., 2009), though the exact nature of the interneuronal and CA1-PC connections that participate in the θγ mid circuit modulated by midazolam remain undefined. Unlike its effect on CFC at the somatic recording site, midazolam did not significantly alter the CFC at the distal apical dendrite for either γ slow or γ mid . This finding provides a further indication that the different sub-bands are produced by distinct biophysical mechanisms.
Since inhibitory circuitry is thought to contribute to generation of oscillations, and oscillations are important for memory function, it is curious that at least some types of oscillations are enhanced by midazolam, a drug that disrupts memory. How can these observations be reconciled? The answer may lie in the need for precise timing. Phase-amplitude coupling, which is the most widely recognized and studied form of CFC, is thought to arise by the interaction of slow spiking interneurons (e.g., O-LM interneurons) with fast spiking interneurons (e.g., basket cells) (White et al., 2000;Rotstein et al., 2005;Tort et al., 2007). This coupling was proposed to be a mechanism by which information processing and encoding occurs in neural assemblies (Lisman and Idiart, 1995;Lakatos et al., 2005;Senior et al., 2008), and to be important for memory in humans (Axmacher et al., 2010), as it predicts learning (Tort et al., 2009) and overall memory performance (Shirvalkar et al., 2010;Friese et al., 2013). We found that when midazolam was administered at the ED 50 dose for amnesia in mice, there were changes in CFC of θ-γ slow and θ-γ mid at stratum pyramidale. These results indicate that midazolam and other drugs that produce a state of "conscious amnesia" might impair memory by disrupting essential timing mechanisms rather than suppressing overall circuit activity.

MODULATION OF γ-OSCILLATIONS BY ATROPINE
Cholinergic agonists such as carbachol are used to induce gamma oscillations in hippocampal slices, and atropine blocks these in vitro oscillations (Fellous and Sejnowski, 2000;Traub et al., 2000;Whittington et al., 2000;Mann et al., 2005). However we found that atropine increased the amplitude of γ-oscillations at all locations in exploring mice. As somatic inputs are predominantly (or exclusively) inhibitory, the increase in CSD amplitude at the soma indicates that atropine increased inhibition, at least for γ slow . The increased inhibition at the soma is likely a result of increased GABA release by the PV-BC's that targets the soma of CA1-PC, mediated by atropine block of the m2 subunit of mAChRs, which are expressed presynaptically on the axon terminals of PV-BC's (Levey et al., 1995;Hájos et al., 1997;Rouse et al., 2000).
In the dendrites, both excitatory and inhibitory inputs are present, so atropine may have altered either inhibition or excitation. We are unable to ascribe changes to specific mechanisms, but the possibilities include blockade of acetylcholine-mediated presynaptic inhibition on glutamatergic and GABAergic synapses (Valentino and Dingledine, 1981;Levey et al., 1995;Qian and Saggau, 1997;Leung and Péloquin, 2010), or reduced feed forward inhibition from CA1 interneurons excited by γ-frequency input from CA3 (Zemankovics et al., 2013). The lack of significant effect on the CFC by atropine suggests that mAChR's are not placed in the crucial position where the slow (θ) and fast (γ) networks interact. The effect of atropine could also be a reflection of changes in the ECtx transmitted to CA1 by changing input strength. However, cholinergic blockade was found to decrease the power of γ-oscillations in medial entorhinal cortex (Newman et al., 2013), making this explanation less likely.

IMPLICATIONS
Multiple mechanisms have been proposed for generation of gamma oscillations, including interneuron network gamma (ING), pyramidal-interneuron network gamma (PING) and persistent gamma (White et al., 2000;Whittington et al., 2000;Kopell et al., 2010b). Similarly, a number of different models for θ-γ cross frequency coupling have been proposed (Wulff et al., 2009;Kopell et al., 2010a;Buzsáki and Wang, 2012;Buzsáki and Watson, 2012;Tort et al., 2012;Lisman and Jensen, 2013). Given the evidence that multiple γ-oscillations co-exist even within a single brain region, and that they are produced by distinct biophysical and circuit mechanisms, it is possible that each of the different models applies to a specific subset of oscillators, and that there is not a single universally applicable mechanism. Rather, there may be a degree of overlap between these different models acting in tandem to bring about the features of LFPs seen in vivo.