Thèse présentée pour obtenir le grade de docteur

Sorbonne Université
❧
Institut des Neurosciences Paris-Saclay
École doctorale ED 158 - Cerveau, cognition, comportement

Discipline : Neurosciences

Responsiveness and dynamics of

asynchronous states and Gamma oscillations
in cerebral cortex

Par : Eduarda DEMORI SUSIN

Sous la direction de Alain Destexhe, Directeur de Recherche au

CNRS, Directeur adjoint de NeuroPSI, Université Paris-Saclay

Membres du jury:
Rapporteur : Laure Buhry, Université de Lorraine.

Rapporteur : María Victoria Sánchez-Vives, Universitat de Barcelona

Examinateur : Diego Contreras, University of Pennsylvania, School of Medicine

Examinateur : Demian Battaglia, Aix-Marseille l’Université

Président du jury : Stéphane Charpier, Sorbonne Université

Directeur de Thèse : Alain Destexhe, Université Paris-Saclay

Date de soutenance : 24 novembre 2021

À minha querida mãe e a meu querido pai que partiu deste mundo sem poder ver este
trabalho concluído.
 Résumé

Comprendre la cognition et ses mécanismes sous-jacents (comment l’information est codée,

décodée et traitée) est l’un des objectifs des neurosciences. Certaines observations in-
diquent l’implication des oscillations Gamma (30-90 Hz) dans le traitement de l’information
en étant associées à différentes fonctions cognitives de haut niveau, telles que la mémoire,
la perception, l’attention, l’éveil focalisé et la prédiction. Néanmoins, malgré ces corréla-
tions, le rôle exact des oscillations Gamma est encore débattu. Afin d’étudier les principes
neuronaux de la transmission de l’information dans le cerveau, nous avons profité de
l’analyse des données électrophysiologiques acquises chez l’homme. Cette analyse a indiqué
comment des neurones individuels participent aux rythmes Gamma, nous permettant
d’utiliser ces caractéristiques pour construire et contraindre des modèles de réseaux infor-
matiques. Nous avons construit trois modèles de réseaux générant des rythmes Gamma
par trois mécanismes différents : soit par l’interaction exclusive entre neurones inhibiteurs
[Interneuron Gamma (ING)], soit par l’interaction de neurones inhibiteurs et excitateurs
via Pyramidal-Interneuron Gamma (PING) ou via un mécanisme lié à la présence d’un
type particulier de neurone excitateur - appelée Chattering cell en anglais, le Chattering
Induced Gamma (CHING). Ces modèles ont été explorés et comparés à des états électro-
physiologiques asynchrones, afin de déterminer quelles sont les conséquences de chaque
mécanisme oscillatoire dans la manière dont l’information est traitée par le réseau. La
transmission de l’information a été évaluée à travers de la mesure de la réactivité du
réseau (responsiveness), qui a été acquise dans des conditions saines et dans des con-
ditions pathologiques, imitant des états hallucinatoires ou semblables à des conditions
observées dans des patients schizophréniques. Ce travail apporte alors une compréhension
du rôle des oscillations Gamma non seulement dans des conditions saines, mais aussi dans
des conditions pathologiques.

v
 Abstract

Understanding cognition and its underlying mechanisms (how information is encoded,

decoded and processed) is one of the purposes of neuroscience. Gamma oscillations (30-90
Hz) are believed to be involved in information processing, and have been associated to
different high-level cognitive functions, such as memory, perception, attention, focused
arousal and prediction. Regardless of these observations the exact role of these oscilla-
tions is still debated. In order to study the neuronal principles of information transmission
in the brain, we took advantage of multi-unit electrophysiological data acquired in hu-
man. The analysis of this data indicated how individual neurons participate to Gamma
rhythms, and these features were used to build and constrain computational network
models. We built three network models generating Gamma rhythms by three different
mechanisms: either by the exclusive interaction between inhibitory neurons [Interneuron
Gamma (ING)] or by the interaction of inhibitory and excitatory neurons via Pyramidal-
Interneuron Gamma (PING) or via Chattering Induced Gamma (CHING). These models
were explored and compared to asynchronous states, in order to determine what are the
consequences of each oscillatory mechanism in the way information is processed by the
network. Information transmission was assessed by means of the measurement of net-
work responsiveness, which was acquired in healthy and in pathological states mimicking
hallucinatory states or schizophrenia disease. This work provides insight into the role of
Gamma oscillations during healthy and pathological conditions.

vii
 Remerciements

Ce travail n’aurait pas été possible sans le soutien de l’École des Neurosciences de Paris
(ENP), la Fondation pour la Recherche Médicale (FRM) et la Fondation des Treilles, qui
m’ont permis, grâce à une allocation de recherches et diverses aides financières, de me
consacrer sereinement à l’élaboration de ma thèse.

Je remercie Alain Destexhe, Directeur de Recherche au CNRS, qui m’a encadré tout au
long de cette thèse et qui m’a fait partager ses brillantes intuitions. Qu’il soit aussi remercié
pour sa gentillesse, sa disponibilité permanente et pour le soutien qu’il m’a prodigué.

J’adresse tous mes remerciements au président du comité de cette thèse Stéphane Charpier
(Professeur Sorbonne Université et Chef d’équipe à l’Institut du Cerveau et de la Moelle
épinière), à Laure Buhry (Maître de Conférence - HDR et Professeur associé à l’Institut
des sciences du digital Management & Cognition, Université de Lorraine) et María Vic-
toria Sánchez-Vives, (Professeur chercheur à l’Institut d’Investigacions Biomèdiques
August Pi i Sunyer, Universitat de Barcelona), ainsi qu’à Diego Contreras (Professeur
chercheur au département de Neurosciences à University of Pennsylvania, School of Medicine)
et Demian Battaglia (Chercheur au CNRS à l’Institut de Neuro-sciences des Systèmes,
Aix-Marseille l’Université), de l’honneur qu’ils m’ont fait en acceptant d’être rapporteurs
et examinateurs de cette thèse.

Je tiens à remercier Mallory Carlu et Damien Depannemaecker, mes collègues de lab-

oratoire, pour leurs amitié, compagnie, soutien, encouragements et discussions éclairées.

J’exprime ma gratitude à Leonardo Gregory Brunnet, Professor titular de l’Universidade

Federal do Rio Grande do Sul, à Rita Maria Cunha de Almeida, Professora associada
de l’Universidade Federal do Rio Grande do Sul et à Sandra Denise Prado, Professora
associada de l’Universidade Federal do Rio Grande do Sul, pour leur accompagnement
pendant ma Licence et mon Master au Brésil. Ils ont toujours été là pour me soutenir et
c’est grâce à eux que j’ai découvert la recherche.

Je remercie mes amis proches que m’ont apporté le soutien émotionnel pour franchir les
difficultés et toutes les personnes avec qui j’ai partagé mes études, notamment ces années
de thèse.

ix
Remerciements

Finalement, il m’est impossible d’oublier le soutien inconditionnel de ma mère, Luciana

Isabel Demori, de mon père, Enelsir Bernart, de ma marraine Eliane Delegrave et
de mon compagnon Siegfried Nebel. Cette thèse est aussi à vous.

x
 Summary

Résumé v

Abstract vii

Remerciements ix

I Introduction 1

Key Biological Concepts 3

The Brain, Neurons and Synapses . . . . . . . . . . . . . . . . . . . . . . . . . . 3
Chemical Synapses, Neuro-transmitters and Receptors . . . . . . . . . . . . . . 4
Neuronal Heterogeneity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
Cortical Layers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
Brain States: States of Consciousness . . . . . . . . . . . . . . . . . . . . . . . . 8
Brain Rhythms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10

Asynchronous and Irregular States 15

Gamma Oscillations 17
Mechanisms of Generation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17

Gamma Oscillations, NMDA receptors and Schizophrenia 21

NMDA receptor antagonists in sub-anesthetics doses . . . . . . . . . . . . . . . 22
Gamma Oscillations in Schizophrenic Patients . . . . . . . . . . . . . . . . . . . 23
Gamma oscillations, psychotic states and NMDAR hypofunction . . . . . . . . . 24

Computational Models 27
Historical Perspective . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
Individual neuron models . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
Network models and Oscillations . . . . . . . . . . . . . . . . . . . . . . . 28
Gamma oscillations in theoretical models . . . . . . . . . . . . . . . . . . . . . . 29
ING Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
PING Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
CHING Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31

Human recordings Analysis 33

xi
SUMMARY

Developed Models 39

II Research Articles 45

1 Cellular correlates of wakefulness and slow-wave sleep: evidence for a

key role of inhibition 47

2 Integration, coincidence detection and resonance in networks of spik-

ing neurons expressing gamma oscillations and asynchronous states 55

3 A network model of the modulation of Gamma oscillations by NMDA

receptors in cerebral cortex 89

III Discussion 101

IV Complement Research Article 107

4 Bridging Single Neuron Dynamics to Global Brain States 109

xii
 Part I

Introduction

1
 Key Biological Concepts

The Brain, Neurons and Synapses

The brain1 is the center of the nervous system and is the most complex organ in all
vertebrates. In all mammals, the outer layer of the brain is denominated as the cerebral
cortex. In humans, it can be divided into four distinct lobes: parietal, occipital, frontal
and temporal. Each region of the cerebral cortex has its specificities: different cellular
composition with different patterns of connectivity.
The cerebral cortex is composed by two main categories of cells, the glia cells (required
for structural stabilization and energy supply) and another very heterogeneous class of
cells called neurons. Even though recent works have shown the possible implication of glia
cells in information processing [1], the main elementary computing units of the nervous
system are thought to be the neurons.
Neurons, like all livings cells, present an electric membrane potential, due to the dif-
ference between positive and negative charges in the intra- and extra-cellular domain.
On the other hand, unlike most cells, neurons are excitable. This means that the electric
membrane potential of neurons is capable of fluctuating and, as a consequence, generating
and transmitting electrical and chemical stimuli to other cells.
In spite of the large variety of neuronal shapes, a typical neuron can be divided into
three structural parts: soma, dendrites and axon. While the dendrites and the axon can be
seen as input and output "devices" of the neuron, the soma2 can be seen as the "information
processing center" of the cell. This information processing happens in a way that if the total
input leads the membrane potential to exceed a certain threshold, then an output signal
is generated and transmitted to other neurons by the axon. When this signal is generated,
we say that the neuron spiked and that an Action Potential 3 was produced. The place
where this signal exchange takes place is called synapse. Neurons which are sending and
receiving this signal are respectively called presynaptic and postsynaptic neurons. This
process can happen in different ways depending on the nature of the synapse: electrical
1
The brain is the ensemble of three structures: the cerebrum, the cerebellum and the brain stem.
2
The soma, or cell body, is the part of the cell where the genetic information is stored and where most
of the physiological processes take place.
3
The Action Potential is a rapid variation in the neuronal membrane potential, composed of a fast
period of depolarization (due to the opening of N a+ channels), and a slow period of repolarization and
hyperpolarization (due to inactivation of N a+ channels and opening of K + channels). It is known as
the Absolute Refractory Period, the period of time that covers the effects of the increase in the K +
conductivity (with respect to the resting state) and the effects of the residual inactivation of the N a+
channels.

3
 A B C

Figure 1: The central nervous system parts and its units. A) Anatomical parts
of the central nervous system. It can be divided into 7 main parts (as indicated in the
figure). The brain is the ensemble of three structures: the cerebrum, the cerebellum and
the brain stem. The Cerebrum is composed by the cerebral cortex and several subcortical
structures (the hippocampus, the basal ganglia, and the olfactory bulb). The Diencephalon
is composed by the thalamus, the hypothalamus, the epithalamus and the subthalamus.
B) Anatomical division of the cerebral cortex into 4 lobes. C) General structure of a
neuron. Figures A, B and C were adapted from [2] (pages 9 and 25).

or chemical.
In chemical synapses, presynaptic and postsynaptic neurons are completely separated
from each other in space. On the other hand, in electrical synapses, pre- and postsynaptic
cells are directly connected (sharing their cytoplasm) by means of special channels called
gap-junctions.
Gap-junctions have been proposed to contribute to brain rhythms both experimentally
and theoretically [3, 4, 5, 6, 7]. However in the network models developed in this thesis,
we decided to ignore this level of description, and only included chemical synapses in our
models.

Chemical Synapses, Neuro-transmitters and Recep-

tors
At chemical synapses, the synaptic transmission happens by means of vesicles of small
organic molecules, called neurotransmitters, released by the presynaptic cell, due to the
rise of Ca2+ concentration within the synaptic terminal, following an Action Potential.
These vesicles diffuse in the extracellular medium between the pre- and the postsynaptic
cells, the synaptic cleft, and bind to the receptors on the postsynaptic neuron. This binding
leads to the opening of ion channels at the postsynaptic neuron, generating a change in its
membrane potential (see Figure 2). All these processes generate a time difference between

4
the presynaptic spike and the postsynaptic response. This time difference is called synaptic
delay.

A B

Figure 2: Synaptic transmission at chemical synapses. A) Membrane potential vari-

ation during an Action Potential on the presynaptic cell. B) Succession of events com-
posing the chemical synaptic transmission: the action potential arriving at the terminal
of a presynaptic axon causes voltage-gated Ca2+ channels at the presynaptic terminal
to open (left panel). This opening generates an intracellular high concentration of Ca2+
in the presynaptic terminal, allowing vesicles containing neurotransmitter to fuse with
the presynaptic cell membrane and release their contents into the synaptic cleft (middle
panel). Subsequently, the released neurotransmitter molecules diffuse across the synaptic
cleft and bind specific receptors on the postsynaptic membrane. These receptors cause
ion channels to open (or close), thereby changing the membrane conductance and the
membrane potential of the postsynaptic cell (right panel). C) Postsynaptic potential fluc-
tuation due to the synaptic transmission. This figure was adapted from [2].

Conductance changes by means of chemical synapses can either increase the post-
synaptic membrane potential (excitatory synapse) or decrease it (inhibitory synapses),
depending on the type of neurotransmitter released. Glutamate and γ-aminobutyric acid
(GABA) are respectively the main excitatory and inhibitory neurotransmitters in the
brain.
The glutamate-gated channels conduct both N a+ and K + ions, with nearly equal
permeability. The two major glutamate receptors are α-amino-3-hydroxy-5-methyl-4-iso-
xazolepropionic (AMPA) and n-methyl-d-aspartate (NMDA) receptors. Unlike AMPA,
NMDA channels are additionally highly permeable to Ca2+ , which is crucial for induction
of synaptic plasticity [8]. While the AMPA-mediated currents are fast (0.34-11ms), the
NMDA-mediated are considerably slower (6–350 ms) and present a complex relation with
respect to the membrane potential [9, 10, 11, 12]. Both AMPA and NMDA receptors are
ionotropic, meaning that these channels open directly when glutamate binds to them.
However, the opening of NMDA channels depends not only on the biding of neurotrans-
mitters but also on the membrane voltage. At the resting membrane potential extracellular
M g 2+ ions are tightly bond to NMDA channel pores, blocking ionic current. When the

5
membrane is depolarized (for example, by the opening of AMPA receptor-channels), M g 2+
is expelled by electrostatic repulsion, allowing the flux of ions. Due to this M g 2+ block,
NMDA receptor channels rarely initiate neuronal excitation on their own. The magne-
sium block of the NMDA receptor channel (B) is accurately modeled by the following
phenomenological expression as a function of voltage of the membrane potential (V) [13]
:

1
B(V ) = (1)
1 + exp(−0.062V ).([M g 2+ ]o /3.57)
where [M g 2+ ]o is the external magnesium concentration (1 to 2 mM in physiolog-
ical conditions). NMDA-mediated synapses have been reported to participate in both
thalamo-cortical and intracortical synaptic pathways [14] acting on excitatory and in-
hibitory neurons [15].
The main receptors of the inhibitory neurotransmitters GABA are GABAA and GABAB .
While GABAA receptors are ionotropic, directly opening Cl− channels, GABAB recep-
tors are metabotropic, meaning that they activate second-messenger cascades, opening
indirectly other channels (often K + channels). For this reason the current produced by
GABAA receptor channels are fast, while the ones from GABAB are slow. Most ofthe
fast inhibitory postsynaptic potentials are mediated by GABAA receptors in the central
nervous system. In this thesis only GABAA receptors are used in our models. Figure 3
depicts the representative postsynaptic currents due to GABAA , GABAB , AMPA and
NMDA receptor channels.

Neuronal Heterogeneity
Although all neurons inherit the same genes, depending on the particular cell’s develop-
mental history, only a restricted set of them are expressed. As a consequence, an enormous
variety of their enzymes, structural proteins, membrane constituents, ion channels, and
secretory products exists among neurons [2, 17, 18, 19]. This rich repertoire of channels
allows neurons to generate action potentials with a wide range of shapes (spike waveform),
firing rates and duration [20]. These physiological differences can be used to categorized
neurons into different types [21, 22, 23, 24].
McCormick and collaborators were the first to describe the so-called Regular-Spiking
(RS) and Fast-spiking (FS) neurons. When presented with prolonged stimuli of constant
amplitude, RS neurons exhibit pronounced adaptation of the spike frequency while FS
neurons undergo little or no adaptation [21] (see Figure 4A and 4B). In 1985, McCormick
and colleagues related electrophysiological aspects to morphological and immunocyto-
chemical features, identifying neocortical FS neurons as being GABA-mediated interneu-
rons [21] and RS neurons as being pyramidal neurons. Today, other types of cell are known

6
Figure 3: Postsynaptic currents due to different receptors. A) AMPA-mediated cur-
rents. B) NMDA-mediated currents. C) GABAA -mediated currents. D) GABAB -mediated
currents. For all graphs, the averaged recording of the synaptic current (noisy trace) is
represented with the best fit obtained using detailed kinetic models (continuous trace)
[16]. This figure was adapted from [16]. Original works that provided the experimental
curves are cited in [16].

to also display regular spiking activity. This is the case for the Cholecystokinin-positive
basket cells (GABA-mediated interneuron) [25]. Likewise, two types of GABA-mediated
interneurons are known to display fast spiking activity: Chandelier cells4 and Parvalbumin-
positive (PV) Basket cells [25]. In this thesis three types of cells are used in our models:
Regular Spiking pyramidal cells (RS), Fast Spiking basket cells (FS) and Chattering (Ch)
cells.

Chattering Cells were proposed to be a biophysically distinct class of pyramidal neu-

rons because of its distinguished intrinsic firing properties [27]. These neurons, found in
superficial layers of the cortex, intrinsically generate bursts that repeat with a firing fre-
quency of 20 Hz and upwards due to a suprathreshold depolarizing current injection [27,
28, 29] (see Figure 4C).

4
Chandelier cells are a type of GABA-mediated interneurons that differs from Basket cells both mor-
phologically and functionally. While Basket cells form axosomatic synapses (axons terminating on the
cell bodies of target neurons), Chandelier cells form axo-axonic synapses (axons terminating exclusively
on the axons of target neurons). Interestingly this morphological difference (inputs onto axon initial
segment) proffer this type of cell an excitatory effect on its postsynaptic neurons [26]. Based on the
classification by means of the expression of specific molecular markers, Chandelier cells are also classified
as Parvalbumin-positive (PV) [25].

7
 A B C
Regular Spiking Fast Spiking Chattering

Figure 4: Intrinsic firing patterns of different cell types. A) Extracellular recording

in mouse thalamocortical slice of a Regular Spiking (RS) pyramidal cell (adapted from
[30, 22]). B) Intracellular recording in guinea pig neocortical slices of Fast Spiking (FS)
inhibitory cells (adapted from [21, 22]). C) Intracellular recording in the cat neocortex in
vivo of a Chattering (Ch) neuron (adapted from [27]).

Cortical Layers
The neocortex, external region of cerebral cortex5 , can be morphologically segmented
based on the density of specific pyramidal neurons, as well as by their afferent and efferent
projections [31]. Most of the neocortex contains six layers, numbered from the outer
surface (pia mater) of the cortex to the inner white matter [2]. Layer L4 receives most
of the stimulus coming from the primary sensory thalamic nuclei, together with L5b and
L6. Layers L1 and L5a receive inputs from associative thalamus, while layer L5b send
projections to subcortical areas. The cortical layers communicate with each other in an
intricate way. A scheme of their connections is depicted in Figure 5. In this thesis we are
particularly interested in layers 2/3.
Inhibitory neurons are present in all layers and usually have axons that remain within
the same area where their cell body are [2]. They are believed to receive inputs from
the same sources as the principal cells. It is estimated that inhibitory neurons constitute
around 20% to 25% of the neurons in the neocortex.

Brain States: States of Consciousness

Different brain states can be categorized based on two important concepts: awareness and
arousal 6 [33]. The level of arousal is associated to the global capacity to respond to stimuli.
For instance, clinicians use scoring systems such as the Glasgow Coma Scale [35] to access
the level of arousal in patients. In this protocol, the patient reactions (eye opening, motor
response and verbal response) are measured with respect to censorial stimuli and scored
accordingly. Scores range from 3 to 15, in which 15 indicates a patient fully alert. The
level of awareness, on the other hand, is related to the ability to perceive and interact
5
The neocortex constitutes 90% of the cortex. The remaining 10% is occupied by the allocortex, which
contains the olfactory system and the hippocampus.
6
The concept of arousal is also associated to the nomenclature of wakefulness or vigilance [33, 34]

8
A B

Figure 5: The six layers of neocortex and their connections. A) Visualization (by
means of 3 different types of stain) of neocortical neurons arrangement in 6 different layers.
The Golgi stain unveil a subset of neuronal cell bodies, axons, and dendritic trees, while
the Nissl method allows only the visualization of cell bodies and proximal dendrites. The
Weigert stain method, on the other hand, enable the visualization of myelinated fibers
(like axons). Figure A was adapted from [2] (page 345) and originally extracted from [32].
B) Simplified diagram illustrating localization and input–output connectivity patterns of
excitatory neurons in the 6 cortical layers (adapted from [25] ). WM stands for white
matter.

with the environment in a complex and conscious way. In this perspective, awake states
are seen as the state in which the patient is fully aroused and fully aware. While a patient
in vegetative state is considered to be in a state in which there is some level of arousal,
but no level of awareness. The scheme depicted in Figure 6, illustrates these concepts.
In mammals and in most bird species, sleep is divided in two distinguishable neuro-
physiological states: rapid eye movement (REM) and non-rapid eye movement (NREM)
sleep [36]. REM sleep is characterized by inhibited muscular tone (as measured by elec-
tromyography) and involuntary saccadic eye movements (from where its name is derived).
In humans, NREM sleep can be divided into three gradual sleep depth levels [37]: stage
N1 (transition between wake and sleep), Stage N2 (light sleep) and Stage N3 (deep sleep,
also known as Slow Wave Sleep, SWS). The SWS receives this name because of its char-
acteristic extracellular activity, which exhibits slow oscillations (as will be discussed in
the next section).
In general, the level of arousal increases with the increase of awareness. This is what
is observed when we compare the different stages of NREM sleep with conscious wake-
fulness. However, an interesting exception exists: REM sleep. This sleep stage presents

9
Figure 6: Scheme illustrating different brain states. States of Consciousness are
classified with respect to the level of awareness and to the level of wakefulness (arousal).
Figure adapted from [33]

electrophysiological activity similar to the one observed during Awake states: single units
firing tonically, together with extracellular recordings capturing spontaneous bursts of
low-amplitude oscillations in the range of 15-90 Hz [38, 39, 40, 41, 42]. For this reason,
REM sleep is seen as a state with high awareness but no arousal.
Regardless of the difficulties to rigorously define awareness, some authors tried to quan-
tify it, by means of certain concepts of complexity theory. The Complement Research
Article inserted in the end of this thesis reviews some of them.

Brain Rhythms
The discovering of brain oscillations dates from 1929, when a controversial psychiatrist,
Hans Berger, trying to prove the existence of psychic phenomena, observed electrical
oscillations coming from electrodes placed at the head of one of his patients. This was the
first electroencephalogram (EEG) recording registered [43].
Brain rhythms is the generic therm used to describe oscillatory electro-magnetic phe-
nomena generated by the collective activity of neurons in the brain in a mesoscopic scale
(few millimeters to centimeters). This type of activity can be measured by methods such
as ECoG (electrocorticography), EEG (electroencephalography), LFP (local field poten-
tials) and MEG (magnetoencephalography), with temporal precision varying from the
order of milliseconds to seconds [44, 45]. See Figure 7.
The canonical classification of these oscillations has been defined in the following
frequency bands: Slow-oscillations (0.3-1 Hz), Delta (1–4 Hz), Theta (4–8 Hz), Alpha
(8–12 Hz), Beta (15–30 Hz), Gamma (30–90 Hz), High Gamma (90-140 Hz). In addition,

10
 A B C

C
D

Figure 7: Brain Rhythm Measurements. A) Schematic illustration of different tech-

niques of brain rhythm recordings. B) EEG is the less invasive method beeing recorded on
the scalp surface. It allows stable recordings over time but presents low spatial resolution
(covering a large area) and a low signal-to-noise ratio. C) ECoG is more invasive than
EEG, being recorded intracranially, in the surface of the brain. It allows a high signal-
to-noise ratio and stable recordings over time. D) Multi-electrode arrays (left). The LFP
is measured by means of penetrating electrodes directly in the brain tissue (right). This
technique is highly invasive and doesn’t allow very long recordings since the neuronal en-
vironment deteriorates over time leading to instabilities in the recordings. Its advantage
is that it allows a high signal-to-noise ratio close to the neurons. Some types of arrays
allow not only the measurement of LFP but also unit activity. This is the case of the
experimental data used in this thesis. Figure A was adapted from [44]. Figures B was
adapted from [46] and Figures C, D and E were adapted from [47].

other important rhythms are the Spindles and the Ripples oscillations, which are not
exclusively defined by their frequency band (see below). In principle, all these rhythms
could emerge in the brain at any given time and region, but some correlations between
certain rhythms regions and behavior have been established throughout the last decades.
Bellow we highlight some of these aspects related to each brain rhythm. These correlations
suggest possible mechanistic roles for each oscillation on cognitive phenomena, but the
exact function of these oscillations is still debated.

Brain Rhythms Features:

• Delta and Slow oscillations (SO): Mostly observed during deep non-REM sleep
or during anesthesia. Different evidences suggests that these rhythms, SO (0.3-1 Hz)
and Delta (1-4 Hz), have an important role on the formation and consolidation of
memories acquired during wakefulness [48, 49, 50, 51, 52]. SO and Delta oscillations
have been classically classified as two distinct neurophysiological phenomena with
different spatial and temporal properties [53, 51]. For example, while some obser-
vations indicate that SO are generated exclusively by cortical networks [54, 55, 56],
Delta rhythms generators were found in both cortex and thalamus, with mechanism

11
 of generation specific to each structure [57]. More recent studies indicate that SO
and Delta oscillations have competing roles in sleep-dependent memory consolida-
tion [58].

• Theta : In Hippocampus, it is observed during spatial navigation, and the precise

timing of spikes with respect to the Theta cycle is believed to partially encode spatial
information [59]. In addittion, Theta is also believed to play a role in the formation
and retrieval of episodic and spatial memory [60] in the hippocampus. In cortex,
Theta rhythms have been associated to working memory tasks, both in Human [61]
and in Monkey [62] .

• Alpha : Frequently observed in the occipital lobe, alpha rhythms are associated
with wakeful rest with eyes closed [63].

• Beta : Beta rhythms are more frequently associated to the inhibitory control in
the motor system. They occur during the preparation for the movement ("readiness
state") and cease at the onset of movement execution. Other studies, additionally
indicate that Beta oscillations are more generally involved in sensory-motor integra-
tion and top-down signaling [64].

• Gamma : Gamma oscillations occur in all cortical areas, thalamus and hippocam-
pus [65, 66, 67, 68, 69, 70, 71, 72, 73]. This type of activity has been associated to
different high-level cognitive functions, such as memory [74, 75, 76], perception [77,
78, 79, 80], attention [81, 82, 83, 84], focused arousal [85] and prediction [86].

• High Gamma : The terms fast Gamma [87] and high Gamma [88] were coined to
describe a frequency band between 90 and 140 Hz in the cortex. Two hypothesis
about high Gamma oscillations have been debated. One states that high Gamma
play the same role as Gamma oscillations but operate at a shorter timescales [75],
while other authors propose that high Gamma power is simply related to a spiking
activity increase [89, 90]. Furthermore, different works suggests [91, 92] that Gamma
and high Gamma have different mechanisms of generation, which would justify a
difference in the nomenclature for rhythms in the 30–90 Hz and the 90–140 Hz
bands.

• Ripple Oscillations : This type of high frequency oscillation is typically observed

during non-REM sleep, anesthesia and during consummatory behaviors (immobility,
drinking, eating and grooming). It has been traditionally studied in the hippocam-
pus [93, 67, 94], but recent work has proposed that ripples can also be recorded in
the cortex [57, 95, 96]. In hippocampus ripples are observed together with sharp
waves. Sharp wave ripples in hippocampus are typically characterized by periods
of approximately 50 to 100 ms of a large negative sharp wave with superimposed

12
 high frequency oscillatory activity (80–200 Hz) called ripples, followed by a period
of approximately 200 ms of a positive wave [94]. Ripple oscillations have been pro-
posed to play an important role in memory consolidation during non-REM sleep
or consummatory behaviors [97, 93], and, more recently, to be involved in memory
retrieval during active awake states [97, 95, 96].

• Spindle Oscillations : Also known as sleep spindles, it is a characteristic rhythm

that occurs during non-REM sleep (stages 2 and 3 [98]). It is typified as oscillations
(11-16 Hz) lasting for 0.5 to 3 seconds, whose amplitudes wax and wane [98, 99].
Spindles are the product of the interaction between thalamic reticular, thalamocor-
tical, and cortical pyramidal networks [100, 53]. It has been proposed that spindles
may be an essential element for memory formation [48] and short- and middle term
synaptic plasticity [51].

13
14
 Asynchronous and Irregular States

The spontaneous cortical activity observed in vivo is characterized by highly stochastic

and irregular neuronal spiking with low pairwise spike correlations in different brain states
[101, 102, 103, 104].
One of the first works to approach the spontaneous cortical activity in a theoretical
point of view, demonstrating it to be an emergent property of large networks of excitatory
and inhibitory neurons sparsely connected by strong synapses, was developed by Carl Van
Vreeswijk and Haim Sompolinsky [105]. In this seminal work based on binary neurons
(see the section Computational Models), the stochastic feature of spiking activity was
associated to the approximate balance between its excitatory and inhibitory inputs which
was dynamically adjusted by the own network intrinsic properties. This balanced state has
also been experimentally observed in vitro [106] and in vivo works [107, 108].
Subsequently, Nicolas Brunel and collaborators [109, 110] extended these results by
working with systems in which synaptic time scales were included. By means of spiking
networks of leaky integrate-and-fire neurons and mean field models (see the section Com-
putational Models), they showed that irregular firing activity in low firing rates could
be achieved in different conditions in networks with sparsely connected neurons with
strong synapses composed of excitatory and inhibitory units. In networks in which the
time constants of the membrane potentials of excitatory and inhibitory neurons were the
same, this stochastic firing behavior could only emerge when inputs were closely balanced.
On the other hand, if inhibitory neurons could depolarize faster than excitatory neurons
(through smaller membrane time constants), then this condition could be relaxed, leading
to irregular spiking activity close to what is observed in vivo, even if excitatory inputs
dominated [109]. Sub-sequential works have enlarged this view to more complex neuronal
and synaptic dynamics [111, 112]. More generally, this early analysis [110] showed that
different activity regimes could be achieved depending on the balance between inhibition
and excitation, and on the magnitude of external inputs. These regimes were classified
generically due to two main aspects: global activity (synchronous or asynchronous) and
individual neuronal activity (regular or irregular). Following this criteria, the observed
spontaneous activity observed in cortical regions in vivo has been classified to be in an
Asynchronous and Irregular (AI) state.
The AI state is precisely defined as a stationary global activity with strongly irregular
individual firing at low rates. However, spontaneous awake activity is far from being
stationary. In vitro studies have shown that to be able to correctly explain the inter-spike
intervals (ISI) of neuronal activity observed in vivo, factors such as synchrony need to

15
be taken into consideration [113]. In addition, human ECoG measurements, in conditions
in which the subjects are awake and immobile, show the occurrences of different brain
rhythms in different brain regions [114]. Using the terminology developed by [110], resting
awake activity could be interpreted as the dynamical switch between Asynchronous and
Irregular states and Synchronous and Irregular states (brain oscillations). This switch,
according to these simplified models [110], could be controlled by the strength of an
external drive and by the balance between excitation and inhibition. The theoretical
description of brain oscillations will be detailed on chapter: Computational Models.
Different brain states are characterized by different levels of occurrences of different
brain rhythms [115]. In particular, during Slow-wave sleep (SWS), neuronal activity re-
mains irregular while it is strongly modulated by a Delta rhythm, which is characterized
by periods of sustained firing and periods of very low spiking activity called the Up and
Down states [102].
It has been proposed that the neuronal activity displayed during Up states of SWS are
similar to the activity displayed during awake states both in terms of spiking and rhyth-
mic activity observed on the LFP [38]. In Work 1, we review several works that have
shown similarities and differences between these two brain states. We compare previously
published results with our own analysis of human data provided by one of our collabora-
tors, focusing on differences of spiking patterns (firing rate and spiking correlations) and
on different neuronal participation on Gamma rhythms in awake and SWS.
Other properties of Awake-like states (Asynchronous and Irregular) are investigated
in Work 2, in which network responsiveness is measured and compared to periods of
increased synchrony where the networks present Gamma oscillations in a Synchronous
and Irregular manner.

16
 Gamma Oscillations

The first study characterizing Gamma-band activity (30-90 Hz) was developed by Edgar
Adrian in 1942 [116], who reported induced activity, due to olfactory stimulus, as si-
nusoidal oscillations between 30 and 60 Hz lasting for the duration of the stimulation.
Nevertheless, the study of Gamma oscillations gained the attention of the general scien-
tific community only in 1987, when Gray and Singer discovered 40 Hz oscillations at the
cellular level [66]. A complete review of historical hallmark studies is given by [117].
Gamma rhythms are observed in many brain regions, during both awake and sleep
states [118, 38, 69, 119, 74, 47], in different species. They have been reported to appear
spontaneously or due to external stimulus, being either induced, evoked, or emitted [120].
Interestingly, the spiking activity, measured concomitantly with Gamma activity on the
LFP, is highly irregular. Examples of this neuronal activity were observed in vivo in the
neocortex and hippocampus, where gamma oscillations were associated with stochastic
and sparse spiking activity of individual cells [68, 75, 121, 81, 122, 87, 69, 38, 123, 124].
In the last decades, Gamma oscillations have been associated to different high-level
cognitive functions, which lead to the hypotheses that Gamma oscillations are important
for information processing and coding. The most popular theories are the Binding-by-
synchronization Hypothesis [125, 126], the Phase Coding Theory [127, 128], the Commu-
nication Through Coherence Theory [129, 130] and Communication through Resonance
Theory [131].
In Work 2, we take advantage of the human data provided by one of our collabora-
tors to further characterize individual neuronal activity during Gamma oscillations. We
typify neurons according to their phase-locking and firing rate changes inside and outside
Gamma bursts, inquiring about behavior changes. In Work 2 we also compare the result
of this analysis with computational models putting into perspective the different theo-
ries of information processing and coding involving Gamma previously cited. In addition,
we use these computational models to explore three mechanisms of Gamma generation
(detailed bellow), and examine their effects on the spiking patterns of individual cells.

Mechanisms of Generation
The mechanisms underlying Gamma oscillations have been widely studied experimentally,
both in vivo and in vitro [70, 132, 133, 134, 135]. Three important mechanisms are the
Interneuron Gamma (ING), in which Gamma oscillations are generated by the exclusively
interaction among inhibitory neurons; the Pyramidal-Interneuron Gamma (PING), in

17
which Gamma is due to the interaction between inhibitory and excitatory neurons; and
the Chattering Induced Gamma (CHING)7 , in which the rhythm production relies on
the presence of pacemaker excitatory cells known as Chattering neurons [27, 133]. The
theoretical approach to study the underlying mechanisms of Gamma oscillations will be
discussed in the next chapter.
Several experimental models have been used to examine oscillation dynamics as an
emergent property of neuronal networks. These protocols include the application of ago-
nists or antagonists of certain receptors, and/or the electrical activation of specific cells
types [135]. We describe in Figure 8 three important works that reported, by means of
extracellular recordings, how Gamma oscillations can be initiated or blocked in vitro by
means of these techniques.
In the work [136], Gamma oscillations are generated on Hippocampus (CA3) by means
of Carbachol, a cholinergic agonist. In this preparation Gamma rhythms are blocked either
by NBQX (AMPA receptor antagonist) or by Bicuculline (GABAA receptor antagonist),
showing that both excitatory and inhibitory synapses are involved in Gamma generation,
an in vitro demonstration of PING mechanism (Figure 8A).
The work [137] reported Gamma oscillation in CA3 due to Kainate (Kainate receptors
agonist). In this study, Gamma oscillations were abolished by Bicuculline, but were not
affected by the AMPA receptor antagonist GYKI 53655. This results indicate that the
observed rhythmic activity was mainly a consequence of inhibitory synaptic activity, an
illustration of ING mechanism (Figure 8B).
At last, the work [133] reported the importance of Chattering cells on Gamma rhythm
generation (CHING mechanism) on Cortex. In this study, Gamma oscillations were evoked
by Kainate and abolish after phenytoin application. Phenytoin is an anticonvulsant drug
that blocks persistent sodium conductance, acting preferentially on Chattering cells be-
cause of its electrophysiological behavior. It is known that Chattering cells are absent at
Hippocampus. The work of [133] showed that Kainate-induced Gamma oscillations were
only extinct by phenytoin on cortex and not in Hippocampus, indicating the importance
of Chattering cells on the rhythm generation, illustrating the CHING mechanism (Figure
8C).
Even-though these three mechanisms can be observed in cortex in vitro, it is still con-
troversial which of these mechanisms are engaged in different cortical regions in different
conditions in vivo [138, 132, 139, 140].

7
The term Chattering Induced Gamma (CHING) was coined in this thesis to designate the Gamma
generation mechanism that relies on the presence of Chattering neurons.

18
 A PING B ING C CHING
Pyramidal-Interneuron Interneuron Chattering Induced
Gamma Gamma Gamma
Cortex
400 nM kainate

50μM phenytoin

Hippocampus
kainate

phenytoin
0.1 mV

200 ms

Figure 8: Extracellular recordings demonstrating different Gamma generation

mechanisms in vitro. A) Gamma oscillations were evoked on hippocampal region CA3
by means of Carbachol and are blocked by both AMPA receptor antagonist NBQX and the
GABAA receptor antagonist Bicuculline, illustrating PING mechanism. Figure adapted
from [136]. B) Kainate-induced Gamma oscillations in CA3 are abolished by Bicuculline
but suffer no effect from AMPA receptor antagonist GYKI 53655 application, illustrat-
ing ING mechanism. Figure adapted from [137]. C) Phenytoin blocks Kainate-induced
Gamma oscillations in cortex, but not in Hippocampus. Since Chattering cells are sensi-
tive to Phenytoin and are absent in Hippocampus, this preparation illustrates the CHING
mechanism. Figure adapted from [133].

19
20
 Gamma Oscillations, NMDA
receptors and Schizophrenia

Schizophrenia is a mental disorder that affects around 1% of global population [141]. It is

characterized by three classes of symptoms: positive symptoms, which include delusions,
hallucinations and disordered thoughts or speech; negative symptoms, which comprehend
poverty of speech and deficits of normal emotional response, such as lack of motivation and
inability to experience pleasure; and cognitive deficits involving dysfunction of working
memory, long-term memory, attention and learning [142, 143, 144].
Several abnormalities have been identified in schizophrenic patients. These abnor-
malities extend from important differences in neurotransmitters systems to anatomical
deficits8 and are reflected in abnormal neuronal activity.
In healthy subjects, NMDA receptor antagonists can induce a psychotic state that re-
sembles all three classes of schizophrenic symptoms [147, 148, 149]. Likewise, in schizophrenic
patients, NMDA receptor antagonists can exacerbate the symptoms [148, 150]. These ob-
servations lead to the conception of the NMDA hypofunction paradigm [151, 152], which
states that the hypofunction of NMDA receptors (NMDAR) might be one of the causes
of schizophrenia9 . Furthermore, a reduced expression or binding of NMDAR in thalamus
and in cortex have been identified in postmortem studies [156].
NMDA receptors are critical to the development and to the adult function of GABAer-
gic interneurons. NMDA receptors regulate the expression of GAD67 [157] (GABA syn-
thesizing enzyme) and the expression of parvalbumin10 (PV) [157]. In vitro studies have
shown that NMDA receptor antagonists, such as ketamine or phencyclidine, reduce the
expression of GAD67 and PV [157, 160, 161], and reduce inhibitory synaptic transmission
in cortical slices [162]. Furthermore, post-mortem analysis of schizophrenic patient’s brain
have shown a reduced expression of PV and GAD67 [163, 164, 165, 166, 167, 142].
8
Schizophrenia is associated with widespread reductions in the volume of grey matter [145]. Since the
overall number of neurons is largely preserved, these volume reductions are believed to reflect a reduction
of synaptic connectivity [146].
9
Another important paradigm used to study schizophrenia is the one of the hyperdopaminergic hypothe-
sis. This is supported by the fact that dopamine D2 receptor antagonists are able to provide anti-psychotic
effects [153, 154]. D2 dopamine receptor antagonists are one of the most common antipsychotic drug type
used to treat schizophrenia. On the other hand, this type of treatment are not effective on negative
symptoms or cognitive deficits [155], which has been one of the reasons to focus research attention on the
NMDA hypofunction paradigm.
10
Parvalbumin, commonly known as PV, is a protein that modulates neuronal firing properties [158]
and neuronal plasticity in inhibitory neurons [159]. It is used as criterion to categorize inhibitory neurons
[25].

21
 The reduction of GAD67 in schizophrenic patients have been shown to be principally
pronounced in a particular type of inhibitory neurons, the PV-positive neurons [168],
which receive an important amount of glutamatergic drive from NMDA channels [157].
In agreement, genetic ablation of NMDA receptors in PV-positive interneurons in rodents
mimics important phenotypical (reduction of GAD67 [169], increase of neuronal excitabil-
ity [169] and increase of spontaneous Gamma power [170, 171, 172]) and behavioral fea-
tures of schizophrenia [173]. These observations support the idea that the hypofunction
of NMDA receptors in PV-positive interneurons are specially important in this illness11 .
An important question, which is still under debate, is whether the alterations observed
in Gabaergic neurons are the consequence or the cause of the modifications in the NMDA
neurotransmitter system.

NMDA receptor antagonists in sub-anesthetics doses

generate increase in neural activity and increase Gamma
power
As stated in previous sections, NMDA receptors mediate excitatory synaptic transmission.
Hence, it is intuitively expected that when applying NMDAR antagonists neural activity
be decreased. In agreement, NMDA receptor antagonists, such as Ketamine, have been
extensively used as anesthetics [176]. However, several preparations with sub-anesthetics
doses of NMDAR antagonists produce neural excitation [177, 178, 179, 180, 181, 182].
Several explanations have been proposed to explain this apparent paradox [144]. The in-
terpretation we adopt in this thesis, which is supported by experimental observations [15,
183, 162, 184], is that NMDAR antagonists in sub-anesthetics doses act preferentially on
inhibitory neurons, therefore increasing network activity indirectly by means of desinhi-
bition12 . For example, the work reported in [15] showed, on medial prefrontal-cortex of
freely moving rats injected with an NMDAR antagonist dizocilpine maleate (MK801),
that approximately 69% of inhibitory FS neurons decreased their firing, while 86% of
excitatory RS neurons increased their firing after MK801 injection, leading to a global
excitatory effect.
Besides of producing neural excitation, Ketamine in sub-anesthetic doses has been
associated to the increase of self-generated and evoked Gamma oscillations. Works in
human healthy subjects, under sub-anesthetic doses of Ketamine, reported an increase
of induced Gamma amplitude during sensory tasks in primary motor and visual cortices
11
Some works reported conflicting results and have questioned the hypothesis that PV-positive Fast
Spiking neurons play a role in Schizophrenia [174, 175]. Furthermore NMDA receptors are expressed
in both GABAergic and glutamatergic neurons, and it still remains unclear in which types of cells the
NMDA receptor hypofunction cause schizophrenia [175, 144].
12
One in vitro study has reported conflicting results [174].

22
[185] and an increase of evoked Gamma amplitude in auditory cortex, during steady-
state response [186] and in a paired-click paradigm [187]. In agreement, same results were
reproduced in animals in vivo. In rat neocortex, sub-anesthetic doses of Ketamine increase
Gamma power of spontaneous generated Gamma oscillations during free movement [188,
189, 190, 191] and were accompanied by a marked increase in locomotion, hyperactivity,
and ataxic behavior. Additionally, experiments in monkey with similar doses of Ketamine
also reported robustly induced spontaneous Gamma in primary motor cortex [192].

Gamma Oscillations in Schizophrenic Patients

Power and temporal correlations of neural oscillations during the resting state have been
demonstrated in studies with healthy twins to be highly heritable [193]. Following this
observation, dysfunctional neural oscillations could represent an endophenotype of certain
mental illness, guiding the search for its genetic contributions. Regarding schizophrenia,
important evidence for the relationship between anomalous neural oscillations and genetic
predisposition have been reported [194, 195, 196, 197]. Different abnormalities in Gamma
oscillations have been observed in schizophrenic patients, both in sensory-driven (evoked
oscillations) and in self-generated oscillations (spontaneous or induced) [198, 199].
Several works have reported evoked Gamma oscillation in schizophrenic patients pre-
senting reduced amplitude and reduced phase synchronization with respect to healthy
subjects during cognitive tasks. These studies include observations either during complex
tasks, such as during mental arithmetic calculations [200], or simple ones in which Gamma
band activity responses are measured due to auditory stimuli (steady-state evoked poten-
tials [201, 202, 203, 204, 205, 206, 207] or auditory oddball paradigm [208]), due to visual
stimuli (stimulus-locked evoked oscillation) [209, 210], or due to Transcranial Magnetic
Stimulation (TMS) [211]. On the other hand, in protocols measuring spontaneous (not
evoked) Gamma, several studies have found an increased power in Gamma band [212,
213, 214, 215, 199].
However, contrasting findings reporting an increase of evoked Gamma activity and a
decrease in spontaneous activity in schizophrenic patients also exist. Increases of Gamma
activity were reported in response to stimulus during both complex and simple tasks,
involving working memory [216] for example, or the auditory oddball selective attention
task [217]. While a decrease in spontaneous Gamma activity have been reported during
some cognitive tasks (not stimulus locked) [218, 219] or during resting conditions (when
the subject in not engaged in any cognitive task) [220]. In addition, other works have re-
ported no significant changes between schizophrenic patients and healthy subjects during
resting conditions [221, 214]. Similar discrepancies were detailed in [222, 223].
It has been proposed that part of these incompatibilities with respect of evoked Gamma
oscillations could be explained by technical reasons. Indeed, several studies [224, 209, 225,

23
205, 209, 226] analyzed changes on the power of evoked Gammma activity relative to a
pre-stimulus baseline rather than relative to the absolute level of Gamma power. However,
differences in baseline levels influence directly the calculation of post-stimulus responses.
For example, the study [217], showed that the absolute magnitude of Gamma synchrony
was enhanced when the reference to this baseline period was removed.
While part of the discrepancies in evoked Gamma could be explained by the observed
increase of Gamma activity on stimulus base-line, the discrepant results with respect to
spontaneous Gamma activity could be in part related to the stage of the schizophrenic
disease of the patients used in each study. Since it has been recently shown that the resting-
state Gamma band activity is significantly different in each of the stages of schizophrenia
disease [227].
This recently study [227] observed, for example, a widespread decreased of resting-
state Gamma activity in frontal, temporal and sensorimotor areas in chronic schizophrenic
patients. On the other hand, in first-episode patients13 , an increased activity was observed
on the occipital cortex (significant in the band 30-46 Hz, and strong in the band 64-90
Hz), while an activity decrease was only detected on the occipital cortex (significant in
the band 30-46 Hz, and moderate in the band 64-90 Hz). This increase of Gamma activity
in first-episode schizophrenic patients have also been detect in early works [217], which
showed, for example, an increased Gamma activity in first-episode patients in both resting
state and under an attention task. Several other works indicating this phenomena have
been systematically reviewed by Perrottelli and colleagues [223]. With this review, the
authors argue that the increase of Gamma activity could be used as a marker of the
illness onset, since the Gamma increase in the early phases of schizophrenia is a well
established fact.

Gamma oscillations, psychotic states and NMDAR hy-

pofunction
As stated in previous sections, the administration of sub-anesthetic doses of NMDAR
antagonists generate a significant increase of Gamma power on neuronal activity and
induces several schizophrenia-relevant symptoms, including hallucinations and delusions
[214, 228, 188, 191, 189, 190]
The increase of Gamma activity, even though not well established as a robust feature
of schizophrenia disease in chronic patients [198], has been frequently observed in patients
in early phases of the disease [223]. In addition, it was recently shown [229] that other
features of the early phases of schizophrenia (but not chronic phases), such as hyper-
13
The term first-episode patients refers to patients in the early-course schizophrenia disease, who have
experienced only one or few psychotic episodes.

24
connectivity, could be reproduced by the administration of an NMDAR antagonist 14 .
Other evidences in favor of the hypofunction of NMDA receptors as one of the elements
related to schizophrenia can be found in the review [144].
In parallel, it has been shown [230] that first-episode patients presented 9.3% higher
scores than chronic patients on the Positive And Negative Syndrome Scale (PANSS) and
16.3% lower on the negative symptom scale. Concurrently, Gamma-band responses in
schizophrenic patients often correlate with positive psychotic symptoms, with higher
gamma-band activity corresponding to increased symptom load [225, 205, 209, 226].
On the other hand, despite of these independent observations, the role of the increase of
Gamma activity on the occurrence of psychotic states in schizophrenic patients is largely
controversial. In Work 3, we use network models endowed with NMDA synaptic con-
ductances to mimic schizophrenic states in which the hypo-function of NMDA synapses
generate an increase of excitation (due to inhibitory desinhibition) and an increase of
the Gamma power generated by the global network activity. In Work 3 we also explore
the effect of NMDA hypo-function on the capability of the network to respond to ex-
ternal stimuli, providing, as a consequence, a possible interpretation with respect to the
type of neuronal information processing occurring during hallucinations and/or delusions
occurring during psychotic states.

14
In the referred work [229], a robust hyper-connectivity of the prefrontal cortex was observed in healthy
volunteers administered with sub-anesthetic doses of ketamine, similarly to the one observed in patients
in early schizophrenic stages. Interestingly this hyper-connectivity was not observed in chronic patients
[229].

25
26
 Computational Models

Historical Perspective

Individual neuron models

The mathematical study of neural activity dates the earliest eighteen century, when Louis
Lapicque, in 1907, published a paper on the excitability of nerve cells. Lapicque modeled
neurons by means of an analogy to electrical circuits, describing the neuronal membrane
as a circuit consisting of a parallel capacitor and a resistor. Lapique’s insights about the
fundamental relationship between the membrane parameters and its excitability gave rise
only decades later (1960s) to the simplified model known today as the Leaky Integrate-
and-Fire (LIF) [231].

A B

Figure 9: The neuron as a RC circuit. A) Equivalent circuit in which the neuronal

membrane is seen as capacitor C, coupled in parallel with a resistor R. V is the membrane
potential, Vrest is the resting membrane potential and I indicates an injected current. B)
Scheme representing the voltage trajectory of the integrate-and-fire model imagined by
Lapique, indicating that when V reaches a threshold value, an action potential is generated
and V is reset to a subthreshold value. Figure adapted from [232].

In 1943, Warren McCulloch and Walter Pitts developed an even more simplified model,
the formal neuron [233], in which neurons would be either active or inactive, presenting
as outputs 0 or 1. This type of model allowed the construction of neural networks which
could be simulated computationally with the technology available at the time and gave
important insights on the study of associative memory [234]. Subsequently, in 1952, Alan
Lloyd Hodgkin and Andrew Fielding Huxley explained and described mathematically for

27
the first time the ionic mechanisms underlying the initiation and propagation of action
potentials in the squid giant axon, receiving for this work the Nobel Prize in Physiology
or Medicine in 1963.
Today this model is known as the Hodgkin-Huxley model (HH) [235] and a huge diver-
sity of variations of it is available: from extended versions, describing neurons with several
compartments (spines, dendrites, soma, axon segments) [236, 237] to simplified versions,
such as the Adaptive Exponential Integrate-and-Fire (AdEx) [238]. In addition, several
other models have been developed though-out the decades, the most importants being the
FitzHugh–Nagumo Model, [239], the Hindmarsh–Rose Model, [240], the Izhikevich Model
[241] and the Quadratic Integrate-and-Fire (QIF) [242]. See the reviews [243, 244, 237].
The network models developed in this thesis used the AdEx model to describe their
units. In this model, neurons are described by their membrane potential, that thanks to
a supplementary equation, is capable of displaying spike-frequency adaptation 15 . Further-
more, neurons in the AdEx model are point neurons, that is, their membrane potential is
assumed to be homogeneous in the whole cell body, which is the same as assuming that
the neuron have no structure: no axon and no dendrites, only a punctual soma. More
details about this model are described in the methods section of Work 2 and Work 3.

Network models and Oscillations

Yoshiki Kuramoto started a general theory to treat chemical and biological systems whose
elements behave as oscillators [247]. This theory has been proven to be useful in several
domains of knowledge, including neuroscience [248]. Nancy Kopell, Bard Ermentrout and
collaborators were some of the pioneers to apply Kuramoto’s ideas on the field. This
branch of theoretical neuroscience got known as the Coupled Oscillators Theory, in which
neurons are depicted as intrinsically rhythmic units that spike regularly and are coupled
by weak synaptic connections 16 .
Great amount of our understanding on neural oscillations have been given by the Cou-
pled Oscillators Theory, which allowed the investigation of the effect of noise, couplings,
and delays on the synchrony and regularity of network dynamics [252, 253, 135, 254, 255,
15
Spike-frequency adaptation is the name given for a phenomenon in which neurons, when stimulated
with a step current, display a gradual decrease of their firing frequency (after an initial increase). This
phenomena is caused by several biophysical mechanisms, which are reviewed in [245, 246].
16
Couple Oscillators Theory is a general term assembling different types of description: from simplified
models, in which neurons are described by its Phase Response Curve (PRC) [249, 250, 251], to more
detailed models, in which the ideas from the PRC approach are extended, keeping the assumption of
weak synaptic disorder and weak noise. In these more detailed models, neurons fire synchronously with
the same firing rate as the network, acting as oscillators [252, 253, 135, 254, 255, 256, 257, 258, 247, 259,
260]. The Phase Response Curve (PRC) is a function that describes neuronal phase, characterizing the
way a neuron shifts its spike times with respect to the timing of depolarizing currents [249, 250, 251].
Neuronal phase is defined based on neuronal spikes: the phase oscillates between 0 to 2π, being the spike
time usually associated to the phase zero. PRC of individual neurons can be coupled and used to build
network models [261, 262, 263]

28
256, 257, 258, 259, 260] and gave important qualitative predictions about the mechanisms
of oscillation generation for different neuron types [242, 257, 252, 64]. However, the condi-
tions under which neurons can be considered as coupled oscillators, e.g weak synapses and
weak noise, are frequently not fulfilled [264, 70, 265]. Furthermore, the assumed spiking
regularity of individual neurons during global oscillatory activity is typically not observed,
since, as stated earlier, in general even when coherent oscillations are detected in extra-
cellular recordings in vivo, neurons present a stochastic and sparse pattern of firing [68,
75, 121, 81, 122, 87, 69, 38, 123, 124].
Nicolas Brunel and collaborators developed models able of capturing both phenomena:
fast network oscillations combined with irregular neuron firing. These models assume fast
synapses with strong synaptic weights and synaptic delays [266, 110, 111, 112, 267]. This
type of description allows neurons to spike with firing rates much smaller than the network
oscillation frequency, in a stochastic way, similarly to what it is observed in vivo. This
type of dynamical regime is known as the firing rate regime [111], in contrast to fully
synchronized modes, like the ones observed in coupled oscillators models, known as the
spike-to-spike regime.
Another important theoretical approach, in the study of neural oscillations, is the
field of the so-called rate models (also known as mean-field models or mass models). This
approach describes the collective properties of a large numbers of neurons instead of
focusing on individual neural dynamics, by modeling the average firing rate of a particular
population of neurons. These models are obtained by means of heuristic arguments [268,
269, 270], not being directly derived from spiking neurons. The first firing rate models
were derived under the assumption that individual neurons spike stochastically [271, 272]
being able to characterize (stationary) oscillations in the firing rate regime [266, 110, 111,
112, 267]. These models were usually characterized by one single variable (the average
population firing rate), evolving according to an differential equation described by a time
constant τ and a steady-state input-output transfer function (f-I curve). These models, on
the other hand, could not describe neuronal activity if a significant portion of the neurons
fired synchronously (like in spike-to-spike regime). Recent firing rate models [273, 274]
have incorporated a second variable (describing the network synaptic activity) being able
to model oscillations in both regimes (spike-to-spike regime and firing rate regime).
In Work 2, we developed different network structures generating Gamma oscillations,
working in the firing rate regime, using Adaptive exponential integrate-and-fire neurons
[238]. These models were improved (with the addition of NMDA channels) in Work 3.

Gamma oscillations in theoretical models

Even though several experimental protocols have been used to examine the generation
mechanisms of Gamma oscillations, a more complete understanding requires the use of

29
theoretical models [132, 70]. Gamma oscillations have been extensively modeled in the
literature with different neuronal models and networks structures [132, 70, 64]. In Work
2, three types of Gamma generation mechanisms were explored: the Interneuron Gamma
(ING), the Pyramidal-Interneuron Gamma (PING) and the Chattering Induced Gamma
(CHING).
In this section, we describe theoretically each of these mechanisms and reference some
of the most important theoretical hallmarks of the literature.

ING Model
The ING mechanism can be qualitatively understood as a succession of events: (1) The
rhythm starts when a subset of inhibitory neurons discharge together, generating syn-
chronous inhibitory post-synaptic potentials (IPSPs). (2) The neurons who receive these
inputs, have their spike times constrained, being able to spike again only when these
inhibitory currents decay. (3) These postsynaptic neurons are forced to discharge their
next spike closer in time to their pre-synaptic neurons, which will also spike when the
global inhibitory input decays. This spike time restriction increases network synchrony
progressively, generating the rhythm.
It has been shown theoretically that Gamma oscillations emerge spontaneously in a
networks exclusively composed of mutually connected inhibitory neurons if a sufficient
external drive is provided [275, 276, 277, 259, 278, 266, 111, 112] and certain conditions
are fulfilled. However, these conditions are different depending on the oscillatory regime
studied (firing rate regime or spike-to-spike regime). While in the spike-to-spike regime
the oscillation frequency strongly depends on synaptic decay time, allowing synchrony to
emerge even without any synaptic latency [275, 276, 277, 259], in the firing rate regime,
synchronization is generated by the presence of a neuronal phase-lag (either due to synap-
tic delays or due to the time a conductance-based neuron takes to respond to synaptic
currents because of its intrinsic properties) and depends weakly on synaptic decay times
[111, 112]. In addition, the frequency of Gamma oscillations generated by the ING mech-
anism is also differently determined in these two regimes. In the spike-to-spike regime the
oscillation frequency is mainly determined by the inhibitory synaptic decay time [276,
277], while in the firing rate regime it depends much more on the fastest time scales, like
latency and the rise times, than on the synaptic decay times.

PING Model
The PING model, based on the reciprocal interaction between pyramidal neurons and
interneurons, is the earliest proposed model to explain Gamma oscillations [271, 279, 280,
281, 111, 112]. This mechanism can be understood qualitatively as a succession of cycles
of picks of excitation and inhibition that alternate. In these cycles, excitatory neurons

30
recruit inhibitory neurons that in turn reduce network activity. The new cycle starts
when inhibition decays.
PING models can also describe Gamma oscillations in both spike-to-spike regime and
firing rate regime. Furthermore, the parameters defining the oscillation frequency in each
of these conditions are similar. In both regimes, the oscillation frequency depends strongly
on excitatory and inhibitory synaptic time decays [282, 111].
The phase relationship between excitation and inhibition is an important aspect to be
discussed, since it has been suggested to be a marker of the type of Gamma generation
mechanism [140]. In theoretical models of PING mechanism, excitatory neurons can spike
both before and after inhibitory neurons [283, 284, 112] during Gamma oscillations.
It has been shown theoretically by [112] that in the firing rate regime, in models
composed of conductance based neurons (neurons that include non-linear spike generation
mechanisms in their equations), the spiking order of excitatory and inhibitory populations
depends exclusively on single-cell characteristics. When the IAM P A /IGABA ratio is the
same in excitatory and inhibitory neurons, excitatory cells tend to follow the inhibitory
ones in most of the physiologically plausible parameter space. On the other hand, when
the ratio of excitation to inhibition is weaker in excitatory cells than in inhibitory ones,
excitatory cells tend to precede inhibitory neurons [111, 112]. Furthermore, [283, 284] have
studied neuron phase differences during ING and PING (on the spike-to-spike regime) and
the conditions in which each mechanism take place when both of them are allowed by the
network structure. These works have indicated, in addition, in which conditions during
spike-to-spike Gamma oscillations inhibitory neurons proceed excitatory ones when the
PING mechanism take place.

CHING Model
Chattering cells, also known as Fast Rhythmic Bursting (FRB) cells, were proposed
decades ago to be Gamma generators in cortical networks due to certain experimental
observations [66, 27, 118]. Nonetheless very few computational models, in our knowledge,
were proposed to study the emergence of Gamma oscillations due to the presence of this
type of neurons.
By means of multi-compartment neurons and voltage-dependent currents described
by the Hodgkin-Huxley formalism, the study of [285] showed that Chattering cells have
an important role in recruiting feedback inhibition through their pacemaker rhythmic
spiking. They observed that this role is not only due to the increase of firing rate, but
specifically due to their bursting firing pattern. The work of [285] argues that Chattering
cells would be important on Gamma generation in brain regions in which the synaptic
transmission to interneurons is not highly reliable. In agreement, experimental studies
have shown that Chattering cells work as amplifiers of Gamma activity due to their

31
intrinsic suprathreshold properties and to their profuse axonal projections to other chat-
tering cells and to other neurons on the network [27, 118, 133, 286]. In addition, other
computational multi-compartment models have shown the importance of the connection
via axonal gap junctions of Chattering cells with other pyramidal cells in the generation
of Gamma oscillations [133, 287].
In this thesis we developed a single compartment model of Adaptive exponential
integrate-and-fire neurons composed of Chattering cells, Regular spiking cells, Fast Spik-
ing cells, in which Gamma oscillations are generated exclusively thanks to presence of
Chattering Cells. We coined the term Chattering Induced Gamma (CHING) to refer to
Gamma oscillations generated thanks to the presence of these pacemaker excitatory cells.
This model is discussed in Work 2.

32
 Human Recordings Analysis

The data used in this thesis to constrain our models, both during Gamma and AI states,
has been already published in other works [288, 47] and was donated by one of our
collaborators, Sydney S. Cash (Department of Neurology, Massachusetts General Hospital
and Harvard Medical School, Boston).
In this thesis, we extend these prior analysis, focusing on awake states. The data
was acquired extracellularly in patients suffering of intractable epilepsy, who had had
multi-electrode arrays implanted during therapeutic procedures. The arrays registered
simultaneously local field potentials (LFP) and single-unit activity. We considered the
data of one patient for which the recording was very stable, and in which several periods
of wakefulness could be analyzed. Figure 10A indicates the structure of the recording, in
which several brain states during the night could be observed: slow wave sleep (SWS), light
sleep, REM sleep and awake periods. Only long and consecutive periods of wakefulness,
containing several seconds of Gamma bursts, were considered (Segments 1, 2, 3, 4 and 5
indicated in 10A).
Eighty one electrodes were implanted on the temporal cortex of this patient (layer 2/3)
and 91 neurons could be identified (some neurons were recorded by the same electrode).
These neurons have been already previously [288, 47] classified as Regular Spiking Cell
(RS), putative excitatory, and Fast Spiking Cells (FS), putative inhibitory. From the 91
neurons, 23 were recognized as FS and 68 were recognized as RS. This classification was
performed by means of clustering based on the spike shape and functional interactions
(determined using cross-correlograms) [289, 288].
In each data part, periods of Gamma oscillations (Gamma bursts) were identified and
neural activity was characterized with respect to the Gamma cycles. The duration of these
data parts are indicated in Figure 10B, together with the total amount of Gamma periods
identified inside of each of them. The identification of Gamma bursts was done separately
for each electrode. We considered as Gamma bursts periods in which the amplitude of
Hilbert Transform envelope (absolute value) differed from the mean, by at least 2 standard
deviations, for a minimum duration of 3 Gamma cycles (See Figure 10C). These criteria
were not enough to identify all Gamma bursts (some Gamma bursts were ignored). On
the other hand, no false positives were included in the analysis. All the Gamma bursts
automatically identified by the algorithm were individually confirmed visually. The oscil-
lation phase was acquired using the angle of the imaginary part of the Hilbert Transform
(See Figure 10C).
Individual neuron spike times were analyzed with respect to Gamma periods, both in

33
regard to their spiking phase and to their spiking frequency inside and outside Gamma
bursts. To identify neurons that were particularly phase-locked to Gamma oscillations, we
constructed individual phase distributions, which were tested for circular uniformity using
a Bonferroni-corrected Rayleigh test [61, 290]. A neuron was considered phase-locked if
we could reject circular uniformity at P < 0.01 (Z > Zc ). Figure 11 depicts in A and B,
two examples of distributions of phase-locked neurons.
To identify firing rate changes inside Gamma bursts, the average firing rates inside
(fγ ) and outside (fout ) of Gamma bursts were computed for each neuron. A neuron was
considered to increase its firing significantly if the observed average number of spikes
inside Gamma bursts was higher than the percent point function of a 95% interval of
confidence of a Poissonian distribution with average firing rate fout . Cells which had firing
rates smaller than 0.1 Hz or cells whose electrode measured less than 1 second of Gamma
bursts, in the respective data segment, were classified as inconclusive. Figure 12 indicates
an example of this procedure.
The results of this experimental data analysis, which inquired the relationship between
individual neuron activity and Gamma rhythms measured on LFP, are characterized in
Work 2. This study showed, by means of this analysis, that the participation of individual
cells of Gamma oscillations is very sparse. During Gamma bursts, only a small percent-
age of the recorded neurons have been identified to be phase-locked to the rhythm and
only some cells increased their firing. Furthermore, throughout the night (different data
segments), neurons changed of behavior. That is, neurons that were previously changing
their firing during Gamma bursts kept their firing the same and/or neurons that were
previously identified to be phase-locked were not phase-locked in other periods of time.
In addition, this analysis showed that FS cells presented significant higher level of phase-
locking and firing rate increase in comparison to RS cells. Likewise, FS cells changed
its behavior less than RS cells. Figure 13 indicates that this neuronal behavior changes
across the five data segments used on our analysis. The models described in Work 2 were
capable of displaying most of these features.

34
 A

Data
Seg. 3

Data
Data Seg. 4
Seg. 1 Data
Seg. 2 Data
Seg. 5

B C

Data
Seg. 4

Figure 10: Human electrophysiological data. A) Structure of the recordings of one

patient during the night. Thirty eight periods of wakefulness (whose duration are indi-
cated) could be recorded. These periods were interspersed with sleep states (SWS, light
sleep and REM), represented by colored lines, without the indication of their duration. In
our analysis, only long and consecutive periods of wakefulness were considered: Segments
1, 2, 3, 4 and 5. B) Total duration of each of the 5 data parts selected for the analysis and
the respective total duration of Gamma bursts inside of each of them. Averages are indi-
cated by the dashed lines. C) Upper graph: Gamma periods detection. Raw LFP (black),
band-pass filtered LFP (yellow) and Hilbert Transform Envelope (red) are shown. Gamma
bursts were detected by means of the deviation from the average of the Hilbert Transform
envelope (dashed red line) of at least 2 SDs (dotted red line), with a minimum duration
of 3 Gamma cycles. The gray shaded region indicates one example of identified Gamma
burst. Lower graph: Oscillation Phase extraction. The oscillation phase were obtained by
the angle of the imaginary part of the Hilbert Transform. Phase distributions per neuron
were computed based on the oscillation phases where each neuron spiked.

35
 A Neuron 66 B Neuron 9

Figure 11: Phase-locking statistical test. Phase distribution of two randomly picked
cells from the human recordings during Data segment 1 : one excitatory (A) and one
inhibitory (B). The phase distribution of each cell was fitted to a Von Mises curve, which
allowed the estimation of its preferred phase θV M . The phase distribution of each neuron
was tested for circular uniformity using a Bonferroni-corrected Rayleigh test [61, 290]. A
neuron was considered phase-locked if the circular uniformity at P < 0.01, (Z > Zc ) could
be rejected.

Figure 12: Firing rate change statistical test. A: Activity of two randomly picked
cells during several Gamma bursts: neuron 13 (inhibitory, left) and neuron 75 (excitatory,
right). The graphs display the firing patter around Gamma bursts (indicated by the black
doted lines). Each point corresponds to one spike in the correspondent tuple of time and
burst ID (y-axis). B: Histogram computing the distributions of all spikes inside all Gamma
bursts of neuron 13 (left) and neuron 75 (right). C: Exemplification of firing rate change
statistical test. The Poissonian distribution of these two neurons is constructed based on
their average firing rate calculated outside of Gamma bursts. The critical number of spikes
nc , indicated by the dotted lines, is calculated based on the percent point function of the
respective Poissonian Distribution for a period T, with an 95% interval of confidence.
The observed number of spikes nobsv is depict as a dot over the curve. According to this
procedure, only neuron 75 is considered to increase its firing, since nobsv > nc .

36
 Inconclusive

Figure 13: Change of Cell Behavior across time during Gamma in experimental
data. Middle panel represents each cell by a circle in each of the 5 data segments. FS and
RS phase-locked cells are depicted respectively as red and green circles, while not phase-
locked or inconclusive cells (with respect to phase locked) of both types are depicted as
blue and gray circles respectively. Superposed to each cell circle, pointing up and down
triangles were added to indicate if the cell increased (△) or decreased (▽) its firing. If
the cell didn’t change its firing significantly a minus sign (-) was added. Side box plots
indicate, on the left, the percentage of phase-locked FS (red) and RS (green) cells in each
of the 5 data segments, and, on the right, the percentage of firing rate increase. Dotted
lines indicate the average value between the 5 data segments (phase-locking level: left and
firing rate increase: right). The box plot at the bottom depicts the superposed counts of
phase-locking and firing rate increase behavior of each individual cell, computed in the 5
data segments.

37
38
 Developed Models

The models developed in this thesis had their parameters constrained by electrophysio-
logical data acquired in human as described in the previous chapter. We exploit differ-
ent network structures to investigate three well-known mechanisms of Gamma generation
[138, 132, 139, 140, 27, 133]: either by the exclusive interaction between inhibitory neurons
[Interneuron Gamma (ING)] or by the interaction of inhibitory and excitatory neurons
via Pyramidal-Interneuron Gamma (PING) or via Chattering Induced Gamma (CHING).
While networks with a structure similar to the PING Network have been largely used in
the literature, the structures of ING and CHING Networks were developed exclusively in
this thesis.

A PING B ING C CHING

Pyramidal-Interneuron Interneuron Chattering Induced
Gamma Gamma Gamma

Figure 14: Network structures of the three developed models.

Each of the three developed networks used the Adaptive Exponential Integrate-And-Fire
Model (Adex) [238] for its neural units, and were composed of different types of cells:
Regular Spiking Cells (RS), Chattering Cells (Ch) and Fast Spiking Cells (FS). Network
structures and individual cell parameters are described in detail in Work 2.

Parameter search

An extensive parameter search was performed to identify network and neuronal parameter
sets which could, in each model, reproduce experimental features. Physiologically plausible
distributions of firing rates, of membrane conductance and of inter-spike intervals were
used to constrain these models, the same way as the level of phase-locking and firing rate
changes during Gamma oscillations extracted from our data analysis.

39
Parameter search in PING Network

For PING network, a parameter search with respect to the synaptic decays was performed.
Parameters were chosen in such a way that oscillations in the Gamma band would be
present in a network composed of neurons with neuronal parameters chosen in accordance
with previous publications [291, 292]. Figure 15 indicates this parameter search.

Parameter search in ING Network

The ING network is composed by the mixture of two networks: one being composed by
RS and FS neurons (which can not generate oscillations due to the choice of synaptic
conductances, τe =τi = 5 ms, like it is shown on Figure 15D), and one network being
composed by FS neurons that are highly connected (FS2 neurons17 ). This network of
highly connected FS neurons can generate Gamma oscillations by its own, thanks to its
dense connections, like it is shown in Figure 16.
For the ING network, a parameter search was performed to identify which interaction
between each population (RS, FS and FS2 neurons) could allow the model to reproduce
experimental features. Among these experimental features are the level of phase-locking
and the level of firing rate changes during Gamma oscillations, extracted from our data
analysis. Because this parameter search was highly multidimensional, it could not be
performed as systematically as the previous ones depicted in Figures 15 and 16. For this
reason, we present a scheme indicating the qualitative conclusions from this parameter
search, which justify the choice of probabilities of connections between each neuronal
population, used in this model (Figure 17).

Parameter search in CHING Network

The CHING network is exactly the same network displayed in Figure 15D (with synaptic
conductances: τe =τi = 5 ms), with the difference that 5% of its RS neurons were replaced
by Chattering cells. For this reason, the rhythms it presents are exclusively generated by
the Chattering activity. Consequently, the parameter search performed in this model was
entirely focused on the choice of the intrinsic cellular parameters of this neuronal type.
Again, since this parameter search was highly multidimensional, it could not be performed
in a systematical way. Several parameter combinations were tested taking into account
the following criteria: 1) cells should keep their chattering properties, 2) the oscillation
should be on Gamma band, 3) the oscillation should be on the firings rate regime, with
sparse and irregular firing, 4) cells should present physiologically plausible distributions of
firing rates, of membrane conductance and of inter-spike intervals, 5) cells should present
17
We decided to rename this population as FS2, to stress the different pattern of connectivity between
this group of highly connected FS neurons, from the other FS neurons in the network.

40
levels of phase-locking and firing rate changes during Gamma oscillations similar as the
one extracted from our data analysis.

41
A

B C
Synchrony index

Synchrony index
Δt

Δt

Figure 15: Synaptic time scale parameter search of a network composed of RS

and FS neurons randomly connected. The network used to produce this figure was
composed of 20000 excitatory Regular Spiking and 5000 inhibitory Fast Spiking neurons
connected randomly with a probability of connection of 2%. All synapses were delayed
by a time delay of 1.5 ms, and had reference synaptic strengths of QR e = 1 nS or Qi = 5
R

nS and reference synaptic time scales of τeR =τiR = 5 ms. Synaptic strengths (Qe,i ) were
normalized at each tested time scale (τe,i ) to keep the same synaptic gain, such that:
Qe,i = (QR e,i .τe,i )/τe,i . A) Network oscillation frequency depicted in a color scheme as a
R

function of excitatory and inhibitory synaptic time scales. White color corresponds to
regions in which no oscillation was identified in RS population. B) Synchrony Index of
RS population (top) and network balance (bottom) as a function of synaptic time scales.
The Synchrony Index (SI) is based on the auto-correlation of the population frequency of
RS cells. To be calculated, the autocorrelation of the population frequency was fitted by a
damped cosine function and the value of this fitted function at zero time lag was defined
as the SI. If the exponential decay rate was higher than 100, it was considered that there
was no global oscillation at the population scale. The network balance was defined as the
⟨Iexc ⟩N
rate between the average excitatory and inhibitory synaptic currents, ⟨Iinh ⟩N
, in which
t
⟨⟩N stand for average among neurons and ⟨⟩t average on time. White squares indicate the
two different parameter sets used in our simulations ( τe =τi = 5 ms for AI Network, and
τe =1 ms , τi = 7.5 ms for PING Network, as decribed in Work 2). C) Same as B but
calculated for the FS population. D) Population frequency autocorrelation of RS (green
dots) and FS population (blue dots) neurons of the two used parameter sets. Solid lines
indicate the damped cosine fitted function.
 A C E
e1 e3

B D e2 e4

Figure 16: Gamma Network parameter search. The network connectivity (p) vs. inhibitory
synaptic strengths (Qi ) parameter space of the Gamma Network are displayed as color-plots. A)
Average spiking frequency. B) Network oscillation frequency.
 C) Network
 balance: rate between
⟨Iexc ⟩N
the average excitatory and inhibitory synaptic currents, ⟨IInh ⟩N , in which ⟨⟩N stand for
t
average among neurons and ⟨⟩t stand for average on time. D) Membrane Potential Synchrony
σV2
1 N
(χ), calculated by means of the equation: χ2 = 1
ΣN 2 , in which V (t) = N Σi Vi (t), σV 2 =
N i σVi

⟨[V (t)]2 ⟩t − [⟨V (t)⟩t ]2 and σVi 2 = ⟨[Vi (t)]2 ⟩t −

[⟨Vi (t)⟩t ]2 .
The set of parameters which allowed
the Gamma Network to oscillate in the Gamma range are indicated by a star symbol. The white
and yellow curves depict parameter choices in which the product between p and Qi are the
same. The yellow curve indicates all parameters equivalent to a choice of p=60% and Qi = 5 nS
(Q′i = 3/p′ ), while the white curve indicates all parameters equivalent to a choice of p=10% and
Qi = 5 nS (Q′i = 0.5/p′ ), like it is usually used in other works [111]. Every point in each graph
is given by the average output of 10 simulations of 5 seconds each. In these simulations, each
neuron of the Gamma Network received 400 independent and identically distributed excitatory
Poissonian spike trains with a spiking frequency µExt =5 Hz and a synaptic strength of QExt =
1 nS (that decayed with synaptic time constant of τE =5 ms). E) Network activity for the
parameters indicated with a start in A, B, C and D (p=60% and Qi = 5 nS). The raster plot of
the whole network (e1), the population frequency (e2), the membrane potential of 3 randomly
chosen neurons (e3) and the power spectrum of the population frequency (e4) are indicated. The
population frequency is calculated as the total number of spikes (spikes of the whole network) in
a time bin of 1 ms, divided by the duration of this time bin. Because of the exclusive presence of
inhibitory neurons and its high level of recurrent inhibition, this network is capable of generating
Gamma rhythms with frequencies around 70Hz by means of an ING mechanism.

ING Network

Figure 17: Qualitative scheme indicating the role of each population interaction on
the network dynamics.
 Part II

Research Articles

45
 Work 1

Cellular correlates of wakefulness

and slow-wave sleep: evidence for a
key role of inhibition

Reference: Susin, Eduarda, and Alain Destexhe. "Cellular correlates of wakefulness and
slow-wave sleep: evidence for a key role of inhibition." Current Opinion in Physiology 15
(2020): 68-73.

French Abstract:

Dans cet article, nous avons considéré et discuté certaines études récentes à propos
de corrélations cellulaires, dans des états de veille et de sommeil, basés sur des enreg-
istrements d’unités multiples. Ces études concernaient, d’entre autres, une nouvelle forme
d’homéostasie observée pendant le sommeil lent profond (SWS de l’anglais), où les cel-
lules les plus actives (haut taux de décharge) ont montré une baisse dans leur taux de
décharge pendant le SWS, tandis que, inversement, les cellules les moins actives ont mon-
tré une tendance à décharger plus rapidement dans le SWS. Également, autres références
ont montré une implication plus forte des neurones inhibiteurs par rapport aux neurones
excitateurs lors des oscillations Gamma dans le SWS, et une augmentation des corréla-
tions croisées par paires exclusivement entre les neurones inhibiteurs pendant le SWS.
Nous comparons les résultats de ces études (chez des modèles animaux) avec nos analyses
effectuées sur des enregistrements chez l’homme. Nous confirmons certaines similitudes
dans la dynamique cellulaire observées entre l’état de veille et le SWS, et soulignons les
différences qui sont principalement apparentes dans la dynamique et dans les corrélations
des neurones inhibiteurs. Ces résultats suggèrent que les réseaux inhibiteurs sont essen-
tiels dans la dynamique du sommeil, et devraient être une cible principale dans les études
futures.

47
 Available online at www.sciencedirect.com

ScienceDirect

Cellular correlates of wakefulness and slow-wave sleep: evidence

for a key role of inhibition$
Eduarda Susin and Alain Destexhe

Recent studies have identified interesting cellular dynamics in In natural SWS, extracellularly recorded neurons fire in
wakefulness and slow-wave sleep (SWS), as we review here for coincidence with the depth-negative component of slow-
unit recordings in animals and human. First, a novel form of wave complexes, whereas the depth-positive component
homeostasis was observed during sleep, where high-firing cells is associated with neuronal silence [4–6]. Similar relations
in wake tend to fire slower during SWS, while low-firing cells in were found for delta waves in the intact brain [7], or
wake tend to fire faster in SWS. Second, there seems to be a isolated cortex [8]. Thus it seems that slow-wave com-
stronger involvement of inhibitory cells compared to excitatory plexes are characterized by alternating periods of sus-
cells during gamma oscillations in SWS. Third, pairwise cross- tained firing and neuronal silence usually called “Up” and
correlations between cells seem to increase specifically during “Down” states, respectively. It was further shown that, in
SWS, but only for inhibitory neurons. We compare these results natural sleep, Up and Down states appear synchronously
between animal and human unit recordings, and confirm the in multiple cells recorded extracellularly [6], and the
similarities in cellular dynamics in wake and SWS, and highlight relation between the slow-wave and cellular hyperpolari-
that differences are mostly apparent in the dynamics and zation was later confirmed by intracellular recordings of
correlations of inhibitory cells. These results suggest that neurons in naturally sleeping animals [9]. Note that
inhibitory networks are key in the dynamics of sleep, and Down states are not periods of complete silence. A
should be a main target in future studies. possible relation between the residual activity during
Down states and memory consolidation has been pro-
Address posed based on extracellular recordings in prefrontal
Department of Integrative and Computational Neuroscience (ICN), Paris- cortical areas in rats [10]. This residual activity during
Saclay Institute of Neuroscience (NeuroPSI), Centre National de la
Recherche Scientifique (CNRS), 91198 Gif-sur-Yvette, France
down states was also observed in human recordings [11].

Corresponding authors: Susin, Eduarda ([email protected]), Early studies also showed that, in aroused states, cortical
Destexhe, Alain ([email protected]) cells fire tonically [12–14], and the EEG is dominated by
low-amplitude fast activity in the beta/gamma (15–75 Hz)
Current Opinion in Physiology 2020, 15:68–73 frequency range. Extracellular studies in natural SWS
found that all characteristics from the spiking activity,
This review comes from a themed issue on Physiology of Sleep
LFP activity or LFP coherence, are identical between the
Edited by Vladyslav Vyazovskiy and Jenny Morton
Up-states of SWS, wakefulness or rapid-eye movement
For a complete overview see the Issue and the Editorial (REM) sleep [6]. These seemingly identical properties
Available online 27th December 2019 led to the suggestion that SWS Up-states constitute small
https://doi.org/10.1016/j.cophys.2019.12.006 episodes of activity replayed from wakefulness, possibly
involved in memory consolidation [15]. This view was
2468-8673/ã 2019 Elsevier Ltd. All rights reserved.
challenged recently by studies who found subtle differ-
ences in firing activity between natural SWS and wake-
fulness [16], as we review in this paper.

Cellular correlates of wakefulness and slow-

Introduction wave sleep
Slow-wave sleep (SWS) is characterized by the preva- Generic features of brain activity during wake and sleep
lence of slow waves in the delta (0.5–4 Hz) frequency states in humans are illustrated in Figure 1. We used a
range, which are visible in the electroencephalogram dataset where units could be separated into fast-spiking
(EEG) and local field potentials (LFPs). Early studies (FS), presumed inhibitory, and regular-spiking (RS), pre-
using intracellular recordings in anesthetized sleeping sumed excitatory cells; some of these were confirmed by
animals [1–3] have shown that the depth-positive (sur- direct functional identification [19]. In these human
face-negative) EEG components of slow waves are asso- recordings, during Wake and REM states, the activity
ciated with cellular hyperpolarization and pause of firing. of both RS and FS cells is sustained, asynchronous and

$
Research supported by Centre National de la Recherche Scientifique (CNRS), the European Community (Human Brain Project, H2020-785907),
and École des Neurosciences de Paris (ENP).

Current Opinion in Physiology 2020, 15:68–73 www.sciencedirect.com

48
 Cellular correlates of wake and sleep states Susin and Destexhe 69

Figure 1

SWS Wake REM

250 μV

0 1 2 3 4 5 6 7 8 9 0 1 2 3 4 5 6 7 8 9 0 1 2 3 4 5 6 7 8 9
Time [s] Time [s] Time [s]
Current Opinion in Physiology

Features of different brain states in Human. Selected parts of 10 s of recordings of LFP (top) and spiking activity (bottom) of SWS, Wake and
REM are shown. The data were recorded from the same human subject which had a multielectrode array implanted in temporal cortex before a
therapeutic surgery. LFP and spiking activity were simultaneously measured and the spikes were sorted allowing the identification of 68 putative
excitatory (blue) and 23 putative inhibitory (red). Data sample from [19].

irregular, and the LFP is dominated by high-frequency interneuron activity follow the relation: Wake = REM
activity. In contrast, during SWS, the activity is domi- > SWS, while somatostatin-positive interneurons follow:
nated by slow waves in the LFP, which are paralleled Wake > SWS > REM. In human recordings, the firing
with synchronized silences (Down states) where almost rates were comparable in all brain states [21], which would
all cells cease firing. Outside these periods (Up states), tend to support that the recorded FS cells are PV+ cells,
the dynamics is sustained, asynchronous and irregular, although this should be confirmed with better statistics.
similar to Wake and REM states. These observations lead
to the statement that the Down states is what makes SWS In addition, a recent analysis [16] found important dif-
dynamics more synchronized, and in a sense, the higher ferences between SWS and Wake by focusing on Up
synchrony during SWS does not come from the firing of states. By means of silicon probes implanted in frontal
units, but rather comes from these periods of non-firing. cortical areas of rats, measurements of extracellular sig-
nals allowed simultaneous identification of brain state and
Consistent with this interpretation, a previous extracellu- the assessment of the firing pattern of putative excitatory
lar study in marmoset cortical areas [18], found that there and inhibitory cells. This analysis [16] showed that
was no significant increase of neural correlation during during SWS pyramidal cells that displayed high-firing
SWS when recording spontaneous activity, providing during Wake tended to decrease their activity during
another evidence that SWS Up-states display firing SWS, while cells that displayed low-firing during wake
dynamics almost identical to wakefulness. tended to increase their firing during SWS. We confirm
these findings also in humans through similar recordings
However, several observations have shown differences in and data processing performed previously [19]. Figure 2
spiking activity during wake states and SWS. For compares the results from [16] with our own analysis
instance, a relevant feature that has been taken into based on the data coming from [19]. Despite the small
consideration is the firing level of single units in different number of cells available in the human data set, in a
states. A recent work using wide-field calcium imaging in confidence interval of 95%, it is possible to observe a
cortical superficial and deep layers in mice [17] showed small slope with respect to the identity (0.67–0.94). These
that the activity of excitatory neurons was highest during results (in both data sets) indicate that a new type of
Wake, followed by that of SWS and subsequently fol- homeostasis, as argued by [16], in which a homogeniza-
lowed by that of REM sleep (Wake > SWS > REM). tion of firing is observed thorough the differential action
Using two-photon imaging in layer 2/3, the same study in cells with different levels of activity (in opposition to
[17] also showed that paralvabumin-positive (PV+) previous homeostatic models [22,23]). On the other hand,

www.sciencedirect.com Current Opinion in Physiology 2020, 15:68–73

49
70 Physiology of Sleep

Figure 2

Putative Excitatory Cells Putative Inhibitory Cells

Rat
(a)
SWS Firing Rate [Hz]

SWS Firing Rate [Hz]

WAKE Firing Rate [Hz] WAKE Firing Rate [Hz]

Human
(b)
SWS Firing Rate [Hz]

SWS Firing Rate [Hz]

WAKE Firing Rate [Hz] WAKE Firing Rate [Hz]

Current Opinion in Physiology

Comparison of firing rate during Wake and SWS states in rat and human cortex. Each point in the graph characterizes the firing rate (log
scale) of an individual cell in the two states. Putative excitatory cells, both in rat and humans, suffer a change in their firing rate depending of their
level of activity. Neurons that are highly active during wake decrease their firing during SWS, while neurons that have very low firing rates increase
their firing (see arrows). Diverging results were found in the two data sets when looking to putative inhibitory cells. (A) Putative excitatory and
inhibitory neurons in rat (slopes, 95% confidence interval: 0.66-0.71 and 0.996-1.005). Figure adapted with permission from [16]. (B) Putative
excitatory and inhibitory neurons in human (slopes, 95% confidence interval: 0.67–0.94 and 0.68–0.94). Data sample from [19].

when considering putative inhibitory cells the result is Indicating an increase of FS participation in gamma
divergent in the two data sets (rat and human). No firing generation during SWS.
alteration is observed in the firing of inhibitory cells
during Wake and SWS in rat [16], while a tendency Another key aspect to differentiate the network state
close to that which is observed with excitatory cells is between Wake and SWS is the pairwise correlations
observed in human inhibitory cells. This result should be displayed by unit spikes. Figure 4A shows recordings
checked in data sets where more neurons are available. of multielectrode arrays implanted in two monkey cortical
areas (premotor dorsal — PMd — and motor neocortical
A further difference observed during Up states in SWS -MI, layers II/III) [11], in which spiking activity and LFP
with respect to Wake states is the involvement of inhibi- were simultaneously recorded. Cross-correlations
tory cells in gamma oscillations. As recently shown in between spike trains were normalized by the geometric
human recordings [11], both putative excitatory regular- mean of each cell pair’s average firing rates to avoid
spiking and putative inhibitory fast-spiking populations spurious effects from cell-intrinsic firing levels [24].
increase their firing during gamma oscillations. Here, by Remarkably, comparing the correlation matrices between
analyzing the pre-processed data from [19], we noted that Wake and SWS, one sees that pairs of cells become
there is a significant increase on the average firing of the correlated specifically during SWS (Figure 4). These
inhibitory population during SWS with respect to its sleep-specific correlations appear even between distant
activity during Wake (see Figure 3D). While no signifi- cortical areas. Interestingly, these pairwise spike train
cant increase in firing with respect to Wake is exhibited correlations only concern pairs of FS cells. With this
by the excitatory population during SWS (see Figure 3E). observation we emphasize one more time that there

Current Opinion in Physiology 2020, 15:68–73 www.sciencedirect.com

50
 Cellular correlates of wake and sleep states Susin and Destexhe 71

Figure 3

(a) Wake (b) SWS (c) 30-50 Hz

200 ms

FS Average Frequency [Hz]

200 µV
LFP

(d) FS

RS Average Frequency [Hz]

Time [s]
(e) RS

Frequency [Hz]
Time [s]
Time [s] Time [s]
Current Opinion in Physiology

Difference of firing rates during g events in RS and FS cells in Human Wake and SWS states. The color-maps indicating the average firing
rate per cell around g events in both Wake and SWS states are indicated in (A) and (B). The firing of regular spiking (RS) are displayed in the
bottom and the fast spiking (FS) cells is displayed in the top. Each line indicates the color-coded firing rate of each neuron in time averaged
between all the gamma periods detected. For each g event the firing rate was calculated in bins of 20 ms around the centered g period. The white
dashed lines indicate the beginning and the average end of g periods identified in the data. Gamma periods were identified by the variation of the
envelope obtained through the Hilbert Transform of the filtered LFP in the range of 30-50 Hz as is schematically shown in (C). Neurons were
ordered by their discharge probability during g. The average firing rate per population during g events in Wake and SWS are displayed in (D) and
(E). The color-coded firing rates shown in A and B were average in between neurons for each population: FS (D) and RS (E). The shadowed parts
indicate the standard error of the mean. Data sample from [19].

seems to exist a particularly high involvement of inhibi- certain time window. This work showed that a model
tory cells specifically during SWS. In agreement, a previ- based exclusively on pairwise couplings can successfully
ous study [18] also reported an increase of correlations predict the activity of both neuron types during Wake
during SWS in the response to auditory stimuli, but no states (see the measures of the Kullback–Leibler diver-
separation between RS and FS cells was attempted in that gence, DKL , in Figure 4b1). However, it also demon-
study. strated that pairwise interactions alone are not enough
to describe the probability distribution displayed by the
Furthermore, this particular involvement of inhibition FS population during SWS (Figure 4b2). This result
can also be detected from correlated patterns analyzed implies that, in contrast to the RS population, higher-
using maximum entropy models. Such methods can infer order, or even population-wide, interactions govern the
correlations from large data sets with the advantage that behavior of FS cells during SWS. This indicates another
they can uncover collective dynamics with very few ad hoc important difference between Wake and SWS states with
assumptions. Recently, spiking activity in human record- respect to FS cells.
ings were used to infer the parameters of a Maximum
Entropy model that takes into account only and exactly Discussion
the single neurons’ spiking probability, and the pairwise In this paper, we briefly reviewed a few of recent studies
covariances observed in the data [20]. In this work, about the cellular correlates of wake and sleep states
illustrated in Figure 4B, the empirical probability distri- based on multiple unit recordings. Similar to a previous
butions of population activity of FS and RS cells, during study in cats [6], it was found that in human and monkey
Wake and SWS, were compared to the probability dis- multielectrode recordings, wake and sleep states globally
tributions predicted by the model. Population activity display similar levels of firing. Most units fire irregularly
was defined as the average number of active cells in a and asynchronously, and this pattern is seen in all brain

www.sciencedirect.com Current Opinion in Physiology 2020, 15:68–73

51
72 Physiology of Sleep

Figure 4

(a) (b)

(a1) (b1)
Wake

Probability distribution
Population activity
(b2)
(a2) (a3) SWS
Probability distribution

Population activity
Current Opinion in Physiology

Difference in pairwise spike train cross-correlations among RS and FS cells during Wake and SWS. (A) Spiking activity measured during b
oscillations through multielectrod arrays in two cortical areas in monkey (premotor dorsal, PMd, and motor neocortical, MI). Neurons could be
identified as RS and FS. a1) Cell-cell correlations matrices containing cells from the two areas (PMd and MI). For cells 57/69 (during SWS) the
type of target cells is indicated with colors (red, FS; and blue, RS). a2) Spatial location of different cells in the implanted arrays. Note that FS cells
(exemplified by cells 57 and 69) display synchronous activity even with cell recorded in another cortical area (M1). a3) Spike cross-correlogram
between cells 57 and 69 during SWS and Wake states. A significant peak is observed only during SWS. Figure A was adapted from [11]. (B)
Empirical probability distributions of human spiking activity are compared with the ones predicted by Maximum Entropy model (see main text),
both in Wake (b1) and SWS (b2) for two type of cells (FS and RS). Measures of the Kullback-Leibler divergence, DKL , are indicated in the graphs.
Figure B was adapted from [20].

states (Figure 1). Note that, in mice, a different level of Taken together, these results show that inhibition seems
activity is observed between Wake and sleep depending to be a key to differentiate the network state between
on the cell type. In the light of recent studies [11], we Wake and SWS, even though the differences are subtle.
also found differences between Wake and SWS in human: This is true for the level of firing of inhibitory cells, which
(1) there seems to exist a weak homeostasis of firing seems to be maximally involved during SWS gamma
activity, where high-firing cells in wake tend to fire lower oscillations. It is also true for the level of correlation
during SWS, and vice-versa for low-firing cells, as found in between inhibitory cells, which show remarkably high
rat excitatory neurons [16] and which we confirm here in correlations that specifically appear during SWS, some-
human (Figure 2) for RS and FS cells. (2) There is a times extending to large cortical distances [11]
stronger participation of FS cells during gamma oscilla- (Figure 4A).
tions in SWS compared to Wake (Figure 3). (3) Besides
the level of firing, there is also an increase of the spiking We presently do not know why inhibitory networks seem
correlation between pairs of FS cells, specifically during different during sleep. A possibility is that the different
sleep [11] (Figure 4). levels of cholinergic drive in Wake and SWS may

Current Opinion in Physiology 2020, 15:68–73 www.sciencedirect.com

52
 Cellular correlates of wake and sleep states Susin and Destexhe 73

up-regulate some of the inhibitory neurons, resulting in this paper investigated the firing of excitatory and inhibitory cells in human
and monkey during wake and sleep states, during gamma oscillations. A
an increase of participation in gamma oscillations, as we higher level of participation was found for inhibitory cells, as well as high
observed. However, this does not explain why inhibitory correlations, specific to SWS, and only between inhibitory cells.
cells are so correlated during SWS, which necessarily 12. Hubel DH: Single unit activity in striate cortex of unrestrained
cats. J Physiol 1959, 147:226-238.
comes from an excitatory drive yet to be identified.
13. Evarts EV: Temporal patterns of discharge of pyramidal tract
neurons during sleep and waking in the monkey. J Neurophysiol
Acknowledgements 1964, 27:152-171.
This research was supported by the Centre National de la Recherche
Scientifique (CNRS) and the European Community (Human Brain Project, 14. Steriade M, Deschenes M: Inhibitory processes and
H2020-785907). E.S. acknowledges a PhD fellowship from the École des interneuronal apparatus in motor cortex during sleep and
Neurosciences de Paris (ENP). waking. II. recurrent and afferent inhibition of pyramidal tract
neurons. J Neurophysiol 1974, 37:1093-1113.
15. Destexhe A, Hughes SW, Rudolph M, Crunelli V: Are
corticothalamic ‘up’ states fragments of wakefulness? Trends
References Neurosci 2007, 30:334-342.
16. Watson BO, Levenstein D, Greene JP, Gelinas JN, Buzsáki G:
1. Steriade M, Nunez A, Amzica F: Intracellular analysis of relations  Network homeostasis and state dynamics of neocortical
between the slow (< 1 Hz) neocortical oscillation and other sleep. Neuron 2016, 90:839-852
sleep rhythms of the electroencephalogram. J Neurosci 1993, this paper showed clear evidence for a regulation of the firing of neurons
13:3266-3283. between Wake and SWS states, where neurons with high levels of firing
2. Steriade M, Contreras D, Dossi RC, Nunez A: The slow (<1 hz) tend to fire lower during SWS, and conversely, neurons with low levels of
oscillation in reticular thalamic and thalamocortical neurons: firing tend to fire higher during SWS. This can be interpreted as a form of
scenario of sleep rhythm generation in interacting thalamic homeostasis of firing dynamics between wake and sleep states.
and neocortical networks. J Neurosci 1993, 13:3284-3299. 17. Niethard N, Hasegawa M, Itokazu T, Oyanedel CN, Born J,
3. Contreras D, Steriade M: Cellular basis of eeg slow rhythms: a  Sato TR: Sleep-stage-specific regulation of cortical excitation
study of dynamic corticothalamic relationships. J Neurosci and inhibition. Curr Biol 2016, 26:2739-2749
1995, 15:604-622. this paper showed features of the neural activity of wake and sleep states
in mice. This was done for excitatory cells, which fired more during Wake
4. Steriade M, Contreras D, Amzica F, Timofeev I: Synchronization than the two sleep states, while interneurons also tended to fire more, but
of fast (30–40 Hz) spontaneous oscillations in intrathalamic with some differences according to the type of inhibitory cell.
and thalamocortical networks. J Neurosci 1996, 16:2788-2808.
18. Issa EB, Wang X: Increased neural correlations in primate
5. Amzica F, Steriade M: The k-complex: its slow (<1-hz) auditory cortex during slow-wave sleep. J. Neurophysiol. 2013,
rhythmicity and relation to delta waves. Neurology 1997, 109:2732-2738.
49:952-959.
19. Peyrache A, Dehghani N, Eskandar EN, Madsen JR, Anderson WS,
6. Destexhe A, Contreras D, Steriade M: Spatiotemporal analysis of Donoghue JA, Hochberg LR, Halgren E, Cash SS, Destexhe A:
local field potentials and unit discharges in cat cerebral cortex Spatiotemporal dynamics of neocortical excitation and
during natural wake and sleep states. J Neurosci 1999, 19:4595- inhibition during human sleep. Proc Natl Acad Sci 2012,
4608. 109:1731-1736.
7. Buzsaki G, Bickford RG, Ponomareff G, Thal L, Mandel R, 20. Nghiem T-A, Telenczuk B, Marre O, Destexhe A, Ferrari U:
Gage FH: Nucleus basalis and thalamic control of neocortical  Maximum-entropy models reveal the excitatory and inhibitory
activity in the freely moving rat. J Neurosci 1988, 8:4007-4026. correlation structures in cortical neuronal activity. Phys Rev E
2018, 98:012402
8. Frost JD Jr, Kellaway P, Gol A: Single-unit discharges in this study confirmed that inhibitory neurons follow a very particular
isolated cerebral cortex. Exp Neurol 1966, 14:305-316. dynamics during sleep, pairwise correlations cannot account for their
firing statistics, and higher-order correlations are needed.
9. Steriade M, Timofeev I, Grenier F: Natural waking and sleep
 states: a view from inside neocortical neurons. J Neurophysiol 21. Dehghani N, Peyrache A, Telenczuk B, Le Van Quyen M, Halgren E,
2001, 85:1969-1985 Cash SS, Hatsopoulos NG, Destexhe A: Dynamic balance of
first paper having shown the dynamics of slow-wave sleep intracellularly excitation and inhibition in human and monkey neocortex.
in cats, where the Down-states are clearly associated with cellular Sci Rep 2016, 6:23176.
hyperpolarization, while Up-states are intracellularly very close to the
dynamics found during wakefulness. 22. Tononi G, Cirelli C: Sleep and the price of plasticity: from
synaptic and cellular homeostasis to memory consolidation
10. Todorova R, Zugaro M: Isolated cortical computations during and integration. Neuron 2014, 81:12-34.
delta waves support memory consolidation. Science 2019,
366:377-381. 23. Turrigiano GG, Nelson SB: Homeostatic plasticity in the
developing nervous system. Nat Rev Neurosci 2004, 5:97.
11. Le Van Quyen M, Muller LE, Telenczuk B, Halgren E, Cash S,
 Hatsopoulos NG, Dehghani N, Destexhe A: High-frequency  K, Reyes A:
24. De La Rocha J, Doiron B, Shea-Brown E, Josic
oscillations in human and monkey neocortex during the wake- Correlation between neural spike trains increases with firing
sleep cycle. Proc Natl Acad Sci 2016, 113:9363-9368 rate. Nature 2007, 448:802.

www.sciencedirect.com Current Opinion in Physiology 2020, 15:68–73

53
54
 Work 2

Integration, coincidence detection

and resonance in networks of spiking
neurons expressing gamma
oscillations and asynchronous states

Reference: Susin, Eduarda, and Alain Destexhe. "Integration, coincidence detection and
resonance in networks of spiking neurons expressing Gamma oscillations and asynchronous
states." Plos Computational Biology (2021).

French Abstract:

Pendant l’état de veille, l’activité neuronale observée dans le cerveau est typiquement
asynchrone et irrégulière. Cependant, cette activité est également accompagnée de péri-
odes où des oscillations modulent l’activité des neurones. Une de ces oscillations, connues
comme des Ondes Gamma (30-90 Hz), a été associée au traitement de l’information. Dans
cette étude, nous utilisons des modèles informatiques pour étudier comment les circuits
cérébraux génèrent des oscillations d’une manière cohérente avec les enregistrements de
microélectrodes chez l’homme. Nous étudions ensuite comment ces réseaux réagissent aux
stimuli externes, en comparant les états asynchrones avec les états oscillatoires. Ceci est
testé selon plusieurs paradigmes: un mode intégrative où des entrées qui variant lentement
sont progressivement intégrées; un mode de détection de coïncidence où les entrées brèves
sont traitées en fonction de la phase des oscillations; et un mode de résonance où le réseau
est sondé avec des stimuli oscillatoires. Étonnamment, nous constatons que dans tous les
cas, la présence d’oscillations Gamma diminue la réactivité aux stimuli externes. Nous
discutons des implications possibles de cette diminution de réactivité sur le traitement de
l’information et proposons de nouvelles directions pour une exploration plus poussée.

55
PLOS COMPUTATIONAL BIOLOGY

RESEARCH ARTICLE

Integration, coincidence detection and

resonance in networks of spiking neurons
expressing Gamma oscillations and
asynchronous states
Eduarda Susin ID*, Alain Destexhe ID

Institute of Neuroscience (NeuroPSI), Paris-Saclay University, Centre National de la Recherche Scientifique

(CNRS), Gif-sur-Yvette, France

a1111111111 * [email protected]
a1111111111
a1111111111
a1111111111
a1111111111 Abstract
Gamma oscillations are widely seen in the awake and sleeping cerebral cortex, but the
exact role of these oscillations is still debated. Here, we used biophysical models to examine
how Gamma oscillations may participate to the processing of afferent stimuli. We con-
OPEN ACCESS structed conductance-based network models of Gamma oscillations, based on different cell
Citation: Susin E, Destexhe A (2021) Integration, types found in cerebral cortex. The models were adjusted to extracellular unit recordings in
coincidence detection and resonance in networks
humans, where Gamma oscillations always coexist with the asynchronous firing mode. We
of spiking neurons expressing Gamma oscillations
and asynchronous states. PLoS Comput Biol 17(9): considered three different mechanisms to generate Gamma, first a mechanism based on
e1009416. https://doi.org/10.1371/journal. the interaction between pyramidal neurons and interneurons (PING), second a mechanism
pcbi.1009416 in which Gamma is generated by interneuron networks (ING) and third, a mechanism which
Editor: Arvind Kumar, Royal Institute of relies on Gamma oscillations generated by pacemaker chattering neurons (CHING). We
Technology (KTH), SWEDEN find that all three mechanisms generate features consistent with human recordings, but that
Received: May 4, 2021 the ING mechanism is most consistent with the firing rate change inside Gamma bursts
Accepted: September 2, 2021 seen in the human data. We next evaluated the responsiveness and resonant properties of
these networks, contrasting Gamma oscillations with the asynchronous mode. We find that
Published: September 16, 2021
for both slowly-varying stimuli and precisely-timed stimuli, the responsiveness is generally
Peer Review History: PLOS recognizes the
lower during Gamma compared to asynchronous states, while resonant properties are
benefits of transparency in the peer review
process; therefore, we enable the publication of similar around the Gamma band. We could not find conditions where Gamma oscillations
all of the content of peer review and author were more responsive. We therefore predict that asynchronous states provide the highest
responses alongside final, published articles. The responsiveness to external stimuli, while Gamma oscillations tend to overall diminish
editorial history of this article is available here:
responsiveness.
https://doi.org/10.1371/journal.pcbi.1009416

Copyright: © 2021 Susin, Destexhe. This is an

open access article distributed under the terms of
the Creative Commons Attribution License, which
permits unrestricted use, distribution, and Author summary
reproduction in any medium, provided the original
author and source are credited. In the awake and attentive brain, the activity of neurons is typically asynchronous and
irregular. It also occasionally displays oscillations in the Gamma frequency range (30–90
Data Availability Statement: Codes written in
support of this publication are publicly available at
Hz), which are believed to be involved in information processing. Here, we use computa-
ModelDB: http://modeldb.yale.edu/267039. tional models to investigate how brain circuits generate oscillations in a manner

PLOS Computational Biology | https://doi.org/10.1371/journal.pcbi.1009416 September 16, 2021 1 / 33

56
PLOS COMPUTATIONAL BIOLOGY Integration, coincidence detection and resonance in networks of spiking neurons

Funding: A.D was supported by the Centre National

de la Recherche Scientifique (CNRS) and the consistent with microelectrode recordings in humans. We then study how these networks
European Community (Human Brain Project, respond to external input, comparing asynchronous and oscillatory states. This is tested
H2020-785907). E.S. acknowledges a PhD
according to several paradigms, an integrative mode, where slowly varying inputs are pro-
fellowship from the École des Neurosciences de
Paris (ENP) and from the Fondation pour la
gressively integrated, a coincidence detection mode, where brief inputs are processed
Recherche Médicale (FRM) - grant according to the phase of the oscillations, and a resonance mode where the network is
FDT202012010566 - and the financial support probed with oscillatory inputs. Surprisingly, we find that in all cases, the presence of
from La Fondation des Treilles. The funders had no Gamma oscillations tends to diminish the responsiveness to external inputs. We discuss
role in study design, data collection and analysis, possible implications of this responsiveness decrease on information processing and pro-
decision to publish, or preparation of the
pose new directions for further exploration.
manuscript. CNRS: www.cnrs.fr Human Brain
Project: www.humanbrainproject.eu ENP: www.
paris-neuroscience.fr FRM: www.frm.org La
Fondation des Treilles: https://www.les-treilles.
com/la-recherche/le-prix-jeune-chercheur/.

Competing interests: The authors have declared Introduction

that no competing interests exist. Gamma oscillations appear in many brain states and brain regions [1] and are detectable
mostly from the local field potential (LFP) as oscillations in the 30–90 Hz frequency range.
During sensory responses, oscillations in this frequency range were initially proposed to serve
as a mechanism for coordination of neural activity among cells coding for different aspects of
the same stimulus [2–5]. Strengthening of synaptic input due to temporal summation led to
the hypothesis that Gamma synchrony was necessary to effectively transmit specific sets of
information across cortical networks in the very noisy conditions in which the brain operates.
This concept was later expanded by proposing that synchronous Gamma also engages inhibi-
tion in target networks. Phase-locked inhibition creates strong suppression around the excit-
atory drive and creates windows of low and high neuronal excitability. Such observations led
to hypotheses that Gamma oscillations are important for information processing and coding.
The most popular theories are the Binding-by-synchronization Hypothesis [4, 5], the Phase
Coding Theory [6, 7], the Communication Through Coherence Theory [8, 9] and Communi-
cation through Resonance Theory [10].
An alternative hypothesis, instead of relying on oscillations for efficient cortical communi-
cation, posits that desynchronized states are optimal for the transfer of signals between cortical
networks [11, 12]. Desynchronized states, called Asynchronous-Irregular (AI) [13] because of
its features, are characterized in cortical cells in vivo by irregular firing with very weak correla-
tions and stationary global activity [14–18]. This type of activity can be modeled by networks
with balanced excitatory and inhibitory inputs [19].
In the present work, we aim at testing these two discrepant points of view using computa-
tional models. We take advantage of previously published electrophysilogical data, measured
extracellularly in human temporal cortex [20, 21], to characterize the behavior of individual
neurons during Gamma oscillations in resting awake states, and to compare such experimental
features to spiking neural networks generating Gamma. We exploit different network struc-
tures to investigate three well-known mechanisms of Gamma generation [22–27]: either by the
exclusive interaction between inhibitory neurons [Interneuron Gamma (ING)] or by the inter-
action of inhibitory and excitatory neurons via Pyramidal-Interneuron Gamma (PING) or via
Chattering Induced Gamma (CHING). First we compare to what degree each mechanism can
reproduce the observed experimental features of human Gamma oscillations and what are the
specificities of each mechanism, in the way neurons behave during Gamma. Subsequently,
we examine network responsiveness due to three types of stimulus: Gaussian slowly-varying
inputs (integration mode), precisely-timed Gaussian inputs (coincidence detection mode) and a
sinusoidal varying Poissonian input (resonance).

PLOS Computational Biology | https://doi.org/10.1371/journal.pcbi.1009416 September 16, 2021 2 / 33

57
PLOS COMPUTATIONAL BIOLOGY Integration, coincidence detection and resonance in networks of spiking neurons

Materials and methods

Neuron and network models
Each of the three networks developed in this work uses the Adaptive Exponential Integrate-
And-Fire Model (Adex) [28] for its neural units. In this model, each neuron i is described by its
membrane potential Vi, which evolves according to the following equations:
� �
dVi ðtÞ ðVi ðtÞ Vth Þ
C ¼ gL ðVi EL Þ þ gL Dexp wi ðtÞ ISyni ðtÞ
dt D

ISyni ðtÞ ¼ gEi ðtÞðVi ðtÞ EE Þ þ gIi ðtÞðVi ðtÞ EI Þ

ð1Þ
dgE;Ii ðtÞ X
tE;I ¼ gE;Ii ðtÞ þ QE;Ii dðt tk Þ
dt k

dwi ðtÞ X
twi ¼ aðVi ðtÞ EL Þ wi ðtÞ þ b dðt tj Þ
dt j

where C is the membrane capacitance, gL is the leakage conductance, EL is the leaky membrane
potential, Vth is the effective threshold and Δ is the threshold slope factor. The synaptic current
(ISyni (t)) received from other neurons to neuron i is taken into account as conductance based:
every time a presynaptic neuron spikes at time tk, the excitatory (gEi ) or the inhibitory (gIi ) syn-
aptic conductance increase by a discrete amount QE or QI (excitatory or inhibitory synaptic
strength), depending on the nature of the presynaptic neuron. Synaptic conductances subse-
quently decay exponentially with a time constant τE or τI. EE and EI are the reversal potential
of excitatory (EE) and inhibitory (EI) synapses. The ∑k runs over all the presynaptic excitatory
or inhibitory neurons spike times. During the simulations, the equation characterizing the
membrane potential Vi is numerically integrated until a spike is generated. Formally this hap-
pens when Vi grows rapidly toward infinity. In practice, the spiking time is defined as the
moment in which Vi reaches a certain threshold (Vth). When Vi = Vth the membrane potential
is reset to Vrest, which is kept constant until the end of the refractory period Tref. After the
refractory period the equations start being integrated again. The adaptation current is
described by the variable wi. It increases by an amount b every time neuron i emits a spike at
times tj and decays exponentially with time scale τw. The parameter a indicates the subthresh-
old adaptation.
Three types of cells were used in our models: Regular Spiking Cells (RS), Chattering Cells
(Ch) and Fast Spiking Cells (FS). The cell specific activities are displayed in Fig 1 and their
parameters are indicated in Table 1.
Each of the three developed networks are composed of N = 25000 neurons, 80% excitatory
and 20% of inhibitory. All neurons are connected randomly. Additionally to recurrent connec-
tions, each neuron receive an external drive (noise). This noise was implemented as NExt =
20000 independent and identically distributed excitatory Poissonian spike trains with a spiking
frequency μExt, being sent to the network with a 2% probability of connection. These spike
trains were computed inside of the synaptic current term Isyn(t), by means of a discontinuous
increase of the excitatory synaptic conductance gE by an amount QExt (at every spike time).
This type of implementation adds to the network a low degree of correlation, since some neu-
rons share the same drive. Nevertheless, this extra correlation does not affect our results,
which kept being qualitatively the same when a drive with no correlations was applied. The

PLOS Computational Biology | https://doi.org/10.1371/journal.pcbi.1009416 September 16, 2021 3 / 33

58
PLOS COMPUTATIONAL BIOLOGY Integration, coincidence detection and resonance in networks of spiking neurons

Fig 1. Neuronal response to an external current. A: External drive fluctuation. External current, in each neuron, varied from 0 to 0.5 nA
in a linear way, was kept constant for 500 ms, subsequently decreasing to 0 nA in a linear way. B: Isolated RS cell in response to the
external drive presented in A. C: Isolated FS cell in response to the external drive presented in A. D: Activity of one Ch cell, in a network
exclusively composed of 1000 Ch cells connected randomly with a probability of 2%.
https://doi.org/10.1371/journal.pcbi.1009416.g001

patterns of connection and neuron type composition of each network model, as well as the
specific values of Poissonian stimulation (μExt and QExt), are described bellow.
• PING Network: It is composed of 25000 Adex neurons (20000 excitatory Regular Spiking
and 5000 inhibitory Fast Spiking cells). All neurons are connected randomly with a probabil-
ity of connection of 2%. All synapses are delayed by a time delay of 1.5 ms. The synaptic
excitatory (inhibitory) time scales are τE = 1.5 ms (τI = 7.5 ms), with synaptic strengths of

PLOS Computational Biology | https://doi.org/10.1371/journal.pcbi.1009416 September 16, 2021 4 / 33

59
PLOS COMPUTATIONAL BIOLOGY Integration, coincidence detection and resonance in networks of spiking neurons

Table 1. Specific neuron model parameters.

Parameter RS FS Ch
Vth -40 mV -47.5 mV -47.5 mV
Δ 2 mV 0.5 mV 0.5 mV
Tref 5 ms 5 ms 1 ms
τw 500 ms 500 ms 50 ms
a 4 nS 0 nS 80 nS
b 20 pA 0 pA 150 pS
C 150 pF 150 pF 150 pF
gL 10 nS 10 nS 10 nS
EL -65 mV -65 mV -58 mV
EE 0 mV 0 mV 0 mV
EI −80 mV −80 mV −80 mV
Vrest -65 mV -65 mV -65 mV
https://doi.org/10.1371/journal.pcbi.1009416.t001

QE = 5 nS (QI = 3.34 nS). Synaptic time scales were chosen accordingly to the parameter
search indicated in S1 Fig. For Gamma activity, the network was stimulated with an external
noise of μExt = 3 Hz and QExt = 4 nS. For an activity similar to an Asynchronous and Irregu-
lar activity (AI-like), the network was stimulated with an external noise of μExt = 2 Hz and
QExt = 4 nS.
• Asynchronous and Irregular (AI) Network: The AI Network was used in this work as one
of the building blocks for the ING and the CHING Network. It is composed of 25000 neu-
rons (20000 excitatory Regular Spiking and 5000 inhibitory Fast Spiking). All neurons are
connected randomly with a probability of connection of 2%. All synapses have synaptic
strengths of QE = 1 nS or QI = 5 nS, and are delayed by a time delay of 1.5 ms. This network,
independently of the strength of the the external noise, can not generate Gamma rhythms.
This is the case because the chosen synaptic excitatory and inhibitory time scales (τE = τI = 5
ms) are in a region of the parameter space in which the regime is asynchronous and irregu-
lar. See S1 Fig. Because of this feature, the AI Network was used as a control to study network
responsiveness (see Results section).
• Gamma Network: The Gamma Network was used in this work as one of the building blocks
for the ING Network. It is composed of 1000 inhibitory Fast Spiking neurons, highly con-
nected between each other. All neurons are connected randomly with a probability of con-
nection of 60%. All synapses have synaptic strengths of QI = 5 nS and synaptic time constant
of τI = 5 ms, and are delayed by a time delay of 1.5 ms. This network is capable of generating
oscillations by its own due to the exclusive presence of inhibitory neurons excited by an
external drive [29, 30]. Low oscillation frequencies in the Gamma range (�70 Hz) are possi-
ble thanks to the high connectivity patterns used (60%). S2 Fig displays the parameter space
of network connectivity vs. inhibitory synaptic strengths for this network. The parameters
chosen in our simulations (p = 60% and QI = 5 nS) are indicated.
• ING Network: The ING Network is constructed as a mixture of AI network with the Gamma
Network. It is composed of 25000 neurons: 20000 RS and 4000 FS from the AI network plus
1000 FS neurons from the Gamma Network. The Fast Spiking neurons in the original AI
network and the ones in the Gamma Network share all the same parameters of FS cells in
Table 1. The only difference among them is their pattern of connectivity. To make it clear,
we call as FS2, the FS neurons that were part of the Gamma Network, and we keep calling as
FS the ones that were part of the AI Network. In the ING Network, FS2 cells send and receive

PLOS Computational Biology | https://doi.org/10.1371/journal.pcbi.1009416 September 16, 2021 5 / 33

60
PLOS COMPUTATIONAL BIOLOGY Integration, coincidence detection and resonance in networks of spiking neurons

random connections to RS neurons with a probability of 15%, FS2 cells send random con-
nections to FS neurons with a probability of 15% while FS cells send random connections to
FS2 neurons with a probability of 3%. This combination of the Gamma network with the AI
Network allows the oscillation frequency to slow down further, reaching � 55 Hz. All synap-
ses have synaptic strengths of QE = 1 nS or QI = 5 nS and synaptic time scales of τE = τI = 5
ms. Synapses are delayed by a time of 1.5 ms. For Gamma activity the network was stimu-
lated with an external noise of μExt = 3 Hz, while for Asynchronous and Irregular activity,
the network was stimulated with an external noise of μExt = 2 Hz. The external noise used
had a synaptic strength of QExt = 0.9 nS.
• CHING Network: The CHING Network is constructed the same way as the AI network, with
the difference that 5% of the RS cells were replaced by Chattering Cells (Ch). This way, the
CHING Network is composed of 25000 neurons: 19000 RS, 1000 Ch and 5000 FS. All cells in
the network are randomly connected to each other with a probability of 2%. All synapses
have synaptic time scales of τE = τI = 5 ms and are delayed by a time delay of 1.5 ms. Excit-
atory synapses have synaptic strengths of QE = 1 nS, while inhibitory synapses from FS cells
to Ch or to RS have synaptic strengths of QI = 7 nS. Synapses from FS to FS have synaptic
strengths of QI = 5 nS. The network receives external noise with synaptic strength of QExt = 1
nS in excitatory cells (RS and Ch) and QExt = 0.75 nS in FS cells. For Gamma, external noise
of μExt = 2 was used, while for Asynchronous and Irregular activity, μExt = 1 Hz.

Simulations
All neural networks were constructed using Brian2 simulator [31]. All equations were numeri-
cally integrated using Euler Methods and dt = 0.1 ms as integration time step. The codes for
each one of the three developed networks are available at ModelDB platform: http://modeldb.
yale.edu/267039.

LFP model
To model the LFP generated by each of the three developed networks, we used a recent method
developed by [32]. This approach calculates the contribution of individual neurons to the LFP
by means of the convolution of individual neuron spike trains (generated by the networks)
with a phenomenological Kernel K, which had its parameters fitted from unitary LFPs (the
LFP generated by a single axon, uLFP) measured experimentally [32]. Each neuron spike train
is convoluted with a particular Kernel Kp that depends on the particular neuron position x~p in
a 2-D space.
2
Kp ð~
x ; tÞ ¼ Að~
x Þexp½ ðt tpick Þ =ð2s2 Þ�

tpick ¼ x x~p j=va

t0 þ d þ j~ ð2Þ

Að~
xÞ ¼ x x~p j=l�
A0 exp½ j~

in which σ is the standard deviation in time, tpick is the peak time of the uLFP, t0 is the time of
the spike of a particular cell p, d is a constant delay, va is the axonal speed, and j~x x~p j is the
distance between the position of particular cell (~ xp ) and the position of the electrode (~
x ). Að~
xÞ
gives the space-dependent amplitude, in which A0 is the maximal amplitude, and λ is the
space constant of the decay. These parameters were estimated separately for excitatory and

PLOS Computational Biology | https://doi.org/10.1371/journal.pcbi.1009416 September 16, 2021 6 / 33

61
PLOS COMPUTATIONAL BIOLOGY Integration, coincidence detection and resonance in networks of spiking neurons

inhibitory contributions (KpE and KpI ) [32, 33]. The LFP, at a particular electrode position ~
x , is
given by the sum of all individual neuron contributions:
LFPð~
x ; tÞ ¼ ð3Þ
! !
XZ p
X XZ p
X
K ð~
x; t
E tÞ dðt j
t Þ dt þ
p K ð~
Ix; t tÞ dðt j
t Þ dt
p ð4Þ
p j p j

In which ∑p runs over neurons p, and ∑j runs over all spike times of neuron p. To be able to
apply this method to our simulations (which don’t presume any neuronal localization in
space), we randomly displaced the network neurons in 2-D grid, assuming that the electrode
was displaced on its center and was measuring the LFP in the same layer as neuronal soma.
The program code of the kernel method is available in ModelDB (http://modeldb.yale.edu/
266508), using python 3 or the hoc language of NEURON.

Detection of Gamma rhythms and Gamma phase

In both, experimental and simulated signals, Gamma rhythms were detected by means of the
Hilbert transform of the band-filtered LFP. The identification of Gamma bursts was done sepa-
rately for each electrode. We considered as Gamma bursts periods in which the amplitude of
Hilbert Transform envelope (absolute value) differed from the mean, by at least 2 standard devi-
ations for the experimental data, and by at least 1 standard deviation for the numerical ones, for
a minimum duration of 3 Gamma cycles. This criteria were not enough to identify all Gamma
bursts (some Gamma bursts were ignored). On the other hand, no false positives were included
in the analysis. All the Gamma bursts automatically identified by the algorithm were individu-
ally confirmed visually. The oscillation phase was acquired using the angle of the imaginary
part of the transform. The LFP was band-pass filtered in the band of 30–50 Hz (unless indicated
otherwise). To band-pass the LFP signals, we used a FIR (Finite Impulse Response) filter using
the Kaiser window method with a 60 dB stop-band attenuation and a 5Hz width from pass to
stop transition [34]. To implement the filter we used the following functions from the Python-
based ecosystem Scipy: signal.kaiserord, signal.lfilter and signal.firwin [35].

Spike-LFP phase-locking
Every time a Gamma period was identified, in both experimental and simulated signals, the
spiking times of each neuron was stored and compared to the Gamma rhythm phase. This
information allowed the construction of the phase distribution of each neuron. For the experi-
mental data, considering that the identification of Gamma bursts was done separately for each
electrode, neurons measured in particular electrode, had their phases and firing rates analyzed
exclusively with respect to the rhythm measured in this electrode. Neuron phases were calcu-
lated from -π to π. In this way neurons with negative phases should be interpreted as spiking
preferentially before than neurons with positive phases. The phase distribution of each neuron
was tested for circular uniformity using a Bonferroni-corrected Rayleigh test [36, 37]. A neuron
was considered phase-locked if we could reject circular uniformity at P < 0.01. See S3 Fig. Neu-
rons that spiked less then 5 times inside Gamma bursts, or neurons whose electrode measured
less then 1 second of Gamma, in the respective data segment, were classified as inconclusive.

Firing rate change

The average firing rate of each neuron outside Gamma bursts (fout) was computed based in the
total time, excluding the activity inside Gamma bursts and their duration. In accordance, the

PLOS Computational Biology | https://doi.org/10.1371/journal.pcbi.1009416 September 16, 2021 7 / 33

62
PLOS COMPUTATIONAL BIOLOGY Integration, coincidence detection and resonance in networks of spiking neurons

average firing rate inside Gamma bursts (fγ) was calculated based on the total Gamma duration
and the activity occurring exclusively inside Gamma bursts. A neuron was considered to
increase its firing significantly if the observed number of spikes in the measured time was
higher than the percent point function of a 95% Interval of Confidence of a Poissonian distribu-
tion with average firing rate fout. Cells that had firing rates smaller then 0.1 Hz or cells whose
electrode measured less then 1 second of Gamma bursts, in the respective data segment, were
classified as inconclusive. See S4 Fig.

Responsiveness
The level of responsiveness (R) of a network, due to a stimulus (S) in a time window of duration
T, is defined as the difference between the total number of spikes generated by the whole net-
S
work due to a stimulus (Nspikes ) and the total number of spikes generated in the absence of the
stimulus (Nspikes), normalized by the network size (total number of neurons Nn) and the dura-
tion of the time window T.
S
Nspikes Nspikes
R¼ ð5Þ
TNn

Phase-dependent responsiveness
The Phase-dependent responsiveness of a network R(θ), in a time window of duration T, due to
a stimulus S presented to the network in a particular phase θ of the Gamma cycle, is defined as
the difference between the total number of spikes generated by the whole network due to a
S
stimulus at the θ phase, Nspikes ðyÞ, and the total number of spikes generated in the absence of
the stimulus at the θ phase, Nspikes(θ), normalized by the network size (total number of neurons
Nn) and the time window T.
S
Nspikes ðyÞ Nspikes ðyÞ
RðyÞ ¼ ð6Þ
TNn

Human recordings
In one epileptic patient with intractable seizures, 10x10 Neuroprobe silicon multielectrode
arrays (400-μm inter-electrode separation, 1 mm electrode length, Blackrock Microsystems)
were implanted in the middle temporal gyrus (layers II/III). Electrodes were implanted in
regions expected to be removed, and after the monitoring session, the implant area was
excised. The patient consented to the procedure, which was approved by the Massachusetts
General Hospital Institutional Review Board in accordance with the ethical standards of the
Declaration of Helsinki. This data set have already been published previously [20, 21]. Neurons
could be classified through clustering based on the spike shape and functional interactions
(determined using cross-correlograms) [20, 38] as Regular Spiking Cell (RS), putative excit-
atory, and Fast Spiking Cells (FS), putative inhibitory. From 81 electrodes, 91 neurons could
be detected: 23 FS and 68 RS.

Results
We first analyze Gamma oscillations from human recordings, then examine network models
of Gamma oscillations and compare them to the experimental data. Finally, we examine the

PLOS Computational Biology | https://doi.org/10.1371/journal.pcbi.1009416 September 16, 2021 8 / 33

63
PLOS COMPUTATIONAL BIOLOGY Integration, coincidence detection and resonance in networks of spiking neurons

responsiveness and resonant properties of these networks, comparing Gamma and asynchro-
nous states.

Human recordings analysis

In this paper, aiming to constrain our computational models to observed experimental fea-
tures, we extend the human data analysis performed in [20, 21], focusing on awake states. The
data was acquired extracellularly in patients suffering of intractable epilepsy, who had multi-
electrode arrays implanted during therapeutic procedures. The arrays registered simulta-
neously local field potential (LFP) and unit activity. We considered here one patient for which
the recording was very stable, and in which several periods of wakefulness could be analyzed.
In each electrode, Gamma rhythms were identified and neural activity was characterized
with respect to the Gamma cycles. Fig 2A illustrates a specific instant in which Gamma bursts
were observed in most of the electrodes (spiking activity and the respective electrode band-fil-
tered LFP are shown). Gamma rhythms were determined through the Hilbert transform of the
filtered LFP (30–50 Hz). Fig 2B and 2C give an example of how Gamma is detected and how
neural phase with respect to the oscillation is extracted (see Detection of Gamma rhythms and
Gamma phase in Materials and methods Section). The data were acquired during the night.
Five awake periods could be recorded, having a mean duration of 27 minutes, containing on
average 13 seconds of Gamma (Fig 2D). During these periods the patient was in a resting
awake condition.
In accordance with other studies, the spiking activity during Gamma bursts was observed
to be very irregular and close to a Poissonian process, with a spiking frequency much smaller
than the population frequency [21, 39–41]. Moreover, conformable to [21], on average, only
4% of RS cells and 17% of FS cells were Phase-Locked (Fig 2E), with RS cells having a phase
preference later in the cycle than the FS cells (see S5 Fig). Furthermore, by measuring the firing
rate change of each cell inside and outside Gamma bursts (Fig 2F), we encountered on average
47% of FS cells that increased their firing inside Gamma bursts, while only 17% of RS cells did.
These observations suggest that Gamma oscillations modulate spiking activity in two manners:
by means of firing rate increase and by defining time windows were some neurons are more
likely to spike (phase-locking).
Contrary to the intuition that all neurons in a network generating Gamma would be partici-
pating to the rhythm, this analysis indicates that, only a small percentage of neurons has its
activity modulated by the oscillation (either by phase-locking or by firing rate increase). We
call this group of neurons as Gamma participating cells.
To better characterize the non-participation to Gamma rhythms, we followed each cell in
each of the 5 waking periods present in the recordings, searching for behavioral changes. We
observed that in different data segments, different groups of neurons were identified to partici-
pate to Gamma, indicating that the group of Gamma participating cells varies with time (see
S6 Fig). Furthermore, cells that were classified as phase-locked in different data segments, had
their preferred phase changed from one recording to the other (see cells 65 and 22 in S5 Fig).
We called this feature as dynamical phase preference. Fig 3 indicates the individual cell behavior
consistency, that is, how frequently a cell keeps being identified to a certain behavior: either
being phase-locked or to have its firing rate changed inside Gamma bursts in a particular data
segment. Stacked bars of Fig 3A and 3B indicate a color-coded behavior distribution of indi-
vidual neurons, inside of the 5 data segments, with respect to firing rate change and phase-
locking respectively. Neurons are ordered in a way in which inhibitory cells are displayed in
the beginning. Red neuron indexes stand for FS cells and green neuron indexes stand for RS
cells. Fig 3C and 3D depict the distribution among all recorded neurons of each behavior (C:

PLOS Computational Biology | https://doi.org/10.1371/journal.pcbi.1009416 September 16, 2021 9 / 33

64
PLOS COMPUTATIONAL BIOLOGY Integration, coincidence detection and resonance in networks of spiking neurons

Fig 2. Human electrophysiological data. A: Simultaneously recorded LFP and multi-units activity. The Filtered LFP (30–50 Hz) of the 81 electrodes
are shown together with the spiking times of 91 neurons. Some neurons were recorded by the same electrode, which had its LFP duplicated in the
figure. The identification of Gamma bursts was done separately for each electrode. This way, neurons measured in a particular electrode, had their
phases and firing rates analyzed exclusively with respect to the rhythm measured in its respective electrode. Spikes of Fast Spiking (FS) neurons,
presumably inhibitory, are shown in red, and spikes from Regular Spiking (RS) neurons, presumably excitatory, are shown in green. B: Gamma periods
detection. Raw LFP (black), band-pass filtered LFP (yellow) and Hilbert Transform Envelope (red) are shown. Gamma bursts were detected by means
of the deviation from the average of the Hilbert Transform envelope (dashed red line) of at least 2 SDs (dotted red line), with a minimum duration of 3
Gamma cycles. The gray shaded region indicates one example of identified Gamma burst. C: Oscillation Phase extraction. The oscillation phases were
obtained by the angle of the imaginary part of the Hilbert Transform. The phase distributions of each neuron were computed based on the oscillation
phases where each neuron spiked. D: Data organization. Five awake periods could be recorded during one night. Each period had a different total time
duration (yellow bars in minutes) and a different average duration of total Gamma occurrences (orange bars in seconds). Since each electrode was
analyzed individually, the average indicated in the bars is the average among all the electrodes in the respective segment. E: Percentage of neurons
identified as phase-locked in each data segment. The average amount of Phase-locked neurons in the five data segments was of 4% in RS and 17% in FS.
RS neurons are shown in green and FS neuron in red. F: Percentage of neurons that increased their firing during Gamma, in each data segment. The
average amount neurons in the five data segments which increased their firing during Gamma was of 17% in RS and 47% in FS. Same color scheme as
in E.
https://doi.org/10.1371/journal.pcbi.1009416.g002

Firing Rate Increase, D: Phase-Locking). A behavior consistency of zero denotes that the indi-
cated percentage of neurons never presented that behavior, while a behavior consistency of 5
denotes that the indicated percentage of neurons presented that behavior in all 5 data seg-
ments. FS cells tended to participate of Gamma bursts with higher consistency than RS cells.

PLOS Computational Biology | https://doi.org/10.1371/journal.pcbi.1009416 September 16, 2021 10 / 33

65
PLOS COMPUTATIONAL BIOLOGY Integration, coincidence detection and resonance in networks of spiking neurons

Fig 3. Individual neural behavior consistency on human recordings. Stacked bars indicating the color-coded distribution, inside of the 5 data
segments, of individual neural behavior relative to firing rate change (A) and phase-locking (B). Neurons are ordered in a way in which inhibitory
neurons are displayed in the beginning of the graph. Red neuron indexes stand for FS cells and green neuron indexes stand for RS cells. Items C and D
indicate respectively the statistics of the consistency indexes among the recorded neurons for Firing Rate Increase and Phase-Locking.
https://doi.org/10.1371/journal.pcbi.1009416.g003

While 34.8% of FS increased their firing inside Gamma bursts in at least 4 of the 5 data seg-
ments, only 4.4% of RS cells did the same. Moreover 8.7% of FS cells kept being phase-locked
in at least 4 data segments, in comparison to only 1.5% in RS population (see S7 Fig). Likewise,
we call the reader’s attention to the significant number of cells that never increase their firing
rate inside Gamma bursts (Fig 3C, � 40% of the recorded neurons) and to the significant
number of cells that never presented phase-locking (Fig 3D, � 80% of the recorded neurons).
The behavior of individual cells during Gamma is quantified in S8 Fig.
Furthermore, another important aspect to be acknowledge is a possible correlation between
high firing rate cells (inside Gamma bursts) and those cells that show higher phase locking.
Nonetheless, the human data set used in this study is too small to be able to arrive to any con-
clusion. In our analysis cells with high firing rates were observed to be not phase-locked (or
phase-locked), the same way as cells with lower firing rates were observed to phase-locked
(or not phase-locked). See S9 Fig. The same is true if we try do drive conclusions about the co-
occurrences of firing rate increase and phase-locking (see S6 Fig).
In summary our analysis shows that, during Gamma bursts, only a small percentage of the
recorded neurons participate of the rhythm. This participation takes place in two ways: phase-
locking and/or firing rate increase. FS cells presented significant higher level of phase-locking
and firing rate increase in comparison to RS cells. Likewise the level of consistency behavior
were also more marked in FS cells than RS cells. Our analysis further indicates that, the group
of Gamma participating cells changes with time as well as their phase-preference.

Network models of Gamma oscillations

Gamma oscillations have been extensively modeled in the literature with different neuronal
models and networks structures [23, 42]. The low and irregular firing rates observed during
Gamma oscillations have been reproduced in recurrent networks of spiking neurons [13, 30,
43–45] by means of strong recurrent inhibition and strong noise (due to external inputs and/
or due to synaptic disorder). Networks displaying this type of activity are known to be in the

PLOS Computational Biology | https://doi.org/10.1371/journal.pcbi.1009416 September 16, 2021 11 / 33

66
PLOS COMPUTATIONAL BIOLOGY Integration, coincidence detection and resonance in networks of spiking neurons

firing rate regime [30]; in contrast to models fully synchronized, in which neurons behave as
periodic oscillators. In this last regime, known as an spike-to-spike regime, neurons spike at
every cycle (or once every two cycles), with an average firing rate close to the frequency of
oscillatory network activity [46–54]
It is well established, experimentally and theoretically, that inhibition plays a crucial role in
generating Gamma rhythms [21, 23, 29, 42, 55–59]. Nonetheless, it is still controversial [22–
25] whether Gamma oscillations are generated by the exclusively interaction among inhibitory
neurons [Interneuron Gamma (ING)] or via the interaction of inhibitory and excitatory neu-
rons [Pyramidal-Interneuron Gamma (PING)]. Furthermore, a third mechanism, less explored
in the literature, relies on the presence of pacemaker excitatory cells known as Chattering neu-
rons [26, 27]. We named this third mechanism as Chattering Induced Gamma (CHING).
To compare to what degree each of three previously mentioned mechanisms can reproduce
the observed experimental features, and what are the consequences of each mechanism, we
constructed three neural networks working in the firing rate regime, adapted to generate
Gamma by means of ING, PING or CHING. Network and neuronal parameters were chosen
in a way to allow each model to reproduce experimental features as well as possible, with physi-
ologically plausible firing rates and membrane conductance distributions (see S10 and S11
Figs). We call the reader’s attention to the fact that, while networks with a structure similar to
our PING Network have been largely used in the literature, the structures of ING and CHING
Networks were developed exclusively for this study.
Like in previous works [60], all three networks are capable of generating spontaneous
Gamma bursts. These Gamma bursts are controlled by fluctuations of recurrent drive gener-
ated by the network dynamics, which for this reason occur irregularly and in an unpredictable
fashion. However, more predictable Gamma bursts can be obtained by increasing the external
drive (in all three networks). Fig 4 shows the behavior of the three networks when a fluctuation
on the Poissonian input generates Gamma, mimicking the Gamma bursts observed experi-
mentally. Note however that, outside of Gamma bursts (low input amplitude), the networks do
not necessarily display a pure AI state: all three networks display reminiscent low-amplitude
oscillations. In all cases, the firing dynamics remained irregular and with low synchrony, so we
called them AI-like states.
We next performed on the network models an equivalent analysis as in the human data
recordings. Each cell was followed in 5 different simulations containing on average 13 seconds
of Gamma bursts (same duration as in the experimental recordings, mimicking the five experi-
mental data segments) and statistical tests to identify phase-locking and firing rate changes
were performed. Fig 5A, 5B and 5C display respectively the quantification of behavior consis-
tency for PING, ING and CHING Networks. Accordingly to the unit recordings [21], the cells
were generally more depolarized and increased their firing during Gamma. On the other
hand, within the three models, only the ING Network (Fig 5Bc) is capable of describing the
appropriate amount of neurons that increase their firing in different data segments, during
Gamma. The PING and CHING networks predict an over-estimation of this number. The
presence of a sub-population of highly connected inhibitory neurons, capable of generating
Gamma rhythms by their own (see Neuron and Network Models: Gamma Network in Materials
and methods Section), allows the ING Network to provide a compensation for external excit-
atory fluctuations: whenever there is an augmentation of input in the network (generating
Gamma), there is in addition a concomitant augmentation of inhibition thanks to the FS2
population.
In comparison to the experimental data analysis performed previously, all three models are
capable of correctly describing the frequency of re-occurrence of phase-locking inside of a
group of neurons in different data segments. That is, all three models predict the same the

PLOS Computational Biology | https://doi.org/10.1371/journal.pcbi.1009416 September 16, 2021 12 / 33

67
PLOS COMPUTATIONAL BIOLOGY Integration, coincidence detection and resonance in networks of spiking neurons

Fig 4. Neural activity of different Gamma generation mechanisms networks. PING Network (left), ING Network (middle) and CHING Network
(right). A: Scheme of each network structure and pattern of connectivity. B: External Poissonian noise fluctuation generating Gamma bursts. C: Raster
plot of network activity inside and outside Gamma bursts. Only 1000 neurons of each cell type are shown. D: Membrane potential activity of randomly
picked neurons of each type. Pay attention to the well defined subthreshold oscillation exclusively present in the ING Network. E: Simulated LFP (raw—
in black) and its filtered version (yellow).
https://doi.org/10.1371/journal.pcbi.1009416.g004

same intensity of phase-locking consistency as the one observed on the human recordings (Fig
3D). On the other hand, regardless of the mechanisms of Gamma generation, all networks pre-
dict an over estimated phase-locking level (total number of phase-locked neurons per data seg-
ment) (see S12 Fig). With respect to the human data set, the PING and ING networks predict a
comparable level of phase-locking in the excitatory population but an exaggerated level in the
inhibitory population. In contrast, the CHING Network predicts a comparable level of phase-
locking in the inhibitory population but an exaggerated level in the excitatory one. Side by
side, the CHING Network is the one that still captures the best the level of phase-locking in
both populations (excitatory and inhibitory).
The right prediction of phase-locking consistency can be explained by the type of activity
regime in which each network works: the fluctuation-driven regime. Since this regime allows
neurons to spike with low firing rates in an irregular fashion, participating of the global
Gamma oscillation only in certain cycles due to the subthreshold randomness. Nonetheless,
the over-estimation of phase-locking level, indicates that the simple fact of being in the fluctua-
tion-driven regime is not enough to capture all levels of description. We hypothesize that the
network structure play a key role in the way neurons behave during oscillations. Fig 5 illus-
trates how different network structures (different connectivities in the ING Network or differ-
ent neuron types in the CHING Network) influence network activity.

PLOS Computational Biology | https://doi.org/10.1371/journal.pcbi.1009416 September 16, 2021 13 / 33

68
PLOS COMPUTATIONAL BIOLOGY Integration, coincidence detection and resonance in networks of spiking neurons

Fig 5. Individual neural behavior consistency in computational models. A: PING Network. B: ING Network. C: CHING Network. Same
analysis and color codes used in Fig 3. To mimic the five experimental independent data segments in the Human data recordings (Fig 3) on
the network models, five simulations (per model) were performed, containing on average the same amount of total Gamma bursts duration
as in the experimental data (13 seconds). In addition, to match the number of recorded neurons in the experimental data, in the models a
subset of 100 randomly picked neurons were selected in each case.
https://doi.org/10.1371/journal.pcbi.1009416.g005

In the presented human recordings, inhibitory neurons tended to spike earlier in the cycle
than excitatory neurons. Fig 6 shows the phase preference with respect to the Gamma cycle of
all the neurons considered phase-locked in the human data recordings (Fig 6A) and in each of
the three developed networks (Fig 6B, 6C and 6D). The ING and CHING networks predict the

PLOS Computational Biology | https://doi.org/10.1371/journal.pcbi.1009416 September 16, 2021 14 / 33

69
PLOS COMPUTATIONAL BIOLOGY Integration, coincidence detection and resonance in networks of spiking neurons

Fig 6. Phase preference of phase-locked cells. A: Human Data (Data segment 2). B: PING Network Data. C: ING Network Data. D: CHING Network
Data. The preferred phases of each phase-locked cell are displayed in polar graph representation. Note that, since phases were calculated from −π to π
(see Spike-LFP phase-locking in Materials and methods Section), these polar graphs should be interpreted clockwise with time. The vector size gives a
measure of the phase distribution of each cell. Big amplitude vectors indicate very concentrated distributions while small amplitude vectors indicate less
concentrated ones (see S3 Fig). The color of each vector encodes the type of the cell of whom it represents the phase: red (FS), dark red (FS2), green (RS)
and dark green (Ch). Cell number IDs are indicated. Dark colored vectors indicate the average phase among each neuron type and Δθ the phase
difference among them. Data segment 2 presented 43 minutes of recordings, containing 14 seconds of Gamma activity.
https://doi.org/10.1371/journal.pcbi.1009416.g006

same relationship as observed in the human recordings (inhibition preceding excitation) while
the PING Network predicts the opposite. Moreover, in the same way as the human data set (S5
Fig), cells that were classified as phase-locked, have their preferred phase changed from one
simulation to other (dynamical phase preference). We argue that this feature is also a conse-
quence of the fluctuation-driven regime.
The phase relationship between excitation and inhibition is an important aspect to be dis-
cussed, since it has been suggested to be a marker of the type of Gamma generation mecha-
nism [25]. It has been shown theoretically by [45] that, in models composed of conductance
based neurons (neurons that include non-linear spike generation mechanisms on their equa-
tions) the spiking order of excitatory and inhibitory populations depends exclusively on sin-
gle-cell characteristics. Based on their analysis, when the IAMPA/IGABA ratio is the same in
excitatory and inhibitory neurons, excitatory cells tend to follow the inhibitory ones in most of
the physiologically plausible parameter space. On the other hand, when the ratio of excitation
to inhibition is weaker in excitatory cells than in inhibitory ones, excitatory cells tend to pre-
cede inhibitory neurons [30, 45]. In our simulations, the only network in which this theory
can be directly applied (because of the network structure) is the PING Network, in which the
IAMPA/IGABA ratio in excitatory cells is weaker than in inhibitory cells. Interesting discussions
about neural properties and population phase-differences can also be found on [61, 62].
Concluding this section, we showed that network models working in the firing rate regime,
regardless of the mechanism of Gamma generation, can reproduce qualitatively some of the

PLOS Computational Biology | https://doi.org/10.1371/journal.pcbi.1009416 September 16, 2021 15 / 33

70
PLOS COMPUTATIONAL BIOLOGY Integration, coincidence detection and resonance in networks of spiking neurons

most important features of experimental neural activity during Gamma: phase-locking consis-
tency and dynamical phase preference. On the other hand, all models predict an overestima-
tion of the phase-locking levels. Additionally, only the ING Network model was capable of
describing a reasonable level of firing rate increase inside Gamma bursts, as found in the
human recordings. We advocate that just the simple fact of being in the fluctuation-driven
regime is not enough to capture all levels of description of Gamma oscillations, and hypothe-
size that the network structure play a key role in the way neurons behave during oscillations.
Considering that the different types of spontaneous activity exhibited by the three presented
models could greatly influence how the network processes external input, we have investigated
this issue of responsiveness to external input in the next section.

Responsiveness and resonance during Gamma oscillations

Responsiveness. The way information is encoded and processed in the brain is still a
largely investigated enigma. Several ways of encoding information have been considered, such
as firing rates [63, 64], pairwise correlations [65, 66], spike pattern irregularity [67–70] and
spike packets [71], among others. In particular, two main theories have been dominating the
debate: Temporal Coding in which individual neurons encode information by means of precise
spike timings (working as coincidence detectors), and the Rate Coding in which neurons
encode information by means of changes in their spike rates (working as temporal integra-
tors). Regardless of the encoded strategy used to encode information, the way the network is
capable of responding to a certain stimulus is of prime importance. To identify how Gamma
rhythms change the response properties of a network to an external stimulus with respect to
AI, in this section we applied two protocols, investigating the effect of Gamma in both, the
coincidence detection mode and in the integration mode [72, 73].
In the integration mode protocol, we compared how each of the three developed models
responded to slowly-varying inputs (occurring in a time window much bigger than the
Gamma period). In this protocol, each network received Poissonian drive (spikes from an
external network) with firing rates varying in time, in a Gaussian manner, both during
Gamma and AI-like states. The applied Gaussian inputs had a standard deviation of 50 ms,
allowing the stimulus to interact with different Gamma cycles. Several amplitudes of slowly-
varying Gaussian were tested, and the responsiveness of excitatory and inhibitory populations
were measured separately. Responsiveness (see Eq 5) was defined as the difference between the
total number of spikes (in a time window of duration T) generated by the whole network in
the presence and in the absence of the stimulus (normalized by the network size and the time
window duration T).
Fig 7 shows the responsiveness of the PING Network, the ING Network and the CHING Net-
work, when the integration mode protocol was applied. To be able to measure the real impact
of Gamma oscillations on network responsiveness, we used as a control the responsiveness
curves from the AI-Network, in which no oscillation is generated, independently of the level of
external drive. See Neuron and Network Models: AI Network in Materials and methods Section
and S13 Fig. All models, regardless of the mechanism of Gamma generation, were less respon-
sive during Gamma bursts in comparison with their baseline responsiveness during AI-like
states. Furthermore, the responsiveness of a network in a real AI-state (gray curve generated
by the AI Network) is equal or higher to the AI-like responsiveness, in each of the networks,
and is always higher then the Gamma state responsiveness in all cases. In addition, to further
investigate this result, we examined the responsiveness of individual cells (S14 Fig). Due to the
previous finding that only a restricted group of cells participate to Gamma, one could imagine
that there could still be few cells (Gamma participating cells) that would be more responsive,

PLOS Computational Biology | https://doi.org/10.1371/journal.pcbi.1009416 September 16, 2021 16 / 33

71
PLOS COMPUTATIONAL BIOLOGY Integration, coincidence detection and resonance in networks of spiking neurons

Fig 7. Network responsiveness to a Gaussian input with varying amplitude. The responsiveness, in different states (Gamma and AI-
like), was measured in the three developed networks and compared to the responsiveness of the AI Network as a control. A:
Responsiveness protocol scheme for Gamma state. A Gamma burst is generated due to fluctuations of the external drive (black dashed
line). During the Gamma activity, a Gaussian input (green line) is applied. The total number of spikes due to the stimulus, in time
window of 500 ms, is measured. To measure the total number of spikes in the absence of the stimulus, another drive fluctuation is created
generating Gamma. The total number of spikes inside of a time window of 500 ms is measured again (this time, without the Gaussian
input). Only the situation in response to a stimulus is depicted in the scheme. This procedure was repeated 100 times per each Gaussian
amplitude input. To measure responsiveness in AI and AI-like states no drive fluctuation was applied (the black dashed line in the figure

PLOS Computational Biology | https://doi.org/10.1371/journal.pcbi.1009416 September 16, 2021 17 / 33

72
PLOS COMPUTATIONAL BIOLOGY Integration, coincidence detection and resonance in networks of spiking neurons

was kept constant). B: Input Amplitude Variation. The stimulus consisted of a Gaussian fluctuation in the firing rate of the external drive.
The Gaussian amplitude varied from 0.05 Hz to 2.5 Hz (step of 0.05 Hz) and had a standard deviation of 50 ms. Items C, D and E display
respectively the responsiveness of the PING Network, the ING Network and the CHING Network, inside Gamma bursts (green for
excitatory cells, red for inhibitory cells), and outside Gamma bursts—AI-like activity—(black for both types of cells). Every point
corresponds to the average responsiveness measured in 100 simulations. Standard error of the mean are indicated by the shaded region
around each curve. The responsiveness of the AI-Network was added as a control in each case (gray curve in C, D and E). To implement
the responsiveness protocol the AI-Network received a constant drive with μExt = 3 Hz, in addition to the Gaussian inputs. See S13 Fig.
https://doi.org/10.1371/journal.pcbi.1009416.g007

while all others (Gamma non-participating cells) would be less responsive, leading to a yet
overall less prominent responsiveness. Nonetheless, S14 Fig shows the contrary. All cells seem
to follow the same decrease of responsiveness during Gamma oscillations, and we found no
evidence that some subset of cells would be more responsive, for all amplitudes tested.
In the coincidence detection mode protocol, the responsiveness at different Gamma
phases was measured. To do this, precisely-timed inputs (occurring in a time window much
smaller than the Gamma period) were applied and related to Gamma cycles in each of the
three developed networks. In this protocol the amplitude of the stimulation was kept con-
stant, while the time of the application of the Gaussian stimulus changed with respect to the
phase of the Gamma oscillation. This procedure allowed each network to be stimulated at
different Gamma phases (see S15 Fig). Fig 8 indicates the network response of excitatory
cells per Gamma phase, in different states: Gamma state (blue), AI-like (black) and AI-
like modulated by a control external current oscillating at Gamma frequency (gray). All
responses were normalized by the average response of AI-like states without external cur-
rent modulation (black).
AI-like states, when modulated by an external oscillatory current, displayed, in all network
models, preferred phases in which the network response was higher in comparison to the non-
modulated AI-like state. This constitutes an important control, because the external current
creates periods of higher and lower excitability in the network, which is translated in a phase-
dependent response (as shown by the gray curves in Fig 8). Likewise, when generating
Gamma, our models (PING and ING) demonstrate an equivalent type of phase-dependence
response (even-tough with a narrow amplitude range). On the other hand, in agreement with
the integration mode protocol, our simulations show that the responsiveness during Gamma
states at all phases are less or equal to that during AI-like states.
Resonance. In Physics, when dealing with an oscillatory system, one of the first features
to be explored is its resonant properties. In general, resonance describes the phenomenon of
increased amplitude in a system, that occurs due to the application of an oscillatory stimulus
whose frequency is equal or close to the natural frequency of the system. It has been shown
experimentally that this phenomenon can also be observed in inhibitory [56] and excitatory
[74] neuronal populations. Furthermore, theoretical studies [75] have shown that resonance is
a fundamental property of spiking networks composed of excitatory and inhibitory neurons.
Resonance has also been proposed as a mechanism to gate neuronal signals [76] and to com-
municate information [10].
We tested the resonant properties of each of our networks in AI-like and Gamma states. In
this protocol, each network received Poissonian drive with firing rates varying in time in a
sinusoidal manner, with different frequencies (Fig 9A). Fig 9B, 9C and 9D depict, for each fre-
quency and oscillation phase, the average number of spikes per RS neuron and time bin, dur-
ing Gamma and AI-like states, for the PING, ING and CHING Networks. All values were
normalized by the average firing inside of each state to exclude the state dependent firing rate
level (which is higher on Gamma). To enhance the comprehension of the responsive proper-
ties of each network, a linear version of the color maps depicted in Fig 9 (amplitude vs. phase)

PLOS Computational Biology | https://doi.org/10.1371/journal.pcbi.1009416 September 16, 2021 18 / 33

73
PLOS COMPUTATIONAL BIOLOGY Integration, coincidence detection and resonance in networks of spiking neurons

Fig 8. Phase-dependent network response. A: External oscillatory current applied at AI-like state as function of its oscillation phases (gray curve)
and the filtered LFP measured during Gamma states as function of its oscillation phases (blue curve). All networks received a current oscillating
from 0 to 0.1 nA in a sinusoidal manner with a Gamma frequency Fγ. To match the Gamma oscillation frequency generated by each network, the
frequency of the external current applied to PING and CHING networks was Fγ = 40 Hz, while the one applied to ING network was Fγ = 55 Hz. The
LFP depicted is the one from PING network. ING and CHING also displayed a similar LFP pattern. B: PING Network phase-dependent response C:
ING Network phase-dependent response. D:CHING Network phase-dependent response. The phase-dependent network response was calculated
according to Eq 6, in a time window of duration T equal to one Gamma cycle (T = 25ms for the PING and CHING Networks and T = 18ms for

PLOS Computational Biology | https://doi.org/10.1371/journal.pcbi.1009416 September 16, 2021 19 / 33

74
PLOS COMPUTATIONAL BIOLOGY Integration, coincidence detection and resonance in networks of spiking neurons

ING). Responses measured inside AI-like activity (outside Gamma bursts) are shown in black, and in gray when the networks received a
supplementary oscillatory external current. Responses measured inside Gamma bursts are displayed in blue. All curves were normalized by the
average response inside AI-like activity without external current modulation. Solid lines indicate the average, and the shaded region indicates the
standard error of the mean. The curves were calculated based on the output of 12000 simulations (120 positions of the Gaussian stimulus in 100
numerical seeds for external Poissonian drive). The Gaussian stimulus used had an amplitude of 50 Hz and standard deviation of 1 ms.
https://doi.org/10.1371/journal.pcbi.1009416.g008

is provided in S16 Fig. In addition, S17 Fig depicts the resonant properties in other cell types
(FS, FS2 or Ch) during Gamma state for each one of the networks.
We observe that, in both AI-like and Gamma states, all models display resonant properties
around the Gamma band, with the main difference in between these two states being a shift of
the resonance frequency center. In this protocol we detect a similar level of responsiveness per
phase (reflected in the measured number of spikes per time bin) in AI and Gamma, indicating
that networks receiving oscillatory inputs have the same latent potential to resonate at Gamma
ranges regardless if they are displaying AI or Gamma oscillations. One should note that each
model presents its own particularities. While the PING network presents just a shift of the cen-
ter frequency of resonance, the ING network presents an enlarged potential of resonance in AI
(in addition to the frequency shift). During AI, the ING network presents an equal resonance
in several bands other than Gamma. Moreover, when a Gamma oscillation is triggered in this
network, this resonance is shrunk and becomes more concentrated in the Gamma band. The
CHING network, on the other hand, presents a strong resonance in the 15–25 Hz frequency
range during AI, while during Gamma this resonance is lost.
Concluding this section, we investigated three dynamical properties (Responsiveness,
Phase-dependent-responsiveness and Resonance) in different states (AI-like and Gamma) of
each of the three developed networks. We encounter that, regardless of Gamma generating
mechanism (PING, ING or CHING), the network responsiveness, in both coincidence detec-
tion and integrative mode, is decreased at Gamma states with respect to AI. On the other hand,
the resonant properties around the Gamma band in all networks did not change significantly
from one state to the other. The main resonant properties changes between AI and Gamma
states in each model were most prominent around other bands. The implications of these
observations on the role Gamma rhythms in neural computations and information transfer
will be discussed in the next section.

Discussion
In this paper, we have examined the genesis and responsiveness of Gamma oscillations con-
strained by human recordings. We analyzed Gamma oscillations from previous studies [20,
21], where the recordings were stable, and in which RS and FS cells were discriminated. We
compared the results of this analysis to conductance-based network models implementing
three different mechanisms that were proposed for Gamma oscillations, PING, ING and
CHING. We next examined these three networks with respect to their responsiveness and res-
onance to external inputs. We discuss these aspects below.

Human data analysis

Compared to a previous analysis of the cellular correlates of Gamma oscillations [21], we con-
firm here the low level of cellular engagement and a greater participation of FS cells during
Gamma, either through phase-locking or through firing rate increase. FS cells not only pre-
sented a higher percentage of phase-locking or firing rate increase during Gamma, but they
also presented a more consistent behavior compared to RS cells which were much more vari-
able. Our analysis further indicates that, the group of Gamma participating cells changes with

PLOS Computational Biology | https://doi.org/10.1371/journal.pcbi.1009416 September 16, 2021 20 / 33

75
PLOS COMPUTATIONAL BIOLOGY Integration, coincidence detection and resonance in networks of spiking neurons

Fig 9. Resonant properties of computational models. A: Representation of external Poissonian noise varying in time in a
sinusoidal manner around μnoise. In this protocol sinusoidal frequencies varied from 5 Hz to 100 Hz (step of 5 Hz). Two oscillatory
frequencies are depicted: 20Hz (blue) and 40 Hz (black), together with their phases (second axis) and time bins (vertical line). For
all frequencies the average Poissonian noise (μnoise) was kept the same, varying from μnoise − Δnoise and μnoise + Δnoise. The bins were
chosen in a way in which the oscillatory phases (from -π to π) were divided into 25 intervals (for all frequencies), resulting in time
bins of different duration for each oscillatory frequency. B: Resonant properties of PING Network. C: Resonant properties of ING
Network. D: Resonant properties of CHING Network. The color maps displayed in B, C and D depict, for each oscillatory
frequency and oscillation phase, the average number of spikes per RS neuron per time bin, during Gamma and AI-like states. All

PLOS Computational Biology | https://doi.org/10.1371/journal.pcbi.1009416 September 16, 2021 21 / 33

76
PLOS COMPUTATIONAL BIOLOGY Integration, coincidence detection and resonance in networks of spiking neurons

values were normalized by the average firing inside of each state to exclude the state dependent firing rate level (which is higher on
Gamma). Δnoise = 0.5 Hz in all network models but μnoise varied in each case. For AI, in PING and ING Networks μnoise = 2 Hz and
in CHING Network μnoise = 1 Hz, while for Gamma, μnoise = 3 Hz in in PING and ING Networks and μnoise = 2 Hz in CHING
Network.
https://doi.org/10.1371/journal.pcbi.1009416.g009

time as well as their phase-preference. The analysis performed on this work is very qualitative,
since it was based on a single patient. Nonetheless, this very sparse participation of RS and FS
cells during Gamma was seen in different patients, and the same was observed in monkey for
beta oscillations [21].

Responsiveness
The occurrence of Gamma rhythms have been correlated with conscious perception [77–81]
and several authors support these rhythms as being a suitable marker of consciousness. On the
other hand, it has been proposed that the Asynchronous and Irregular activity, observed dur-
ing awake and aroused states, due to its specific responsiveness properties, is an ideal setting
for integrating multiple external inputs [12]. In support of this, it was concluded in a review
that asynchronous states constitute the most reliable correlate of conscious states [82].
Previous work [12] has compared the responsiveness of a fully synchronized network
(spike-to-spike regime) with a network in AI state, showing that the AI state is the best state to
integrate multiple external inputs. It was also shown that, in rate-based networks, the most
chaotic states could display the highest responsiveness, as measured using Shannon informa-
tion [83]. In the present work, we compared the responsive properties of AI state with Gamma
states generated by means of three different mechanism: PING, ING and CHING. Each of
these networks were submitted to two types of inputs. First, a slowly-varying input integrated
by the population of neurons over a substantial period of time (integrative mode). Second, we
examined precisely-timed inputs, occurring in a time window smaller than the Gamma period
(coincidence detection mode). For the integrative mode, we systematically found that the
Gamma oscillations yielded less responsiveness than the AI-like states and even lesser respon-
siveness than real AI states (generated by the AI Network, used as a control). In the coincidence
detection mode, we found that the response was only weakly modulated by the phase of the
Gamma. This was assessed by comparing the Gamma oscillation to a sinusoidal control input,
in which case the response was clearly phase-dependent. In agreement with the integrative
mode, the responsiveness measured in the coincidence detection mode protocol was generally
higher for the AI-like states. In addition, in the coincidence detection mode, among the three
models, the ING Network is the only one that presents a similar responsiveness between
Gamma and AI states, which stresses again the importance of network topology on networks
behaviors.
A smaller responsiveness during Gamma states is somehow surprising since neurons are in
general more depolarized in this state and additionally increase their firing, as we showed in
our data analysis. On the other hand this observation is intuitively easy to understand, if we
take into account the fact that Gamma oscillation are composed of successions of periods of
high inhibition, which define time windows in which neurons are less likely to spike. While
during Gamma states, these time windows of high inhibition constrain the times a certain neu-
ron can spike, during AI states neurons can spike at all moments with the same probability.
Indeed, we observed that the response during Gamma oscillations is phase-dependent, while
there is no phase preference during AI states. However, although there was a phase depen-
dence, Gamma oscillations did not provide a preferred phase where the network is more
responsive than during AI states. The fact that higher levels of inhibition during Gamma could

PLOS Computational Biology | https://doi.org/10.1371/journal.pcbi.1009416 September 16, 2021 22 / 33

77
PLOS COMPUTATIONAL BIOLOGY Integration, coincidence detection and resonance in networks of spiking neurons

explain their diminished responsiveness should be testable experimentally using intracellular

recordings in vivo.
Given our model results, what this decrease of responsiveness could be useful for, and
what are the advantages of a higher responsive state in AI? This questions can be approached
in the light of the Phase Coding Theory (PC). This theory was initially formulated with
respect to Theta rhythm [6], but lately extended to Gamma [7]. This theory states that, within
the Gamma cycle, the excitatory input to pyramidal cells is converted into a temporal code
whereby the amplitude of excitation is re-coded in the time of occurrence of output spikes
relative to the cycle [7]. In this view, the cells that are most excited fire earlier in the cycle,
while cells that are not excited enough are prohibited to spike due the new wave of inhibition
composing the cycle. This process can be seen as a winner-take-all phenomena (or more pre-
cisely a few-winners-take-all phenomena, since it involves several neurons neurons) [7].
Such a coding strategy enables transmission and read out of amplitude information within a
single Gamma cycle without requiring rate integration, proving a fast processing and readout
by means of coincidence detection, rather than on rate integration [84], in agreement with
more recent work [60, 76, 85]. Furthermore, this type of encoding strategy would, in princi-
ple, allow an improvement of signal-to-noise ratios, since neurons not conveying informa-
tion would be hindered to spike. In this perspective, according to our models, Gamma
oscillations would allow a network to respond quicker at the expense of decreasing the
strength of its response. On the other hand, more responsive states such as AI, would be bet-
ter suited to respond to low amplitude stimulus (due to their high sensitivity) at the cost of
loosing temporal precision. Thus, AI states, because of their high responsiveness, seem well
suited to detect inputs, while gamma oscillations, due to their tighter time precision, seem
better suited to transmit timing information. Such possibilities constitute interesting direc-
tions to explore by future models.

Resonance
In this work we reproduced previous results [75] showing that resonance is a fundamental
property of spiking networks composed of excitatory and inhibitory neurons. We compared
the resonant properties during AI and Gamma states generated by three different mechanism
(ING, PING and CHING) and verified that, apart from a shift on the resonant frequency cen-
ter, the resonant properties around the Gamma band in all networks did not change signifi-
cantly from one state to the other. We call the reader attention to the particularities of each
network model, especially the enlarged potential of resonance of ING network during AI.
Even though previous work proposed the importance of resonance in information transfer
and processing in the brain [10], this aspect has been left aside until recently [86]. The most
popular view, known as the Communication Through Coherence (CTC) Theory [8, 9], pro-
poses a mechanistic explanation for how different neural regions could communicate by
means of coherence [60]. This theory advocates that, since oscillations generate a rhythmic
modulations in neuronal excitability (defining time windows in which neurons are capable to
respond), only coherently oscillating groups can effectively communicate. In contrast, a recent
work [86] present results indicating that, to the contrary, coherence is a consequence of com-
munication, not a cause of it. This study shows that if an oscillating network is connected to
another network that owns resonant properties around this same frequency, these two net-
works present coherent activity, and that the presence of these resonant interactions could
explain more than 50% of the observed coherence. Furthermore, they show that the oscillating
network sends information to the resonant one (the Granger-causality between field potentials
is dominated by oscillatory synchronization in the sending area).

PLOS Computational Biology | https://doi.org/10.1371/journal.pcbi.1009416 September 16, 2021 23 / 33

78
PLOS COMPUTATIONAL BIOLOGY Integration, coincidence detection and resonance in networks of spiking neurons

In this perspective, the enlarged potential of resonance of ING network in different bands
during AI, indicates that this type of network structure (with heterogeneous connectivity pat-
terns in between inhibitory neurons) could potentially convey information equally well in
several bands. This stress the importance of network topology for neuronal information pro-
cessing and also constitutes interesting directions to further explore.

Supporting information
S1 Fig. Synaptic time scale parameter search of a network composed of RS and FS neurons
randomly connected. The network used to produce this figure was composed of 20000 excit-
atory Regular Spiking and 5000 inhibitory Fast Spiking neurons connected randomly with a
probability of connection of 2%. All synapses were delayed by a time delay of 1.5 ms, and had
reference synaptic strengths of QRe ¼ 1 nS or QRi ¼ 5 nS and reference synaptic time scales of
tRe ¼ tRi ¼ 5 ms. Synaptic strengths (Qe,i) were normalized at each tested time scale (τe,i) to
keep the same synaptic gain, such that: Qe;i ¼ ðQRe;i :tRe;i Þ=te;i . A: Network oscillation frequency
depicted in a color scheme as a function of excitatory and inhibitory synaptic time scales.
White color corresponds to regions in which no oscillation was identified in RS population. B:
Synchrony Index of RS population (top) and network balance (bottom) as a function of synap-
tic time scales. The Synchrony Index (SI) is based on the auto-correlation of the population
frequency of RS cells. To be calculated, the autocorrelation of the population frequency was
fitted by a damped cosine function and the value of this fitted function at zero time lag was
defined as the SI. If the exponential decay rate was higher then 100, it was considered that
there was no global oscillation at the population scale. The network balance was defined as the
rate between the average excitatory and inhibitory synaptic currents, hhIhIinh exc iN
i
it , in which hiN
N

stand for average among neurons and hit average on time. White squares indicate the two
different parameter sets used in our simulations (τe = τi = 5 ms for AI Network, and τe = 1 ms,
τi = 7.5 ms for PING Network). C: Same as B but calculated for the FS population. D: Popula-
tion frequency autocorrelation of RS (green dots) and FS population (blue dots) neurons of
the two used parameter sets. Solid lines indicate the damped cosine fitted function.
(TIF)
S2 Fig. Gamma Network parameter search. The network connectivity (p) vs. inhibitory syn-
aptic strengths (Qi) parameter space of the Gamma Network are displayed as color-plots. A:
Average spiking frequency. B: Network oscillation frequency. C: Network balance: rate
between the average excitatory and inhibitory synaptic currents, hhIhIInh
exc iN
i
it , in which hiN stand
N

for average among neurons and hit average on time. D: Membrane Potential Synchrony
s2V
(χ), calculated by means of the equation: w2 ¼ 1 N 2 , in which VðtÞ ¼ N1 SNi Vi ðtÞ,
N Si sVi
2 2 2 2
sV 2 ¼ h½VðtÞ� it ½hVðtÞit � and sVi 2 ¼ h½Vi ðtÞ� it ½hVi ðtÞit � . The set of parameter which
allowed Gamma Network to oscillate in the Gamma range are indicated by a star symbol. The
white and yellow curves depict parameter choices in which the product between p and Qi
are the same. The yellow curve indicates all parameters equivalent to a choice of p = 60% and
Qi = 5 nS (Q0i ¼ 3=p0 ), while the white curve indicates all parameters equivalent to a choice of
p = 10% and Qi = 5 nS (Q0i ¼ 0:5=p0 ), like it is usually used in other works [30]. Every point in
each graph is given by the average output of 10 simulations of 5 seconds each. In this simula-
tions each neuron of the Gamma Network received 400 independent and identically distrib-
uted excitatory Poissonian spike trains with a spiking frequency μExt = 5 Hz and a synaptic
strength of QExt = 1 nS that decayed with synaptic time constant of τE = 5 ms. E: Network activ-
ity for the parameters indicated with a start in A, B, C and (p = 60% and Qi = 5 nS). The raster

PLOS Computational Biology | https://doi.org/10.1371/journal.pcbi.1009416 September 16, 2021 24 / 33

79
PLOS COMPUTATIONAL BIOLOGY Integration, coincidence detection and resonance in networks of spiking neurons

plot of the whole network (e1), the population frequency (e2), the membrane potential of 3
randomly chosen neurons (e3) and the power spectrum of the population frequency (e4) are
indicated. The population frequency is calculated as the total number of spikes (spikes of the
whole network) in a time bin of 1 ms, divided by the duration of this time bin. Because of the
exclusive presence of inhibitory neurons and its high level of recurrent inhibition, this network
is capable of generating Gamma rhythms with frequencies around 70Hz by means of an ING
mechanism.
(TIF)
S3 Fig. Phase-locking statistical test. A and B: Phase distribution of two randomly picked
cells from the human recordings (Data segment 1): one excitatory (A, green) and one inhibi-
tory (B, red). The phase distribution of each cell was fitted to a Von Mises curve, which allowed
the estimation of its preferred phase yVM . The phase distribution of each neuron was tested for
circular uniformity using a Bonferroni-corrected Rayleigh test [36, 37]. C and D: Rayleigh Z
calculated for all recorded neurons: excitatory (C, green) and inhibitory (D, red). A neuron
was considered phase-locked if the circular uniformity at P < 0.01, (Z > Zc) could be rejected.
In these plots, neurons were ordered according to their Z value and not according to their
original indexes. E: Preferred phases, yVM , of each phase-locked cell, displayed in polar graph
representation. Dark colored vectors indicate the average phase among each neuron type and
Δθ the phase difference among RS and FS. Data segment 1 presented 22 minutes of recordings,
containing 9 seconds of Gamma activity.
(TIF)
S4 Fig. Firing rate change statistical test. A: Activity of two randomly picked cells during sev-
eral Gamma bursts: neuron 13 (inhibitory, left) and neuron 75 (excitatory, right). The graphs
display the firing patter around Gamma bursts (indicated by the black doted lines). Each point
corresponds to one spike in the correspondent tuple of time and burst ID (y-axis). B: Histo-
gram computing the distributions of all spikes inside all Gamma bursts of neuron 13 (left)
and neuron 75 (right). C: Exemplification of firing rate change statistical test. The Poissonian
distribution of these two neurons is constructed based on their average firing rate calculated
outside of Gamma bursts. The critical number of spikes nc, indicated by the dotted lines, is
calculated based on the Percent Point Function of the respective Poissonian Distribution for a
period T, with an 95% Interval of Confidence. The observed number of spikes nobsv is depict as
a dot over the curve. According to this procedure, only neuron 75 is considered to increase its
firing, since nobsv > nc.
(TIF)
S5 Fig. Phase preference of phase-locked cells per data segment in the human record-
ings. A: Data segment 1—containing 22 minutes of recordings and 9 seconds of total
Gamma activity. B: Data segment 2—containing 43 minutes of recordings and 14 seconds of
total Gamma activity. C: Data segment 3—containing 28 minutes of recordings and 16 sec-
onds of total Gamma activity. D: Data segment 4—containing 26 minutes of recordings and
13 seconds of total Gamma activity. E: Data segment 5—containing 16 minutes of record-
ings and 11 seconds of total Gamma activity. The preferred phases of each phase-locked cell
are displayed in polar graph representation. Phases were calculated from −π to π. The vector
size gives a measure of the phase distribution of each cell. Big amplitude vectors indicate
very concentrated distributions while small amplitude vectors indicate less concentrated
ones. The color of each vector encodes the type of the cell of whom it represents the phase:
red (FS), and green (RS). Cell number IDs are indicated. Dark colored vectors indicate the

PLOS Computational Biology | https://doi.org/10.1371/journal.pcbi.1009416 September 16, 2021 25 / 33

80
PLOS COMPUTATIONAL BIOLOGY Integration, coincidence detection and resonance in networks of spiking neurons

average phase among each neuron type and Δθ the phase difference among them.
(TIF)
S6 Fig. Change of Gamma participating cells with time in experimental data. The middle
panel represents each cell by a circle in each of the 5 data segments. FS and RS phase-locked
cells are depicted respectively as red and green circles, while not phase-locked or inconclusive
(with respect to phase locked) cells of both types are depicted as blue and gray circles respec-
tively. Superposed to each cell circle, pointing up and down triangles were added to indicate if
the cell increased (4) or decreased (5) its firing. If the cell didn’t change its firing significantly
a minus sign (-) was added. Side box plots indicate, on the left, the percentage of phase-locked
FS (red) and RS (green) cells in each of the 5 data segments, and, on the right, the percentage
of firing rate increase. Dotted lines indicate the average value (phase-locking level: left and
firing rate increase: right) between the 5 data segments. The bottom box plot depicts the super-
posed counts of phase-locking or firing rate increase behavior of each individual cell, com-
puted in the 5 data segments.
(TIF)
S7 Fig. Behavior consistency of RS and FS cells in human recordings. Distributions of con-
sistency indexes among the recorded neurons with respect to to firing rate increase are dis-
played respectively in A and B for RS cells and FS cells, while C and D display the consistency
indexes distribution of phase-locking for RS and FS.
(TIF)
S8 Fig. Neural behavior time distribution in the human data. The activity of each neuron
inside and outside Gamma bursts in all 5 data segments were quantified. Taking into account
that each data segment had a different duration, containing a different total Gamma duration,
and that some neurons were silent in some data segments, each neuron was analyzed individu-
ally, taking into account the percentage of the total amount of time in which the neuron was
active. A: Phase-locking time distribution. The grid plot in the middle displays the amount of
time (with respect to the total recording time) in which each neuron was considered phase-
locked (A, y axis), and the the amount of time in which each neuron was considered not
phase-locked (A, x axis). RS neurons are depicted in green and FS neurons in red, together
with their ID number. Neurons lying outside of the diagonal are neurons of whom statistical
analysis was inconclusive at some data segments, due to the reduced number of spikes. At the
top left corner, lie neurons that were always considered phase-locked, while neurons that were
never considered phase-locked are placed at the bottom right corner. Pie plots indicate the per-
centage of neurons that passed at least 50% of the total time being either phase-locked or not
phase-locked (neurons that fall inside of the colored quadrants) and the neurons lying on the
left white quadrant. B: Same analysis as A but displaying the firing rate change time distribu-
tion. This analysis indicates that only a small percentage of neurons passed at least 50% of the
total time being either phase-locked (RS: 4.4%, FS: 13%) or increasing its firing (RS: 20.6%, FS:
52.2%). Moreover, even though no cell was 100% of the time phase-locked to Gamma, some
cells were 100% of the time not phase-locked to Gamma (RS: 22.1%, FS: 13%) and others never
increased their firing (RS: 41.2%, FS: 17.4%).
(TIF)
S9 Fig. Firing rate distribution of individual neurons inside Gamma bursts and their
phase-locking classification in human data recordings. The average firing rate of each neu-
ron in each of the 5 data segments (inside Gamma bursts) is depicted as a point in this graph
(each neuron presents 5 points). The color of each point corresponds to the neuron classifica-
tion with respect to phase-locking in the correspondent data segment (purple: phase-locked,

PLOS Computational Biology | https://doi.org/10.1371/journal.pcbi.1009416 September 16, 2021 26 / 33

81
PLOS COMPUTATIONAL BIOLOGY Integration, coincidence detection and resonance in networks of spiking neurons

red: not phase-locked and gray: inconclusive). The average firing rate inside Gamma bursts
was calculated based on the total Gamma duration (recorded by the electrode, that also
recorded the particular neuron, in the respective data segment) and the total number of spikes
emitted by this particular neuron exclusively inside the Gamma bursts of the respective data
segment. Cells classified as inconclusive are cells that spiked less then 5 times inside Gamma
bursts, or cells whose electrode measured less then 1 second of Gamma bursts in the respective
data segment. FS neurons are depicted on the left and RS neurons on the right. Box plots refer-
ent to each neuron distribution are added to help in the visualization (regardless of the reduced
number of points). The box extends from the lower to upper quartile values of the data, with
a line at the median. The whiskers extend from the box to show the range of the data. Flier
points are those past the end of the whiskers and are depicted with black circle together with
the color point. This graph illustrates the fact that phase-locking and not phase-locking behav-
iors are observed both in cells with high and low firing rates.
(TIF)
S10 Fig. Firing rate distributions. Firing rate distributions of different neuron types (inside
and outside Gamma bursts) are depicted in A, B, C and D for each studied system. A: Human
recordings. B: PING Network. C: ING Network and D: CHING Network. Average firing rates of
each cell type is indicated by the dotted line.
(TIF)
S11 Fig. Average excitatory and inhibitory synaptic conductances. A: Illustration of the ana-
lyzed system: PING Network, ING Network and CHING Network. B: Ratio between excitatory
conductance (Ge) and leakage conductance (GL). C: Ratio between inhibitory conductance
(Gi) and leakage conductance (GL). Averages are indicated by the dotted line. The distributions
fall inside of the physiological range observed experimentally [87].
(TIF)
S12 Fig. Average level of phase-locking. The average level of phase-locking is defined as the
averaged percentage of cells in the network considered to be phase-locked, across the 5 seg-
ments of data recorded. The analysis was done separately for excitation and inhibition. A:
Human Data recordings, B:PING Network, C: ING Network and D: CHING Network. The per-
centage of cells signaled as inconclusive relates to cells in which the number of spikes inside
Gamma burst were too small to allow statistical significant phase-locking.
(TIF)
S13 Fig. Network responsiveness of a network composed of RS and FS neurons randomly
connected with different synaptic time scales. A: AI network receiving a Poissonian drive of
3Hz. B: PING network receiving a Poissonian drive of 3Hz (inducing Gamma). C: PING net-
work receiving a Poissonian drive of 2Hz (not inducing Gamma). In addition to the drive
each network received a Gaussian stimulus of 2Hz pick and a standard deviation of 50 ms. The
drive and stimulus are depicted in each case in a1, b1 and c1. The raster plot of each network
during the stimulation is depicted in each case in a2, b2 and c2. The membrane potential of 3
randomly picked neurons are depicted in each case in a3, b3 and c3. The raw and the filtered
simulated LFP are depicted in each case in a4, b4 and c4.
(TIF)
S14 Fig. Responsiveness of individual cells in computational models. A: PING Network. B:
ING Network. C: CHING Network. To estimate the individual cell responsiveness, we calcu-
lated the average spiking frequency of each cell inside (y-axis) and outside stimulus (x-axis)
during AI-like states (left) and Gamma states (right). RS cells are displayed in green and FS

PLOS Computational Biology | https://doi.org/10.1371/journal.pcbi.1009416 September 16, 2021 27 / 33

82
PLOS COMPUTATIONAL BIOLOGY Integration, coincidence detection and resonance in networks of spiking neurons

cells in red. In each plot the linear regression from the points is depicted with the identity. We
observe that all cells follow the same rule of responsiveness (proportional to their firing outside
the stimulus). No difference can be seen between the responsiveness of neurons classified as
Gamma participating and the Gamma non-participating cells.
(TIF)
S15 Fig. Phase-dependent network response protocol. A: Protocol scheme in ING Network
when it displays Gamma oscillations (45-65 Hz). Top: stimulus used to measure network
phase-dependent response. The stimulus consisted of fast Gaussian fluctuation (standard devi-
ation of 1 ms) which modulated the firing rate of the external Poissonian spike trains injected
into network from 0 to 50 Hz. Middle: Raster plot indicating the network response to the
Gaussian stimulus. The network responsiveness was calculated according to Eq 6, in a time
window T = 18ms (shaded gray area). Bottom: Gamma oscillation phase around the the stimu-
lus pick. The phase at the time the stimulus was applied is indicated. The Phase-dependent net-
work responsiveness was measured in three different network states: B: AI state (Poissonian
noise = 2Hz, no external current). C: AI-modulated states (Poissonian noise = 1Hz, with sinu-
soidal external current). D: Gamma state (Poissonian noise = 3Hz, no external current). Items
A, B and C display the Raster activity of ING Network without the Gaussian stimulation. Only
20% of network is shown.
(TIF)
S16 Fig. Linear representation of color maps depicted in Fig 9. A: Resonant properties of
PING Network. B: Resonant properties of ING Network. C: Resonant properties of CHING
Network. The curves displayed in B, C and D depict, for each oscillatory frequency (color
scheme) the amplitude (average number of spikes per neuron per time bin) as a function of
the oscillation phase, during Gamma and AI-like states. All values were normalized by the
average firing inside of each state to exclude the state dependent firing rate level (which is
higher on Gamma). Δnoise = 0.5 Hz in all network models but μnoise varied in each case. For
AI, in PING and ING Networks μnoise = 2 Hz and in CHING Network μnoise = 1 Hz, while for
Gamma, μnoise = 3 Hz in in PING and ING Networks and μnoise = 2 Hz in CHING Network.
(TIF)
S17 Fig. Resonant properties of computational models during Gamma in each cell type. A:
Resonant properties of PING Network. B: Resonant properties of ING Network. C: Resonant
properties of CHING Network. The color maps displayed in A, B and C depict, for each oscil-
latory frequency and oscillation phase, the average number of spikes per cell type (RS, FS,
FS2 or Ch) and time bin, during Gamma state. Differently than Fig 9 no normalization was
applied. Δnoise = 0.5 Hz in all network models but μnoise varied in each case. In PING and ING
Networks μnoise = 3 Hz and in CHING Network μnoise = 2 Hz.
(TIF)

Acknowledgments
We thank Damien Depannemaecker and Mallory Carlu for enlightening discussions and
insights.

Author Contributions
Conceptualization: Eduarda Susin, Alain Destexhe.
Data curation: Eduarda Susin.

PLOS Computational Biology | https://doi.org/10.1371/journal.pcbi.1009416 September 16, 2021 28 / 33

83
PLOS COMPUTATIONAL BIOLOGY Integration, coincidence detection and resonance in networks of spiking neurons

Formal analysis: Eduarda Susin.

Funding acquisition: Eduarda Susin, Alain Destexhe.
Investigation: Eduarda Susin.
Methodology: Eduarda Susin, Alain Destexhe.
Project administration: Eduarda Susin, Alain Destexhe.
Resources: Alain Destexhe.
Software: Eduarda Susin.
Supervision: Alain Destexhe.
Visualization: Eduarda Susin.
Writing – original draft: Eduarda Susin.
Writing – review & editing: Alain Destexhe.

References
1. Buzsaki G. Rhythms of the Brain. Oxford University Press; 2006.
2. Gray CM, Singer W. Stimulus-specific neuronal oscillations in orientation columns of cat visual cortex.
Proceedings of the National Academy of Sciences. 1989; 86(5):1698–1702. https://doi.org/10.1073/
pnas.86.5.1698 PMID: 2922407
3. Gray CM, König P, Engel AK, Singer W. Oscillatory responses in cat visual cortex exhibit inter-columnar
synchronization which reflects global stimulus properties. Nature. 1989; 338(6213):334–337. https://
doi.org/10.1038/338334a0 PMID: 2922061
4. Singer W, Gray CM. Visual feature integration and the temporal correlation hypothesis. Annual review
of neuroscience. 1995; 18(1):555–586. https://doi.org/10.1146/annurev.ne.18.030195.003011 PMID:
7605074
5. Singer W. Neuronal synchrony: a versatile code for the definition of relations? Neuron. 1999; 24(1):49–
65. https://doi.org/10.1016/S0896-6273(00)80821-1 PMID: 10677026
6. O’Keefe J, Recce ML. Phase relationship between hippocampal place units and the EEG theta rhythm.
Hippocampus. 1993; 3(3):317–330. https://doi.org/10.1002/hipo.450030307 PMID: 8353611
7. Fries P, Nikolić D, Singer W. The gamma cycle. Trends in neurosciences. 2007; 30(7):309–316. https://
doi.org/10.1016/j.tins.2007.05.005 PMID: 17555828
8. Fries P. A mechanism for cognitive dynamics: neuronal communication through neuronal coherence.
Trends in cognitive sciences. 2005; 9(10):474–480. https://doi.org/10.1016/j.tics.2005.08.011 PMID:
16150631
9. Fries P. Rhythms for cognition: communication through coherence. Neuron. 2015; 88(1):220–235.
https://doi.org/10.1016/j.neuron.2015.09.034 PMID: 26447583
10. Hahn G, Bujan AF, Frégnac Y, Aertsen A, Kumar A. Communication through resonance in spiking neu-
ronal networks. PLoS Comput Biol. 2014; 10(8):e1003811. https://doi.org/10.1371/journal.pcbi.
1003811 PMID: 25165853
11. Destexhe A, Contreras D. Neuronal computations with stochastic network states. Science. 2006; 314
(5796):85–90. https://doi.org/10.1126/science.1127241 PMID: 17023650
12. Zerlaut Y, Destexhe A. Enhanced responsiveness and low-level awareness in stochastic network
states. Neuron. 2017; 94(5):1002–1009. https://doi.org/10.1016/j.neuron.2017.04.001 PMID: 28595044
13. Brunel N. Dynamics of sparsely connected networks of excitatory and inhibitory spiking neurons. Jour-
nal of computational neuroscience. 2000; 8(3):183–208. https://doi.org/10.1023/A:1008925309027
PMID: 10809012
14. Softky WR, Koch C. The highly irregular firing of cortical cells is inconsistent with temporal integration of
random EPSPs. Journal of neuroscience. 1993; 13(1):334–350. https://doi.org/10.1523/JNEUROSCI.
13-01-00334.1993 PMID: 8423479
15. Holt GR, Softky WR, Koch C, Douglas RJ. Comparison of discharge variability in vitro and in vivo in cat
visual cortex neurons. Journal of neurophysiology. 1996; 75(5):1806–1814. https://doi.org/10.1152/jn.
1996.75.5.1806 PMID: 8734581

PLOS Computational Biology | https://doi.org/10.1371/journal.pcbi.1009416 September 16, 2021 29 / 33

84
PLOS COMPUTATIONAL BIOLOGY Integration, coincidence detection and resonance in networks of spiking neurons

16. Shadlen MN, Newsome WT. The variable discharge of cortical neurons: implications for connectivity,
computation, and information coding. Journal of neuroscience. 1998; 18(10):3870–3896. https://doi.
org/10.1523/JNEUROSCI.18-10-03870.1998 PMID: 9570816
17. Destexhe A, Rudolph M, Paré D. The high-conductance state of neocortical neurons in vivo. Nature
reviews neuroscience. 2003; 4(9):739–751. https://doi.org/10.1038/nrn1198 PMID: 12951566
18. Henrie JA, Shapley R. LFP power spectra in V1 cortex: the graded effect of stimulus contrast. Journal of
neurophysiology. 2005; 94(1):479–490. https://doi.org/10.1152/jn.00919.2004 PMID: 15703230
19. Van Vreeswijk C, Sompolinsky H. Chaos in neuronal networks with balanced excitatory and inhibitory
activity. Science. 1996; 274(5293):1724–1726. https://doi.org/10.1126/science.274.5293.1724 PMID:
8939866
20. Peyrache A, Dehghani N, Eskandar EN, Madsen JR, Anderson WS, Donoghue JA, et al. Spatiotempo-
ral dynamics of neocortical excitation and inhibition during human sleep. Proceedings of the National
Academy of Sciences. 2012; 109(5):1731–1736. https://doi.org/10.1073/pnas.1109895109 PMID:
22307639
21. Le Van Quyen M, Muller LE, Telenczuk B, Halgren E, Cash S, Hatsopoulos NG, et al. High-frequency
oscillations in human and monkey neocortex during the wake–sleep cycle. Proceedings of the National
Academy of Sciences. 2016; 113(33):9363–9368. https://doi.org/10.1073/pnas.1523583113 PMID:
27482084
22. Traub RD, Jefferys JG, Jefferys JG, Whittington MA. Fast oscillations in cortical circuits. MIT press;
1999.
23. Buzsáki G, Wang XJ. Mechanisms of gamma oscillations. Annual review of neuroscience. 2012;
35:203–225. https://doi.org/10.1146/annurev-neuro-062111-150444 PMID: 22443509
24. Traub RD. Fast oscillations. Scholarpedia. 2006; 1(12):1764. https://doi.org/10.4249/scholarpedia.
1764
25. Tiesinga P, Sejnowski TJ. Cortical enlightenment: are attentional gamma oscillations driven by ING or
PING? Neuron. 2009; 63(6):727–732. https://doi.org/10.1016/j.neuron.2009.09.009 PMID: 19778503
26. Gray CM, McCormick DA. Chattering cells: superficial pyramidal neurons contributing to the generation
of synchronous oscillations in the visual cortex. Science. 1996; 274(5284):109–113. https://doi.org/10.
1126/science.274.5284.109 PMID: 8810245
27. Cunningham MO, Whittington MA, Bibbig A, Roopun A, LeBeau FE, Vogt A, et al. A role for fast rhyth-
mic bursting neurons in cortical gamma oscillations in vitro. Proceedings of the National Academy of
Sciences. 2004; 101(18):7152–7157. https://doi.org/10.1073/pnas.0402060101 PMID: 15103017
28. Brette R, Gerstner W. Adaptive exponential integrate-and-fire model as an effective description of neu-
ronal activity. Journal of neurophysiology. 2005; 94(5):3637–3642. https://doi.org/10.1152/jn.00686.
2005 PMID: 16014787
29. Van Vreeswijk C, Abbott L, Ermentrout GB. When inhibition not excitation synchronizes neural firing.
Journal of computational neuroscience. 1994; 1(4):313–321. https://doi.org/10.1007/BF00961879
PMID: 8792237
30. Brunel N, Wang XJ. What determines the frequency of fast network oscillations with irregular neural dis-
charges? I. Synaptic dynamics and excitation-inhibition balance. Journal of neurophysiology. 2003; 90
(1):415–430. https://doi.org/10.1152/jn.01095.2002 PMID: 12611969
31. Stimberg M, Brette R, Goodman DF. Brian 2, an intuitive and efficient neural simulator. eLife. 2019; 8:
e47314. https://doi.org/10.7554/eLife.47314 PMID: 31429824
32. Telenczuk B, Telenczuk M, Destexhe A. A kernel-based method to calculate local field potentials from
networks of spiking neurons. Journal of Neuroscience Methods. 2020; 344:108871. https://doi.org/10.
1016/j.jneumeth.2020.108871 PMID: 32687850
33. Teleńczuk B, Dehghani N, Le Van Quyen M, Cash SS, Halgren E, Hatsopoulos NG, et al. Local field
potentials primarily reflect inhibitory neuron activity in human and monkey cortex. Scientific reports.
2017; 7:40211. https://doi.org/10.1038/srep40211 PMID: 28074856
34. Widmann A, Schröger E, Maess B. Digital filter design for electrophysiological data–a practical
approach. Journal of neuroscience methods. 2015; 250:34–46. https://doi.org/10.1016/j.jneumeth.
2014.08.002 PMID: 25128257
35. Virtanen P, Gommers R, Oliphant TE, Haberland M, Reddy T, Cournapeau D, et al. SciPy 1.0: Funda-
mental Algorithms for Scientific Computing in Python. Nature Methods. 2020; 17:261–272. https://doi.
org/10.1038/s41592-019-0686-2 PMID: 32015543
36. Jacobs J, Kahana MJ, Ekstrom AD, Fried I. Brain oscillations control timing of single-neuron activity in
humans. Journal of Neuroscience. 2007; 27(14):3839–3844. https://doi.org/10.1523/JNEUROSCI.
4636-06.2007 PMID: 17409248

PLOS Computational Biology | https://doi.org/10.1371/journal.pcbi.1009416 September 16, 2021 30 / 33

85
PLOS COMPUTATIONAL BIOLOGY Integration, coincidence detection and resonance in networks of spiking neurons

37. Le Van Quyen M, Staba R, Bragin A, Dickson C, Valderrama M, Fried I, et al. Large-scale microelec-
trode recordings of high-frequency gamma oscillations in human cortex during sleep. Journal of Neuro-
science. 2010; 30(23):7770–7782. https://doi.org/10.1523/JNEUROSCI.5049-09.2010 PMID:
20534826
38. Barthó P, Hirase H, Monconduit L, Zugaro M, Harris KD, Buzsaki G. Characterization of neocortical
principal cells and interneurons by network interactions and extracellular features. Journal of neuro-
physiology. 2004; 92(1):600–608. https://doi.org/10.1152/jn.01170.2003 PMID: 15056678
39. Destexhe A, Contreras D, Steriade M. Spatiotemporal analysis of local field potentials and unit dis-
charges in cat cerebral cortex during natural wake and sleep states. Journal of Neuroscience. 1999; 19
(11):4595–4608. https://doi.org/10.1523/JNEUROSCI.19-11-04595.1999 PMID: 10341257
40. Fries P, Reynolds JH, Rorie AE, Desimone R. Modulation of oscillatory neuronal synchronization by
selective visual attention. Science. 2001; 291(5508):1560–1563. https://doi.org/10.1126/science.
1055465 PMID: 11222864
41. Pesaran B, Pezaris JS, Sahani M, Mitra PP, Andersen RA. Temporal structure in neuronal activity dur-
ing working memory in macaque parietal cortex. Nature neuroscience. 2002; 5(8):805–811. https://doi.
org/10.1038/nn890 PMID: 12134152
42. Bartos M, Vida I, Jonas P. Synaptic mechanisms of synchronized gamma oscillations in inhibitory inter-
neuron networks. Nature reviews neuroscience. 2007; 8(1):45–56. https://doi.org/10.1038/nrn2044
PMID: 17180162
43. Brunel N, Hakim V. Fast global oscillations in networks of integrate-and-fire neurons with low firing
rates. Neural computation. 1999; 11(7):1621–1671. https://doi.org/10.1162/089976699300016179
PMID: 10490941
44. Tiesinga P, José JV. Robust gamma oscillations in networks of inhibitory hippocampal interneurons.
Network: Computation in Neural Systems. 2000; 11(1):1–23. https://doi.org/10.1088/0954-898X_11_
1_301 PMID: 10735526
45. Geisler C, Brunel N, Wang XJ. Contributions of intrinsic membrane dynamics to fast network oscillations
with irregular neuronal discharges. Journal of neurophysiology. 2005; 94(6):4344–4361. https://doi.org/
10.1152/jn.00510.2004 PMID: 16093332
46. Abbott L, van Vreeswijk C. Asynchronous states in networks of pulse-coupled oscillators. Physical
Review E. 1993; 48(2):1483. https://doi.org/10.1103/PhysRevE.48.1483 PMID: 9960738
47. Gerstner W. Time structure of the activity in neural network models. Physical review E. 1995; 51(1):738.
https://doi.org/10.1103/PhysRevE.51.738 PMID: 9962697
48. Gerstner W, Van Hemmen JL, Cowan JD. What matters in neuronal locking? Neural computation.
1996; 8(8):1653–1676. https://doi.org/10.1162/neco.1996.8.8.1653 PMID: 8888612
49. Hansel D, Mato G, Meunier C. Synchrony in excitatory neural networks. Neural computation. 1995; 7
(2):307–337. https://doi.org/10.1162/neco.1995.7.2.307 PMID: 8974733
50. Kopell N, Ermentrout G. Symmetry and phaselocking in chains of weakly coupled oscillators. Communica-
tions on Pure and Applied Mathematics. 1986; 39(5):623–660. https://doi.org/10.1002/cpa.3160390504
51. Kuramoto Y. Chemical oscillations, waves, and turbulence. Courier Corporation; 2003.
52. Marder E. From biophysics to models of network function. Annual review of neuroscience. 1998; 21
(1):25–45. https://doi.org/10.1146/annurev.neuro.21.1.25 PMID: 9530490
53. Traub R, Whittington M, Colling S, Buzsaki G, Jefferys J. Analysis of gamma rhythms in the rat hippo-
campus in vitro and in vivo. The Journal of physiology. 1996; 493(2):471–484. https://doi.org/10.1113/
jphysiol.1996.sp021397 PMID: 8782110
54. Wang XJ, Buzsáki G. Gamma oscillation by synaptic inhibition in a hippocampal interneuronal network
model. Journal of neuroscience. 1996; 16(20):6402–6413. https://doi.org/10.1523/JNEUROSCI.16-20-
06402.1996 PMID: 8815919
55. Whittington MA, Traub RD, Jefferys JG. Synchronized oscillations in interneuron networks driven by
metabotropic glutamate receptor activation. Nature. 1995; 373(6515):612–615. https://doi.org/10.1038/
373612a0 PMID: 7854418
56. Cardin JA, Carlén M, Meletis K, Knoblich U, Zhang F, Deisseroth K, et al. Driving fast-spiking cells
induces gamma rhythm and controls sensory responses. Nature. 2009; 459(7247):663–667. https://doi.
org/10.1038/nature08002 PMID: 19396156
57. Sohal VS, Zhang F, Yizhar O, Deisseroth K. Parvalbumin neurons and gamma rhythms enhance corti-
cal circuit performance. Nature. 2009; 459(7247):698–702. https://doi.org/10.1038/nature07991 PMID:
19396159
58. Zemankovics R, Veres JM, Oren I, Hájos N. Feedforward inhibition underlies the propagation of cholin-
ergically induced gamma oscillations from hippocampal CA3 to CA1. Journal of Neuroscience. 2013; 33
(30):12337–12351. https://doi.org/10.1523/JNEUROSCI.3680-12.2013 PMID: 23884940

PLOS Computational Biology | https://doi.org/10.1371/journal.pcbi.1009416 September 16, 2021 31 / 33

86
PLOS COMPUTATIONAL BIOLOGY Integration, coincidence detection and resonance in networks of spiking neurons

59. Susin E, Destexhe A. Cellular correlates of wakefulness and slow-wave sleep: evidence for a key role of
inhibition. Current Opinion in Physiology. 2020; 15:68–73. https://doi.org/10.1016/j.cophys.2019.12.
006
60. Palmigiano A, Geisel T, Wolf F, Battaglia D. Flexible information routing by transient synchrony. Nature
neuroscience. 2017; 20(7):1014. https://doi.org/10.1038/nn.4569 PMID: 28530664
61. Viriyopase A, Memmesheimer RM, Gielen S. Cooperation and competition of gamma oscillation mech-
anisms. Journal of neurophysiology. 2016; 116(2):232–251. https://doi.org/10.1152/jn.00493.2015
PMID: 26912589
62. Viriyopase A, Memmesheimer RM, Gielen S. Analyzing the competition of gamma rhythms with
delayed pulse-coupled oscillators in phase representation. Physical Review E. 2018; 98(2):022217.
https://doi.org/10.1103/PhysRevE.98.022217 PMID: 30253475
63. Barlow HB. Single units and sensation: a neuron doctrine for perceptual psychology? Perception. 1972;
1(4):371–394. https://doi.org/10.1068/p010371 PMID: 4377168
64. Churchland MM, Byron MY, Cunningham JP, Sugrue LP, Cohen MR, Corrado GS, et al. Stimulus onset
quenches neural variability: a widespread cortical phenomenon. Nature neuroscience. 2010; 13
(3):369–378. https://doi.org/10.1038/nn.2501 PMID: 20173745
65. Riehle A, Grün S, Diesmann M, Aertsen A. Spike synchronization and rate modulation differentially
involved in motor cortical function. Science. 1997; 278(5345):1950–1953. https://doi.org/10.1126/
science.278.5345.1950 PMID: 9395398
66. Vaadia E, Haalman I, Abeles M, Bergman H, Prut Y, Slovin H, et al. Dynamics of neuronal interactions
in monkey cortex in relation to behavioural events. Nature. 1995; 373(6514):515–518. https://doi.org/
10.1038/373515a0 PMID: 7845462
67. Thorpe S, Delorme A, Van Rullen R. Spike-based strategies for rapid processing. Neural networks.
2001; 14(6-7):715–725. https://doi.org/10.1016/S0893-6080(01)00083-1 PMID: 11665765
68. Deneve S. Bayesian spiking neurons I: inference. Neural computation. 2008; 20(1):91–117. https://doi.
org/10.1162/neco.2008.20.1.91 PMID: 18045002
69. Shinomoto S, Kim H, Shimokawa T, Matsuno N, Funahashi S, Shima K, et al. Relating neuronal firing
patterns to functional differentiation of cerebral cortex. PLoS Comput Biol. 2009; 5(7):e1000433. https://
doi.org/10.1371/journal.pcbi.1000433 PMID: 19593378
70. Maimon G, Assad JA. Beyond Poisson: increased spike-time regularity across primate parietal cortex.
Neuron. 2009; 62(3):426–440. https://doi.org/10.1016/j.neuron.2009.03.021 PMID: 19447097
71. Luczak A, McNaughton BL, Harris KD. Packet-based communication in the cortex. Nature Reviews
Neuroscience. 2015; 16(12):745–755. https://doi.org/10.1038/nrn4026 PMID: 26507295
72. Maršálek P, Koch C, Maunsell J. On the relationship between synaptic input and spike output jitter in
individual neurons. Proceedings of the National Academy of Sciences. 1997; 94(2):735–740. https://
doi.org/10.1073/pnas.94.2.735 PMID: 9012854
73. Rudolph M, Destexhe A. Tuning neocortical pyramidal neurons between integrators and coincidence
detectors. Journal of computational neuroscience. 2003; 14(3):239–251. https://doi.org/10.1023/
A:1023245625896 PMID: 12766426
74. Lepousez G, Lledo PM. Odor discrimination requires proper olfactory fast oscillations in awake mice.
Neuron. 2013; 80(4):1010–1024. https://doi.org/10.1016/j.neuron.2013.07.025 PMID: 24139818
75. Ledoux E, Brunel N. Dynamics of networks of excitatory and inhibitory neurons in response to time-
dependent inputs. Frontiers in computational neuroscience. 2011; 5:25. https://doi.org/10.3389/fncom.
2011.00025 PMID: 21647353
76. Akam T, Kullmann DM. Oscillations and filtering networks support flexible routing of information. Neu-
ron. 2010; 67(2):308–320. https://doi.org/10.1016/j.neuron.2010.06.019 PMID: 20670837
77. Melloni L, Molina C, Pena M, Torres D, Singer W, Rodriguez E. Synchronization of neural activity across
cortical areas correlates with conscious perception. Journal of neuroscience. 2007; 27(11):2858–2865.
https://doi.org/10.1523/JNEUROSCI.4623-06.2007 PMID: 17360907
78. McCarley RW. Brainstem control of Wakefulness and Sleep. New York: Plenum Press; 1990.
79. Fries P, Roelfsema PR, Engel AK, König P, Singer W. Synchronization of oscillatory responses
in visual cortex correlates with perception in interocular rivalry. Proceedings of the National Acad-
emy of Sciences. 1997; 94(23):12699–12704. https://doi.org/10.1073/pnas.94.23.12699 PMID:
9356513
80. Uhlhaas PJ, Singer W. Abnormal neural oscillations and synchrony in schizophrenia. Nature reviews
neuroscience. 2010; 11(2):100–113. https://doi.org/10.1038/nrn2774 PMID: 20087360
81. Kulli J, Koch C. Does anesthesia cause loss of consciousness? Trends in neurosciences. 1991; 14
(1):6–10. https://doi.org/10.1016/0166-2236(91)90172-Q PMID: 1709532

PLOS Computational Biology | https://doi.org/10.1371/journal.pcbi.1009416 September 16, 2021 32 / 33

87
PLOS COMPUTATIONAL BIOLOGY Integration, coincidence detection and resonance in networks of spiking neurons

82. Koch C, Massimini M, Boly M, Tononi G. Neural correlates of consciousness: progress and problems.
Nature Reviews Neuroscience. 2016; 17(5):307–321. https://doi.org/10.1038/nrn.2016.22 PMID:
27094080
83. Destexhe A. Oscillations, complex spatiotemporal behavior, and information transport in networks of
excitatory and inhibitory neurons. Phys Rev E Stat Phys Plasmas Fluids Relat Interdiscip Topics. 1994;
50(2):1594–1606. PMID: 9962131
84. VanRullen R, Thorpe SJ. Surfing a spike wave down the ventral stream. Vision research. 2002; 42
(23):2593–2615. https://doi.org/10.1016/S0042-6989(02)00298-5 PMID: 12446033
85. Sherfey J, Ardid S, Miller EK, Hasselmo ME, Kopell NJ. Prefrontal oscillations modulate the propagation
of neuronal activity required for working memory. Neurobiology of learning and memory. 2020;
173:107228. https://doi.org/10.1016/j.nlm.2020.107228 PMID: 32561459
86. Schneider M, Dann B, Sheshadri S, Scherberger H, Vinck M. A general theory of coherence between
brain areas. bioRxiv. 2020.
87. Rudolph M, Pospischil M, Timofeev I, Destexhe A. Inhibition determines membrane potential dynamics
and controls action potential generation in awake and sleeping cat cortex. Journal of neuroscience.
2007; 27(20):5280–5290. https://doi.org/10.1523/JNEUROSCI.4652-06.2007 PMID: 17507551

PLOS Computational Biology | https://doi.org/10.1371/journal.pcbi.1009416 September 16, 2021 33 / 33

88
 Work 3

A network model of the modulation

of Gamma oscillations by NMDA
receptors in cerebral cortex

French Abstract:

Des médicaments psychotiques, tels que la kétamine, induisent des symptômes proches
de la schizophrénie et stimulent l’augmentation d’oscillations Gamma, similaires à ce que
s’observe chez les patients touches par cette maladie. Toutefois, les mécanismes sous-
jacents à ces effets ne sont toujours pas bien décrits. Dans cette étude, nous avons utilisé
des modèles informatiques de réseaux de neurones corticaux générant des oscillations
Gamma, où nous avons intégré l’action de médicaments tels que la kétamine pour bloquer
partiellement les récepteurs de n-méthyl-d-aspartate (NMDA). Le modèle développé décrit
la modulation des oscillations Gamma par des antagonistes des récepteurs de NMDA, en
supposant que les antagonistes affectent davantage les récepteurs de NMDA sur les in-
terneurones inhibiteurs. Le modèle a ensuite été utilisé pour comparer la réactivité du
réseau aux stimuli externes. Nous avons constaté que lorsque les canaux de NMDA sont
bloqués, une augmentation de la puissance des oscillations Gamma est observée, conjoin-
tement avec une augmentation de la réactivité du réseau. Cependant, cette augmentation
de la réactivité s’applique non seulement aux états oscillatoires dans la fréquence Gamma,
mais également aux états asynchrones sans oscillations apparentes. Nous concluons que
les antagonistes des canaux de NMDA induisent un état d’excitabilité accru, qui produit
une réponse aux stimuli externes exacerbée, et qui peut ou non produire des oscillations
Gamma. Nous cogitons que cet effet stimulant peut expliquer des phénomènes tels qu’une
perception altérée ou des hallucinations.

89
PREPRINT 1

A network model of the modulation of gamma

oscillations by NMDA receptors in cerebral cortex
Eduarda Susin1,* , Alain Destexhe1
1 Paris-Saclay University, Centre National de la Recherche Scientifique (CNRS), Institute of Neuroscience
(NeuroPSI), Gif-sur-Yvette, France * [email protected]

Abstract—Psychotic drugs such as ketamine induce symptoms In this study we investigate by means of computational
close to schizophrenia, and stimulates the production of gamma models how NMDAR antagonists, such as ketamine, affect the
oscillations, as also seen in patients, but the underlying mecha- dynamics of neural networks and how the generated boosting
nisms are still unclear. Here, we have used computational models
of cortical networks generating gamma oscillations, and have of Gamma activity affects the network response, providing
integrated the action of drugs such as ketamine to partially an interpretation for the observed correlation between Gamma
block n-methyl-d-Aspartate (NMDA) receptors. The model can Power and psychotic episodes.
reproduce the modulation of gamma oscillations by NMDA-
receptor antagonists, assuming that antagonists affect NMDA
receptors predominantly on inhibitory interneurons. We next II. R ESULTS
used the model to compare the responsiveness of the network Computational model reproduces experimental features
to external stimuli, and found that when NMDA channnels are
blocked an increase of Gamma power is observed altogether Several preparations with sub-anesthetics doses of NMDAR
with an increase of network responsiveness. However, this re- antagonists have reported to produce neural excitation [25–30].
sponsiveness increase applies not only to gamma states, but also Since NMDAR mediate excitatory synaptic transmission, this
to asynchronous states with no apparent gamma. We conclude
that NMDA antagonists induce increased excitability state, which behavior is intriguing. Several hypothesis have been proposed
may or may not produce gamma oscillations, but the response to explain this apparent paradox [3]. One of the possible
to external inputs is exacerbated, which may explain phenomena explanations is that NMDAR antagonists in sub-anesthetics
such as altered perception or hallucinations. doses act preferentially on inhibitory neurons, increasing
Index Terms—Schizophrenia, NMDAR hypofunction, Gamma network activity indirectly by means of desinhibition. Even
oscillations, Network Model, Psychosis though some contrasting results have been observed [31], this
interpretation has been supported experimentally by several
works [32–35]. Network excitability have also been reported
I. I NTRODUCTION
to increase in schizophrenic patients [36, 37], and its increase

S CHIZOPHRENIA is a mental disorder characterized by

three classes of symptoms: positive symptoms (such
as delusions, hallucinations and disordered thoughts or
in sensory and association cortex have been correlated with
hallucinations [38, 39].
Another important effect of NMDAR antagonists in sub-
speech), negative symptoms (comprehending poverty of anesthetics doses is the increase of Gamma-band activity.
speech and deficits of normal emotional response), and cogni- These observations were reported in human [17–19], monkey
tive deficits [1–3]. Several abnormalities have been identified [24] and rats [20–23], both during cognitive tasks or free
in schizophrenic patients, including important differences in movement.
neurotransmitters systems, anatomical deficits and abnormal The network model developed in the present work (see
neural rhythms [4, 5]. Methods) is able to reproduce both of these features (increase
Gamma oscillations (30-90 Hz) in early-course schizophre- of network excitability and increase of Gamma power). Fig-
nia patients are commonly reported to present increased power ure 1 depicts the network behavior with respect to the to
and/or phase synchronization [6–8]. In parallel, positive cor- different NMDA synaptic strengths, QN M DA , in excitatory
relation between psychotic symptoms and the Gamma power Regular Spiking (RS) and in inhibitory Fast Spiking (FS)
have been identified in schizophrenic patients, in which higher cells. We mimic the block of NMDA channels due to the
Gamma-band activity corresponded to increased symptom action of NMDAR antagonists by decreasing QN M DA in
load [9–12]. These findings indicate that hallucinations and RS and FS cells according to Figure 1A (see Methods).
delusions could be related to an excess of oscillatory synchro- Points of higher synaptic strengths are associated with healthy
nization in the Gamma band. conditions, while points with lower synaptic strengths are
NMDA receptor (NMDAR) antagonists, commonly used in associated to pathological conditions supposedly similar to the
sub-anesthetic doses as animal and human models to study schizophrenic brain. The network dynamics for two sets of
Schizophrenia [13], induce a psychotic state that resembles all NMDA synaptic strengths are shown in Figure 1B and Figure
three classes of symptoms of the disease [14–16]. Furthermore, 1C by means of a Raster Plot. As the synaptic strengths of
NMDAR antagonists also increase Gamma power amplitude, NMDA channels decreased (higher concentration of NMDAR
both in human and in animal models [17–24]. antagonists), the firing rate of excitatory RS cells increased

90
PREPRINT 2

while the firing rate of inhibitory FS cells decreased (Figure In the present study, we compare AI and Gamma states at
1D). In addition, the Gamma power of the population activity different levels of NMDAR block. Figure 4 depicts the respon-
(see Methods) presented an increase (Figure 1E and F). siveness of RS neurons, with respect to different stimulus am-
plitudes (same protocol as Figure 2), for different ensembles
of NMDA synaptic strengths. In agreement with Figure 2, pa-
Network Responsiveness during Gamma rhythms in different
rameter sets in which NMDA synaptic strengths are decreased
levels of NMDAR block
(mimicking the action NMDAR antagonists) correspond to
We investigated how the decrease of NMDA synaptic regions of the parameter space with higher responsiveness. For
strength changed the network dynamics and its capacity to example, QN FS
M DA
= 0.4 nS and QN FS
M DA
=0.36 nS displayed
respond to external stimulus. higher responsiveness then the networks in which the NMDA
While network excitability is related to an overall increase of synaptic strengths wrere QNFS
M DA
= 1 nS and QNFS
M DA
=0.8 nS.
spiking activity, network responsiveness relates to the network Interestingly, in both conditions, responsiveness in AI states
capacity to react to a certain stimulus, producing additional were always superior to the one in Gamma. This result was
spikes then the ones generated by spontaneous activity. These also observed in a similar model in the obscene of NMDA
two dynamical measurements (excitability and responsiveness) channels [40]. This example illustrates a general tendency,
are not always congruent, meaning that it is possible to observe which was also observed with other parameter sets.
an increase in excitability but a concomitant decrease in
responsiveness [40].
III. D ISCUSSION
Network responsiveness was defined as the difference be-
tween the total number of spikes generated by the whole In this work, we used computational models to investigate
network in the presence and in the absence of the stimulus (see the effect of psychotic drugs such as ketamine in cerebral
Eq 6). We measured network responsiveness at different levels cortex, and how gamma oscillations relate to these effects.
of NMDAR block for different stimulus amplitudes (Figure 2). Our findings are (1) NMDA receptors antagonists modulate
The stimulus consisted of a variation in time of the external the rhythms produced by a simple network model consisting
Poissonian drive, in a Gaussian manner (see Methods). of two distinct cell types, RS and FS cells, which generate
Network responsiveness in RS cells increased with the Gamma oscillations by means of the PING mechanism [68].
increased level of NMDAR block, while the responsiveness of This modulation is obtained assuming that the NMDAR block
FS neurons decreased. In this case, both, network excitability predominantly affects interneurons. (2) The boosted gamma
and network responsiveness, behave in the same direction. oscillations following partial block of NMDA receptors, was
The increase of network responsiveness can be understood accompanied by an increased responsiveness to external in-
from Figure 3. The NMDA receptors block depolarizes RS puts. (3) This increase of responsiveness could also be seen
cells, while FS neurons are overall hyperpolarized. For weak for asynchronous states, with no apparent gamma. We discuss
levels of NMDA receptors block, no or weak depolarization is below the implications of these findings.
observed in FS cells, while for strong levels of NMDA block A first prediction of the model is that it was necessary
a significant hyperpolarization is observed. that the antagonism affects predominantly NMDAR receptors
on interneurons. This feature is supported by a number of
observations. Intuitively, if the NMDAR block would occur
Gamma states vs. AI states
predominantly on excitatory cells, then it is difficult to see
Gamma oscillations (30-90 Hz) are believed to be involved how diminishing excitation could augment the activity and
in information processing [41–46], and have been associated to excitability of the network. This long-standing question was
different high-level cognitive functions, such as memory [47– resolved recently by finding that indeed, NMDAR antagonists
49], perception [50–53], attention [54–57], focused arousal primarily affects NMDA receptors on interneurons. It was
[58] and prediction [59]. In parallel, studies with schizophrenic observed that the application of Ketamine or MK-801 in sub-
patients have reported a positive correlation between psychotic anesthetic doses leads to an increased activity of glutamatergic
symptoms and the power of Gamma oscillations [9–12]. neurons both in cortex [25, 35] and in hippocampus [33], and
In contrast, Asynchronous-and-Irregular (AI) states [60] are that this increase of glutamatergic activity is a consequence of
usually associated to conscious states [61], being observed the disinhibition of GABAergic neurons [32, 34]. In addition,
during awake and aroused states [62]. This regime are char- it has also been reported in hippocampus that inhibitory
acterized by irregular and sustained firing with very weak neurons are more sensitive to NMDAR antagonists than glu-
correlations [63–67]. tamatergic neurons [69, 70]. Thus, our model completely
In a previous study [40] we reported that AI states, in supports these findings, and could reproduce the increase of
comparison to oscillatory states in Gamma band, provide Gamma power induced by NMDA receptor antagonists. On
the highest responsiveness to external stimuli, indicating that the other hand, contrasting results also exist. For example,
Gamma oscillations tend to overall diminish responsiveness. [31] argue that NMDAR have less impact on the activity of
This observation could indicate that Gamma rhythms present inhibitory neurons than on the one of excitatory neurons, since
a masking effect, conveying information in its cycles on spike they and other authors observed that NMDAR block depressed
timing at the expense of decreasing the strength of the network large EPSP–spike coupling more strongly in excitatory than in
response. inhibitory neurons [31, 71, 72].

91
PREPRINT 3

A D

Healthy

Pathological

B E

FS
C F

Fig. 1. Network dynamics with respect to different levels of NMDA channels block in the network. A) Possible trajectory in the parameter space of
QN M DA vs. QN M DA , mimicking the action of NMDA receptor (NMDAR) antagonists (the higher the intensity of the NMDAR antagonists, the smaller the
RS FS
NMDA synaptic strengths). The thin line indicates the identity for reference. The arrow indicates the sens of action of NMDAR antagonists. Points of higher
synaptic strengths are associated with healthy conditions, while points with lower synaptic strengths are associated to pathological conditions supposedly
similar to the schizophrenic brain. B) and C) Raster plots indicating the activity of only 1000 cells of each type (FS in red and RS in green), for two parameter
sets. B: QN M DA = 0.8 nS and QN M DA = 1 nS, and C: QN M DA = 0.213 nS and QN M DA = 0.2 nS. D) Average firing rate of RS (green) and FS cells
RS FS RS FS
(red) with respect to the trajectory in parameter space depicted in A. Only the values of QN FS
M DA are indicated in the x axis. Standard errors of the mean
(SEM) are indicated as error bars. E) Average normalized Power Spectrum of the network LFP for different NMDA synaptic strength. Like in D, the synaptic
strengths follow the curve indicated in A, but only the values of QN M DA are indicated in the color scheme. Notice the shift of the Power Spectrum pick
FS
toward smaller frequencies with the increase of NMDA channel block. F) Power Spectrum peak amplitude with respect to the levels of NMDA channels block
(following the synaptic strengths indicated in A). The color scheme (presented for better visualization) are the same as in E. Standard errors of the mean
(SEM) are indicated as error bars. Results expressed in D, E and F are the outcome of 50 simulations average. The arrows indicate the sense of the behavior
according to amount of block of NMDA channels.

The second finding, which is probably the main finding FS cells. Consequently, there is an increase of responsiveness
of our study, is that the network has a marked increased of RS cells, with a corresponding decrease for FS cells, as
responsiveness under the boosted Gamma condition. This we observed. In this model, the increase of responsiveness is
increased responsiveness could be tested experimentally either due to the depolarizing effect on RS cells, and are not due
in vitro, by testing the response of cortical slices with and to gamma oscillations. Indeed, the highest responsiveness was
without application of NMDAR antagonists, or in vivo, by seen for asynchronous states, also in agreement with a previous
monitoring their response following administration of NMDA modeling study [40].
antagonists.
The third finding is that the increase of responsiveness is not Possible implications to understand brain pathologies
specific to gamma oscillations, because it was also present for Our model exhibits several interesting properties that can
asynchronous states with no apparent gamma. The underlying be related to pathologies. First, the model provides a possible
mechanism is that the antagonism of NMDA receptors produce explanation for the symptoms associated to ketamine and
an overall depolarization of RS cells, and hyperpolarization of others NMDA receptor antagonists, such as hallucinations. The

92
PREPRINT 4

A B

Stimulus
Spiking Activity
Responsiveness
Time Window 500 ms 500 ms
Filtered LFP

C D

Fig. 2. Network responsiveness to broad Gaussian inputs in different levels of NMDA channel blocked during Gamma rhythms. A) Responsiveness
protocol scheme. The total number of spikes generated by the network were measured during an external stimulus and in its absence in a time window of 500
ms. The stimulus consisted of a Gaussian fluctuation in the firing rate of the external noise input. Responsiveness was calculated according to Equation 6. B)
Gaussian input amplitude variation. The Gaussian amplitude varied from 0.05 Hz to 2.5 Hz (step of 0.05 Hz), always keeping the same standard deviation of
50 ms. C and D depict respectively the responsiveness of RS (C) and FS (D) neurons for different Gaussian amplitudes in different levels of NMDAR block,
when the network was displaying Gamma activity. The color-scheme indicates the synaptic weights of NMDA synapses (QN M DA ) in RS and FS cells. The
arrow indicates the sense of the simulated action of NMDA antagonist (decreasing synaptic strength). Every point corresponds to the average responsiveness
measured in 15 simulations. Standard error of the mean are indicated by the shaded region around each curve.

Fig. 3. Membrane potential polarization as a function of NMDA receptor

block. The average membrane potential of RS (green, left y-axis) and FS (red,
right y-axis) is expressed as function of NMDA synaptic weights of RS and
FS cells. The values of QN RS
M DA and QN M DA follow the trajectory in the
FS Fig. 4. Network responsiveness to broad Gaussian inputs of different
parameter space indicated in Figure 1A. Only the values of QN FS
M DA are
amplitudes during Gamma and AI states. The responsiveness of RS
indicated in the x axis. The average was performed first in between neurons neurons, due to different Gaussian amplitudes stimuli (same as in the protocol
(hiN ), obtaining an average curve as a function of time, and subsequently with of Figure 2), was measured in different states AI and Gamma for NMDA
respect time (hit ). The values plotted correspond to the average of hhV iN it synaptic parameter sets: QN M DA = 0.8 nS and QN M DA = 1 nS (Gamma:
RS FS
in between 10 simulations. The error bars indicate the standard error of the black, AI: gray), and QN M DA = 0.36 nS and QN M DA = 0.4 nS (Gamma:
RS FS
mean between these simulations. blue, AI: light blue). The Gaussian amplitude varied from 0.05 Hz to 2.5 Hz
(step of 0.05 Hz), always keeping the same standard deviation of 50 ms.

enhanced responsiveness produced by antagonizing NMDA re-

ceptors may explain exacerbated responses to sensory stimuli, sistent with the previously reported role for FS neurons in
which may be related to phenomena such as altered perception schizophrenia. Post-mortem analysis of schizophrenic patient
or hallucinations. Indeed, it is well documented that ketamine brains have shown a reduced expression of parvalbumin (PV)
produces hallucinations together with a marked increase of and GAD67 [1, 76–80]. In parallel, genetic ablation of NMDA
gamma oscillations [73–75]. receptors in PV-positive interneurons in rodents mimics impor-
Besides hallucinations, the model seems also a priori con- tant behavioral [81] and phenotypical features of of the disease

93
PREPRINT 5

(reduction of GAD67 [82], increase of neuronal excitability Vi = Vth the membrane potential is reset to Vrest , which is
[82] and increase of spontaneous Gamma power [83–85]). kept constant until the end of the refractory period Tref . After
These observations support the idea that the hypofunction the refractory period the equations start being integrated again.
of NMDA receptors in PV-positive interneurons are specially In the developed network two types of cells were used:
important in this illness. Regular Spiking (RS) excitatory cells and Fast Spiking (FS)
However, NMDA receptors are expressed in both GABAer- inhibitory cells. The cell specific parameters are indicated in
gic and glutamatergic neurons [32], and it still remains unclear Table I.
in which types of cells the NMDA receptor hypofunction TABLE I
causes schizophrenia [3, 86]. Some works reported conflicting Specific Neuron Model Parameters
results and have questioned the hypothesis that PV-positive
Fast Spiking neurons play a role in Schizophrenia [31, 86]. Parameter RS FS
In our model, the effect of NMDAR antagonists is to Vth -40 mV -47.5 mV
∆ 2 mV 0.5 mV
increase excitability due to desinhibition, consistent with a Tref 5 ms 5 ms
number of experimental observations [25–30]. This increased τw 500 ms 500 ms
excitability is accompanied by a Gamma power increase, a 4 nS 0 nS
b 20 pA 0 pA
as also found in experiments with ketamine [17–19] or in C 150 pF 150 pF
schizophrenic patients [6–12]. The model could reproduce gL 10 nS 10 nS
all these experimental observations only assuming a larger EL -65 mV -65 mV
decrease of the NMDA synaptic strengths in FS cells than EE 0 mV 0 mV
EI 80 mV 80 mV
in RS cells (see Figure 1A). These results support the idea Vrest -65 mV -65 mV
sustained by some authors [87], that PV-positive Fast Spiking
inhibitory neurons play a key role in schizophrenia. Another
modeling study also stressed the importance of NMDA chan-
nels into FS neurons [88]. Thus, models support the view Synaptic Models
that the hypofunction of NMDA receptors on FS cells could
explain a number of features typical of schizophrenia, such as The post-synaptic current received by each neuron i is
anomalous responses and boosted gamma oscillations. composed by three components: two excitatory, referent to
AM P A and N M DA synaptic channels, and one inhibitory,
referent to GABAA channels.
IV. M ETHODS
Neuronal Model IiSyn (t) = IiAM P A (t) + IiGABAA (t) + IiN M DA (t)
Neural units are described by the Adaptive Exponential
Integrate-And-Fire Model (Adex) [89]. In this model, each
neuron i is described by its membrane potential Vi , which
in which
evolves according to the following equations:

  IiAM P A (t) = GAM

i
PA
(t)(Vi (t) − E AM P A )
dVi (t) (Vi (t) − Vth )
C = − gL (Vi − EL ) + gL ∆exp
dt ∆
IiGABAA (t) = GGABAA
(t)(Vi (t) − E GABAA ) (2)
− wi (t) − IiSyn (t) i

dwi (t) X IiN M DA (t) = GN

i
M DA
(t)(Vi (t) − E N M DA )B(Vi (t))
τwi = a(Vi (t) − EL ) − wi (t) + b δ(t − tj )
dt j
(1) E AM P A = 0 mV, E GABAA = -80 mV and E N M DA = 0 mV
where C is the membrane capacitance, gL is the leakage are the reversal potentials of AM P A, GABAA and N M DA
conductance, EL is the leaky membrane potential, Vth is the channels. While the AM P A and GABAA -mediated currents
effective threshold, ∆ is the threshold slope factor and IiSyn (t) are fast, the NMDA-mediated are considerably slower and
is postsynaptic current received by the neuron i (see next present a complex relation with respect to the membrane
section). The adaptation current, described by the variable wi , potential [90–93]. This complex relation , due to magnesium
increases by an amount b every time the neuron i emits a spike block, is accurately modeled by the phenomenological expres-
at times tj and decays exponentially with time scale τw . The sion B(V) [94] :
subthreshold adaptation is governed by the parameter a.
1
During the simulations, the equation characterizing the B(V ) = (3)
membrane potential Vi is numerically integrated until a spike 1 + exp(−0.062V ).([M g 2+ ]o /3.57)
is generated. Formally this happens when Vi grows rapidly where [M g 2+ ]o = 1 mM is the external magnesium
toward infinity. In practice, the spiking time is defined as the concentration (1 to 2 mM in physiological conditions).
moment in which Vi reaches a certain threshold (Vth ). When

94
PREPRINT 6

Because of the fast dynamicas of AM P A and GABAA

channels, their synaptic conductances (GX with X=AM P A
, GABAA ) are usually modeled to increase discontinuously
by a discrete amount QX , every time a presynaptic neuron
A
spikes at time tk , and to subsequently decay exponentially
with a decay time constant τdecay
X
according to the following
equation:

dGX i (t) X
X
τdecay = −GX i (t) + QX δ(t − tk ) (4)
dt
k
P
In which, k runs over all the presynaptic spike times. The
synaptic time constantes used for AM P A and GABAA
GABAA
synapses are τdecay
AM P A
= 1.5 ms and τdecay = 7.5 ms.
B
NMDA channels synaptic conductances, GN M DA , be-
cause of their slow dynamics, are usually modeled as a bi-
exponential function characterized by a rise time constant,
N M DA
τrise = 2 ms, and a decay time constant τdecay
N M DA
= 200 ms,
according to the following equation:

GN
i
M DA
= QN
i
M DA
si (t)N M DA

dsi (t)N M DA si (t)N M DA

= − N M DA + αxi (t)(1 − si (t)N M DA )
dt τdecay C LFP

dxi (t) xi (t) X

= − N M DA + δ(t − tk )
dt τrise k
(5)

In which, QN i
M DA
is the synaptic strength of the NMDA
synapse towards the neuron
P i, α= 0.5/ms and x(t) is an
auxiliary variable. The k runs over all the presynaptic spike
times. Both, s(t)N M DA and x(t), are adimensional.
D
Synaptic strenghs of N M DA synapses (towards RS and
FS neurons) were chosen according to the parameter search
expressed in Figure 5, while the synaptic parameters of
AM P A and GABAA synapses were chosen according to
previous works [40, 95] (QAM P A = 5 nS and QGABAA = 3.34
nS). All synapses (AM P A, GABAA and N M DA) were
delayed by time of 1.5 ms. With these choice of parameters
the NMDA/AMPA charge ratio in the network is on average
higher in RS cells then in FS cells (see Figure IV), in agree-
ment with experimental measurements in prefrontal cortex of
adult mice [31] and rat [96]. Fig. 5. Parameter space of NMDA synaptic weights in RS and FS cells.
A) Average spiking rate in RS cells. B) Average spiking rate in FS cells. C)
LFP Power Spectrum pick. D) LFP Power Spectrum amplitude. The parameter
Network Structure space of NMDA synaptic weights (QN M DA ) was explored for RS and FS
cells in the developed network model. QN RS
M DA and QN M DA varied from 0
FS
The network developed in this work is composed of 5000 nS to 1 nS in steps of 0.05 nS. Each point in the color maps corresponds to the
neurons (4000 RS and 1000 FS). Each neuron (RS or FS) average of 10 simulations of 5 seconds. Points in which QN RS
M DA =QN M DA
FS
are highlighted. Small squares indicate a possible trajectory in the parameter
was connected randomly to every other neurons in the network space (in the direction of the arrow) generated by the action of NMDAR
with a probability of 10%, receiving on average 500 excitatory antagonists. This is the same trajectory indicated in Figure 1A.
synapses (mediated by both AM P A and N M DA channels)
and 100 inhibitory synapses (mediated by GABAA channels).

95
PREPRINT 7

External Input
In addition to recurrent connections, each neuron received A
an external drive to keep the network active. This external
drive consisted of Next = 5000 independent and identically
distributed excitatory Poissonian spike trains with a spiking
frequency µext . These spike trains were sent to the network
with a 10% probability of connection and were computed
inside of the synaptic current term I AM P A , with a synaptic
strength of QAM
Ext
PA
= 0.8 nS. For Gamma activity, the network
was stimulated with a drive with µext = 3 Hz. For Asyn-
chronous and Irregular activity, the network was stimulated B
with a drive with µext = 2 Hz. The external drive mimicked
cortical input, like if the network was embedded in a much
bigger one.
To test network responsiveness, an additional external input
was included in the simulations. This external input, similar
to the external drive, also consisted of Next = 5000 inde-
pendent and identically distributed excitatory Poissonian spike
trains, connected to the network with a 10% probability. The
difference of this input was its firing rate time dependence C
(µext (t)). The spiking frequency of the spike trains varied in
a Gaussian manner, with a standard deviation of 50 ms and
variable amplitude. These spike trains were computed inside
of both synaptic current terms I AM P A and I N M DA , with
a synaptic strengths of QAM Ext
PA
= 0.8 nS, and QN ExtRS and
M DA

QExtF S as indicated in each case.

N M DA

Block of NMDA channels: effect of NMDAR antagonists

In this work, we mimic the effect of NMDAR antagonists by Fig. 6. Excitatory synaptic currents. A) Average AMPA current of one
changing the value of the NMDA synaptic weights QN M DA . randomly picked RS (green) and one randomly picked FS (red) neuron.
In Figure 5 a possible trajectory in the parameter space B) Average NMDA current of one randomly picked RS (green) and one
randomly picked FS (red) neuron. C) Ratio of NMDA and AMPA charges
generated by the action of NMDAR antagonists is depicted. for RS and FS cells. The synaptic charge ratio of each neuron was calculated
This is the same trajectory indicated in Figure 1A. separately. The Bars indicate the mean and the standard deviation among the
RS and FS population. The NMDA synaptic strengths in RS and FS cells are
QN M DA =0.8 nS and QN M DA = 1 nS (which, in our model, describes a
Simulations RS
healthy condition).
FS

All neural networks were constructed using Brian2 simula-

tor [97]. All equations were numerically integrated using Euler
Methods and dt=0.1 ms as integration time step. The codes Power Spectrum
for each one of the three developed networks are available at The Power spectrum of the simulated LFP was calculated
ModelDB platform. by means of the Welch’s method, using a Hamming window
of length 0.25 seconds and 125 overlapping points. We used
Population activity: LFP model the Python-based ecosystem Scipy function signal.welch to do
To measure the global behavior of the neuronal population, our calculations.
we used a simulated Local Field Potential (LFP). This LFP
was generated by the network, by means of a recent method Synaptic Charge
developed by [98]. This approach calculates the LFP by
The synaptic charge (AMPA or NMDA) of each neuron is
convolving the spike trains of the network with a Kernel that
defined as the area under the curve of the average synaptic
have been previously estimated from unitary LFPs (the LFP
current (shaded areas of Figure A or B), which was calculated
generated by a single axon, uLFP) measured experimentally.
from the presynaptic input time until 10 ms after it.
Since this method assumes a spatial neuronal displacement,
to be able to apply it to our simulations, we randomly
displaced part of the network (50 neurons) in 2-D grid, Responsiveness
assuming that the electrode was displaced on its center and was The level of responsiveness (R) of a network, due to an
measuring the LFP in the same layer as neuronal soma. The stimulus (S) in a time window of duration T , is defined as
program code of the kernel method is available in ModelDB the difference between the total number of spikes generated by
(http://modeldb.yale.edu/266508), using python 3 or the hoc the whole network due to an stimulus (Nspikes
S
) and the total
language of NEURON. number of spikes generated in the absence of the stimulus

96
PREPRINT 8

(Nspikes ), normalized by the network size (total number of [10] K. M. Spencer, D. F. Salisbury, M. E. Shenton, and
neurons Nn ) and the duration of the time window T . R. W. McCarley, “γ-band auditory steady-state responses
are impaired in first episode psychosis,” Biological
psychiatry, vol. 64, no. 5, pp. 369–375, 2008.
S
Nspikes − Nspikes [11] K. M. Spencer, P. G. Nestor, R. Perlmutter, M. A.
R= (6) Niznikiewicz, M. C. Klump, M. Frumin, M. E. Shen-
T Nn
ton, and R. W. McCarley, “Neural synchrony indexes
disordered perception and cognition in schizophrenia,”
Proceedings of the National Academy of Sciences, vol.
101, no. 49, pp. 17 288–17 293, 2004.
ACKNOWLEDGMENTS
[12] K. M. Spencer, M. A. Niznikiewicz, P. G. Nestor,
This research was supported by the Centre National de la M. E. Shenton, and R. W. McCarley, “Left auditory
Recherche Scientifique (CNRS) and the European Community cortex gamma synchronization and auditory hallucination
(Human Brain Project, H2020-785907). E.S. acknowledges a symptoms in schizophrenia,” BMC neuroscience, vol. 10,
PhD fellowship from the École des Neurosciences de Paris no. 1, pp. 1–13, 2009.
(ENP) and from the Fondation pour la Recherche Médicale [13] H. Gunduz-Bruce, “The acute effects of nmda antago-
(FRM) - grant FDT202012010566 and the financial support nism: from the rodent to the human brain,” Brain research
from La Fondation des Treilles. reviews, vol. 60, no. 2, pp. 279–286, 2009.
[14] D. C. Javitt and S. R. Zukin, “Recent advances in the
R EFERENCES phencyclidine model of schizophrenia.” The American
[1] D. A. Lewis, T. Hashimoto, and D. W. Volk, “Cortical journal of psychiatry, 1991.
inhibitory neurons and schizophrenia,” Nature Reviews [15] J. H. Krystal, L. P. Karper, J. P. Seibyl, G. K. Freeman,
Neuroscience, vol. 6, no. 4, pp. 312–324, 2005. R. Delaney, J. D. Bremner, G. R. Heninger, M. B.
[2] V. P. Bozikas and C. Andreou, “Longitudinal studies of Bowers, and D. S. Charney, “Subanesthetic effects of
cognition in first episode psychosis: a systematic review the noncompetitive nmda antagonist, ketamine, in hu-
of the literature,” Australian & New Zealand Journal of mans: psychotomimetic, perceptual, cognitive, and neu-
Psychiatry, vol. 45, no. 2, pp. 93–108, 2011. roendocrine responses,” Archives of general psychiatry,
[3] T. Su, Y. Lu, Y. Geng, W. Lu, and Y. Chen, “How could vol. 51, no. 3, pp. 199–214, 1994.
n-methyl-d-aspartate receptor antagonists lead to exci- [16] S. S. Kalsi, D. M. Wood, and P. I. Dargan, “The epi-
tation instead of inhibition?” Brain Science Advances, demiology and patterns of acute and chronic toxicity
vol. 4, no. 2, pp. 73–98, 2018. associated with recreational ketamine use,” Emerging
[4] P. J. Uhlhaas and W. Singer, “Abnormal neural oscilla- Health Threats Journal, vol. 4, no. 1, p. 7107, 2011.
tions and synchrony in schizophrenia,” Nature reviews [17] A. D. Shaw, N. Saxena, L. E. Jackson, J. E. Hall, K. D.
neuroscience, vol. 11, no. 2, pp. 100–113, 2010. Singh, and S. D. Muthukumaraswamy, “Ketamine ampli-
[5] M. E. Shenton, C. C. Dickey, M. Frumin, and R. W. fies induced gamma frequency oscillations in the human
McCarley, “A review of mri findings in schizophrenia,” cerebral cortex,” European Neuropsychopharmacology,
Schizophrenia research, vol. 49, no. 1-2, pp. 1–52, 2001. vol. 25, no. 8, pp. 1136–1146, 2015.
[6] G. Flynn, D. Alexander, A. Harris, T. Whitford, [18] G. Plourde, J. Baribeau, and V. Bonhomme, “Ketamine
W. Wong, C. Galletly, S. Silverstein, E. Gordon, increases the amplitude of the 40-hz auditory steady-
and L. M. Williams, “Increased absolute magni- state response in humans.” British journal of anaesthesia,
tude of gamma synchrony in first-episode psychosis,” vol. 78, no. 5, pp. 524–529, 1997.
Schizophrenia research, vol. 105, no. 1-3, pp. 262–271, [19] L. E. Hong, A. Summerfelt, R. W. Buchanan,
2008. P. O’donnell, G. K. Thaker, M. A. Weiler, and A. C.
[7] T. Grent, J. Gross, J. Goense, M. Wibral, R. Gajwani, Lahti, “Gamma and delta neural oscillations and as-
A. I. Gumley, S. M. Lawrie, M. Schwannauer, sociation with clinical symptoms under subanesthetic
F. Schultze-Lutter, T. N. Schröder, et al., “Resting-state ketamine,” Neuropsychopharmacology, vol. 35, no. 3, pp.
gamma-band power alterations in schizophrenia reveal 632–640, 2010.
e/i-balance abnormalities across illness-stages,” Elife, [20] D. Pinault, “N-methyl d-aspartate receptor antagonists
vol. 7, p. e37799, 2018. ketamine and mk-801 induce wake-related aberrant γ
[8] A. Perrottelli, G. M. Giordano, F. Brando, L. Giuliani, oscillations in the rat neocortex,” Biological psychiatry,
and A. Mucci, “Eeg-based measures in at-risk mental vol. 63, no. 8, pp. 730–735, 2008.
state and early stages of schizophrenia: A systematic [21] B. Kocsis, “Differential role of nr2a and nr2b subunits
review,” Frontiers in psychiatry, vol. 12, p. 582, 2021. in n-methyl-d-aspartate receptor antagonist-induced aber-
[9] C. Mulert, V. Kirsch, R. Pascual-Marqui, R. W. rant cortical gamma oscillations,” Biological psychiatry,
McCarley, and K. M. Spencer, “Long-range syn- vol. 71, no. 11, pp. 987–995, 2012.
chrony of gamma oscillations and auditory hallucination [22] J. Wood, Y. Kim, and B. Moghaddam, “Disruption of
symptoms in schizophrenia,” International Journal of prefrontal cortex large scale neuronal activity by dif-
Psychophysiology, vol. 79, no. 1, pp. 55–63, 2011. ferent classes of psychotomimetic drugs,” Journal of

97
PREPRINT 9

Neuroscience, vol. 32, no. 9, pp. 3022–3031, 2012. [34] Y. Zhang, M. M. Behrens, and J. E. Lisman, “Prolonged
[23] M. J. Nicolás, J. López-Azcárate, M. Valencia, M. Ale- exposure to nmdar antagonist suppresses inhibitory
gre, M. Pérez-Alcázar, J. Iriarte, and J. Artieda, synaptic transmission in prefrontal cortex,” Journal of
“Ketamine-induced oscillations in the motor circuit of neurophysiology, vol. 100, no. 2, pp. 959–965, 2008.
the rat basal ganglia,” PloS one, vol. 6, no. 7, p. e21814, [35] V. D. Lazzaro, A. Oliviero, P. Profice, M. Pennisi,
2011. F. Pilato, G. Zito, M. Dileone, R. Nicoletti, P. Pasqualetti,
[24] M. Slovik, B. Rosin, S. Moshel, R. Mitelman, E. Schecht- and P. Tonali, “Ketamine increases human motor cortex
man, R. Eitan, A. Raz, and H. Bergman, “Ketamine in- excitability to transcranial magnetic stimulation,” The
duced converged synchronous gamma oscillations in the Journal of physiology, vol. 547, no. 2, pp. 485–496,
cortico-basal ganglia network of nonhuman primates,” 2003.
Journal of neurophysiology, vol. 118, no. 2, pp. 917– [36] R. E. Hoffman and I. Cavus, “Slow transcranial magnetic
931, 2017. stimulation, long-term depotentiation, and brain hyperex-
[25] B. Moghaddam, B. Adams, A. Verma, and D. Daly, “Ac- citability disorders,” American Journal of Psychiatry, vol.
tivation of glutamatergic neurotransmission by ketamine: 159, no. 7, pp. 1093–1102, 2002.
a novel step in the pathway from nmda receptor blockade [37] Z. J. Daskalakis, P. B. Fitzgerald, and B. K. Christensen,
to dopaminergic and cognitive disruptions associated “The role of cortical inhibition in the pathophysiology
with the prefrontal cortex,” Journal of Neuroscience, and treatment of schizophrenia,” Brain research reviews,
vol. 17, no. 8, pp. 2921–2927, 1997. vol. 56, no. 2, pp. 427–442, 2007.
[26] A. C. Lahti, H. H. Holcomb, D. R. Medoff, and C. A. [38] R. E. Hoffman, K. A. Hawkins, R. Gueorguieva, N. N.
Tamminga, “Ketamine activates psychosis and alters lim- Boutros, F. Rachid, K. Carroll, and J. H. Krystal, “Tran-
bic blood flow in schizophrenia.” Neuroreport, vol. 6, scranial magnetic stimulation of left temporoparietal
no. 6, pp. 869–872, 1995. cortex and medication-resistant auditory hallucinations,”
[27] A. Breier, A. K. Malhotra, D. A. Pinals, N. I. Weisen- Archives of general psychiatry, vol. 60, no. 1, pp. 49–56,
feld, and D. Pickar, “Association of ketamine-induced 2003.
psychosis with focal activation of the prefrontal cortex in [39] L. B. Merabet, M. Kobayashi, J. Barton, and A. Pascual-
healthy volunteers.” The American journal of psychiatry, Leone, “Suppression of complex visual hallucinatory ex-
1997. periences by occipital transcranial magnetic stimulation:
[28] F. Vollenweider, K. Leenders, I. Øye, D. Hell, a case report,” Neurocase, vol. 9, no. 5, pp. 436–440,
and J. Angst, “Differential psychopathology and 2003.
patterns of cerebral glucose utilisation produced [40] E. Susin and A. Destexhe, “Integration, coincidence
by (s)-and (r)-ketamine in healthy volunteers using detection and resonance in networks of spiking neurons
positron emission tomography (pet),” European expressing gamma oscillations and asynchronous states,”
Neuropsychopharmacology, vol. 7, no. 1, pp. 25– bioRxiv, 2021.
38, 1997. [41] W. Singer and C. M. Gray, “Visual feature integration
[29] Y. Suzuki, E. Jodo, S. Takeuchi, S. Niwa, and Y. Kayama, and the temporal correlation hypothesis,” Annual review
“Acute administration of phencyclidine induces tonic of neuroscience, vol. 18, no. 1, pp. 555–586, 1995.
activation of medial prefrontal cortex neurons in freely [42] W. Singer, “Neuronal synchrony: a versatile code for the
moving rats,” Neuroscience, vol. 114, no. 3, pp. 769–779, definition of relations?” Neuron, vol. 24, no. 1, pp. 49–
2002. 65, 1999.
[30] M. E. Jackson, H. Homayoun, and B. Moghaddam, [43] J. O’Keefe and M. L. Recce, “Phase relationship between
“Nmda receptor hypofunction produces concomitant fir- hippocampal place units and the eeg theta rhythm,”
ing rate potentiation and burst activity reduction in the Hippocampus, vol. 3, no. 3, pp. 317–330, 1993.
prefrontal cortex,” Proceedings of the National Academy [44] P. Fries, D. Nikolić, and W. Singer, “The gamma cycle,”
of Sciences, vol. 101, no. 22, pp. 8467–8472, 2004. Trends in neurosciences, vol. 30, no. 7, pp. 309–316,
[31] D. C. Rotaru, H. Yoshino, D. A. Lewis, G. B. Er- 2007.
mentrout, and G. Gonzalez-Burgos, “Glutamate recep- [45] P. Fries, “A mechanism for cognitive dynamics: neuronal
tor subtypes mediating synaptic activation of prefrontal communication through neuronal coherence,” Trends in
cortex neurons: relevance for schizophrenia,” Journal of cognitive sciences, vol. 9, no. 10, pp. 474–480, 2005.
Neuroscience, vol. 31, no. 1, pp. 142–156, 2011. [46] ——, “Rhythms for cognition: communication through
[32] H. Homayoun and B. Moghaddam, “Nmda receptor coherence,” Neuron, vol. 88, no. 1, pp. 220–235, 2015.
hypofunction produces opposite effects on prefrontal [47] B. Pesaran, J. S. Pezaris, M. Sahani, P. P. Mitra, and
cortex interneurons and pyramidal neurons,” Journal of R. A. Andersen, “Temporal structure in neuronal activity
Neuroscience, vol. 27, no. 43, pp. 11 496–11 500, 2007. during working memory in macaque parietal cortex,”
[33] A. J. Widman and L. L. McMahon, “Disinhibition of ca1 Nature neuroscience, vol. 5, no. 8, pp. 805–811, 2002.
pyramidal cells by low-dose ketamine and other antago- [48] L. L. Colgin, T. Denninger, M. Fyhn, T. Hafting, T. Bon-
nists with rapid antidepressant efficacy,” Proceedings of nevie, O. Jensen, M.-B. Moser, and E. I. Moser, “Fre-
the National Academy of Sciences, vol. 115, no. 13, pp. quency of gamma oscillations routes flow of information
E3007–E3016, 2018. in the hippocampus,” Nature, vol. 462, no. 7271, pp. 353–

98
PREPRINT 10

357, 2009. global brain states,” Frontiers in systems neuroscience,

[49] M. F. Carr, M. P. Karlsson, and L. M. Frank, “Transient vol. 13, p. 75, 2019.
slow gamma synchrony underlies hippocampal memory [63] W. R. Softky and C. Koch, “The highly irregular firing of
replay,” Neuron, vol. 75, no. 4, pp. 700–713, 2012. cortical cells is inconsistent with temporal integration of
[50] A. Rougeul-Buser, J. Bouyer, and P. Buser, “From atten- random epsps,” Journal of neuroscience, vol. 13, no. 1,
tiveness to sleep. a topographical analysis of localized” pp. 334–350, 1993.
synchronized” activities on the cortex of normal cat and [64] G. R. Holt, W. R. Softky, C. Koch, and R. J. Douglas,
monkey,” Acta Neurobiol Exp (Warsz), vol. 35, no. 5-6, “Comparison of discharge variability in vitro and in vivo
pp. 805–19, 1975. in cat visual cortex neurons,” Journal of neurophysiology,
[51] J. Bouyer, M. Montaron, and A. Rougeul, “Fast fronto- vol. 75, no. 5, pp. 1806–1814, 1996.
parietal rhythms during combined focused attentive [65] M. N. Shadlen and W. T. Newsome, “The variable
behaviour and immobility in cat: cortical and thala- discharge of cortical neurons: implications for connec-
mic localizations,” Electroencephalography and clinical tivity, computation, and information coding,” Journal of
neurophysiology, vol. 51, no. 3, pp. 244–252, 1981. neuroscience, vol. 18, no. 10, pp. 3870–3896, 1998.
[52] E. Rodriguez, N. George, J.-P. Lachaux, J. Martinerie, [66] A. Destexhe, M. Rudolph, and D. Paré, “The high-
B. Renault, and F. J. Varela, “Perception’s shadow: conductance state of neocortical neurons in vivo,” Nature
long-distance synchronization of human brain activity,” reviews neuroscience, vol. 4, no. 9, pp. 739–751, 2003.
Nature, vol. 397, no. 6718, pp. 430–433, 1999. [67] J. A. Henrie and R. Shapley, “Lfp power spectra in v1
[53] L. Melloni, C. Molina, M. Pena, D. Torres, W. Singer, cortex: the graded effect of stimulus contrast,” Journal of
and E. Rodriguez, “Synchronization of neural activity neurophysiology, vol. 94, no. 1, pp. 479–490, 2005.
across cortical areas correlates with conscious percep- [68] P. Tiesinga and T. J. Sejnowski, “Cortical enlightenment:
tion,” Journal of neuroscience, vol. 27, no. 11, pp. 2858– are attentional gamma oscillations driven by ing or
2865, 2007. ping?” Neuron, vol. 63, no. 6, pp. 727–732, 2009.
[54] P. Fries, J. H. Reynolds, A. E. Rorie, and R. Desimone, [69] D. S. Ling and L. S. Benardo, “Activity-dependent de-
“Modulation of oscillatory neuronal synchronization by pression of monosynaptic fast ipscs in hippocampus:
selective visual attention,” Science, vol. 291, no. 5508, contributions from reductions in chloride driving force
pp. 1560–1563, 2001. and conductance,” Brain research, vol. 670, no. 1, pp.
[55] G. G. Gregoriou, S. J. Gotts, H. Zhou, and R. Desimone, 142–146, 1995.
“High-frequency, long-range coupling between prefrontal [70] H. C. Grunze, D. G. Rainnie, M. E. Hasselmo, E. Barkai,
and visual cortex during attention,” science, vol. 324, no. E. F. Hearn, R. W. McCarley, and R. W. Greene, “Nmda-
5931, pp. 1207–1210, 2009. dependent modulation of ca1 local circuit inhibition,”
[56] M. Vinck, T. Womelsdorf, E. A. Buffalo, R. Desi- Journal of Neuroscience, vol. 16, no. 6, pp. 2034–2043,
mone, and P. Fries, “Attentional modulation of cell-class- 1996.
specific gamma-band synchronization in awake monkey [71] D. Ling and L. S. Benardo, “Recruitment of gabaa inhibi-
area v4,” Neuron, vol. 80, no. 4, pp. 1077–1089, 2013. tion in rat neocortex is limited and not nmda dependent,”
[57] S. Rouhinen, J. Panula, J. M. Palva, and S. Palva, “Load Journal of Neurophysiology, vol. 74, no. 6, pp. 2329–
dependence of β and γ oscillations predicts individual 2335, 1995.
capacity of visual attention,” Journal of Neuroscience, [72] T. Karayannis, I. Huerta-Ocampo, and M. Capogna,
vol. 33, no. 48, pp. 19 023–19 033, 2013. “Gabaergic and pyramidal neurons of deep cortical layers
[58] D. E. Sheer, “Focused arousal and the cognitive 40- directly receive and differently integrate callosal input,”
hz event-related potentials: differential diagnosis of Cerebral Cortex, vol. 17, no. 5, pp. 1213–1226, 2007.
alzheimer’s disease.” Progress in clinical and biological [73] N. C. Jones, M. Reddy, P. Anderson, M. R. Salzberg, T. J.
research, vol. 317, pp. 79–94, 1989. O’Brien, and D. Pinault, “Acute administration of typical
[59] T. Womelsdorf, P. Fries, P. P. Mitra, and R. Desimone, and atypical antipsychotics reduces eeg gamma power,
“Gamma-band synchronization in visual cortex predicts but only the preclinical compound ly379268 reduces the
speed of change detection,” Nature, vol. 439, no. 7077, ketamine-induced rise in gamma power,” International
pp. 733–736, 2006. Journal of Neuropsychopharmacology, vol. 15, no. 5, pp.
[60] N. Brunel, “Dynamics of sparsely connected networks 657–668, 2012.
of excitatory and inhibitory spiking neurons,” Journal of [74] T. Hakami, N. C. Jones, E. A. Tolmacheva, J. Gaudias,
computational neuroscience, vol. 8, no. 3, pp. 183–208, J. Chaumont, M. Salzberg, T. J. O’Brien, and D. Pinault,
2000. “Nmda receptor hypofunction leads to generalized and
[61] C. Koch, M. Massimini, M. Boly, and G. Tononi, “Neural persistent aberrant γ oscillations independent of hyper-
correlates of consciousness: progress and problems,” locomotion and the state of consciousness,” PloS one,
Nature Reviews Neuroscience, vol. 17, no. 5, pp. 307– vol. 4, no. 8, p. e6755, 2009.
321, 2016. [75] K.-H. Lee, L. M. Williams, M. Breakspear, and E. Gor-
[62] J. S. Goldman, N. Tort-Colet, M. Di Volo, E. Susin, don, “Synchronous gamma activity: a review and
J. Bouté, M. Dali, M. Carlu, T.-A. Nghiem, T. Górski, contribution to an integrative neuroscience model of
and A. Destexhe, “Bridging single neuron dynamics to schizophrenia,” Brain Research Reviews, vol. 41, no. 1,

99
PREPRINT 11

pp. 57–78, 2003. vol. 34, no. 5, pp. 944–961, 2008.

[76] S. Akbarian, J. J. Kim, S. G. Potkin, J. O. Hagman, [88] K. M. Spencer, “The functional consequences of cor-
A. Tafazzoli, W. E. Bunney, and E. G. Jones, “Gene ex- tical circuit abnormalities on gamma oscillations in
pression for glutamic acid decarboxylase is reduced with- schizophrenia: insights from computational modeling,”
out loss of neurons in prefrontal cortex of schizophren- Frontiers in human neuroscience, vol. 3, p. 33, 2009.
ics,” Archives of general psychiatry, vol. 52, no. 4, pp. [89] R. Brette and W. Gerstner, “Adaptive exponential
258–266, 1995. integrate-and-fire model as an effective description of
[77] D. W. Volk, M. C. Austin, J. N. Pierri, A. R. Sampson, neuronal activity,” Journal of neurophysiology, vol. 94,
and D. A. Lewis, “Decreased glutamic acid decarboxy- no. 5, pp. 3637–3642, 2005.
lase67 messenger rna expression in a subset of prefrontal [90] D. S. Faber and H. Korn, “Single-shot channel activation
cortical γ-aminobutyric acid neurons in subjects with accounts for duration of inhibitory postsynaptic poten-
schizophrenia,” Archives of general psychiatry, vol. 57, tials in a central neuron,” Science, vol. 208, no. 4444,
no. 3, pp. 237–245, 2000. pp. 612–615, 1980.
[78] S. Akbarian and H.-S. Huang, “Molecular and cellu- [91] M. Perouansky and Y. Yaari, “Kinetic properties of nmda
lar mechanisms of altered gad1/gad67 expression in receptor-mediated synaptic currents in rat hippocampal
schizophrenia and related disorders,” Brain research pyramidal cells versus interneurones.” The Journal of
reviews, vol. 52, no. 2, pp. 293–304, 2006. Physiology, vol. 465, no. 1, pp. 223–244, 1993.
[79] M. Inan, T. J. Petros, and S. A. Anderson, “Losing [92] T. Götz, U. Kraushaar, J. Geiger, J. Lübke, T. Berger,
your inhibition: linking cortical gabaergic interneurons and P. Jonas, “Functional properties of ampa and nmda
to schizophrenia,” Neurobiology of disease, vol. 53, pp. receptors expressed in identified types of basal ganglia
36–48, 2013. neurons,” Journal of Neuroscience, vol. 17, no. 1, pp.
[80] D. Eyles, J. McGrath, and G. Reynolds, “Neu- 204–215, 1997.
ronal calcium-binding proteins and schizophrenia,” [93] M. C. Bellingham, R. Lim, and B. Walmsley, “Develop-
Schizophrenia research, vol. 57, no. 1, pp. 27–34, 2002. mental changes in epsc quantal size and quantal content
[81] T. Korotkova, E. C. Fuchs, A. Ponomarenko, J. von at a central glutamatergic synapse in rat,” The Journal of
Engelhardt, and H. Monyer, “Nmda receptor ablation on Physiology, vol. 511, no. 3, pp. 861–869, 1998.
parvalbumin-positive interneurons impairs hippocampal [94] C. E. Jahr and C. F. Stevens, “Voltage dependence
synchrony, spatial representations, and working mem- of nmda-activated macroscopic conductances predicted
ory,” Neuron, vol. 68, no. 3, pp. 557–569, 2010. by single-channel kinetics,” Journal of Neuroscience,
[82] J. E. Belforte, V. Zsiros, E. R. Sklar, Z. Jiang, G. Yu, vol. 10, no. 9, pp. 3178–3182, 1990.
Y. Li, E. M. Quinlan, and K. Nakazawa, “Postnatal nmda [95] Y. Zerlaut and A. Destexhe, “Enhanced responsiveness
receptor ablation in corticolimbic interneurons confers and low-level awareness in stochastic network states,”
schizophrenia-like phenotypes,” Nature neuroscience, Neuron, vol. 94, no. 5, pp. 1002–1009, 2017.
vol. 13, no. 1, pp. 76–83, 2010. [96] H.-X. Wang and W.-J. Gao, “Cell type-specific devel-
[83] M. Carlen, K. Meletis, J. Siegle, J. Cardin, K. Fu- opment of nmda receptors in the interneurons of rat
tai, D. Vierling-Claassen, C. Ruehlmann, S. R. Jones, prefrontal cortex,” Neuropsychopharmacology, vol. 34,
K. Deisseroth, M. Sheng, et al., “A critical role for nmda no. 8, pp. 2028–2040, 2009.
receptors in parvalbumin interneurons for gamma rhythm [97] M. Stimberg, R. Brette, and D. F. Goodman, “Brian 2,
induction and behavior,” Molecular psychiatry, vol. 17, an intuitive and efficient neural simulator,” eLife, vol. 8,
no. 5, pp. 537–548, 2012. p. e47314, Aug. 2019.
[84] E. N. Billingslea, V. M. Tatard-Leitman, J. Anguiano, [98] B. Telenczuk, M. Telenczuk, and A. Destexhe, “A kernel-
C. R. Jutzeler, J. Suh, J. A. Saunders, S. Morita, R. E. based method to calculate local field potentials from
Featherstone, P. I. Ortinski, M. J. Gandal, et al., “Parval- networks of spiking neurons,” bioRxiv, 2020.
bumin cell ablation of nmda-r1 causes increased resting
network excitability with associated social and self-care
deficits,” Neuropsychopharmacology, vol. 39, no. 7, pp.
1603–1613, 2014.
[85] K. Nakao and K. Nakazawa, “Brain state-dependent ab-
normal lfp activity in the auditory cortex of a schizophre-
nia mouse model,” Frontiers in neuroscience, vol. 8, p.
168, 2014.
[86] G. Gonzalez-Burgos and D. A. Lewis, “Nmda receptor
hypofunction, parvalbumin-positive neurons, and corti-
cal gamma oscillations in schizophrenia,” Schizophrenia
bulletin, vol. 38, no. 5, pp. 950–957, 2012.
[87] ——, “Gaba neurons and the mechanisms of net-
work oscillations: implications for understanding cortical
dysfunction in schizophrenia,” Schizophrenia bulletin,

100
 Part III

Discussion

101
 Discussion

One of the main goals of neuroscience is to understand the underlying processes of

cognition, that is, how information is encoded, decoded and processed in the brain [293].
Several ways of encoding information have been hypothesized, such as firing rates [294,
295], pairwise correlations [296, 297], spike pattern irregularity [298, 299, 300, 301] and
spike packets [302], among others.
An important aspect to be taken into consideration, when exploring cognitive mecha-
nisms and their related information processes of coding, decoding and transmission, is the
environment in which the information is treated. It has been shown, in different studies
and protocols, that the brain state promotes major changes in the way the external and
internal stimuli are processed. For example, it is well known that attention modulates
how neurons respond to external stimuli throughout visual cortex [303]. The same way,
the level of responses vary with sleep stage [304] and level of anesthesia [305].
In this thesis we explored in different ways how the activity of individual neurons
and their collective behavior interfere in information transmission in the network. We
devoted our attention to three particular brain states: Slow-wave sleep, Asynchronous-
and-Irregular activity (AI) and Gamma activity. In particular, Gamma and AI activity
were studied both in normal and in pathological conditions associated to schizoid brain
conditions. To do this, we analyzed LFP and single units in the human cortex, provided
by one of our collaborators Sydney S. Cash (Department of Neurology, Massachusetts
General Hospital and Harvard Medical School, Boston). In this analysis, we evaluated
spiking patterns (firing rate, spike regularity and spiking correlations) and the neuronal
participation on Gamma rhythms in Awake and Slow-wave sleep states. In Work 1, we
compared previously published results with our own analysis, stressing differences and
similarities between these two states, with respect to neuronal activity and its contribu-
tions on Gamma rhythms.
To investigate the principles of information transmission in the brain, we developed
network models constrained by this analysis. These models were capable of generating
Gamma rhythms and Asynchronous-and-Irregular activity with spiking activity similar
to experimental observations, like it is described in Work 2. We developed three different
models, each one capable of generating Gamma rhythms by a different mechanism: PING,
ING or CHING. The transfer of information was studied in these models, both during
Gamma and AI states. Information transmission was measured by means of network
responsiveness, that is, the the capacity level of the network to produce additional spikes
in response to a certain stimulus. This type of measurement, associated to the change in

103
spiking probability after a stimulus, is easily comparable to experimental measurements,
obtained, for example, by EEG, LFP, Calcium imaging, etc. In addition, networks with
different level of responsiveness can exhibit drastic changes in their behavior. For example,
a decrease in responsiveness imposes important restrictions on the conditions for a cell
to fire with respect to the number of necessary coincidental inputs, forbidding network
responses due to weak stimuli. Conversely, networks with high levels of responsiveness can
easily detect single afferent stimuli, that would be subthreshold otherwise.
In Work 2, we compared network responsiveness during Gamma and AI states in
different paradigms, and found that in all cases, the presence of Gamma oscillations tends
to diminish the responsiveness to external inputs. Our models allowed the interpretation
that Asynchronous-and-Irregular states would be better suited to stimulus detection, al-
lowing the perception of even low amplitude stimuli (due to their high sensitivity), while
Gamma oscillating states, at the expense of loosing stimulus sensitivity, would be able to
encode temporal information. This interpretation comes from the Phase Coding Theory
[128], which states that input strengths could be converted into spike-timing relative to
the cycle (most excited fire earlier in the cycle, while cells that are not excited enough are
prohibited to spike). This was observed in our models by the phase-dependence respon-
siveness present in Gamma states, and absent in AI states.
In a third study, Work 3, we compared network responsiveness in healthy and patho-
logical conditions related to the schizoid brain, during Gamma oscillations and AI states.
The schizoid condition was simulated in our networks by means of different levels of
NMDA receptors block in the respective channels of the models. In this work, as in Work
2, a similar result with respect to Gamma and AI states responsiveness was observed. In
healthy and pathological conditions (with all levels of NMDA receptors block), AI states
displayed a bigger responsiveness than Gamma states. The interesting observation, on
the other hand, was that responsiveness increased with the severity of NMDA receptors
block, meaning that pathological states were in general, both during Gamma and AI
states, more responsive than healthy states. This observation led us to the interpretation
that the hallucinations observed during psychotic episodes in schizophrenic patients could
be potentially related to an incapacity of the neuronal network to process all the stimuli
in the environment, since the network would be excessively responsive, responsive even to
stimuli that in normal conditions would be ignored.
Furthermore, our experimental analysis, compared to previous studies on cellular cor-
relates of Gamma oscillations [47], confirmed a sparse participation of neuronal activity
during Gamma rhythms. We observed two forms of neuronal participation to the rhythms:
phase-locking and/or firing rate increase. FS cells presented significant higher level of
phase-locking, firing rate increase, and behavior consistency than RS cells. In addition,
we identified that the group of Gamma participating cells changes with time as well as
their phase-preference. The combination of this analysis with the developed theoretical

104
work brought important theoretical insights. In Work 2, while most of these features
could be obtained by choosing parameters which set the network dynamics into a firing
rate regime, the right level of firing rate increase could only be reproduced with a partic-
ular network configuration, in the ING Network. This network was also the one in which
the difference in responsiveness between AI and Gamma was smaller. Furthermore, we
showed (Work 2) that the resonant properties of each of the developed models (PING,
ING and CHING) strongly depended on the patterns of connections and on the neuronal
composition. These examples illustrate how network structure drastically change network
dynamics and consequently how the information is transmitted by a network.
Another important aspect explored in the thesis is the question if Gamma oscilla-
tions have an active role in information processing or alternatively, if they are just an
epiphenomenon of this process. Several hypotheses concerning the Gamma role have been
proposed. The most popular theories are the Binding-by-synchronization Hypothesis [125,
126], the Phase Coding Theory [127, 128], the Communication Through Coherence Theory
[129, 130] and Communication through Resonance Theory [131]. Even though exploring
these hypothesis was one of the initial ambitions of this thesis, our models could not
provide any evidences in favor or against them. On the contrary, these theories guided us
to allow the interpretation of our results. In this perspective, for example, if the Commu-
nication through Resonance Theory [131] is indeed one of the mechanism of information
transmission, network structure and network resonant properties should be explored in
future models.
Albeit no conclusion could be made with respect to the role of Gamma oscillations
per se, in Work 3 our models could give a possible interpretation for the role of Gamma
oscillations increase in pathological conditions. In our study, we observed an increase of
responsiveness in conditions associated to the hyper-production of Gamma, such as in
early schizophrenic patients. This increase of responsiveness was accompanied by Gamma
power increase. However, since the same effects were also observed during AI states, this
increase of responsiveness was not a direct effect of this Gamma increase, but rather, an
effect of the increase of excitability generated by alteration of NMDA receptors. In this
situation, the Gamma power increase appeared to be a simple side effect of altered or
anomalous NMDA function, suggesting that Gamma oscillations have no direct role per
se in the pathological condition.

In summary, in this thesis we performed an analysis of experimental data from the

human cortex, analyzing how neurons participate to Gamma rhythms. We observed that:
(1) there are two types of neuronal participation, phase-locking and firing rate increase,
that (2) neuronal participation of Gamma rhythms is sparse and that (3) inhibitory
neurons participate to the rhythm more expressively than excitatory neurons. Based on

105
these features, we built and constrained three network models generating Gamma by
three different mechanisms: PING, ING and CHING. These models were used as a tool
to explore the principles of information transmission in the brain, which was accessed by
the measurement of network responsiveness. Our models indicated that network structure
should be the center of future studies, since it affects different dynamical properties, such
as responsiveness. We explored network responsiveness in healthy and in pathological
states similar to that seen in schizophrenia disease, and found that Gamma oscillations
diminish network responsiveness in comparison to AI, in both, healthy and pathological
conditions. Furthermore, our models indicated that schizoid pathological states were more
responsive than healthy states, providing a possible comprehension of positive symptoms
associated to schizophrenia. In our interpretation, these symptoms are possibly caused
by the hypofunction of NMDA channels, rather than by the the excess of Gamma. In
this perspective, we suggest that the increase of Gamma oscillations, observed during
early phases of the disease, represents a side effect of the increase of neuronal synchrony,
generated by the NMDA anomalies present in the disease.

106
 Part IV

Complement Research Article

107
 Work 4

Bridging Single Neuron Dynamics to

Global Brain States

Reference:
Goldman, J. S., Tort-Colet, N., Di Volo, M., Susin, E., Bouté, J., Dali, M., Carlu, M.
Nghiem, T., Górski, T. & Destexhe, A. (2019). Bridging single neuron dynamics to global
brain states. Frontiers in systems neuroscience, 13, 75.

French Abstract:

Les réseaux de neurones biologiques produisent de l’information de fonds spontanés à

plusieurs échelles. L’activité, qui encapsule ces informations, devient plus complexe dans
les états cérébraux qui affichent des capacités plus élevées pour la cognition. Dans cette
perspective, les états d’éveil attentif sont plus complexes que les états endormis ou anesthésiés,
par exemple. Dans cet article, nous couvrant dès les courants des canaux ioniques (micro-
échelle) à l’échelle du cerveau entier (macro-échelle), en exposant différentes études à
propos des mécanismes qui contrôlent les différents états dynamiques du cerveau. D’une
façon similaire dont les interactions microscopiques entre les molécules sont liées à des
structures macroscopiques de la matière, en utilisant la physique statistique, la dynamique
des phénomènes neuronaux peuvent être liés à la dynamique macroscopique du cerveau
à travers des échelles mésoscopiques. Pendant des états de conscience, on observe que
des stimuli externes provoquent des effondrements de complexité au-delà des dynamiques
spontanées, accompagnées d’une dynamique multi-dimensionnel, asynchrone et irrégulier.
En revanche, la complexité des états d’inconscience ne peut pas être davantage réduite au-
delà de la synchronie et de la régularité caractéristiques des activité spontanée observées
pendant ces états. Dans cet article, nous proposons que l’augmentation de dimensionnal-
ité des états dynamique spontanée observées pendant les états de conscience promeut la
réactivité du cerveau, améliorant la capacité émergente des réseaux de neurones à encoder
des informations sur plusieurs échelles.

109
 PERSPECTIVE
published: 06 December 2019
doi: 10.3389/fnsys.2019.00075

Bridging Single Neuron Dynamics to

Global Brain States
Jennifer S. Goldman 1*, Núria Tort-Colet 1 , Matteo di Volo 1† , Eduarda Susin 1 , Jules Bouté 1 ,
Melissa Dali 1 , Mallory Carlu 1 , Trang-Anh Nghiem 2 , Tomasz Górski 1 and Alain Destexhe 1
1
Department of Integrative and Computational Neuroscience (ICN), Centre National de la Recherche Scientifique (CNRS),
Paris-Saclay Institute of Neuroscience (NeuroPSI), Gif-sur-Yvette, France, 2 Department of Physics, Ecole Normale
Supérieure, Paris, France

Biological neural networks produce information backgrounds of multi-scale spontaneous

activity that become more complex in brain states displaying higher capacities for
cognition, for instance, attentive awake versus asleep or anesthetized states. Here, we
review brain state-dependent mechanisms spanning ion channel currents (microscale)
to the dynamics of brain-wide, distributed, transient functional assemblies (macroscale).
Not unlike how microscopic interactions between molecules underlie structures formed
in macroscopic states of matter, using statistical physics, the dynamics of microscopic
neural phenomena can be linked to macroscopic brain dynamics through mesoscopic
scales. Beyond spontaneous dynamics, it is observed that stimuli evoke collapses of
complexity, most remarkable over high dimensional, asynchronous, irregular background
dynamics during consciousness. In contrast, complexity may not be further collapsed
Edited by:
Per E. Roland, beyond synchrony and regularity characteristic of unconscious spontaneous activity. We
University of Copenhagen, Denmark propose that increased dimensionality of spontaneous dynamics during conscious states
Reviewed by: supports responsiveness, enhancing neural networks’ emergent capacity to robustly
Mario Rosanova,
University of Milan, Italy
encode information over multiple scales.
Sacha Jennifer van Albada,
Keywords: computational neuroscience, neural network models, mean-field models, membrane biophysics,
Julich Research Centre, Germany
low-dimensional manifold, cerebral cortex, coupling, desynchronized
*Correspondence:
Jennifer S. Goldman
[email protected] INTRODUCTION
† Presentaddress:
Matteo di Volo,
Brain activity transitions between healthy states, including stages of sleep, restful and aroused
Laboratoire de Physique Théorique et waking, as well as pathological states such as epilepsy, coma, and unresponsive wakefulness
Modelisation, Université de syndrome. From such a diversity of brain states, phenomenological categories encompassing
Cergy-Pontoise, Cergy-Pontoise, similar spatio-temporal activity patterns can roughly, but usefully, be defined: unconscious (e.g.,
France sleep and anesthesia) and conscious (e.g., waking and dreaming) brain states. At the macroscopic,
global scale, unconscious brain states are dominated by high voltage, low frequency oscillatory
Received: 28 June 2019 brain activity related to the microscopic alternation of synchronous neuronal spiking and near
Accepted: 19 November 2019 silence (Steriade et al., 1993; Brown et al., 2010). Conversely, conscious states are macroscopically
Published: 06 December 2019
characterized by low voltage, high frequency, complex “disorganized” dynamics resulting from
Citation: more asynchronous irregular (AI) microscopic network activity (Tsodyks and Sejnowski, 1995;
Goldman JS, Tort-Colet N, di Volo M,
Van Vreeswijk and Sompolinsky, 1996; Brunel, 2000), thought to be important for neural coding
Susin E, Bouté J, Dali M, Carlu M,
Nghiem T-A, GórskiT and Destexhe A
(Skarda and Freeman, 1987; Van Vreeswijk and Sompolinsky, 1996; Tononi and Edelman, 1998;
(2019) Bridging Single Neuron Zerlaut and Destexhe, 2017).
Dynamics to Global Brain States. Much as different states of matter like solids, liquids, and gases emerge from interactions
Front. Syst. Neurosci. 13:75. between populations of molecules, different brain states may emerge from the interactions
doi: 10.3389/fnsys.2019.00075 between populations of neurons. Statistical physics provides a mathematical framework to

Frontiers in Systems Neuroscience | www.frontiersin.org 1 December 2019 | Volume 13 | Article 75

110
Goldman et al. Neural Codes; Order From Complexity

uncover structures of microscopic interactions underlying during unconscious states indicate a relative simplicity of
macroscopic properties. In this sense, macroscopically observed signals compared to conscious states. In unconscious states,
high synchrony, low complexity brain signals recorded from synchronous activity slowly sweeps across the cortex (Massimini
unconscious states may be accounted for by an increased et al., 2004) along paths formed by cortical tracts (Capone
coupling in the system’s components, behaving more like a et al., 2017). In both conscious resting and unconscious states,
solid (Peyrache et al., 2012; Le Van Quyen et al., 2016; Olcese the default mode network (Raichle et al., 2001; Boly et al.,
et al., 2016; Nghiem et al., 2018a). In contrast, conscious brain 2008) establishes a pattern of synchronization between brain
states may be described as higher complexity (Sitt et al., 2014; areas, producing correlations in ultra-slow (< 1 Hz) dynamics
Engemann et al., 2018; Nghiem et al., 2018a), perhaps liquid-like. (Brookes et al., 2011). Sustained, slow oscillations were initially
Though quantitative expressions directly linking order and reported in the thalamocortical system (Steriade, 2003), but
complexity are not straightforward, various definitions and are also observed experimentally in isolated cortex, without
metrics of complexity have been described to vary between thalamus (Sanchez-Vives and McCormick, 2000; Timofeev et al.,
brain states. Reports of enhanced complexity in conscious 2000). Thalamocortical connections shape slow wave dynamics
compared to unconscious states may be understood as increased (Destexhe et al., 2007; Poulet et al., 2012; David et al., 2013;
dimensionality (El Boustani and Destexhe, 2010), namely the Crunelli et al., 2015; Zucca et al., 2019) although slow oscillations
number of degrees of freedom needed to capture a system’s appear to be the default state of cortical networks (Sanchez-Vives
dynamics. Intuitively, dimensionality relates, though is not and McCormick, 2000; Sanchez-Vives et al., 2017).
reducible to, algorithmic complexity which quantifies the length Patterns of neocortical regions activated in resting state
of a deterministic algorithm required to reproduce an exact networks have been successfully retrieved using eigenmodes of
signal. For a random signal resulting from purely stochastic the structural connectivity matrix, i.e., the possible oscillatory
dynamics (similar to neural activity observed during conscious patterns at frequencies allowed by white matter tract lengths
states), the length of the algorithm would be as long as the (Atasoy et al., 2016). In active states, the executive control
signal itself. In contrast, a purely oscillatory signal (reminiscent network replaces the default mode (Fox et al., 2005), and the co-
of unconscious brain dynamics) can be recapitulated by a activation of different cortical regions is more strongly controlled
shorter algorithm, easily described by a periodic trajectory in by correlations in external stimuli than by white matter structural
few dimensions. connectivity (Gilson et al., 2018), with patterns of activity
Here, we aim to connect spatial scales from microscopic propagating recurrently between low-level, sensory areas and
(nanometers to micrometers—molecules to whole neurons) to high-level, associative areas.
macroscopic brain activity (centimeters to meters—brain areas During conscious states, on the background of globally
to individual subjects’ brains), describing both spontaneous and disorganized neural activity, transient patterns emerge (Duncan-
evoked dynamics. Toward linking interpretations of studies Johnson and Donchin, 1982; Goodin and Aminoff, 1984; Sur and
between scales, mesoscopic data (micrometers to millimeters— Sinha, 2009; Uhlhaas et al., 2009; Luck and Kappenman, 2011;
populations of thousands to tens of thousands of neurons) Churchland et al., 2012; Sato et al., 2012; Singer, 2013; Chemla
have been useful to inform models of neuronal assemblies. et al., 2019). Under an interpretation of brain states in analogy to
The perspective concludes by discussing a hypothesis best states of matter, microscopic changes in the interactions between
tested with a multi-scale understanding of brain function: the neurons could permit the emergence of larger-scale structures in
global complexity of neural activity increases in conscious brain activity.
brain states so as to enhance responsiveness to stimuli. We
suggest responsiveness may depend on the capacity of neural
networks to transiently collapse the dimensionality of collective MICROSCOPIC MECHANISMS;
dynamics—in particular neural assemblies sensitive to stimulus BIOPHYSICS OF BRAIN STATES
features—into evoked low-dimensional trajectories supporting
neural codes (Figure 1A). Experiments have demonstrated that during unconscious brain
states, the membrane potential (Vm ) of single cells slowly
oscillates between hyperpolarized and depolarized potentials
MACROSCOPIC SIGNALS VARY associated with alternating periods of silence (Down states,
ROBUSTLY BETWEEN BRAIN STATES also termed “OFF periods”) and AI-like firing (Up states, also
termed “ON periods”) (Steriade et al., 1993) (Figure 2A). During
Both spontaneous and evoked (Figures 1A,B) neural signals conscious brain states, neurons show sustained but sparse and
vary macroscopically across brain states, as demonstrated irregular AI firing patterns (Vreeswijk and Sompolinsky, 1998;
in electroencephalography (EEG), magnetoencephalography Destexhe et al., 1999; Brunel, 2000; Steriade, 2000; Renart et al.,
(MEG), and functional magnetic resonance imaging (fMRI). 2010; Dehghani et al., 2016; di Volo and Torcini, 2018). It
In unconscious states, neural activity is dominated by low- was found that, during AI states, excitatory (E) and inhibitory
frequency, high-amplitude signals (Niedermeyer and Lopes da (I) synaptic inputs are near-balanced (Dehghani et al., 2016),
Silva, 2005). Accordingly, analyses of entropy (Sitt et al., 2014; as predicted theoretically (Van Vreeswijk and Sompolinsky,
Engemann et al., 2018), complexity (Tononi and Edelman, 1996). In AI states, voltage fluctuations drive neurons over
1998), and dimensionality (El Boustani and Destexhe, 2010) the threshold for firing action potentials, resulting in irregular

Frontiers in Systems Neuroscience | www.frontiersin.org 2 December 2019 | Volume 13 | Article 75

111
Goldman et al. Neural Codes; Order From Complexity

FIGURE 1 | Complex dynamics associated with conscious brain states provide a potential substrate for neural coding. (A) Schematics of spontaneous (top) and
evoked (bottom) dynamics in connected neuronal assemblies encoding different related concepts (different colors) in unconscious (left) and conscious (right) brain
states. In unconscious brain states, slow, synchronous, large amplitude oscillations are observed. Stimuli delivered during unconscious states evoke large amplitude,
transient responses similar to spontaneous activity. In contrast, during conscious states, asynchronous, irregular firing of neurons results in macroscopically
desynchronized, low amplitude signals. Only networks recruited by the perturbation (here, a rabbit) produce lower-dimensional patterns that propagate relatively
further in time and space. (B) Global mean-field power (GMFP) recorded with EEG in response to transcranial magnetic stimulation, during deep, non-rapid eye
movement (NREM) sleep versus wakefulness. Mean EEG signal is represented by black traces. Background colors represent temporal latency (light blue, 0 ms; red,
300 ms) of maximum current sources, also shown in cortical space on the right, where yellow crosses represent the location of stimulation (right dorsolateral premotor
cortex). Reprinted with permission from AAAS (Massimini et al., 2005). If brain dynamics between states may be described in analogy to states of matter, perturbing
unconscious brains results in large, brief signals perhaps akin to a small perturbation of a solid, which can displace the solid briefly, but will not modify its internal
structure. In contrast, the same perturbation delivered during conscious, liquid-like brain states results in smaller but more complex patterns that propagate further in
time and space. Under this interpretation, in coding networks, responses evoked during conscious states could represent a form of transient “crystallization,”
consistent with neural trajectories lying on low-dimensional manifolds.

spiking dynamics, also known as fluctuation-driven regimes neurons (McCormick, 1992) and thus transitions between
(Kuhn et al., 2004; Destexhe, 2007; Destexhe and Rudolph-Lilith, AI and slow oscillatory dynamics through the regulation of
2012). To understand mechanisms at work during fluctuation- membrane currents (Hill and Tononi, 2005). Neuromodulators
driven dynamics, computational models have further shown that are present at higher concentrations during conscious states
three parameters are important to capture neuronal responses in (McCormick, 1992; Jones, 2003) and, most generally, inhibit
this regime, the average membrane voltage Vm , the amplitude potassium (activity-dependent and leak K + ) channels, which
of Vm fluctuations, and the conductance state of the membrane leads to depolarization of cells and suppression of spike-
(Reig et al., 2015; Zerlaut et al., 2016). frequency adaptation. At low neuromodulatory concentrations,
Neuromodulators, including acetylcholine, play important during unconscious states, K + leak channels are constitutively
biological roles in modulating the membrane properties of open and activity-dependent K + channels open when neurons

Frontiers in Systems Neuroscience | www.frontiersin.org 3 December 2019 | Volume 13 | Article 75

112
Goldman et al. Neural Codes; Order From Complexity

FIGURE 2 | Simple, high-amplitude signals in unconscious brain states are associated with synchronous regular neuronal firing, whereas complex, low-amplitude
signals in conscious brain states emerge from asynchronous irregular firing. (A) Data sample from Peyrache et al. (2012), Dehghani et al. (2016),
(Continued)

Frontiers in Systems Neuroscience | www.frontiersin.org 4 December 2019 | Volume 13 | Article 75

113
Goldman et al. Neural Codes; Order From Complexity

FIGURE 2 | Le Van Quyen et al. (2016), Teleńczuk et al. (2017), and Nghiem et al. (2018b), containing local field potential (LFP; top), spike times (action potentials;
middle), and spike counts (bottom) recorded from a human subject during NREM sleep (left) and wakefulness (right). Spikes from inhibitory (orange) and excitatory
(blue) neurons were separated and spike counts were calculated in bins of 5ms. Up states shaded in the left panel. (B) Transition between slow-wave (unconscious)
and activated (conscious) state dynamics in vivo (top) and in silico (bottom). Experimentally the transition is generated by electrical stimulation of acetylcholine neurons
in the pedunculopontine tegmentum (PPT) of anesthetized cat (Volgushev et al., 2011), triggering awake-like, desynchronized dynamics in cortex (Rudolph et al.,
2005). A prominent consequence of enhancing cholinergic signaling in cortex is a reduction of spike-frequency adaptation (McCormick, 1992). In silico, a similarly
desynchronizing effect can be generated by reducing the parameter responsible for spike-frequency adaptation. Simulated traces shown in the bottom were modified
from Destexhe (2009), which used a network of adaptive exponential integrate-and-fire neurons. The average Vm of the network, the Vm of a randomly chosen neuron,
and the raster plot of the network are shown. Reproduced with permission from Destexhe (2009). (C) State dependence of network responsiveness. The
responsiveness of two spiking networks to a Gaussian pulse is shown. Raster plots display spike times of excitatory (blue) and inhibitory (orange) neurons connected
by conductance-based synapses. Population activity (spike counts, thin line), as well as mean-field model (thick lines), and standard deviation (shaded area) of
population firing rate generated by a mean-field model developed in di Volo et al. (2019). Responsiveness is found to vary between different network states, obtained
by changing the ratio of the time-averaged global excitatory conductance (GE ) (Destexhe et al., 2003) to membrane leakage conductance (GL ).

spike, allowing K + ions to exit cells thus hyperpolarizing the graphoelements) are observed. Cash et al. have elegantly shown
membrane. Accumulating self-inhibition in the form of spike- that K-complexes (graphoelements characteristic of sleep stage
frequency adaptation during Up periods results in the transition 2) are complementarily observed both at microscopic and
to Down states. Conversely, spike-frequency adaptation wears macroscopic scales (Cash et al., 2009). Other identifiable patterns
off during Down states, allowing noise fluctuations (present also begin to emerge in empirical and theoretical data, including
ubiquitously; Destexhe and Rudolph-Lilith, 2012) to trigger phase cones (Freeman and Barrie, 2000) and interacting traveling
transitions to Up states (Destexhe, 2009; Jercog et al., 2017; waves (Sato et al., 2012; Chemla et al., 2019). Since statistical
Nghiem T.-A. E. et al., 2018; di Volo et al., 2019) (Figure 2B). physics has successfully described neuronal interactions for
Computationally speaking, for high values of spike-frequency different brain states, we ask next whether mesoscale methods
adaptation, bistability can be observed, with solutions at firing from statistical physics can help represent spontaneous and
rate zero (Down state) and non-zero (Up state) values (Holcman evoked dynamics of neuronal populations, thus formally linking
and Tsodyks, 2006; di Volo et al., 2019). The more chaotic knowledge between micro- and macroscopic scales.
dynamics of AI states associated with consciousness allows for
more reliable stimulus encoding (D’Andola et al., 2017), more
reliable propagation (Zerlaut and Destexhe, 2017), and more MESOSCALE BRIDGES; POPULATIONS OF
sustained responses (Nghiem T.-A. E. et al., 2018) to stimuli over NEURONS
time. In contrast, during unconscious states, neuronal responses
are more unreliable and vary greatly depending on the stimulus Brain dynamics at mesoscopic scales, describing thousands of
amplitude and whether cells receive inputs in Up or Down neurons, are investigated empirically by electrophysiology and
periods (Rosanova and Timofeev, 2005; Reig et al., 2015). more recently, voltage-sensitive dyes (Arieli et al., 1996; Chemla
The Ising model for spin glasses (Jaynes, 1982) fitted to neural and Chavane, 2010). At mesoscales, brain activity follows the
data (Schneidman et al., 2006) has revealed divergent types trend of increasing complexity of spontaneous activity with
of emergent neuronal dynamics in conscious and unconscious consciousness (Figure 2A). Studying the effects of inputs at
states. While neuronal interactions are pairwise in wakefulness the mesoscale, studies have shown that perturbations during
(Nghiem et al., 2017), coupling becomes population-wide in deep sleep states induce slow waves, but, during waking states,
deep sleep (Tavoni et al., 2017; Nghiem et al., 2018b). In perturbations can result in chains of phase-locked activity
particular, inhibitory neurons organize synchronous activity (Pigorini et al., 2015) leading to causal global interactions
across populations (Nghiem et al., 2018b; Zanoci et al., 2019), (Rosanova et al., 2018).
especially during deep sleep (Peyrache et al., 2012; Le Van Quyen Mean-field models offer a formalism for scaling up
et al., 2016; Olcese et al., 2016) where inhibitory neurons regulate microscopic detail to collective macroscopic dynamics using few
rhythms of slow wave dynamics (Compte et al., 2008; Funk et al., equations, offering a computational advantage for simulations.
2017; Zucca et al., 2017, 2019). In describing states of matter, mean-field models simplify
To summarize, between unconscious and conscious the probabilistic behavior of molecules to the relatively more
brain states, microscopic data appear intuitively related to predictable behavior of macroscopic states (Kadanoff, 2009).
macroscopic data: synchronous microscopic Up and Down A rich literature has begun to develop mean-field models of
states resulting from constitutive and activity-dependent, neuronal populations, showing that global variables describing
hyperpolarizing currents due to reduced neuromodulation population activity can be usefully derived from the biophysics
correspond to relatively simple, high-amplitude macroscopic of neurons and their interactions (Ohira and Cowan, 1993;
dynamics observed in unconscious states. Active, disorganized, Ginzburg and Sompolinsky, 1994; El Boustani and Destexhe,
desynchronized, AI, low adaptation, high neuromodulation 2009; Buice et al., 2010; Dahmen et al., 2016). Mean-field models
conditions correspond to low amplitude, complex, conscious have qualitatively reproduced temporal features of spontaneous
brain signals. On backgrounds of differing spontaneous dynamics including AI (El Boustani and Destexhe, 2009), Up
dynamics, generalizable patterns of activity (a.k.a. neural and Down dynamics (Compte et al., 2003; Jercog et al., 2017;

Frontiers in Systems Neuroscience | www.frontiersin.org 5 December 2019 | Volume 13 | Article 75

114
Goldman et al. Neural Codes; Order From Complexity

Tartaglia and Brunel, 2017; di Volo et al., 2019), and transitions networks display both excessively high conductance and
between these states (di Volo et al., 2019; Tort-Colet et al., strongly synchronized, regular dynamics (El Boustani
2019). In addition, connecting mean-fields provides a tool and Destexhe, 2010). Further, brain signals in coma are
for simulating the propagation of patterns through time and both low-amplitude and low-complexity, in contrast to
space, across mesoscale structures. For example, recent work high-amplitude signals observed in other unconscious
deriving mean-field models of networks with conductance-based states, but also to complex signals observed in conscious
synapses has reproduced the suppressive interaction between states (El Boustani and Destexhe, 2010). Such anomalous
traveling waves observed in visual cortex during conscious deviations from the overall trend of coordinated changes
states, a biological phenomenon that could not be captured by in complexity and amplitude may illuminate mechanisms
current-based networks (Chemla et al., 2019). underlying disease-causing deviations from healthy brain
Mean-field models have highlighted that, while complicated states (Mackey and Glass, 1977).
to apply mathematically in the framework of conductance- To characterize brain states, it has been useful to consider
based models (di Volo et al., 2019), voltage-dependent not only spontaneous dynamics but also patterns evoked by
interactions constitute a significant non-linearity in the perturbations. It was found that macroscopic responsiveness
membrane evolution equations. Voltage-dependent interactions highly depends on brain state and different patterns of responses
appear to be important for explaining non-trivial responses are evoked in conscious versus unconscious states (Massimini
of biological neurons, through the mean and fluctuations et al., 2005). Such state-dependent responsiveness can also be
of the cells’ membrane voltage (Reig et al., 2015). In seen at the level of local networks in vivo and in silico, for
fact, while these results do not imply that differences in example in the different reliability of responses to perturbations
responsiveness are due only to conductances, they show given during Up and Down periods of slow waves (Reig et al.,
that voltage dependent synapses play a role in the nonlinear 2015; Zerlaut and Destexhe, 2017). In simulations, different
state-dependent response of a neural network. As shown in responsiveness could be accounted for by three parameters:
Figure 2C, various levels of membrane conductance, regulated membrane voltage, voltage fluctuation amplitude, and membrane
by voltage-dependent synapses, are shown to differently shape conductance (Reig et al., 2015). These parameters could be well
population responses. described by mean-field models (di Volo et al., 2019), able
Finally, renormalization group theory, a method of coarse- to capture fundamental properties of spontaneous dynamics
graining microscopic detail to obtain macroscopic laws helping and also state-dependent responses at mesoscales. As such, the
to understand how order can emerge from apparent disorder data-driven coupling of such mean-field models may serve as
(Wilson, 1979; Cardy, 1996; Goldenfeld, 2018) has recently begun natural candidates for modeling the emergence of mesoscopic
to be applied to neural assemblies (Meshulam et al., 2019), and macroscopic-scale patterns.
laying further foundation for the formal connection of our Transient collapses of dimensionality found in encoding
understanding of brain function across scales. networks were also discussed as substrates potentially supporting
neural codes. Such collapses in complexity have been observed
in active ensembles at scales spanning microscopic (Churchland
DISCUSSION et al., 2010; Fairhall, 2019) to macroscopic (Quiroga et al., 2001;
Zang et al., 2004) activity. This echoes recent work studying
In this paper, we briefly reviewed work on the measurement recordings of neural populations which highlighted that neural
and modeling of brain states at different scales, from single representations of stimuli may lie on low-dimensional manifolds
neurons to cell assemblies and global brain activity, considering (Churchland et al., 2012; Sadtler et al., 2014; Gallego et al.,
both spontaneous and evoked dynamics. In particular, we 2017; Zhao and Park, 2017; Golub et al., 2018; Chaudhuri
highlighted that increased complexity in the dynamics of et al., 2019; Recanatesi et al., 2019; Stringer et al., 2019).
conscious brain states relates to changes in single-neuron Indeed neurons do not fire independently, which would yield
biophysics, tuned by neuromodulation. In unconscious states, dynamics of dimensionality as high as the number of neurons,
reduced neuromodulation promotes activity-dependent self- but instead follow constrained trajectories of activity that can
inhibition of excitatory neurons as they spike, leading to be captured by descriptions of much lower dimensionality that
alternating, synchronous transients of silence and firing, depend on spontaneous and evoked dynamics. For example, a
that produce high-amplitude, low-complexity, synchronous neural population firing in synchrony could be fully described
signals, on resonant frequencies of the structural connectome. by a periodic orbit trajectory constrained to a low-dimensional
During conscious states, neuronal discharges are asynchronous, space (Churchland et al., 2012). Since spontaneous global
irregular and fluctuation-driven, resulting from sustained network activity increases in dimensionality during conscious
membrane depolarization in cortical neurons, promoting states, we ask whether the transient collapse of complexity
effective neural communication. in specific networks, translating the emergence of simpler
Beyond conscious and unconscious categories proposed dynamical structures from disorder, may be associated to
here for the sake of brevity, important differences exist neural codes.
within categories of unconscious and conscious states (Brown As an analogy, windmills facing all in one direction
et al., 2010; El Boustani and Destexhe, 2010; Nghiem et al., display low complexity, but can only be synchronously
2018a). Unlike healthy wakefulness and sleep, epileptic active or inactive. Windmills facing in random directions,

Frontiers in Systems Neuroscience | www.frontiersin.org 6 December 2019 | Volume 13 | Article 75

115
Goldman et al. Neural Codes; Order From Complexity

in contrast, are a higher complexity configuration able to FUNDING

represent wind from any direction through the activity of
a subset. The activity of an ensemble of windmills tuned This research was supported by the Centre National de la
to a particular direction of wind could represent a collapse Recherche Scientifique (CNRS), the European Community
of complexity and the generation of information by that (Human Brain Project, H2020-785907), and by École des
subset (in this case, about the direction of wind). Similarly, Neurosciences de Paris (ENP).
enhanced dimensionality associated with conscious states
could subserve neural information through the collapse ACKNOWLEDGMENTS
of complexity in neural assemblies tuned to encode
particular representations. The authors would like to thank Simone Blanco Malerba,
Cristiano Capone, Stephen E. Clarke, Damien Depannemaecker,
AUTHOR CONTRIBUTIONS Anton Filipchuk, Enrique Hansen, J. N. K. Jaynes, Vicente Medel,
Archibald Parsons, Mattias Peuvrier, Wolf Singer, Kat Uesat,
All authors listed have made a substantial, direct and intellectual Bahar Hazal Yalçnkaya, and Yann Zerlaut for useful discussion
contribution to the work, and approved it for publication. of the manuscript.

REFERENCES Churchland, M. M., Cunningham, J. P., Kaufman, M. T., Foster, J. D., Nuyujukian,
P., Ryu, S. I., et al. (2012). Neural population dynamics during reaching. Nature
Arieli, A., Sterkin, A., Grinvald, A., and Aertsen, A. (1996). Dynamics of ongoing 487, 51–56. doi: 10.1038/nature11129
activity: explanation of the large variability in evoked cortical responses. Science Compte, A., Reig, R., Descalzo, V. F., Harvey, M. A., Puccini, G. D., and Sanchez-
273, 1868–1871. doi: 10.1126/science.273.5283.1868 Vives, M. V. (2008). Spontaneous high-frequency (10–80 hz) oscillations
Atasoy, S., Donnelly, I., and Pearson, J. (2016). Human brain networks during up states in the cerebral cortex in vitro. J. Neurosci. 28, 13828–13844.
function in connectome-specific harmonic waves. Nat. Commun. 7:10340. doi: 10.1523/JNEUROSCI.2684-08.2008
doi: 10.1038/ncomms10340 Compte, A., Sanchez-Vives, M. V., McCormick, D. A., and Wang, X.-J. (2003).
Boly, M., Phillips, C., Tshibanda, L., Vanhaudenhuyse, A., Schabus, M., Dang-Vu, Cellular and network mechanisms of slow oscillatory activity (< 1 HZ) and
T. T., et al. (2008). Intrinsic brain activity in altered states of consciousness: how wave propagations in a cortical network model. J. Neurophysiol. 89, 2707–2725.
conscious is the default mode of brain function? Ann. N. Y. Acad. Sci. 1129:119. doi: 10.1152/jn.00845.2002
doi: 10.1196/annals.1417.015 Crunelli, V., David, F., Lőrincz, M. L., and Hughes, S. W. (2015). The
Brookes, M. J., Woolrich, M., Luckhoo, H., Price, D., Hale, J. R., Stephenson, thalamocortical network as a single slow wave-generating unit. Curr. Opin.
M. C., et al. (2011). Investigating the electrophysiological basis of resting state Neurobiol. 31, 72–80. doi: 10.1016/j.conb.2014.09.001
networks using magnetoencephalography. Proc. Natl. Acad. Sci. U.S.A. 108, Dahmen, D., Bos, H., and Helias, M. (2016). Correlated fluctuations in
16783–16788. doi: 10.1073/pnas.1112685108 strongly coupled binary networks beyond equilibrium. Phys. Rev. X 6:031024.
Brown, E. N., Lydic, R., and Schiff, N. D. (2010). General anesthesia, sleep, and doi: 10.1103/PhysRevX.6.031024
coma. New Engl. J. Med. 363, 2638–2650. doi: 10.1056/NEJMra0808281 D’Andola, M., Rebollo, B., Casali, A. G., Weinert, J. F., Pigorini, A., Villa, R., et al.
Brunel, N. (2000). Dynamics of sparsely connected networks of excitatory (2017). Bistability, causality, and complexity in cortical networks: an in vitro
and inhibitory spiking neurons. J. Comput. Neurosci. 8, 183–208. perturbational study. Cereb. Cortex 28, 2233–2242. doi: 10.1093/cercor/bhx122
doi: 10.1023/A:1008925309027 David, F., Schmiedt, J. T., Taylor, H. L., Orban, G., Di Giovanni, G., Uebele, V. N.,
Buice, M. A., Cowan, J. D., and Chow, C. C. (2010). Systematic fluctuation et al. (2013). Essential thalamic contribution to slow waves of natural sleep. J.
expansion for neural network activity equations. Neural Comput. 22, 377–426. Neurosci. 33, 19599–19610. doi: 10.1523/JNEUROSCI.3169-13.2013
doi: 10.1162/neco.2009.02-09-960 Dehghani, N., Peyrache, A., Telenczuk, B., Le Van Quyen, M., Halgren, E., Cash,
Capone, C., Rebollo, B., Muñoz, A., Illa, X., Del Giudice, P., Sanchez-Vives, S. S., et al. (2016). Dynamic balance of excitation and inhibition in human and
M. V., et al. (2017). Slow waves in cortical slices: how spontaneous activity is monkey neocortex. Sci. Rep. 6:23176. doi: 10.1038/srep23176
shaped by laminar structure. Cereb. Cortex 29, 319–335. doi: 10.1093/cercor/ Destexhe, A. (2007). High-conductance state. Scholarpedia 2:1341.
bhx326 doi: 10.4249/scholarpedia.1341
Cardy, J. (1996). Scaling and Renormalization in Statistical Physics. Vol. 5. Destexhe, A. (2009). Self-sustained asynchronous irregular states and up–
Cambridge: Cambridge University Press. down states in thalamic, cortical and thalamocortical networks of
Cash, S. S., Halgren, E., Dehghani, N., Rossetti, A. O., Thesen, T., Wang, C., nonlinear integrate-and-fire neurons. J. Comput. Neurosci. 27, 493–506.
et al. (2009). The human k-complex represents an isolated cortical down-state. doi: 10.1007/s10827-009-0164-4
Science 324, 1084–1087. doi: 10.1126/science.1169626 Destexhe, A., Contreras, D., and Steriade, M. (1999). Spatiotemporal
Chaudhuri, R., Gerçek, B., Pandey, B., Peyrache, A., and Fiete, I. (2019). analysis of local field potentials and unit discharges in cat cerebral
The intrinsic attractor manifold and population dynamics of a canonical cortex during natural wake and sleep states. J. Neurosci. 19, 4595–4608.
cognitive circuit across waking and sleep. Nat. Neurosci. 22, 1512–1520. doi: 10.1523/JNEUROSCI.19-11-04595.1999
doi: 10.1038/s41593-019-0460-x Destexhe, A., Hughes, S. W., Rudolph, M., and Crunelli, V. (2007). Are
Chemla, S., and Chavane, F. (2010). Voltage-sensitive dye imaging: technique corticothalamic ‘up’states fragments of wakefulness? Trends Neurosci. 30, 334–
review and models. J. Physiol. 104, 40–50. doi: 10.1016/j.jphysparis.2009.11.009 342. doi: 10.1016/j.tins.2007.04.006
Chemla, S., Reynaud, A., di Volo, M., Zerlaut, Y., Perrinet, L., Destexhe, A., Destexhe, A., Rudolph, M., and Paré, D. (2003). The high-conductance
et al. (2019). Suppressive traveling waves shape representations of illusory state of neocortical neurons in vivo. Nat. Rev. Neurosci. 4, 739–751.
motion in primary visual cortex of awake primate. J. Neurosci. 39, 4282–4298. doi: 10.1038/nrn1198
doi: 10.1523/JNEUROSCI.2792-18.2019 Destexhe, A., and Rudolph-Lilith, M. (2012). Neuronal Noise. New York, NY:
Churchland, M. M., Byron, M. Y., Cunningham, J. P., Sugrue, L. P., Cohen, Springer.
M. R., Corrado, G. S., et al. (2010). Stimulus onset quenches neural di Volo, M., Romagnoni, A., Capone, C., and Destexhe, A. (2019). Biologically
variability: a widespread cortical phenomenon. Nat. Neurosci. 13, 369–378. realistic mean-field models of conductance-based networks of spiking neurons
doi: 10.1038/nn.2501 with adaptation. Neural Comput. 31, 653–680. doi: 10.1162/neco_a_01173

Frontiers in Systems Neuroscience | www.frontiersin.org 7 December 2019 | Volume 13 | Article 75

116
Goldman et al. Neural Codes; Order From Complexity

di Volo, M., and Torcini, A. (2018). Transition from asynchronous to oscillatory Mackey, M. C., and Glass, L. (1977). Oscillation and chaos in physiological control
dynamics in balanced spiking networks with instantaneous synapses. Phys. Rev. systems. Science 197, 287–289. doi: 10.1126/science.267326
Lett. 121:128301. doi: 10.1103/PhysRevLett.121.128301 Massimini, M., Ferrarelli, F., Huber, R., Esser, S. K., Singh, H., and Tononi, G.
Duncan-Johnson, C. C., and Donchin, E. (1982). The p300 component of the (2005). Breakdown of cortical effective connectivity during sleep. Science 309,
event-related brain potential as an index of information processing. Biol. 2228–2232. doi: 10.1126/science.1117256
Psychol. 14, 1–52. doi: 10.1016/0301-0511(82)90016-3 Massimini, M., Huber, R., Ferrarelli, F., Hill, S., and Tononi, G. (2004).
El Boustani, S., and Destexhe, A. (2009). A master equation formalism for The sleep slow oscillation as a traveling wave. J. Neurosci. 24, 6862–6870.
macroscopic modeling of asynchronous irregular activity states. Neural doi: 10.1523/JNEUROSCI.1318-04.2004
Comput. 21, 46–100. doi: 10.1162/neco.2009.02-08-710 McCormick, D. A. (1992). Neurotransmitter actions in the thalamus and cerebral
El Boustani, S., and Destexhe, A. (2010). Brain dynamics at multiple scales: can cortex and their role in neuromodulation of thalamocortical activity. Prog.
one reconcile the apparent low-dimensional chaos of macroscopic variables Neurobiol. 39, 337–388. doi: 10.1016/0301-0082(92)90012-4
with the seemingly stochastic behavior of single neurons? Int. J. Bifur. Chaos Meshulam, L., Gauthier, J. L., Brody, C. D., Tank, D. W., and Bialek, W. (2019).
20, 1687–1702. doi: 10.1142/S0218127410026769 Coarse graining, fixed points, and scaling in a large population of neurons. Phys.
Engemann, D. A., Raimondo, F., King, J.-R., Rohaut, B., Louppe, G., Faugeras, F., Rev. Lett. 123:178103. doi: 10.1103/PhysRevLett.123.178103
et al. (2018). Robust EEG-based cross-site and cross-protocol classification of Nghiem, T.-A., Lina, J.-M., di Volo, M., Capone, C., Evans, A. C., Destexhe, A.,
states of consciousness. Brain 141, 3179–3192. doi: 10.1093/brain/awy251 et al. (2018a). State equation from the spectral structure of human brain activity.
Fairhall, A. L. (2019). Whither variability? Nat. Neurosci. 22, 329–330. arXiv [preprint]. arXiv:1806.07365.
doi: 10.1038/s41593-019-0344-0 Nghiem, T.-A., Marre, O., Destexhe, A., and Ferrari, U. (2017). “Pairwise
Fox, M. D., Snyder, A. Z., Vincent, J. L., Corbetta, M., Van Essen, D. C., and ising model analysis of human cortical neuron recordings,” in International
Raichle, M. E. (2005). The human brain is intrinsically organized into dynamic, Conference on Geometric Science of Information (New York, NY: Springer),
anticorrelated functional networks. Proc. Natl. Acad. Sci. U.S.A. 102, 9673– 257–264. doi: 10.1007/978-3-319-68445-1_30
9678. doi: 10.1073/pnas.0504136102 Nghiem, T.-A., Telenczuk, B., Marre, O., Destexhe, A., and Ferrari, U.
Freeman, W. J., and Barrie, J. M. (2000). Analysis of spatial patterns of (2018b). Maximum-entropy models reveal the excitatory and inhibitory
phase in neocortical gamma EEGs in rabbit. J. Neurophysiol. 84, 1266–1278. correlation structures in cortical neuronal activity. Phys. Rev. E 98:012402.
doi: 10.1152/jn.2000.84.3.1266 doi: 10.1103/PhysRevE.98.012402
Funk, C. M., Peelman, K., Bellesi, M., Marshall, W., Cirelli, C., and Nghiem, T.-A. E., Tort-Colet, N., Gorski, T., Ferrari, U., Moghimyfiroozabad, S.,
Tononi, G. (2017). Role of somatostatin-positive cortical interneurons Goldman, J. S., et al. (2018). Cholinergic switch between two different types of
in the generation of sleep slow waves. J. Neurosci. 37, 9132–9148. slow waves in cerebral cortex. bioRxiv [Preprint]. doi: 10.1101/430405
doi: 10.1523/JNEUROSCI.1303-17.2017 Niedermeyer, E., and Lopes da Silva, F. H. (2005). Electroencephalography: Basic
Gallego, J. A., Perich, M. G., Miller, L. E., and Solla, S. A. (2017). Principles, Clinical Applications, and Related Fields. Philadelphia, PA: Lippincott
Neural manifolds for the control of movement. Neuron 94, 978–984. Williams & Wilkins.
doi: 10.1016/j.neuron.2017.05.025 Ohira, T., and Cowan, J. D. (1993). Master-equation approach to stochastic
Gilson, M., Deco, G., Friston, K. J., Hagmann, P., Mantini, D., Betti, V., et al. neurodynamics. Phys. Rev. E 48:2259. doi: 10.1103/physreve.48.2259
(2018). Effective connectivity inferred from fMRI transition dynamics during Olcese, U., Bos, J. J., Vinck, M., Lankelma, J. V., van Mourik-Donga, L. B.,
movie viewing points to a balanced reconfiguration of cortical interactions. Schlumm, F., et al. (2016). Spike-based functional connectivity in cerebral
Neuroimage 180, 534–546. doi: 10.1016/j.neuroimage.2017.09.061 cortex and hippocampus: loss of global connectivity is coupled to preservation
Ginzburg, I., and Sompolinsky, H. (1994). Theory of correlations in stochastic of local connectivity during non-REM sleep. J. Neurosci. 36, 7676–7692.
neural networks. Phys. Rev. E 50, 3171–3191. doi: 10.1103/PhysRevE.50.3171 doi: 10.1523/JNEUROSCI.4201-15.2016
Goldenfeld, N. (2018). Lectures on Phase Transitions and the Renormalization Peyrache, A., Dehghani, N., Eskandar, E. N., Madsen, J. R., Anderson, W. S.,
Group. Boca Raton, FL: CRC Press. Donoghue, J. A., et al. (2012). Spatiotemporal dynamics of neocortical
Golub, M. D., Sadtler, P. T., Oby, E. R., Quick, K. M., Ryu, S. I., Tyler-Kabara, excitation and inhibition during human sleep. Proc. Natl. Acad. Sci. U.S.A. 109,
E. C., et al. (2018). Learning by neural reassociation. Nat. Neurosci. 21, 607–616. 1731–1736. doi: 10.1073/pnas.1109895109
doi: 10.1038/s41593-018-0095-3 Pigorini, A., Sarasso, S., Proserpio, P., Szymanski, C., Arnulfo, G., Casarotto,
Goodin, D. S., and Aminoff, M. J. (1984). The relationship between the evoked S., et al. (2015). Bistability breaks-off deterministic responses to
potential and brain events in sensory discrimination and motor response. Brain intracortical stimulation during non-rem sleep. Neuroimage 112, 105–113.
107, 241–251. doi: 10.1093/brain/107.1.241 doi: 10.1016/j.neuroimage.2015.02.056
Hill, S., and Tononi, G. (2005). Modeling sleep and wakefulness Poulet, J. F., Fernandez, L. M., Crochet, S., and Petersen, C. C. (2012). Thalamic
in the thalamocortical system. J. Neurophysiol. 93, 1671–1698. control of cortical states. Nat. Neurosci. 15, 370–372. doi: 10.1038/nn.3035
doi: 10.1152/jn.00915.2004 Quiroga, R. Q., Rosso, O. A., Başar, E., and Schürmann, M. (2001). Wavelet entropy
Holcman, D., and Tsodyks, M. (2006). The emergence of up and down states in in event-related potentials: a new method shows ordering of EEG oscillations.
cortical networks. PLoS Comput. Biol. 2:e23. doi: 10.1371/journal.pcbi.0020023 Biol. Cybern. 84, 291–299. doi: 10.1007/s004220000212
Jaynes, E. T. (1982). On the rationale of maximum-entropy methods. Proc. IEEE Raichle, M. E., MacLeod, A. M., Snyder, A. Z., Powers, W. J., Gusnard, D. A., and
70, 939–952. doi: 10.1109/PROC.1982.12425 Shulman, G. L. (2001). A default mode of brain function. Proc. Natl. Acad. Sci.
Jercog, D., Roxin, A., Bartho, P., Luczak, A., Compte, A., and de la Rocha, J. (2017). U.S.A. 98, 676–682. doi: 10.1073/pnas.98.2.676
Up-down cortical dynamics reflect state transitions in a bistable network. Elife Recanatesi, S., Ocker, G. K., Buice, M. A., and Shea-Brown, E. (2019).
6:e22425. doi: 10.7554/eLife.22425 Dimensionality in recurrent spiking networks: global trends in activity
Jones, B. E. (2003). Arousal systems. Front. Biosci. 8:438–451. doi: 10.2741/1074 and local origins in connectivity. PLoS Comput. Biol. 15:e1006446.
Kadanoff, L. P. (2009). More is the same; phase transitions and mean field theories. doi: 10.1371/journal.pcbi.1006446
J. Stat. Phys. 137:777. doi: 10.1007/s10955-009-9814-1 Reig, R., Zerlaut, Y., Vergara, R., Destexhe, A., and Sanchez-Vives, M. V. (2015).
Kuhn, A., Aertsen, A., and Rotter, S. (2004). Neuronal integration of Gain modulation of synaptic inputs by network state in auditory cortex in vivo.
synaptic input in the fluctuation-driven regime. J. Neurosci. 24, 2345–2356. J. Neurosci. 35, 2689–2702. doi: 10.1523/JNEUROSCI.2004-14.2015
doi: 10.1523/JNEUROSCI.3349-03.2004 Renart, A., De La Rocha, J., Bartho, P., Hollender, L., Parga, N., Reyes, A.,
Le Van Quyen, M., Muller, L. E., Telenczuk, B., Halgren, E., Cash, S., Hatsopoulos, et al. (2010). The asynchronous state in cortical circuits. Science 327, 587–590.
N. G., et al. (2016). High-frequency oscillations in human and monkey doi: 10.1126/science.1179850
neocortex during the wake–sleep cycle. Proc. Natl. Acad. Sci. U.S.A. 113, Rosanova, M., Fecchio, M., Casarotto, S., Sarasso, S., Casali, A., Pigorini, A., et al.
9363–9368. doi: 10.1073/pnas.1523583113 (2018). Sleep-like cortical off-periods disrupt causality and complexity in the
Luck, S. J., and Kappenman, E. S. (2011). The Oxford Handbook of Event-Related brain of unresponsive wakefulness syndrome patients. Nat. Commun. 9:4427.
Potential Components. Oxford: Oxford University Press. doi: 10.1038/s41467-018-06871-1

Frontiers in Systems Neuroscience | www.frontiersin.org 8 December 2019 | Volume 13 | Article 75

117
Goldman et al. Neural Codes; Order From Complexity

Rosanova, M., and Timofeev, I. (2005). Neuronal mechanisms mediating the Tononi, G., and Edelman, G. M. (1998). Consciousness and complexity. Science
variability of somatosensory evoked potentials during sleep oscillations in cats. 282, 1846–1851. doi: 10.1126/science.282.5395.1846
J. Physiol. 562, 569–582. doi: 10.1113/jphysiol.2004.071381 Tort-Colet, N., Capone, C., Sanchez-Vives, M. V., and Mattia, M. (2019). Attractor
Rudolph, M., Pelletier, J.-G., Paré, D., and Destexhe, A. (2005). Characterization competition enriches cortical dynamics during awakening from anesthesia.
of synaptic conductances and integrative properties during electrically-induced bioRxiv [Preprint]. doi: 10.1101/517102
EEG-activated states in neocortical neurons in vivo. J. Neurophysiol. 94, 2805– Tsodyks, M. V., and Sejnowski, T. (1995). Rapid state switching in balanced cortical
2821. doi: 10.1152/jn.01313.2004 network models. Network 6, 111–124. doi: 10.1088/0954-898X_6_2_001
Sadtler, P. T., Quick, K. M., Golub, M. D., Chase, S. M., Ryu, S. I., Tyler-Kabara, Uhlhaas, P., Pipa, G., Lima, B., Melloni, L., Neuenschwander, S., Nikolić, D., et al.
E. C., et al. (2014). Neural constraints on learning. Nature 512, 423–426. (2009). Neural synchrony in cortical networks: history, concept and current
doi: 10.1038/nature13665 status. Front. Integr. Neurosci. 3:17. doi: 10.3389/neuro.07.017.2009
Sanchez-Vives, M. V., Massimini, M., and Mattia, M. (2017). Shaping the Van Vreeswijk, C., and Sompolinsky, H. (1996). Chaos in neuronal networks
default activity pattern of the cortical network. Neuron 94, 993–1001. with balanced excitatory and inhibitory activity. Science 274, 1724–1726.
doi: 10.1016/j.neuron.2017.05.015 doi: 10.1126/science.274.5293.1724
Sanchez-Vives, M. V., and McCormick, D. A. (2000). Cellular and network Volgushev, M., Chauvette, S., and Timofeev, I. (2011). “Long-range correlation
mechanisms of rhythmic recurrent activity in neocortex. Nat. Neurosci. 3, of the membrane potential in neocortical neurons during slow oscillation,” in
1027–1034. doi: 10.1038/79848 Progress in Brain Research, Vol. 193 (Amsterdam: Elsevier), 181–199.
Sato, T. K., Nauhaus, I., and Carandini, M. (2012). Traveling waves in visual cortex. Vreeswijk, C. V., and Sompolinsky, H. (1998). Chaotic balanced state
Neuron 75, 218–229. doi: 10.1016/j.neuron.2012.06.029 in a model of cortical circuits. Neural Comput. 10, 1321–1371.
Schneidman, E., Berry II, M. J., Segev, R., and Bialek, W. (2006). Weak pairwise doi: 10.1162/089976698300017214
correlations imply strongly correlated network states in a neural population. Wilson, K. G. (1979). Problems in physics with many scales of length. Sci. Am. 241,
Nature 440, 1007–1012. doi: 10.1038/nature04701 158–179. doi: 10.1038/scientificamerican0879-158
Singer, W. (2013). Cortical dynamics revisited. Trends Cogn. Sci. 17, 616–626. Zang, Y., Jiang, T., Lu, Y., He, Y., and Tian, L. (2004). Regional
doi: 10.1016/j.tics.2013.09.006 homogeneity approach to fMRI data analysis. Neuroimage 22, 394–400.
Sitt, J. D., King, J.-R., El Karoui, I., Rohaut, B., Faugeras, F., Gramfort, A., doi: 10.1016/j.neuroimage.2003.12.030
et al. (2014). Large scale screening of neural signatures of consciousness in Zanoci, C., Dehghani, N., and Tegmark, M. (2019). Ensemble inhibition and
patients in a vegetative or minimally conscious state. Brain 137, 2258–2270. excitation in the human cortex: an ising-model analysis with uncertainties.
doi: 10.1093/brain/awu141 Phys. Rev. E 99:032408. doi: 10.1103/PhysRevE.99.032408
Skarda, C. A., and Freeman, W. J. (1987). How brains make chaos in Zerlaut, Y., and Destexhe, A. (2017). Enhanced responsiveness and low-
order to make sense of the world. Behav. Brain Sci. 10, 161–173. level awareness in stochastic network states. Neuron 94, 1002–1009.
doi: 10.1017/S0140525X00047336 doi: 10.1016/j.neuron.2017.04.001
Steriade, M. (2000). Corticothalamic resonance, states of vigilance and mentation. Zerlaut, Y., Teleńczuk, B., Deleuze, C., Bal, T., Ouanounou, G., and Destexhe,
Neuroscience 101, 243–276. doi: 10.1016/S0306-4522(00)00353-5 A. (2016). Heterogeneous firing rate response of mouse layer V pyramidal
Steriade, M. (2003). Neuronal Substrates of Sleep and Epilepsy. Cambridge: neurons in the fluctuation-driven regime. J. Physiol. 594, 3791–3808.
Cambridge University Press. doi: 10.1113/JP272317
Steriade, M., Nunez, A., and Amzica, F. (1993). A novel slow (< 1 Hz) oscillation Zhao, Y., and Park, I. M. (2017). Variational latent gaussian process for recovering
of neocortical neurons in vivo: depolarizing and hyperpolarizing components. single-trial dynamics from population spike trains. Neural Comput. 29, 1293–
J. Neurosci. 13, 3252–3265. doi: 10.1523/JNEUROSCI.13-08-03252.1993 1316. doi: 10.1162/NECO_a_00953
Stringer, C., Pachitariu, M., Steinmetz, N., Carandini, M., and Harris, K. D. (2019). Zucca, S., D’Urso, G., Pasquale, V., Vecchia, D., Pica, G., Bovetti, S., et al.
High-dimensional geometry of population responses in visual cortex. Nature (2017). An inhibitory gate for state transition in cortex. Elife 6:e26177.
571, 361–365. doi: 10.1038/s41586-019-1346-5 doi: 10.7554/eLife.26177
Sur, S., and Sinha, V. (2009). Event-related potential: an overview. Ind. Psychiatry Zucca, S., Pasquale, V., de Leon Roig, P. L., Panzeri, S., and Fellin,
J. 18, 70–73. doi: 10.4103/0972-6748.57865 T. (2019). Thalamic drive of cortical parvalbumin-positive interneurons
Tartaglia, E. M., and Brunel, N. (2017). Bistability and up/down state alternations during down states in anesthetized mice. Curr. Biol. 29, 1481–1490.
in inhibition-dominated randomly connected networks of lif neurons. Sci. Rep. doi: 10.1016/j.cub.2019.04.007
7:11916. doi: 10.1038/s41598-017-12033-y
Tavoni, G., Ferrari, U., Battaglia, F. P., Cocco, S., and Monasson, R. (2017). Conflict of Interest: The authors declare that the research was conducted in the
Functional coupling networks inferred from prefrontal cortex activity absence of any commercial or financial relationships that could be construed as a
show experience-related effective plasticity. Netw. Neurosci. 1, 275–301. potential conflict of interest.
doi: 10.1162/NETN_a_00014
Teleńczuk, B., Dehghani, N., Le Van Quyen, M., Cash, S. S., Halgren, E., Copyright © 2019 Goldman, Tort-Colet, di Volo, Susin, Bouté, Dali, Carlu, Nghiem,
Hatsopoulos, N. G., et al. (2017). Local field potentials primarily reflect Górski and Destexhe. This is an open-access article distributed under the terms
inhibitory neuron activity in human and monkey cortex. Sci. Rep. 7:40211. of the Creative Commons Attribution License (CC BY). The use, distribution or
doi: 10.1038/srep40211 reproduction in other forums is permitted, provided the original author(s) and the
Timofeev, I., Grenier, F., Bazhenov, M., Sejnowski, T., and Steriade, M. (2000). copyright owner(s) are credited and that the original publication in this journal
Origin of slow cortical oscillations in deafferented cortical slabs. Cereb. Cortex is cited, in accordance with accepted academic practice. No use, distribution or
10, 1185–1199. doi: 10.1093/cercor/10.12.1185 reproduction is permitted which does not comply with these terms.

Frontiers in Systems Neuroscience | www.frontiersin.org 9 December 2019 | Volume 13 | Article 75

118
 Bibliography

[1] Andrea Volterra, Pierre Julius Magistretti, and Philip G Haydon. The tripartite
synapse: glia in synaptic transmission. BOOK. Oxford University Press, 2002.
[2] Eric R Kandel et al. Principles of neural science. Vol. 4. McGraw-hill New York,
2013.
[3] Roger D Traub et al. “A model of gamma-frequency network oscillations induced
in the rat CA3 region by carbachol in vitro”. In: European Journal of Neuroscience
12.11 (2000), pp. 4093–4106.
[4] Mario Galarreta and Shaul Hestrin. “Spike transmission and synchrony detection
in networks of GABAergic interneurons”. In: Science 292.5525 (2001), pp. 2295–
2299.
[5] Roger D Traub et al. “Gap junctions between interneuron dendrites can enhance
synchrony of gamma oscillations in distributed networks”. In: Journal of Neuro-
science 21.23 (2001), pp. 9478–9486.
[6] Sheriar G Hormuzdi et al. “Impaired electrical signaling disrupts gamma frequency
oscillations in connexin 36-deficient mice”. In: Neuron 31.3 (2001), pp. 487–495.
[7] Derek L Buhl et al. “Selective impairment of hippocampal gamma oscillations in
connexin-36 knock-out mouse in vivo”. In: Journal of Neuroscience 23.3 (2003),
pp. 1013–1018.
[8] Robert S Zucker. “Calcium-and activity-dependent synaptic plasticity”. In: Cur-
rent opinion in neurobiology 9.3 (1999), pp. 305–313.
[9] Donald S Faber and Henri Korn. “Single-shot channel activation accounts for dura-
tion of inhibitory postsynaptic potentials in a central neuron”. In: Science 208.4444
(1980), pp. 612–615.
[10] MICHAEL Perouansky and YOEL Yaari. “Kinetic properties of NMDA receptor-
mediated synaptic currents in rat hippocampal pyramidal cells versus interneu-
rones.” In: The Journal of Physiology 465.1 (1993), pp. 223–244.
[11] Thomas Götz et al. “Functional properties of AMPA and NMDA receptors ex-
pressed in identified types of basal ganglia neurons”. In: Journal of Neuroscience
17.1 (1997), pp. 204–215.

119
[12] Mark C Bellingham, Rebecca Lim, and Bruce Walmsley. “Developmental changes
in EPSC quantal size and quantal content at a central glutamatergic synapse in
rat”. In: The Journal of Physiology 511.3 (1998), pp. 861–869.
[13] Craig E Jahr and Charles F Stevens. “Voltage dependence of NMDA-activated
macroscopic conductances predicted by single-channel kinetics”. In: Journal of
Neuroscience 10.9 (1990), pp. 3178–3182.
[14] Ziv Gil and Yael Amitai. “Adult thalamocortical transmission involves both NMDA
and non-NMDA receptors”. In: Journal of neurophysiology 76.4 (1996), pp. 2547–
2554.
[15] Houman Homayoun and Bita Moghaddam. “NMDA receptor hypofunction pro-
duces opposite effects on prefrontal cortex interneurons and pyramidal neurons”.
In: Journal of Neuroscience 27.43 (2007), pp. 11496–11500.
[16] Alain Destexhe, Zachary F Mainen, and Terrence J Sejnowski. “Kinetic models of
synaptic transmission”. In: Methods in neuronal modeling 2 (1998), pp. 1–25.
[17] Yasuo Kawaguchi and Yoshiyuki Kubota. “GABAergic cell subtypes and their
synaptic connections in rat frontal cortex.” In: Cerebral cortex (New York, NY:
1991) 7.6 (1997), pp. 476–486.
[18] Carsten K Pfeffer et al. “Inhibition of inhibition in visual cortex: the logic of
connections between molecularly distinct interneurons”. In: Nature neuroscience
16.8 (2013), pp. 1068–1076.
[19] Yoshiyuki Kubota. “Untangling GABAergic wiring in the cortical microcircuit”.
In: Current opinion in neurobiology 26 (2014), pp. 7–14.
[20] Bruce P Bean. “The action potential in mammalian central neurons”. In: Nature
Reviews Neuroscience 8.6 (2007), pp. 451–465.
[21] David A McCormick et al. “Comparative electrophysiology of pyramidal and sparsely
spiny stellate neurons of the neocortex”. In: Journal of neurophysiology 54.4 (1985),
pp. 782–806.
[22] Barry W Connors and Michael J Gutnick. “Intrinsic firing patterns of diverse
neocortical neurons”. In: Trends in neurosciences 13.3 (1990), pp. 99–104.
[23] Petilla Interneuron Nomenclature Group (PING et al. “Petilla terminology: nomen-
clature of features of GABAergic interneurons of the cerebral cortex”. In: Nature
reviews. Neuroscience 9.7 (2008), p. 557.
[24] Javier DeFelipe et al. “New insights into the classification and nomenclature of
cortical GABAergic interneurons”. In: Nature Reviews Neuroscience 14.3 (2013),
pp. 202–216.

120
 BIBLIOGRAPHY

[25] Camille Allene, Joana Lourenço, and Alberto Bacci. “The neuronal identity bias
behind neocortical GABAergic plasticity”. In: Trends in neurosciences 38.9 (2015),
pp. 524–534.
[26] János Szabadics et al. “Excitatory effect of GABAergic axo-axonic cells in cortical
microcircuits”. In: Science 311.5758 (2006), pp. 233–235.
[27] Charles M Gray and David A McCormick. “Chattering cells: superficial pyramidal
neurons contributing to the generation of synchronous oscillations in the visual
cortex”. In: Science 274.5284 (1996), pp. 109–113.
[28] Mircea Steriade et al. “Dynamic properties of corticothalamic neurons and local
cortical interneurons generating fast rhythmic (30–40 Hz) spike bursts”. In: Journal
of Neurophysiology 79.1 (1998), pp. 483–490.
[29] Mircea Steriade. The intact and sliced brain. MIT press, 2001.
[30] Ariel Agmon and Barry W Connors. “Correlation between intrinsic firing patterns
and thalamocortical synaptic responses of neurons in mouse barrel cortex”. In:
Journal of Neuroscience 12.1 (1992), pp. 319–329.
[31] Gordon M Shepherd. The synaptic organization of the brain. Oxford university
press, 2004.
[32] Lennart Heimer. The human brain and spinal cord: functional neuroanatomy and
dissection guide. Springer Science & Business Media, 2012.
[33] Steven Laureys. “The neural correlate of (un) awareness: lessons from the vegeta-
tive state”. In: Trends in cognitive sciences 9.12 (2005), pp. 556–559.
[34] Florian Mormann and Christof Koch. “Neural correlates of consciousness”. In:
Scholarpedia 2.12 (2007), p. 1740.
[35] Cathy Jones. Glasgow coma scale. 1979.
[36] Olivier Le Bon. “Relationships between REM and NREM in the NREM-REM sleep
cycle: a review on competing concepts”. In: Sleep medicine 70 (2020), pp. 6–16.
[37] Conrad Iber. “The AASM manual for the scoring of sleep and associated events:
Rules”. In: Terminology and Technical Specification (2007).
[38] Alain Destexhe, Diego Contreras, and Mircea Steriade. “Spatiotemporal analysis
of local field potentials and unit discharges in cat cerebral cortex during natural
wake and sleep states”. In: Journal of Neuroscience 19.11 (1999), pp. 4595–4608.
[39] David H Hubel. “Single unit activity in striate cortex of unrestrained cats”. In:
The Journal of physiology 147.2 (1959), pp. 226–238.
[40] Edward V Evarts. “Temporal patterns of discharge of pyramidal tract neurons dur-
ing sleep and waking in the monkey”. In: Journal of neurophysiology 27.2 (1964),
pp. 152–171.

121
[41] M Steriade, M Deschênes, and G Oakson. “Inhibitory processes and interneuronal
apparatus in motor cortex during sleep and waking. I. Background firing and re-
sponsiveness of pyramidal tract neurons and interneurons.” In: Journal of neuro-
physiology 37.5 (1974), pp. 1065–1092.
[42] Rodolfo Llinas and Urs Ribary. “Coherent 40-Hz oscillation characterizes dream
state in humans.” In: Proceedings of the National Academy of Sciences 90.5 (1993),
pp. 2078–2081.
[43] H Berger. “On the use of the encephalogram in humans”. In: Arch Psychiatr Ner-
venkr 87 (1929), pp. 527–570.
[44] Marie Engelene J Obien et al. “Revealing neuronal function through microelectrode
array recordings”. In: Frontiers in neuroscience 8 (2015), p. 423.
[45] Sylvain Baillet. “Magnetoencephalography for brain electrophysiology and imag-
ing”. In: Nature neuroscience 20.3 (2017), pp. 327–339.
[46] Paul L Nunez and Ramesh Srinivasan. “Electroencephalogram”. In: Scholarpedia
2.2 (2007), p. 1348.
[47] Michel Le Van Quyen et al. “High-frequency oscillations in human and monkey
neocortex during the wake-sleep cycle”. In: Proceedings of the National Academy
of Sciences 113.33 (2016), pp. 9363–9368.
[48] Steffen Gais et al. “Early sleep triggers memory for early visual discrimination
skills”. In: Nature neuroscience 3.12 (2000), pp. 1335–1339.
[49] Robert Stickgold, LaTanya James, and J Allan Hobson. “Visual discrimination
learning requires sleep after training”. In: Nature neuroscience 3.12 (2000), pp. 1237–
1238.
[50] Pierre Maquet. “The role of sleep in learning and memory”. In: science 294.5544
(2001), pp. 1048–1052.
[51] Mircea Steriade and Igor Timofeev. “Neuronal plasticity in thalamocortical net-
works during sleep and waking oscillations”. In: Neuron 37.4 (2003), pp. 563–576.
[52] Reto Huber et al. “Local sleep and learning”. In: Nature 430.6995 (2004), pp. 78–
81.
[53] Mircea Steriade, David A McCormick, and Terrence J Sejnowski. “Thalamocor-
tical oscillations in the sleeping and aroused brain”. In: Science 262.5134 (1993),
pp. 679–685.
[54] Mircea Steriade, Angel Nunez, and Florin Amzica. “Intracellular analysis of rela-
tions between the slow (< 1 Hz) neocortical oscillation and other sleep rhythms
of the electroencephalogram”. In: Journal of Neuroscience 13.8 (1993), pp. 3266–
3283.

122
 BIBLIOGRAPHY

[55] I Timofeev and M Steriade. “Low-frequency rhythms in the thalamus of intact-

cortex and decorticated cats”. In: Journal of neurophysiology 76.6 (1996), pp. 4152–
4168.
[56] Maria V Sanchez-Vives and David A McCormick. “Cellular and network mecha-
nisms of rhythmic recurrent activity in neocortex”. In: Nature neuroscience 3.10
(2000), pp. 1027–1034.
[57] Maxim Bazhenov and Igor Timofeev. “Thalamocortical oscillations”. In: Scholar-
pedia 1.6 (2006), p. 1319.
[58] Jaekyung Kim, Tanuj Gulati, and Karunesh Ganguly. “Competing roles of slow
oscillations and delta waves in memory consolidation versus forgetting”. In: Cell
179.2 (2019), pp. 514–526.
[59] Caroline Geisler et al. “Hippocampal place cell assemblies are speed-controlled
oscillators”. In: Proceedings of the National Academy of Sciences 104.19 (2007),
pp. 8149–8154.
[60] Michael E Hasselmo. “What is the function of hippocampal theta rhythm?—Linking
behavioral data to phasic properties of field potential and unit recording data”. In:
Hippocampus 15.7 (2005), pp. 936–949.
[61] Joshua Jacobs et al. “Brain oscillations control timing of single-neuron activity in
humans”. In: Journal of Neuroscience 27.14 (2007), pp. 3839–3844.
[62] Han Lee et al. “Phase locking of single neuron activity to theta oscillations during
working memory in monkey extrastriate visual cortex”. In: Neuron 45.1 (2005),
pp. 147–156.
[63] Omkar N Markand. “Alpha rhythms”. In: Journal of Clinical Neurophysiology 7.2
(1990), pp. 163–190.
[64] Xiao-Jing Wang. “Neurophysiological and computational principles of cortical rhythms
in cognition”. In: Physiological reviews 90.3 (2010), pp. 1195–1268.
[65] György Buzsáki and Andreas Draguhn. “Neuronal oscillations in cortical net-
works”. In: science 304.5679 (2004), pp. 1926–1929.
[66] Charles M Gray and Wolf Singer. “Stimulus-specific neuronal oscillations in orien-
tation columns of cat visual cortex”. In: Proceedings of the National Academy of
Sciences 86.5 (1989), pp. 1698–1702.
[67] György Buzsáki, Cornelius H Vanderwolf, et al. “Cellular bases of hippocampal
EEG in the behaving rat”. In: Brain Research Reviews 6.2 (1983), pp. 139–171.
[68] Anatol Bragin et al. “Gamma (40-100 Hz) oscillation in the hippocampus of the
behaving rat”. In: Journal of Neuroscience 15.1 (1995), pp. 47–60.

123
[69] Jozsef Csicsvari et al. “Mechanisms of gamma oscillations in the hippocampus of
the behaving rat”. In: Neuron 37.2 (2003), pp. 311–322.
[70] Marlene Bartos, Imre Vida, and Peter Jonas. “Synaptic mechanisms of synchro-
nized gamma oscillations in inhibitory interneuron networks”. In: Nature reviews
neuroscience 8.1 (2007), pp. 45–56.
[71] Christof Brücke et al. “Thalamic gamma oscillations correlate with reaction time
in a Go/noGo task in patients with essential tremor”. In: Neuroimage 75 (2013),
pp. 36–45.
[72] Christine Pedroarena and Rodolfo Llinás. “Dendritic calcium conductances gener-
ate high-frequency oscillation in thalamocortical neurons”. In: Proceedings of the
National Academy of Sciences 94.2 (1997), pp. 724–728.
[73] D Pinault and M Deschenes. “Control of 40-Hz firing of reticular thalamic cells by
neurotransmitters”. In: Neuroscience 51.2 (1992), pp. 259–268.
[74] Bijan Pesaran et al. “Temporal structure in neuronal activity during working mem-
ory in macaque parietal cortex”. In: Nature neuroscience 5.8 (2002), pp. 805–811.
[75] Laura Lee Colgin et al. “Frequency of gamma oscillations routes flow of information
in the hippocampus”. In: Nature 462.7271 (2009), p. 353.
[76] Margaret F Carr, Mattias P Karlsson, and Loren M Frank. “Transient slow gamma
synchrony underlies hippocampal memory replay”. In: Neuron 75.4 (2012), pp. 700–
713.
[77] A Rougeul-Buser, JJ Bouyer, and P Buser. “From attentiveness to sleep. A topo-
graphical analysis of localized" synchronized" activities on the cortex of normal cat
and monkey”. In: Acta Neurobiol Exp (Warsz) 35.5-6 (1975), pp. 805–19.
[78] JJ Bouyer, MF Montaron, and A Rougeul. “Fast fronto-parietal rhythms during
combined focused attentive behaviour and immobility in cat: cortical and thalamic
localizations”. In: Electroencephalography and clinical neurophysiology 51.3 (1981),
pp. 244–252.
[79] Eugenio Rodriguez et al. “Perception’s shadow: long-distance synchronization of
human brain activity”. In: Nature 397.6718 (1999), pp. 430–433.
[80] Lucia Melloni et al. “Synchronization of neural activity across cortical areas corre-
lates with conscious perception”. In: Journal of neuroscience 27.11 (2007), pp. 2858–
2865.
[81] Pascal Fries et al. “Modulation of oscillatory neuronal synchronization by selective
visual attention”. In: Science 291.5508 (2001), pp. 1560–1563.

124
 BIBLIOGRAPHY

[82] Georgia G Gregoriou et al. “High-frequency, long-range coupling between pre-

frontal and visual cortex during attention”. In: science 324.5931 (2009), pp. 1207–
1210.
[83] Martin Vinck et al. “Attentional modulation of cell-class-specific gamma-band syn-
chronization in awake monkey area v4”. In: Neuron 80.4 (2013), pp. 1077–1089.
[84] Sateri Rouhinen et al. “Load dependence of β and γ oscillations predicts individual
capacity of visual attention”. In: Journal of Neuroscience 33.48 (2013), pp. 19023–
19033.
[85] Daniel E Sheer. “Focused arousal and the cognitive 40-Hz event-related potentials:
differential diagnosis of Alzheimer’s disease.” In: Progress in clinical and biological
research 317 (1989), pp. 79–94.
[86] Thilo Womelsdorf et al. “Gamma-band synchronization in visual cortex predicts
speed of change detection”. In: Nature 439.7077 (2006), pp. 733–736.
[87] Jozsef Csicsvari et al. “Oscillatory coupling of hippocampal pyramidal cells and
interneurons in the behaving rat”. In: Journal of Neuroscience 19.1 (1999), pp. 274–
287.
[88] Ryan T Canolty et al. “High gamma power is phase-locked to theta oscillations in
human neocortex”. In: science 313.5793 (2006), pp. 1626–1628.
[89] Supratim Ray et al. “Neural correlates of high-gamma oscillations (60–200 Hz) in
macaque local field potentials and their potential implications in electrocorticog-
raphy”. In: Journal of Neuroscience 28.45 (2008), pp. 11526–11536.
[90] Kevin Whittingstall and Nikos K Logothetis. “Frequency-band coupling in surface
EEG reflects spiking activity in monkey visual cortex”. In: Neuron 64.2 (2009),
pp. 281–289.
[91] Mariano A Belluscio et al. “Cross-frequency phase–phase coupling between theta
and gamma oscillations in the hippocampus”. In: Journal of Neuroscience 32.2
(2012), pp. 423–435.
[92] Supratim Ray and John HR Maunsell. “Different origins of gamma rhythm and
high-gamma activity in macaque visual cortex”. In: PLoS biology 9.4 (2011), e1000610.
[93] Xi Jiang, Jorge Gonzalez-Martinez, and Eric Halgren. “Coordination of human
hippocampal sharpwave ripples during NREM sleep with cortical theta bursts,
spindles, downstates, and upstates”. In: Journal of Neuroscience 39.44 (2019),
pp. 8744–8761.
[94] György Buzsáki. “Hippocampal sharp wave-ripple: A cognitive biomarker for episodic
memory and planning”. In: Hippocampus 25.10 (2015), pp. 1073–1188.

125
 [95] Dion Khodagholy, Jennifer N Gelinas, and György Buzsáki. “Learning-enhanced
coupling between ripple oscillations in association cortices and hippocampus”. In:
Science 358.6361 (2017), pp. 369–372.
[96] Alex P Vaz et al. “Coupled ripple oscillations between the medial temporal lobe
and neocortex retrieve human memory”. In: Science 363.6430 (2019), pp. 975–978.
[97] Nikos K Logothetis et al. “Hippocampal–cortical interaction during periods of sub-
cortical silence”. In: Nature 491.7425 (2012), pp. 547–553.
[98] Richard B Berry et al. “The AASM manual for the scoring of sleep and associ-
ated events”. In: Rules, Terminology and Technical Specifications, Darien, Illinois,
American Academy of Sleep Medicine 176 (2012), p. 2012.
[99] Oren M Weiner and Thien Thanh Dang-Vu. “Spindle oscillations in sleep disorders:
a systematic review”. In: Neural plasticity 2016 (2016).
[100] Thien Thanh Dang-Vu et al. “Interplay between spontaneous and induced brain
activity during human non-rapid eye movement sleep”. In: Proceedings of the Na-
tional Academy of Sciences 108.37 (2011), pp. 15438–15443.
[101] Hiroharu Noda and W Ross Adey. “Firing variability in cat association cortex
during sleep and wakefulness”. In: Brain research 18.3 (1970), pp. 513–526.
[102] Mircea Steriade, Igor Timofeev, and F Grenier. “Natural waking and sleep states: a
view from inside neocortical neurons”. In: Journal of neurophysiology 85.5 (2001),
pp. 1969–1985.
[103] Alexander S Ecker et al. “Decorrelated neuronal firing in cortical microcircuits”.
In: science 327.5965 (2010), pp. 584–587.
[104] Alfonso Renart et al. “The asynchronous state in cortical circuits”. In: science
327.5965 (2010), pp. 587–590.
[105] Carl Van Vreeswijk and Haim Sompolinsky. “Chaos in neuronal networks with
balanced excitatory and inhibitory activity”. In: Science 274.5293 (1996), pp. 1724–
1726.
[106] Yousheng Shu, Andrea Hasenstaub, and David A McCormick. “Turning on and off
recurrent balanced cortical activity”. In: Nature 423.6937 (2003), pp. 288–293.
[107] Bilal Haider et al. “Neocortical network activity in vivo is generated through a
dynamic balance of excitation and inhibition”. In: Journal of Neuroscience 26.17
(2006), pp. 4535–4545.
[108] Nima Dehghani et al. “Dynamic balance of excitation and inhibition in human and
monkey neocortex”. In: Scientific reports 6.1 (2016), pp. 1–12.

126
 BIBLIOGRAPHY

[109] Daniel J Amit and Nicolas Brunel. “Model of global spontaneous activity and local
structured activity during delay periods in the cerebral cortex.” In: Cerebral cortex
(New York, NY: 1991) 7.3 (1997), pp. 237–252.
[110] Nicolas Brunel. “Dynamics of sparsely connected networks of excitatory and in-
hibitory spiking neurons”. In: Journal of computational neuroscience 8.3 (2000),
pp. 183–208.
[111] Nicolas Brunel and Xiao-Jing Wang. “What determines the frequency of fast
network oscillations with irregular neural discharges? I. Synaptic dynamics and
excitation-inhibition balance”. In: Journal of neurophysiology 90.1 (2003), pp. 415–
430.
[112] Caroline Geisler, Nicolas Brunel, and Xiao-Jing Wang. “Contributions of intrin-
sic membrane dynamics to fast network oscillations with irregular neuronal dis-
charges”. In: Journal of neurophysiology 94.6 (2005), pp. 4344–4361.
[113] Charles F Stevens and Anthony M Zador. “Input synchrony and the irregular firing
of cortical neurons”. In: Nature neuroscience 1.3 (1998), pp. 210–217.
[114] David M Groppe et al. “Dominant frequencies of resting human brain activity as
measured by the electrocorticogram”. In: Neuroimage 79 (2013), pp. 223–233.
[115] Priyanka A Abhang, Bharti W Gawali, and Suresh C Mehrotra. Introduction to
EEG-and speech-based emotion recognition. Academic Press, 2016.
[116] Edgar D Adrian. “Olfactory reactions in the brain of the hedgehog”. In: The Jour-
nal of physiology 100.4 (1942), pp. 459–473.
[117] Canan Başar-Eroglu et al. “Gamma-band responses in the brain: a short review of
psychophysiological correlates and functional significance”. In: International jour-
nal of psychophysiology 24.1-2 (1996), pp. 101–112.
[118] Mircea Steriade, Florin Amzica, and Diego Contreras. “Synchronization of fast
(30-40 Hz) spontaneous cortical rhythms during brain activation”. In: Journal of
Neuroscience 16.1 (1996), pp. 392–417.
[119] Andreas Gabriel and Reinhard Eckhorn. “A multi-channel correlation method de-
tects traveling γ-waves in monkey visual cortex”. In: Journal of neuroscience meth-
ods 131.1-2 (2003), pp. 171–184.
[120] Robert Galambos. “A comparison of certain gamma band (40-Hz) brain rhythms
in cat and man”. In: Induced rhythms in the brain. Springer, 1992, pp. 201–216.
[121] Albert Compte et al. “Spontaneous high-frequency (10–80 Hz) oscillations during
up states in the cerebral cortex in vitro”. In: Journal of Neuroscience 28.51 (2008),
pp. 13828–13844.

127
[122] Nikos K Logothetis et al. “Neurophysiological investigation of the basis of the fMRI
signal”. In: Nature 412.6843 (2001), p. 150.
[123] Reinhard Eckhorn et al. “Coherent oscillations: A mechanism of feature linking in
the visual cortex?” In: Biological cybernetics 60.2 (1988), pp. 121–130.
[124] Le Bon-Jego, Rafael Yuste, et al. “Persistently active, pacemaker-like neurons in
neocortex”. In: Frontiers in neuroscience 1 (2007), p. 9.
[125] Wolf Singer and Charles M Gray. “Visual feature integration and the temporal
correlation hypothesis”. In: Annual review of neuroscience 18.1 (1995), pp. 555–
586.
[126] Wolf Singer. “Neuronal synchrony: a versatile code for the definition of relations?”
In: Neuron 24.1 (1999), pp. 49–65.
[127] John O’Keefe and Michael L Recce. “Phase relationship between hippocampal
place units and the EEG theta rhythm”. In: Hippocampus 3.3 (1993), pp. 317–330.
[128] Pascal Fries, Danko Nikolić, and Wolf Singer. “The gamma cycle”. In: Trends in
neurosciences 30.7 (2007), pp. 309–316.
[129] Pascal Fries. “A mechanism for cognitive dynamics: neuronal communication through
neuronal coherence”. In: Trends in cognitive sciences 9.10 (2005), pp. 474–480.
[130] Pascal Fries. “Rhythms for cognition: communication through coherence”. In: Neu-
ron 88.1 (2015), pp. 220–235.
[131] Gerald Hahn et al. “Communication through resonance in spiking neuronal net-
works”. In: PLoS Comput Biol 10.8 (2014), e1003811.
[132] György Buzsáki and Xiao-Jing Wang. “Mechanisms of gamma oscillations”. In:
Annual review of neuroscience 35 (2012), pp. 203–225.
[133] Mark O Cunningham et al. “A role for fast rhythmic bursting neurons in cortical
gamma oscillations in vitro”. In: Proceedings of the National Academy of Sciences
101.18 (2004), pp. 7152–7157.
[134] Charles M Gray. “Synchronous oscillations in neuronal systems: mechanisms and
functions”. In: Journal of computational neuroscience 1.1-2 (1994), pp. 11–38.
[135] Miles A Whittington et al. “Inhibition-based rhythms: experimental and math-
ematical observations on network dynamics”. In: International journal of psy-
chophysiology 38.3 (2000), pp. 315–336.
[136] Andre Fisahn et al. “Cholinergic induction of network oscillations at 40 Hz in the
hippocampus in vitro”. In: Nature 394.6689 (1998), pp. 186–189.
[137] André Fisahn et al. “Distinct roles for the kainate receptor subunits GluR5 and
GluR6 in kainate-induced hippocampal gamma oscillations”. In: Journal of Neu-
roscience 24.43 (2004), pp. 9658–9668.

128
 BIBLIOGRAPHY

[138] Roger D Traub et al. Fast oscillations in cortical circuits. MIT press, 1999.
[139] Roger D Traub. “Fast oscillations”. In: Scholarpedia 1.12 (2006), p. 1764.
[140] Paul Tiesinga and Terrence J Sejnowski. “Cortical enlightenment: are attentional
gamma oscillations driven by ING or PING?” In: Neuron 63.6 (2009), pp. 727–732.
[141] David A Lewis. “GABAergic local circuit neurons and prefrontal cortical dysfunc-
tion in schizophrenia”. In: Brain Research Reviews 31.2-3 (2000), pp. 270–276.
[142] David A Lewis, Takanori Hashimoto, and David W Volk. “Cortical inhibitory
neurons and schizophrenia”. In: Nature Reviews Neuroscience 6.4 (2005), pp. 312–
324.
[143] Vasilis P Bozikas and Christina Andreou. “Longitudinal studies of cognition in
first episode psychosis: a systematic review of the literature”. In: Australian &
New Zealand Journal of Psychiatry 45.2 (2011), pp. 93–108.
[144] Tonghui Su et al. “How could N-methyl-D-aspartate receptor antagonists lead to
excitation instead of inhibition?” In: Brain Science Advances 4.2 (2018), pp. 73–98.
[145] Martha E Shenton et al. “A review of MRI findings in schizophrenia”. In: Schizophre-
nia research 49.1-2 (2001), pp. 1–52.
[146] Lynn D Selemon and Patricia S Goldman-Rakic. “The reduced neuropil hypothe-
sis: a circuit based model of schizophrenia”. In: Biological psychiatry 45.1 (1999),
pp. 17–25.
[147] Daniel C Javitt and Stephen R Zukin. “Recent advances in the phencyclidine model
of schizophrenia.” In: The American journal of psychiatry (1991).
[148] John H Krystal et al. “Subanesthetic effects of the noncompetitive NMDA antag-
onist, ketamine, in humans: psychotomimetic, perceptual, cognitive, and neuroen-
docrine responses”. In: Archives of general psychiatry 51.3 (1994), pp. 199–214.
[149] Sarbjeet S Kalsi, David M Wood, and Paul I Dargan. “The epidemiology and
patterns of acute and chronic toxicity associated with recreational ketamine use”.
In: Emerging Health Threats Journal 4.1 (2011), p. 7107.
[150] John W Newcomer et al. “Ketamine-induced NMDA receptor hypofunction as a
model of memory impairment and psychosis”. In: Neuropsychopharmacology 20.2
(1999), pp. 106–118.
[151] Joseph T Coyle. “Glutamate and schizophrenia: beyond the dopamine hypothesis”.
In: Cellular and molecular neurobiology 26.4 (2006), pp. 363–382.
[152] Craig W Lindsley et al. “Progress towards validating the NMDA receptor hypo-
function hypothesis of schizophrenia”. In: Current topics in medicinal chemistry
6.8 (2006), pp. 771–785.

129
[153] Herbert Y Meltzer. “Dopamine receptors and average clinical doses”. In: Science
194.4264 (1976), pp. 545–546.
[154] Solomon H Snyder. “Dopamine receptors, neuroleptics, and schizophrenia.” In: The
American journal of psychiatry (1981).
[155] Stephanie M Perez and Daniel J Lodge. “New approaches to the management of
schizophrenia: focus on aberrant hippocampal drive of dopamine pathways”. In:
Drug design, development and therapy 8 (2014), p. 887.
[156] B Pakkenberg, J Scheel-Krüger, and LV Kristiansen. “Schizophrenia; from struc-
ture to function with special focus on the mediodorsal thalamic prefrontal loop”.
In: Acta Psychiatrica Scandinavica 120.5 (2009), pp. 345–354.
[157] Jefferson W Kinney et al. “A specific role for NR2A-containing NMDA receptors
in the maintenance of parvalbumin and GAD67 immunoreactivity in cultured in-
terneurons”. In: Journal of Neuroscience 26.5 (2006), pp. 1604–1615.
[158] Lavinia Albéri et al. “The calcium-binding protein parvalbumin modulates the
firing 1 properties of the reticular thalamic nucleus bursting neurons”. In: Journal
of neurophysiology 109.11 (2013), pp. 2827–2841.
[159] Olivier Caillard et al. “Role of the calcium-binding protein parvalbumin in short-
term synaptic plasticity”. In: Proceedings of the National Academy of Sciences
97.24 (2000), pp. 13372–13377.
[160] M Margarita Behrens et al. “Ketamine-induced loss of phenotype of fast-spiking in-
terneurons is mediated by NADPH-oxidase”. In: Science 318.5856 (2007), pp. 1645–
1647.
[161] Susan M Cochran et al. “Induction of metabolic hypofunction and neurochemical
deficits after chronic intermittent exposure to phencyclidine: differential modula-
tion by antipsychotic drugs”. In: Neuropsychopharmacology 28.2 (2003), pp. 265–
275.
[162] Yuchun Zhang, M Margarita Behrens, and John E Lisman. “Prolonged exposure
to NMDAR antagonist suppresses inhibitory synaptic transmission in prefrontal
cortex”. In: Journal of neurophysiology 100.2 (2008), pp. 959–965.
[163] Schahram Akbarian et al. “Gene expression for glutamic acid decarboxylase is
reduced without loss of neurons in prefrontal cortex of schizophrenics”. In: Archives
of general psychiatry 52.4 (1995), pp. 258–266.
[164] David W Volk et al. “Decreased glutamic acid decarboxylase67 messenger RNA
expression in a subset of prefrontal cortical γ-aminobutyric acid neurons in subjects
with schizophrenia”. In: Archives of general psychiatry 57.3 (2000), pp. 237–245.

130
 BIBLIOGRAPHY

[165] Schahram Akbarian and Hsien-Sung Huang. “Molecular and cellular mechanisms
of altered GAD1/GAD67 expression in schizophrenia and related disorders”. In:
Brain research reviews 52.2 (2006), pp. 293–304.
[166] Melis Inan, Timothy J Petros, and Stewart A Anderson. “Losing your inhibition:
linking cortical GABAergic interneurons to schizophrenia”. In: Neurobiology of
disease 53 (2013), pp. 36–48.
[167] DW Eyles, JJ McGrath, and GP Reynolds. “Neuronal calcium-binding proteins
and schizophrenia”. In: Schizophrenia research 57.1 (2002), pp. 27–34.
[168] Takanori Hashimoto et al. “Gene expression deficits in a subclass of GABA neurons
in the prefrontal cortex of subjects with schizophrenia”. In: Journal of Neuroscience
23.15 (2003), pp. 6315–6326.
[169] Juan E Belforte et al. “Postnatal NMDA receptor ablation in corticolimbic in-
terneurons confers schizophrenia-like phenotypes”. In: Nature neuroscience 13.1
(2010), pp. 76–83.
[170] Marie Carlen et al. “A critical role for NMDA receptors in parvalbumin interneu-
rons for gamma rhythm induction and behavior”. In: Molecular psychiatry 17.5
(2012), pp. 537–548.
[171] Eddie N Billingslea et al. “Parvalbumin cell ablation of NMDA-R1 causes increased
resting network excitability with associated social and self-care deficits”. In: Neu-
ropsychopharmacology 39.7 (2014), pp. 1603–1613.
[172] Kazuhito Nakao and Kazu Nakazawa. “Brain state-dependent abnormal LFP ac-
tivity in the auditory cortex of a schizophrenia mouse model”. In: Frontiers in
neuroscience 8 (2014), p. 168.
[173] Tatiana Korotkova et al. “NMDA receptor ablation on parvalbumin-positive in-
terneurons impairs hippocampal synchrony, spatial representations, and working
memory”. In: Neuron 68.3 (2010), pp. 557–569.
[174] Diana C Rotaru et al. “Glutamate receptor subtypes mediating synaptic activation
of prefrontal cortex neurons: relevance for schizophrenia”. In: Journal of Neuro-
science 31.1 (2011), pp. 142–156.
[175] Guillermo Gonzalez-Burgos and David A Lewis. “NMDA receptor hypofunction,
parvalbumin-positive neurons, and cortical gamma oscillations in schizophrenia”.
In: Schizophrenia bulletin 38.5 (2012), pp. 950–957.
[176] Madhuri S Kurdi, Kaushic A Theerth, and Radhika S Deva. “Ketamine: current
applications in anesthesia, pain, and critical care”. In: Anesthesia, essays and re-
searches 8.3 (2014), p. 283.

131
[177] Bita Moghaddam et al. “Activation of glutamatergic neurotransmission by ke-
tamine: a novel step in the pathway from NMDA receptor blockade to dopamin-
ergic and cognitive disruptions associated with the prefrontal cortex”. In: Journal
of Neuroscience 17.8 (1997), pp. 2921–2927.
[178] Adrienne C Lahti et al. “Ketamine activates psychosis and alters limbic blood flow
in schizophrenia.” In: Neuroreport 6.6 (1995), pp. 869–872.
[179] Alan Breier et al. “Association of ketamine-induced psychosis with focal activa-
tion of the prefrontal cortex in healthy volunteers.” In: The American journal of
psychiatry (1997).
[180] FX Vollenweider et al. “Differential psychopathology and patterns of cerebral glu-
cose utilisation produced by (S)-and (R)-ketamine in healthy volunteers using
positron emission tomography (PET)”. In: European Neuropsychopharmacology
7.1 (1997), pp. 25–38.
[181] Y Suzuki et al. “Acute administration of phencyclidine induces tonic activation
of medial prefrontal cortex neurons in freely moving rats”. In: Neuroscience 114.3
(2002), pp. 769–779.
[182] Mark E Jackson, Houman Homayoun, and Bita Moghaddam. “NMDA receptor hy-
pofunction produces concomitant firing rate potentiation and burst activity reduc-
tion in the prefrontal cortex”. In: Proceedings of the National Academy of Sciences
101.22 (2004), pp. 8467–8472.
[183] Allie J Widman and Lori L McMahon. “Disinhibition of CA1 pyramidal cells by
low-dose ketamine and other antagonists with rapid antidepressant efficacy”. In:
Proceedings of the National Academy of Sciences 115.13 (2018), E3007–E3016.
[184] V Di Lazzaro et al. “Ketamine increases human motor cortex excitability to tran-
scranial magnetic stimulation”. In: The Journal of physiology 547.2 (2003), pp. 485–
496.
[185] Alexander D Shaw et al. “Ketamine amplifies induced gamma frequency oscilla-
tions in the human cerebral cortex”. In: European Neuropsychopharmacology 25.8
(2015), pp. 1136–1146.
[186] Gilles Plourde, Jacinthe Baribeau, and Vincent Bonhomme. “Ketamine increases
the amplitude of the 40-Hz auditory steady-state response in humans.” In: British
journal of anaesthesia 78.5 (1997), pp. 524–529.
[187] L Elliot Hong et al. “Gamma and delta neural oscillations and association with
clinical symptoms under subanesthetic ketamine”. In: Neuropsychopharmacology
35.3 (2010), pp. 632–640.

132
 BIBLIOGRAPHY

[188] Didier Pinault. “N-methyl d-aspartate receptor antagonists ketamine and MK-801
induce wake-related aberrant γ oscillations in the rat neocortex”. In: Biological
psychiatry 63.8 (2008), pp. 730–735.
[189] Bernat Kocsis. “Differential role of NR2A and NR2B subunits in N-methyl-D-
aspartate receptor antagonist-induced aberrant cortical gamma oscillations”. In:
Biological psychiatry 71.11 (2012), pp. 987–995.
[190] Jesse Wood, Yunbok Kim, and Bita Moghaddam. “Disruption of prefrontal cortex
large scale neuronal activity by different classes of psychotomimetic drugs”. In:
Journal of Neuroscience 32.9 (2012), pp. 3022–3031.
[191] María Jesús Nicolás et al. “Ketamine-induced oscillations in the motor circuit of
the rat basal ganglia”. In: PloS one 6.7 (2011), e21814.
[192] Maya Slovik et al. “Ketamine induced converged synchronous gamma oscillations
in the cortico-basal ganglia network of nonhuman primates”. In: Journal of neu-
rophysiology 118.2 (2017), pp. 917–931.
[193] Klaus Linkenkaer-Hansen et al. “Genetic contributions to long-range temporal
correlations in ongoing oscillations”. In: Journal of Neuroscience 27.50 (2007),
pp. 13882–13889.
[194] L Elliot Hong et al. “Sensory gating endophenotype based on its neural oscillatory
pattern and heritability estimate”. In: Archives of general psychiatry 65.9 (2008),
pp. 1008–1016.
[195] L Elliot Hong et al. “Evoked gamma band synchronization and the liability for
schizophrenia”. In: Schizophrenia research 70.2-3 (2004), pp. 293–302.
[196] Mei-Hua Hall et al. “The early auditory gamma-band response is heritable and a
putative endophenotype of schizophrenia”. In: Schizophrenia bulletin 37.4 (2011),
pp. 778–787.
[197] Gregor Leicht et al. “Alterations of the early auditory evoked gamma-band re-
sponse in first-degree relatives of patients with schizophrenia: hints to a new in-
termediate phenotype”. In: Journal of psychiatric research 45.5 (2011), pp. 699–
705.
[198] Peter J Uhlhaas and Wolf Singer. “Abnormal neural oscillations and synchrony in
schizophrenia”. In: Nature reviews neuroscience 11.2 (2010), pp. 100–113.
[199] Peter J Uhlhaas and Wolf Singer. “High-frequency oscillations and the neurobiology
of schizophrenia”. In: Dialogues in clinical neuroscience 15.3 (2013), p. 301.
[200] Johanna Kissler et al. “MEG gamma band activity in schizophrenia patients and
healthy subjects in a mental arithmetic task and at rest”. In: Clinical Neurophys-
iology 111.11 (2000), pp. 2079–2087.

133
[201] Jun Soo Kwon et al. “Gamma frequency–range abnormalities to auditory stimula-
tion in schizophrenia”. In: Archives of general psychiatry 56.11 (1999), pp. 1001–
1005.
[202] Colleen A Brenner et al. “EEG synchronization to modulated auditory tones in
schizophrenia, schizoaffective disorder, and schizotypal personality disorder”. In:
American Journal of Psychiatry 160.12 (2003), pp. 2238–2240.
[203] Gregory A Light et al. “Gamma band oscillations reveal neural network cortical
coherence dysfunction in schizophrenia patients”. In: Biological psychiatry 60.11
(2006), pp. 1231–1240.
[204] Dorea Vierling-Claassen et al. “Modeling GABA alterations in schizophrenia: a
link between impaired inhibition and altered gamma and beta range auditory en-
trainment”. In: Journal of neurophysiology 99.5 (2008), pp. 2656–2671.
[205] Kevin M Spencer et al. “γ-Band auditory steady-state responses are impaired in
first episode psychosis”. In: Biological psychiatry 64.5 (2008), pp. 369–375.
[206] Tony W Wilson et al. “Cortical gamma generators suggest abnormal auditory
circuitry in early-onset psychosis”. In: Cerebral cortex 18.2 (2008), pp. 371–378.
[207] Giri P Krishnan et al. “Steady state and induced auditory gamma deficits in
schizophrenia”. In: Neuroimage 47.4 (2009), pp. 1711–1719.
[208] Kwang-Hyuk Lee et al. “An integration of 40 Hz Gamma and phasic arousal:
novelty and routinization processing in schizophrenia”. In: Clinical neurophysiology
112.8 (2001), pp. 1499–1507.
[209] Kevin M Spencer et al. “Neural synchrony indexes disordered perception and cogni-
tion in schizophrenia”. In: Proceedings of the National Academy of Sciences 101.49
(2004), pp. 17288–17293.
[210] Kevin M Spencer et al. “Sensory-evoked gamma oscillations in chronic schizophre-
nia”. In: Biological psychiatry 63.8 (2008), pp. 744–747.
[211] Jürgen Gallinat et al. “Reduced oscillatory gamma-band responses in unmedicated
schizophrenic patients indicate impaired frontal network processing”. In: Clinical
Neurophysiology 115.8 (2004), pp. 1863–1874.
[212] Kevin M Spencer. “Baseline gamma power during auditory steady-state stimula-
tion in schizophrenia”. In: Frontiers in human neuroscience 5 (2012), p. 190.
[213] Noah C Venables, Edward M Bernat, and Scott R Sponheim. “Genetic and disorder-
specific aspects of resting state EEG abnormalities in schizophrenia”. In: Schizophre-
nia bulletin 35.4 (2009), pp. 826–839.
[214] Yoji Hirano et al. “Spontaneous gamma activity in schizophrenia”. In: JAMA psy-
chiatry 72.8 (2015), pp. 813–821.

134
 BIBLIOGRAPHY

[215] Georg Winterer et al. “Prefrontal broadband noise, working memory, and genetic
risk for schizophrenia”. In: American Journal of Psychiatry 161.3 (2004), pp. 490–
500.
[216] Canan Basar-Eroglu et al. “Working memory related gamma oscillations in schizophre-
nia patients”. In: International journal of psychophysiology 64.1 (2007), pp. 39–45.
[217] Gary Flynn et al. “Increased absolute magnitude of gamma synchrony in first-
episode psychosis”. In: Schizophrenia research 105.1-3 (2008), pp. 262–271.
[218] Corinna Haenschel et al. “Cortical oscillatory activity is critical for working mem-
ory as revealed by deficits in early-onset schizophrenia”. In: Journal of Neuro-
science 29.30 (2009), pp. 9481–9489.
[219] RY Cho, RO Konecky, and Cameron S Carter. “Impairments in frontal cortical γ
synchrony and cognitive control in schizophrenia”. In: Proceedings of the National
Academy of Sciences 103.52 (2006), pp. 19878–19883.
[220] Lindsay Rutter et al. “Magnetoencephalographic gamma power reduction in pa-
tients with schizophrenia during resting condition”. In: Human brain mapping
30.10 (2009), pp. 3254–3264.
[221] Balaji Narayanan et al. “Resting state electroencephalogram oscillatory abnormal-
ities in schizophrenia and psychotic bipolar patients and their relatives from the
bipolar and schizophrenia network on intermediate phenotypes study”. In: Biolog-
ical psychiatry 76.6 (2014), pp. 456–465.
[222] Erol Başar. “A review of gamma oscillations in healthy subjects and in cognitive
impairment”. In: International Journal of Psychophysiology 90.2 (2013), pp. 99–
117.
[223] Andrea Perrottelli et al. “EEG-Based Measures in At-Risk Mental State and Early
Stages of Schizophrenia: A Systematic Review”. In: Frontiers in psychiatry 12
(2021), p. 582.
[224] Matthew B Symond et al. ““Gamma synchrony” in first-episode schizophrenia:
a disorder of temporal connectivity?” In: American Journal of Psychiatry 162.3
(2005), pp. 459–465.
[225] C Mulert et al. “Long-range synchrony of gamma oscillations and auditory halluci-
nation symptoms in schizophrenia”. In: International Journal of Psychophysiology
79.1 (2011), pp. 55–63.
[226] Kevin M Spencer et al. “Left auditory cortex gamma synchronization and audi-
tory hallucination symptoms in schizophrenia”. In: BMC neuroscience 10.1 (2009),
pp. 1–13.

135
[227] Tineke Grent et al. “Resting-state gamma-band power alterations in schizophrenia
reveal E/I-balance abnormalities across illness-stages”. In: Elife 7 (2018), e37799.
[228] Maciej T Lazarewicz et al. “Ketamine modulates theta and gamma oscillations”.
In: Journal of cognitive neuroscience 22.7 (2010), pp. 1452–1464.
[229] Alan Anticevic et al. “N-methyl-D-aspartate receptor antagonist effects on pre-
frontal cortical connectivity better model early than chronic schizophrenia”. In:
Biological psychiatry 77.6 (2015), pp. 569–580.
[230] Zheng Wang et al. “Comparison of first-episode and chronic patients diagnosed
with schizophrenia: Symptoms and childhood trauma”. In: Early intervention in
psychiatry 7.1 (2013), pp. 23–30.
[231] Nicolas Brunel and Mark CW Van Rossum. “Lapicque’s 1907 paper: from frogs to
integrate-and-fire”. In: Biological cybernetics 97.5-6 (2007), pp. 337–339.
[232] Larry F Abbott. “Lapicque’s introduction of the integrate-and-fire model neuron
(1907)”. In: Brain research bulletin 50.5-6 (1999), pp. 303–304.
[233] Frederic B Fitch. “McCulloch Warren S. and Pitts Walter. A logical calculus of
the ideas immanent in nervous activity. Bulletin of mathematical biophysics, vol.
5, pp. 115–133”. In: (1944).
[234] Andreas Knoblauch. “Neural associative memory for brain modeling and informa-
tion retrieval”. In: Information Processing Letters 95.6 (2005), pp. 537–544.
[235] Alan L Hodgkin and Andrew F Huxley. “A quantitative description of membrane
current and its application to conduction and excitation in nerve”. In: The Journal
of physiology 117.4 (1952), pp. 500–544.
[236] Catherine Morris and Harold Lecar. “Voltage oscillations in the barnacle giant
muscle fiber”. In: Biophysical journal 35.1 (1981), pp. 193–213.
[237] W Rall et al. “Matching dendritic neuron models to experimental data”. In: Phys-
iological Reviews 72.suppl_4 (1992), S159–S186.
[238] Romain Brette and Wulfram Gerstner. “Adaptive exponential integrate-and-fire
model as an effective description of neuronal activity”. In: Journal of neurophysi-
ology 94.5 (2005), pp. 3637–3642.
[239] Richard FitzHugh. “Impulses and physiological states in theoretical models of nerve
membrane”. In: Biophysical journal 1.6 (1961), pp. 445–466.
[240] RM Rose and JL Hindmarsh. “The assembly of ionic currents in a thalamic neuron
III. The seven-dimensional model”. In: Proc. R. Soc. Lond. B 237.1288 (1989),
pp. 313–334.
[241] Eugene M Izhikevich. “Simple model of spiking neurons”. In: IEEE Transactions
on neural networks 14.6 (2003), pp. 1569–1572.

136
 BIBLIOGRAPHY

[242] Bard Ermentrout. “Type I membranes, phase resetting curves, and synchrony”. In:
Neural computation 8.5 (1996), pp. 979–1001.
[243] Eugene M Izhikevich. “Which model to use for cortical spiking neurons?” In: IEEE
transactions on neural networks 15.5 (2004), pp. 1063–1070.
[244] Romain Brette. “What is the most realistic single-compartment model of spike
initiation?” In: PLoS computational biology 11.4 (2015), e1004114.
[245] Jan Benda and Andreas VM Herz. “A universal model for spike-frequency adap-
tation”. In: Neural computation 15.11 (2003), pp. 2523–2564.
[246] Boris Gutkin and Fleur Zeldenrust. “Spike frequency adaptation”. In: Scholarpedia
9.2 (2014), p. 30643.
[247] Yoshiki Kuramoto. Chemical oscillations, waves, and turbulence. Courier Corpo-
ration, 2003.
[248] Michael Breakspear, Stewart Heitmann, and Andreas Daffertshofer. “Generative
models of cortical oscillations: neurobiological implications of the Kuramoto model”.
In: Frontiers in human neuroscience 4 (2010), p. 190.
[249] Jianxia Cui, Carmen C Canavier, and Robert J Butera. “Functional phase re-
sponse curves: a method for understanding synchronization of adapting neurons”.
In: Journal of Neurophysiology 102.1 (2009), pp. 387–398.
[250] Arthur T Winfree. The geometry of biological time. Vol. 12. Springer Science &
Business Media, 2001.
[251] Leon Glass and Michael C Mackey. From clocks to chaos: the rhythms of life.
Princeton University Press, 1988.
[252] Xiao-Jing Wang and György Buzsáki. “Gamma oscillation by synaptic inhibition
in a hippocampal interneuronal network model”. In: Journal of neuroscience 16.20
(1996), pp. 6402–6413.
[253] Anastasiia Panchuk et al. “Synchronization of coupled neural oscillators with het-
erogeneous delays”. In: International Journal of Bifurcation and Chaos 23.12 (2013),
p. 1330039.
[254] LF Abbott and Carl van Vreeswijk. “Asynchronous states in networks of pulse-
coupled oscillators”. In: Physical Review E 48.2 (1993), p. 1483.
[255] Wulfram Gerstner. “Time structure of the activity in neural network models”. In:
Physical review E 51.1 (1995), p. 738.
[256] Wulfram Gerstner, J Leo Van Hemmen, and Jack D Cowan. “What matters in
neuronal locking?” In: Neural computation 8.8 (1996), pp. 1653–1676.

137
[257] David Hansel, Germán Mato, and Claude Meunier. “Synchrony in excitatory neural
networks”. In: Neural computation 7.2 (1995), pp. 307–337.
[258] Nancy Kopell and GB Ermentrout. “Symmetry and phaselocking in chains of
weakly coupled oscillators”. In: Communications on Pure and Applied Mathematics
39.5 (1986), pp. 623–660.
[259] RD Traub et al. “Analysis of gamma rhythms in the rat hippocampus in vitro and
in vivo.” In: The Journal of physiology 493.2 (1996), pp. 471–484.
[260] Alessandro Treves. “Mean-field analysis of neuronal spike dynamics”. In: Network:
Computation in Neural Systems 4.3 (1993), pp. 259–284.
[261] Theoden I Netoff et al. “Synchronization in hybrid neuronal networks of the hip-
pocampal formation”. In: Journal of neurophysiology 93.3 (2005), pp. 1197–1208.
[262] Michael A Schwemmer and Timothy J Lewis. “The theory of weakly coupled os-
cillators”. In: Phase response curves in neuroscience. Springer, 2012, pp. 3–31.
[263] Klaus M Stiefel and G Bard Ermentrout. “Neurons as oscillators”. In: Journal of
neurophysiology 116.6 (2016), pp. 2950–2960.
[264] Roy M Smeal, G Bard Ermentrout, and John A White. “Phase-response curves
and synchronized neural networks”. In: Philosophical Transactions of the Royal
Society B: Biological Sciences 365.1551 (2010), pp. 2407–2422.
[265] MR Jarvis and PP Mitra. “Sampling properties of the spectrum and coherency of
sequences of action potentials”. In: Neural Computation 13.4 (2001), pp. 717–749.
[266] Nicolas Brunel and Vincent Hakim. “Fast global oscillations in networks of integrate-
and-fire neurons with low firing rates”. In: Neural computation 11.7 (1999), pp. 1621–
1671.
[267] Nicolas Brunel and Vincent Hakim. “Sparsely synchronized neuronal oscillations”.
In: Chaos: An Interdisciplinary Journal of Nonlinear Science 18.1 (2008), p. 015113.
[268] Stephen Coombes, Peter beim Graben, and Roland Potthast. “Tutorial on neural
field theory”. In: Neural fields. Springer, 2014, pp. 1–43.
[269] G Bard Ermentrout and David H Terman. Mathematical foundations of neuro-
science. Vol. 35. Springer Science & Business Media, 2010.
[270] Wulfram Gerstner et al. Neuronal dynamics: From single neurons to networks and
models of cognition. Cambridge University Press, 2014.
[271] Hugh R Wilson and Jack D Cowan. “Excitatory and inhibitory interactions in
localized populations of model neurons”. In: Biophysical journal 12.1 (1972), pp. 1–
24.

138
 BIBLIOGRAPHY

[272] Hugh R Wilson and Jack D Cowan. “A mathematical theory of the functional
dynamics of cortical and thalamic nervous tissue”. In: Kybernetik 13.2 (1973),
pp. 55–80.
[273] Federico Devalle, Alex Roxin, and Ernest Montbrió. “Firing rate equations require
a spike synchrony mechanism to correctly describe fast oscillations in inhibitory
networks”. In: PLoS computational biology 13.12 (2017), e1005881.
[274] Stephen Keeley et al. “Firing rate models for gamma oscillations”. In: Journal of
neurophysiology 121.6 (2019), pp. 2181–2190.
[275] Xiao-Jing Wang and John Rinzel. “Alternating and synchronous rhythms in re-
ciprocally inhibitory model neurons”. In: Neural computation 4.1 (1992), pp. 84–
97.
[276] Miles A Whittington, Roger D Traub, and John GR Jefferys. “Synchronized oscil-
lations in interneuron networks driven by metabotropic glutamate receptor activa-
tion”. In: Nature 373.6515 (1995), pp. 612–615.
[277] Xiao-Jing Wang and György Buzsáki. “Gamma oscillation by synaptic inhibition
in a hippocampal interneuronal network model”. In: Journal of neuroscience 16.20
(1996), pp. 6402–6413.
[278] Carl Van Vreeswijk, LF Abbott, and G Bard Ermentrout. “When inhibition not
excitation synchronizes neural firing”. In: Journal of computational neuroscience
1.4 (1994), pp. 313–321.
[279] LS Leung. “Nonlinear feedback model of neuronal populations in hippocampal CAl
region”. In: Journal of neurophysiology 47.5 (1982), pp. 845–868.
[280] G Bard Ermentrout and Nancy Kopell. “Fine structure of neural spiking and syn-
chronization in the presence of conduction delays”. In: Proceedings of the National
Academy of Sciences 95.3 (1998), pp. 1259–1264.
[281] Christoph Börgers and Nancy Kopell. “Synchronization in networks of excitatory
and inhibitory neurons with sparse, random connectivity”. In: Neural computation
15.3 (2003), pp. 509–538.
[282] David Hansel and Germán Mato. “Asynchronous states and the emergence of syn-
chrony in large networks of interacting excitatory and inhibitory neurons”. In:
Neural computation 15.1 (2003), pp. 1–56.
[283] Atthaphon Viriyopase, Raoul-Martin Memmesheimer, and Stan Gielen. “Cooper-
ation and competition of gamma oscillation mechanisms”. In: Journal of neuro-
physiology 116.2 (2016), pp. 232–251.

139
[284] Atthaphon Viriyopase, Raoul-Martin Memmesheimer, and Stan Gielen. “Analyz-
ing the competition of gamma rhythms with delayed pulse-coupled oscillators in
phase representation”. In: Physical Review E 98.2 (2018), p. 022217.
[285] X-J Wang. “Fast burst firing and short-term synaptic plasticity: a model of neo-
cortical chattering neurons”. In: Neuroscience 89.2 (1999), pp. 347–362.
[286] Roger D Traub et al. “Single-column thalamocortical network model exhibiting
gamma oscillations, sleep spindles, and epileptogenic bursts”. In: Journal of neu-
rophysiology 93.4 (2005), pp. 2194–2232.
[287] Roger D Traub et al. “Persistent gamma oscillations in superficial layers of rat
auditory neocortex: experiment and model”. In: The Journal of physiology 562.1
(2005), pp. 3–8.
[288] Adrien Peyrache et al. “Spatiotemporal dynamics of neocortical excitation and in-
hibition during human sleep”. In: Proceedings of the National Academy of Sciences
109.5 (2012), pp. 1731–1736.
[289] Peter Barthó et al. “Characterization of neocortical principal cells and interneurons
by network interactions and extracellular features”. In: Journal of neurophysiology
92.1 (2004), pp. 600–608.
[290] Michel Le Van Quyen et al. “Large-scale microelectrode recordings of high-frequency
gamma oscillations in human cortex during sleep”. In: Journal of Neuroscience
30.23 (2010), pp. 7770–7782.
[291] Yann Zerlaut and Alain Destexhe. “Enhanced responsiveness and low-level aware-
ness in stochastic network states”. In: Neuron 94.5 (2017), pp. 1002–1009.
[292] Yann Zerlaut et al. “Modeling mesoscopic cortical dynamics using a mean-field
model of conductance-based networks of adaptive exponential integrate-and-fire
neurons”. In: Journal of computational neuroscience 44.1 (2018), pp. 45–61.
[293] Robert Friedman. “Themes of advanced information processing in the primate
brain”. In: AIMS neuroscience 7.4 (2020), p. 373.
[294] Horace B Barlow. “Single units and sensation: a neuron doctrine for perceptual
psychology?” In: Perception 1.4 (1972), pp. 371–394.
[295] Mark M Churchland et al. “Stimulus onset quenches neural variability: a widespread
cortical phenomenon”. In: Nature neuroscience 13.3 (2010), pp. 369–378.
[296] Alexa Riehle et al. “Spike synchronization and rate modulation differentially in-
volved in motor cortical function”. In: Science 278.5345 (1997), pp. 1950–1953.
[297] E Vaadia et al. “Dynamics of neuronal interactions in monkey cortex in relation
to behavioural events”. In: Nature 373.6514 (1995), pp. 515–518.

140
 BIBLIOGRAPHY

[298] Simon Thorpe, Arnaud Delorme, and Rufin Van Rullen. “Spike-based strategies
for rapid processing”. In: Neural networks 14.6-7 (2001), pp. 715–725.
[299] Sophie Deneve. “Bayesian spiking neurons I: inference”. In: Neural computation
20.1 (2008), pp. 91–117.
[300] Shigeru Shinomoto et al. “Relating neuronal firing patterns to functional differen-
tiation of cerebral cortex”. In: PLoS Comput Biol 5.7 (2009), e1000433.
[301] Gaby Maimon and John A Assad. “Beyond Poisson: increased spike-time regularity
across primate parietal cortex”. In: Neuron 62.3 (2009), pp. 426–440.
[302] Artur Luczak, Bruce L McNaughton, and Kenneth D Harris. “Packet-based com-
munication in the cortex”. In: Nature Reviews Neuroscience 16.12 (2015), pp. 745–
755.
[303] Sabine Kastner. “Attentional Response Modulation in the Human Visual System.”
In: (2004).
[304] Thomas Andrillon et al. “Neural markers of responsiveness to the environment in
human sleep”. In: Journal of Neuroscience 36.24 (2016), pp. 6583–6596.
[305] SL Shafer and DR Stanski. “Defining depth of anesthesia”. In: Modern Anesthetics
(2008), pp. 409–423.

141