Advanced Review

Multiscale modeling of brain

dynamics: from single neurons and
networks to mathematical tools
Constantinos Siettos1* and Jens Starke2

The extreme complexity of the brain naturally requires mathematical modeling

approaches on a large variety of scales; the spectrum ranges from single neuron
dynamics over the behavior of groups of neurons to neuronal network activity.
Thus, the connection between the microscopic scale (single neuron activity) to
macroscopic behavior (emergent behavior of the collective dynamics) and vice
versa is a key to understand the brain in its complexity. In this work, we attempt
a review of a wide range of approaches, ranging from the modeling of single
neuron dynamics to machine learning. The models include biophysical as well as
data-driven phenomenological models. The discussed models include Hodgkin–
Huxley, FitzHugh–Nagumo, coupled oscillators (Kuramoto oscillators, Rössler
oscillators, and the Hindmarsh–Rose neuron), Integrate and Fire, networks of
neurons, and neural field equations. In addition to the mathematical models,
important mathematical methods in multiscale modeling and reconstruction of
the causal connectivity are sketched. The methods include linear and nonlinear
tools from statistics, data analysis, and time series analysis up to differential
equations, dynamical systems, and bifurcation theory, including Granger causal
connectivity analysis, phase synchronization connectivity analysis, principal com-
ponent analysis (PCA), independent component analysis (ICA), and manifold
learning algorithms such as ISOMAP, and diffusion maps and equation-free
techniques. © 2016 Wiley Periodicals, Inc.
How to cite this article:
WIREs Syst Biol Med 2016. doi: 10.1002/wsbm.1348

Love looks not with the eyes but with the mind. A knowledge goes deep to such a detailed level, ranging
Midsummer Night’s Dream (I, i, 234). from the molecular, anatomical, and functional prop-
erties to the neurons–genes interactions. To date, we
know many details about how neurons transmit and
receive messages to other neurons through different
INTRODUCTION chemical molecules (neurotransmitters) and electrical
signals, how a neuron’s dendrites decode neurotrans-
T he rapid technological and theoretical progress
achieved in the past decades in biology has dras-
tically enhanced our arsenal in understanding the
mitters through special molecules called receptors,
how the messages pass from dendrites to the neu-
structure and physiology of neuronal cells. Our ron’s body, how an electrical potential emanates and
propagates with the aid of sodium and potassium
*Correspondence to: [email protected]
channels in the cell membrane through the neuronal
1 membrane to reach the axon and its terminal junc-
School of Applied Mathematics and Physical Sciences, National
Technical University of Athens, Athens, Greece tions, the synapses, which trigger the release of neu-
2
School of Mathematical Sciences, Queen Mary University of
rotransmitters to the other neurons. However, while
London, London, UK the physiology of neurons relies on the same ‘first
principles’, their anatomical and functional diversity
Conflict of interest: The authors have declared no conflicts of inter-
est for this article. is immense. Studies have revealed thousands of

© 2016 Wiley Periodicals, Inc.

 Advanced Review wires.wiley.com/sysbio

specific types of neurons and interneurons (which act ~10 ms–1 min), where the organization of brain
as ‘mediators’ for the establishment of the communi- activity emerges.
cation between sensory and motor neurons) provid- In this adventure of discovering how the brain
ing a basis of cognitive behavior.1 works, mathematical models, medical and biological
However, despite all the theoretical and tech- studies are going hand-in-hand in enhancing our
nological advances, we are still scratching the sur- understanding and leading to breakthroughs in the
face in understanding how the interplay between field of neuroscience. Over the years, a large number
molecular, cellular, chemical, anatomical, and neu- of models have been proposed, trying to shed light
ronal wiring/interconnection gives rise to complex from different perspectives. In general, these can be
brain functions such as vision, cognition, learning, classified in two main directions10–12: (1) the
and emotions on one side and dysfunctions such as bottom–up and (2) the top–down modeling
Alzheimer, epilepsy, Parkinson, and schizophrenia approach. The starting point of the bottom–up
on the other side. For example, only recently, studies approach is a set of dynamical equations that
have shown that the total memory capacity of the approximate the dynamics of neurons and/or net-
brain may have been underestimated by an order of works of neurons to the analysis of the complex
magnitude,2 that the N-methyl-D-aspartate (NMDA) emergent behavior. The top–down approach is data-
receptor antagonists that inhibit the transmission of driven. Based on the emergence of the brain dynam-
current to the so-called cortical GABAergic inter- ics as measured by e.g., EEG, MEG, fMRI (func-
neurons and pyramidal neurons is linked to patho- tional magnetic resonance imaging), builds models
logical features of schizophrenia,3 that brain based on systems theory, model identification, and
function (or dysfunction) is governed by the interac- nonlinear time series analysis. Both approaches
tions and organization between both neurons and weave a big and generally diverse literature. Here, we
genes rather than just neurons,4–6 or how the net- try to sketch the map of these categories and try to
work topology shapes brain responses to damages describe their basic underpinning concepts and dis-
and disorders.7 cuss their couplings as addressed in recent studies.
There is no doubt, that we have learned a great These categories are illustrated in Figure 1 and will
deal about the neuronal anatomical and functional be explained in what follows.
diversity. Yet, there is so much more to discover. The
function of brain with about 1012 neurons and 1015
neurons interconnections, packing densities (e.g.,
HISTORY AND DEVELOPMENT
within a cubic millimeter of the primate cortex, there
are about 105 neurons and about 109 synapses), and One of the very first attempts within the bottom–up
cortical folding (which results in a much larger sur- approach in trying to systematically quantify the
face area of the confined volume) in about 1.5 kg of dynamics of nerve cells in mathematical terms can be
mass is much more than a sum of single activities of traced back to 1907 when Lapicque introduced the
genes and neurons. It is more about the emergent modeling of a neuron’s activity by a simple integrate
behavior that rises from the complex interactions of and fire process using an electric circuit with a capac-
its subunits: all in all, what we call brain complex- itor and resistor in parallel, driven by a time-varying
ity.8 This emergent collective—macroscopic— current (see also e.g., Ref 14). McCulloch and Pitts
complex behavior is far from intuitive and cannot be in 194315 introduced the concept of the artificial neu-
straightforwardly predicted.9 From a dynamics point ron, while in 1952, Hodgkin and Huxley16 proposed
of view, the emergent behavior is characterized by the cornerstone model in the field describing the
unique spatiotemporal patterns and functionalities. dynamics of action potentials. This work was
Sustained oscillations, traveling waves, multiplicity of awarded the Nobel Prize in Physiology and Medicine
stationary states and spatiotemporal patterns are in 1963. Another early cornerstone that shaped the
some examples of the rich nonlinear dynamics emer- field is the work of Rall17 who originated the devel-
ging in the EEG/MEG-like level.10 For the systematic opment of compartmental models incorporating spa-
analysis and deeper understanding of such complex tial electro-physiological characteristics. On the
responses, one of the most critical issues in contem- phenomenological side, the celebrated FitzHugh–
porary neuroscience revolves around the bridging of Nagumo (FHN) model18,19 describing the propaga-
the different space and time scales8: the microscale tion of nerve pulses in the form of traveling waves
(~10−6–10−3 m, ~1 ms), where the single neuron has set the basis for further developments. Wiener
operates and interacts with other neurons, and that and Rosenbluth20 and Greenberg and Hastings intro-
of the meso-and macroscopic scales (~10−3–10−1 m, duced Cellular Automata for the phenomenological

© 2016 Wiley Periodicals, Inc.

 WIREs System Biology and Medicine Multiscale modeling of brain dynamics

Macro-level

Linear Non-linear Biophysical Phenomenological

Data-driven
Dimension Dimension
Equation-free
Reduction Reduction
Networks
PCA ISOMAP of regions
ICA Diffusion maps
Bifurcation analysis

Connectivity Phase Regions Neural field

Synchronization Networks equations
Analysis
PLV Kuramoto
Correlation

Model-based
Individualistic
PLI Wilson-Cowan
Coherence models
PSI (coupled
Granger Lyapunov oscillators)
Causality
-spectral G Mutual information
-PDC Transfer entropy
-dDTF Single-neuron
State-space Rössler
H-H
reconstruction Hindmarsh–Rose
Morris-Lecar
DCM integrate and fire
FHN

Micro-level

FI GU RE 1 | Overview of the presented modeling methods and mathematical tools for brain dynamics. The brain images were visualized with
the BrainNet Viewer.13

modeling of complex spatial patterns in excitable brain functionality or dysfuntionality by the informa-
media with interconnected neurons.21 Early models tion acquired at the macroscopic-emergent scale, i.e.,
of spatially localized neural populations can be found at the level of EEG, MEG, fMRI, or intracranial EEG
in Refs 22–24. Advances in nonlinear dynamics and (iEE) measurements. Hence, top–down modeling cov-
chaos theory as founded in the pioneering work of ers a wide range of scales extending from the recon-
Lorenz25,26 have facilitated the design, modeling, and struction of functional networks between specific
analysis of systems of coupled nonlinear oscillators brain anatomical regions to the circuit of specific
such as Rössler’s system27,28 and Hindmarsh–Rose groups of neurons down to the analysis of the impact
neuron.29 The dynamical behavior of these models of individual neurons. No doubt Hebb’s neuropsy-
contain all the dynamics observed both at the level of chological theory, known also as Hebbian learning,
real neurons, such as regular and irregular spiking that addressed the concepts of Hebb synapse, cell-
and bursting, and at the level of the collective brain assembly, and sequential activation of assemblies of
dynamics.30 Reviews on recent developments on the neurons, which fire together, strengthening their con-
modeling of single neuron as well as multiscale col- nection, has set the foundations in this direction,35
lective dynamics can be found among others in thus introducing the importance of causality. In the
Refs 31–34. 1960s, Granger defined causality in his seminal
On the other hand, the roots of the top–down paper36 by means of vector autoregressive models.
approach spring from systems identification, nonlin- The approach was exploited and boosted further
ear time series analysis, dynamics, and chaos theory. developments in computational neuroscience.
Within this framework, the aim is to build models Another key concept that is used and exploited to
and reveal the mechanisms that pertain to a specific reveal causal dependences between stochastic

© 2016 Wiley Periodicals, Inc.

 Advanced Review wires.wiley.com/sysbio

variables is that of mutual information By tuning the values of the model para-
(MI) addressed by Shannon and Weaver in 1949 meters67,68 and the structure of the neuronal
(see, e.g., Ref 37 for a review of this work), which is connectivity,69,70 one can get a variety of nonlinear
a nonparametric modeling approach related to spatiotemporal dynamics, which may in principle
entropy and conditional entropy. Finally, in the last approximate the characteristics of different types of
decades, the arsenal of nonlinear dynamics theory behaviors both at the individual and macroscopic
has enhanced our understanding and shed light on scale. For example, Lago-Fernández71 studied the dif-
the mechanisms that govern the emergence of com- ferences in the oscillatory patterns of neurons with
plex brain phenomena such as spatiotemporal respect to several types of network structures, and
chaos, pattern formation, and phase transitions. Terman et al.72 studied several types of networks
Measures of phase synchronization,34,35 transfer between globus pallidus, external, and subthalamic
entropy (TE) and MI,37–39 and embedding dimen- nucleus to approximate the dynamics of Parkinson
sion40,41 are used to quantify and reconstruct a mul- disease. Netoff et al.63 constructed small-world net-
tidimensional embedding space that geometrically work models of various types of neurons (Poisson
approximates the original dynamics. By now, the spike train, leaky integrate-and-fire, stochastic H–H)
benefits of using nonlinear dynamics theory in com- to study the link between epileptic seizures and the
putational neuroscience have been clearly triggering properties of the hippocampus CA3
demonstrated,10,42,43 and more promising develop- regions (exhibiting short bursts of activity) and CA1
ments are yet to come. regions (exhibiting seizure like activity lasting for sec-
onds). For increasing values of the rewiring probabil-
ity, the models were able to predict normal brain
BOTTOM–UP APPROACH: FROM function, transients of seizures, and continuous burst-
SINGLE NEURON MODELS TO ing, i.e., phase transitions driven by criticalities
observed in real brain signals of normal and abnor-
MULTISCALE COLLECTIVE DYNAMICS
mal activity. A recent review describing both the
Studies have proceeded to the development of dynamics of single neuronal and neural mass models
detailed state-of-the-art models aspiring to approxi- can be found in Ref 73.
mate the complex dynamics underlying the physics
and mechanisms of a wide range of problems stretch-
ing from the description of the behavior of certain Single Neuron
neural tissues, such as the cerebral neocortex Here, we have chosen to describe briefly four of the
(involved in higher cognitive functions such as per- most commonly used models in the field, namely the
ception, motor commands, and spatial leaky integrate and fire, the H–H, the FHN,
reasoning),44,45 cortical and thalamic (involving Hindmarsh–Rose and Rössler oscillators. A detailed
somatosensory functions as well as higher functions presentation and categorization of single neuron
such attention, language, and emotions),46–48 and the dynamics can be found in Refs 33,43.
hippocampus (involved in many functions including
the storage of long-term memory)49,50 to visual Leaky Integrate and Fire
hallucinations,51,52 phase transitions in human hand In its simplest form, this phenomenological model is
movements,53–55 and working-memory an extension of the Lapicque’s14 simple Integrate and
mechanisms56–58 and from neurological disorders Fire model reading:
such as schizophrenia59–61 and epileptic seizures62,63
to therapeutic surgical procedures such as deep-brain
τV_ = ðI + a −bV Þ for V < Vc ð1Þ
stimulation process for Parkinson disease64 to name
just a few.
When the voltage V of neuron’s membrane reaches a
These models are constructed on the individual-
certain threshold Vc, then its voltage is reset to its
neuronal level; they are composed of a large number
resting potential (about −70 mV), and a new cycle
of individual neurons interacting through a graph
begins with the integration of Eq. (1).
with known properties, which is of course a carica-
ture of the connections circuitry.65 The dynamics of
each single neuron are described by biophysical mod- Hodgkin–Huxley
els such as the Hodgkin–Huxley (H–H)16 and The H–H model16 is the cornerstone of biophysical
Morris–Lecar66 models or phenomenological repre- modeling in neuroscience. It was introduced in 1952
sentations such as the FHN equations.18,19 to explain the ionic mechanisms underlying the

© 2016 Wiley Periodicals, Inc.

 WIREs System Biology and Medicine Multiscale modeling of brain dynamics

generation and propagation of action potential in the study both individual and collective dynamics in
squid giant axon. The H–H model considers three computational neuroscience74,75. Its dynamics is
currents, namely sodium (Na+), potassium (K+), and given by the following set of equations:
a leakage current, which is primarily due to the chlo-
ride (Cl−): X_ = − Y −Z + I
Y_ = X + aY
CV_ = I − gNa m3 hðV − ENaÞ − gK n4 ðV − EKÞ − gCl ðV− ECl Þ Z_ = b + ZðX −cÞ ð4Þ
_ = ðm ∞ ðV Þ −mÞ
τm m
τn n_ = ðn ∞ ðV Þ− nÞ The Hindmarsh–Rose model29 extends the FHN
τh h_ = ðh ∞ ðV Þ− hÞ ð2Þ equations by a third dimension:

The voltage difference V describing a relatively slow X_ = Y − aX3 + bX2 −Z + I

behavior across the neuron’s membrane is assumed Y_ = c −dX2 −Y
to be explicitly dependent from Na+ and K+ channels Z_ = r½sðX − XR Þ −Z ð5Þ
that govern the flow of those ions through the cell
membrane, while the leakage current, which is pri- In both models, X approximates the (slow) dynamics
marily due to Cl−, represents the action of all other of the membrane’s potential, while Y and Z in anal-
channels. The constants Ei, i = Na, K, Cl are the ogy to the H–H model simulate the (fast) ions trans-
reversal potentials, and gi are conductances. The vari- portation dynamics across the neurons’ membrane.
ables m, n, and h represent the probabilities that a As in H–H, I is the externally applied current, such
channel is activated or deactivated, while the con- as the one that may come from other neurons in the
stants τi are their corresponding time constants. Na+ network. The different selection of the model para-
gates are described by m and h, while n describes the meters results in a rich variety of temporal pat-
K+ channel; the index ∞ denotes values at equilib- terns.76,77 The H–H, Rössler, and Hindmarsh–Rose
rium, which are functions of V whose parameters are a kind of universal models that can capture the
have been fitted to experimental data. C denotes the essential dynamics observed in single neurons,
membrane’s capacity, and I is the externally applied groups of neurons, and brain areas. A comparison of
current, which may come from other neurons in the their efficiency in approximating specifics such as
network. spiking, bursting, and chaotic patterning of different
types of real neurons can be found in Refs 78,79.
FitzHugh–Nagumo These specific dynamics can be systematically identi-
The FHN model18,19 is a phenomenological model, fied and categorized with the aid of bifurcation
representing a simplification of the H–H equations, analysis.39
given by:

 Multiscale Dynamics: Collective
V3
V_ = τ I + V − +W Dynamics, Mean Field Approximations,
3
1 and Beyond
_ = − ðV −a + bW Þ
W ð3Þ However, even if we know the details about how
τ
neurons function at the molecular level, it is the
The voltage V of the neuron’s membrane represents higher level of complexity emerging in the collective
the slow dynamics, while the recovery dynamics of groups of neurons or networks of
W approximates the hyperpolarizing action potential groups of neurons that governs the behavior of the
in the neuron’s membrane (representing the fast brain and its functions. One then typically faces the
dynamics). Note that the FHN model for a = b = 0 not-trivial task of extracting macroscale dynamics
reduces to the nonlinear Van der Pol oscillator.18 (at the level of a region of the brain) from the micro-
scale information (at the neuronal level). However,
the collective (macroscopic) behavior of such multi-
Rössler Oscillator and the Hindmarsh–Rose cellular systems is far from intuitive and cannot, most
Neuron of the time, be straightforwardly predicted (see e.g.,
The Rössler oscillator27,28 is a simplification of the the discussion in Ref 9).
Lorenz equations25,26 targeted on climate modeling. Hence, one of the most challenging problems
However, Rössler oscillators have also been used to revolves around the bridging of the different scales:

© 2016 Wiley Periodicals, Inc.

 Advanced Review wires.wiley.com/sysbio

the microscopic scale, where the interactions of the between the ith and jth subpopulation, while P(t)
neurons take place, and that of the macroscopic and Q(t) are the average external stimuli to the excit-
scale, where the overall dynamics of the brain atory and inhibitory subpopulations, respectively.
emerges. Traditionally, the gap between high- Wilson and Cowan46 extended the above
dimensional (microscopic) and low-dimensional model to incorporate spatial interactions between
(coarse-grained level) scales is bridged through clo- groups of localized aggregates of excitatory and
sures, which relate fast higher dimensions to a few inhibitory neurons, which in a simplified form are
slow dimensions of the underlying detailed dynamics given by Refs 92–94:
(see e.g., Refs 80–84).
dEðx,tÞ
τE = − E + ð1− rE EÞFE ½cEE *E−cEI *I + Pðx, tÞ
Neural (Mean) Field Models dt
Mean-field approximations are typically the first dIðx, tÞ
τI = −I + ð1 − rI IÞFI ½cIE *E −cII *I + Qðx, t Þ
mesoscopic equations one tries to derive in the form dt
of ordinary differential equations (ODEs) or partial ð6bÞ
differential equations (PDEs) in a closed form,
describing the coarse-grained dynamics of a popula- where * is the convolution operator defined as
tion of neurons with similar characteristics. Within
ð
this formulation, networks of different types of neu-
ral masses provide the basis for approximating the cjk Sk = cjk ðx− yÞSðy,t Þdy ð6cÞ
macroscopic dynamics of brain activity.33,85–87 One Ω
of the very first successful paradigms is the famous
Wilson–Cowan model of interacting neurons24,46 Ω represents the spatial domain, while typical con-
and the Kuramoto model of coupled nonlinear nections strengths in the spatial network are given by
oscillators.88–91 Gaussian functions defining exponential decrease in
the neuronal strength spatial connectivity:
The Wilson–Cowan Model
ðx− yÞ2
The original Wilson–Cowan’s model consists of two −
σ2

nonlinear ODEs approximating the mean evolution cjk ðx− yÞ = wjk e jk ð6dÞ
of the activation densities of the two main types of
neurons in the brain: the excitatory and inhibitory
neurons that release neurotransmitters that favor or
inhibit the generation of nerve impulses (action The Kuramoto Model
potentials). The lumped (spatially constant/localized) The Kuramoto model approximates the dynamics of
neuronal aggregate activity reads24: a population of N neurons by means of phase
oscillators:
dE
τE = − E + ð1 − rE EÞFE ½cEE E− cEI I + PðtÞ X
N  
dt
ð6aÞ θ_ i = ωi + Kij sin θj −θi , i = 1, 2, …, N ð7aÞ
dI j=1
τI = − I + ð1 − rI IÞFI ½cIE E− cII I + Qðt Þ
dt
θ stands for the phase and ω for the natural fre-
where E(t) and I(t) are the densities of excitatory and quency. By assuming homogeneous (mean-field) cou-
inhibitory neuronal subpopulations firing per unit K
pling with an average coupling strength, Kij = N , and
time at time instant t; τE,I are the corresponding syn-
a collective (mean-field) phase, say Ψ , one obtains:
aptic delays, i.e., the times required from the release
of neurotransmitters at the synapse to the onset of an
KX N
action potential at the membrane; rE,I is the absolute θ_ i = ωi + sin ðΨ − θi Þ, i = 1, 2, …, N ð7bÞ
refractory period, i.e., the period during which a N j=1
stimulus to the neuron will not lead to a second
action potential; and FE, FI are the nonlinear (sig- Defining the phase coherence, 0 ≤ r ≤ 1, quantifying
moid) subpopulation response functions giving the how much the oscillators maintain a common phase
mean proportion of cells in a subpopulation that
1X N
would respond to a given level of excitation. The as re iΨ = e iθj , the above equation becomes:
constants cij represent the average strength of links N j=1

© 2016 Wiley Periodicals, Inc.

 WIREs System Biology and Medicine Multiscale modeling of brain dynamics

Ref 117). Good closures in the form of ODEs or

θ_ i = ωi + rKðΨ −θi Þ, i = 1, 2, …,N: ð7cÞ
PDEs describing the coarse-grained dynamics may,
in principle, be extracted from the individualistic
In the limit of infinite population, i.e., N ! ∞, we interactions but are usually unavailable or extremely
can define the behavior of the density of the oscilla- difficult to derive. The problem is that due to het-
tors by considering the continuity equation: erogeneities, nonlinearities, and stochastic interac-
tions, serious limitations come up when trying to
∂ρðω, θ, t Þ ∂ _ systematically bridge the different time and spatial
+ ðvρðω, θ, tÞÞ = 0, where v ≡ θ: ð7dÞ
∂t ∂θ scales in order to extract effective macroscopic
equations.
Based on the above, the dynamics of the oscillators Without the availability of such good closures,
in the continuum can be described by the following what is usually done with such detailed microscopic
mean-field equation: simulators is the exploration of the coarse-grained
dynamics by performing long simulations over time,
∂ρ ∂ starting from different initial configurations and aver-
+ ð½ω + rKðΨ −θÞρÞ = 0; ð7eÞ
∂t ∂θ aging over large ensembles.
However, this approach is both computation-
that has toðbe solved together with the normalization ally consuming and inadequate for important tasks
π
constraint ρðω, θ, t Þ = 1. such as the systematic detection of coarse-grained cri-
−π ticalities, phase transitions, stability analysis, and the
The microscopic basis of other mean-field mod- derivation of solution branches (including unstable
els are different types of nonlinear oscillators,95,96 ones) in parameter space.
H–H,97,98 FHN,99 and Integrate and Fire83,100–104 as
well as Cellular Automata105–107 models. Sustained Equation-Free Multiscale Modeling
oscillations, traveling waves, multiple stationary Over the last years, the Equation-Free computa-
states, and spatiotemporal chaos are some examples tional framework introduced in Refs 118–120 for
of the rich nonlinear dynamics that can emerge from multiscale computations has been used to efficiently
such approximations at the coarse-grained popula- extract ‘systems-level’ information from detailed
tion level. These neural field models are formed by microscopic simulators, bypassing the need for
assuming explicit equations (called closures) relating obtaining analytical closure approximations, thus
higher- with lower-dimensional state descriptions. A providing an ‘on-demand’ model identification and
review of such multiscale modeling approaches and reduction. The main assumption is that a macro-
closures can, for example, be found in Refs scopic model in the form of ODEs or PDEs exists in
33,108,109. principle and closes in terms of relatively few mac-
Having such mean-field approximations, one roscopic variables, such as the densities of mem-
can systematically investigate the dynamics through brane voltages or populations of exhibitory and
the tools of bifurcation analysis. Tracing both sta- inhibitory neurons. Typically, these are the low-
ble and unstable branches of fixed points and order moments of the underlying microscopically
nonlinear periodic oscillations, the detection of crit- evolving distributions. The existence of a coarse-
ical points that mark the onset of changes in grained model implies that the higher-order
qualitative behavior can be accomplished with the moments, such as the densities of the ions channels,
use of state-of-the-art computational tools (such as become relatively quickly over the time scales of
AUTO,110 MATCONT,111 and COCO112) for such interest, determined (slaved) by a few macroscopic
system-level analysis. For applications of such tools variables. The Equation-Free methodology provides
in the analysis of problems in neuroscience, see these closures on demand: relatively short bursts of
e.g., Refs 113–116. fine scale simulations naturally establish this slaving
However, closures are often based on ad hoc relationship (Figure 2). Branches of both equilibria
assumptions such as uniformity of neurons, homo- and limit cycles, stability analysis, detection of criti-
geneous networks, and infinite populations. These cal points, and rare event analysis for the quantifica-
may introduce systematic biases and possibly miss tion of phase transitions can be computed in a
important qualitative and quantitative information computationally efficient manner within this frame-
at the coarse-grained level (see, e.g., the critical dis- work. There are two fundamental assumptions of
cussion on the influence of certain closure approxi- the approach, namely: (1) time scale separation
mations applied to integrate and fire networks in between the micro and macro dynamics and (2) the

© 2016 Wiley Periodicals, Inc.

 Advanced Review wires.wiley.com/sysbio

TOP–DOWN APPROACH: FROM

BRAIN SIGNALS TO NEURONAL
CIRCUITS
Lift Restrict
U(t0)
Neuronal
U(t0 + T) The current view of the brain as a complex self-
 M
simulator organizing network that constantly modifies its
functional architecture to achieve different ‘missions’
has replaced the simplistic model of the brain
u(t0 + T)
viewed as a complicated machine in which certain
u(t0) Coarse timestepper regions of the brain have unique functions and are
u(t0 +T) = FT(u(t0),p) thus related by a one-to-one relationship to specific
cognitive processes.133 For example, the accomplish-
ment of tasks like visual perception or motor pro-
Stationary/ Bifurcation
time-dependent parameter gramming and execution, which are continuously
solutions and
stability analysis
carried out by the human brain, requires the coordi-
nation of neural resources that are anatomically dis-
Bifurcation tributed all over the brain. A fundamental question
anlysis tools
that then arises is how these neuronal elements are
functionally integrated into the cognitive system (for
F I G U R E 2 | Schematic of the Equation-free multiscale modeling a review see, e.g., Ref 134). In a way, the top–down
framework.119,120 Once the appropriate macroscopic observables have
approach aspires to give an answer to the inverse
been identified, the following steps provide a bridge between the
microscopic and macroscopic modeling scale: (1) Prescribe a
problem, i.e., how macroscale phenomena observed
macroscopic initial condition u(t0). Define also a restriction operator M at the, say, EEG, MEF or fMRI scale can be
mapping the microscopic-level description to the macroscopic one U; exploited to infer about this brain organization and
that is, u = M[U]. (2) Transform the macroscopic initial condition u(t0) ultimately explain the underlying functional mechan-
through a lifting operator, μ, into consistent microscopic realizations: isms and causal influences among anatomical
U(t0) = μ[u(t0)]. (3) Evolve these realizations in time using the regions. Toward this aim, a variety of approaches
microscopic simulator for a short macroscopic time T generating the have been proposed, ranging from simple correla-
microscopic distribution U(t0 + T). The choice of T is associated with tion and coherence135–138 to Granger causality-
the (estimated) spectral gap between micro and macro level. based methods36,139–141 and from phase-synchrony
(4) Obtain the coarse-grained values using the restriction operator detection142,143 and MI-based graphs41,144 to mani-
M : uðt0 + TÞ = M½Uðt0 + TÞ. The above procedure defines a so-called
fold or machine learning algorithms such as inde-
coarse timestepper from time t0 to t0 + T between macroscopic state
dynamics: uðt0 + TÞ = FT ½uðt0 Þ, p where p 2 Rq is the parameter
pendent component analysis (ICA),145–147
148,149
vector of the system. At this point, one can utilize bifurcation analysis ISOMAP, and Diffusion Maps150 for the
tools as shells ’wrapped around’ the coarse timestepper to trace reconstruction of embedded networks. These meth-
branches of coarse-grained equilibria or periodic solutions (even ods can be categorized as linear, nonlinear, paramet-
unstable ones) and perform stability and rare-events analysis (e.g., ric or nonparametric.
compute transition rates between the apparent coarse-grained states).
Coupling Equation-Free with manifold learning algorithms such as
Diffusion Maps allows the identification of the dominant macroscopic
states.121 LINEAR METHODS
Nonparametric Approaches
Two of the simplest measures for the quantification
a priori knowledge of the coarse-grained ‘slow’ vari- of interdependencies among time series are correla-
ables. The Equation-Free approach has been coupled tion (in the time domain) and coherence (in the fre-
with advanced nonlinear machine-learning techni- quency domain). Let us assume two discrete zero-
ques such as Diffusion Maps122,123 that can suggest mean stationary times series x(t), y(t). Correlation at
the right coarse-grained observables.121 The a time lag τ is defined as:
Equation-Free multiscale modeling approach has
been proven very useful by successfully handling X
∞
complexity in a wide range of models, including H– Rxy ðτÞ = xðtÞyðt + τÞ ð8aÞ
H,124 Integrate and Fire,125–128 coupled −∞
129,130
oscillators, and Markov chains on
networks.131,132 which is usually normalized in [−1, 1] to read:

© 2016 Wiley Periodicals, Inc.

 WIREs System Biology and Medicine Multiscale modeling of brain dynamics

Rxy ðτÞ where e1ja12, j = 0 is the uncertainty when the para-

rxy ðτÞ = pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi ð8bÞ
Rxx ðτÞRyy ðτÞ meters a12,j are zero. Granger-causality has also been
extended to model interdependencies in the frequency
where Rxx(τ), Ryy(τ) are the autocorrelations. domain by taking the Fourier transformation of auto-
The spectral coherence Γ xy(f ) is defined as: regressive models.140,141,173 Due to the construction
of the method, Granger-causality can infer just linear
 
Sxy ð f Þ interdependencies in stationary signals. Another limi-
Γ xy ð f Þ = qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
 ffi ð8cÞ tation of Granger-causality is that it cannot distin-
Sxx ð f ÞSyy ð f Þ guish between direct and indirect influences in
multivariate systems. In addition, as the emergent sig-
where Sxy(f ), Sxx(f ), Syy(f ) are the cross-spectral and nal produced by a source in the brain may be recorded
autospectral densities of x(t), y(t), respectively. from several (neighboring) channels, the method may
One of the first applications of coherence in identify fake interconnections. This phenomenon is
neuroscience was the investigation of activity of sei- known as volume conduction effect. A contemporary
zures discharge in epilepsy over a range of anatomi- review on Granger causality can be found in Ref 139.
cal connections,135 while in Ref 136, the concept of To deal with the last two drawbacks, several other
correlation was used to detect functional groups of methods based on the Granger-causality concept, such
neurons based on the observation of impulses. Since as the partial direct coherence (PDC)174 and the direct
then, both measures continue to hold a significant transfer function (dDTF),175 have been addressed. A
share in the analysis of functional networks137,138, comparison of the efficiency of various linear
especially on the basis of blood-oxygen-level depend- approaches to detect directed interactions in multivar-
ent (BOLD) signal fluctuations as measured by an iate systems can be found in Ref 176.
fMRI.151–153 Reviews of these methods can be found
in Refs 137,154. Nonlinear Methods

Parametric Approaches Parametric Approaches

The key concept that underlies linear parametric meth- Several nonlinear extensions of the parametric Gran-
ods is causality as introduced by Granger36: a variable ger-causality approach have been proposed.177–179
X Granger-causes another variable Y, if past values of Signals are approximated by nonlinear functions,
X help to forecast the evolution of Y in a better way for example, in a two dimensional system as:
than using just the past values of Y. Since then, Gran-
ger-causality has been applied in several studies, mainly xðtÞ = c10 + c11 f11 ðXÞ + c12 f12 ðY Þ + c13 f13 ðX, Y Þ + e12 ðtÞ
in fMRI155–164 and in MEG and EEG.165–172 Within yðt Þ = c20 + c21 f21 ðXÞ + c22 f22 ðXÞ + c23 f23 ðX, Y Þ + e21 ðt Þ
this framework, the dynamics of the time series are ð10Þ
approximated by (multivariate) autoregressive mod-
els.36,139 For a two-dimensional system, this reads as: where X = [x(t − 1), x(t − 2), …, x(t − p)], Y = [y
(t − 1), y(t − 2), …, y(t − p)].
X
p X
p
In analogy to linear causality, y causes x if the
x ðt Þ = a11, j xðt − jÞ + a12, j yðt −jÞ + e1 ðt Þ prediction error e12 is lower than the prediction error
j=1 j=1
ð9aÞ for c12 = c13 = 0 when taking the full model with
Xp Xp
y ðt Þ = a21, j xðt −jÞ + a22, j yðt −jÞ + e2 ðtÞ c10, c11, c12, c13 6¼ 0.
j=1 j=1 However, such nonlinear parametric
approaches lack a systematic way of choosing appro-
where p is the order of the model, and e1 and e2 are priate nonlinear functions on one hand, and on the
the uncertainties in the modeling. Let us consider the other hand, they suffer from the ‘curse of dimension-
prediction of x in terms of past values of x and past ality’ related to the identification of the model order
values of y. If the variance of e1 is reduced by for (real-world) multivariate problems.
encountering y in the prediction of x, then one says
y Granger-causes x. This influence is measured as: Nonparametric Approaches

var e1ja12, j = 0 Phase Synchronization
G2!1 = ln ð9bÞ Brain activity can be measured by either increases
varðe1 Þ or decreases of the signal amplitude at a specific

© 2016 Wiley Periodicals, Inc.

 Advanced Review wires.wiley.com/sysbio

frequency range or by measuring the phase relation Mutual Information and Transfer Entropy
between different (EEG/MEG) channels.180 The Over the last years, the concept of MI and that of the
changes in the signal amplitude are measured locally Shanon TE37–39 have been exploited and advanced
at the level of each MEG/EEG channel reflecting the to infer about directed flow of information between
activity of groups of neurons, while the phase syn- two random processes also in the presence of
chronization quantifies the interaction between nonlinearity.74,144,198–201
MEG/EEG channels (groups of neurons) and there- The MI of two (discrete) random processes
fore can be used as a measure of their interdepend- X, Y with probability distributions p(X), p(Y) is
ency.181 Hence, a fundamental concept in identifying defined as:
functional interdependencies is the one of phase syn-
chronization. This concept is related to the theory of XX pðyjxÞ
chaotic oscillators. It has been shown that synchroni- IðX, Y Þ = pðyjxÞ log ; ð12aÞ
χ y
pðyÞ
zation among segregated brain areas through phase
synchronization characterizes memory
where the conditional probability distribution p(y|x)
processes,182,183 conscious perception,184–186 and
has to be estimated from large numbers of repeti-
attention.187 The same concept underpins the phe-
tions. In Xterms of the Shannon entropy, defined as
nomenon of phase desynchronization, also known as
phase scattering.188,189 H ðX Þ = − pðxÞ logpðxÞ, the MI can be written as:
x
One measure for phase synchronization is the
phase-locking value (PLV),142 defined as:
IðX, Y Þ = H ðXÞ− H ðXjY Þ = H ðY Þ− H ðY jXÞ: ð12bÞ
 
1 XN 
ið φi ðtÞ − φj ðt ÞÞ  The notion of TE39,144 in turn that measures directed
PLVij ðt Þ =  e : ð11aÞ
N n = 1  flow of information between two random processes
is related to the conditional entropy H(Y|X) as:
The above measure reflects the average phase dif-  
ference φi(t) − φj(t) for a pair of channels i, j over TEX!Y = H Yi + 1 jYi, i− 1, …, i− k
N trials. The above measure ranges from zero  
(random phase relation) to one (perfect phase −H Yi + 1 jYi, i− 1, …, i − k , Xi, i −1, …, i −l : ð12cÞ
locking).
A problem encountered by the implementation
of PLV is that it is also affected by the volume con-
Embedding Dimension and Reconstruction
duction effect.142 To overcome this issue, another
of the State-Space
measure was introduced in Ref 143 called phase lag
The celebrated Takens embedding theorem,202 which
index (PLI), defined as:
can be considered the ‘heart’ of nonlinear time series
  theory, provides the basis for quantifying couplings
1 XN h i
 in nonlinear dynamical systems.40,41 The problem
PLIij ðt Þ =  sign φi ðtÞ − φj ðtÞ  ð11bÞ
N n = 1  here is to reconstruct the state-space X 2 Rn by
means of m time series measurements S 2 Rm, i.e., to
In comparison to the PLV, the phase synchroniza- construct a topological equivalence relation between
tion index PLI discards phase synchrony that is X(t) and S(t). There are two different embedding pro-
centered around 0 mod π. Hence, PLI identifies cedures.10,203 The first approach finds an appropriate
‘shifted’ synchronization, i.e., synchronization state-space representation from a single time series
where the phase difference has a constant non-zero using the concept of coordinate delay. The second
value. An evaluation of its performance, for brain one, on the basis of a generalization of the original
network inference in EEG, can be found in Takens theorem,204 uses spatial embedding where
Ref 190. the embedding dimension is taken equal to the num-
Other phase synchronization estimators include ber of channels. Within the time delay, the state-
the mean phase coherence,191 the phase-slope index space is reconstructed by considering the basis
(PSI),192 the weighted phase lag index (WPLI),193 vectors:
and the use of Lyapunov exponents.194–197 Statisti-
cally significant values of phase synchronization are vðt Þ = ½sðtÞ, sðt −τÞ, …, sðt − ðm −2ÞτÞ, sðt − ðm −1ÞτÞ;
determined by means of surrogate data. ð13aÞ

© 2016 Wiley Periodicals, Inc.

 WIREs System Biology and Medicine Multiscale modeling of brain dynamics

where m is the embedding dimension. For a spatial In many biological applications, particularly in
embedding, the basis vectors are given by: neuroscience, it is relevant to identify and separate
different functional units, such as neuronal groups,
vðtÞ = ½s1 ðt Þ, s2 ðtÞ, …, sm − 1 ðtÞ, sm ðtÞ; ð13bÞ from each other as well as subtract artifacts such as
heartbeat, motion, or measurement noise from the
where si(t) it the signal of the ith channel at time t. original data. This can be achieved with ICA, which
An assessment of the efficiency of Granger- finds a low-dimensional description of the data by
based measures and nonlinear methods based on TE, searching for statistically independent spatial
MI, and embedding reconstruction measures can be components.146,210–214 ICA is the most used method
found in Ref 205. to analyze fMRI data. For this example, the method
can be represented as:
Manifold Learning Algorithms
Neuroimaging techniques, especially fMRI analysis, X
M

provide signals that are contained in a high- sðt, x, y, zÞ = Ai ðt ÞCi ðx, y, zÞ ð14Þ
i=1
dimensional space. In a typical 4D functional connec-
tivity fMRI scan, more than 100,000 voxels (which
contain the activity information of millions of neu- where s(t, x, y, z) is the measured signal fluctuation
rons) are recorded at each time instance. Each voxel (e.g., BOLD) at time t at position (x, y, z). The coeffi-
has various (noisy) data; thus in the four dimensions cient Ai(t) represents the temporal amplitude, and
of space and time, one gets several million data mea- Ci(x, y, z) is the spatial magnitude of the ith ICA
surements. In addition, experiments are repeated over component.
a group (typically between 10 and 50) of subjects to ICA has also been used to extract information
facilitate statistical analysis. Hence, one is confronted about the connectivity of the brain by measuring the
with two challenges: that of dimensionality reduction strength of interaction between the components Ai(t)
and the problem of data mining, i.e., extracting fea- as this is computed, for example, from a correlation
tures (biomarkers) that pertain to the investigated analysis.148,209,215–217
brain function within each subject and/or that clas- When the linear methods such as PCA and ICA
sify differences among subjects or groups of subjects. do not suffice to produce a good low-dimensional
Usually, a first step to reduction in dimensional- data representation, nonlinear manifold learning
ity and noise is the implementation of linear techni- algorithms such as the ISOMAP149 and Diffusion
ques such as the principal component analysis (PCA) Maps122,123 that preserve the relative distance
and ICA, which are often used in brain imaging [such between neighboring nodes can be used to find the
as fMRI and positron emission tomography (PET)] low-dimensional embedded connectivity
to reduce the high-dimensional signal space into a network.148,150,218–220
lower-dimensional one.206–209 The ISOMAP construction of the embedded
PCA searches for a low-dimensional representa- network can be described briefly by the following
tion defined by a few orthogonal vectors (compo- three steps:
nents), which successively account for the maximum
possible variability such that the error is minimized 1. Construction of the connectivity graph based
when reconstructing the original data with a low- on the k-nearest neighbors of the components.
dimensional approximation. The PCA can be used 2. Computation of the shortest path between each
among other methods to find components that repre- pair of nodes in the graph. This will result in a
sent dominant functional connectivity patterns hid- square matrix D, with elements representing
den in the correlation matrix C, which contains the the length of the shortest paths (called geodesic
resulting M × M correlation matrices among sub- distances) among nodes.
jects.208 These matrices are usually formed by taking 3. Construct the d-dimensional embedding space.
the maximum of |rxy(τ)|, measuring the strength of The embedded space is spanned by the first
interaction between the high-dimensional data. These d eigenvectors corresponding to the largest
dominant connectivity patterns correspond to the eigenvalues of D.
eigenvectors of the leading eigenvalues of the covari-
T
ance matrix CC , where C results from C by sub- Diffusion Maps create time-dependent diffusion pro-
tracting the mean of each row from each matrix cesses to embed the high dimensionality data space
element in the corresponding row.208 to a lower one. Here, connectivity between two

© 2016 Wiley Periodicals, Inc.

 Advanced Review wires.wiley.com/sysbio

(stochastic) time series is defined as the probability of stimuli. The vector p defines the model parameters
jumping in one step from one point to another by related to functional connectivity between neurons,
means of a random walk. This probability is analo- groups of neurons, or brain regions. Instead of using
gous with what is called diffusion kernel, reflecting the nonlinear model (15a) per se, its bilinear approxi-
the similarity between points with respect to the mation is often used:
underlying geometric topology of the state-space.
These probabilities define the weights of the network X
M
links. Assuming a high dimensional space X 2 RN, _ = AX +
X Bi ui X + Cu; ð15bÞ
the procedure reads as follows122,123: i=1

∂f
1. Compute the NxN similarity matrix K based where A = ∂X reflects the fixed connectivity in the
2
∂ f
on a kernel k(Xi, Xj). For example, one can use absence of input and the matrices; Bi = ∂X∂u reflects
i
  j Xi − Xj j the change in coupling strength induced by the jth
the Gaussian kernel k Xi , Xj = e − ε , where
input; and the matrix C reflects the impact of stimuli.
ε is the kernel scale.
The objective is to estimate the model’s parameters,
2. Create the Markovian matrix by K0 = D− 1K, i.e., the elements of the matrices A, Bi, C, by the infor-
where D is the matrix resulting from a summa- mation acquired on the macroscopic level. These para-
tion of the rows of K. meters represent the functional interactions among
3. Construct the d-dimensional embedding space. brain regions at a neuronal level.229–231 DCM has been
The embedded space is spanned by the first used to unfold the neuronal circuitry between and
d eigenvectors corresponding to the largest within regions involved in several key brain processes
eigenvalues of K0 . and responses ranging from visual perception,232
emotion,233,234 and anxiety235 to synaptic-level analy-
A review at the intersection between statistical sis236 as well as dysfunctions such as depression,237
machine-learning techniques, including graph embed- schizophrenia,238 Parkinson’s disease,239and bipolar
ding techniques, can be found in Ref 221. disorder240 to name just a few. In most of these studies,
DCM is coupled with Bayesian Model Selection to
infer the best linear of the nonlinear model that per-
BOTTOM–UP AND TOP–DOWN tains to the specific brain function.235,240,241
APPROACHES At this point, we should also mention frame-
works that are based on machine learning such as
In recent years, there is an intense effort to build
LAMINART, a real-time probabilistic decision cir-
mathematical approaches that provide a bridge
cuit modeling of visual processes,45 and the hierarchi-
between neuroimaging phenomenological recordings
cal temporal memory (HTM) framework242
and the biophysical (mechanistic) level on neuron
composed of learning algorithms such as Bayesian
dynamics.11 These studies incorporate advances of
networks and clustering as well as theories that
both bottom–up and data-driven top–down
model functions of the neocortex.
approaches. Examples of such attempts include the
approximation of the dynamics of the hemodynamic
activity in visual cortex222 and the dynamics of disor-
ders such as Parkinson’s disease,223–225 dopamine’s CONCLUSIONS
effect in schizophrenia,226 and epilepsy.227,228 The huge progress achieved in neuroscience in the
A general modeling approach that targets the last decades was possible by multidisciplinary activ-
reconstruction of effective connectivity networks is ities where experimental investigations were going
dynamic causal modeling (DCM).229 In particular, hand in hand with a massive support by mathemati-
DCM treats the brain as a network of deterministic cal modeling and analysis on multiple scales. This is
nonlinear dynamic subsystems (groups of neurons, or particularly remarkable as the brain represents one
brain regions) whose dynamics can be described by a of the most complex systems we currently investigate.
set of nonlinear ODEs of the form: Mathematical tools facilitate a deeper understanding
and allow the testing of hypotheses in the functional
_ = f ðX, u, pÞ
X ð15aÞ context of biological subsystems such that they
uncover structure and connectivity. The obtained
where X 2 RN represents the states of the N different results and insights where multiscale modeling math-
subsystems, and u 2 RM represents the signals of ematical methods have proven to give a lot of

© 2016 Wiley Periodicals, Inc.

 WIREs System Biology and Medicine Multiscale modeling of brain dynamics

additional insight to mechanisms that pertain to spe- ‘biomarkers’ of specific cognitive mechanisms and
cific brain functions range from single neurons to net- early diagnosis purposes. This has been benefited a
works. It has also been possible to bridge between lot by the development of computational methods
the dynamics of individual-based models evolving on that facilitate a deeper mechanistic understanding of
a micro level and the emergent dynamics on a macro realistic complex biophysical models. This has been
level. This led to a step change towards viewing the possible due to advances in computational power
brain as a complex self-organizing network that con- that allow the simulation and state-of-the-art analysis
stantly modifies its architecture to achieve its sophis- of very-large-scale complex models. In particular,
ticated functionality. modern methods that can handle nonlinearity and
Advanced statistical methods allow utilizing complexity such as data-driven nonlinear time-series
data in order to perform tasks such as linear and analysis, machine learning, ISOMAP and Diffusion
nonlinear system identification. Methods from non- Maps and multiscale Equation-Free techniques facili-
linear dynamics help us to clarify and understand the tate the modeling and analysis of more and more
surprising and counter intuitive nonlinearity and realistic biophysical systems whose complexity posed
complexity of systems in neuroscience. This includes in the past obstacles to our ability to analyze them in
the analysis and understanding of parameter- a systematic and efficient way.
dependent changes of the qualitative behavior as There is no doubt that we will see more fruitful
observed on different scales, the reconstruction of results from multiscale modeling and corresponding
causality functional networks that govern the infor- mathematical tools adding essential/indispensable
mation flow of the brain signals, the identification of insight in combination with experimental findings in
distinct spatiotemporal patterns that can serve as the future.

REFERENCES
1. Klausberger T, Somogyi P. Neuronal diversity and 8. Koch C, Laurent G. Complexity and the nervous sys-
temporal dynamics: the unity of hippocampal circuit tem. Science 1999, 284:96–98.
operations. Science 2008, 321:53–57. 9. Destexhe A, Contreras D. Neuronal computations
2. Bartol Jr TM, Bromer C, Kinney J, Chirillo MA, with stochastic network states. Science 2006, 314:
Bourne JN, Harris KM, Sejnowski TJ. Nanoconnec- 85–90.
tomic upper bound on the variability of synaptic plas- 10. Stam CJ. Nonlinear dynamical analysis of EEG and
ticity. eLife 2015, 4. doi:10.7554/eLife.10778. MEG: review of an emerging field. Clin Neurophysiol
3. Coyle JT. NMDA receptor and schizophrenia: a brief Off J Int Fed Clin Neurophysiol 2005, 116:
history. Schizophr Bull 2012, 38:920–926. 2266–2301.

4. Walczak A, Szczepankiewicz AA, Ruszczycki B, 11. Gerstner W, Sprekeler H, Deco G. Theory and simu-
Magalska A, Zamlynska K, Dzwonek J, Wilczek E, lation in neuroscience. Science 2012, 338:60–65.
Zybura-Broda K, Rylski M, Malinowska M, 12. van Ooyen A. Using theoretical models to analyse
et al. Novel higher-order epigenetic regulation of the neural development. Nat Rev Neurosci 2011,
Bdnf gene upon seizures. J Neurosci 2013, 33: 12:311–326.
2507–2511.
13. Xia M, Wang J, He Y. BrainNet Viewer: a network
5. Hawrylycz MJ, Lein ES, Guillozet-Bongaarts AL, visualization tool for human brain connectomics.
Shen EH, Ng L, Miller JA, van de Lagemaat LN, PLoS One 2013, 8:e68910.
Smith KA, Ebbert A, Riley ZL, et al. An anatomically
14. Abbott LF. Lapicque’s introduction of the integrate-
comprehensive atlas of the adult human brain tran-
and-fire model neuron (1907). Brain Res Bull 1999,
scriptome. Nature 2012, 489:391–399.
50:303–304.
6. Robinson GE. Brains work via their genes just as 15. McCulloch WS, Pitts W. A logical calculus of the
much as their neurons. Conversation October 14, ideas immanent in nervous activity. Bull Math Bio-
2015. https://theconversation.com/brains-work-via- phys 1943, 5:115–133.
their-genes-just-as-much-as-their-neurons-47522.
16. Hodgkin AL, Huxley AF. A quantitative description
7. Fornito A, Zalesky A, Breakspear M. The connec- of membrane current and its application to conduc-
tomics of brain disorders. Nat Rev Neurosci 2015, tion and excitation in nerve. J Physiol 1952, 117:
16:159–172. 500–544.

© 2016 Wiley Periodicals, Inc.

Advanced Review wires.wiley.com/sysbio

17. Rall W. Electrophysiology of a dendritic neuron 35. Hebb DO. The Organization of Behavior: A Neuro-
model. Biophys J 1962, 2:145–167. psychological Theory. New ed. Mahwah, NJ: Psy-
18. FitzHugh R. Mathematical models of threshold phe- chology Press; 2002.
nomena in the nerve membrane. Bull Math Biophys 36. Granger CWJ. Investigating causal relations by econ-
1955, 17:257–278. ometric models and cross-spectral methods. Econo-
metrica 1969, 37:424–438.
19. Nagumo J, Arimoto S, Yoshizawa S. An active pulse
transmission line simulating nerve axon. Proc IRE 37. Hockett CF. Review of the mathematical theory of
1962, 50:2061–2070. communication. Language 1953, 29:69–93.
38. Borst A, Theunissen FE. Information theory and neu-
20. Wiener N, Rosenblueth A. The mathematical formu-
ral coding. Nat Neurosci 1999, 2:947–957.
lation of the problem of conduction of impulses in a
network of connected excitable elements, specifically 39. Schreiber T. Measuring information transfer. Phys
in cardiac muscle. Arch Inst Cardiol Mex 1946, Rev Lett 2000, 85:461–464.
16:205–265. 40. Janjarasjitt S, Loparo KA. An approach for character-
izing coupling in dynamical systems. Phys Nonlinear
21. Greenberg J, Hastings S. Spatial patterns for discrete
Phenom 2008, 237:2482–2486.
models of diffusion in excitable media. SIAM J Appl
Math 1978, 34:515–523. 41. Vlachos I, Kugiumtzis D. Nonuniform state-space
reconstruction and coupling detection. Phys Rev E
22. Smith DR, Davidson CH. Maintained activity in neu- 2010, 82:016207.
ral nets. J ACM 1962, 9:268–279.
42. Rabinovich MI, Varona P, Selverston AI,
23. Anninos PA, Beek B, Csermely TJ, Harth EM, Abarbanel HDI. Dynamical principles in neurosci-
Pertile G. Dynamics of neural structures. J Theor Biol ence. Rev Mod Phys 2006, 78:1213–1265.
1970, 26:121–148. 43. Izhikevich EM. Dynamical Systems in Neuroscience.
24. Wilson HR, Cowan JD. Excitatory and inhibitory Cambridge, MA: MIT Press; 2007.
interactions in localized populations of model neu- 44. Marr D. A theory for cerebral neocortex. Proc R Soc
rons. Biophys J 1972, 12:1–24. Lond B Biol Sci 1970, 176:161–234.
25. Lorenz EN. Deterministic nonperiodic flow. J Atmos 45. Grossberg S. Towards a unified theory of neocortex:
Sci 1963, 20:130–141. laminar cortical circuits for vision and cognition.
26. Lorenz EN. Atmospheric predictability as revealed by Prog Brain Res 2007, 165:79–104.
naturally occurring analogues. J Atmos Sci 1969, 46. Wilson HR, Cowan JD. A mathematical theory of the
26:636–646. functional dynamics of cortical and thalamic nervous
tissue. Kybernetik 1973, 13:55–80.
27. Rössler OE. An equation for continuous chaos. Phys
Lett A 1976, 57:397–398. 47. Elijah DH, Samengo I, Montemurro MA. Thalamic
neuron models encode stimulus information by burst-
28. Rössler OE. An equation for hyperchaos. Phys Lett A size modulation. Front Comput Neurosci
1979, 71:155–157. 2015, 9:113.
29. Hindmarsh JL, Rose RM. A model of neuronal burst- 48. Safaai H, Neves R, Eschenko O, Logothetis NK,
ing using three coupled first order differential equa- Panzeri S. Modeling the effect of locus coeruleus fir-
tions. Proc R Soc Lond B Biol Sci 1984, 221:87–102. ing on cortical state dynamics and single-trial sensory
30. Korn H, Faure P. Is there chaos in the brain? processing. Proc Natl Acad Sci USA 2015,
II. Experimental evidence and related models. C R 112:12834–12839.
Biol 2003, 326:787–840. 49. Traub RD, Wong RK, Miles R, et al. A model of a
CA3 hippocampal pyramidal neuron incorporating
31. Izhikevich EM. Which model to use for cortical spik-
voltage-clamp data on intrinsic conductances.
ing neurons? IEEE Trans Neural Netw 2004,
J Neurophysiol 1991, 66:635–650.
15:1063–1070.
50. Yamaguti Y, Kuroda S, Fukushima Y, et al. A mathe-
32. Herz AVM, Gollisch T, Machens CK, et al. Modeling matical model for Cantor coding in the hippocampus.
single-neuron dynamics and computations: a balance Neural Netw Off J Int Neural Netw Soc 2011,
of detail and abstraction. Science 2006, 314:80–85. 24:43–53.
33. Deco G, Jirsa VK, Robinson PA, Breakspear M, 51. Ermentrout GB, Cowan JD. A mathematical theory
Friston K, et al. The dynamic brain: from spiking of visual hallucination patterns. Biol Cybern 1979,
neurons to neural masses and cortical fields. PLoS 34:137–150.
Comput Biol 2008, 4:e1000092. 52. Bressloff PC, Cowan JD, Golubitsky M, Thomas PJ,
34. Jirsa VK. Connectivity and dynamics of neural infor- Wiener MC. What geometric visual hallucinations tell
mation processing. Neuroinformatics 2004, 2: us about the visual cortex. Neural Comput 2002,
183–204. 14:473–491.

© 2016 Wiley Periodicals, Inc.

 WIREs System Biology and Medicine Multiscale modeling of brain dynamics

53. Haken H, Kelso JAS, Bunz H. A theoretical model of 71. Lago-Fernández LF, Huerta R, Corbacho F,
phase transitions in human hand movements. Biol Sigüenza JA. Fast response and temporal coherent
Cybern 1985, 51:347–356. oscillations in small-world networks. Phys Rev Lett
54. Kelso JAS. Instabilities and phase transitions in 2000, 84:2758–2761.
human brain and behavior. Front Hum Neurosci 72. Terman D, Ahn S, Wang X, Just W. Reducing neu-
2010, 4. doi:10.3389/fnhum.2010.00023. ronal networks to discrete dynamics. Phys Nonlinear
55. Jirsa VK, Friedrich R, Haken H, Kelso JA. A theoreti- Phenom 2008, 237:324–338.
cal model of phase transitions in the human brain. 73. Ashwin P, Coombes S, Nicks R. Mathematical frame-
Biol Cybern 1994, 71:27–35. works for oscillatory network dynamics in neurosci-
56. Durstewitz D, Seamans JK, Sejnowski TJ. Neurocom- ence. J Math Neurosci 2016, 6:2.
putational models of working memory. Nat Neurosci
74. Fine AS, Nicholls DP, Mogul DJ. Assessing instanta-
2000, 3:1184–1191.
neous synchrony of nonlinear nonstationary oscilla-
57. Laing CR, Troy WC, Gutkin B, Ermentrout GB. Mul- tors in the brain. J Neurosci Methods 2010,
tiple bumps in a neuronal model of working memory. 186:42–51.
SIAM J Appl Math 2002, 63:62–97.
75. Hadjipapas A, Casagrande E, Nevado A, Barnes GR,
58. Durstewitz D, Seamans JK. Beyond bistability: bio- Green G, Holliday IE. Can we observe collective neu-
physics and temporal dynamics of working memory. ronal activity from macroscopic aggregate signals?
Neuroscience 2006, 139:119–133. Neuroimage 2009, 44:1290–1303.
59. Cohen JD, Braver TS, O’Reilly RC. A computational
76. Barrio R, Shilnikov A. Parameter-sweeping techniques
approach to prefrontal cortex, cognitive control and
for temporal dynamics of neuronal systems: case
schizophrenia: recent developments and current chal-
study of Hindmarsh-Rose model. J Math Neurosci
lenges. Philos Trans R Soc Lond B Biol Sci 1996,
2011, 1:6.
351:1515–1527.
60. Hoffman RE, McGlashan TH. Synaptic elimination, 77. Lainscsek C, Weyhenmeyer J, Hernandez ME,
neurodevelopment, and the mechanism of halluci- Poizner H, Sejnowski TJ. Non-linear dynamical clas-
nated “voices” in schizophrenia. Am J Psychiatry sification of short time series of the Rössler system in
1997, 154:1683–1689. high noise regimes. Front Neurol 2013, 4.
doi:10.3389/fneur.2013.00182.
61. Rolls ET, Deco G. A computational neuroscience
approach to schizophrenia and its onset. Neurosci 78. Selverston AI, Rabinovich MI, Abarbanel HD,
Biobehav Rev 2011, 35:1644–1653. Elson R, Szücs A, Pinto RD, Huerta R,
Varona P. Reliable circuits from irregular neurons: a
62. Babloyantz A, Destexhe A. Low-dimensional chaos in
dynamical approach to understanding central pattern
an instance of epilepsy. Proc Natl Acad Sci 1986,
generators. J Physiol Paris 2000, 94:357–374.
83:3513–3517.
63. Netoff TI, Clewley R, Arno S, Keck T, White JA. Epi- 79. Sungar N, Allaria E, Leyva I, Arecchi FT. Compari-
lepsy in small-world networks. J Neurosci 2004, son of single neuron models in terms of synchroniza-
24:8075–8083. tion propensity. Chaos Woodbury N 2008,
18:033108.
64. Rubin JE, Terman D. High frequency stimulation of
the subthalamic nucleus eliminates pathological tha- 80. Treves A. Mean-field analysis of neuronal spike
lamic rhythmicity in a computational model. dynamics. Netw Comput Neural Syst 1993, 4:
J Comput Neurosci 2004, 16:211–235. 259–284.
65. Strogatz SH. Exploring complex networks. Nature 81. Bressloff PC, Coombes S, de Souza B. Dynamics of a
2001, 410:268–276. ring of pulse-coupled oscillators: group-theoretic
66. Morris C, Lecar H. Voltage oscillations in the barna- approach. Phys Rev Lett 1997, 79:2791–2794.
cle giant muscle fiber. Biophys J 1981, 35:193–213. 82. Haskell E, Nykamp DQ, Tranchina D. Population
67. Izhikevich EM. Simple model of spiking neurons. density methods for large-scale modelling of neuronal
Trans Neur Netw 2003, 14:1569–1572. networks with realistic synaptic kinetics: cutting the
68. Boerlin M, Machens CK, Denève S. Predictive coding dimension down to size. Netw Bristol Engl 2001,
of dynamical variables in balanced spiking networks. 12:141–174.
PLoS Comput Biol 2013, 9:e1003258. 83. Casti ARR, Omurtag A, Sornborger A, Kaplan E,
69. Saparov A, Schwemmer MA. Effects of passive den- Knight B, Victor J, Sirovich L. A population study of
dritic tree properties on the firing dynamics of a integrate-and-fire-or-burst neurons. Neural Comput
leaky-integrate-and-fire neuron. Math Biosci 2015, 2002, 14:957–986.
269:61–75. 84. Cai D, Tao L, Shelley M, McLaughlin DW. An effec-
70. Torres JJ, Marro J. Brain performance versus phase tive kinetic representation of fluctuation-driven neu-
transitions. Sci Rep 2015, 5:12216. ronal networks with application to simple and

© 2016 Wiley Periodicals, Inc.

Advanced Review wires.wiley.com/sysbio

complex cells in visual cortex. Proc Natl Acad Sci neural relaxation oscillator. Phys Rev E Stat Phys
USA 2004, 101:7757–7762. Plasmas Fluids Relat Interdiscip Top 2000,
85. Sanz-Leon P, Knock SA, Spiegler A, Jirsa VK. Mathe- 62:4057–4066.
matical framework for large-scale brain network 100. Usher M, Schuster HG, Niebur E. Dynamics of popu-
modeling in The Virtual Brain. Neuroimage 2015, lations of integrate-and-fire neurons, partial synchro-
111:385–430. nization and memory. Neural Comput 1993,
86. Cabral J, Hugues E, Sporns O, Deco G. Role of local 5:570–586.
network oscillations in resting-state functional con- 101. Brunel N, Hakim V. Fast global oscillations in net-
nectivity. Neuroimage 2011, 57:130–139. works of integrate-and-fire neurons with low firing
87. Váša F, Shanahan M, Hellyer PJ, Scott G, Cabral J, rates. Neural Comput 1999, 11:1621–1671.
Leech R. Effects of lesions on synchrony and metasta- 102. Bressloff PC, Coombes S. Travelling waves in chains
bility in cortical networks. Neuroimage 2015, of pulse-coupled integrate-and-fire oscillators with
118:456–467. distributed delays. Phys Nonlinear Phenom 1999,
88. Acebrón JA, Bonilla LL, Pérez Vicente CJ, Ritort F, 130:232–254.
Spigler R. The Kuramoto model: a simple paradigm 103. Omurtag A, Knight BW, Sirovich L. On the simula-
for synchronization phenomena. Rev Mod Phys tion of large populations of neurons. J Comput Neu-
2005, 77:137–185. rosci 2000, 8:51–63.
89. Kuramoto Y. Collective synchronization of pulse- 104. Ermentrout GB, Chow CC. Modeling neural oscilla-
coupled oscillators and excitable units. Phys Nonlin- tions. Physiol Behav 2002, 77:629–633.
ear Phenom 1991, 50:15–30. 105. Wolfram S. Universality and complexity in cellular
90. Kuramoto Y, Nishikawa I. Statistical macrodynamics automata. Phys Nonlinear Phenom 1984, 10:1–35.
of large dynamical systems. Case of a phase transition 106. Kozma R, Puljic M, Balister P, Bollobás B,
in oscillator communities. J Stat Phys 1987, Freeman WJ. Phase transitions in the neuropercola-
49:569–605. tion model of neural populations with mixed local
91. Strogatz SH. From Kuramoto to Crawford: exploring and non-local interactions. Biol Cybern 2005,
the onset of synchronization in populations of 92:367–379.
coupled oscillators. Phys Nonlinear Phenom 2000, 107. Furtado LS, Copelli M. Response of electrically
143:1–20. coupled spiking neurons: a cellular automaton
92. Wilson HR. Spikes, Decisions, and Actions: The approach. Phys Rev E 2006, 73:011907.
Dynamical Foundations of Neuroscience. Oxford and 108. Deco G, Jirsa VK, Robinson PA, Breakspear M,
New York: Oxford University Press; 1999. Friston K. The dynamic brain: from spiking neurons
93. Negahbani E, Steyn-Ross DA, Steyn-Ross ML, to neural masses and cortical fields. PLoS Comput
Wilson MT, Sleigh JW. Noise-induced precursors of Biol 2008, 4:e1000092.
state transitions in the stochastic Wilson-Cowan 109. Borisyuk GN, Borisyuk RM, Kazanovich YB,
model. J Math Neurosci 2015, 5:1–27. Ivanitskiib GR. Models of neural dynamics in brain
94. Kilpatrick ZP. Wilson-Cowan model. In: Jaeger D, information processing — the developments of “the
Jung R, eds. Encyclopedia of Computational Neuro- decade”. Phys-Uspekhi 2002, 45:1073–1095.
science. New York: Springer; 2014. 110. Doedel EJ, Govaerts W, Kuznetsov YA,
95. Ermentrout GB. Stable periodic solutions to discrete Dhooge A. Numerical continuation of branch points
and continuum arrays of weakly coupled nonlinear of equilibria and periodic orbits. Int J Bifurc Chaos
oscillators. SIAM J Appl Math 1992, 52:1665–1687. 2005, 15:841–860.
96. Abbott LF, van Vreeswijk C. Asynchronous states in 111. Dhooge A, Govaerts W, Kuznetsov YA. MATCONT:
networks of pulse-coupled oscillators. Phys Rev E a MATLAB package for numerical bifurcation analy-
Stat Phys Plasmas Fluids Relat Interdiscip Top 1993, sis of ODEs. ACM Trans Math Softw 2003,
48:1483–1490. 29:141–164.
97. Gutkin BS, Laing CR, Colby CL, Chow CC, 112. Dankowicz H, Schilder F. Recipes for Continuation.
Ermentrout GB. Turning on and off with excitation: Philadelphia, USA: SIAM Bookst; 2013.
the role of spike-timing asynchrony and synchrony in 113. Rinzel J, Ermentrout GB. Analysis of neural excitabil-
sustained neural activity. J Comput Neurosci 2001, ity and oscillations. In: Koch C, Segev I, eds. Methods
11:121–134. in Neuronal Modeling. Cambridge, MA, USA: MIT
98. Laing CR, Chow CC. A spiking neuron model for Press; 1989, 135–169.
binocular rivalry. J Comput Neurosci 2002, 114. Rinzel J, Lee YS. Dissection of a model for neuronal
12:39–53. parabolic bursting. J Math Biol 1987, 25:653–675.
99. Coombes S, Osbaldestin AH. Period-adding bifurca- 115. Reidl J, Borowski P, Sensse A, Starke J, Zapotocky
tions and chaos in a periodically stimulated excitable M, Eiswirth M. Model of calcium oscillations due to

© 2016 Wiley Periodicals, Inc.

 WIREs System Biology and Medicine Multiscale modeling of brain dynamics

negative feedback in olfactory cilia. Biophys J 2006, 129. Moon SJ, Ghanem R, Kevrekidis IG. Coarse graining
90:1147–1155. the dynamics of coupled oscillators. Phys Rev Lett
116. Nowotny T, Rabinovich MI. Dynamical origin of 2006, 96:144101.
independent spiking and bursting activity in neural 130. Laing CR, Kevrekidis IG. Periodically-forced finite
microcircuits. Phys Rev Lett 2007, 98:128106. networks of heterogeneous globally-coupled oscilla-
117. Ly C, Tranchina D. Critical analysis of dimension tors: a low-dimensional approach. Phys Nonlinear
reduction by a moment closure method in a popula- Phenom 2008, 237:207–215.
tion density approach to neural network modeling. 131. Spiliotis KG, Siettos CI. Multiscale computations on
Neural Comput 2007, 19:2032–2092. neural networks: from the individual neuron interac-
tions to the macroscopic-level analysis. Int J Bifurc
118. Theodoropoulos C, Qian Y-H, Kevrekidis IG.
Chaos 2010, 20:121–134.
“Coarse” stability and bifurcation analysis using
time-steppers: a reaction–diffusion example. Proc 132. Spiliotis KG, Siettos CI. A timestepper-based
Natl Acad Sci USA 2000, 97:9840–9843. approach for the coarse-grained analysis of micro-
scopic neuronal simulators on networks: bifurcation
119. Kevrekidis IG, Gear CW, Hyman JM, Kevrekidis PG,
and rare-events micro- to macro-computations. Neu-
Runborg O, Theodoropoulos C. Equation-free,
rocomputing 2011, 74:3576–3589.
coarse-grained multiscale computation: enabling
mocroscopic simulators to perform system-level anal- 133. Papo D, Buldú JM, Boccaletti S, Bullmore ET. Com-
ysis. Commun Math Sci 2003, 1:715–762. plex network theory and the brain. Phil Trans R Soc
B 2014, 369:20130520.
120. Kevrekidis IG, Gear CW, Hummer G. Equation-free:
the computer-aided analysis of complex multiscale 134. Sporns O, Tononi G, Edelman GM. Connectivity and
systems. AIChE J 2004, 50:1346–1355. complexity: the relationship between neuroanatomy
and brain dynamics. Neural Netw Off J Int Neural
121. Coifman RR, I. G. Kevrekidis, S. Lafon, M.
Netw Soc 2000, 13:909–922.
Maggioni, and B. Nadler. Diffusion maps, reduction
coordinates, and low dimensional representation of 135. Brazier MA. Spread of seizure discharges in epilepsy:
stochastic systems. Multiscale Model Simul 2008, anatomical and electrophysiological considerations.
7:842–864. Exp Neurol 1972, 36:263–272.
122. Coifman RR, Lafon S, Lee AB, Maggioni M, Nadler 136. Gerstein GL, Perkel DH, Subramanian KN. Identifi-
B, Warner F, Zucker SW. Geometric diffusions as a cation of functionally related neural assemblies. Brain
tool for harmonic analysis and structure definition of Res 1978, 140:43–62.
data: diffusion maps. Proc Natl Acad Sci USA 2005, 137. Zalesky A, Fornito A, Bullmore E. On the use of cor-
102:7426–7431. relation as a measure of network connectivity. Neuro-
123. Coifman RR, Lafon S, Lee AB, Maggioni M, Nadler image 2012, 60:2096–2106.
B, Warner F, Zucker SW. Geometric diffusions as a 138. Siggiridou E, Kugiumtzis D, Kimiskidis VK. Correla-
tool for harmonic analysis and structure definition of tion networks for identifying changes in brain con-
data: multiscale methods. Proc Natl Acad Sci USA nectivity during epileptiform discharges and
2005, 102:7432–7437. transcranial magnetic stimulation. Sensors 2014,
124. Moon SJ, Cook KA, Rajendran K, Kevrekidis IG, 14:12585–12597.
Cisternas J, Laing CR. Coarse-grained clustering 139. Seth AK, Barrett AB, Barnett L. Granger causality
dynamics of heterogeneously coupled neurons. analysis in neuroscience and neuroimaging.
J Math Neurosci JMN 2015, 5:1–20. J Neurosci Off J Soc Neurosci 2015, 35:3293–3297.
125. Laing CR, Frewen TA, Kevrekidis IG. Coarse-grained 140. Seth AK. A MATLAB toolbox for Granger causal
dynamics of an activity bump in a neural field model. connectivity analysis. J Neurosci Methods 2010,
Nonlinearity 2007, 20:2127. 186:262–273.
126. Wasylenko TM, Cisternas JE, Laing CR, 141. Geweke J. Measurement of linear dependence and
Kevrekidis IG. Bifurcations of lurching waves in a feedback between multiple time series. J Am Stat
thalamic neuronal network. Biol Cybern 2010, Assoc 1982, 77:304–313.
103:447–462. 142. Lachaux JP, Rodriguez E, Martinerie J,
127. Marschler C, Faust-Ellsässer C, Starke J, van et al. Measuring phase synchrony in brain signals.
Hemmen JL. Bifurcation of learning and structure Hum Brain Mapp 1999, 8:194–208.
formation in neuronal maps. EPL Europhys Lett 143. Stam CJ, Nolte G, Daffertshofer A. Phase lag index:
2014, 108:48005. assessment of functional connectivity from multi
128. Laing CR, Kevrekidis IG. Equation-free analysis of channel EEG and MEG with diminished bias from
spike-timing-dependent plasticity. Biol Cybern 2015, common sources. Hum Brain Mapp 2007,
109:701–714. 28:1178–1193.

© 2016 Wiley Periodicals, Inc.

 Advanced Review wires.wiley.com/sysbio

144. Vicente R, Wibral M, Lindner M, Pipa G. Transfer 157. Gao Q, Duan X, Chen H. Evaluation of effective con-
entropy—a model-free measure of effective connectiv- nectivity of motor areas during motor imagery and
ity for the neurosciences. J Comput Neurosci 2010, execution using conditional Granger causality. Neu-
30:45–67. roimage 2011, 54:1280–1288.
145. Calhoun VD, Adali T, Pearlson GD, Pekar 158. Liao W, Ding J, Marinazzo D, Xu Q, Wang Z, Yuan
JJ. A method for making group inferences from func- C, Zhang Z, Lu G, Chen H. Small-world directed
tional MRI data using independent component analy- networks in the human brain: multivariate Granger
sis. Hum Brain Mapp 2001, 14:140–151. causality analysis of resting-state fMRI. Neuroimage
2011, 54:2683–2694.
146. Calhoun VD, Liu J, Adali T. A review of group ICA
for fMRI data and ICA for joint inference of imaging, 159. Miao X, Wu X, Li R, Chen K, Yao L. Altered con-
genetic, and ERP data. Neuroimage 2009, 45: nectivity pattern of hubs in default-mode network
S163–S172. with Alzheimer’s disease: an Granger causality model-
ing approach. PLoS One 2011, 6:e25546.
147. Reidl J, Starke J, Omer DB, et al. Independent com-
160. Roebroeck A, Formisano E, Goebel R. Mapping
ponent analysis of high-resolution imaging data iden-
directed influence over the brain using Granger cau-
tifies distinct functional domains. Neuroimage 2007,
sality and fMRI. Neuroimage 2005, 25:230–242.
34:94–108.
161. Sabatinelli D, McTeague LM, Dhamala M,
148. Anderson A, Cohen MS. Decreased small-world func- Frank DW, Wanger TJ, Adhikari BM. Reduced
tional network connectivity and clustering across rest- medial prefrontal-subcortical connectivity in dyspho-
ing state networks in schizophrenia: an fMRI ria: Granger causality analyses of rapid functional
classification tutorial. Front Hum Neurosci magnetic resonance imaging. Brain Connect
2013, 7:520. 2015, 5:1–9.
149. Tenenbaum JB, de Silva V, Langford JC. A global 162. Tang W, Bressler SL, Sylvester CM, Shulman GL,
geometric framework for nonlinear dimensionality Corbetta M. Measuring Granger causality between
reduction. Science 2000, 290:2319–2323. cortical regions from voxelwise fMRI BOLD signals
150. Duncan D, Talmon R, Zaveri HP, Coifman RR. Identi- with LASSO. PLoS Comput Biol 2012, 8:e1002513.
fying preseizure state in intracranial EEG data using 163. Wu GR, Liao W, Stramaglia S, Chen H,
diffusion kernels. Math Biosci Eng MBE 2013, Marinazzo D. Recovering directed networks in neu-
10:579–590. roimaging datasets using partially conditioned Gran-
151. Keller CJ, Bickel S, Honey CJ, Groppe DM, Entz L, ger causality. Brain Connect 2013, 3:294–301.
Craddock RC, Lado FA, Kelly C, Milham M, 164. Zhou Z, Wang X, Klahr NJ, Liu W, Arias D, Liu H,
Mehta AD. Neurophysiological investigation of spon- von Deneen KM, Wen Y, Lu Z, Xu D, et al. A condi-
taneous correlated and anticorrelated fluctuations of tional Granger causality model approach for group
the BOLD signal. J Neurosci Off J Soc Neurosci analysis in functional magnetic resonance imaging.
2013, 33:6333–6342. Magn Reson Imaging 2011, 29:418–433.
152. Megumi F, Yamashita A, Kawato M, Imamizu H, 165. Barrett AB, Murphy M, Bruno MA, Noirhomme Q,
et al. Functional MRI neurofeedback training on con- Boly M, Laureys S, Seth AK. Granger causality analy-
nectivity between two regions induces long-lasting sis of steady-state electroencephalographic signals
changes in intrinsic functional network. Front Hum during propofol-induced anaesthesia. PLoS One
Neurosci 2015, 9:160. 2012, 7:e29072.
166. Barnett L, Seth AK. Behaviour of Granger causality
153. Li X, Wang H. Identification of functional networks
under filtering: theoretical invariance and practical
in resting state fMRI data using adaptive sparse rep-
application. J Neurosci Methods 2011, 201:404–419.
resentation and affinity propagation clustering. Front
Neurosci 2015, 9. doi:10.3389/fnins.2015.00383. 167. Gow DW Jr, Segawa JA, Ahlfors SP, Lin FH. Lexical
influences on speech perception: a Granger causality
154. Shaw JC. Correlation and coherence analysis of the analysis of MEG and EEG source estimates. Neuro-
EEG: a selective tutorial review. Int J Psychophysiol image 2008, 43:614–623.
Off J Int Organ Psychophysiol 1984, 1:255–266.
168. Keil A, Sabatinelli D, Ding M, Lang PJ, Ihssen N,
155. Al-Aidroos N, Said CP, Turk-Browne NB. Top-down Heim S. Re-entrant projections modulate visual cor-
attention switches coupling between low-level and tex in affective perception: evidence from Granger
high-level areas of human visual cortex. Proc Natl causality analysis. Hum Brain Mapp 2009,
Acad Sci USA 2012, 109:14675–14680. 30:532–540.
156. Florin E, Gross J, Pfeifer J, Fink GR, Timmermann 169. Nicolaou N, Hourris S, Alexandrou P, Georgiou
L. The effect of filtering on Granger causality based J. EEG-based automatic classification of “awake”
multivariate causality measures. Neuroimage 2010, versus “anesthetized” state in general anesthesia using
50:577–588. Granger causality. PLoS One 2012, 7:e33869.

© 2016 Wiley Periodicals, Inc.

 WIREs System Biology and Medicine Multiscale modeling of brain dynamics

170. de Tommaso M, Stramaglia S, Marinazzo D, 183. Kitzbichler MG, Henson RNA, Smith ML, Nathan
Trotta G, Pellicoro M. Functional and effective con- PJ, Bullmore ET. Cognitive effort drives workspace
nectivity in EEG alpha and beta bands during inter- configuration of human brain functional networks.
mittent flash stimulation in migraine with and J Neurosci Off J Soc Neurosci 2011, 31:8259–8270.
without aura. Cephalalgia Int J Headache 2013, 184. Melloni L, Molina C, Pena M, Torres D, Singer W,
33:938–947. Rodriguez E. Synchronization of neural activity
171. Protopapa F, Siettos CI, Evdokimidis I, Smyrnis N. across cortical areas correlates with conscious percep-
Granger causality analysis reveals distinct spatio- tion. J Neurosci Off J Soc Neurosci 2007,
temporal connectivity patterns in motor and percep- 27:2858–2865.
tual visuo-spatial working memory. Front Comput 185. Gaillard R, Dehaene S, Adam C, Clémenceau S,
Neurosci 2014, 8:146. Hasboun D, Baulac M, Cohen L, Naccache L. Con-
172. Protopapa F, Siettos CI, Myatchin I, Lagae L. Chil- verging intracranial markers of conscious access.
dren with well controlled epilepsy possess different PLoS Biol 2009, 7:e61.
spatio-temporal patterns of causal network connectiv-
186. Pockett S, Bold GEJ, Freeman WJ. EEG synchrony
ity during a visual working memory task. Cogn Neu-
during a perceptual-cognitive task: widespread phase
rodyn 2016, 10:1–13.
synchrony at all frequencies. Clin Neurophysiol Off J
173. Kaminski M, Ding M, Truccolo WA, Int Fed Clin Neurophysiol 2009, 120:695–708.
Bressler SL. Evaluating causal relations in neural sys-
187. Gross J, Schmitz F, Schnitzler I, Kessler K, Shapiro K,
tems: Granger causality, directed transfer function
Hommel B, Schnitzler A. Modulation of long-range
and statistical assessment of significance. Biol Cybern
neural synchrony reflects temporal limitations of vis-
2001, 85:145–157.
ual attention in humans. Proc Natl Acad Sci USA
174. Baccalá LA, Sameshima K. Partial directed coherence: 2004, 101:13050–13055.
a new concept in neural structure determination. Biol
Cybern 2001, 84:463–474. 188. Rodriguez E, George N, Lachaux JP, Martinerie J,
Renault B, Varela FJ. Perception’s shadow: long-
175. Korzeniewska A, Ma nczak M, Kami nski M,
distance synchronization of human brain activity.
Blinowska KJ, Kasicki S. Determination of informa-
Nature 1999, 397:430–433.
tion flow direction among brain structures by a modi-
fied directed transfer function (dDTF) method. 189. Mylonas DS, Siettos CI, Evdokimidis I,
J Neurosci Methods 2003, 125:195–207. Papanicolaou AC, Smyrnis N. Modular patterns of
phase desynchronization networks during a simple
176. Winterhalder M, Schelter B, Hesse W, Schwab K,
visuomotor task. Brain Topogr 2016, 29:118–129.
Leistritz L, Klan D, Bauer R, Timmer J,
Witte H. Comparison of linear signal processing tech- 190. Peraza LR, Asghar AU, Green G, Halliday DM. Vol-
niques to infer directed interactions in multivariate ume conduction effects in brain network inference
neural systems. Signal Process 2005, 85:2137–2160. from electroencephalographic recordings using phase
lag index. J Neurosci Methods 2012, 207:189–199.
177. Freiwald WA, Valdes P, Bosch J, Biscay R,
Jimenez JC, Rodriguez LM, Rodriguez V, Kreiter AK, 191. Mormann F, Lehnertz K, David P, Elger CE. Mean
Singer W. Testing non-linearity and directedness of phase coherence as a measure for phase synchroniza-
interactions between neural groups in the macaque tion and its application to the EEG of epilepsy
inferotemporal cortex. J Neurosci Methods 1999, patients. Phys Nonlinear Phenom 2000, 144:
94:105–119. 358–369.
178. Marinazzo D, Liao W, Chen H, Stramaglia S. 192. Nolte G, Ziehe A, Nikulin VV, et al. Robustly esti-
Nonlinear connectivity by Granger causality. Neuro- mating the flow direction of information in complex
image 2011, 58:330–338. physical systems. Phys Rev Lett 2008, 100:234101.
179. Marinazzo D, Pellicoro M, Stramaglia S. Nonlinear 193. Vinck M, Oostenveld R, van Wingerden M,
parametric model for Granger causality of time series. Battaglia F, Pennartz CM. An improved index of
Phys Rev E 2006, 73:066216. phase-synchronization for electrophysiological data
180. Trujillo LT, Peterson MA, Kaszniak AW, Allen JJ. in the presence of volume-conduction, noise and
EEG phase synchrony differences across visual per- sample-size bias. Neuroimage 2011, 55:1548–1565.
ception conditions may depend on recording and 194. Iasemidis LD, Sackellares JC, Zaveri HP,
analysis methods. Clin Neurophysiol Off J Int Fed Williams WJ. Phase space topography and the Lyapu-
Clin Neurophysiol 2005, 116:172–189. nov exponent of electrocorticograms in partial sei-
181. Lopes da Silva FH. Event-related neural activities: zures. Brain Topogr 1990, 2:187–201.
what about phase? Prog Brain Res 2006, 159:3–17. 195. Iasemidis L, Sabesan S, Chakravarthy N, Prasad A,
182. Schack B, Weiss S. Quantification of phase synchroni- Tsakalis K. Brain dynamics and modeling in epilepsy:
zation phenomena and their importance for verbal prediction and control studies. In: Dana SK, Roy PK,
memory processes. Biol Cybern 2005, 92:275–287. Kurths J, eds. Complex Dynamics in Physiological

© 2016 Wiley Periodicals, Inc.

 Advanced Review wires.wiley.com/sysbio

Systems: From Heart to Brain. The Netherlands: 210. Hyvärinen A, Oja E. Independent component analy-
Springer; 2009, 185–214. sis: algorithms and applications. Neural Netw Off J
196. Senthilkumar DV, Lakshmanan M, Kurths J. Transi- Int Neural Netw Soc 2000, 13:411–430.
tion from phase to generalized synchronization in 211. Pruim RH, Mennes M, van Rooij D, Llera A,
time-delay systems. Chaos Interdiscip J Nonlinear Sci Buitelaar JK, Beckmann CF. ICA-AROMA: a robust
2008, 18:023118. ICA-based strategy for removing motion artifacts
197. Pikovsky A, Rosenblum M, Kurths J. Phase synchro- from fMRI data. Neuroimage 2015, 112:267–277.
nization in regular and chaotic systems. Int J Bifurc 212. Robinson SD, Schöpf V. ICA of fMRI studies: new
Chaos 2000, 10:2291–2305. approaches and cutting edge applications. Front Hum
198. Chávez M, Martinerie J, Le Van Quyen M. Statistical Neurosci 2013, 7:724.
assessment of nonlinear causality: application to epi- 213. McKeown MJ, Jung TP, Makeig S, Brown G,
leptic EEG signals. J Neurosci Methods 2003, Kindermann SS, Lee TW, Sejnowski TJ. Spatially
124:113–128. independent activity patterns in functional MRI data
199. Ramanand P, Bruce MC, Bruce EN. Mutual informa- during the stroop color-naming task. Proc Natl Acad
tion analysis of EEG signals indicates age-related Sci USA 1998, 95:803–810.
changes in cortical interdependence during sleep in 214. McKeown MJ, Hansen LK, Sejnowsk TJ. Independ-
middle-aged versus elderly women. J Clin Neurophy- ent component analysis of functional MRI: what is
siol Off Publ Am Electroencephalogr Soc 2010, signal and what is noise? Curr Opin Neurobiol 2003,
27:274–284. 13:620–629.
200. Wan X, Crüts B, Jensen HJ. The causal inference of 215. Jamadar S, Egan G, Calhoun VD, Johnson B,
cortical neural networks during music improvisations. Fielding J. Intrinsic connectivity provides the baseline
PLoS One 2014, 9:e112776. doi:10.1371/journal. framework for variability in motor performance: a
pone.0112776. multivariate fusion analysis of low- and high-
201. Wibral M, Vicente R, Lizier JT, eds. Directed Infor- frequency resting-state oscillations and antisaccade
mation Measures in Neuroscience. Berlin and Heidel- performance. Brain Connect 2016. doi:10.1089/
berg: Springer; 2014. brain.2015.0411.
202. Takens F. Detecting strange attractors in turbulence. 216. Bär KJ, de la Cruz F, Schumann A, Koehler S, Sauer
In: Rand D, Young L-S, eds. Dynamical Systems and H, Critchley H, Wagner G. Functional connectivity
Turbulence, Warwick 1980. Berlin and Heidelberg: and network analysis of midbrain and brainstem
Springer; 1981, 366–381. nuclei. Neuroimage 2016, 134:53–63.
203. Pezard L, Lachaux JP, Thomasson N, Martinerie J. 217. Cassidy CM, Van Snellenberg JX, Benavides C,
Why bother to spatially embed EEG? Comments on Slifstein M, Wang Z, Moore H, Abi-Dargham A,
Pritchard et al., Psychophysiology 1996, 33: Horga G. Dynamic connectivity between brain net-
362–368. Psychophysiology 1999, 36:527–531. works supports working memory: relationships to
204. Sauer T, Yorke JA, Casdagli M. Embedology. J Stat dopamine release and schizophrenia. J Neurosci Off J
Phys 1991, 65:579–616. Soc Neurosci 2016, 36:4377–4388.
205. Papana A, Kyrtsou C, Kugiumtzis D, Diks C. Simula- 218. Verma R, Khurd P, Davatzikos C. On analyzing dif-
tion study of direct causality measures in multivariate fusion tensor images by identifying manifold structure
time series. Entropy 2013, 15:2635–2661. using isomaps. IEEE Trans Med Imaging 2007,
206. Friston KJ, Frith CD, Liddle PF, Frackowiak RS. 26:772–778.
Functional connectivity: the principal-component 219. Kortelainen J, Vayrynen E, Seppanen T. Isomap
analysis of large (PET) data sets. J Cereb Blood Flow approach to EEG-based assessment of neurophysio-
Metab 1993, 13:5–14. logical changes during anesthesia. IEEE Trans Neural
207. Li K, Guo L, Nie J, Li G, Liu T. Review of methods Syst Rehabil Eng Publ IEEE Eng Med Biol Soc 2011,
for functional brain connectivity detection using 19:113–120.
fMRI. Comput Med Imaging Graph Off J Comput 220. Ye AQ, Ajilore OA, Conte G, GadElkarim J,
Med Imaging Soc 2009, 33:131–139. Thomas-Ramos G, Zhan L, Yang S, Kumar A,
208. Leonardi N, Richiardi J, Gschwind M, Simioni S, Magin RL, Forbes AG, et al. The intrinsic geometry
Annoni JM, Schluep M, Vuilleumier P, Van De Ville of the human brain connectome. Brain Inform 2015,
D. Principal components of functional connectivity: a 2:197–210.
new approach to study dynamic brain connectivity 221. Richiardi J, Achard S, Bunke H, Van De Ville D.
during rest. Neuroimage 2013, 83:937–950. Machine learning with brain graphs: predictive mod-
209. Anderson A, Dinov ID, Sherin JE, Quintana J, Yuille eling approaches for functional imaging in systems
AL, Cohen MS. Classification of spatially unaligned neuroscience. IEEE Signal Process Mag 2013,
fMRI scans. Neuroimage 2010, 49:2509–2519. 30:58–70.

© 2016 Wiley Periodicals, Inc.

 WIREs System Biology and Medicine Multiscale modeling of brain dynamics

222. Becker R, Knock S, Ritter P, Jirsa V. Relating alpha Quervain DJ. Dynamic modulation of amygdala-
power and phase to population firing and hemody- hippocampal connectivity by emotional arousal.
namic activity using a thalamo-cortical neural mass J Neurosci Off J Soc Neurosci 2014, 34:
model. PLoS Comput Biol 2015, 11:e1004352. 13935–13947.
223. McCarthy MM, Moore-Kochlacs C, Gu X, Boyden 235. Guhn A, Domschke K, Müller LD, Dresler T, Eff F,
ES, Han X, Kopell N. Striatal origin of the pathologic Kopf J, Deckert J, Reif A, Herrmann MJ. Neuropep-
beta oscillations in Parkinson’s disease. Proc Natl tide S receptor gene variation differentially modulates
Acad Sci USA 2011, 108:11620–11625. fronto-limbic effective connectivity in childhood and
224. Dovzhenok A, Rubchinsky LL. On the origin of adolescence. Cereb Cortex 2015, 10:1730–7.
tremor in Parkinson’s disease. PLoS One 2012, 7: doi:10.1093/cercor/bhv259.
e41598.
236. Gilbert JR, Symmonds M, Hanna MG, Dolan RJ,
225. Sarbaz Y, Pourakbari H. A review of presented math- Friston KJ, Moran RJ. Profiling neuronal ion channe-
ematical models in Parkinson’s disease: black- and lopathies with non-invasive brain imaging and
gray-box models. Med Biol Eng Comput 2015, dynamic causal models: case studies of single gene
54:855–868. doi:10.1007/s11517-015-1401-9. mutations. Neuroimage 2016, 124:43–53.
226. Qi Z, Miller GW, Voit EO. A mathematical model
of presynaptic dopamine homeostasis: implications 237. Musgrove DR, Eberly LE, Klimes-Dougan B, Basgoze
for schizophrenia. Pharmacopsychiatry 2008, Z3, Thomas KM, Mueller BA, Houri A, Lim KO,
41(Suppl 1):S89–S98. Cullen KR. Impaired bottom-Up effective connectivity
between amygdala and subgenual anterior cingulate
227. González-Ramírez LR, Ahmed OJ, Cash SS, cortex in unmedicated adolescents with major depres-
Wayne CE, Kramer MA. A biologically constrained, sion: results from a dynamic causal modeling analy-
mathematical model of cortical wave propagation sis. Brain Connect 2015, 5:608–619.
preceding seizure termination. PLoS Comput Biol
2015, 11:e1004065. 238. Cui LB, Liu J, Wang LX, Li C, Xi YB, Guo F, Wang
228. Stefanescu RA, Shivakeshavan RG, Talathi SS. Com- HN, Zhang LC, Liu WM, He H, Tian P, Yin H, Lu
putational models of epilepsy. Seizure 2012, H. Anterior cingulate cortex-related connectivity in
21:748–759. first-episode schizophrenia: a spectral dynamic causal
modeling study with functional magnetic resonance
229. Friston KJ, Harrison L, Penny W. Dynamic causal imaging. Front Hum Neurosci 2015, 9:589.
modelling. Neuroimage 2003, 19:1273–1302.
230. Stephan KE, Penny WD, Moran RJ, den Ouden HE, 239. Herz DM, Siebner HR, Hulme OJ, Florin E,
Daunizeau J, Friston KJ. Ten simple rules for Christensen MS, Timmermann L. Levodopa reinstates
dynamic causal modeling. Neuroimage 2010, connectivity from prefrontal to premotor cortex dur-
49:3099–3109. ing externally paced movement in Parkinson’s dis-
ease. Neuroimage 2014, 90:15–23.
231. Moran R, Pinotsis DA, Friston K. Neural masses and
fields in dynamic causal modeling. Front Comput 240. Breakspear M, Roberts G, Green MJ, Nguyen VT,
Neurosci 2013, 7:57. Frankland A, Levy F, Lenroot R, Mitchell
232. Youssofzadeh V, Prasad G, Fagan AJ, Reilly RB, PB. Network dysfunction of emotional and cognitive
Martens S, Meaney JF, Wong-Lin K. Signal propaga- processes in those at genetic risk of bipolar disorder.
tion in the human visual pathways: an effective con- Brain J Neurol 2015, 138:3427–3439.
nectivity analysis. J Neurosci Off J Soc Neurosci 241. Stephan KE, Friston KJ. Analyzing effective connec-
2015, 35:13501–13510. tivity with fMRI. Wiley Interdiscip Rev Cogn Sci
233. Nguyen VT, Breakspear M, Hu X, Guo CC. The inte- 2010, 1:446–459.
gration of the internal and external milieu in the
242. Hawkins J, Ahmad S. Why Neurons Have Thousands
insula during dynamic emotional experiences. Neuro-
of Synapses, a Theory of Sequence Memory in Neo-
image 2016, 124:455–463.
cortex. Front. Neural Circuits 2016, doi:10.3389/
234. Fastenrath M, Coynel D, Spalek K, Milnik A, fncir.2016.00023.
Gschwind L, Roozendaal B, Papassotiropoulos A, de

© 2016 Wiley Periodicals, Inc.