QUICK SEARCH:   [advanced] Author: Keyword(s): Year:  Vol:  Page:

HOME  |   SEARCH  |   ARCHIVE  |   SUBSCRIBE  |   CONTACT  |   HELP

The Journal of Neuroscience, July 16, 2003, 23(15):6280-6294

This Article Abstract Full Text (PDF) A correction has been published Submit an eLetter Alert me when this article is cited Alert me when eLetters are posted Alert me if a correction is posted Citation Map Services Email this article to a friend Similar articles in this journal Similar articles in ISI Web of Science Similar articles in PubMed Alert me to new issues of the journal Download to citation manager Citing Articles Citing Articles via HighWire Citing Articles via ISI Web of Science (35) Citing Articles via Google Scholar Google Scholar Articles by Pfeuty, B. Articles by Hansel, D. Search for Related Content PubMed PubMed Citation Articles by Pfeuty, B. Articles by Hansel, D.

 Previous Article  |  Next Article 

Electrical Synapses and Synchrony: The Role of Intrinsic Currents

Benjamin Pfeuty,¹ Germán Mato,² David Golomb,^3,4 and David Hansel^1,5

¹Laboratoire de Neurophysique et Physiologie du Système Moteur, Centre National de la Recherche Scientifique–Unité Mixte de Recherche 8119, Université René Descartes, 75270 Paris Cedex 06, France, ²Comisión Nacional de Energia Atómica and Consejo Nacional de Investigaciones Cientificas y Técnicas, Centro Atómico Bariloche and Instituto Balsiero, Universidad Nacional de Cordoba, 8400 San Carlos de Bariloche, Argentina, ³Department of Physiology and Zlotowski Center for Neuroscience, Faculty of Health Sciences, Ben Gurion University of the Negev, Be'er-Sheva 84105, Israel, ⁴Mathematical Biosciences Institute, Ohio State University, Columbus, Ohio 43210, and ⁵Interdisciplinary Center for Neural Computation, The Hebrew University, Jerusalem 91904, Israel

	Abstract

Top Abstract Introduction Materials and Methods Results Discussion Appendix References

 
Electrical synapses are ubiquitous in the mammalian CNS. Particularly in the neocortex, electrical synapses have been shown to connect low-threshold spiking (LTS) as well as fast spiking (FS) interneurons. Experiments have highlighted the roles of electrical synapses in the dynamics of neuronal networks. Here we investigate theoretically how intrinsic cell properties affect the synchronization of neurons interacting by electrical synapses. Numerical simulations of a network of conductance-based neurons randomly connected with electrical synapses show that potassium currents promote synchrony, whereas the persistent sodium current impedes it. Furthermore, synchrony varies with the firing rate in qualitatively different ways depending on the intrinsic currents. We also study analytically a network of quadratic integrate-and-fire neurons. We relate the stability of the asynchronous state of this network to the phase-response function (PRF), which characterizes the effect of small perturbations on the firing timing of the neurons. In particular, we show that the greater the skew of the PRF toward the first half of the period, the more stable the asynchronous state. Combining our simulations with our analytical results, we establish general rules to predict the dynamic state of large networks of neurons coupled with electrical synapses. Our work provides a natural explanation for surprising experimental observations that blocking electrical synapses may increase the synchrony of neuronal activity. It also suggests different synchronization properties for LTS and FS cells. Finally, we propose to further test our predictions in experiments using dynamic clamp techniques.

Key words: electrical synapses; conductance-based model; synchrony; neuronal network model; intrinsic currents; neocortex

	Introduction

Top Abstract Introduction Materials and Methods Results Discussion Appendix References

 
Electrical synapses are sites at which gap-junctions bridge the membranes of two neurons. They have long been known to exist in invertebrates (Watanabe, 1958; Furshpan and Potter, 1959), but only recently has evidence of their ubiquity been unequivocally found in the mammalian brain. Electrical synapses are present in the inferior olive (Llinas and Yarom, 1986), the hippocampus (Draghun et al., 1998; Venance et al., 2000), the cerebellum (Mann-Metzer and Yarom, 1999), the locus coereleus (Christie et al., 1989; Alvarez et al., 2002), the striatum (Kita et al., 1990), the neocortex (Galarreta and Hestrin, 1999; Gibson et al., 1999), the reticular thalamic nucleus (Landisman et al., 2002), and between motoneurons (Kiehn and Tresch, 2002).

Experiments have revealed that electrical synapses are involved in synchronizing neural activity (Draghun et al., 1998; Mann-Metzer and Yarom, 1999; Beierlein et al., 2000; Perez-Velazquez and Carlen, 2000; Tamas et al., 2000; Deans et al., 2001; Hormuzdi et al., 2001). In contrast to the results of these studies, it has been reported recently that inspiratory motoneurons may display more strongly synchronized activity in presence of carbenoxolone (CBX), a blocker of electrical synapses, than in the control situation (Bou-Flores and Berger, 2001). Therefore, in this case, electrical synapses desynchronize neural activity.

The dynamics of networks of neurons interacting via chemical synapses have been studied extensively (Golomb et al., 2001). However, only a few theoretical studies have addressed the dynamics of networks in which neurons are coupled by electrical synapses. Models for pattern generation in the lobster pyloric system have been investigated (Kepler et al., 1990; Abbott et al., 1991; Meunier, 1992). Stable antiphase locking and transitions between in-phase and antiphase locking were demonstrated in pairs of model neurons coupled by electrical synapses, and its impact on rhythmogenesis was investigated (Sherman and Rinzel, 1992; Cymbalyuk et al., 1994; Han et al., 1995). More recently, the discovery of electrical synapses in the CNS of mammals has motivated simulations of conductance-based models (Traub et al., 2001) and analytical studies in the framework of leaky integrate-and-fire (LIF) models (Chow and Kopell, 2000; Lewis and Rinzel, 2003).

Although it is clear that synchronization properties of neurons depend on their intrinsic currents, the general principles that determine this dependence are still unknown in the case of neurons interacting via electrical synapses. In this paper, we combine numerical simulations of conductance-based network models and analytical investigations of quadratic integrate-and-fire (QIF) neurons (Hansel and Mato, 2003) to discover such principles. In particular, we show that potassium currents promote synchrony, whereas sodium currents impede it. We compare our results with existing experimental data. We suggest that our findings may account for those reported by Bou-Flores and Berger (2001). We also propose experiments to further test the predictions of our work. Part of this work has been published previously in abstract form (Pfeuty et al., 2002).

	Materials and Methods

Top Abstract Introduction Materials and Methods Results Discussion Appendix References

 
The conductance-based model. The membrane potential, V, of the neuron follows the equation:

(1)

where I_L = -g_L(V - V_L) is a leak current and _ionI_ion is the sum of all voltage-dependent ionic currents. The currents I_ext and I_noise are a constant external current and a Gaussian white noise with a zero mean and SD, , respectively.

Our neuron model incorporates an inactivating sodium current, I_Na = g_Nam ³ h(V - V_Na); a delayed rectifier potassium current, I_K = g_Kn ⁴(V - V_K); a slow potassium current, I_Ks = g_Kss ⁴(V - V_K) (Erisir et al., 1999); and a noninactivating (persistent) sodium current (French et al., 1990), I_NaP = g_NaPp(V - V_Na). The kinetics of the gating variable h, n, s are given by:

(2)

with x = h, n, s and _h(V) = 0.21e ^-(V+58)/20, _h(V) = 3/(1 + e ^-(V+28)/10), _n(V) = 0.03(V + 34)/(1 - e ^-(V+34)/10), _n(V) = 0.375e ^-(V+44)/80, _s(V) = 0.07(44 + V)/(1 - e ^-(V+44)/4.6), _s(V) = 0.008e ^-(V+44)/68. The activation functions m and p are given by: m(V) = _m(V)/(_m(V) + _m(V)), where _m(V) = 0.1(V + 35)/(1 - e ^-(V+35)/10), M(V) = 4e ^-(V+60)/18, and p(V) = 1/(1 + e ^-(V+50)/6).

Throughout this work, the parameters g_Na = 35 mS/cm ², V_Na = 55 mV, V_K = -90 mV, g_L = 0.1 mS/cm ², V_L = -65 mV, and C = 1 µF/cm ² are kept constant, and we study the ways in which the network dynamics depend on the conductances g_K, g_Ks, and g_NaP.

The network architecture. The network consists of N neurons randomly connected by bidirectional electrical couplings. The average number of connections per neuron is denoted by M. We assume that all existing synapses have the same conductance, g, and we define a connectivity matrix by J_ij = g if neuron i and j, (i, j = 1..., N), are connected and J_ij = 0 otherwise. The connection between neuron i and j adds a contribution, -g(V_i - V_j), to the total synaptic current received by neuron i, where V_i and V_j are the membrane potential of neurons i and j, respectively. Therefore, to include the effect of the electrical synapses in the dynamics of the network, an additional current I_i^ES = -_jJ_ij(V_i - V_j) is added to the right side of the differential equation satisfied by the membrane potential of neuron i (see Eqs. 1, 9). In our simulations, the network size is N = 1600, the average connectivity is M = 10, and the conductance is the same for all of the synapses: g = 0.005 mS/cm ².

Definition of synchrony and measure of the synchrony level in numerical simulations. By synchrony of the activity of a pair of neurons, we mean the tendency of these neurons to fire spikes at the same time. In this sense, a pair of neurons will be said to fire in synchrony if the cross-correlation of their spike trains displays a central peak that is above chance. The degree of synchrony can be characterized by the amplitude of this peak. According to this definition, if two neurons fire action potentials periodically and in antiphase, they can be considered to fire asynchronously.

In a large network of neurons, the activity is asynchronous if at any time the number of action potentials fired in the network is the same up to some random fluctuations. This implies that in a large network, the asynchronous state is unique. When this state is unstable, the network activity is necessarily synchronous. In this case, neurons tend to fire preferentially in some windows of time.

Note that a stable asynchronous state can coexist with a stable synchronous state if the network dynamics display multistability. In this work we focus on the conditions under which asynchronous states in large networks and antiphase locking states in pairs of identical neurons become unstable because of electrical synaptic interactions.

One can characterize the degree of synchrony in a population of N neurons by measuring the temporal fluctuations of macroscopic observables, as for example, the membrane potential averaged over the population (Hansel and Sompolinsky, 1992; Golomb and Rinzel, 1994). The quantity

(3)

is evaluated over time, and the variance of its temporal fluctuations is computed, where denotes time-averaging. After normalization of _V to the average over the population of the single-cell membrane potentials, , one defines (N):

(4)

which varies between 0 and 1. The central limit theorem implies that in the limit n , (N) behaves as:

(5)

where a > 0 is a constant and O(1/N) means a term of order 1/N. In particular, (N) = 1 if the activity of the network is fully synchronized (i.e., V_i(t) = V(t) for all i), and (N) = O(1/N) if the state of the network activity is asynchronous. In the asynchronous state, () = 0. More generally, the larger (), the more synchronized the population. Note that this measure of synchrony is sensitive not only to the correlations in the spike timing but also to the correlations in the time course of the membrane potentials in the subthreshold range.

Measure of the irregularity of spike trains. The irregularity of the spike trains is characterized by the coefficient of variability (CV) (i.e., the ratio between the SD of the interspike intervals and their mean values averaged over the entire network). For periodic spike trains, CV = 0, whereas CV = 1 for spike trains randomly distributed with Poisson statistics.

Quadratic integrate-and-fire model. It can be shown that near the onset of periodic firing, the detailed subthreshold dynamics of a large class of neurons can be reduced to a simple one-dimensional model in which the dynamical variable, v_red, evolves in time (Ermentrout, 1996; Hansel and Mato, 2003):

(6)

The constants, C_red, A, V^*, I_c^red, the reduced external current, I_ext ^red, and the reduced noise, I_{noise red}, can be computed as functions of the parameters of the full model as shown by Ermentrout (1996) and Hansel and Mato (2003). For I_red > I_c^red, the solution of Equation 6 diverges in finite time. This corresponds to the firing of an action potential. If one supplements Equation 6 with the condition that the variable v_red is reset to V_r < V^* immediately after v_red reaches some threshold, V_T, from below, this yields a reduced model that accurately describes the dynamics of the neuron in the limit I_ext ^red I_c^red. In particular, the current–frequency relationship (I–F curve) of this model and the neuron behave similarly in this limit, for any value of the parameters V_r and V_T. When I_ext ^red-I_c^red is not small, the reduced model is no longer an exact description of the neuronal dynamics. However, one can fit the parameters so that the reduced model provides a good approximation of the I–F curve of the neuron. The difference, V_T - V_r, will be called the reset depth.

It is convenient to rewrite the subthreshold dynamics of the reduced model in terms of the dimensionless variables _red and ^red defined by:

(7)

(8)

where ₀ has the dimension of a time. This yields:

(9)

The variable _red is reset to _r = A(V_r - V^*)₀ /C_red whenever it reaches the threshold value: _T = A(V_T - V^*)₀ /C_red. Note that _r < 0. A similar model has been used to study networks of excitatory and inhibitory neurons (Latham et al., 2000; Hansel and Mato, 2003).

The subthreshold dynamics (Eq. 9) need to be supplemented with a model for the suprathreshold part of the membrane potential time course. Assuming that the width of the action potentials is much smaller than the interspikes, we represent their time course by a -function at each time a spike is fired. Therefore, the (dimensionless) reduced membrane potential of the neuron can be written:

(10)

where measures the integral over time of the suprathreshold part of an action potential.

The model defined by Equations 9 and 10 will be called the QIF model. To simplify notations of the reduced model, we will drop the index ("red") and the tildes.

Phase reduction in the weak coupling limit. Let us consider a neuron firing action potentials periodically with an interspike T. A small and instantaneous perturbation applied at time t after a spike induces a small change in the timing of the subsequent spikes. This change, which depends on t, or equivalently on = 2t/T, can be characterized by a function, Z(), which measures the delay or the advance induced in the firing times after the perturbation. A positive value of Z() indicates that the perturbation advances the subsequent spikes. A negative value of Z() indicates a delay. The effect of noninstantaneous weak perturbations (such as spikelets) can be estimated by convolving the phase-response function (PRF), Z, with the perturbation. It can be shown that the dynamic behavior of a network of weakly interacting neurons can be completely described in terms of the response function in the framework of the phase reduction approach (Kuramoto, 1984; Ermentrout and Kopell, 1986; Hansel et al., 1993, 1995; Golomb and Hansel, 2000; Neltner et al., 2000; Lewis and Rinzel, 2003). In this approach, a phase variable is associated with each neuron in the network. This variable, _i(i = 1,..., N), measures the time elapsed since the last action potential fired by neuron i. For a network of identical neurons coupled with electrical synapses, one can show (see Appendix) that the phase variables follow a set of n first-order coupled differential equations:

(11)

(_i - _j) the phase coupling between neurons i and j given by:

(12)

J_ij is the connectivity matrix, and _i(t) is a white noise with zero mean and variance

with , the SD of the noise of the nonreduced dynamics.

Numerical integration. In the simulations of the conductance-based model, the differential equations were integrated using the second-order Runge-Kutta scheme with fixed-time step: _t = 0.01 msec. Averaged quantities (firing rate, CV, ) were computed over a time period of 1 sec after discarding a transient of 500 msec.

	Results

Top Abstract Introduction Materials and Methods Results Discussion Appendix References

 
Single neuron and coupling properties of the conductance-based model
Firing properties of single neurons
In the absence of persistent sodium and slow potassium currents (the "control model"), our conductance-based model neuron fires tonically for large enough external currents, I_ext > I_c = 0.16 µA/cm². The frequency of the discharge in response to a step of current is plotted as a function of the step amplitude (I–F curve) in Figure 1A. For I_ext larger than but close to I_c, the minimum value of the external current required for the neuron to fire, the firing frequency can be arbitrarily small, because action potentials appear at I_c through a saddle-node bifurcation [for a definition of a saddle-node bifurcation, see Strogatz (1994) and Rinzel and Ermentrout (1998)]. Decreasing g_K does not significantly modify the resting potential and the rheobase of the neuron, but it increases the gain of the I–F curve (Fig. 1B). The slow potassium (I_Ks) and persistent sodium (I_NaP) currents modify the onset of firing, because these currents are activated at rest. As should be expected, the persistent sodium current increases the excitability of the neuron and reduces its rheobase, I_c (Fig. 1C). For g_NaP > 0.08 mS/cm², the neuron is spontaneously active. As shown in Figure 1D, the main effects of adding I_Ks on the I–F curve are a translation to the right, a linearization, and, if g_Ks is large enough, a suppression of the ability of the neuron to fire at low rates (because for large enough g_Ks, a subcritical Hopf bifurcation occurs at the onset of firing [for a definition of a subcritical Hopf bifurcation, see Strogatz (1994) and Rinzel and Ermentrout (1998)].

View larger version (26K): [in this window] [in a new window]   Figure 1. I–F curves of the conductance-based model neuron. Each panel corresponds to a different set of intrinsic conductance values. The firing rate was computed from the steady-state response of the neuron to steps of constant external current of different amplitudes. A, Control case: gK = 9 mS/cm 2, gKs = gNaP = 0. B, gK = 2.5 mS/cm 2, gKs = gNaP = 0. Note that the scale of the x-axis is one-half the scale used in A. The main effect on the I–F curve of the reduction of gK is multiplicative (change in the gain). C, gK = 9 mS/cm 2, gNaP = 0.2 mS/cm 2, gKs = 0. The main effect on the I–F curve of the persistent sodium is subtractive (change in the rheobase). D, gK = 2.5 mS/cm 2, gKs = 0.2 mS/cm 2, gNaP = 0. The slow potassium current reduces the gain of the neuron and prevents low-frequency firing. It also linearizes the I–Fcurve (compare with B).

 

The response of the neuron to a step of external current, whose value is adjusted to give a firing frequency of 50 Hz, is plotted in Figure 2 for different combinations of the intrinsic currents. The action potentials are shown with a higher temporal resolution in Figure 3C. In the control case (Fig. 2A; and Fig. 3C, solid line), action potentials are narrow and are followed by a strong afterhyperpolarization (AHP) that brings the membrane potential 15 mV below its resting value. When g_K is decreased (Fig. 2B), the resting potential of the neuron does not change, but the AHP is reduced and the membrane potential always remains above rest. This conductance also controls the width of the spike and the depth of the AHP [Fig. 3C, compare the solid line (g_K = 9 mS/cm²) with the dotted line (g_K = 2.5 mS/cm²)].

View larger version (27K): [in this window] [in a new window]   Figure 2. The response of the neuron to a step of current. A step of current of 200 msec was injected into the neuron. No noise was included. The membrane potential before and after the current injection is indicated to the left of each panel. In all cases, the external current during the step was chosen to obtain a discharge rate of 50 Hz. A, Control case. External current, Iext, varies from 0 to 1.10 µA/cm 2. Resting potential is -63 mV. B, gK = 2.5 mS/cm 2, gKs = gNaP = 0. Iext varies from 0 to 0.48 µA/cm 2. Note that the concavity of the membrane potential time course is always upward in contrast to A. C, gK = 9 mS/cm 2, gNaP = 0.2 mS/cm 2, gKs = 0. The neuron is hyperpolarized by the injection of a negative current, Iext = -1.55 µA/cm 2, before the current step, to prevent spontaneous firing. The step amplitude is 1.0 µA/cm 2. D, gK = 2.5 mS/cm 2, gKs = 0.2 mS/cm 2, gNaP = 0. The external current varies from 0 to 4.88 µA/cm 2. The resting membrane potential is more hyperpolarized than in B, because of the slow potassium current that is activated at rest. The (shifted) time course of the membrane potential during the discharge is very similar to the trace in B. Note the transient hyperpolarization right after the switching off of the current, which is caused by the slow relaxation of IKs back to its resting value.

 

View larger version (24K): [in this window] [in a new window]   Figure 3. The spikelets induced for different combinations of ionic currents. Solid line, gK = 9 mS/cm 2, gKs = gNaP = 0 (control case). Dotted line, gK = 2.5 mS/cm 2, gKs = gNaP = 0. Dashed line, gK = 9 mS/cm 2, gKs = 0, gNaP = 0.2 mS/cm 2. Dashed-dotted line, gK = 2.5 mS/cm 2, gKs = 0.2 mS/cm 2, gNaP = 0. A, An action potential generated by a short and strong current pulse. The duration and the amplitude A of the pulse are the same for all four traces: = 1 msec and A = 50 µA/cm 2. The neuron was hyperpolarized to prevent firing when persistent sodium was present and to obtain the same initial membrane potential for all four parameter sets. The solid line and the dashed line overlap. B, Spikelet induced by a presynaptic neuron firing an action potential as in A. The synaptic conductance is g = 0.005 mS/cm 2, which corresponds to a coupling coefficient (Amitai et al., 2002) CC5%. The postsynaptic neuron has the same intrinsic properties as the presynaptic neuron. C, A constant external current is injected to make the neuron fire at 50 Hz. The external current is Iext = 1.10 µA/cm 2 for the solid line, Iext = 0.48 µA/cm 2 for the dotted line, Iext = 4.88 µA/cm 2 for the dashed-dotted line, and Iext = -0.55 µA/cm 2 for the dashed line. The subthreshold voltage of the neuron when the persistent sodium is added (dashed line) is below the control (solid line), although the persistent sodium current accelerates the depolarization of the neuron, because here we compare the firing patterns for the same firing rate. The external current is therefore smaller when the persistent sodium is present than in the control case. D, Spikelets induced by a presynaptic neuron firing tonically, as in C. The postsynaptic and the presynaptic neurons have the same intrinsic properties.

 

The persistent sodium and the slow potassium currents significantly modify the resting potential and the external current required to adjust the firing rate. In contrast, they only slightly affect the depth of the repolarization, measured from threshold, after the action potential (Fig. 2C,D). The size and width of the action potentials remain almost unchanged when the conductances of these currents are varied [Fig. 3C, compare the dotted line with the dashed-dotted line (effect of g_Ks) and the solid line with the dashed line (effect of g_NaP)].

Properties of the spikelets
When a neuron fires an action potential, it induces a depolarization, called a spikelet, in the neurons that are connected to it. The spikelets generated by a neuron firing one isolated spike in response to a very brief but strong pulse of current (Fig. 3A) or firing tonically at 50 Hz (Fig. 3C), are shown in Figure 3, B and D, respectively, for different parameters of the ionic currents.

In all four cases displayed, the width of the spikelets is much larger than the width of the presynaptic spikes that generate them. This is because the spikelets are a filtered version of the presynaptic membrane potential profile.

The size of the spikelet is practically unaffected by the presence of the persistent sodium (Fig. 3B,D, compare the solid line with the dashed line), because this current does not greatly affect the shape of the action potential. The only noticeable effect of the persistent sodium on the spikelet is to slow down its decay time course. This is because it reduces the input conductance of the postsynaptic neuron (by 50% at -76 mV).

The modulation of the spikelets by the delayed rectifier potassium current is primarily caused by changes in the presynaptic neuron voltage time course. When g_K is reduced, the action potential is broader, and this induces an increase in the amplitude and the width of the spikelets. Although the slow potassium current does not greatly affect the width of the action potential, it contributes to bringing the membrane of the presynaptic neuron transiently below its holding potential. This explains why in Figure 3B the spikelet is narrower and displays a faster and deeper repolarization in the presence of I_Ks.

The spikelets generated by tonically firing neurons at low firing rates are similar to those in Figure 3B (data not shown). For high-enough firing rates, the slow potassium saturates, affecting the dynamics primarily as would a constant hyperpolarizing current. Therefore, its effect on the voltage traces and on the spikelets becomes less pronounced (Fig. 3, compare B and D).

Synchrony in the conductance-based network model
The goal of this section is to show how synchrony properties of our conductance-based network model depend on the intrinsic currents of the neurons. Further understanding of the effects described in this section will be provided in the next section by investigating analytically a simplified and more abstract model.

Synchrony is modified when intrinsic conductances are changed A pair of conductance-based neurons
The effect of intrinsic currents on neuronal network dynamics can first be demonstrated by studying the dynamics of a pair of neurons. Here, as an example, we show that a persistent sodium current tends to promote antiphase locking, whereas a potassium current promotes in-phase locking. In Figure 4A, the traces of two identical neurons (g_K = 9 mS/cm²; g_Ks = g_NaP = 0) are plotted over a time interval of 100 sec. Both neurons receive a noisy input that makes them fire at 50 Hz. These traces suggest that the neurons fire action potentials in synchrony. This is confirmed by the cross-correlation of the traces computed over a longer time interval (100 sec), which displays a strong peak centered around t = 0 msec, as shown in Figure 4B. In contrast, when the persistent sodium conductance is large enough (g_NaP = 0.4 mS/cm²), neurons tend to fire in antiphase (Fig. 4C), and the peak of the cross-correlogram is shifted to 10 msec (Fig. 4D). If a slow potassium current is added (g_Ks = 0.15 mS/cm²) while keeping the persistent sodium conductance the same, antiphase locking becomes unstable and neurons tend to fire in-phase (Fig. 4E, F). Therefore, a slow potassium current is a promoter of synchronous activity.

View larger version (28K): [in this window] [in a new window]   Figure 4. Effect of persistent sodium and potassium currents on the synchronization of a pair of neurons: phase versus antiphase stability. Results from numerical simulations. The synaptic conductance is g = 0.005 mS/cm 2. The SD of the noise is = 0.3 mV/msec1/2. In all of the simulations, neurons were firing in antiphase at the beginning of the simulation (data not shown). Spike trains were computed with a time bin of 1 msec. The spike cross-correlograms (B, D, F) were calculated from spike trains recorded over 100 sec (after discarding a transient of 5 sec) and were normalized to the expected number of spikes during this time interval. Neurons fire at an average firing rate of 50 Hz. A, B, Traces of the voltage of the two neurons and the cross-correlograms of the spike trains for the control parameters and Iext = 1.08 µA/cm 2. Neurons tend to fire in-phase. C, D, Persistent sodium current stabilizes antiphase locking. Parameters are gK = 9 mS/cm 2, gNaP = 0.4 mS/cm 2, and gKs = 0. Iext = -1.38 µA/cm 2. E, F, Slow potassium conductance destabilizes antiphase locking. Parameters are gK = 9 mS/cm 2, gNaP = 0.4 mS/cm 2, and gKs = 0.15 mS/cm 2. Iext = 0.80 µA/cm 2. Neurons tend to fire in-phase.

 

Large networks of conductance-based neurons
In Figure 5 the synchrony measure, , is plotted against the conductances g_K, g_Ks, and g_NaP. Examples of raster plots of network activity are also displayed in this figure, for particular values of the intrinsic conductances (Fig. 5, insets). In each of the panels, the external input (average deviation and SD) is kept constant. Therefore, when the intrinsic conductances vary, the average firing rate of the neurons and the CV of their interspike intervals change. In particular, the firing rate decreases (respectively increases) when the potassium (respectively the persistent sodium) conductances increase (top right figures in each of the panels in Figure 5).

View larger version (29K): [in this window] [in a new window]   Figure 5. Effect of intrinsic currents on the synchronization of a large network. Results from numerical simulations. Details on the simulation and the network parameters are given in Materials and Methods. The synaptic conductance is g = 0.005 mS/cm 2. In all of the simulations, the SD of the noise was set at = 0.6 mV/msec 1/2. The synchrony level, (left), the average frequency, and the average CV of the neuronal discharges are plotted as a function of an intrinsic conductance that is varied. A, The conductance, gK, varies; gKs = gNaP = 0. The external current is Iext = 0.8 µA/cm 2. For gK > g*K, increases and saturates. Inset, Raster plot representing the spiking times of 100 neurons for gK = 9 mS/cm 2 (neurons are arrayed vertically, and the x-axis corresponds to time where the scale denotes 25 msec). Panels on the right indicate that the average frequency of the neurons decreases when gK increases, because the gain of the I–F curve decreases. The variability of the neuronal discharge varies slightly and nonmonotonically in the investigated range of gK. B, The conductance, gKs, varies; gK = 2.5 mS/cm 2, gNaP = 0. The external current is Iext = 2 µA/cm 2. Inset, Raster plots are for gKs = 0 and for gKs = 0.1 mS/cm 2. At large values of gKs, a slight decrease occurs, but the activity of the network remains strongly synchronized. Panels on the right indicate that the average firing rate of the neurons is strongly reduced when gKs varies, primarily because IKs has a strong subtractive effect on the I–F curve of the neurons (the rheobase increases with gKs). C, The conductance, gNaP, varies; gK = 9 mS/cm 2, gKs = 0. The external current is Iext = 0.8 µA/cm 2. The synchrony is reduced when gNaP increases. Inset, Raster plot for gNaP = 0.2 µA/cm 2. Panels on the right indicate that the average firing rate of the neurons increases with gNaP, primarily because of the subtractive effect of INaP on the I–F curve of the neurons (the rheobase is reduced when gNaP increases).

 

In Figure 5A, the effect of g_K on the synchrony level is shown (g_Ks = g_NaP = 0). When g_K is small enough (g_K < 4.5 mS/cm²), is very close to zero. A detailed study shows that in this region, is on the order of 1/N, where N is the number of neurons in the network, and vanishes in the limit of a very large network (data not shown): non-zero values of are attributable to finite size effects, and the network is asynchronous. For g_K of >4.5 mS/cm², increases rapidly up to 0.35. For example, in the raster plot shown in the right inset of Figure 5A (control situation, g_K = 9 mS/cm²), = 0.34. This corresponds to a state in which the neurons fire together within time windows of approximately one-third of the period.

Figure 5B shows the effect of the slow potassium current for g_NaP = 0 and g_K = 2.5 mS/cm². For g_Ks = 0, the neurons fire asynchronously. At g_Ks 0.06 mS/cm², starts to increase. For large g_Ks, saturates at a value, 0.55, that is substantially larger than in the control situation. This value of corresponds to a tight synchrony of the action potentials fired by the neurons (Fig. 5B inset on the right).

In Figure 5C, is plotted against g_NaP for g_K and g_Ks at their control values (g_K = 9 mS/cm², g_Ks = 0). Clearly, is a decreasing function of g_NaP. When g_NaP is large enough (g_NaP > 0.1 mS/cm²), the neurons fire asynchronously.

In the results presented in Figure 5, the firing rate of the neurons is not controlled. Therefore, the origin of the variation of with the intrinsic current conductances is not clear. It may be attributable to the fact that, in general, synchrony depends on the firing rate for given intrinsic currents. It might also be a result of the fact that dynamic properties of neurons with different intrinsic currents may be different even if they have the same firing rate.

Dependence of synchrony on intrinsic currents at fixed firing rates
To clarify further how potassium and sodium currents affect the synchrony of the network, we performed another set of numerical simulations in which the external input and the noise level were tuned to keep constant both the average firing rate of the neurons and the CV of their spike trains when the conductances of the intrinsic currents were changed. The results are shown in Figure 6 (three panels on the left). In all of the simulations, the average firing rate of the network was 50 Hz and the CV 0.12 (Fig. 6, panels on the right).

View larger version (24K): [in this window] [in a new window]   Figure 6. The synchrony level depends on the intrinsic currents when the frequency is kept constant. Results from numerical simulations. Details on the parameters are given in Materials and Methods. The synaptic conductance is g = 0.005 mS/cm 2. The external current, Iext, and the SD of the noise, , were changed to control the average firing rate and the average CV of the neurons. In all of the simulations, the average firing rate is 50 Hz ± 10%, and the CV 0.12. In each of the panels, the corresponding values of Iext and are plotted versus the conductance that is varied (panels to the right). A, The conductance, gK, varies; gKs = gNaP = 0. B, The conductance, gKs, varies; gK = 2.5 mS/cm 2, gNaP = 0. C, The conductance, gNaP, varies; gK = 9 mS/cm 2, gKs = 0.

 

The level of synchrony increases with the potassium conductances. For small-enough g_K or g_Ks, the activity of the network is weakly correlated. A detailed study of the dependence of with the network size reveals that for g_K (respectively g_Ks) smaller than g^*_K 4 mS/cm² (respectively g_Ks smaller than g^*_Ks 0.075 mS/cm²), is on the order of 1/N (data not shown). Therefore in this range of conductances, the network is in the asynchronous state. At g^*_K and g^*_Ks, a transition to a synchronous state occurs, and beyond these values the network settles into a synchronous state. This occurs although the level of noise has been increased.

In contrast to the potassium conductances, which promote synchrony, the persistent sodium current impedes it, as shown in Figure 6C. For large-enough g_NaP, synchrony is even destroyed. A sharp transition to the asynchronous state occurs for g_NaP 0.15 mS/cm².

Relying on the way intrinsic currents affect spikelet modulation (Fig. 3), one might naively expect that potassium currents that reduce the size of the spikelets should also reduce synchrony, and that persistent sodium current, which increases this size, should promote it. As a matter of fact, we found a trend that was exactly the opposite.

Firing rate affects synchrony differently when different intrinsic currents are involved
We also investigated in what ways the network dynamic state depends on the firing rate of the neurons. The firing rate was changed by varying the average external input, and was plotted as a function of the frequency of the neurons for several combinations of potassium and sodium conductances.

Figure 7 plots the synchrony measure, , versus the average firing rate of the neurons for four sets of conductance parameters and two levels of noise. The qualitative behavior of as a function of the firing rate depends on the intrinsic currents involved.

View larger version (25K): [in this window] [in a new window]   Figure 7. Synchrony varies with the firing rate in a way that depends on the intrinsic currents. Results from numerical simulations. Details on the simulation parameters are given in Materials and Methods. The synaptic conductance is g = 0.005 mS/cm 2. The average firing rate of the neurons is varied by changing the external current. In each of the panels, the results are displayed for two noise levels. Circles, = 0.6 mV/(msec)1/2. Squares, = 0.3 mV/(msec)1/2. A, gK = 9 mS/cm 2, gKs = gNaP = 0. The external current was varied between 0.15 and 2.7 µA/cm 2 (100 Hz). B, gK = 3.5 mS/cm 2, gKs = gNaP = 0. The external current, Iext, was varied between 0.11 and 1.4 µA/cm 2. C, gK = 9 mS/cm 2, gKs = 0, gNaP = 0.15 mS/cm 2. Iext was varied between -1.17 and 0.8 µA/cm 2. D, gK = 3.5 mS/cm 2, gKs = 0.15 mS/cm 2, gNaP = 0. Iext was varied between 2.3 and 7.2 µA/cm 2. The firing rate cannot be reduced to <20 Hz because of the presence of the slow potassium current (see also Fig. 1 D).

 

As expected, is always larger for the smaller level of noise ( = 0.3 mV/msec^1/2) (Fig. 7, circles) than for the larger one ( = 0.6 mV/msec^1/2) (Fig. 7, squares). In particular, for = 0.6 mV/msec^1/2, the noise is strong enough to prevent synchrony in the entire range of firing frequencies for g_K = 3.5 mS/cm² (Fig. 7B) and for g_NaP = 0.15 mS/cm² (Fig. 7C). This level of noise is sufficient to destroy synchrony as well in the control situation, but only at <20 Hz (Fig. 7A).

In the control case, increases monotonously with the firing rate (except for a small decrease followed by a slight increase between 50 and 100 Hz for = 0.3 mV/msec^1/2). In contrast, for smaller g_K (e.g., g_K = 3.5 mS/cm²), we find a nonmonotonous variation of synchrony: increases at low firing rates but decreases again for firing rates of >20 Hz. The network state becomes asynchronous for firing rates of >60 Hz (Fig. 7B, squares). If the slow potassium current is added (g_Ks = 0.15 mS/cm²), again varies monotonously with the firing rate (Fig. 7D). Finally, with enough persistent sodium current and not overly strong noise (g_K = 9 mS/cm², with g_NaP = 0.15 mS/cm², g_Ks = 0) (Fig. 7C, squares), varies nonmonotonously with the firing rate, as in the reduced g_K case (Fig. 7B).

Synchrony in networks of quadratic integrate-and-fire neurons depends on the phase response function
The PRF, which characterizes how neurons respond to small perturbations, is a key concept to understand the relationship between intrinsic properties of neurons and their collective dynamics (Kuramoto, 1984; Hansel et al., 1993; van Vreeswijk et al., 1994; Kopell and Ermentrout, 2002). This function depends on the excitability properties of the neurons and therefore is determined by the intrinsic currents involved in their dynamics. One expects that the differences in the dynamic behavior of our conductance-based network model for different sets of parameters reflect the changes in the excitability properties of the neurons. However, relying on numerical simulations alone, it would very difficult to establish general principles to relate the excitability properties and the neuronal PRF to the synchronization properties. Below, we consider a network of QIF neurons (Eqs. 9, 10) fully connected by electrical synapses. As we show below, the qualitative properties of the excitability of the QIF neurons crucially depend on the parameters (the threshold and the reset voltages and the external current). Thanks to its relative simplicity, this model can be investigated using analytical techniques. It reveals how, in the framework of this model, one can relate the stability of antiphase locking of a pair of neurons and the stability of the asynchronous state of a large network to the shape of the PRF of the neurons. Subsequently, we show that similar rules can be applied to conductance-based models, and that they provide a unified framework explaining the diversity of behavior found in our numerical simulations of the conductance-based model.

The membrane potential and the PRF of QIF neurons
In the Appendix, it is shown that between two successive spikes the subthreshold membrane potential of a QIF neuron reads:

(13)

where t measures the time elapsed since the first spike. The firing period, T, is determined by the condition: v(T) = V_T. Therefore:

(14)

In the weak coupling limit, the dynamics of the network are completely determined by the phase-response function, Z() (see Materials and Methods), which depends on three parameters, namely, the firing frequency of the neurons (or equivalently the external current I_ext), the threshold V_T, and the reset potential V_r. The response function of the QIF model can be derived analytically as shown in the Appendix. One finds:

(15)

Changing the ratio -V_r/V_T for a fixed reset depth, V_T - V_r has a strong influence on the shape of the trajectory of the phase-response function of the neuron, as shown in Figure 8. In Figure 8A, the membrane potential of a QIF neuron firing at 50 Hz is plotted for three values of the ratio V_r/V_T and a fixed reset depth: V_T - V_r = 3. For -V_r/V_T = 0.1, the concavity of the subthreshold trajectory is always directed upward. In contrast, for -V_r/V_T = 10, it is always downward. For -V_r/V_T = 1, the concavity changes from upward to downward.

View larger version (21K): [in this window] [in a new window]   Figure 8. Voltage traces and phase-response functions of the QIF neuron. The voltage traces and the phase-response functions were computed using Equations 13 and 15, respectively. 0 = 10 msec. In A and B, the external current is such that the firing rate is 50 Hz. A, The voltage trace of the neuron in response to a constant injected current for three values of the ratio -Vr/VT. The reset depth is the same in all three cases: VT - Vr = 3. Note the changes in the concavity of the voltage trace between the spikes when -Vr/VT increases. The external current is Iext = 0.97 for -Vr/VT = 1 and for -Vr/VT = 10, and Iext = 0.66 for -Vr/VT = 0.1. B, The phase-response function, Z, for the same cases as in A. C, Same as in B, but for a firing rate of 10 Hz. The phase-response changes less than in B when -Vr/VT varies.

 

It is easy to see that Z() > 0 for all , and that it is with one maximum. The location of the maximum of Z depends on V_r, V_T, and I_ext. This is depicted in Figure 8, where the voltage trace and the response function are plotted for different values of these parameters. For -V_r/V_T = 1, Z() is symmetric and its maximum is always at = . For -V_r/V_T < 1, the maximum of Z is located in the first half of the firing period, whereas for -V_r/V_T > 1, it is located in the second half. The value of Z at the maximum also depends on V_r, V_T. It has smaller values when -V_r/V_T = 1. The larger the firing rate, the stronger the dependence of Z on -V_r/V_T (Fig. 8, compare B, for 50 Hz, and C, for 10 Hz).

A pair of identical QIF neurons without noise
When, in the absence of noise, two identical neurons interact in a symmetric way, they reach, at large time, a phase-locked state in which the two neurons fire action potentials with a fixed phase shift, . In the Appendix, it is shown that is a solution to the equation:

(16)

where the function _-() 1/2(() - (-)), and () is the effective phase coupling (see Materials and Methods). Because of the symmetry of the system and the 2 periodicity of the function _-, there are always at least two solutions to this equation: = 0 and = . The first solution corresponds to the state in which the two neurons fire in-phase (in-phase locking), the second one to the state in which they fire in antiphase (antiphase locking). Other solutions with 0 < < may also exist. However, only solutions that are stable can be reached at large time. Therefore, out of all of the solutions to Equation 16, only those that satisfy the condition (see Appendix):

(17)

correspond to phase-locked states that the two neurons can eventually reach starting from appropriate initial conditions. In the following, we focus on the stability of the antiphase state of a pair of QIF neurons.

Stability conditions of antiphase locking
The voltage trajectory consists of three parts: (1) a subthreshold part, during which the membrane potential increases with time from a value V_r to a threshold value V_T; (2) a spike, modeled by a -function of amplitude , and (3) a resetting, where potential is instantaneously brought back to V_r. Then, as explained in the Appendix, antiphase locking of a pair of QIF neurons is stable if:

(18)

with:

(19)

(20)

(21)

The first term, S_sp, is the contribution of the presynaptic spikes to the destabilization of antiphase locking. The second term, S_r, is the contribution of the instantaneous reset of the membrane potential. The last term, S_sub, corresponds to the effect of the coupling between the two neurons when the presynaptic neuron is subthreshold. These three terms are plotted in Figure 9A as a function of the ratio -V_r/V_T, for = 1, V_T - V_r = 3, and a frequency v = 50 Hz.

View larger version (29K): [in this window] [in a new window]   Figure 9. Stability of the asynchronous state of QIF networks at weak coupling in the absence of noise. Results from the analytical study of the QIF model. Time constant0 = 10 msec. The size of the spikes is = 1. A, A pair of identical neurons. The quantity '() is plotted (solid line). The three contributions to this quantity (Eq. 21) are also plotted: Ssub (dotted line), Sr (dashed-dotted line), and Ssp (dashed line). The double-dotted-dashed line corresponds to Sr + Ssub. All of these quantities were computed from Equations 21, 13, and 15. Note the logarithmic scale on the x-axis. The neurons fire at 50 Hz. B, A large network of all-to-all weakly coupled identical neurons: the quantity µ1. The asynchronous state is stable with respect to the mode n = 1 if µ1 < 0. The quantities µ' 1 (dashed line) and µ ' 1 (double-dotted-dashed line) are also plotted. The firing rate of the neurons is 50 Hz. C, The phase diagram of antiphase locking for a pair of identical neurons in the weak coupling limit. Note the logarithmic scale on the x-axis. The solid lines separate the region in which antiphase locking is unstable (in the middle) from the regions in which it is stable (on the left and in the small region on the right). The parameters Vr and VT are varied, whereas the reset depth is kept constant: VT - Vr = 3. The boundary lines are defined by the condition: '() = 0. D, The phase diagram for the stability of the asynchronous state in a large network of all-to-all coupled neurons. The solid line is the boundary between the domains in which the asynchronous state is stable (to the left) and unstable (to the right).

 

The terms S_r and S_sub (Fig. 9, dotted line and dashed-dotted line, respectively) are invariant under the transformation -V_r/V_T V_T/-V_r. That is why the corresponding curves in Figure 8B are symmetric around -V_r/V_T = 1 (note the logarithmic scale of the x-axis). S_r is always positive and therefore destabilizes antiphase locking. In contrast, S_sub is always negative and leads to stabilized antiphase locking. The sum of the two terms is also plotted in Figure 9A (double-dotted-dashed line) to show that their overall contribution is destabilizing except for very small or very large values of -V_r/V_T.

Because the maximum of Z moves from the first part of the period to the second part of the period when -V_r/V_T crosses 1 (Fig. 8), the sign of the derivative of Z at = also changes from negative to positive, and the term S_sp (dashed line) tends to destabilize antiphase locking for -V_r/V_T > 1. The sum of all three terms of the right side of Equation 18 is plotted in Figure 9A (solid line). Antiphase locking is stable only if the phase at which Z reaches its maximum is small enough.

The phase diagram for the stability of antiphase locking as a function of -V_r/V_T and the firing frequencies is displayed in Figure 9C. The stability of antiphase locking requires that -V_r/V_T be small enough. Moreover, the range of values of -V_r/V_T where it is stable increases with the frequency of the neurons. This is because the contribution of the term S_sp in the right side of Equation 18 grows with the firing rate and tends to dominate the contribution of the two other terms. In contrast, at low frequencies, the response function weakly depends on -V_r/V_T (Fig. 8), and S_sp, which is proportional to the frequency, is negligible. Moreover, S_r + S_sub is positive except for very small or very large values of -V_r/V_T. Therefore, for small frequencies, the domain of stability of antiphase locking is very narrow.

This analysis shows that for a pair of neurons, the stability of antiphase locking depends critically on the shape of the phase-response function and in particular on the sign of the derivative of the response function at the mid-period. The more skewed toward the right (the second part of the firing period) the response function of the neurons, the more unstable antiphase locking becomes.

Large networks of weakly coupled QIF neurons
The stability analysis of the asynchronous state of a large network of neurons coupled all-to-all receiving a noisy input is also simplified if one assumes weak coupling. Using the phase-reduction approach, it can be shown (see Appendix) that the asynchronous state is stable if:

(22)

For all integers, n > 0, with Z_n and v_n the nth-Fourier components of the function Z() and v(), respectively (Eqs. 13, 15) defined by and a similar equation for v_n. If for some n, µ_n is negative (respectively positive), it is said that the mode of order n is stable (respectively unstable). For the asynchronous state to be stable, the modes at any order must be stable. It is unstable when at least one mode is unstable.

The first two terms in Equation 22 correspond to the effect of the interaction. The first term represents the effect of the spikes and can be written:

(23)

The second term, µ ' _n = ng Im(nZ_nv_-n), corresponds to the combined effect of the reset of the membrane potential and its subsequent subthreshold evolution.

The sign of their sum depends on the parameters, V_r, V_T and on the frequency of the neurons, v. The last term in Equation 22 corresponds to the effect of the noise. It is always positive. Therefore, as should be expected, noise increases the stability of the asynchronous state (because of the negative sign in front of this term in Eq. 22). The stability of the asynchronous state depends on the competition between the first two terms and the last term. In particular, the sign of µ_n depends on the ratio ²/g. Note that because of the factor n² in the last term of Equation 22, the stability of the modes increases rapidly with their order.

We first consider the case of a noiseless network (i.e., with = 0). In Figure 9B, we plotted µ' ₁ (dashed line) and µ ' ₁ (dashed-dotted line) against -V_r/V_T. The qualitative behavior of these quantities is similar to those of S_sp and S_sub + S_r, respectively, for a pair of neurons. In particular µ ' ₁ is symmetric (in a semilogarithmic scale) around -V_r/V_T = 1. Moreover, µ' ₁ increases monotonically from negative values to positive values when -V_r/V_T increases, and it changes sign at -V_r/V_T = 1. This can be easily explained. Indeed, for -V_r/V_T > 1 (respectively, -V_r/V_T < 1), the function Z() is skewed toward < (respectively > ) where sin() > 0 [respectively sin() < 0]; therefore, for n = 1, the integral in Equation 23 is negative (respectively positive) and µ' ₁ < 0 (respectively µ' ₁ > 0). In particular, because for -V_r/V_T = 1, Z()is symmetric around = , the integral in Equation 23 vanishes for n = 1 [because sin( + ) = -sin()]. The sign of µ₁ (solid line) is primarily determined by the sign of µ' ₁. It is negative for small -V_r/V_T and becomes positive for -V_r/V_T 1. Therefore the mode n = 1 is stable only if -V_r/V_T is small enough. Similar analysis can be performed for the modes n > 1. It reveals that if -V_r/V_T is not too large, the mode n = 1 determines