Abstract

Learning of sensory cues is believed to rely on synchronous pre- and postsynaptic neuronal firing. Evidence is mounting that such synchronicity is not merely caused by properties of the underlying neuronal network but could also depend on the integrity of gap junctions that connect neurons and astrocytes in networks too. In this perspective, we set out to investigate the effect of astrocytic gap junctions on perceptual learning, introducing a model for coupled neuron-astrocyte networks. In particular, we focus on the fact that astrocytes are rich of GABA transporters (GATs) which can either uptake or release GABA depending on the astrocyte membrane potential, which is a function of local neural activity. We show that GABAergic signaling is a crucial component of intracolumnar neuronal synchronization, thereby promoting learning by neurons in the same cell assembly that are activated by a shared sensory cue. At the same time, we show that this effect can critically depend on astrocytic gap junctions insofar as these latter could synchronize extracellular GABA levels around many neurons and throughout entire cell assemblies. These results are supported by extensive computational arguments and predict that astrocytic gap junctions could improve perceptual learning by controlling extracellular GABA.

1  Introduction

Perceptual learning refers to our ability to improve perception of external sensory stimuli as we experience the same stimulus repeatedly and has been reported in different sensory systems, most notably including the visual system (Schoups, Vogels, Qian, & Orban, 2001), the auditory system (Ohl & Scheich, 2005), and the somatosensory system (Dinse, Ragert, Pleger, Schwenkreis, & Tegenthoff, 2003). It is a shared belief that the neural correlate of perceptual learning is the change of strength of synaptic connections between neurons by spike-timing-dependent (Hebbian) plasticity (STDP) (Gilbert, 1994, 1996). Accordingly, in the classical STDP formulation, a spike evoked in a presynaptic neuron shortly before or after a spike evoked in a postsynaptic neuron activates the synapse between the two neurons (Bi & Poo, 1998; Markram et al., 2004). The former induces long-term-potentiation (LTP) and the latter long-term depression (LTD).

Precise manipulation in physiological experimental settings allows us to generate coincident firing of neurons, by which STDP can be produced easily. Experiments (Berninger & Bi, 2002) have indicated the existence of biological mechanisms for maintaining coherence in neuronal firing and suggested that certain input stimuli triggered neuronal activity that propagated and reverberated along specific pathways in a recurrent excitatory manner, helping achieve STDP. However in behaving animals, this might not be sufficient insofar as neurons are subjected to a variety of inputs in addition to the sensory cue, and their firing activity is often stochastic (Softky & Koch 1993). Consequently a topic of active research is to understand how the brain develops neural synchrony to deploy STDP.

GABAergic signaling contributes to synchronizing neuronal activity. In the cortex, basket cells can coordinate the timing of spikes and synchronize them (Paulsen & Moser, 1998). In the hippocampus, neuronal synchronization could be triggered by phasic stimulation of basket cells (Bonifazi et al., 2009). Parvalbumin (PV)-expressing, fast-spiking interneurons interact with pyramidal cells and generate cortical oscillations in the rhythm (i.e., 20–40 Hz; Buzsaki & Draghun, 2004) that may coordinate cortical activity during cognitive processing (Volman, Behrents, & Sejnowski, 2011).

In addition to a mere neuronal substrate for STDP, evidence is mounting that astrocytes, the most numerous cortical glial cell type, could also have a role in synchronization of neural activity (Poskanzer & Yuste, 2011, 2016). Astrocytes can control the extracellular environment regulating concentration there of multiple ions and neurotransmitters such as glutamate (Bergles, Diamond, & Jahr, 1999) and GABA (Losi, Mariotti & Carmignoto, 2014). Remarkably, astrocytes are rich in GABA transporters (Losi et al., 2014), whereby they could modulate extracellular (ambient) GABA levels and tonically modulate neuronal receptors ultimately affecting neuronal activity (Serrano, Haddjeri, Lacaille, & Robitaille, 2006; Yoon & Lee, 2014).

We speculated that astrocytic GABA might have an important role in STDP. Astrocytes might regulate local ambient GABA levels by transporters, which could be synchronized by gap junctions between astrocytes, thereby possibly developing synchronization of pyramidal cell activity. To test our hypothesis, we propose here a neural network model that is tightly interwoven with astrocytes. Dynamic cell assemblies express information about sensory cues. Each cell assembly, comprising pyramidal cells, small and large basket cells, and astrocytes, responds to one particular sensory cue provided as a sensory stimulus. Ambient GABA acts on receptors in membranes outside synapses and provides pyramidal cells with inhibitory currents in a tonic manner. We investigate how the neuron-astrocyte network works in STDP-based perceptual learning.

Here we show that the synchronization of local ambient GABA levels promotes coincidental pre- and postsynaptic spike generation in stimulus-sensitive pyramidal cells and facilitates STDP-based perceptual learning. This intracolumnar neuronal synchronization within the same cell assembly strengthens cross-columnar inhibition between different cell assemblies and suppresses stimulus-insensitive pyramidal cells, thereby enhancing STDP. Gap junctions between astrocytes work to synchronize local ambient GABA levels. We conclude that astrocytic gap junctional communication can improve perceptual learning by controlling extracellular GABA concentration.

2  Methods

2.1  Modeling of Networks of Coupled Neurons and Astrocytes

To seek a realistic model of neuron-astrocyte networks shown in Figure 1A, we assume these networks to be constituted by coupled cell assemblies (Buzsaki, 2010). Each assembly is formed by interconnected 20 cells units, and each of these units is made of four cells: a pyramidal cell (P), a large basket cell (), a small basket cell (), and an astrocyte (A). Within the same cell assembly (), each P cell receives excitatory inputs from other P cells and inhibitory inputs from cells. Each cell receives excitatory inputs from P cells belonging to different cell assemblies (). Each cell receives an excitatory input from its accompanying P cell. Each A cell instead receives input from both P and cells and is connected with other astrocytes by gap junctions. In the presence of a sensory cue (Feature n in Figure 1A), P cells receive a further excitatory current that mimics the stimulatory action of the cue.

Figure 1:

Neuronal architecture. (A) The neural network model containing cell assemblies (). Each cell assembly comprises pyramidal cells (P), small and large basket cells (, ), and astrocytes (A). Within cell assemblies, gap junctions connect neighboring A cells. The open and filled triangles denote excitatory and inhibitory synapses, respectively. A constant excitatory current is provided to P cells when presented with Feature as a cue input. (B) A schematic illustration of GABA transport by astrocytes. Transporters on an A cell import or export GABA molecules depending on astrocyte membrane potential. Ambient GABA molecules are accepted by extrasynaptic receptors and tonically inhibit a P cell.

Figure 1:

Neuronal architecture. (A) The neural network model containing cell assemblies (). Each cell assembly comprises pyramidal cells (P), small and large basket cells (, ), and astrocytes (A). Within cell assemblies, gap junctions connect neighboring A cells. The open and filled triangles denote excitatory and inhibitory synapses, respectively. A constant excitatory current is provided to P cells when presented with Feature as a cue input. (B) A schematic illustration of GABA transport by astrocytes. Transporters on an A cell import or export GABA molecules depending on astrocyte membrane potential. Ambient GABA molecules are accepted by extrasynaptic receptors and tonically inhibit a P cell.

Concerning -to-P and -to-A connections, a variety of GABAergic interneurons have been found in the cortex, including large and small basket cells (Wang, Gupta, Toledo-Rodriguez, Wu, & Markram, 2002; Markram et al., 2004). Large basket cells have broad axonal arborizations that can inhibit distant neurons, while small basket cells have limited axonal arborizations. We assume that the cell (as large basket cell) projects to all (nearby to distant) P cells and the cell (as small basket cell) projects to its proximal A cell within the same cell assembly.

-aminobutyric acid (GABA) is the major inhibitory neurotransmitter and mediates phasic inhibition by activating intrasynaptic GABA receptors—GABA receptors in the synaptic cleft. Tonic inhibition occurs when extracellular GABA activates receptors in membranes outside synapses (Semyanov, Walker, Kullmann, & Silver, 2004; Farrant & Nusser, 2005; Ortinski et al., 2006). GABA molecules in extracellular space and GABA receptors in extrasynaptic membrane regions are referred to as ambient GABA and extrasynaptic GABA receptors, respectively. Extrasynaptic receptors have been found in the cerebellum (Somogyi, Takagi, Richards, & Mohler, 1989; Nusser, Roberts, Baude, Richards, & Somogyi, 1995; Brickley, Cull-Candy, & Farrant, 1996; Soltesz & Nusser, 2001) and the cortex (Drasbek & Jensen, 2006; Scimemi et al., 2006).

In the brain, intrasynaptic GABA rises to a millimolar level triggered by a presynaptic action potential (Maconochie, Zempel, & Steinbach, 1994; Jones & Westbrook, 1995). In contrast, ambient GABA is maintained within a range of submicromolar to several micromolar levels (Lerma, Herranz, Herreras, Abraira, & Martin, 1986; Tossman, Jonsson, & Ungerstedt, 1986; Scimemi, Semyanov, Sperk, Kullmann, & Walker, 2005). The lower ambient GABA level is sufficient to activate extrasynaptic but not intrasynaptic receptors. receptors containing the subunit found in extrasynaptic membrane regions (Somogyi et al., 1989; Nusser et al., 1995; Brickley et al., 1996; Soltesz & Nusser, 2001) are known to have high affinity for GABA (Saxena & Macdonald, 1996; Brown, Kerby, Bonnert, Whiting, & Wafford, 2002) and little desensitization to continuous activation by GABA (Bianchim, Haas, & Macdonald, 2001, 2002). This leads to inhibition of neurons at even lower ambient GABA levels, as will be modeled in section 2.2 (see equation 2.6).

Glial membrane transporters, such as GAT-1, GAT-2 and GAT-3, are known to regulate neuronal activity by modulating ambient GABA levels (Barakat & Bordey, 2002; Koch & Magnusson, 2009; Eulenburg & Gomeza, 2010). Experimental and theoretical studies suggest that GABA transporters operate by pursuing an equilibrium point that depends on the transporter stoichiometry, the concentration gradients of substrates, and the astrocyte membrane potential (Richerson & Wu, 2003; Wu, Wang, & Richerson, 2003; Richerson, 2004; Wu, Wang, Diez-Sampedro, & Richerson, 2007). Under normal physiological conditions, a thermodynamic reaction cycle involves coupled translocation of two ions, one ion, and one uncharged GABA molecule. The co-transported molecules (2, , GABA) cross the membrane together. The driving force for the coupled transport is the electrochemical potential, which is the sum of the electropotential and the chemical potential.

The reversal potential of transporter is the equilibrium voltage of astrocytes at which the value of electrochemical potential is equal to 0. Under the normal physiological condition, the reversal potential of GABA transporter GAT-1 was estimated to be 67.16.47 mV for cultured neurons (Wu et al., 2007), which we assumed for astrocytes (Hoshino, 2012, 2014). At astrocyte membrane potentials below the reversal potential, net influx of GABA, called forward transport (i.e., GABA import), takes place. If the astrocyte membrane potential is above the reversal potential, net efflux of GABA, called reverse transport (i.e., GABA export), takes place.

To regulate local ambient GABA levels, we assume P-to-A excitatory synapses and -to-A inhibitory synapses. -to-A signaling increases stimulus-sensitive P cell activity by decreasing ambient GABA levels, while P-to-A signaling suppresses stimulus-insensitive P cell activity by increasing ambient GABA levels. The -to-A and P-to-A signaling improves the selective response property of the network to sensory cues, which is expected to benefit STDP-based perceptual learning. Although the GABAergic gliotransmission mechanism is still in debate, we control extracellular GABA concentration by astrocyte transporters, as will be modeled in section 2.4.

2.2  Neuronal Network Model

Dynamics of membrane potential of the th P, , and cells in cell assembly evolve according to equations 2.1 and 2.2. As soon as the membrane potential reaches a threshold value , a cell fires an action potential with probability (Anderson, Hu, Pittman, & Kiss, 2004) according to equation 2.3, during which we keep the membrane depolarized at = −10 mV for 1 ms. We also consider an absolute refractory period of 1 ms for all cells, immediately following an action potential.
formula
2.1
formula
2.2
formula
2.3
The input current to a P cell is the sum of both inhibitory currents from large basket cells () and extracellular GABA () and excitatory currents from recurrent connections with other P cells in the assembly () or by a direct sensory cue (). The input current to a cell is the current from a P cell (), and the input current to a cell is the current from P cells (). These currents are defined by
formula
2.4
formula
2.5
formula
2.6
formula
2.7
formula
2.8
where , , , and are P-to-P, -to-P, P-to-, and P-to- synaptic connection weights, respectively. is the number of extrasynaptic receptors embedded in P cell membrane. is the fraction of AMPA receptors in the open state triggered by presynaptic action potentials of the th P cell. is the fraction of intrasynaptic receptors in the open state triggered by presynaptic action potentials evoked in the th cell. is the fraction of extrasynaptic receptors, located on the th P cell, in the open state provoked by ambient GABA.
For sensory stimulation, we consider a gaussian-shaped input, which accounts for the fact that each cell assembly in our network has its own receptive field with respect to presented sensory cues (Hubel & Wiesel, 1962). In this fashion, we hypothesized that when the network is presented by a sensory cue characterized by a feature , the cell assembly receives the maximal input current (see equation 2.9, ), while the surrounding assemblies () are only marginally stimulated, with a current that decays with the (Hamming) distance from feature (see equation 2.9, ).
formula
2.9
In the previous equations of currents, receptors' dynamics is modeled as equations 2.10 and 2.11, where and are intrasynaptic glutamate and GABA concentrations released from cell type (), respectively, and is extracellular (ambient) GABA concentration. The intrasynaptic concentrations are modeled by Heaviside functions that are nonzero only during the brief window of neuronal firing (see equations 2.12 and 2.13):
formula
2.10
formula
2.11
formula
2.12
formula
2.13
where and are the quantal synaptic release of neurotransmitters (Destexhe, Mainen, & Sejnowski, 1998).

2.3  Astrocyte Network Model

Dynamics of membrane potential of the th A cell in cell assembly evolve according to equations 2.14 and 2.15. Differently from neurons, astrocytes do not fire action potentials, so their membrane potential is simply governed by a standard resistive capacitive equation of this type with current . Although there is no direct connection between neurons and astrocytes, these latter are stimulated by the activation of AMPA (Lalo, Pankratov, Parpura, & Verkhratsky, 2011), NMDA (Lalo, Pankratov, Kirchhoff, North, & Verkhratsky, 2006) and receptors (Losi et al., 2014) which are activated by neurotransmitter molecules spilled out of the synaptic cleft.
formula
2.14
formula
2.15
In this fashion, ensues from three components: a current caused by glutamate released by P cell synapses (see equation 2.16), and a GABAergic current due to spill out from cell synapses (; see equation 2.17). We do not consider -to-A inhibition, because it interferes with lateral inhibition by reducing ambient GABA concentrations around stimulus-insensitive P cells due to which sensory tuning performance deteriorates fatally. In addition, we also need to consider the fact that astrocytes may be stimulated by recurrent connections by gap junctions (; see equation 2.18).
formula
2.16
formula
2.17
formula
2.18
where index indicates the extent of connectivity. Houades, Koulakoff, Ezan, Seif, & Giaume, (2008) demonstrated that astrocyte connectivity is strong within the same somatosensory barrel but much weaker or absent across different barrels. Based on their study, we assume gap junction coupling between astrocytes in the same assembly but not across different cell assemblies. and are P-to-A and -to-A synaptic connection weights, respectively. is the fraction of intrasynaptic receptors in the open state triggered by presynaptic action potentials evoked in the th cell.

2.4  Neuron-Astrocyte Coupling by Extracellular GABA

The mechanism of regulation of extracellular GABA is schematized in Figure 1B. Astrocytic GABA transporters are responsible for extracellular GABA levels. As mentioned earlier, these transporters can either remove from or add GABA to the extracellular space depending on the value of the astrocyte membrane potential with respect to the transporter reversal potential. Generally hyperpolarizations of the astrocytes such as those due to cells tend to promote extracellular GABA removal by the astrocyte, whereas P cell–mediated depolarizations could promote reverse operation of the transporter, making the astrocyte release GABA into the extracellular space (Wu et al., 2007). We assume that only P cells express extrasynaptic GABA receptors that can be activated by extracellular GABA. This mimics experimental conditions often implemented in the study of STDP in slices (Bright & Smart, 2013) and eases our aim of characterizing the effect of astrocytic gap junction communication on STDP-based perceptual learning.

If basket cells were also affected by ambient GABA, in fact, we could expect that the ensuing reduction of the inhibitory action of these cells on P cells would be detrimental for the network selective and coherent response to sensory cues, ultimately disturbing the learning of these sensory cues by STDP. Finally, we make the simplifying assumption that GABA exchange between cell assemblies is negligible because astrocyte connectivity is quite weak or absent across different barrels in somatosensory cortex (Houades et al., 2008).

Therefore, extracellular GABA concentration within a cell assembly evolves according to equation 2.19:
formula
2.19
where and are a decay constant and the basal ambient GABA concentration, respectively. determines the modulation rate of ambient GABA concentration. and restrict ambient GABA concentration to a maximum and a minimum, respectively. is the reversal potential of the transporter.

2.5  Spike-Timing-Dependent Plasticity

Learning is assumed to occur only on excitatory synapses by any sensory cues out of cues in total. In particular, synaptic weights between P cells change according to a classic STDP paradigm (Hoshino, 2011, 2015),
formula
2.20
where and denote synaptic decay rate and synaptic modification rate, respectively; is a built-in synaptic weight (); and the kernel , which denotes the change of synaptic weight depending on the timing of pre/post spikes (Abbott & Nelson, 2000; Bi & Poo, 1998).
formula
2.21
where and denote potentiation rate and depression rate, respectively. and are potentiation time constant and depression time constant, respectively.

2.6  Parameter Estimation, Numerical Methods, and Data Analysis

The astrocytic capacitance (, equation 2.14) value has been estimated by electrophysiological experiments and ranges between 5 pF and 40 pF, in association with a membrane conductance (, equation 2.14) between 50 nS and 90 nS (Amzica & Neckelmann, 1999). The astrocyte resting potential (, equation 2.14) is reported in the range of 70 to 90 mV (Erlichman, Cook, Schwab, Budd, & Leiter, 2004). The typical conductance value of gap junction channels connecting astrocyte pairs (, equation 2.18) was estimated to be nS (Dermietzel, Hertberg, Kessler, & Spray, 1991). Values of all model parameters used in the simulations of this study are summarized in Table 1.

Table 1:
List of Parameters and Their Values.
SymbolDescriptionValueUnit
 Membrane capacitance for cell type ( P, B B, A)  pF 
 Membrane conductance for cell type ( P, B B, A)  nS 
 Resting potential for cell type ( P, B B, A)  mV 
 Maximal conductance for receptor type Z (Z AMPA, GABA)  nS 
 Maximal conductance for gap junction 20 nS 
 Reversal potential for receptor type Z (Z AMPA, GABA)  mV 
 Number of cell-units within cell assemblies 20  
 Number of cell assemblies  
 Synaptic weight (strength) from to th P cell in cell assembly    
 Synaptic weight from th B to th P cell in cell assembly   
 Synaptic weight from th P to B cell in cell assembly  60  
 Synaptic weight from th P to B cell between different cell assemblies  
 Synaptic weight from th P to A cell between different cell assemblies 0.4  
 Synaptic weight from th B to A cell in cell assembly   
 Amount of extrasynaptic GABA receptors on P cell   
 Sensory input current 600 pA 
 Input broadness  
 Channel opening rate for receptor type Z (Z AMPA, GABA)  Ms 
 Channel closing rate for receptor type Z (Z AMPA, GABA)  s 
 Steepness of sigmoid function for cell type ( P, B B  mV 
 Threshold of sigmoid function for cell type ( P, B B  mV 
 Extent of connectivity  
 Decay constant for ambient GABA concentration 10 s 
 Basal ambient GABA concentration 
 Maximal ambient GABA concentration 
 Minimal ambient GABA concentration 
 GABA transfer coefficient MmV s 
 Reversal potential of GABA transporter  mV 
 Maximal glutamate concentration for presynaptic cell P mM 
 Maximal GABA concentration for presynaptic cell type ( B B  mM 
 Synaptic decay rate   
 Synaptic modification rate  
 Built-in synaptic weight 2.5  
 Potentiation rate  
 Depression rate 0.5  
 Potentiation time constant 10 ms 
 Depression time constant 30 ms 
SymbolDescriptionValueUnit
 Membrane capacitance for cell type ( P, B B, A)  pF 
 Membrane conductance for cell type ( P, B B, A)  nS 
 Resting potential for cell type ( P, B B, A)  mV 
 Maximal conductance for receptor type Z (Z AMPA, GABA)  nS 
 Maximal conductance for gap junction 20 nS 
 Reversal potential for receptor type Z (Z AMPA, GABA)  mV 
 Number of cell-units within cell assemblies 20  
 Number of cell assemblies  
 Synaptic weight (strength) from to th P cell in cell assembly    
 Synaptic weight from th B to th P cell in cell assembly   
 Synaptic weight from th P to B cell in cell assembly  60  
 Synaptic weight from th P to B cell between different cell assemblies  
 Synaptic weight from th P to A cell between different cell assemblies 0.4  
 Synaptic weight from th B to A cell in cell assembly   
 Amount of extrasynaptic GABA receptors on P cell   
 Sensory input current 600 pA 
 Input broadness  
 Channel opening rate for receptor type Z (Z AMPA, GABA)  Ms 
 Channel closing rate for receptor type Z (Z AMPA, GABA)  s 
 Steepness of sigmoid function for cell type ( P, B B  mV 
 Threshold of sigmoid function for cell type ( P, B B  mV 
 Extent of connectivity  
 Decay constant for ambient GABA concentration 10 s 
 Basal ambient GABA concentration 
 Maximal ambient GABA concentration 
 Minimal ambient GABA concentration 
 GABA transfer coefficient MmV s 
 Reversal potential of GABA transporter  mV 
 Maximal glutamate concentration for presynaptic cell P mM 
 Maximal GABA concentration for presynaptic cell type ( B B  mM 
 Synaptic decay rate   
 Synaptic modification rate  
 Built-in synaptic weight 2.5  
 Potentiation rate  
 Depression rate 0.5  
 Potentiation time constant 10 ms 
 Depression time constant 30 ms 

Unitless.

The neural network model contains eight cell assemblies, each cell assembly consists of 20 cell units, and each cell unit comprises 1 P cell, 1 cell, 1 cell, and 1 A cell. Connection structures were uniform within (P-to-P, P-to-, -to-P, -to-A) and between (P-to-, P-to-A) cell assemblies. There were 8 sensory cues (Feature : ). The mechanism of recurrent excitation within cell assemblies and lateral inhibition across cell assemblies is known to function in cortical sensory tuning in vision, audition, and somatosensation (Isaacson & Scanziani, 2011).

Model equations were solved by Euler numerical integration with step 0.1 ms. In all our simulations, the Fano factor (FF) was computed out of 20 ambient GABA concentration values. Cross-correlations (CCs) were computed for raw spike trains and ambient GABA concentration values off their ensemble mean between 10 pairs because more than 10 pairs were mandatory in statistically significant analysis. Both FF and CCs were computed considering a 3 s sliding window. Statistical analysis was by paired testing, and statistical significance was set for (asterisks in figures).

3  Results

3.1  Astrocytic Gap-Junction Communication Improving STDP-Based Perceptual Learning

To assess the effect of gap junctions (GJs) between astrocytes on learning, we investigated how neuronal firing activity at the core of STDP changes in the presence or the absence of GJ communication. The absent GJ communication mimics experimental application of typical GJ blockers like heptanol, octanol, carbenoxolone, flufenamic acid, and mefloquine (Scemes & Giaume, 2006). Accordingly, we consider two scenarios: a control scenario where GJ communication is present ( = 20 nS; see Figure 2A) and the opposite scenario where GJs are blocked so that no astrocyte coupling exists ( = 0 nS; see Figure 2B).

Figure 2:

Change of synaptic weights by STDP. (A) Top: Raster plots of action potentials evoked in P cells belonging to different cell assemblies (). Middle: Cell membrane potentials. Bottom: Ambient GABA concentrations around P cells. In the simulation, one trial was carried out, which included one learning period (500 ms, left horizontal bars) and one perceptual period (500 ms, right horizontal bar). (B) Those obtained in the absence of GJ communication ( nS). (C) Top: Stimulus-evoked P cell activity () after learning. Bottom: Synaptic connection weights between P cells () after learning. The network performed the same task for 20 trials. (D) Top: Spike counts recorded from 20 stimulus-sensitive ( = 4) P cells during the learning period where the GJ communication existed (i.e., = 20 nS) or not (i.e., = 0 nS). A sliding 10 ms window was used for counting. Middle: Cumulative spike counts. Bottom: Time-resolved changes in synaptic weights.

Figure 2:

Change of synaptic weights by STDP. (A) Top: Raster plots of action potentials evoked in P cells belonging to different cell assemblies (). Middle: Cell membrane potentials. Bottom: Ambient GABA concentrations around P cells. In the simulation, one trial was carried out, which included one learning period (500 ms, left horizontal bars) and one perceptual period (500 ms, right horizontal bar). (B) Those obtained in the absence of GJ communication ( nS). (C) Top: Stimulus-evoked P cell activity () after learning. Bottom: Synaptic connection weights between P cells () after learning. The network performed the same task for 20 trials. (D) Top: Spike counts recorded from 20 stimulus-sensitive ( = 4) P cells during the learning period where the GJ communication existed (i.e., = 20 nS) or not (i.e., = 0 nS). A sliding 10 ms window was used for counting. Middle: Cumulative spike counts. Bottom: Time-resolved changes in synaptic weights.

The network was stimulated in two consecutive periods. In the first, “training” period, a sensory cue (Feature 4) was presented, allowing excitatory synaptic weights to change by STDP (see equations 2.20 and 2.21). Firing activity, astrocyte membrane potentials, and local (with respect to the cell assembly) ambient GABA concentrations are shown in the top, middle, and bottom panels respectively. During the second, “input” period, the same cue was presented to the network, but this time STDP was blocked, allowing us to measure extracellular GABA in an assembly as a result of previous learning. In this regard it may be noted that ambient GABA decreases following learning (see Figure 2A, bottom, = 4), and this results in a reduction in tonic inhibition and thus promotes the firing of stimulus-sensitive P cells.

To see how the GJ conductance affects the firing activity and synaptic weight, we ran a simulation in which the GJ conductance was varied between 0 nS and 24 nS. As shown in Figure 2C, the stimulus-evoked firing activity (top) and synaptic weight (bottom) are increased as the GJ conductance increases, indicating that the GJ communication is beneficial for STDP. To see their temporal aspects, we recorded spikes and weights during the learning period, which are shown in Figure 2D. Spike counts in a 10 ms sliding window are shown in the top panel, their cumulative representations are shown in the middle panel, and time-resolved changes in mean weights are shown in the bottom panel. These results indicate that the synaptic weight is gradually increased as learning proceeds and that the enhancement of synaptic strength is greater when the GJ communication exists ( = 20 nS) than when it does not ( = 0 nS).

To see how the GJ communication affects neuronal activity, we computed cross-correlation functions between pairs of local ambient GABA concentrations and pairs of spike trains and averaged over them, which are shown in Figure 3A. Less synchrony in ambient GABA levels (left column, = 0 nS) results in weak synchrony in P cell activity (right column, = 0 nS). The GJ communication synchronizes ambient GABA levels (left column, = 10, 20 nS), leading to strong synchrony in P cell activity (right column, = 10, 20 nS). Notably, the GJ communication does not increase the firing rate (panel B, left). We analyzed the input-output relationship of P cells (right), where the GJ communication existed (filled circles) or not (open circles). Synaptic connection weights were the same. We found that the two input-output curves are nearly identical and confirmed no significant difference between them by paired -test. These results suggest that the potentiation of synaptic weights could result from synchrony in spiking activity mediated by local ambient GABA level synchronization but not from an increase in firing rate.

Figure 3:

Control of local ambient GABA concentrations by GJ communication promotes neuronal synchrony. (A) Cross-correlation functions between local ambient GABA concentrations (left) and between spike trains evoked in P cells (right) recorded from the stimulus-sensitive cell assembly ( = 4). was changed between 0 nS and 20 nS. (B) Left: Dependence of firing rate on (meanSD). Right: Input-output relationship of a stimulus-sensitive P cell before learning. GJs existed (filled) or not (open).

Figure 3:

Control of local ambient GABA concentrations by GJ communication promotes neuronal synchrony. (A) Cross-correlation functions between local ambient GABA concentrations (left) and between spike trains evoked in P cells (right) recorded from the stimulus-sensitive cell assembly ( = 4). was changed between 0 nS and 20 nS. (B) Left: Dependence of firing rate on (meanSD). Right: Input-output relationship of a stimulus-sensitive P cell before learning. GJs existed (filled) or not (open).

To understand how the GJ conductance affects A cell activity and extracellular GABA concentration, we recorded membrane potentials and local ambient GABA concentrations, where the GJ communication existed (see Figure 4A, = 20 nS) or not (see Figure 4B, = 0 nS). The variability of local ambient GABA concentrations is evaluated by the variance (middle) and the Fano factor (bottom) in Figure 4C, where the GJ communication existed ( = 20 nS, solid) or not ( = 0 nS, dashed). It may be noted that GJ communication correlates with a reduced variability of extracellular GABA concentration as can be seen in the bottom panel where was set at 20 nS (solid traces).

Figure 4:

Dynamic and statistical nature of astrocyte membrane potential and local ambient GABA concentration. (A) GJ communication existed ( = 20 nS). (B) GJ communication did not exist ( = 0 nS). (C) Mean (top), variance (middle), and Fano factor (variance/mean, bottom) of local ambient GABA concentrations shown in panels A and B recorded in the stimulus-sensitive cell assembly = 4. Solid and dashed traces respectively denote values in the presence of and absence of GJ communication.

Figure 4:

Dynamic and statistical nature of astrocyte membrane potential and local ambient GABA concentration. (A) GJ communication existed ( = 20 nS). (B) GJ communication did not exist ( = 0 nS). (C) Mean (top), variance (middle), and Fano factor (variance/mean, bottom) of local ambient GABA concentrations shown in panels A and B recorded in the stimulus-sensitive cell assembly = 4. Solid and dashed traces respectively denote values in the presence of and absence of GJ communication.

To see how astrocytic GJs affect such variability of ambient GABA, in Figure 5A we show the mean (top), variance (middle), and Fano factor (bottom) of ambient GABA concentrations recorded in different cell assemblies () as a function of (see equation 2.18). It may be noted that while the mean ambient GABA concentration is left almost unchanged, the decrease of variance of ambient GABA is dramatic. It reflects an overall synchronization of ambient GABA levels around different P cells in the same assembly and explains why the FF is smaller for larger values of .

Figure 5:

Relevance of variability of ambient GABA concentration in STDP. (A) Dependence of the mean (top), variance (middle), and Fano factor (bottom) of local ambient GABA concentrations recorded in each cell assembly on GJ conductance (values are time-average representations of those in Figure 4C). (B) Top: Dependence on of synaptic weight after learning. Bottom: Relation between synaptic weight change and Fano factor.

Figure 5:

Relevance of variability of ambient GABA concentration in STDP. (A) Dependence of the mean (top), variance (middle), and Fano factor (bottom) of local ambient GABA concentrations recorded in each cell assembly on GJ conductance (values are time-average representations of those in Figure 4C). (B) Top: Dependence on of synaptic weight after learning. Bottom: Relation between synaptic weight change and Fano factor.

Figure 5B further shows how such reduction in GABA variability could be beneficial for STDP-mediated synaptic potentiation (i.e., LTP), where it is seen that the increase of GJ conductance leads to an increase ( = 4) or a decrease ( = 3, 5) in synaptic weight (top panel). This suggests that LTP is strengthed in the stimulus-sensitive cell assembly ( = 4) while it is weakened in stimulus-insensitive cell assembles ( = 3, 5). Plotting the mean synaptic weight for different Fano factor values (bottom) from Figure 5A also reveals that such LTP strengthening and weakening correlate with the variability of ambient GABA concentrations. In the presence of GJ communication, cross-columnar interaction strengthens LTP ( = 4) by lateral inhibition between cell assemblies. On the contrary, the cross-columnar suppression, which is enhanced by the increase of neuronal synchronization via astrocytic GJ communication (as revealed by the reduction of FFs in Figure 5A), results in a decreased firing rate in those columns, thereby weakening LTP there ( = 3, 5).

3.2  Influence of Cross-Columnar Interaction

We next asked how the cross-columnar interaction affects STDP-based perceptual learning. To address this, we implemented learning in the absence of P-to- and P-to-A cross-columnar projections (see Figure 6A). As shown in Figure 6B, the strength of GJ conductance does not contribute to neuronal firing insofar as the average firing rate is essentially independent of as in the presence of cross-columnar interactions (see Figure 3B). Yet what dramatically changes are P and A cell activities and local extracellular GABA (see Figure 6C) as it may be noted that the selective responsiveness of P cells is deteriorated (right, top panel); A cell activities (right, middle panel) and ambient GABA concentrations (right, bottom panel) do not synchronize, which will be clearly indicated by cross-correlation analyses in Figure 6E.

Figure 6:

Influence of cross-columnar interaction on STDP-based perceptual learning. (A) Neuron-astrocyte circuitry where P-to-A and P-to- projections were cut. (B) Dependence of firing rate on (meanSD). (C) Analysis of network activity in the presence (left) or absence of GJ communication (right). Raster plots of action potentials evoked in P cells belonging to different cell assemblies () (top), A cell membrane potentials recorded in the stimulus-sensitive cell assembly (middle, = 4), and ambient GABA concentrations around P cells recorded in the stimulus-sensitive cell assembly (bottom, = 4). A learning process took place during the input period. (D) Top: Input-output relationship of a stimulus-sensitive P cell after learning. The network performed the same task for 20 trials, where = 20 nS (filled circles) or = 0 nS (open circles). An asterisk indicates statistical significance (). Bottom: Synaptic connection weights between P cells (meanSD) after learning. (E) Cross-correlation functions between local ambient GABA concentrations (left) and between spike trains evoked in P cells (right) recorded in the stimulus-sensitive cell assembly ( = 4). was changed between 0 nA and 20 nS. Values were averaged out of 10 cell pairs.

Figure 6:

Influence of cross-columnar interaction on STDP-based perceptual learning. (A) Neuron-astrocyte circuitry where P-to-A and P-to- projections were cut. (B) Dependence of firing rate on (meanSD). (C) Analysis of network activity in the presence (left) or absence of GJ communication (right). Raster plots of action potentials evoked in P cells belonging to different cell assemblies () (top), A cell membrane potentials recorded in the stimulus-sensitive cell assembly (middle, = 4), and ambient GABA concentrations around P cells recorded in the stimulus-sensitive cell assembly (bottom, = 4). A learning process took place during the input period. (D) Top: Input-output relationship of a stimulus-sensitive P cell after learning. The network performed the same task for 20 trials, where = 20 nS (filled circles) or = 0 nS (open circles). An asterisk indicates statistical significance (). Bottom: Synaptic connection weights between P cells (meanSD) after learning. (E) Cross-correlation functions between local ambient GABA concentrations (left) and between spike trains evoked in P cells (right) recorded in the stimulus-sensitive cell assembly ( = 4). was changed between 0 nA and 20 nS. Values were averaged out of 10 cell pairs.

If there is a cross-columnar effect on learning, we may expect that this is reflected by the ensuing receptive field of individual P cells with respect to stimulation. To this extent, when the network was repeatedly stimulated and GJ communication blocked, plotting of firing rates (see Figure 6D, top) as a function of input current reveals its shift (open circles), implying a decrease in neuronal sensitivity to the input. The bottom panel of Figure 6D shows that the sensitivity decrease is due to a decrease in synaptic weight (open circle at = 4). To see how the GJ communication promotes the neuronal synchronization, we computed cross-correlation functions between pairs of local ambient GABA concentrations and pairs of spike trains, and averaged over them; they are shown in Figure 6E. Less synchrony in ambient GABA levels (left column, = 0 nS) results in weak synchrony in P cell activity (right column, = 0 nS). The GJ communication synchronizes ambient GABA levels (left column, = 10, 20 nS), leading to strong synchrony in P cell activity (right column, = 10, 20 nS). These results suggest that the neuronal synchronization mediated by GJ communication strengthens LTP.

As previously noted in section 3.1, the synchronization of ambient GABA levels can be revealed by the analysis of mean and variance of GABA concentration surrounding different P cells. Figure 7A shows the mean (top), variance (middle), and Fano factor (bottom) of local ambient GABA concentrations around stimulus-sensitive ( = 4) P cells, which is compared to those obtained under the intact condition (see Figure 7B). The elimination of the cross-columnar interaction decreases the mean concentration. The blocking of GJ communication ( = 0 nS) results in an increase in the variability of local ambient GABA concentrations, which is not significantly influenced by the cross-columnar interaction. This result points to the fact that GJ communication is an important factor for the reduction of GABA variability.

Figure 7:

Influence of cross-columnar interaction on variability of ambient GABA concentration. (A) Mean (top), variance (middle), and Fano factor (bottom), of local ambient GABA concentrations around stimulus-sensitive ( = 4) P cells. The solid and dashed traces denote = 20 nS and = 0 nS, respectively. The cross-columnar interaction was eliminated by cutting P-to-A and P-to- projections. (B) Those obtained under the original condition, that is, the cross-columnar interaction existed. (C) Phase relationship between P and A cell activities. Membrane potentials of P (top), (second), A (third) cells, and ambient GABA concentrations (bottom) were recorded in the stimulus-sensitive cell assembly ( = 4). The GJ communication existed (left, = 20 nS) or not (right, = 0 nS), where the cross-columnar interaction was eliminated by cutting P-to-A and P-to- projections. (D) Average cross-correlation function between membrane potentials out of 10 pairs of P and A cells. The solid and dashed traces denote = 20 nS and = 0 nS, respectively. They were computed after subtracting the mean.

Figure 7:

Influence of cross-columnar interaction on variability of ambient GABA concentration. (A) Mean (top), variance (middle), and Fano factor (bottom), of local ambient GABA concentrations around stimulus-sensitive ( = 4) P cells. The solid and dashed traces denote = 20 nS and = 0 nS, respectively. The cross-columnar interaction was eliminated by cutting P-to-A and P-to- projections. (B) Those obtained under the original condition, that is, the cross-columnar interaction existed. (C) Phase relationship between P and A cell activities. Membrane potentials of P (top), (second), A (third) cells, and ambient GABA concentrations (bottom) were recorded in the stimulus-sensitive cell assembly ( = 4). The GJ communication existed (left, = 20 nS) or not (right, = 0 nS), where the cross-columnar interaction was eliminated by cutting P-to-A and P-to- projections. (D) Average cross-correlation function between membrane potentials out of 10 pairs of P and A cells. The solid and dashed traces denote = 20 nS and = 0 nS, respectively. They were computed after subtracting the mean.

We next investigated why neuronal firing does not decrease by blocking the GJ communication ( = 0 nS), which was shown in Figures 3B and 6B. We focused on the hyperpolarization of A cells by cell single spikes. The strong transient hyperpolarization due to the blockade of GJ, shown in the right panel of Figure 7C (third row), promotes a reduction in extracellular GABA concentration (bottom row) and thus excites P cells (top row). The activation of P cells leads to an increase in the probability of action potential generation, thereby keeping firing nearly at the same rate as in the presence of GJ communication (left, top row).

To see how P and A cells interact, we computed cross-correlation functions between pairs of their membrane potentials in the stimulus-sensitive cell assembly (panel C, = 4), which are shown in Figure 7D. The GJ communication existed (solid trace) or not (dashed trace). Presynaptic P cell firing triggers hyperpolarization in the A cell (filled arrows). The A cell hyperpolarization decreases ambient GABA level and triggers P cell firing (open arrow). Notably, the A cell is strongly hyperpolarized (see Figure 7C, third trace in the right column) when the gap-junction communication does not exist, leading to a strong negative correlation between P and A cells (arrows). The apparent stronger synchrony observed for nS might arise from such a stronger negative correlation.

We showed in Figures 6 and 7 how the intracolumnar interaction contributed to perceptual learning. We next investigated how these intra- and cross-columnar interactions affect STDP-mediated synaptic potentiation (i.e., LTP). The top panel of Figure 8 shows weight changes resulting from intracolumnar neuronal synchronization mediated by GJs. The middle and bottom panels show weight changes resulting from tonic and phasic cross-columnar interactions mediated by P-to-A and P-to- projections, respectively. These results suggest that the tonic lateral inhibition between cell assemblies (middle) contributes to the synaptic enhancement as well ( = 4). The phasic lateral inhibition (bottom) prohibits the synaptic enhancement for stimulus-insensitive cell assemblies (e.g., = 3 and 5), which is necessary for the network to tune to the sensory input. Without this inhibition, not only stimulus-sensitive but also stimulus-insensitive P cells tend to respond (see Figure 6C).

Figure 8:

Influence of intra- and cross-columnar interactions on perceptual learning. Top: Weight changes resulting from intracolumnar neuronal synchronization mediated by gap-junctional communication. Middle: Weight changes resulting from tonic cross-columnar inhibition mediated by P-to-A projections. Bottom: Weight changes resulting from phasic cross-columnar inhibition mediated by P-to- projections.

Figure 8:

Influence of intra- and cross-columnar interactions on perceptual learning. Top: Weight changes resulting from intracolumnar neuronal synchronization mediated by gap-junctional communication. Middle: Weight changes resulting from tonic cross-columnar inhibition mediated by P-to-A projections. Bottom: Weight changes resulting from phasic cross-columnar inhibition mediated by P-to- projections.

In this section, we showed that the potentiation of synaptic weights could result from synchrony in spiking activity mediated by ambient GABA-level synchronization, which is partially reflected in the reduced variability of extracellular GABA concentration. As shown in Figure 6C, the sensory input broadness (; see equation 2.9) affects responses of P cells and thus presumably LTP. This will be investigated in section 3.3 in association with astrocyte connectivity.

3.3  Influences of Sensory Input Broadness and Astrocyte Connectivity

We next asked how the input broadness (; see equation 2.9) affects STDP-based perceptual learning. To address this, we stimulated the network repeatedly with different input current distributions by varying and implemented multiple learning trials. In each trial, the same stimulus was presented at a random stimulus-onset time. Each panel of Figures 9A and 9B shows eight sets of mean weights overlaid for individual cell assemblies, where the GJ communication existed (panel A) or not (panel B). Changes in synaptic connection weights caused by GJ communication are shown in Figure 9C. Positive values indicate that astrocytic GJ communication strengthens LTP, while negative values stand for weakening LTP. It may be noted that LTP is strengthened in the stimulus-sensitive cell assembly ( = 4) and weakened in stimulus-insensitive cell assemblies ( = 3, 5). The LTP weakening is remarkable when the stimulus is broad (e.g., ).

Figure 9:

Influence of input broadness on perceptual learning. (A) Synaptic connection weights between P cells (meanSD) after learning with GJ communication (i.e., = 20 nS). The input broadness was varied from 1 (top) to 5 (bottom). We implemented 10 trials. (B) Synaptic connection weights after learning without GJ communication (i.e., = 0 nS). (C) Changes in learning-modified synaptic connection weights (meanSD) caused by GJ communication, which are defined by . An asterisk indicates statistical significance (). (D) Synaptic weight changes caused by GJs as a function of in stimulus-sensitive cell assembly = 4.

Figure 9:

Influence of input broadness on perceptual learning. (A) Synaptic connection weights between P cells (meanSD) after learning with GJ communication (i.e., = 20 nS). The input broadness was varied from 1 (top) to 5 (bottom). We implemented 10 trials. (B) Synaptic connection weights after learning without GJ communication (i.e., = 0 nS). (C) Changes in learning-modified synaptic connection weights (meanSD) caused by GJ communication, which are defined by . An asterisk indicates statistical significance (). (D) Synaptic weight changes caused by GJs as a function of in stimulus-sensitive cell assembly = 4.

To clearly see that LTP nonmonotonically (i.e., nonlinearly) depends on input broadness, we plotted the change of synaptic weight as a function of , which is shown in Figure 9D. This result suggests that the GJ communication contributes to strengthening LTP in stimulus-sensitive cell assembly ( = 4) when the stimulus is broad (), for which the weakening of LTP in stimulus-insensitive cell assemblies might be crucial (panel C, = 4, = 3, 5). Noteworthy is that this LTP weakening is not observed if the stimulus is too broad (panel C, = 5, = 3), because the broad input activates not only stimulus-sensitive but also stimulus-insensitive P cells.

To show the significance of these weight changes in network performance, we evaluated the gain function (input-output relation) of P cells, where the input broadness was varied between 1 and 4 (see Figure 10A). The number of activated P cells is limited (i.e., local activation) when the stimulus is narrow (e.g., ). In contrast, the activation of P cells is diffused rather than selective (e.g., ). These gain functions guarantee the robustness of the network model to sensory stimulation. Their percentage changes caused by GJ communication shown in Figure 10B indicate that the network sensitivity to broad sensory input () can be increased.

Figure 10:

Influence of weight change on network performance. (A) Input-output relationship of a stimulus-sensitive P cell after learning, where the input broadness was varied between 1 and 4. The network performed the same task for 20 trials, where = 20 nS (filled circles) or = 0 nS (open circles). An asterisk indicates statistical significance (). (B) Percentage increase of stimulus-evoked P cell activity by GJ communication. (C) Changes of stimulus-evoked neuronal activity caused by GJs as a function of in stimulus-sensitive cell assembly = 4. The input current was too small (top) or large (bottom).

Figure 10:

Influence of weight change on network performance. (A) Input-output relationship of a stimulus-sensitive P cell after learning, where the input broadness was varied between 1 and 4. The network performed the same task for 20 trials, where = 20 nS (filled circles) or = 0 nS (open circles). An asterisk indicates statistical significance (). (B) Percentage increase of stimulus-evoked P cell activity by GJ communication. (C) Changes of stimulus-evoked neuronal activity caused by GJs as a function of in stimulus-sensitive cell assembly = 4. The input current was too small (top) or large (bottom).

To clearly indicate the nonmonotonic dependence of these gain functions, we plotted stimulus-evoked neuronal activity as a function of , shown in Figure 10C. The input current was varied between 300 pA (top) and 600 pA (bottom). We found an increase in neuronal activity by GJs at a large value (), which is maximal for small input current (400 pA, second panel). This result suggests that the GJ communication contributes to increasing the sensitivity of the network to low intensity and broad sensory input. This might be beneficial in catching weak sensory signals in noisy environments.

Finally, we look at how astrocyte connectivity contributes to synchronization of local ambient GABA levels, shown in Figure 11. In each cell assembly, the th A cell () was not connected to any A cell ( = 0) but to two neighboring (i 1 and i 1) A cells ( = 1) or eight (i 4, i 3, i 2, i 1, i 1, i 2, i 3, i 4) A cells ( = 4). In the stimulus-sensitive cell assembly ( = 4), local ambient GABA levels completely synchronize when these A cells form a dense network ( = 4). Dependence of the time-averaged mean (top), variance (middle), and Fano factor (bottom) of local ambient GABA concentrations on connectivity is shown in Figure 11B, indicating that the variability is reduced as the number of GJ connection sites increases (i.e., as increases). The relationship between mean weight and shown in panel C (top) suggests that the astrocytic network strengthens LTP, and panel C (bottom) reveals correlations of LTP strengthening ( = 4) and weakening ( = 3, 5) to the Fano factor. The sharp transition from = 0 to = 1 indicates that a sparse astrocyte connectivity pattern is enough to synchronize local ambient GABA levels (panel A), reduce variability in ambient GABA concentration (panel B), and strengthen LTP (panel C).

Figure 11:

Influence of astrocyte connectivity on perceptual learning. (A) Cross-correlation functions between eight pairs of local ambient GABA concentrations in the stimulus-sensitive cell assembly ( = 4) for three different connectivities: unconnected ( = 0), two connections ( = 1) and eight connections ( = 4). (B) Dependence of the mean (top), variance (middle), and Fano factor (bottom) of local ambient GABA concentrations on . (C) Top: Dependence of learning-modified synaptic connection weight on . Bottom: Relation between synaptic weight and Fano Factor.

Figure 11:

Influence of astrocyte connectivity on perceptual learning. (A) Cross-correlation functions between eight pairs of local ambient GABA concentrations in the stimulus-sensitive cell assembly ( = 4) for three different connectivities: unconnected ( = 0), two connections ( = 1) and eight connections ( = 4). (B) Dependence of the mean (top), variance (middle), and Fano factor (bottom) of local ambient GABA concentrations on . (C) Top: Dependence of learning-modified synaptic connection weight on . Bottom: Relation between synaptic weight and Fano Factor.

4  Discussion

Perceptual learning refers to our ability to improve perception of external sensory stimuli as we experience the same stimulus repeatedly. It is a shared belief that the neural correlate of perceptual learning is the change of strength of synaptic connections between neurons by spike-timing-dependent (Hebbian) plasticity (STDP) (Gilbert, 1994, 1996). Precise manipulation in physiological experimental settings allows us to generate coincident firing of neurons, by which STDP can be produced easily. However, in behaving animals, this might not be sufficient insofar as neurons are subjected to a variety of inputs in addition to sensory cues, and their firing activity is often stochastic (Softky & Koch 1993). Consequently, a topic of active research is to understand how the brain develops neural synchrony to deploy STDP.

To address this, we have proposed a novel neuron-astrocyte network model and studied the effect of astrocytic gap-junctional communication on perceptual learning. The neural network was tightly interwoven with astrocytes. Astrocyte transporters regulated local ambient (extracellular) GABA levels. Ambient GABA molecules acted on receptors in membranes outside synapses and provided pyramidal cells with inhibitory currents in a tonic manner. We showed that the synchronization of local ambient GABA levels promoted coincidental pre- and postsynaptic spike generation in stimulus-sensitive pyramidal cells and facilitated STDP-based perceptual learning. This intracolumnar neuronal synchronization within the same cell assembly strengthened cross-columnar inhibition between different cell assemblies and suppressed stimulus-insensitive pyramidal cells, thereby enhancing STDP. These results indicate the importance of GABAergic signaling in neuronal synchronization and perceptual learning. We also showed that gap junctions between astrocytes worked to synchronize local ambient GABA levels, indicating the importance of astrocytic signaling and connectivity in orchestrating the GABAergic network dynamics.

Electrophysiological recordings from brain slices demonstrated that astrocytes are highly coupled with each other by gap junctions (Rouach, Koulakoff, & Giaume, 2004). Poskanzer and Yuste (2011, 2016) demonstrated that astrocyte networks played a causal role in regulating synchronized activation of neuronal ensembles in the cortex. Lee et al. (2014) demonstrated that glutamate release from astrocytes contributed to maintaining cortical -oscillations. Pannasch et al. (Pannasch et al., 2011; Pannasch & Rouach, 2013) demonstrated that astrocyte networks modulated neural network activity by intercellular exchange of ions, metabolites, and neuroactive substances, and astrocytic gap junctions played an important role in synaptic plasticity. Perisynaptic astrocyte processes modulate synaptic transmission both pre- and postsynaptically (De Pittá & Brunel, 2016; De Pittá, Brunel, & Volterra, 2016). The researchers demonstrated that activity-dependent glutamate release from astrocytes could change synaptic plasticity. Our study highlights a possible important role of astrocytic gap junctions in STDP: development of neuronal synchrony to deploy STDP by synchronizing extracellular GABA levels.

Experimental studies (Houades et al., 2008; Giaume, Koulakoff, Roux, Holcman, & Rouach, 2010) revealed nearest-neighbor connections in a barrel cortex astrocytic network. Simulation studies (Lallouette, De Pitt, Ben-Jacob, & Berry, 2014; Wallach et al., 2014) demonstrated that short-distance connections enhanced intercellular calcium wave (ICW) propagation in an astrocyte network. Such ICW propagation is known to trigger the release of signaling molecules from astrocytes (e.g., glutamate, ATP, and D-serine), thereby modulating neuronal activity and synaptic plasticity. It is noteworthy that astrocytes can control ambient (extracellular) GABA concentration, thereby modulating neuronal circuits in a tonic manner by activating extrasynaptic receptors expressed by neurons (Yoon & Lee, 2014; Serrano et al., 2006). Astrocyte networks may control extracellular GABA concentration in order to improve STDP-based perceptual learning.

In our study, astrocytes contributed to improving STDP-based perceptual learning under noisy environmental conditions. An experimental study (Beste & Dinse, 2013) investigated how distracting stimuli affect perceptual learning in vision. Irrelevant distractor information was presented by varying the orientation of bars (vertical or horizontal) and the salience of the distractor information (e.g., by changing the length to width ratio of the bars). Subjects passively viewed the bars and made a fixation task. Perceptual learning deteriorated when the distractor was salient. Astrocytes may control extracellular GABA concentration in a spatiotemporal manner in order to suppress distracting sensory information, thereby improving perceptual learning.

Acknowledgments

We express our gratitude to Takeshi Kambara and Takami Matsuo for their helpful discussions. We express our sincere gratitude to the reviewers for providing their valuable insights, thoughtful suggestions, and constructive comments that undoubtedly strengthened this study.

References

Abbott
,
L. F.
, &
Nelson
,
S. B.
(
2000
).
Synaptic plasticity: Taming the beast
.
Nat. Neurosci.
,
3
,
1178
1183
.
Amzica
,
F.
, &
Neckelmann
,
D.
(
1999
).
Membrane capacitance of cortical neurons and glia during sleep oscillations and spike-wave seizures
.
J. Neurophysiol.
,
82
,
2731
2746
.
Anderson
,
T.
,
Hu
,
B.
,
Pittman
,
Q.
, &
Kiss
,
Z. H.
(
2004
).
Mechanisms of deep brain stimulation: An intracellular study in rat thalamus
.
J. Physiol.
,
559
,
301
313
.
Barakat
,
L.
, &
Bordey
,
A.
(
2002
).
GAT-1 and reversible GABA transport in Bergmann glia in slices
.
J. Neurophysiol.
,
88
,
1407
1419
.
Berninger
,
B.
, &
Bi
,
G. Q.
(
2002
).
Synaptic modification in neural circuits: A timely action
.
Bioessays
,
24
,
212
222
.
Bergles
,
D. E.
,
Diamond
,
J. S.
, &
Jahr
,
C. E.
(
1999
).
Clearance of glutamate inside the synapse and beyond
.
Curr. Opin. Neurobiol.
,
9
,
293
298
.
Beste
,
C.
, &
Dinse
,
H. R.
(
2013
).
Learning without training
.
Curr. Biol.
,
23
,
R489
R499
.
Bi
,
G. Q.
, &
Poo
,
M. M.
(
1998
).
Synaptic modifications in cultured hippocampal neurons: dependence on spike timing, synaptic strength, and postsynaptic cell type
.
J. Neurosci.
,
18
,
10464
10472
.
Bianchim
,
M. T.
,
Haas
,
K. F.
, &
Macdonald
,
R. L.
(
2001
).
Structural determinants of fast desensitization and desensitization-deactivation coupling in GABAa receptors
.
J. Neurosci.
,
21
,
1127
1136
.
Bianchim
,
M. T.
,
Haas
,
K. F.
, &
Macdonald
,
R. L.
(
2002
).
Alpha1 and alpha6 subunits specify distinct desensitization, deactivation and neurosteroid modulation of GABA(A) receptors containing the delta subunit
.
Neuropharmacology
,
43
,
492
502
.
Bonifazi
,
P.
,
Goldin
,
M.
,
Picardo
,
M. A.
,
Jorquera
,
I.
,
Cattani
,
A.
,
Bianconi
,
G.
, …
Cossart
,
R.
(
2009
).
GABAergic hub neurons orchestrate synchrony in developing hippocampal networks
.
Science
,
326
,
1419
1424
.
Brickley
,
S. G.
,
Cull-Candy
,
S. G.
, &
Farrant
,
M.
(
1996
).
Development of a tonic form of synaptic inhibition in rat cerebellar granule cells resulting from persistent activation of GABAA receptors
.
J. Physiol.
,
497
,
753
759
.
Bright
,
D. P.
, &
Smart
,
T. G.
(
2013
).
Methods for recording and measuring tonic GABAA receptor-mediated inhibition
.
Front. Neural Circuits
,
7
, 193.
Brown
,
N.
,
Kerby
,
J.
,
Bonnert
,
T. P.
,
Whiting
,
P. J.
, &
Wafford
,
K. A.
(
2002
).
Pharmacological characterization of a novel cell line expressing human alpha(4)beta(3)delta GABA(A) receptors
.
Br. J. Pharmacol.
,
136
,
965
974
.
Buzsaki
,
G.
(
2010
).
Neural syntax: Cell assemblies, synapsembles, and readers
.
Neuron
68
,
362
385
.
Buzsaki
,
G.
, &
Draguhn
,
A.
(
2004
).
Neuronal oscillations in cortical networks
.
Science
,
304
,
1926
1929
.
De Pittá
,
M.
, &
Brunel
,
N.
(
2016
).
Modulation of synaptic plasticity by glutamatergic gliotransmission: A modeling study
.
Neural Plast.
,
7607924
.
De Pittá
,
M.
,
Brunel
,
N.
, &
Volterra
,
A.
(
2016
).
Astrocytes: Orchestrating synaptic plasticity
?
Neuroscience
,
323
,
43
61
.
Dermietzel
,
R.
,
Hertberg
,
E. L.
,
Kessler
,
J. A.
, &
Spray
,
D. C.
(
1991
).
Gap junctions between cultured astrocytes: Immunocytochemical, molecular, and electrophysiological analysis
.
J. Neurosci.
,
11
,
1421
1432
.
Destexhe
,
A.
,
Mainen
,
Z. F.
, &
Sejnowski
,
T. J.
(
1998
). Kinetic models of synaptic transmission. In
C.
Koch
&
I.
Segev
(Eds.),
Methods in neuronal modeling
(pp.
1
25
).
Cambridge, MA
:
MIT Press
.
Dinse
,
H. R.
,
Ragert
,
P.
,
Pleger
,
B.
,
Schwenkreis
,
P.
, &
Tegenthoff
,
M.
(
2003
).
Pharmacological modulation of perceptual learning and associated cortical reorganization
.
Science
,
301
,
91
94
.
Drasbek
,
K. R.
, &
Jensen
,
K.
(
2006
).
THIP, a hypnotic and antinociceptive drug, enhances an extrasynaptic GABAA receptor-mediated conductance in mouse neocortex
.
Cereb. Cortex
,
16
,
1134
1141
.
Erlichman
,
J. S.
,
Cook
,
A.
,
Schwab
,
M. C.
,
Budd
,
T. W.
, &
Leiter
,
J. C.
(
2004
).
Heterogeneous patterns of pH regulation in glial cells in the dorsal and ventral medulla
.
Am. J. Physiol. Regul. Integr. Comp. Physiol.
,
286
,
R289
R302
.
Eulenburg
,
V.
, &
Gomeza
,
J.
(
2010
).
Neurotransmitter transporters expressed in glial cells as regulators of synapse function
.
Brain Res. Rev.
,
63
,
103
112
.
Farrant
,
M.
, &
Nusser
,
Z.
(
2005
).
Variations on an inhibitory theme: Phasic and tonic activation of GABA(A) receptors
.
Nat. Rev. Neurosci.
,
6
,
215
229
.
Giaume
,
C.
,
Koulakoff
,
A., Roux
, L.,
Holcman
,
D.
, &
Rouach
,
N.
(
2010
).
Astroglial networks: A step further in neuroglial and gliovascular interactions
.
Nat. Rev. Neurosci.
,
11
,
87
99
.
Gilbert
,
C. D.
(
1994
).
Neural dynamics and perceptual learning
.
Curr. Biol.
,
4
,
627
629
.
Gilbert
,
C. D.
(
1996
).
Plasticity in visual perception and physiology
.
Curr. Opin. Neurobiol.
,
6
,
269
274
.
Hoshino
,
O.
(
2011
).
Subthreshold membrane depolarization as memory trace for perceptual learning
.
Neural Comput.
,
23
,
3205
3231
.
Hoshino
,
O.
(
2012
).
Regulation of ambient GABA levels by neuron-glia signaling for reliable perception of multisensory events
.
Neural Comput.
,
24
,
2964
2993
.
Hoshino
,
O.
(
2014
).
Balanced crossmodal excitation and inhibition essential for maximizing multisensory gain
.
Neural Comput.
,
26
,
1362
1385
.
Hoshino
,
O.
(
2015
).
Regulation of local ambient GABA levels via transporter-mediated GABA import and export for subliminal learning
.
Neural Comput.
,
27
,
1223
1251
.
Houades
,
V.
,
Koulakoff
,
A.
,
Ezan
,
P.
,
Seif
,
I.
, &
Giaume
,
C.
(
2008
).
Gap junction-mediated astrocytic networks in the mouse barrel cortex
.
J. Neurosci.
,
28
,
5207
5217
.
Hubel
,
D. H.
, &
Wiesel
,
T. N.
(
1962
).
Receptive fields, binocular interaction and functional architecture in the cat's visual cortex
.
J. Physiol.
,
160
,
106
154
.
Isaacson
,
J. S.
, &
Scanziani
,
M.
(
2011
).
How inhibition shapes cortical activity
.
Neuron
,
72
,
231
243
.
Jones
,
M. V.
, &
Westbrook
,
G. L.
(
1995
).
Desensitized states prolong GABAA channel responses to brief agonist pulses
.
Neuron
,
15
,
181
191
.
Koch
,
U.
, &
Magnusson
,
A. K.
(
2009
).
Unconventional GABA release: Mechanisms and function
.
Curr. Opin. Neurobiol.
,
19
,
305
310
.
Lallouette
,
J.
,
De Pitt
,
M.
,
Ben-Jacob
,
E.
, &
Berry
,
H.
(
2014
).
Sparse short-distance connections enhance calcium wave propagation in a 3D model of astrocyte networks
.
Front. Comput. Neurosci.
,
8
, 45.
Lalo
,
U.
,
Pankratov
,
Y.
,
Kirchhoff
,
F.
,
North
,
R. A.
, &
Verkhratsky
,
A.
(
2006
).
NMDA receptors mediate neuron-to-glia signaling in mouse cortical astrocytes
.
J. Neurosci.
,
26
,
2673
2683
.
Lalo
,
U.
,
Pankratov
,
Y.
,
Parpura
,
V.
, &
Verkhratsky
,
A.
(
2011
).
Ionotropic receptors in neuronal-astroglial signalling: What is the role of “excitable” molecules in non-excitable cells
.
Biochim. Biophys. Acta.
,
1813
,
992
1002
.
Lee
,
H. S.
,
Ghetti
,
A.
,
Pinto-Duarte
,
A.
,
Wang
,
X.
,
Dziewczapolski
,
G.
,
Galimi
,
F.
, …
Heinemann
S. F.
(
2014
).
Astrocytes contribute to gamma oscillations and recognition memory
.
Proc. Natl. Acad. Sci. USA
,
111
,
E3343
E3352
.
Lerma
,
J.
,
Herranz
,
A. S.
,
Herreras
,
O.
,
Abraira
,
V.
, &
Martin
,
D. R.
(
1986
).
In vivo determination of extracellular concentration of amino acids in the rat hippocampus: A method based on brain dialysis and computerized analysis
.
Brain Res.
,
384
,
145
155
.
Losi
,
G.
,
Mariotti
,
L.
, &
Carmignoto
,
G.
(
2014
).
GABAergic interneuron to astrocyte signalling: A neglected form of cell communication in the brain
.
Philos. Trans. R. Soc. Lond. B
,
369
,
20130609
.
Maconochie
,
D. J.
,
Zempel
,
J. M.
, &
Steinbach
,
J. H.
(
1994
).
How quickly can GABAA receptors open?
Neuron
,
12
,
61
71
.