Modulation of Synaptic Plasticity by Glutamatergic Gliotransmission: A Modeling Study.

Glutamatergic gliotransmission, that is, the release of glutamate from perisynaptic astrocyte processes in an activity-dependent manner, has emerged as a potentially crucial signaling pathway for regulation of synaptic plasticity, yet its modes of expression and function in vivo remain unclear. Here, we focus on two experimentally well-identified gliotransmitter pathways, (i) modulations of synaptic release and (ii) postsynaptic slow inward currents mediated by glutamate released from astrocytes, and investigate their possible functional relevance on synaptic plasticity in a biophysical model of an astrocyte-regulated synapse. Our model predicts that both pathways could profoundly affect both short- and long-term plasticity. In particular, activity-dependent glutamate release from astrocytes could dramatically change spike-timing-dependent plasticity, turning potentiation into depression (and vice versa) for the same induction protocol.

1. Introduction

In recent years, astrocytes have attracted great interest for their capacity to release neuroactive molecules, among which are neurotransmitters like glutamate, because these molecules could modulate neural activity and lead to a possible role for astrocytes in neural information processing [1-3]. Indeed, astrocyte-derived neurotransmitters, also called “gliotransmitters” for their astrocytic origin [4], have been shown to act on neurons and to regulate synaptic transmission and plasticity through a variety of mechanisms [5]. The binding of receptors located on either pre- or postsynaptic terminals by astrocyte-released glutamate has historically been the first pathway for gliotransmission to be discovered and, arguably, the most studied one experimentally for its several possible functional implications [6].

Activation of extrasynaptic receptors on presynaptic terminals by astrocytic glutamate modulates the probability of neurotransmitter release from those terminals [6]. In particular, depending on receptor type, such modulation may be either toward an increase or toward a decrease of the frequency of spontaneous [7-11] and evoked neurotransmitter release in both excitatory [2, 8, 10, 12] and inhibitory synapses [13-15]. Because synaptic release probability characterizes how a synapse filters or, in other words, “processes” presynaptic action potentials [16, 17], modulations of synaptic release probability by astrocytic glutamate are suggested to alter the computational properties of neural circuits [18].

Glutamate released by astrocytes may also bind to extrasynaptically located postsynaptic NMDA receptors, evoking slow inward currents (SICs) in nearby neurons [11, 19–26]. The depolarizing action of these currents modulates neural excitability with the potential to affect neuronal action potential firing [27]. Moreover, because single astrocytes are in close proximity to a large number (~100) of neurons [28], it has been suggested that an inward current can be generated in many adjacent neurons, thereby promoting synchrony of neuronal firing [19-21].

Although modulations of both synaptic release and SICs mediated by glutamatergic gliotransmission have been recorded in the cortex and the hippocampus, as well as in several other brain regions [5], their physiological relevance remains elusive. In particular, beyond regulation of synaptic filtering and neuronal firing, theoretical arguments support a further possible role for both pathways in the regulation of NMDAR-mediated spike-timing-dependent plasticity (STDP) [29]. Both pathways have the potential to regulate activation of postsynaptic NMDA receptors and, in doing so, glutamatergic gliotransmission could ultimately regulate the STDP outcome, that is, either potentiation (LTP) or depression (LTD) [30, 31]. Consistent with this hypothesis, experiments have reported a lower threshold for LTP induction at hippocampal synapses when synaptic release is increased by astrocytic glutamate [9]. Moreover, long-term potentiation responses of neurons in the primary visual cortex by cholinergic activation of surrounding astrocytes has also been reported to be correlated with an increase of SIC frequency in those neurons [32].

While the potential impact on STDP of pre- or postsynaptic activity-dependent modulations of synaptic efficacy has widely been addressed both experimentally [33] and theoretically [34, 35], the possible effect on plasticity of the regulation of these modulations by glutamatergic gliotransmission (and by gliotransmission in general) has been investigated by very few theoretical studies. These studies suggest a potential role in LTP induction both for large increases of synaptic release and for large SICs mediated by astrocytic glutamate [36, 37]. This scenario seems however at odds with the majority of recent experimental observations that report modest signaling magnitudes for these two routes of gliotransmission. It is thus not clear under what biophysical conditions modulations of synaptic release or SICs mediated by glutamatergic gliotransmission could affect STDP. Astrocyte-mediated SICs, for example, are known to occur sporadically, being recorded in single neurons only as often as <5/min [26, 32], raising the question of whether and how, by occurring at such low rates, they could effectively play a role in STDP.

We thus set to investigating what conditions are required for glutamatergic gliotransmission to affect STDP by presynaptic modulations of neurotransmitter release or through postsynaptic SICs. We extend the model of an astrocyte-regulated synapse originally introduced by De Pittà et al. [73] to include a biophysically realistic description of synaptically evoked gliotransmitter release by the astrocyte as well as a mechanism for the generation of postsynaptic SICs and STDP. Extensive numerical investigations of our model lead to two major predictions. First, glutamatergic gliotransmission could change the nature of STDP by modifying the parameter ranges for LTP and LTD induction. Second, this effect crucially depends on the nature of gliotransmission, that is, whether it is release-increasing or release-decreasing, its strength, and its rate of occurrence and when it occurs with respect to pre/post pairs. Thus, while glutamatergic gliotransmission could potentially play a role in STDP and learning, in practice this effect must satisfy several biophysical and activity-dependent constraints, supporting the existence of specialized dynamic interactions between astrocytes and neurons.

2. Biophysical Modeling of a Gliotransmitter-Regulated Synapse

Although there may be several possible routes by which astrocytes release glutamate [38-40], Ca2+-dependent glutamate release is likely the main one in physiological conditions [41, 42]. From a modeling perspective, as illustrated in Figure 1, Ca2+-dependent glutamatergic gliotransmission consists of three distinct signaling pathways. One pathway (black arrows) initiates the release-triggering Ca2+ signal in the astrocyte and may be either exogenous or heterosynaptic or be triggered by the very synapses that are modulated by glutamatergic gliotransmission in a homosynaptic fashion. The other two pathways are instead represented by the two recognized routes for the action of glutamatergic gliotransmission on synaptic terminals: the presynaptic pathway whereby astrocytic glutamate modulates synaptic release (magenta arrows) and the postsynaptic pathway which mediates SICs in nearby neurons (orange arrows). Although both pathways could coexist at the same synapse in principle [11], their functional regulation is probably through different Ca2+-dependent pathways [26], both in terms of spatiotemporal Ca2+-dynamics [24] and in terms of pools of releasable glutamate resources and/or mechanism of release for these latter [43]. Thus, in the following, we set to investigating the effect of synaptic transmission of each pathway independently of the other.

Figure 1

Pathways of glutamatergic gliotransmission. Perisynaptic astrocytic processes in several brain areas and different excitatory (but also inhibitory) synapses may release glutamate in a Ca2+-dependent fashion. In turn, released astrocytic glutamate may increase (or decrease) synaptic neurotransmitter release by activating extrasynaptically located presynaptic receptors (magenta arrows) or contribute to postsynaptic neuronal depolarization by binding to extrasynaptic NMDA receptors (orange arrows) which mediate slow inward currents (SICs). These receptors often (but not always) contain NR2B subunits and are thus different with respect to postsynaptic NMDARs. Glutamate release by the astrocyte could be triggered either by activity from the same synapses that are regulated by the astrocyte (homosynaptic scenario) or by other synapses that are not directly reached by glutamatergic gliotransmission (heterosynaptic scenario).

2.1. Calcium-Dependent Gliotransmitter Release

We begin our study by a description of a biophysically realistic model of synaptically evoked Ca2+-dependent glutamate release from an astrocyte. At excitatory [44] and inhibitory synapses [45], astrocytes can respond to synaptically released neurotransmitters by intracellular Ca2+ elevations and release glutamate in turn [6]. Although morphological and functional details of the coupling between synaptic terminals and the surrounding astrocytic processes remain to be fully elucidated, the current hypothesis is that synaptically evoked glutamate-releasing astrocytic Ca2+ signaling is mainly by spillover of synaptic neurotransmitters and/or other factors, which bind to high-affinity astrocytic G protein-coupled receptors (GPCRs) [5] and thereby trigger inositol 1,4,5-trisphosphate (IP3) production and Ca2+ release from the endoplasmic reticulum (ER) [46-48]. While early work mainly monitored somatic Ca2+ increases concluding that astrocytes respond only to intense neuronal firing patterns [49], recent experiments in astrocytic processes revealed that astrocytes may also respond to low levels of synaptic activity by Ca2+ elevations confined in subcellular regions of their processes [10, 48, 50], suggesting that the profile of astrocytic Ca2+ signaling and thus glutamate release could span the whole spectrum of neuronal (synaptic) activity [5].

To realistically describe synaptic release in the whole spectrum of neuronal firing, we consider the model of an activity-dependent synapse first introduced by Tsodyks and Markram [51]. This model captures the dependence of synaptic release on past activity, that is, presynaptic short-term plasticity, which substantially influences synaptic transmission at high enough rates of neuronal firing [52]. In particular, synaptic release results from the product of two quantities: (i) the probability of neurotransmiter-containing vesicles to be available for release and (ii) the probability of such vesicles to be effectively released by an action potential [53], which correlates with intrasynaptic Ca2+ [54]. At rest, it is assumed that all vesicles are available for release. The arrival of an action potential opens presynaptic voltage-dependent Ca2+ channels that trigger a transient increase of intrasynaptic Ca2+ which promotes release of a fraction u of available vesicles. Following release, the emptied vesicles are refilled in some characteristic time τ , while intrasynaptic Ca2+ and thus vesicle release probability decay to zero with a different time constant τ . For multiple action potentials incoming at time intervals of the order of these two time constants, neither vesicle replenishment nor intrasynaptic Ca2+ are restored to their resting values, so that the resulting synaptic release depends on the history of synaptic activity [55].

We illustrate the response of the synapse model to a train of action potentials in Figures 2(a)–2(c). The low rate of stimulation of the first four action potentials (Figure 2(a)) allows for the reintegration of most of the released neurotransmitter in between action potentials thereby keeping vesicle depletion limited (Figure 2(b), orange trace). In parallel, intrasynaptic Ca2+ grows, and so does vesicle release probability (Figure 2(b), blue trace), resulting in progressively larger release of neurotransmitter per action potential or, in other words, in short-term facilitation of synaptic release (Figure 2(c), t < 500 ms). On the contrary, the presentation of a series of action potentials in rapid succession at t = 500 ms results in a sharp increase of vesicle release probability to a value close to saturation (i.e., Nt. Rel. Pr.≃1) which causes exhaustion of neurotransmitter resources (i.e., Avail. Nt. Pr.≃0). In this scenario, therefore, from one spike to the next one, progressively fewer neurotransmitter resources are available for release and the amount of released resources decreases with incoming action potentials, leading to depression of synaptic transmission. Such depression is short-lived, since synaptic release tends to recover after a sufficiently long period in which no action potentials occur, that is, the case, for example, of the last action potential at t = 800 ms.

Figure 2

Biophysical modeling of a gliotransmitter-regulated synapse. ((a)–(c)) Model of synaptic release. Incoming presynaptic spikes (a) increase intrasynaptic Ca2+ levels which directly control the probability of release of available neurotransmitter resources ((b), Nt. Rel. Pr.) and decrease, upon release, the fraction (or probability) of neurotransmitter-containing vesicles available for release (Avail. Nt. Pr.). Each spike results in release of a quantum of neurotransmitter from the synapse ((c), Released Nt.) whose concentration in the perisynaptic space decays exponentially. Synapse parameters: τ = 0.5 s, τ = 0.3 s, and U 0 = 0.6. Stimulation by Poisson-distributed APs with an average rate of 5 Hz. ((d)–(f)) Model for astrocyte activation. Synaptically released neurotransmitter in the perisynaptic space (d) binds astrocytic receptors ((e), Bound Ast. Rec.), resulting in IP3 production which triggers Ca2+ signaling in the astrocyte (f). This latter also depends on the fraction of deinactivated IP3 receptors/Ca2+ channels (Deinact. IP3Rs) on the astrocyte ER membrane (see Appendix A.1). ((g)–(i)) Model for gliotransmitter release. The increase of astrocytic Ca2+ beyond a threshold concentration ((g), cyan dashed line) results in the release of a quantum of gliotransmitter, which decreases the probability of further release of gliotransmitter ((h), Avail. Gt. Pr.) while transiently increasing extracellular gliotransmitter concentration ((i), Released Gt.). Model parameters as in Table 1.

Once released into the synaptic cleft, synaptic neurotransmitter is rapidly cleared by diffusion as well as by other mechanisms, including uptake by transporters and/or enzymatic degradation [56, 57]. In the simplest approximation, the contribution of these mechanisms can be modeled by a first-order reaction [58] which accounts for the exponentially decaying profile of neurotransmitter concentration in Figure 2(c) after synaptic release at each action potential. A fraction of released neurotransmitter molecules also spills out of the synaptic cleft to the perisynaptic space (Figure 2(d)) where it binds to GPCRs on the astrocyte (Figure 2(e)), therein triggering Ca2+ signaling (Figure 2(f)). To quantitatively describe this process, we modify the model of GPCR-mediated Ca2+ signaling originally introduced by De Pittà et al. [141] to account for dynamic regulation of astrocytic receptors by synaptic activity (see Appendix A, Section A.1). Accordingly, as illustrated in Figure 2(f), GPCR-mediated Ca2+ signaling is a result of the nonlinear interplay of three processes: (i) IP3 production by GPCRs bound by synaptic neurotransmitter (magenta trace), (ii) Ca2+ release from the ER into the cytosol, which is triggered by IP3-bound Ca2+ channels (IP3Rs) and also modulates cytosolic IP3 (black trace), and (iii) the effective fraction of available or, more exactly, “deinactivated” IP3Rs [59] that can take part in Ca2+ release from the ER (yellow trace). Depending on the choice of parameter values, the astrocyte model may display both large, long-lasting somatic Ca2+ elevations and smaller and shorter Ca2+ increases, akin to those reported in astrocytic processes [47] (see Appendix B).

Glutamate release from the astrocyte is then assumed to occur every time that Ca2+ increases beyond a threshold concentration (Figure 2(g), cyan dotted line), in agreement with experimental observations [60, 61]. Although different mechanisms for glutamate release by the astrocyte could be possible, a large amount of evidence points to vesicular exocytosis as the main one to likely occur on a physiological basis [62]. Because astrocytic glutamate exocytosis bears several similarities with its synaptic homologue (reviewed in De Pittà et al. [29]), we model it in the same fashion. Thus, in line with experimental observations [63, 64], we postulate the existence of an astrocytic vesicular compartment that is competent for regulated glutamate exocytosis. Then, upon a suprathreshold Ca2+ elevation, a fixed fraction of astrocytic glutamate-containing vesicles releases glutamate into the extracellular space. Glutamate is then reintegrated into the astrocyte with some characteristic time constant (Figure 2(h)). In this fashion, glutamate concentration in the extracellular space abruptly increases by exocytosis from the astrocyte and then exponentially decays akin to neurotransmitter concentration in the synaptic cleft, yet, in general, at a different rate (Figure 2(h)) (Appendix B).

The description of gliotransmitter release hitherto introduced ignores the possible stochastic nature of astrocytic glutamate release [65] and reproduces the total amount of glutamate released, on average, by a single Ca2+ elevation beyond the release threshold. This description provides a simplified general framework to realistically capture synaptically evoked glutamate release by the astrocyte independently of the underlying mechanism of astrocytic exocytosis, which may either be in the form of a burst of synchronous vesicle fusion events that peaks within the first 50–500 ms from the Ca2+ rise underneath the plasma membrane [61, 65, 66] or occur at slower fusion rates in an asynchronous fashion [67, 68].

2.2. Gliotransmitter-Mediated Regulation of Synaptic Release and Short-Term Synaptic Plasticity

Once released, astrocyte-derived glutamate can diffuse in the extracellular space and bind extrasynaptic receptors located on presynaptic terminals. In particular, ultrastructural evidence suggests colocalization of glutamate-containing vesicles in perisynaptic astrocytic processes with those receptors [8], hinting a focal action of astrocytic glutamate on these latter. Such action is likely spatially confined and temporally precise, akin to that of a neurotransmitter on postsynaptic receptors, and is not affected by synaptic neurotransmitters [6]. Both ionotropic and metabotropic presynaptic receptors may be activated by astrocytic glutamate, yet their differential recruitment likely depends on developmental, regional, physiological, and cellular (synaptic) factors (reviewed in [29]). The details of the biochemical mechanisms of action of these receptors on synaptic physiology are not fully understood [69], but the simplest explanation is that they all modulate intrasynaptic Ca2+ levels, eventually increasing or decreasing synaptic release probability [18], although in a receptor-specific fashion [52, 69, 70].

From a modeling perspective, as originally proposed by De Pittà et al. [73], the common effect on synaptic release shared by different receptors allows expressing, in the simplest approximation, the synapse's resting release probability proportionally to the fraction of presynaptic receptors activated by astrocytic glutamate (Appendix A, Section A.1). In this fashion, as illustrated in Figure 3, the time evolution of the fraction of activated presynaptic receptors ensuing from a series of glutamate release events by the astrocyte (Figures 3(a) and 3(b)) is reflected by the dynamics of synaptic release probability at rest averaged across different trials (Figures 3(c) and 3(e)). The value of the coefficient of proportionality for the dependence of synaptic release probability on receptor activation sets the type of modulation of synaptic release by astrocytic glutamate which can be either release-decreasing (Figure 3(c)), such as in the case of astrocytic glutamate-binding presynaptic kainate receptors or group II/III metabotropic receptors (mGluRs) [13, 14, 71], or release-increasing (Figure 3(e)), when astrocytic glutamate binds NMDARs or group I mGluRs [7–9, 11, 12, 26, 72]. The functional implications of these modulations of synaptic release by glutamatergic gliotransmission on synaptic transmission have been widely addressed in a series of previous studies [18, 29, 73], and the remainder of this section reviews and extends the main results from those studies about short-term synaptic plastic and synaptic filtering.

Figure 3

Presynaptic pathway of gliotransmission. Gliotransmitter released from the astrocyte (a) binds extrasynaptically located presynaptic receptors (b) thereby decreasing or increasing synaptic release depending on the type of gliotransmitter and receptor. In the release-decreasing case, synaptic release probability could approach zero by gliotransmission ((c), red trace, ξ = 0), although, in practice, less dramatic reductions are more likely to be measured with respect to the original value in the absence of gliotransmission (black dashed line). The reduction in synaptic release probability changes pair pulse plasticity increasing the pair pulse ratio (d). In the case of release-increasing gliotransmission, synaptic release probability could instead increase up to one ((e), green trace, ξ = 1). In turn, pair pulse plasticity changes towards a decrease of the ensuing pair pulse ratio (f). Parameters as in Table 1 except for ϱ = 10−4, O = 0.6 μM−1 s−1, τ = 30 s, ζ = 0.54, J = 3 mV, and R in = 60 MΩ.

Figure 3(d) (left panel) shows how postsynaptic currents (PSCs) change in the presence of release-decreasing glutamatergic gliotransmission when elicited by two consecutive action potentials arriving to the resting synapse 20 ms after the onset of gliotransmission at t = 5 s (Figure 3(c)). Two differences with respect to the case without gliotransmission (black trace) may be observed. First the PSC amplitude overall decreases (red trace), consistent with a decrease of synaptic efficacy caused by the reduction of synaptic release by astrocytic glutamate. Then, the second PSC is larger than the first one, which is the opposite of what would be measured in the absence of gliotransmission. In other words, in agreement with experimental observations [13], the release-decreasing effect of astrocytic glutamate results in an increased pair pulse ratio (PPR) with respect to the case without gliotransmission (PPR0). Notably, as shown in Figure 3(d) (right panel), this change in the PPR ratio is only transient and vanishes together with the effect of gliotransmission on synaptic release. Similar considerations also hold in the case of a release-increasing effect of astrocytic glutamate on synaptic transmission [8]: while PSC amplitude increases (Figure 3(f), left panel, green trace), this occurs to the detriment of PPR, which decreases instead (Figure 3(f), right panel). Thus, synapses whose release probability is increased by glutamatergic gliotransmission are likely to run out of neurotransmitter faster, exhibiting rapid onset of short-term depression, consistent with lower PPR values. On the contrary, synapses whose release probability is reduced by astrocyte-released glutamate deplete their neurotransmitter resources slower and may exhibit progressive facilitation (i.e., potentiation) of their efficacy to transmit action potentials and so larger PPR values [74]. That is, the plasticity mode of a synapse, namely, whether it is depressing or facilitating, may not be fixed but rather be modulated by glutamatergic gliotransmission by surrounding astrocytes in an activity-dependent fashion [29, 73].

An important consequence of short-term synaptic dynamics is that synapses can act as filters [16, 17, 75]. Hence, modulations of synaptic dynamics by glutamatergic gliotransmission are also expected to affect the synapse's filtering characteristics [18]. This scenario is illustrated in Figure 4 where the effect of release-decreasing versus release-increasing glutamatergic gliotransmission, respectively, on depressing and facilitating synapses, is shown in terms of changes of the filtering characteristics of these synapses, that is, their steady-state release as a function of the frequency of presynaptic stimulation [17]. In the absence of gliotransmission, depressing synapses, which are characterized by intermediate-to-high initial probability of release [74] (Figure 4(a), black circles), predominantly act as low-pass filters (Figure 4(b), black circles) that are most effective at transmitting low frequency presynaptic spike trains (Figure 4(c), black traces). On the contrary, facilitating synapses, with a low-to-intermediate initial probability of neurotransmitter release [74] (Figure 4(a), black circles), function as high-pass or band-pass filters (Figure 4(b), black circles); that is, they are mostly effective at transmitting action potentials in an intermediate range of presynaptic activity (Figure 4(c), black trace).

Figure 4

Gliotransmitter-mediated modulation of synaptic frequency response. Decrease (a) or increase (d) of synaptic release probability by gliotransmission modulates the average per-spike synaptic release, resulting in a change of the synapse frequency response. Monotonically decreasing frequency responses that are typical of depressing synapses could be flattened by release-decreasing gliotransmission ((b), black versus red points), and vice versa, almost nonmonotonic ones, characteristic of facilitating synapses, could turn into monotonically decreasing responses by release-increasing gliotransmission ((e), black versus green points). Changes in frequency response depend on whether gliotransmission impinges on the very synapse that is triggered by (homosynaptic/closed-loop scenario) or not (heterosynaptic/open-loop scenario). In the homosynaptic scenario, the synaptic response is expected to change only for presynaptic firing rates that are sufficiently high to trigger gliotransmitter release from the astrocyte ((b), (e), cyan points). Data points and error bars: mean ± STD for n = 20 (no gliot. and heterosyn. gliot.) or n = 200 simulations (homosyn. gliot.) with 60 s long Poisson-distributed presynaptic spike trains. ((c), (f)) The change of synaptic frequency response mediated by gliotransmission (three consecutive gliotransmitter releases at the time instants marked by triangles) leads to changes in how presynaptic firing rates (top panels) are transmitted by the synapse (bottom panels). Simulated postsynaptic currents (PSCs) are shown as average traces of n = 1000 simulations for gliotransmitter release at 1 Hz. Release-decreasing gliotransmission was achieved for ξ = 0, whereas ξ = 1 was used for release-increasing gliotransmission. Depressing synapse in ((a), (b)): τ = 0.5 s, τ = 0.3 s, and U 0 = 0.6; facilitating synapse in ((d), (e)): τ = 0.5 s, τ = 0.5 s, and U 0 = 0.15. Other model parameters as in Figure 3 except for R in = 300 MΩ.

In the presence of glutamate release by the astrocyte, these two scenarios could be reversed. Consider indeed the simple heterosynaptic case where glutamatergic gliotransmission is stimulated by other means compared to by the very synapses it impinges on. It may be noted that release-decreasing gliotransmission flattens the synaptic steady-state release towards zero for all frequencies of stimulation (Figure 4(b), red circles), ensuing in synaptic transmission that resembles the one of a facilitating, band-pass synapse (compare the red PSC trace in Figure 4(c) with the black PSC trace in Figure 4(f)). Vice versa, release-increasing gliotransmission could turn band-pass features of transmission by a facilitating synapse (Figure 4(e), green circles) into low-pass, reminiscent of a more depressing synapse (compare the green PSC trace in Figure 4(f) with the black PSC trace in Figure 4(c)). On the other hand, when gliotransmission is stimulated by the same synapses that it modulates, that is, in the homosynaptic scenario of gliotransmission, inspection of the ensuing synaptic filtering characteristics (Figures 4(b) and 4(e), cyan circles) reveals that these latter coincide with those obtained in the absence of gliotransmission for low frequencies of presynaptic activity, while they tend to equal those observed with heterosynaptic gliotransmission as the frequency of stimulation increases. This coexistence of mixed features from apparently opposite scenarios, that is, no gliotransmission versus heterosynaptic gliotransmission, can be explained by the fact that the release of glutamate from the astrocyte requires intracellular Ca2+ to cross a threshold concentration. Hence, in the homosynaptic scenario, synapses that impinge on the astrocyte must be stimulated at rate sufficiently high to allow astrocytic Ca2+ to increase beyond such a threshold.

The modulation of synaptic filtering by glutamatergic gliotransmission offers the possibility that the same stimulus could be differently filtered (i.e., processed) and transmitted by a synapse in the presence (or not) of glutamate release by surrounding astrocytic processes, ultimately endowing that synapse with processing versatility with respect to incoming action potentials. Moreover, to the extent that synaptic dynamics critically shape the computations performed by the neural circuitry, such versatility could also be reflected at the network level, leading to the possibility that the same neuron-glia network could be involved in different computational tasks defined, time by time, by activity-dependent gliotransmitter release by astrocytes in the network.

2.3. Astrocyte-Mediated Slow Inward Currents

Induction of slow inward (i.e., depolarizing) currents (SICs) by activation of extrasynaptically located postsynaptic NMDA receptors is the other mechanism considered in this study whereby glutamatergic gliotransmission could affect synaptic information transfer. While astrocyte-mediated SICs have been reported in several brain regions, the pathway underlying glutamate release by astrocytes has not been fully elucidated [76, 77]. It is likely that, similar to the presynaptic route for glutamatergic gliotransmission discussed above, multiple pathways for glutamate release could be used by the same astrocyte [39], but their deployment depends on developmental, regional, and physiological factors [27]. Astrocytic Ca2+ activity seems to be a crucial factor in the regulation of astrocyte-mediated SICs [19–23, 25, 78]. In particular, SIC frequency and amplitude have been shown to increase upon Ca2+ elevations mediated by GPCRs on astrocytes such as mGluRs [19–23, 72, 79], the metabotropic purinergic P2Y1 receptor [25], the endocannabinoid CB1 receptor [80], or the protease-activated receptor 1 (PAR1) [24]. Remarkably, stimulation of PAR1s on hippocampal astrocytes was shown to trigger, under physiological conditions, Ca2+-dependent glutamate release from these cells through Bestrophin-1 anion channel [81, 82], and this pathway of glutamate release has been suggested as a candidate mechanism for SICs [83]. Channel-mediated glutamate release is expected to account for prolonged (>10 s) release of transmitter but in small amounts per unit time [82] thus ensuing in modest, very slow rising and decaying inward currents. While similar SICs have indeed been recorded [71, 84], most experiments reported SICs within a wide range of amplitudes to last only few seconds at most and rise in correlation with astrocytic Ca2+ increases, with rise time much shorter than their decay [20–22, 24, 26, 32, 85, 86] akin to currents that would ensue from a quantal mechanism of gliotransmitter release [62].

Based on these arguments, we assume glutamate exocytosis as the candidate mechanism for glutamate release by astrocytes that mediates SICs. Accordingly, we adopt the description of astrocytic glutamate exocytosis previously introduced (Figures 2(g)–2(i)) to also model astrocyte-mediated SICs. In this fashion, glutamate exocytosis by the astrocyte into the extracellular space (Figure 5(a)) results in activation of extrasynaptically located NMDARs on nearby neuronal dendrites which trigger SICs (Figure 5(b)) that cause slow depolarizing postsynaptic potentials (PSP, Figure 5(c)).

Figure 5

Postsynaptic pathway of gliotransmission by slow inward currents. The transient increase of gliotransmitter concentration in the perisynaptic space (a) triggers a slow inward (depolarizing) current (SIC) in the postsynaptic neuron ((b), (c)). Such SIC adds to postsynaptic currents triggered by presynaptic spikes ((d), (e), cyan triangle marks gliotransmitter release/SIC onset) and may dramatically alter postsynaptic firing (f). In general postsynaptic firing frequency increases with both SIC amplitude (g) and frequency (h). In this latter case, however, SICs as ample as 30 pA (similar to what reported in several experiments) need to impinge on the postsynaptic neuron at unrealistically high rates (≫0.1 Hz) in order to trigger a sensible change in the neuron's firing rate (black data points). Lower, more realistic SIC rates may affect neuronal firing only for larger SIC amplitudes (e.g., 45 pA, grey data points). The entity of SIC-mediated increase of postsynaptic neuronal firing further depends on the neuron's state of depolarization at SIC timings which is set by synaptic inputs (blue and cyan data points). Data points and error bars: mean ± STD out of n = 50 simulations with presynaptic Poisson-distributed spike trains. Parameters as in Table 1 except for ϱ = 10−4, τ = 200 ms, τ = 10 ms, and R in = 150 MΩ.

An important functional consequence of SIC-mediated depolarizations is that they can modulate neuronal excitability [21–23, 85]. As illustrated in Figures 5(d) and 5(e), astrocyte-mediated SICs (cyan trace) may add to regular synaptic currents (black trace) resulting in depolarizations of postsynaptic neurons closer to their firing threshold [23]. In turn, these larger depolarizations could dramatically change generation and timing of action potentials by those neurons in response to incoming synaptic stimuli (Figure 5(f)). This could ultimately affect several neurons within the reach of glutamate released by an astrocyte, leading to synchronous transient increases of their firing activity [21]. Remarkably, this concerted increase of neuronal excitability has often been observed in correspondence with large amplitude (i.e., >100 pA) SICs [21, 25, 85, 87], but experiments report the majority of SICs to be generally smaller, with amplitudes < 80 pA [11, 21, 22, 26, 32, 87]. It is therefore unclear whether SIC-mediated increase of neuronal excitability could occur [88] or not [87, 89, 90] in physiological conditions.

In Figure 5(g), we consider postsynaptic firing in a standard leaky integrate-and-fire neuron model [91, 92] as a function of presynaptic activity for SICs of different amplitudes (30–45 pA, see Appendix B) randomly occurring at an average rate of 1 Hz based on a binomial process for glutamate release from astrocytes as suggested by experiments [65] (see Appendix A). In line with experimental evidence [93], the input-output transfer function in the absence of gliotransmission has a typical sigmoidal shape (black dots) which reflects the following: (i) gradual emergence of firing for low (>10 Hz) fluctuating synaptic inputs; (ii) the progressive, quasi-linear increase of the firing rate for presynaptic activity beyond ~30 Hz; and finally, (iii) saturation of the firing rate for sufficiently strong synaptic inputs such that timing of action potential generation approaches the neuron's refractory period (which was fixed at 2 ms in the simulations, Appendix B) [92]. The addition of astrocyte-mediated SICs alters the firing characteristics of the neuron due to the ensuing larger depolarization. In particular the neuron could generate action potentials for lower rates of presynaptic activity (cyan/blue dots). Clearly, the larger the SIC is, the more the postsynaptic firing increases with respect to the case without SICs, for a given level of presynaptic activity.

As previously mentioned, these results assume an average 1 Hz rate for astrocyte-mediated SICs. While such a rate cannot be excluded, it seems unlikely for the following reasons. The weak correlation of SIC amplitude with somatic Ca2+ elevations observed in experiments favors indeed the idea that glutamate-mediated SICs are highly localized events, occurring within subcellular domains at astrocytic processes [22]. In turn, Ca2+ elevations in astrocytic processes could be as short-lived as ~0.5 s [10, 50], thus in principle allowing for glutamate release rates of the order of 1 Hz. However, in practice, reported SIC frequency is much lower, that is, <5/min (i.e., ~0.08 Hz) [11, 22]. Hence, it may be expected that the effect of SICs on neuronal firing could be considerably reduced with respect to the case in Figure 5(g).

We consider this possibility more closely in Figure 5(h), where we analyze postsynaptic firing in function of the average frequency of astrocyte-mediated SICs, both in the absence of synaptic activity (black and dark blue dots) and in the case of presynaptic activity at an average rate ~ 1 Hz, which corresponds to typical levels of spontaneous activity in vivo [94] (grey and light blue dots). It may be noted that the effect of SICs of typical amplitudes on postsynaptic firing rate is generally small, that is, <0.5 Hz, except for unrealistic (>0.1 Hz) SIC rates, while it gets stronger in association with synaptic activity. In this latter case however the possible increase in postsynaptic firing by astrocyte-mediated SICs is limited by the rate of reintegration of released glutamate resources in the astrocyte (fixed at ~1 Hz, Appendix B). Analogously to short-term synaptic depression in fact, our description of gliotransmitter release predicts that, for release rates that exceed the rate of reintegration of released glutamate by the astrocyte, exhaustion of astrocytic glutamate resources available for further release will result in SICs of smaller amplitude. In this fashion, due to depletion of astrocytic glutamate, the effect of large rates of glutamate release and thus of SICs on neuronal firing tends to be equivalent to that of considerably lower ones.

Taken together, the above results do not exclude a possible role of SICs in modulation of neuronal excitability and firing but suggest that such modulation could effectively occur only in coincidence with proper levels of synaptic activity. In this fashion, astrocyte-mediated SICs could be regarded to operate a sort of coincidence detection between synaptic activity and astrocytic glutamate release [22], whose readout would then be a temporally precise, cell-specific increase of neuronal firing (Figure 5(f)).

3. Astrocyte-Mediated Regulation of Long-Term Plasticity

The strength of a synaptic connection between two neurons can be modified by activity, in a way that depends on the timing of neuronal firing on both sides of the synapse, through a series of processes collectively known as spike-timing-dependent plasticity (STDP) [95]. As both pre- and postsynaptic pathways of glutamatergic gliotransmission potentially change EPSC magnitude, thereby affecting postsynaptic firing, it may be expected that they could also influence STDP.

Although the molecular mechanisms of STDP remain debated, and different mechanisms could be possible depending on type of synapse, age, and induction protocol [34], at several central excitatory synapses postsynaptic calcium concentration has been pointed out as a necessary factor in induction of synaptic changes by STDP [31, 96–99]. Remarkably, amplitude and, likely, time course of postsynaptic Ca2+ could control the direction of plasticity: smaller, slower increases of postsynaptic Ca2+ give rise to spike-timing-dependent long-term depression (LTD), whereas larger, more rapid increases cause spike-timing-dependent long-term potentiation (LTP) [31, 96, 97]. In calcium-based STDP models, this is also known as the “Ca2+-control hypothesis” [35, 100, 101]. According to this hypothesis, no modification of synaptic strength occurs when Ca2+ is below a threshold θ that is larger than the resting Ca2+ concentration. If calcium resides in an intermediate concentration range, between θ and a second threshold θ > θ , the synaptic strength is decreased. Finally, if calcium increases above the second threshold, θ , the synaptic strength is potentiated.

Figures 6(a1) and 6(b1) exemplify the operational mechanism of the Ca2+-control hypothesis within the framework of a nonlinear Ca2+-based model for STDP at glutamatergic synapses originally introduced by Graupner and Brunel [102]. At most glutamatergic synapses, postsynaptic Ca2+ is mainly regulated by two processes: (i) postsynaptic Ca2+ entry mediated by NMDARs [103] and (ii) Ca2+ influx by voltage-dependent Ca2+ channels (VDCCs) [31, 33, 99, 104]. In this fashion, each presynaptic action potential generates a long-lasting Ca2+ transient by opening NMDAR channels, while postsynaptic firing results in a short-lasting Ca2+ transient due to opening of VDCCs by dendritic depolarization through back-propagating action potentials (bAPs) [95]. Presynaptic action potentials alone do not trigger changes in synaptic strength, but they do so in correlation with postsynaptic bAPs [105]. Notably [106], in a typical STDP induction pairing protocol, LTD is induced if the postsynaptic neuron fires before the presynaptic one, that is, post → pre pairing at negative spike-timing intervals Δt (Figure 6(a1)). Contrarily, LTP is induced when the presynaptic cell fires before the postsynaptic cell, that is, for pre → post pairing at positive Δt intervals (Figure 6(a1)). This is possible because, when a presynaptic action potential is followed shortly after by a postsynaptic bAP, the strong depolarization by this latter drastically increases the voltage-dependent NMDAR-mediated Ca2+ current due to removal of the NMDAR magnesium block [107, 108], thereby resulting in supralinear superposition of the NMDAR- and VDCC-mediated Ca2+ influxes.

Figure 6

STDP modulation by gliotransmitter regulation of synaptic release. ((a), (b)) Rationale of LTD and LTP without ((a1), (b1)) and with either release-decreasing ((a2), (b2), ξ = 0) or release-increasing gliotransmission ((a3), (b3), ξ = 1) setting on at the red/green marks. (c) Percentage of time spent by postsynaptic Ca2+ transients (left panel) above depression (dashed lines) and potentiation thresholds (solid lines) for spike-timing intervals (Δt) within ±100 ms and resulting STDP curves (right panel) in the absence of gliotransmission (no gliot., black curve) and with maximal release-decreasing (RD, red curve) or release-increasing gliotransmission (RI, green circles). (d) In general, strength and direction (i.e., “type”) of gliotransmission may dramatically modulate STDP. For example, synaptic changes are attenuated when synaptic release is decreased by gliotransmission (area below the black dashed line). Conversely, for sufficiently strong release-increasing gliotransmission (area above the black dotted line), the LTP window shrinks and LTD may be measured for all Δt < 0, as well as for sufficiently large Δt > 0. (e) A closer inspection of STDP curves indeed reveals that LTD (yellow curve) increases for larger synaptic release accounted by gliotransmission, while the ratio between areas underneath the LTP and LTD (magenta curve), initially in favor of the former (i.e., for release-decreasing gliotransmission), reduces to zero for large enough release-increasing gliotransmission, when two open LTD windows appear outside a small LTP window center for small Δt > 0 (hatched area). Synaptic parameters: τ = 0.33 s, τ = 0.33 s, and U = 0.5 s. Other parameters as in Table 1 except for ϱ = 10−4, τ = 1 ms, W = 78.7, τ = 5 s in ((a), (b)), and τ = 30 s otherwise.

In the framework of the Ca2+-control hypothesis, these observations may be summarized as follows. For large Δt, pre- and postsynaptic Ca2+ transients do not interact, and the contributions from potentiation and depression by pre/post pairs (or vice versa) cancel each other, leading to no synaptic changes on average (Figure 6(c), black curves). For short, negative Δt, the presynaptically evoked Ca2+ transient rises instead above the depression threshold (θ ) but not beyond the potentiation threshold (θ ). Consequently, depression increases whereas potentiation remains constant, which leads to LTD induction. For short, positive Δt however the postsynaptically evoked calcium transient rises on top of the presynaptic transient by the NMDAR nonlinearity and increases activation of both depression and potentiation. Because the rate of potentiation is larger than the rate of depression (Appendix C), this results in LTP induction.

For the same number of pre/post pairs (or vice versa), mapping of the average synaptic modification as a function of the spike-timing interval Δt ultimately provides an STDP curve that qualitatively resembles the classic curve originally described by Bi and Poo [123] (Figure 6(c), right panel, black curve). In the following, we will focus on parameters that lead to such a STDP curve and investigate how this curve is affected in the presence of glutamatergic gliotransmission, through the pre- and postsynaptic pathways of regulation discussed above.

3.1. Presynaptic Pathway

The very nature of synaptic transmission crucially depends on the synapse's initial probability of neurotransmitter release, insofar as this latter sets both the tone of synaptic transmission, that is, how much neurotransmitter is released per action potential by the synapse on average, and whether the synapse displays short-term depression or facilitation [17]. Synapses with low-to-intermediate values of initial neurotransmitter release probability, for example, Schaffer collateral synapses [74], or some cortical synapses [16], are indeed prone to display facilitation, whereas synapses that are characterized by large release probability are generally depressing [16]. Because synaptic release probability also dictates the degree of activation of NMDARs and consequently the magnitude of postsynaptic Ca2+ influx, it is expected that both the tone of synaptic transmission and its short-term dynamics could affect STDP [34]. The relative weight of these two factors in shaping synaptic changes however likely depends on the protocol for STDP induction. Short-term plasticity could indeed sensibly regulate STDP induction only for rates of presynaptic action potentials high enough to allow facilitation or depression of synaptic release from one AP to the following one [109, 110]. In this study, we consider low pre/post frequencies of 1 Hz. At such frequencies we expect short-term plasticity to have a negligible effect, and thus we only focus on how changes in the tone of synaptic transmission by glutamatergic gliotransmission affect STDP.

Figures 6(a2) and 6(b2), respectively, show the outcome of LTD and LTP induction for two consecutive pre → post and pre → post pairings preceded by the onset of release-decreasing gliotransmission at 0.1 s (top panels, black marks). A comparison of the ensuing postsynaptic Ca2+ dynamics with respect to the case without gliotransmission (Figures 6(a1) and 6(b1)) reveals that the strong decrease of synaptic release probability (SRP, top panels, red curves) caused by gliotransmission remarkably reduces the NMDAR-mediated contribution to postsynaptic Ca2+ influx (middle panels), resulting in smaller variations of synaptic strength (bottom panels). In this fashion, at the end of the pairing protocol, release-decreasing gliotransmission accounts for less time spent by Ca2+ above both LTD and LTP thresholds (Figure 6(c), left panel, red traces). The resulting STDP curve thus displays strongly attenuated LTD and LTP (Figure 6(c), right panel, red curve), with LTP windows spanning a considerably smaller range of Δts values than in the curve obtained without gliotransmission (black curve).

Similar considerations apply to the case of release-increasing gliotransmission (Figures 6(a3) and 6(b3)). In this case, the NMDAR component of postsynaptic Ca2+ could increase by gliotransmission even beyond the θ threshold (dashed blue line), thus favoring depression while reducing potentiation (bottom panels). In particular, for short positive Δt, the maximal LTP does not change but the Δt range for LTP induction shrinks. For Δt > 40 ms in fact, the time that Ca2+ spends above the LTD threshold increases with respect to the time spent by Ca2+ above the LTP threshold, thereby resulting in LTD induction (Figure 6(c), left panel, green traces). In this fashion, the STDP curve in the presence of release-increasing gliotransmission displays a narrow 0–40 ms LTP window outside which LTD occurs instead (Figure 6(c), right panel, green curve).

Figure 6(d) summarizes how the STDP curve changes for the whole spectrum of glutamatergic gliotransmission. In this figure, a y-axis value of “gliotransmission type” equal to 0 corresponds to maximum release-decreasing gliotransmission (red curve in Figure 3(c)); a value equal to 1 stands instead for maximum release-increasing gliotransmission (as in the case of the green curve in Figure 3(c)); finally, a value of 0.5 corresponds to no effect of gliotransmission on synaptic release (black curve in Figure 3(c)). It may be noted that gliotransmission may affect the STDP curve in several ways, changing both strength of plastic changes (color code) and shape and areas of LTP and LTD windows. In particular, as revealed by Figure 6(e), maxima of LTP (cyan circles) and LTD (yellow circles) decrease with decreasing values of gliotransmission type, consistently with smaller postsynaptic Ca2+ influx for larger decreases of synaptic release by gliotransmission. This suggests that release-decreasing gliotransmission (red-shaded area) could attenuate STDP yet in a peculiar fashion, counteracting LTD more than LTP induction, as reflected by increasing values of LTP/LTD area ratio (magenta curve).

On the contrary, the effect of release-increasing gliotransmission (Figure 6(e), green-shaded area) could be dramatically different. For sufficiently strong increases of synaptic release by gliotransmission in fact, the LTP/LTD area ratio drops to zero (hatched area) in correspondence with the appearance of two “open” LTD windows, one for Δt < 0 and the other for sufficiently large positive spike-timing intervals. In parallel, consistently with the fact that release-increasing gliotransmission tends to increase the fraction of time spent by postsynaptic Ca2+ above the threshold for LTD thereby promoting this latter (Figure 6(c)), the range for LTP induction also tends to shrink to lower Δt values as release-increasing gliotransmission grows stronger (Figure 6(d), red color-coded areas for gliotransmission type > 0.5).

In summary, our analysis reveals that modulation of synaptic release by glutamatergic gliotransmission could change STDP both quantitatively and qualitatively, from hindering its induction for release-decreasing modulations to altering both shape and existence of LTD windows for release-increasing modulations. However, whether and how this could effectively be observed in experiments remain to be investigated. Supported both by experimental evidence and theoretical arguments is the notion that regulations of the tone of synaptic transmission by glutamatergic gliotransmission likely require specific morphological and functional constraints to be fulfilled by the nature of astrocyte-synapse coupling [5, 18]. Similar arguments may ultimately hold true also for modulation of STDP; insofar as for this modulation to be measured in our simulations, we required both a sufficiently strong increase/decrease of synaptic release by gliotransmission and a decay time of such increase/decrease long enough for this latter to be present during the induction protocol. Should these two aspects not have been fulfilled in our simulations, then modulation of STDP by gliotransmitter-mediated changes of synaptic release would likely have been negligible or even undetectable.

3.2. Postsynaptic Pathway

We now turn our analysis to the possible impact of astrocyte-mediated SICs on STDP. Because SICs are through extrasynaptic NMDA receptors and these receptors are mainly permeable to Ca2+ ions [111], then SICs could contribute to postsynaptic Ca2+ thereby affecting STDP. Nevertheless, we should note that it is unclear whether and how extrasynaptic NMDARs contribute to plasticity, independently of the occurrence of SICs [77]. For example, theta-burst LTP induction in CA1 neurons of rat hippocampal slices is turned into LTD when extracellular NMDARs are selectively stimulated [112], but it is unknown whether these receptors have a role in STDP [113]. In general, for a given STDP induction protocol, two factors that could crucially regulate how Ca2+ transients mediated by extrasynaptic NMDARs are involved in STDP are the location of these receptors on the spine and the morphology of this latter in terms of spine head and neck [114, 115]. Unfortunately both these factors remain unknown in the current knowledge of SIC-mediating extrasynaptic NMDARs and, for the remainder of this study, we assume that, in spite of their possible location away from the postsynaptic density along the spine neck or the dendritic shaft [116], SIC-mediating extrasynaptic NMDARs could still regulate spine Ca2+ dynamics [27].

Based on the above rationale, we thus model SICs as slow postsynaptic Ca2+ transients that will add to presynaptically and postsynaptically triggered ones and study their effect on the induction of SDTP by classic pairing protocols. For the sake of generality, we express the peak of SIC-mediated Ca2+ transients in units of NMDAR-mediated EPSCs. However, since in our STDP description individual EPSCs do not trigger any synaptic modification [102], then we may expect that only SICs sufficiently larger than EPSCs could effectively affect STDP. On the other hand, smaller SICs could also combine with Ca2+ transients by pre/post pairings, resulting in Ca2+ elevations beyond either LTD or LTP thresholds that would ultimately cause synaptic changes (Figures 7(a) and 7(b)). Hence, based on these considerations, we deem amplitude and timing of SICs, in terms of both frequency of occurrence and onset with respect to STDP-inducing stimuli, to be crucial factors in shaping how SICs affect STDP, and thus we set to analyzing these three factors separately.

Figure 7

STDP modulation by gliotransmitter-mediated SICs. ((a), (b)) Inspection of postsynaptic Ca2+ in the initial part of a pairing protocol that includes a gliotransmitter-mediated slow inward current (SIC) arriving to the postsynaptic neuron at t = 0.1 s illustrates how SICs have the potential to modulate postsynaptic Ca2+ thereby regulating LTD and LTP. (c) The magnitude of modulation depends on how large SICs are with respect to synaptic inputs (EPSCs) as well as at (d) what rate they occur. ((c), (d)) STDP curves were calculated for 60 pre/post pairings at 1 Hz and included SICs starting 0.1 s before the first pairing and occurring at 0.1 Hz. Synaptic parameters: τ = 0.33 s, τ = 0.33 s, and U 0 = 0.5 s. Other parameters as in Table 1 except for ϱ = 10−4, τ = 1 ms, τ = 200 ms, τ = 5 ms, and τ = 100 ms.

Figure 7(c) summarizes the results of our simulations for SICs as large as 0.5, 1, or 1.5 times typical EPSCs, occurring at a fixed rate of 0.1 Hz and starting 100 ms before the delivery of 60 STDP-inducing pre/post pairings at 1 Hz. As illustrated in Figures 7(a) and 7(b), for the same SIC kinetics, these simulations guarantee superposition between Ca2+ influxes mediated by SICs and pre/post pairings such that the extension of the ensuing Ca2+ transient beyond LTD and LTP thresholds (dashed lines) merely depends on SIC amplitude. In this fashion, it may be noted that SICs of amplitude smaller than or equal to typical EPSCs (Figure 7(c), turquoise circles and black circles, resp.), which alone would not produce any synaptic modification, do not sensibly change the STDP curve with respect to the previously considered case of an alike synapse in the absence of gliotransmission (Figure 6(c), black circles). Conversely, large SICs could dramatically affect STDP, shifting the STDP curve towards negative synaptic changes (blue circles), and this negative shift increases the larger SICs grow beyond the θ threshold (results not shown). In this case, STDP generally results in LTD with the exception of a LTP window that is comprised between ~0 ms and positive Δt values that are smaller than those in the absence of gliotransmission (Figure 6(c), green circles). This resembles what was previously observed for STDP curves in the presence of release-increasing gliotransmission, with the only difference that, for large |Δt| values, LTD strength in the presence of astrocyte-mediated SICs is found to be the same, regardless of Δt (compare the blue curve in Figure 7(c) with the green curve in Figure 6(c)).

In Figure 7(d) we consider the alternative scenario where only SICs as large as typical EPSCs impinge on the postsynaptic neuron at different rates, yet always 100 ms before STDP-inducing pairings. Akin to what happens for SIC amplitudes, the larger the SIC frequency is, the more the STDP curve changes. Indeed, as SIC frequency increases above SIC decay rate (i.e., 1/τ , Appendix A, Section A.1.4), SIC-mediated Ca2+ transients start adding up, so that the fraction of time spent by Ca2+ beyond the LTD threshold increases favoring LTD induction. In this fashion, the ensuing STDP curve, once again, consists of a narrow LTP window for Δt ≥ 0, outside which only LTD is observed (red curve). In practice however, because SICs occur at rates that are much slower than their typical decay (Appendix B), they likely affect STDP in a more subtle fashion. This may be readily understood considering the pink STDP curve obtained for SICs at 0.1 Hz, that is, the maximum rate experimentally recorded for these currents [22]. Inspection of this curve indeed suggests that SICs could effectively modulate LTD and LTP maxima as well as the outer sides of the LTD/LTP windows, which dictate how fast depression/potentiation decay for large |Δt|, but overall the qualitative features of the STDP curve are preserved with respect to the case without gliotransmission (black curve).

Clearly, the extent of the impact of SIC amplitude and frequency on STDP discussed in Figures 7(c) and 7(d) ultimately depends on when SICs occur with respect to ongoing STDP-inducing pairings. Had we set SICs to occur ~200 ms after pre/post Ca2+ transients in our simulations, then, as illustrated in Figures 8(a) and 8(b), we would have not detected any sensible alteration of STDP, unless SICs were larger than typical EPSCs and/or occurred at sufficiently high rate to generate Ca2+ transients beyond the plasticity thresholds (results not shown). To seek understanding of how timing of SICs versus pre/post pairings could alter LTD and/or LTP, we simulated STDP induction by pairing as the time interval (Δς) between SIC and pre/post pairs was systematically varied (with SIC rate fixed at 0.2 Hz) (Figures 8(c)–8(e)). In doing so, we adopted the convention that negative Δς values stand for SICs preceding pre/post (or post/pre) pairings while positive Δς values refer to the opposite scenario of SICs that follow pairings (Figure 8(c), top schematics). Then, it may be observed that, for Δς approximately in between −300 ms and 0 ms, LTD could be induced for any negative Δt as well as for large positive Δt (Figure 8(c), blue tones), in this latter case to the detriment of the LTP window, whose upper bound moves to lower Δt values (Figure 8(c), red tones). This results in STDP curves (e.g., Figure 8(f), yellow curve for Δς = −75 ms) that bear strong analogy with the blue and red curves in Figures 7(c) and 7(d), respectively, obtained for SICs of large amplitude and frequency and suggest that depression grows as SICs tend to concur with pre/post pairings. An inspection of postsynaptic Ca2+ transients (Figures 8(d) and 8(e)) indeed reveals that coincidence of SICs and pre/post pairings, which occurs at negative Δς of the order of SIC rise time (see Appendix B), corresponds to the longest time spent by Ca2+ above the LTD threshold, thereby resulting in maximum LTD (Figure 8(g)) and thus minimum LTP (Figure 8(h)). Clearly, the Δς range for which coincidence of SICs with pre/post pairings enhances LTD induction ultimately depends on kinetics of SICs, as reflected by their rise (τ ) and/or decay time constants (τ ), and spans Δς values approximately comprised within ± SIC duration (i.e., ≃τ + τ ). As SIC duration increases in fact, because of either larger τ or larger τ or both, so does the Δς range for LTD enhancement, as reflected by the orange and blue curves in Figures 8(f)–8(h).

Figure 8

Effect on STDP of SIC timing with respect to pairings. ((a), (b)) Impact on plasticity of a SIC occurring 0.2 ms after the first pairing instead of 0.1 s before it as previously considered in Figures 7(a), 7(b). (c) STDP curves as a function of the SIC pre/post pair delay (Δς) show how LTD could get stronger while the LTP window shrink for small-to-intermediate Δς ≤ 0 in correspondence with ((d), (e)) a maximum of the duration of Ca2+ transients above the LTD threshold. These results were obtained assuming SIC rise and decay time constants, respectively, equal to ms and ms. ((f), (h)) Peak and range of this LTD increase ultimately depend on SIC kinetics as reflected by the change of sample curves for specific Δς (yellow curve) and spike-timing intervals (cyan and purple curves) when SIC rise and/or decay time constants were slowed down 1.5-fold (orange and blue curves, resp.). ((c), (h)) The same pairing protocol of Figures 7(c) and 7(d) was used but with a SIC frequency of 0.2 Hz and variable SIC onset and kinetics according to Δς, τ , and τ . Parameters as in Figure 7.

In conclusion the simulations in Figures 8(c)–8(h) point to both timing and duration of SICs with respect to pre/post pairing-mediated Ca2+ transients as a further, potentially crucial factor in setting strength and polarity of STDP at glutamatergic synapses. It is noteworthy to emphasize that, however, to appreciate some effect on STDP, we had to assume in those simulations SICs occurring at 0.2 Hz, that is, twofold the maximum SIC rate (i.e., ~0.1 Hz) experimentally observed [22]. Indeed, analogous simulations run with realistic SIC rates ≤ 0.1 Hz did produce only marginal changes to STDP curves, akin to those previously observed for the pink STDP curve in Figure 7(d). The potential functional implications (or lack thereof) of this perhaps puzzling result are addressed in Discussion.

4. Discussion

A large body of evidence has accumulated over the last years suggesting an active role of astrocytes in many brain functions. Collectively, these data fuelled the concept that synapses could be tripartite rather than bipartite, since in addition to the pre- and postsynaptic terminals, the astrocyte could be an active element in synaptic transmission [1, 49, 117]. Using a computational modeling approach, we showed here that glutamatergic gliotransmission could indeed play several roles in synaptic information transfer, either modulating synaptic filtering or controlling postsynaptic neuronal firing, as well as regulating both short- and long-term forms of synaptic plasticity. Supported by experimental observations [8, 9, 13, 23, 118], these results complement and extend previous theoretical work on astrocyte-mediated regulations of synaptic transmission and plasticity [18, 29] and pinpoint biophysical conditions for a possible role of glutamatergic gliotransmission in spike-timing-dependent plasticity.

An important prediction of our model indeed is that both pathways of regulation of synaptic transmission by astrocytic glutamate considered in this study, presynaptic modulation of transmitter release and postsynaptic SICs, could affect STDP, potentially altering induction of LTP and LTD. This alteration could encompass changes in the timing between pre- and postsynaptic firing that is required for plasticity induction, as well as different variations of synaptic strength in response to the same stimulus. With this regard, the increase of LTP observed in our simulations, when moving from release-decreasing to release-increasing gliotransmission (Figure 6(e)), agrees with the experimental observation that LTP induction at hippocampal synapses requires weaker stimuli in the presence of endogenous glutamatergic gliotransmission rather than when gliotransmission is inhibited thereby decreasing synaptic release probability [9].

Notably, spike-timing-dependent plasticity in the hippocampus is not fully understood insofar as STDP induction by pairing protocols has produced a variety of seemingly contradicting observations for this brain region [119]. Recordings in hippocampal slices, for example, showed that pairing of single pre- and postsynaptic action potentials at positive spike-timing intervals could trigger LTP [120-122], as effectively expected by the classic STDP curve [123], but also induce either LTD [124] or no plasticity at all [121]. Although different experimental and physiological factors could account for these diverse observations [119, 125], we may speculate that glutamatergic gliotransmission by astrocytes, which in those experiments was not explicitly taken into account, could also provide an alternative explanation. For example, the prediction of our model that release-increasing glutamatergic gliotransmission could account for multiple LTD windows, at either positive or negative spike-timing intervals (Figure 6), indeed supports the possibility that LTD in the hippocampus could also be induced by proper presentations of pre → post pairings sequences [124]. On the same line of reasoning, the possibility that astrocyte-mediated SICs could transiently increase postsynaptic firing (Figure 5(f)) could explain why, in some experiments, precise spike timing in the induction of synaptic plasticity in the hippocampus could exist only when single EPSPs are paired with postsynaptic bursts [124, 126]. Moreover, it was also shown that postsynaptic firing is relatively less important than EPSP amplitude for the induction of STDP in the immature hippocampus compared to the mature network, possibly due to a reduced backpropagation of somatic APs in juvenile animals [121]. Remarkably, these diverse modes of plasticity induction could also ensue from different dynamics of glutamatergic gliotransmission, as likely mirrored by the developmental profile of somatic Ca2+ signals in hippocampal astrocytes [47], which have been reported to be much more frequent in young mice [127]. Insofar as somatic Ca2+ signals may result in robust astrocytic glutamate release that could trigger, in turn, similar increases of synaptic release and/or SICs [5, 62], the frequent occurrence of these latter could then ultimately guarantee a level of dendritic depolarization sufficient to produce LTP in mice pups [128].

High amplitude/rate SICs, or large increases of synaptic release mediated by glutamatergic gliotransmission, result, in our simulations, in LTD induction for any spike-timing interval except for a narrow LTP window at small-to-intermediate Δt > 0. This is in stark contrast with STDP experiments, where the observed plasticity always depends, to some extent, on the coincidence of pre- and postsynaptic activity, as EPSPs or postsynaptic action potentials fail to induce plasticity by their own [95, 105]. Apart from the consideration that large SIC amplitudes/rates and large increases of synaptic release by astrocytic glutamate may not reflect physiological conditions [76, 90], this contrast may be further resolved on the basis of the following arguments.

A first consideration is that we simulated plasticity induction assuming either persistent occurrence of SICs or continuous modulations of synaptic release during the whole induction protocol. While this rationale proved useful to identify the possible mechanisms of regulation of STDP by glutamatergic gliotransmission, it may likely not reflect what occurs in reality. Indeed, modulations of synaptic release by glutamatergic gliotransmission could last only few tens of seconds [7, 8] and thus be short-lived with respect to typical induction protocols which are of the order of minutes [33, 105, 129]. Moreover, the morphology of astrocytic perisynaptic processes is not fixed but likely undergoes dynamical reshaping in an activity-dependent fashion during plasticity induction [118, 130], thereby potentially setting time and spatial range of action of gliotransmission on nearby synaptic terminals [18]. In this fashion, LTD for large spike-timing intervals could be induced only transiently and at selected synapses, focally targeted by glutamatergic gliotransmission, while leaving unchanged the qualitative features of the classic STDP curve obtained by somatic recordings in the postsynaptic neuron [129].

A further aspect that we did not take into account in our simulations is also the possible voltage dependence of astrocyte-triggered SICs. The exact nature of this dependence remains to be elucidated and likely changes with subunit composition of NMDA receptors that mediate SICs in different brain regions and at different developmental stages [77]. Regardless, it may be generally assumed that slow inward currents through NMDA receptors become substantial only for intermediate postsynaptic depolarizations when the voltage-dependent Mg2+ block of these receptors is released [108]. In this fashion, the possible effect of SICs on STDP would be confined in a time window around Δt ≥ 0 for which coincidence with pre- and postsynaptic spikes allows for robust depolarization of postsynaptic spines. Outside this window instead, SICs would be negligible, and plasticity induction would essentially depend on mere pre- and postsynaptic spiking rescuing the experimental observation of no synaptic modification for large spike-timing intervals [95, 105].

On the other hand, even without considering voltage dependence of SIC-mediating NMDARs, the precise timing of SICs with respect to pre/post pairs is predicted by our analysis, to be potentially critical to determine strength and sign of plasticity. Similar considerations could also hold for the onset time and duration of modulations of synaptic release triggered by gliotransmission with respect to the temporal features of plasticity-inducing stimuli [29]. This ultimately points to timing of glutamate release by the astrocyte (and its downstream effects on synaptic transmission) as a potential additional factor for associative (Hebbian) learning, besides sole correlation between pre- and postsynaptic activities [131, 132]. Remarkably, this could also provide a framework to conciliate the possibility that modest, sporadic SICs that we predict would not substantially affect STDP (Figure 7) could do so instead [32]. Indeed our predictions are based on the average number of SICs within a given time window, as documented in literature, rather than on the precise timing of those SICs in that time window. In this fashion, for example, there is no distinction in terms of effect on STDP in our simulations, between a hypothetical scenario of three SICs randomly occurring on average every ~10 s in a 30 s time frame and the alternative scenario of three SICs taking place within the same time frame but in rapid succession (Figure 5b in [22]), as could happen following an exocytic burst of glutamate release by the astrocyte [61, 62, 65]. Yet the latter case could result in a dramatically different plasticity outcome with respect to the former. While individual SICs likely fail to induce synaptic modification alone in fact, their occurrence in rapid succession would instead allow postsynaptic Ca2+ levels to quickly increase beyond one of the thresholds for plasticity induction. Furthermore, this increase could further be boosted by coincidence of SICs with pre- and postsynaptic activity, ultimately accounting for robust LTP, as indeed predicted by other theoretical investigations [36]. However, to complicate this intriguing scenario is the observation that glutamatergic gliotransmission [65] and even more so astrocyte-mediated SICs [19, 25] are likely not deterministic but rather stochastic processes. Therefore, it would ultimately be interesting to understand how this stochasticity could affect neuronal activity and shape learning [133].

To conclude, our analysis provides theoretical arguments in support of the hypothesis that, beyond neuronal firing, astrocytic gliotransmission could represent an additional factor in the regulation of activity-dependent plasticity and learning [18, 134, 135]. This could occur in a variegated fashion by both presynaptic and postsynaptic elements targeted by glutamatergic gliotransmission, with possibly diverse functional consequences. Nonetheless, the practical observation in future experiments of a possible mechanism of action of glutamatergic gliotransmission on activity-dependent plasticity will depend on the implementation of novel specific plasticity-inducing protocols that match possible stringent temporal and spatial dynamical constraints defining the complex nature of neuron-astrocyte interactions.

Table 1

Symbol	Description	Range	Value	Units
Synaptic dynamics
τ d	Depression time constant	>0.01–2	s.s.	s
τ f	Facilitation time constant	>0.5–2	s.s.	s
U 0	Resting synaptic release probability	<0.09–0.9	s.s.	—
Neurotransmitter release and time course
Y T	Total vesicular glutamate concentration	300–1000	500	mM
ϱ c	Vesicular versus mixing volume ratio		0.005	—
τ c	Glutamate clearance time constant	2–100	25	ms
ζ	Efficacy of synaptic transmission	0-1	0.75	—
Astrocyte GPCR kinetics
O A	Agonist binding rate		0.3	μM−1s−1
τ A	Agonist unbinding time		0.55	s
IP3R kinetics
O 2	Inact. Ca2+ binding rate (with Ca2+ act.)	0.04–0.18	0.2	μM−1s−1
d 1	IP3 binding affinity	0.1–0.15	0.13	μM
d 2	Inact. Ca2+ binding affinity (Ca2+ act.)		1.05	μM
d 3	IP3 binding affinity (Ca2+ inact.)		0.9434	μM
d 5	Act. Ca2+ binding affinity		0.08	μM
Calcium fluxes
ϱ A	ER-to-cytoplasm volume ratio	0.4–0.7	0.18	—
C T	Total ER Ca2+ content	3–5	2	μM
ΩL	Max. Ca2+ leak rate	0.05–0.1	0.1	s−1
ΩC	Max. Ca2+ release rate by IP3Rs	>6	6	s−1
K P	Ca2+ affinity of SERCA pumps	0.05–0.1	0.05	μM
O P	Max. Ca2+ uptake rate	0.4–1.3	0.9	μM s−1
IP3 production
O β	Max. rate of IP3 production by PLCβ	0.05–2	1	μM s−1
K δ	Ca2+ affinity of PLCδ	0.1–1	0.5	μM
κ δ	Inhibiting IP3 affinity of PLCδ	1–1.5	1	μM
O δ	Max. rate of IP3 production by PLCδ	<0.8	0.05	μM s−1
IP3 degradation
Ω5P	Max. rate of IP3 degradation by IP-5P	>0.05–0.25	0.1	s−1
K D	Ca2+ affinity of IP3-3K	0.4–0.5	0.5	μM
K 3K	IP3 affinity of IP3-3K	0.7–1	1	μM
O 3K	Max. rate of IP3 degradation by IP3-3K	>0.6	4.5	μM s−1
Gliotransmitter release and time course
C θ	Ca2+ threshold for exocytosis	0.15–0.8	0.5	μM
τ G	Glutamate recycling time constant	0.003–1.5	1.66	s
U A	Resting glutamate release probability	<0.9	0.6	—
ϱ e	Vesicular versus mixing volume ratio		6.5 · 10−4	—
τ e	Glutamate clearance time constant	≤300	200	ms
Presynaptic receptors
O P	Activation rate	>0.3	1.5	μM−1s−1
τ P	Inactivation time constant	>30–180	120	s
ξ	Gliotransmission type	0-1	s.s.	—
Postsynaptic neuron
τ m	Membrane time constant	20–70	40	ms
τ r	Refractory period	1–5	2	ms
E L	Resting potential	−78.2–−54.8	−60	mV
v θ	Firing threshold	−55–−51	−55	mV
v r	Reset potential	−58–−53	−57	mV
v p	Peak AP amplitude	29.8–41.2	30	mV
R in	Input resistance	60–150	s.s.	MΩ
Postsynaptic currents
τ N r	EPSC rise time	0.4–0.6	0.5	ms
τ N	EPSC decay time	2.7–11.6	10	ms
J S	Synaptic efficacy		4.3	—
I S	EPSP amplitude	0.5–7.5	2	mV
Slow inward currents
τ S r	SIC rise time	20–70	20	ms
τ S	SIC decay time	100–800	600	ms
J A	SIC efficacy		68	—
I A	SIC amplitude	1–10	4.5	mV
Spike-timing-dependent plasticity
C pre	NMDAR-mediated Ca2+ increase per AP		1.0	—
τ pre r	NMDAR Ca2+ rise time		10	ms
τ pre	NMDAR Ca2+ decay time		30	ms
W N	Synaptic weight		39.7	—
C post	VDCC-mediated Ca2+ increase per AP		2.5	—
τ post r	VDCC Ca2+ rise time		2	ms
τ post	VDCC Ca2+ decay time		12	ms
C sic	SIC-mediated Ca2+ increase per AP		1.0	—
τ sic r	SIC Ca2+ rise time		5	ms
τ sic	SIC Ca2+ decay time		100	ms
W A	SIC weight		10.6	—
η	Amplification of NMDAR-mediated Ca2+		4	—
θ d	LTD threshold		1.0	—
θ p	LTP threshold		2.2	—
γ d	LTD learning rate		0.57	s−1
γ p	LTP learning rate		2.32	s−1
ρ ⋆	Boundary between UP/DOWN states		0.5	—
τ ρ	Decay time of synaptic change		1.5	s
σ	Noise amplitude		0.1	—
β	Fraction of synapses in the DOWN state		0.5	—
b	UP/DOWN Synaptic strength ratio		4	—