From membrane receptors to protein synthesis and actin cytoskeleton: mechanisms underlying long lasting forms of synaptic plasticity

Abstract

Synaptic plasticity, the activity dependent change in synaptic strength, forms the molecular foundation of learning and memory. Synaptic plasticity includes structural changes, with spines changing their size to accomodate insertion and removal of postynaptic receptors, allowing for functional changes. Of particular relevance for memory storage are the long lasting forms of synaptic plasticity which are protein synthesis dependent. Given the importance of spine structural plasticity and protein synthesis, this review focuses on the signaling pathways that connect synaptic stimulation with regulation of protein synthesis and remodeling of the actin cytoskeleton. We also review computational models that implement novel aspects of molecular signaling in synaptic plasticity, such as the role of neuromodulators and spatial microdomains, as well as highlight the need for computational models that connect activation of memory kinases with spine actin dynamics.

1.0. Introduction

Long term synaptic plasticity is the activity-dependent, long-lasting change in synaptic strength including both strengthening, called long term potentiation (LTP) and weakening, called long term depression (LTD). This change in synaptic strength is called functional plasticity as it influences the neuron’s output. A related phenomenon is spine morphological plasticity, an activity-dependent, long-lasting change in spine size, which is correlated with functional plasticity [1]. In fact, functional plasticity often involves enlargement and remodeling of the postsynaptic density of the spine (PSD) to allow glutamate receptor insertion. Thus, for the remainder of this article we include both morphological change and functional plasticity in the terms synaptic plasticity, LTP and LTD.

The two most commonly used methods to induce synaptic plasticity include frequency dependent protocols and spike timing dependent protocols. In frequency dependent protocols [2], the frequency of stimulation determines the outcome, e.g. high frequency stimulation produces LTP and low frequency stimulation may produce LTD. Spike timing dependent plasticity (STDP) involves pre-synaptic stimulation paired with precisely timed postsynaptic action potentials [3]. The timing of the pre- and postsynaptic stimulation determines the plasticity outcome: presynaptic stimulation preceeding postsynaptic action potential induces LTP; whereas the opposite timing causes timing dependent LTD. Numerous variations on these stimulation protocols have been developed, and the frequency or timing dependence of LTP and LTD vary by brain region and cell type [4]. One striking example is the observation that high frequency stimulation elicits LTP in hippocampal CA3 to CA1 pyramidal neuron synapses [5,6] versus LTD in striatal spiny projection neurons [7]. Another example is synaptic plasticity of parallel fiber synapses to cerebellar Purkinje cells, which is opposite that of CA3 to CA1 synapses, with low frequency stimulation producing LTP and higher frequency stimulation producing LTD [8].

Long term synaptic plasticity is an interesting phenomenon because of its relationship to learning and memory storage [9,10]. Several properties of synaptic plasticity resemble properties of learning and memory, such as the persistence of both memory and long term synaptic plasticity. The specificity of synaptic plasticity allows the neuron to change its response to the stimulation protocol (specific set or sequence of activated synaptic inputs) used to induce the plasticity, with minimal change in response to a different set of activated synaptic inputs [11]. This property is analogous to stimulus specificity observed in various types of associative learning [12].

Synaptic plasticity is often subdivided into three processes: induction, expression and maintenance [13] (Fig 1). Induction refers to the brief time frame (in the range of minutes) during which the stimulation protocol is applied and also to the biochemical mechanisms activated during the stimulation that produce the plasticity measured at a later time frame. Expression refers to the measurement of the change in synaptic strength anytime after the induction period. Maintenance refers to mechanisms occurring sometime after induction, which allow the synaptic plasticity to persist.

Figure 1.

Three processes involved in L-LTP. Induction of L-LTP refers to molecules transiently activated during the stimulation protocol and is blocked by inhibitors applied during stimulation, yielding E-LTP. Maintenance of L-LTP refers to molecules persistently activated after the stimulation protocol, which can be blocked even after L-LTP has been observed. Expression refers to the molecules required for measurement of the change in synaptic strength, and re-appears after inhibitors are removed.

In addition to these three processes involved in synaptic plasticity, LTP is also subdivided into early phase (E-LTP) versus late phase (L-LTP), which differ in their underlying biochemical mechanisms. The former is a shorter lasting form of LTP, independent of protein synthesis [14]. The latter form typically endures for hours to days and is dependent on protein synthesis [15]. Induction of L-LTP is sometimes confused with maintenance of LTP. A distinguishing feature is that experiments that investigate mechanisms involved in induction of L-LTP apply inhibitor molecules during the induction period [15,16], whereas experiments that investigate mechanisms involved in maintenance of L-LTP apply inhibitor molecules as long as an hour after the induction period [17].

2.0. Signaling pathways underlying long term synaptic plasticity

Across the brain the biochemical mechanisms underlying long term synaptic plasticity differ by region, cell type, synapse type and plasticity polarity (i.e. depression versus potentiation). In this review we summarize a few of the better known forms for which some recent computational modeling papers highlight novel spatial aspects of synaptic plasticity mechanisms.

2.1. Calcium and coincidence detection

Almost all forms of synaptic plasticity at excitatory synapses are initiated by activation of glutamatergic synapses and subsequent calcium elevation [2]. In most cell types, the source of calcium is influx through NMDA receptors (e.g. hippocampal CA1, striatum, neocortex). The voltage dependence of the NMDA receptor, mediated by the magnesium block [18], provides a concise explanation of the frequency dependence and spike-timing dependence of LTP. Both high frequency stimulation and back-propagating action potentials (bAP) occuring shortly after glutamate binding cause sufficient depolarization to relieve the magnesium block and allow calcium influx through the receptor [2]. In some cases, LTD requires influx of calcium through voltage dependent calcium channels, either for frequency dependent LTD [19] or timing dependent LTD [20]. In some cell types, such as cerebellar Purkinje cells, the co-activation of metabotropic glutamate receptors and calcium influx through voltage dependent calcium channels is required for calcium release from intracellular stores [21,22].

Understanding the voltage dependence of LTP and LTD is the subject of numerous computational models [23], though only recently have spatial models begun to explain the role of inhibitory synapses, or how the timing dependence is influenced by location of the synapse on the dendrite [24]. Computational models of the striatum have demonstrated that, in the absence of GABA, NMDARs are activated during pre-post protocols that produce LTP, whereas VDCC are preferentially recruited during post-pre protocols that produce LTD [25,26]. These models illustrate that the distance dependence of STDP is caused by the back-propagating action potential. Both striatal [25,27] and neocortical [28] models show that inhibitory inputs modify the distance dependence of calcium concentration. In neocortical models, inhibitory inputs can block timing LTP or transform it to LTD depending on spatial location. In striatal models, GABA can switch the polarity of STDP both by shunting the bAP during pre-post pairings, and contributing to NMDAR depolarization during post-pre pairings.

2.2. Signaling pathways downstream of calcium

The observation that LTP and LTD both require calcium elevations has inspired numerous experimental and computational studies to investigate how calcium controls polarity of synaptic plasticity (reviewed in [29,30]). Key computational models highlight an issue with a commonly held view that high calcium concentrations produce LTP and moderate calcium elevations produce LTD. If plasticity depends solely on the calcium amplitude, STDP protocols using a long pre-then-post interval would exhibit LTD [31], which is rarely seen experimentally. Since several molecules downstream of calcium are sensitive to the dynamics of calcium elevation, a plausible refinement of that hypothesis is that direction of plasticity is determined by both amplitude and duration of calcium transients. The requirement by LTD for a calcium elevation of sufficiently long duration has experimental support (reviewed in [29]) and several computational models implement various forms of duration detection [25,31–33].

Calcium activates several key molecules that are critical for synaptic plasticity (Fig 2). The main calcium binding protein is calmodulin [34], which has numerous enzymatic and non-enzymatic targets that are required for various forms of synaptic plasticity. An important non-enzymatic target for synaptic plasticity is neurogranin [35,36]. Because neurogranin binds to apo-calmodulin with a higher affinity than to calcium-bound calmodulin, neurogranin acts as a source of calmodulin during high calcium conditions [37]. Among enzymatic targets the most prominent ones are [38,39]: calcium-calmodulin dependent protein kinase type II (CaMKII), adenylyl cyclase (AC) [40], and protein phosphatase 2B, also called calcineurin. CaMKII phosphorylates AMPA type glutamate receptor channels [41,42], producing a change in channel properties [43], as well as other molecules required for protein synthesis dependent L-LTP [44]. In contrast, calcineurin opposes some of the actions of CaMKII by dephosphorylating AMPA receptors [45]. AC produces the second messenger cAMP, which activates protein kinase A (PKA). PKA phosphorylates a number of targets, including AMPA receptor channels, increasing the number of such channels at the synapse [46]; DARPP32; and inhibitor 1, both of which inhibit protein phosphatase 1 (PP1)[47,48] to enhance net kinase activity.

Figure 2.

Calcium activates several key molecules that are critical for synaptic plasticity. Calcium directly binds to calmodulin, calpain and PKC. Calcium bound calmodulin activates adenylyl cyclase (AC1/8), CaMKII, RasGRF, and calcineurin. Adenylyl cyclase, synergistically activated by Gs coupled receptors, produces cAMP, which activates PKA and Epac. Both RasGRF and Epac activate monomeric GTP binding proteins. CaMKII, PKA and PKC all phosphorylate AMPA receptors, whereas PP1 and calcineurin dephosphorylate AMPA receptors. PKAc: catalytic subunit of PKA.

Several other calcium activated molecules play a role in various forms of synaptic plasticity. One is diacylglycerol lipase, required for LTD in many brain regions [7,49]. Diacylglycerol lipase produces the endocannabinoid 2-arachidonylgylcerol, which retrogradely targets the pre-synaptic cannabinoid type 1 receptor to reduce probability of neurotransmitter release [50]. Another calcium activated protein is protein kinase C (PKC), required for LTP in the striatum [51], LTD in the cerebellum [22,52] and LTP maintenance in the hippocampus [17]. The typical PKC isoform requires both calcium and diacylglycerol for activation [53], whereas a constitutively active form of PKC, called PKMζ, is the only kinase shown to be involved in LTP maintenance [17] and is produced either by synthesis or cleavage of other PKC forms. A third and more recently implicated calcium activated molecule is calpain, which is a protease activated by high calcium concentrations. Calpain has several targets, including molecules controlling protein synthesis [54] and actin cytoskeleton remodeling [55].

Though calcium is clearly critical for synaptic plasticity, signaling pathways downstream of G protein coupled receptors, activated by neuromodulators, are additionally required for synaptic plasticity in several brain regions, including amygdala [56], auditory cortex [57], and prefrontal cortex [58,59]. In vivo, neuromodulators are critically important as they carry information on reward, novelty, salience, emotion, and other factors that influence learning [60,61]. Although the neuromodulators dopamine, norepinephrine and acetylcholine are released in numerous brain regions, the downstream intracellular signaling molecules depend on the identity of the postsynaptic receptors and the type of triheteromeric GTP binding protein activated.

In the striatum, LTP is critically dependent on the dopamine D1 receptor in the direct pathway spiny projection neurons (SPN) and the adenosine A2A receptor in the indirect pathway SPNs [62]. Both types of receptors are coupled to Gs proteins, which lead to PKA activation, which is required for striatal LTP [51,63,64].

In the hippocampus, β-adrenergic receptor (βAR) stimulation and subsequent Gs activation is required for several forms of LTP [65]. In addition, βAR can switch their coupling from Gs to Gi [66], recruiting extracellular signal-regulated kinase types 1 and 2 (ERK) through an atypical signaling pathway. Experiments using novel βAR antagonists together with spatial stochastic modeling suggest that βAR may be required for all forms of hippocampal LTP [37,67].

Neuromodulators also are required in the neocortex. For example, STDP in visual cortex requires stimulation of both Gs and Gq coupled receptors [68,69]. Independent of spike timing, stimulation of βAR alone yields LTP, whereas stimulation of muscarinic M1 or adrenerginc α1 receptors produces LTD. Stimulation of both receptors simultaneously is required for the polarity of synaptic plasticity to be dependent on spike timing. In the prefrontal cortex dopamine is needed for both frequency dependent plasticity and STDP. Dopamine can change the temporal window of STDP [59], and with frequency dependent protocols, phasic dopamine release determines whether LTP will occur, whereas tonic dopamine controls the direction and magnitude of synaptic plasticity [70].

2.3. Models of signaling pathways: neuromodulators, spatial aspects and stochasticity

An early hypothesis proposed that information in the brain is stored in the form of molecular switches [71]: a molecular network that can enter (or exit) a long-lasting stable state in response to transient input such as calcium influx. In consequence, molecular activity is bistable with one state of low molecular activity (e.g. corresponding to no information stored or possibly corresponding to LTD) and one state of high molecular activity (e.g. corresponding to LTP). Support for this hypothesis was provided by CaMKII, a multi-subunit molecule highly expressed in the forebrain [72] and activated by calmodulin. More importantly, CaMKII can autophosphorylate itself at Thr-286 (i.e., an active CaMKII subunit can phosphorylate an adjacent calmodulin-bound subunit) preserving its activity after calcium concentration decreases and calmodulin dissociates. This property allows CaMKII to form a molecular switch even in the presence of phosphatases [73]. The ability to phosphorylate adjacent calmodulin bound subunits also conveys to CaMKII a strong dependence on frequency of calcium transients [74], which may explain how the direction of synaptic plasticity depends on stimulation frequency.

Numerous models have investigated the frequency dependence and other properties of CaMKII. Some recent models of synaptic plasticity [75–78] focus on understanding the biochemical mechanisms underlying the dependence of synaptic plasticity on stimulation frequency. These models show that with high calcium concentration (produced by high frequency stimulation), kinase activity dominates and elicits LTP, whereas for moderate calcium concentration (produced by low frequency stimulation) calmodulin preferentially binds to calcineurin and produces LTD [79]. More sophisticated models incorporate more complex interactions between CaMKII and phosphatases, such as calcineurin or PP1, as the foundation of a molecular switch [79–82]. All of these models can qualitatively reproduce results of plasticity outcomes in response to STDP and frequency dependent stimulation protocols. Some recent CaMKII models have incorporated an experiment that demonstrates activity of sub-saturated calmodulin. Though calmodulin requires four calcium ions for full activity, calmodulin bound to two calcium ions has partial activity, e.g. calmodulin bound to two calcium ions can activate CaMKII [83]. In the physiological range of calcium concentration, this sub-saturated calmodulin contributes significantly to CaMKII activation [75,76]. Since there are more calmodulin binding sites than calmodulin molecules, the competition for fully and partly calcium-bound calmodulin influences which downstream molecules are active [77]. Another question is how CaMKII could maintain its activity despite protein turnover, which would exchange a phosphorylated subunit for an unphosphorylated subunit. A recent discovery is that CaMKII holoenzymes exchange dimer subunits with unactivated holoenzymes [84,85]. A computational model [86] confirmed that CaMKII subunit exchange improves the stability of the active state in the PSD, and slows the decay of CaMKII activity in the spine cytosol, even in the presence of protein turnover. Further theoretical implications of diverse timescales of CaMKII activity in the synapse remain to be investigated.

Despite the elegance of these models, there is some question whether CaMKII actually functions as a bistable switch [87,88] and even the first models of CaMKII switching stated that only in the PSD could high CaMKII activity be maintained after calcium concentration returned to its resting value [65]. More importantly, experimental evidence, described below, shows that other molecules can play a pivotal role in synaptic plasticity, and that stochastic flucutations, molecule spatial location or spatial pattern of inputs are critical for accurate simulations of biochemical pathways in neurons.

Stochastic fluctuations are significant because the number of molecules in dendritic spines is quite small; thus the biochemical reactions underlying most forms of synaptic plasticity involve interactions between small numbers of molecules. Under these conditions, molecule concentration has less meaning, and interactions occur randomly, requiring stochastic simulation techniques. A stochastic model of the signaling pathways implicated in cerebellar plasticity [89] demonstrates the all-or-none nature of LTD in Purkinje cells and serves as an example of the critical role played by stochastic fluctuations. The switch like behavior is produced by a feedback loop involving phospholipase A2, PKC, and ERK (discussed more below). Simulations demonstate switch activation exhibits all or none behavior, with the probability of activation increasing with calcium concentration [90,91].

Research is increasingly showing that scaffolding or adaptor proteins are crucial for producing multi-protein complexes and co-localizing key molecules in specific spatial compartments [67,92,93]. Two recent models of cerebellar LTD have shown that molecules involved in AMPAR trafficking [94] or AMPAR interactions with PKC [91] can regulate LTD induction. Computational models of striatal [95] and hippocampal LTP [96] demonstrate that the role of A Kinase Anchoring Proteins is to colocalize the site of cAMP near its target PKA, due to the strong inactivation mechanisms for cAMP that prevent it from diffusing too far. These models also investigate spatial specificity of signaling models by simulating a dendrite with numerous spines. In the striatal signaling pathway models, LTP requires PKA and PKC, the latter of which binds to membrane lipids (e.g. diacylglcerol) as part of its activation; whereas LTD requires production of diffusible endocannabinoids. Model simulations reveal that activation of molecules required for LTP is mostly confined to the stimulated spines, whereas endocannabinoids are spatially diffuse, suggesting that spatially specific LTP induction would be accompanied by non-specific LTD.

2.4. Signaling pathways converge onto MAPK/ERK

ERK, a member of the mitogen activated protein kinase (MAPK) family, seems to be the key integrative molecule required for all forms of LTP [97]. The canonical ERK pathway (Fig 3) is initiated by brain derived neurotrophic factor (BDNF) binding to a receptor tyrosine kinase (TrkB) [98], which, through several intermediary molecules, recruits a guanosine exchange factor (GEF) that activates a monomeric GTP binding protein called Ras [99]. A common motif in ERK activation is activation of a Ras family protein, followed by a series of three kinases: Ras binds to a MAP kinase kinase kinase (e.g. CRaf or BRaf) which phosphorylates a MAP kinase kinase (i.e. MEK) which then phosphorylates ERK itself [100]. In addition to BDNF activation of ERK, several other signaling pathways bypass TrkB receptors but nonetheless converge on ERK phosphorylation (Fig 3).

Figure 3.

Numerous signaling pathways contribute to ERK activation. cAMP contributes to ERK through activation of Epac and PKA. Calcium activated molecules such as RasGRF and PKC can activate ERK through activation of Ras family proteins or phosphorylation of Raf isoforms. BDNF binds to TrkB receptors and then activates Ras family proteins through several scaffolding proteins. In all cases, a series of three events activates ERK: Ras family protein binding to the kinase Raf, Raf phosphorylating MEK at two sites, and then MEK phosphorylating ERK at two sites. Dashed lines indicate that additional molecules may be intermediaries.

Several signaling pathways involving cAMP have been shown to contribute to ERK activation and play a role in synaptic plasticity. PKA can activate ERK through phosphorylation of the monomeric GTP binding protein Rap1 [101], src [102], Rap1Gap [103] or βAR [66], which then switches coupling to Gi and recruits ERK without BDNF. In addition, cAMP activates exchange factor activated by cAMP (Epac) which in turn activates Rap [104]. In Schaffer collateral synapses, Epac activation influences LTD [105]; and also contributes to βAR-dependent L-LTP [106]. Calcium activated pathways also contribute to ERK activation. CaMKII can enhance ERK through phosphorylation of SynGAP, which reduces its GTPase activity against Ras and Rap [107]. Calmodulin activates another guanosine exchange factor, called RasGRF contributing to ERK phosphorylation through Ras activation [108].

Several critical questions regarding ERK activation and spatial specificity of ERK signaling have been addressed with computational models. The first two questions address opposing aspects of spatial specificity of ERK signaling: how can spatial specificity of ERK phosphorylation be maintained with diffusible molecules and how can phospho-ERK propagate from the site of stimulation along the dendrite [109]. A model implementing PKA activation of ERK [110] showed that dendritic geometry and phosphodiesterase concentration controlled the steepness of the phospho-ERK gradient: with low phosphodiesterases, cAMP could diffuse widely, producing a small PKA gradient resulting in a small phospho-ERK gradient. A model of the feedback loop between phospho-ERK and cell excitability [111] showed that the diffusion of phospho-ERK is slow and for phospho-ERK to propagate along the dendrite (analogous to the dynamics of action potential propagation) a positive feedback loop is required: calcium leading to phospho-ERK, which then phosphorylates potassium channels increasing cell excitability and enhancing calcium influx. A later spatial deterministic model of these signaling pathways [112] showed that the same ultrasensitive behavior of ERK coupled with diffusion of upstream activators produced more ERK phosphorylation when nearby spines were stimulated in a spatial order as compared to a random order. Note that none of these models has evaluated the role of BDNF, RasGRF, SynGAP, Epac or βAR switching from Gs to Gi in ERK activation, yet experimental results suggest that the specific pathway recruiting ERK depends on stimulation pattern. Furthermore, it is likely that the different pools of ERK are activated by different signaling pathways [113,114], emphasizing the importance of modeling spatial aspects of neurons to fully understand the role played by ERK in synaptic plasticity.

3.0. From memory kinase to translation

Numerous experiments demonstrate that memory storage requires protein synthesis; thus to understand the synaptic plasticity underlying memory storage, it is necessary to elucidate how memory kinases regulate protein synthesis. Experiments have identified several points of control in both initiation and chain elongation, which are modified by ERK, CaMKII and PKMζ. ERK phosphorylates MAPK interacting kinase (MNK1), which phosphorylates several components of the critical protein complex eukaryotic initiation factor 4F (eIF4F) and S6 Kinase [115]. A second pathway regulating eIF4F is PI3 kinase, which activates Akt, which activates mTOR. Similar to ERK, PI3 kinase is activated by BDNF binding to TrkB receptors and subsequent Ras activation, and both mTOR and ERK additionally lead to phosphorylation of eukaryotic elongation factor 2 Kinase (eEF2K) which modifies the rate of chain elongation. CaMKII regulates protein synthesis through phosphorylation of cytoplasmic polyadenylation element (CPE)-binding protein (CPEB), which stimulates initiation of translation of some mRNAs in dendrites [116].

Jain and Bhalla [117] investigated how three activity dependent kinases: ERK, mTOR and CaMKIII, could control protein synthesis without producing the unstable situation of new proteins causing yet additional protein synthesis. Simulations using their carefully constructed model show that BDNF gates protein synthesis induced by LTP stimuli, and that calcium-calmodulin mediated inhibition of eEF2 (via CaMKIII) prevents more than transient increases in protein synthesis in the absence of BDNF. Another model [118] also investigates interactions between calcium activation of ERK, mTOR and eEF2K. In contrast to [117], ERK and mTOR inhibit eEF2K, whereas calcium-calmodulin directly activates eEF2K. Despite these differences, a similar conclusion is reached, that activation of these pathways produces only transient increases in protein synthesis. This second model further suggests that an additional regulator, i.e. AMP kinase, is required, demonstrating that the cell achieves tight control of protein synthesis by requiring coordinate regulation.

Several models have evaluated the biochemical mechanisms underlying the bistability of PKMζ, the constitutively active form of PKC that is responsible for LTP maintenance [17]. The candidate mechanisms include PKMζ initiating its own synthesis [119]; PKMζ autophosphorylation prolonging its residence time in synapses [120]; and phosphorylation of AMPAR, which increases AMPAR insertion in the membrane, which decreases either the degradation or diffusion of PKMζ [121]. All three mechanisms can reproduce experiments investigating time dependent effects of protein synthesis and PKMζ inhibitors. These models suggest that the PKMζ positive feedback loop is a very robust mechanism for plasticity maintenance. Only one model investigating the PKMζ feedback loop has implemented multiple spatial compartments [122] explaining mechanisms of synaptic tagging and capture [123], in which the synthesis of plasticity related proteins are targeted to stimulated spines (and not to non-stimulated spines). Simulations showed that the local, spine-specific increase in PKMζ maintains itself only in stimulated spines, whereas the concentration of PKMζ in non-stimulated spines never reaches the concentration required to switch to the high activity state.

4.0. From memory kinases to actin cytoskeleton

How is functional plasticity related to morphological plasticity? A well supported hypothesis is that LTP involves insertion of new AMPAR [124] at the PSD, into “slots” comprised of various structural and anchoring proteins, such as PSD-95. Thus, a change in spine size is needed to provide space for new AMPAR, and this requires remodeling of the branched network of actin filaments to re-build the spine actin cytoskeleton [125]. Synaptic plasticity induction protocols do indeed induce rapid remodelling of the actin cytoskeleton in the dendritic spines [126] to allow for increases in postsynaptic responses. A critical question remains unanswered: how do memory kinases or plasticity-related neuromodulators control remodeling of the actin cytoskeleton?

Actin filaments (f-actin) are dynamic structures, with actin monomers (known as g-actin) being added to the barbed end (polymerization) while simultaneously actin monomers are being removed from the pointed end (depolymerization) [127]. When these rates are balanced, the lengths of the actin filaments are maintained. f-actin is controlled directly by several sets of molecules including severing proteins such as cofilin, capping proteins such as capZ, branching proteins such as Arp2/3, and cross-linking/bundling proteins such as CaMKIIβ [127] (Fig 4). Both severing proteins and capping proteins accelerate f-actin disassembly: cofilin cuts the actin filament creating two shorter filaments; capZ attaches to the barbed end, and prevents addition of g-actin monomers, leading to shortening of f-actin because monomers continue to be removed from the pointed end. Paradoxically, cofilin can also promote f-actin lengthening by severing the capped actin and revealing new barbed ends to restart the polymerization process. The branching protein A complex (Arp2/3) allows for initiation of a new branch at an angle to an existing actin filament; thus it is instrumental in shaping the spine heads. Lastly, several cross-linking proteins, such as CamKIIβ and drebrin, stabilize the actin network by binding two filaments (or a filament and proteins at the PSD) and preventing binding of other actin binding proteins.

Figure 4.

Several memory kinases interact with actin binding proteins to remodel the spine actin cytoskeleton. CaMKIIβ both cross-links f-actin and buffers g-actin monomers. Binding of CaMKII to calmodulin decreases g-actin binding making g-actin monomers available and facilitating cofilin cleavage of f-actin. Both CaMKII and PKA enhance Kalirin7 activity through phosphorylation, allowing it to activate monomeric GTP binding proteins Rac1 and cdc42. Their targets include WAVE and WASP members of the ARP2/3 complex (needed for new actin branches) and PAK. Both PAK and PKA phosphorylate (to inactivate) LIMK, which phosphorylates (to inactivate) cofilin. Calcineurin dephosphorylates (to activate) SSH which then dephosphorylates (to activate) cofilin. PKAc: catalytic subunit of PKA.

How can actin filaments be dynamically stable -- maintaining their length while actin monomers are continually being added and removed? This key question was investigated by several computational models of actin dynamics (reviewed in [128]. Roland et al [129] reveal that both the concentration of actin monomers and the severing activity of cofilin are critical determinants of the steady state filament length. Berro et al. [130] demonstrate that binding of capping protein and Arp2/3 must be faster than measured in vitro to account for in vivo measurements of actin filament assembly and disassembly. Another model [131] explains spatial gradients of actin polymerization in lamelipodia of motile cells, showing that, when active WASP is confined to the leading edge of a lamelipodium, then Arp2/3 and its dependent branches exhibit elevated dissociation just behind the leading edge.

How do memory molecules remodel the actin cytoskeleton and exert control over spine morphology during and after plasticity induction? Live cell imaging has revealed transient changes in spine head concentration of several kinases and actin binding proteins during structural LTP induction [132]. The first step in remodeling the spine actin network is unbundling the actin filaments, which are normally crosslinked by unphosphorylated CamKIIβ [133], which also serves as a source of g-actin monomers. During LTP induction, CaMKIIβ is activated by calmodulin binding and CaMKIIα phosphorylation. CaMKIIβ then dissociates from g-actin making monomers available for growing the network. Dissociation of CaMKIIβ and α-actinin4, another cross-linking protein binding to actin and PSD proteins [134], allow access to actin by other actin binding proteins.

The next step after f-actin unbundling is f-actin severing by cofilin, which is controlled by LTP signaling networks through several points. Cofilin is activated during plasticity induction via dephosphorylation by slingshot phosphatase (SSH), which is a target of calcineurin. Subsequent to its activation and enhanced severing activity, cofilin is inactivated by phosphorylation by LIM Kinase (LIMK), either to allow more efficient actin branch creation (for LTP) or to stabilize the spine at a smaller size (for LTD). LIMK activity is controlled by both CaMKII and PKA, both directly and indirectly. PKA can directly phosphorylate (to activate) LIMK, and both PKA and CaMKII phosphorylate Kalirin7 and other GEFs [135,136], which in turn activate monomeric GTPases Rac1 and cdc42, which lead to LIMK phosphorylation through PAK [136].

Rac1 and cdc42 have the additional role of initiating the third part of actin remodling by activating the WAVE/WASP/IRSp35 complex directly upstream of Arp2/3 [137], which is necessary to increase polymerization and branching of the spine actin cytoskeleton. The fact that these two GTPases initiate both actin severing and actin branching suggests that the timing of activation of the different pathways is crucial for increases in spine size. A third monomeric GTPase protein, RhoA, opposing cdc42 and Rac1, activates RhoA Kinase to activate myosin [138] leading to decrease in spine size, due to the motor actions of myosin.

Following remodelling, the spine actin cytoskeleton needs to be stabilized. CaMKIIβ, inactivated due to the decrease in calcium-bound calmodulin, and α-actinin crosslink the new actin filaments. If CaMKIIβ has been phosphorylated on Thr287, then calmodulin may be trapped, and dephosphorylation would be required for crosslinking. Oligophrenin and SynGAP, both GTPase accelerating proteins enriched in the spine, inactivate Rac1 and cdc42 [139]. The importance of SynGAP has been demonstrated experimentally: spine enlargement can be mediated by CaMKII triggered SynGAP dispersion from dendritic spines, and the degree of SynGAP dispersion correlates with the maintenance of spine enlargement [140].

Even in the small volume of the spine, the spatial localization of molecules is critical. SynGAP interacts with PSD-95 to form highly concentrated liquid-like droplets representing a separation of the SynGAP/PSD-95 complex from the cytosol, reminiscent of the formation of ‘membrane-less’ intracellular compartments [141]. Localization of SynGAP to this compartment is critical for proper activity-dependent SynGAP dispersion from the PSD, AMPAR insertion and spine morphological plasticity. In addition, at the PSD CaMKII dephosphorylation may be impaired if PP1 is kept away from CaMKII [142].

Interaction between memory molecules and actin is not a one-way street: the state of the actin skeleton also influences the activation of signaling pathway molecules. Reaction rates between molecules are often limited by the diffusion rates, which control the frequency of encounter between reacting molecules. Diffusion within the cytosol can be dramatically hindered if the environment is crowded with numerous other molecules. For example, in the PSD, protein crowding can impede the diffusion of glutamate receptors [143]. Similarly, a dense actin network can impede diffusion of signaling molecules within the spine. Alternatively, molecular crowding may facilitate reactions by colocalizing reacting molecules and preventing them from diffusing away. Experimental evidence suggests that unbundling of actin transiently modifies diffusion of molecules in and out of the spine [132], with an increase in cofilin diffusion but a decrease in CaMKIIβ diffusion, and highlights the complexity of molecular interactions within a spine. Computational modeling of molecular crowding is a critical but nascent field, with new algorithms being developed to facilitate investigation [144–148].

The pioneering computational model bridging activation of signaling pathways with dynamics of the actin cytoskeleton [149] demonstrates how a single molecule, CaMKII, by virtue of different time constants for activation versus inhibition reactions, can produce LTP and a sustained increase in spine size. Extension of this model, to include control by cAMP activated pathways or to investigate heterosynaptic effects, promises to contribute greatly to our understanding of how synaptic plasticity depends on the interaction between memory kinases and the actin cytoskeleton.

5.0. Conclusions

Models of signaling pathways underlying long-term synaptic plasticity have revealed or confirmed interesting integrative and non-linear properties of diverse sets of molecules. However, more extensive modeling of the control of protein synthesis and actin reorganization is necessary to better understand the processes underlying memory storage. Increased efforts in modeling of spatial aspects of signaling pathways are especially important, such as interactions and differences between spine and dendrite, signaling from the synapse to the soma, and even emergence and function of sub-micron “membrane-less” compartments within the spine. Both experiment and theoretical studies have demonstrated that a crowded environment significantly changes both diffusion and reaction rates, which makes it crucial to include physical aspects of the spine environment such as molecular crowding in compulational models. The challenge to future models is to increase biophysical fidelity, as experiments continue to demonstrate increased complexity of neurons, while maintaining the elegance of simple models that illuminate principles of neuron function.