An emerging view of structure–function relations of synapses in central spiny neurons asserts that larger spines produce large synaptic currents and that these large spines are persistent (‘memory’) compared to small spines which are transient. Furthermore, ‘learning’ involves enlargement of small spine heads and their conversion to being large and stable. It is also assumed that the number of spines, hence the number of synapses, is reflected in the frequency of miniature excitatory postsynaptic currents (mEPSCs). Consequently, there is an assumption that the size and number of mEPSCs are closely correlated with, respectively, the physical size of synapses and number of spines. However, several recent observations do not conform to these generalizations, necessitating a reassessment of the model: spine dimension and synaptic responses are not always correlated. It is proposed that spines are formed and shaped by ongoing network activity, not necessarily by a ‘learning’ event, to the extent that, in the absence of such activity, new spines are not formed and existing ones disappear or convert into thin filopodia. In the absence of spines, neurons can still maintain synapses with afferent fibers, which can now terminate on its dendritic shaft. Shaft synapses are likely to produce larger synaptic currents than spine synapses. Following loss of their spines, neurons are less able to cope with the large synaptic inputs impinging on their dendritic shafts, and these inputs may lead to their eventual death. Thus, dendritic spines protect neurons from synaptic activity-induced rises in intracellular calcium concentrations.
It has been postulated that dendritic spines underlie the neuronal locus of plasticity, in which long-term alterations in synaptic strength (‘memory’) are converted into persistent morphological changes. While ongoing studies attempt to characterize the nature of these morphological changes and the molecular cascades leading to them (Bhatt et al., 2009; Yoshihara et al., 2009; Holtmaat & Svoboda, 2009; Yang & Zhou, 2009; Segal, 2005), it is still not clear what constitutes a ‘memory’ at the spine level, if at all such a function can be assigned to a single dendritic spine. One major issue is whether spine morphology follows changes in ambient network activity, or does it genuinely store ‘memory’ which can be formed even after a single association between two neurons firing, irrespective of the ongoing background activity. Recent studies have addressed this issue in either time-lapse imaging of the same spine arrays over extended periods of time, or in a detailed electron-microscopic analysis of 3-D-reconstructed dendritic spines in control and following a learning experience or a long-term potentiation (LTP)-inducing tetanic stimulation. In the former group of studies, individual dendritic spines could be activated by electrical stimulation or by photoconversion of caged glutamate (Harvey & Svoboda, 2007; Lee et al., 2009), to find that stimulation of a single spine can cause a nearly immediate expansion of the spine head volume by 3–4-fold (Matsuzaki et al., 2004). This effect was dependent on activation of the NMDA receptor and its maintenance, at a lower level than the original expansion, was dependent on activation of kinases (Yang et al., 2008). The spine expansion preceded the electrophysiological change, which progressed at a slower time course, and the change in spine volume was much smaller than in the original report (Yang et al., 2008). These studies illustrate the ability of spines to change their volume over a short period of time after exposure to a massive excitatory stimulation. On the other hand, such a massive increase in spine volume was not seen by others, who found a slow change in volume following a massive activation of glutamate receptors (Sapoznik et al., 2006), or no change at all, even in conditions in which the activation of the spine followed a pairing protocol for induction of LTP (Nevian & Sakmann, 2006). The difference between such observations on spine head expansion may have to do with the insertion of glutamate receptors into the spine heads such that only the spines to which glutamate receptors are added into their heads will expand (Kopec et al., 2007; Korkotian & Segal, 2007) while others will not. Even this expansion is rather slow, and cannot underlie the nearly immediate expansion of spine heads reported previously (Matsuzaki et al., 2004).
More recently, a persistent change in spine number (but not in their volume) in the mouse neocortex has been seen following extensive motor learning; the change lasted over many days after initial training (Yang et al., 2009; Xu et al., 2009). While these results are technologically impressive they do not relate specific spines to specific neuronal activity, to the extent that beyond the correlation between performance and spine number there is no clear indication that these additional spines participate in the enhanced network activity resulting from the training. Still, these studies did not show a dramatic change in spine volume as predicted by the earlier studies.
The other approach, which involves comparisons of populations of spines using 3-D electron-microscopic reconstructions of spines, was used extensively both in vivo and in vitro (Stewart et al., 2005; Medvedev et al., 2010). The advantage of this approach is in the ability to measure accurately the size of the pre- and postsynaptic compartments of the synapse, to verify that the spine is indeed innervated by a presynaptic terminal and to measure a large number of spines in different parts of the neuron at a given time point. These studies employed either electrical stimulation, which produces LTP in a selective pathway, or chemical LTP, which is likely to activate most if not all of the synapses. In general, these studies did not reveal massive changes in spine head volume, although changes in postsynaptic density and changes in the proportion of thin to mushroom spines were noted (Medvedev et al., 2010). In all, these studies demonstrate that populations of spines can shift to having larger spine heads following a tetanic stimulation of an afferent pathway, and it is possible that large changes in spine volume take place in a small subset of spines, although this is not seen in the averaged data.
Assuming that spine volume does change after a specific intense stimulation, it is still not clear what are the relations between spine shape, size and density and ambient network activity: do spine shapes vary in a dynamic fashion as a function of ambient activity, such that an increase in activity results in an increase in spine size or density and, conversely, a decrease in activity results in elongation of spines and a collapse of their heads. Alternatively, if spines model ‘memory’ irrespective of ambient activity, then once a spine is formed following a specific ‘pairing’ it should persistent irrespective of ongoing activity. These two conditions assume opposite demands on the spines, to constantly change their morphology or be stable and store a ‘memory’. This issue is difficult to address directly, but some of the following studies are relevant to this issue.
One of the factors that contribute to the difficulty in generalizing some rules that govern the behavior of spines is the different preparations, ages and imaging conditions used. Obviously, when one images remote dendrites of young cortical neurons in vivo, where spine density is rather low, one cannot expect to generalize a priori to mature, highly spiny neurons recorded in an acute slice or in a cultured slice. The heterogeneity is built into the spine, and any attempt to produce a ‘rule’ has to consider different conditions, ages and preparations. The following provides some illustrations of this complexity.
Spine formation is dependent on active afferent innervation
The role of ambient activity in formation and maturation of dendritic spines can be learned from the order of events that take place during spine formation and maturation. One yet unresolved issue is, do mature spines form from transient filopodia, which are likely to produce a smaller synapse and a smaller impact on the parent neuron than spines, or are spines formed from dendritic shaft synapses? If indeed spines are formed from shaft synapses, then one can assume that moving the synapse away from the shaft may reduce its efficacy, compared to a shaft synapse, hence a spine synapse would influence the parent neuron less than would a shaft synapse, and thus the spine can be considered as having a neuroprotective value. It is generally assumed that in the developing neuron a filopodium is formed first; following establishment of contact with an afferent fiber, it retracts and becomes a spine (Fiala et al., 1998; Sorra & Harris, 2000). In this case the outcome would be viewed as an increase in the efficacy of synaptic transmission. However, stable synaptic connections leading to spontaneous network activity have also been seen in young neurons (3–4 days in vitro) even before the formation of spines, and these synapses are formed primarily on dendritic shafts (Lauri et al., 2003). Likewise, it is not entirely clear that the process of conversion of filopodia to spines is a necessary step in an already mature neuron, where filopodia are rare and spines can form and dissolve within hours, as shown in estrus-cycling female rats (Woolley & McEwen, 1993) and during recovery from hibernation (Popov & Bocharova, 1992; Popov et al., 2007), as well as in time-lapse microscopy in adult mice (Xu et al., 2009; Yang et al., 2009). On the other hand, within hours following activity blockade with tetrodotoxin (TTX), filopodia grow off existing spines, indicating that they are being used as a means of searching for glutamate-releasing presynaptic terminals (Richards et al., 2005). Thus, with a few exceptions, it can be concluded that spines can be formed from shaft synapses, and the presence of spines reduces rather than enhances the impact of an individual synapse on the activity of the parent neuron.
A corollary issue is whether a neuron loses its synapses when spines are pruned, just to regain them when the spines reappear, or whether it retains the synapses with its afferent terminals, which may form shaft synapses? Intuitively, a synapse which is rich in adhesion molecules crossing between pre- and postsynaptic membranes has a bond strong enough to resist mechanical dissociation of the tissue (e.g. during preparation of synaptosomes). Why then should the synapse lose the presynaptic partner just because it retracts by a few micrometers only to reappear a day later, as is the case with the estrus cycle? Recent electron-microscopic data indicate that spine-pruned cortical neurons do lose their connection with afferent inputs (Knott et al., 2006). On the other hand, in hibernating animals there is a marked decrease in spine density during hibernation but there is an increase in shaft synapses (Popov et al., 2007; von der Ohe et al., 2006), and when the animals wake up from hibernation they regain the spines and appear to remember tasks learnt before hibernation, indicating that regardless of the persistence of spines, memories are retained (Clemens et al., 2009). In fact, if trained 24 h after arousal from hibernation, they remember better than controls (Weltzin et al., 2006). Likewise, female rats in the estrus phase of the cycle, when their spine density is down by 30%, are not less capable of remembering items learnt previously. In cultured neurons, pruning or lack of formation of dendritic spines is actually associated with enhancement of the frequency and amplitude of spontaneous synaptic currents (Fishbein & Segal, 2007, Segal et al., 2003). It is therefore not likely that these neurons lose their afferents once their spines disappear or are not formed. The reverse case has been documented in vivo; when the cell loses its afferents the relevant spines disappear, only to reappear when a new pathway innervates the vacated region on the dendritic shaft (Frotscher et al., 2000). Once again, this reported formation of new spines is not associated with an increase in filopodia extension, indicating that spines can form anew or extend from existing shaft synapses.
The need for ongoing activity in the maintenance of dendritic spines has also been demonstrated in cultured slices, where chronic blockade of AMPA receptors led to disappearance of spines, but this was apparently compensated for by the appearance of shaft synapses and by an increase in efficacy of synaptic transmission (Mateos et al., 2007), similar to our observations in dissociated cultures of cortical neurons (Fishbein & Segal, 2007).
There is no consistent relationship between spine formation and afferent activity. In some cases (e.g. cerebellum) the lack of afferent innervation does not deter formation of spines, which seem to develop naturally in a preprogrammed fashion (Cesa & Strata, 2005). On the other hand, we have shown that striatal neurons, about the spiniest cells in the brain, do not form dendritic spines if grown in culture in the absence of excitatory cortical afferents. Only the addition of such afferents enables the formation of dendritic spines in striatal neurons (Segal et al., 2003). Furthermore, blockade of electrical activity in these co-cultured striatal andd cortical neurons chronically exposed to TTX also prevents formation of spines, indicating that ongoing network activity is necessary for the formation and maintenance of dendritic spines in at least these striatal neurons (Segal et al., 2003). An interesting deviation from this tentative rule is the finding that long-term sensory deprivation prevents rather than enhances spine pruning (Zuo et al., 2005). The interpretation of this disparity is complicated by the fact that sensory deprivation produced four synapses away from the monitored neuron in the barrel cortex is not equivalent to a local continuous blockade of activity with TTX, especially as the extrinsic sensory afferents constitute only a fraction of the excitatory innervation of the cortical neuron. In another study, blockade of activity with TTX increased rather than decreased the number of functional synapses, as demonstrated by the density of presynaptic marker proteins, but these studies were conducted with 2- to 3-day-old hippocampal slices, at an age when most of the synapses are still located on the dendritic shaft, and so the developmental rules may be different at this age (Lauri et al., 2003).
Can spine morphology affect synaptic charge?
A pivotal issue, that has only recently begun to be addressed systematically, concerns the contribution of spine size and length to the charge produced at the synapse and recorded at the dendrite or soma. While theoretical assumption was that the spine was not a barrier to the transfer of the synaptic potential to the parent dendrite (Segev et al., 1995), experimental evidence for this issue is rather scarce, for the simple reason that such a comparison is difficult to obtain in view of the many different factors that contribute to the size of the synaptic current. However, tentative evidence suggests that a shaft synapse makes a larger synaptic current recorded at the soma than a spine synapse. In our experiments (Fishbein & Segal, 2007), exposure of cultured cortical neurons to TTX for a period of 7–10 days caused dendritic spine pruning although synapses on the dendritic shafts were retained (Fig. 1). In such cases miniature excitatory synaptic currents are nearly twice as large as those of controls. In a similar set of experiments (Segal et al., 2003), treatment of striatal–cortical cultures with TTX prevented the appearance of dendritic spines on striatal neurons, yet caused an almost two-fold increase in miniature excitatory postsynaptic current (mEPSC) amplitudes in these neurons compared to innervated control striatal–cortical cultures. Finally, transfection of cultured hippocampal neurons with constitutively active Rho GTPase caused elimination of spines and shrinkage of dendrites, yet synapses were still present on dendrites of these neurons and they produced larger mEPSCs than did controls (Pilpel & Segal, 2004). These experiments indicate that shaft synapses are likely to produce larger synaptic currents than spine synapses. In other series of experiments, we (Korkotian & Segal, 2007) and others (Araya et al., 2006) found that long spines produce smaller EPSCs evoked by local flash photolysis of caged glutamate than do short ones (Fig. 2). Similar studies also indicate that the spine neck may act as a barrier for the delivery of synaptic current from the synapse on the spine head to the parent dendrite (Ashby et al., 2006), which may explain the reduction in the synaptic current with distance from the spine head, and the observation that synapses on filopodia are less effective than spine synapses.
Spines and local calcium-handling machinery
A major impetus for the proposal that spines are the locus of synaptic plasticity originates in the early observations that spines constitute unique calcium compartments, able to raise [Ca2+]i levels locally to high concentrations that are not ‘seen’ in the parent dendrite and that such [Ca2+]i rises cannot be reached in an open-ended dendritic compartment. These high concentrations are probably needed for activation of calcium-dependent, plasticity-related kinases. In this context it is interesting to note that the spine does not contain a mitochondrion (except for the large multisynaptic dendritic excrescences apposing mossy fiber terminals in CA3 neurons) which may react to a rise in [Ca2+]i by triggering apoptotic processes. Obviously, the spine is a modifiable compartment whose neck length can be controlled by afferent activity and which can regulate the spread of the [Ca2+]i rise evoked at the synapse and perhaps prevent further spreading into the parent dendrite (Korkotian & Segal, 2007); it can also control the access of synaptic molecules into the sphere of the spine head. One category of molecule which is delivered into and out of the synapse in relation to activity is the ionotrophic AMPA-subtype glutamate receptor. LTP is assumed to involve the addition of glutamate receptors into the postsynaptic density, and LTD results from the removal of AMPA receptors from the spine head. Recently it has been suggested (Korkotian & Segal, 2007; Ashby et al., 2006) that the spine neck is a barrier to the diffusion of glutamate receptors into the synapse. Whether this barrier is determined by the calcium signal delivered to the dendrite or by the diffusion of receptor molecules is less critical; the outcome is that spine neck restricts access of glutamate receptors to the synapse. Consequently, synapses on the parent dendritic shaft should produce larger synaptic currents than those in the spine head, and the length of the spine will determine synapse efficacy.
Calcium stores and dendritic spines
In addition to the influx of calcium through NMDA-gated channels, voltage-gated calcium channels and GluR1-gated, GluR2-lacking channels, the spines are endowed with calcium stores of the ryanodine type, which are activated by influx of calcium or by direct activation of the ryanodine receptors (e.g. by caffeine). These stores have been linked recently to the spine apparatus, en enigmatic structure in the spine neck, via synaptopodin, a molecule found to be in close association with the spine apparatus (Vlachos et al., 2009). Synaptopodin and the spine apparatus have been found primarily in large, mature spines. Thus, it is likely that synaptopodin regulates the levels of ambient [Ca2+]i, which is raised transiently by influx of calcium ions. It is likely that large spines, where a larger influx of calcium is expected, need the stores in order to regulate excess amount of [Ca2+]i. Whether synaptopodin contributes to the stability of the spine is not entirely clear, as time-lapse imaging of synaptopodin and spines show that neither entity is stable over time (Vlachos et al., 2009).
Is there a need to protect the neuron from excessive afferent activity?
Regardless of their plastic properties, spines have been shown to constitute an independent physical compartment in which [Ca2+]i can rise to high levels, independent of the parent dendrite, suggesting that the spine protects the parent neuron from uncontrolled rises in [Ca2+]i, which may otherwise activate apoptotic processes leading to cell death (Schonfeld-Dado et al., 2009).
In spiny neurons, shaft synapses are more likely to be harmful to the parent neuron than spine synapses. This was tested in a recent series of studies with cultured neurons (Fishbein & Segal, 2007): Initially we found that, prior to the massive cell death of cultured cortical neurons which is caused by long-term (1–2 weeks) exposure to TTX, the cells lose functional dendritic spines. While their spines are pruned, some of their spine synapses are transformed into being shaft synapses on the parent dendrite and synaptic currents in these cells are now twice as large as controls (Fig. 1). These large mEPSCs probably contribute to the cell death, as treatment of the cultures with the AMPA receptor antagonist DNQX rescues the TTX-treated neurons from eventual death (Fishbein & Segal, 2007).
In a second test system, we transfected hippocampal neurons in culture with a constitutively active Rho GTPase (Pilpel & Segal, 2004). These neurons, which are grown together with normal untransfected neurons, lose their dendritic spines and their dendritic morphology is grossly simplified, but they maintain synaptic connectivity with neighboring neurons as indicated by the recording of mEPSCs (Pilpel & Segal, 2004). Exposing these neurons to a conditioning medium which enhances their network activity caused selective death of the Rho-overexpressing, spine-less neurons while not affecting control GFP-transfected neurons (Fishbein and Segal, unpublished observations).
Other studies provide correlative information on the relations between spine density and survival of neurons following an acute insult. Exposure of cultured slices to GABA receptor blockade produces a rapid reduction in dendritic spine density and subsequently a massive cell death (Thompson et al., 1996). Estradiol, shown to increase dendritic spine density in CA1 neurons in vivo, also protects these neurons from degeneration following acute ischemia (Sandstorm & Rowan, 2007).
It is important to emphasize the difficulty of producing direct evidence for a neuroprotective role of dendritic spines, as treatment aimed at eliminating spines is likely to affect other processes as well, including a change in glutamate receptor density in the spine head and detachment of the presynaptic partner from existing spines. Nevertheless, these experiments, conducted with different types of neurons in culture or in vivo, indicate that once they lose their spines, naturally spiny neurons produce larger mEPSCs than control cells when their synapses relocate to the dendritic shaft. The neurons are then more vulnerable to otherwise subtoxic insults, leading to their eventual death under conditions that do not harm normal spiny neurons. This process is counterintuitive, as it would be expected that the affected neurons would activate homeostatic mechanisms (Turrigiano, 2007) that would counteract the tendency to increase synaptic currents in conditions of eliminated dendritic spines, but apparently these mechanisms do not operate in such extreme conditions, leading to cell death, as is also the case with exposure to epileptic seizures (Thompson et al., 1996).
The neuroprotective role of spines can then explain why spine density increases with enhanced synaptic activity, including LTP generation, and shrinks or is replaced by filopodia when synaptic activity is low, as is the case with the deafferented neuron. The different spine sizes differ in their responses to afferent stimulation, indicated by a response to flash photolysis of caged glutamate (Fig. 2; modified from Korkotian & Segal, 2007). Massive stimulation, such as epileptic seizure, leads to extensive shrinkage of the spines and the eventual death of the parent neuron (Thompson et al., 1996). On the other hand, an LTD protocol, resulting in a reduction in strength of synaptic connectivity, is associated with retraction, shrinkage and disappearance of spines as is the case of entry into hibernation. These mechanisms are congruent with the basic assumption that spines protect the parent neurons from potentially hazardous afferent stimulation.
While there is a rapid accumulation of molecules that crowd the spine head, there are still some emerging issues that need to be addressed on the way to a more complete understanding of the roles of dendritic spines in neuronal plasticity and cell survival. One issue involves the great chemical heterogeneity of spines. Most recent studies tend to ignore the likelihood that spines vary in shape, but most likely they contain different subsets of molecules. For example, we (Vlachos et al., 2009) found that < 50% of the spines contain synaptopodin. How would this and similar variations affect the functioning of the spines? Likewise, generalizations are currently made rather carelessly, and there is a tendency to ignore the fact that spines may behave differently in dissociated neurons, in cultured slices and in vivo, and to different degrees in different brain areas. Also, treatments of populations of neurons may produce different changes in the spines of the affected neurons than treatments that are aimed at producing a change in a selected spine of the same neuron. It is not obvious that a certain behavior, monitored in one preparation, is indeed universal. These and similar issues need to be addressed in future experiments before a complete chemical and morphological vocabulary of spine behaviors is developed, but this goal is within reach.
I would like to thank Drs Eduard Korkotian and Ianai Fishbein for their contribution to the work cited in this review. Supported by grant #805/09 from the Israel Science Foundation.