|Home | About | Journals | Submit | Contact Us | Français|
Structural changes in brain circuits active during learning are thought to be important for long-term memory storage. If these changes support long-term information storage they might be expected to be present at distant timepoints after learning, as well as specific to the circuit activated with learning, and sensitive to the contingencies of the behavioral paradigm. Here, we show such changes in the hippocampus as a result of contextual fear conditioning. There were significantly fewer spines specifically on active neurons of fear-conditioned mice. This spine loss did not occur in homecage mice or in mice exposed to the training context alone. Mice exposed to unpaired shocks showed a generalized reduction in spines. These learning-related changes in spine density could reflect a direct mechanism of encoding or alternately could reflect a compensatory adaptation to previously described enhancement in transmission due to glutamate receptor insertion.
Dendritic spines are the primary sites for excitatory synaptic contact on pyramidal neurons in the brain. The morphology and density of dendritic spines are both altered in neuropsychiatric conditions that affect cognitive function (Penzes et al., 2011). The spine abnormalities found in these illnesses may underlie cognitive impairment by interfering with the intricate structural plasticity that has been observed for dendritic spines. For instance, the size and density of spines have been found to change in a number of synaptic and behavioral plasticity paradigms, leading to the suggestion that they may form a structural basis for long-term memory (Moser et al., 1994; Leuner et al., 2003; LeVay et al., 1980; Izquierdo et al., 2009; Kelsch et al., 2009; Restivo et al., 2009 ; Lai et al, 2012}.
Spine changes that result from long-term potentiation (LTP) and long-term depression (LTD) suggest an important role for neural activity in synaptic modifications that may be important for long-term memory. LTP leads to a proliferation of spines while LTD is conversely associated with spine elimination (Trommald et al., 1996; Desmond and Levy, 1986; Bastrikova et al., 2008). Whether spine changes exist on the active circuits that are proposed to subserve long-term memory, however, is unclear (De Roo et al., 2008). Spine proliferation as a result of LTP tends to be transient, reverting to baseline values shortly after LTP induction (De Roo et al., 2008; Wosiski-Kuhn and Stranahan, 2012). This observation challenges the proposition that there are durable and activity-specific changes in spine number that are characteristic of long-term memory.
In order to study whether those neurons active at learning exhibit an enduring spine change with long-term memory, we used a transgenic mouse, GFP-GluR1cfos. This mouse expresses a long-lasting dendritic marker (GFP-GluR1) of neural activity, allowing us to assess structural changes in neurons activated at the time of learning relative to inactive neurons in the CA1 region of the hippocampus. We found that spines on active neurons of fear-conditioned animals, but not homecage animals or animals exposed to context alone or unpaired shock, were significantly reduced 24-hours following training relative to spines on inactive neurons. These data point to a mechanism of enduring activity-related synaptic refinement as a component of long-term memory.
GFP-GluR1cfos mice were generated and genotyped as previously described (Matsuo et al., 2008). All colony maintenance and procedures were conducted acoording to TSRI guidelines for the humane care and use of laboratory animals. Animals were bred on a 12h/12h light-dark cycle and were provided with food and water ad libitum. Animals carrying both the cfos-tTA and GFP-GluR1 transgenes were used for experiments. Mice were kept off of doxycycline (dox) during postnatal development and placed on dox (40mg/kg) at weaning. Prior data from our laboratory have shown that GFP-GluR1 is degraded within two weeks of GFP-GluR1cfos being placed on dox allowing us to eliminate any expression from the early postnatal period with this protocol (Matsuo et al., 2008). Only males were used since estrogen is known to affect dendritic spine density (Murakami et al., 2006). Experimental mice were two to four months of age.
Four experimental groups of GFP-GluR1cfos mice were used; HC=homecage control (n=9 mice) CT= contextual control training (n=10 mice), FC= contextual fear-conditioning (FC, n=11 mice), and UP= unpaired shock training (n=8 mice). Mice were individually housed for two days and handled for 2 minutes, twice per day, during this time. Mice were then taken off of dox for four days prior to training. Studies in our laboratory have found that this period is sufficient to allow full clearance of dox.
FC involved placing mice in a wintergreen scented square chamber with black and white checkered pattern and aluminum walls (30-cm length × 24-cm width) and a grid floor that delivers foot shocks (FreezeFrame). A FC session consisted of 120s of free exploration followed by four non-signaled foot shocks (duration 2s, intensity 0.7mA) with an inter-stimulus interval of 60s. Mice remained in the chamber for 60s after the last shock. We performed two FC sessions that were separated by two hours. CT mice were subject to the same procedure without shock.
UP training was conducted as outlined previously (Matsuo et al., 2008). We first placed mice in a novel chamber (context A: square plastic chamber surrounded with white walls, plastic floor with sani-chips and wintergreen scent) for 360s without foot shocks. After context A exposure mice were placed in their home cage and twenty minutes later they underwent four non-signaled foot shocks (duration 2s, intensity 0.7mA) with an interstimulus interval of 4s. These shocks began immediately after mice were placed in the conditioning chamber (context B: rectangular chamber, plexiglass front and back with black and white checkered pattern and aluminum side walls, gridfloor) and mice were immediately returned to their home cages after the last shock. We performed two UP sessions that were separated by two hours. HC animals remained in their cage during these experimental procedures.
Immediately after the final experimental treatment, mice from the HC, CT, FC and UP groups were placed on a high dose dox diet (1g/kg) overnight to suppress further activation of the GFP-GluR1 transgene. Studies in our laboratory have confirmed that overnight feeding with high dose dox will block transactivation.
Twenty-four hours after fear conditioning mice used for dendritic spine analysis were anesthetized with isoflourane and rapidly perfused with cold 1X phosphate buffered saline (PBS) for 1.5 minutes. This was followed by 3.5 minutes of perfusion with cold 4% paraformaldehyde (PFA). Brains were then removed and incubated for one hour in cold 4% PFA and transferred to 30% sucrose where they were stored at 4°C overnight. The next day 100 μm sections were obtained on a vibratome and sections collected in cold 1XPBS. A separate experiment that assessed the induction of GFP-GluR1 to FC used fresh 100 μm sections that were fixed in cold 4% PFA in PBS for 1hr.
Immunohistochemistry was performed as described (Matsuo et al., 2008). Sections were incubated in 5% bovine serum albumin (BSA) in PBS at room temperature (RT) for 30 minutes. Rabbit-anti-GFP (1:2000) was then added and sections were allowed to incubate at 4°C overnight, washed 3×5min in PBS at RT and incubated in Alexa 488-goat-anti-rabbit secondary antibody (1:800) at 4°C overnight. After this incubation they are washed 3×5min in PBS at RT and stained with DiI. Experiments that assessed GFP-GluR1 induction to FC used immunohistochemistry as described above with the addition of 0.15% TritonX-100 for all incubations. This allowed us to visualize the cytoplasmic expression of GFP-GluR1 in order to quantify the number of CA1 neurons activated by FC.
DiI (1,1'-dioctadecyl-3,3,3'3'-tetramethylindocarbocyanine perchlorate) labeling was performed as described (Matsuo et al., 2008). Micropipettes were coated with DiI dissolved in ethanol at 10 mg/ml. The tip of DiI-coated micropipettes was then inserted at several positions in the hippocampal CA1 pyramidal cell layer of fixed slices. Slices were placed in RT PBS for 2h to allow the DiI to spread throughout the dendritic arbor. Sections were mounted with SlowFade mounting medium (Invitrogen) and stored at 4°C.
In order to verify the induction of GFP-GluR1 to FC, we used confocal microscopy to measure the number of CA1 neurons with cytoplasmic expression of GFP-GluR1. The number of cells with cytoplasmic GFP-GluR1 expression was quantified relative to the total population stained with TOPRO-3. The intensity of GFP staining was measured in the hippocampal CA1 dendritic layers of HC versus FC mice.
Spines were imaged with confocal microscopy within one week of mounting sections. Dendrites were randomly imaged using an oil immersion 60× objective with 2× zoom, and the experimenter was blind to experimental condition. Images were acquired using sequential imaging with an Argon (488 nm) and HeNe (543 nm) laser. In order to minimize quenching of fluorescence z-stacks of 5–8 μm thickness consisting of sections at 0.12 μm increments were rapidly scanned within a 107 × 107 μm imaging area. We coupled our sampling of segments from active and inactive neurons within each sampled region. This was to minimize any regional differences in dendrite structure as a source of variability in comparing active versus inactive circuits. A represented image of active segments that were coupled with the sampling of inactive segments within the same image is depicted (Figure 1C). In order for a dendrite to be sampled it had to have consistently bright and continuous DiI labeling throughout it’s course. It also had to be easily separated from neighboring dendrites to avoid obstruction of spine counts. Primary dendrites were not included in the analysis.
Images were analyzed with Neurolucida software (MBF Biosciences) while blinded to experimental condition. Dendrites that were strongly double-labeled with GFP immunoreactive signal and DiI fluorescence were designated as from active neurons. Dendrites that did not show GFP immunoreactivity were designated as from inactive neurons. In order to provide a blinded assessment of spine density by neural activity, GFP immunoreactivity was subtracted from all images prior to spine counting.
Spines were counted only when they were clearly connected to the shaft of the dendrite. The spine depleted zone arising from the soma was excluded in our analysis. In total 14,385 spines were counted across 437, 30 μm dendritc segments. These spines belonged to 80 inactive and 90 active segments from apical dendrites and 126 inactive and 141 active segments from basal dendrites. We counted 2015, 4131, 3998 and 4241 spines within the HC, CT, FC and UP groups, respectively. The HC group had relatively fewer spines counted due to the low expression of GFP-GluR1 in this group and consequentially fewer dendrites from active neurons.
We categorized as mushroom and thin spines only those spines with a head and neck morphology according to established methods for spine classification (Harris et al., 1992). Those spines without a readily observable head and neck morphology were categorized as stubby spines. Mushroom spines were then further differentiated into spines with a calculated head volume of 0.1 μm3 or greater, corresponding to a head diameter of ~ 0.6 μm or greater. Spines that had two heads connected to a single neck were classified as branched spines. This method is in accordance with established methods for spine type classification (Harris et al., 1992; Bourne and Harris, 2008).
The strength of GFP signal within the stratum radiatum and the stratum oriens of the hippocampal CA1 region was measured with Metamorph software and quantified as % increase in GFP signal above HC. The increases in GFP staining were analyzed in the hippocampal neurons with a two-tailed t test comparing HC to FC groups. The effects of our experimental protocols and neural activity on total spine density were analyzed with a Factorial ANOVA and a Bonferroni Post-Hoc test.
The effects of our experimental protocols and neural activity on spine morphology were analyzed with a Factorial ANOVA and Fisher Post-Hoc test After statistical analysis, data examining the relationship of spine changes to neural activity were converted to percentage of the inactive circuit spine density for each learning group.
To investigate the relationship of spine changes in learning to patterns of neural activity associated with the learning event we used the GFP-GluR1cfos transgenic mouse. The mouseexpresses a tetracycline response element linked GFP-GluR1 fusion protein in an activity-dependent and dox dependent manner through regulation by a second c-fos-tTA transgene. Discrete windows for mapping brain activity can be opened by keeping mice on dox for several weeks prior to experiments, and then removing dox from the diet prior to experimental manipulations (Figure 1A).
Subjecting GFP-GluR1cfos to contextual control training alone (CT), contextual fear conditioning (FC) and unpaired shock (UP) protocols results in strong GFP-GluR1 expression above home cage (HC) controls. HC control levels were low, as previously described, but FC resulted in a three-fold increase in the percentage of CA1 neurons expressing GFP-GluR1 (Figure 1B) (Matsuo et al., 2008). This was paralleled by an ~80% increase in GFP-GluR1 signal within the dendritic layers of CA1 (Figure 1B). Co-staining hippocampal sections with GFP and DiI allowed us to identify active neurons and compare dendrites from these and from inactive neurons. We could thereby relate the structure of dendrites to the pattern of neural activity during learning (Figure 1C). Mice kept on dox from weaning through maturity were grouped into HC, CT, FC or UP experimental protocols at maturity (Figure 1D), and removed from dox for four days prior to training. The HC experimental group remained within their homecage throughout the experiment. In a separate group of mice we confirmed that our FC protocol resulted in a strong contextual freezing and that our UP protocol did not (Figure 2A).
Twenty-four hours after training animals were sacrificed and brain sections processed for imaging of dendritic spines. Our experimental protocols significantly affected total spine density (active and inactive neurons combined) with total spine density lower in the experimental groups that received footshocks (F(3, 249) =7.1918, p=0.0001) (HC vs FC, p<0.05; HC vs UP, p<0.001; CT vs UP, p<0.01) (Figure 2B).
In order to determine the relationship of neural activity at the time of training to spine density changes, we dissociated spine changes on dendrites from active (GFP+) and inactive (GFP−) neurons. Neural activity was associated with lower spine density on active compared to inactive neurons (F(1,429)=10.292, p=0.00144). This is presented as percent spine density in active neurons relative of inactive neurons within each group (Figure 2C). The decrease in spine density was found specifically on active relative to inactive neurons of FC animals (**p<0.01). This change was not seen in HC mice or mice exposed to CT or UP (Figure 2C, Figure 2D, Table 1).
Recent findings show that synaptic inputs cluster on active neurons as mechanism of circuit remodeling associated with learning (Takahashi et al., 2012). We therefore investigated whether spine clustering was seen in our experiment and whether this was selective for active neurons. If this occurred on the active cells of our fear conditioned group, we hypothesized that clustered inputs would be identified as a discrete population of dendrite segments enriched in spines, occurring amidst an overall increase in pruned segments.
We did not see clustering but instead found that the FC resulted in a generalized increase in pruned segments in the active circuit without any evidence of a clustered spine population. As opposed to the FC group, the HC, CT and UP groups had similar and overlapping spine density distributions for both active and inactive cells (Figure 3). These data do not suggest that spine changes are anatomically restricted to particular dendritic regions, at least at the 30 μm level of resolution.
We further classified spines into mushroom, branched, stubby, and thin morphologies, since distinctive roles for each of these subtypes are implicated in various plasticity paradigms (Figure 4A). For example, studies have implicated mushroom spines in memory (Bourne and Harris, 2007; Matsuo et al., 2008) and changes in branched spines have been identified in LTP (Desmond and Levy, 1986; Trommald et al., 1996). To assess whether spine elimination is restricted to a specific spine morphology, we analyzed the changes occurring in each morphological type (Figure 4A).
Our experimental protocols significantly affected mushroom spines (Figure 4B, Table 2, F(3,429)=9.9801, p<0.001). CT significantly increased mushroom spines relative to their levels in in the HC group (p<0.05). Relative to the CT group, though, mushroom spines were fewer in the groups that received footshock (FC vs CT, p<0.05; UP vs CT, p<0.001). Between the groups that received a footshock, mushroom spines were lower in the UP compared to the FC group (p<0.01) (Figure 4B).
Our protocols also led to changes in mushroom spines when analyzed by their relative density on active and inactive neurons. There were fewer mushroom spines on active compared to inactive neurons of the FC group (p<0.05), but this activity-related change was not seen in HC mice or mice exposed to UP or CT training (Figure 4B, Table 2). Thus, while both groups that received a foot-shock showed reductions in mushroom-type spines compared to CT, in the UP group this was a generalized decrease across neurons while in the FC group the change was specific to the circuit active during learning. This activity-specific decease in mushroom spines accounted for ~27% of the total spine decrease on active neurons with FC, and was hence representative of activity-specific elimination within the total spine population, but did not account for all of the spine loss. We further examined other spine morphologies for their contribution to activity-specific decreases with learning.
Our protocols significantly affected branched spines (F(3,429)=2.6562, p=0.04803), with more branched spines on the neurons of the UP group compared to mice subject to CT (p<0.05) and FC (p<0.05). Neural activity was associated with lower branched spine density on active compared to inactive neurons (F(1,429)=4.9105. p=0.02722.) The decrease in branched spine density was specifically found on active but not on inactive neurons of FC and UP animals (p<0.001 and p<0.05, respectively), and it was not seen in HC mice or mice exposed to CT (Figure 4B, Table 2).
Experimental protocols led to small but significant changes in stubby and thin spines with fewer stubby spines on neurons of mice in the UP group compared to the HC (p<0.05) and CT groups (p<0.05). Within experimental groups, we did not find activity-specific changes for stubby spine morphology (Figure 4C, Table 2). FC resulted in a decrease in the thin spines relative to HC (p<0.05). There were fewer thin spines on active compared to inactive neurons of the FC group (p<0.05), but this activity-related change was not seen in HC mice or mice exposed to UP or CT training (Figure 4C, Table 2).
Structural changes in dendrites have been proposed as an important mechanism of long-term information storage. (Bailey and Kandel, 1993; Trommald et al., 1996; Maletic-Savatic et al., 1999; Yang et al., 2008). The subtle and transient nature of many of these structural changes, however, has called into question whether they accompany long-term memory within the sparsely distributed circuits that are active during learning (Marrone, 2007; De Roo et al., 2008). To study whether long-term memory is associated with spine changes on the neurons active during learning, we used a transgenic mouse that expresses an enduring dendritic marker (GFP-GluR1) under the control of the c-fos promoter.
Our data show that the stress of foot-shocks, as delivered during FC or UP, resulted in hippocampal CA1 spine loss, consistent with a documented role of stress in decreasing spine density (Michelsen et al., 2007; Chen et al., 2008; Chen et al., 2010; Bloss et al., 2011; Mucha et al., 2011; Chen et al., 2012). However, an important distinction is observed when spine density is analyzed in the active and inactive neural ensembles. With unsignaled foot-shocks in the UP group, the spine loss was generalized across the active and inactive neurons. In the FC group the foot shock is paired with a specific context to develop an associative fear memory, and spine loss was restricted to the neural population that was active at the time of learning.
These findings corroborate the previous data of others that have shown a role for learning in modifying the effects of stress on hippocampal systems (Shors et al., 1989). For instance, LTP is impaired when animals are exposed to the stress of an inescapable footshock. This LTP impairment, however, is much less when this stress is accompanied by escape learning (Shors et al., 1989). The neural processes that subserve these protective effects of learning are poorly understood. Our data suggest that they may be due to a restricted, circuit-specific loss of spines when stress occurs with learning. This is in contrast to a more profound and generalized spine loss that may cause hippocampal impairment when this brain structure experiences an unpredictable and non-associative stressor.
At the level of spine morphology, one might expect this effect to be most apparent within thin spines since this morphology is considered to be susceptible to pruning with stress-related signaling (Chen et al., 2012). The overall modest response of thin spine morphology to the stressful exposure of footshock, though, is in contrast to what may be hypothesized from these prior studies. These differences may be due to in vitro methods that applied corticotropin-releasing hormone (CRH) to hippocampal slices and found a decrease in thin spines (Chen et al., 2012), versus our in vivo experimental paradigm. Alternatively, these differences may be due to different timepoints of analysis or varying methods for classifying thin spine morphologies.
In contrast to the modest alterations that we found in thin and stubby spines, there were more changes in branched and mushroom spines. Branched spines have been found to change as a result of LTP and environmental enrichment, leading to the proposition that they are an important structural substrate for synaptic plasticity. (Desmond and Levy, 1986; Geinisman et al., 1989; Trommald et al., 1996; Johansson and Belichenko, 2002). Our data further suggest a unique role for the branched spine morphology in a structural remodeling that accompanies memory, particularly within active hippocampal circuitry.
Although CT training increased the total population of mushroom spines relative to HC levels, perhaps as an adaptation to spatial learning (Moser et al., 1994), FC and UP were conversely found to result in lower mushroom spines compared to CT. Further analysis of mushroom spines by neural activity found that FC eliminated mushroom spines on active neurons.
The role of such activity-specific spine loss with learning is not known, however it may serve to promote synaptic transmission efficacy during future exposure to learned contingencies. Previous studies have shown a circuit-specific insertion of newly synthesized AMPA-type glutamate receptors into spines using a similar fear-conditioning paradigm (Matsuo et al., 2008) (Figure 5). This AMPA receptor insertion is also induced by artificial stimulation leading to LTP, suggesting that this synaptic strengthening mechanism occurs at concurrently active synapses (Kessels and Malinow, 2009). Conversely, LTD has been associated with spine loss and spike-timing dependent plasticity (STDP) protocols show that LTD occurs when pre- and post-synaptic activity is discordant (Dan and Poo, 2004; Bastrikova et al., 2008).
Together, this would lead to a strengthening of synapses that are strongly activated by the behavioral contingencies (via AMPA receptor insertion) and a weakening of those synapses that are not driven by the stimuli (reflected in synapse loss) due to discordant pre- and post-synaptic activation. These two mechanisms acting together would sculpt the circuit responsiveness to favor activation by the learned contingencies while maintaining a constant level of overall synaptic drive. In favor of this idea, recent findings propose an “inverse” synaptic tagging mechanism whereby weak synapses are silenced within strongly activated circuits to prevent their undesired enhancement (Okuno et al., 2012). Our findings may reflect the structural equivalent of “inverse” synaptic tagging.
Circuit-specific spine elimination as a result of FC points to an important role for the hippocampus in associative processes underlying this form of learning. These observations indicate that a role for the hippocampus in FC extends beyond an encoder of contextual representations that is ancillary to associative processes localized within the amygdala (Anagnostaras et al., 2001). Our data as well as the data of others (Moita et al., 2004), suggests that the hippocampus plays an important role in integrating context representation with shock.
In summary, the subtle and transient structural changes found in synaptic and behavioral plasticity paradigms have challenged whether or not there are enduring spine modifications that underlie memory (Marrone, 2007; De Roo et al., 2008). Our data, point an important locus for such modifications within hippocampal ensembles that our active during learning. These findings are compatible with recent findings showing structural refinement within hippocampal circuitry that is active during learning (Kitanishi et al., 2009a; Kitanishi et al., 2009b). For example, mice that explore a novel enriched environment exhibit fewer spines on activated hippocampal neurons one hour later (Kitanishi et al., 2009b). Our data show a similar circuit-specific spine reduction that persists as an enduring structural signature to strong emotional learning. This spine reduction may reflect a stress-induced spine elimination that is focused into circuit-specific structural modifications in order to refine synaptic connectivity. Collectively, these data imply that dendritic spine perturbations in cognitive disorders impact an intricate system of activity-specific structural reorganization that may be important for long-term memory.
This work was supported by The National Institutes of Health DA028300, MH057368, MH019334.
We have no conflicts of interest to report