Report

Locally Synchronized Synaptic Inputs

See allHide authors and affiliations

Science  20 Jan 2012:
Vol. 335, Issue 6066, pp. 353-356
DOI: 10.1126/science.1210362

Abstract

Synaptic inputs on dendrites are nonlinearly converted to action potential outputs, yet the spatiotemporal patterns of dendritic activation remain to be elucidated at single-synapse resolution. In rodents, we optically imaged synaptic activities from hundreds of dendritic spines in hippocampal and neocortical pyramidal neurons ex vivo and in vivo. Adjacent spines were frequently synchronized in spontaneously active networks, thereby forming dendritic foci that received locally convergent inputs from presynaptic cell assemblies. This precise subcellular geometry manifested itself during N-methyl-d-aspartate receptor–dependent circuit remodeling. Thus, clustered synaptic plasticity is innately programmed to compartmentalize correlated inputs along dendrites and may reify nonlinear synaptic integration.

Cortical microcircuits are nonrandomly intertwined and form cell assemblies that fire in a spatiotemporally orchestrated manner. This patterned activity is decoded by the dendrites of downstream neurons. Dendrites are arborized and electrically active (1), which allows them to exhibit local nonlinear membrane potential dynamics (24) and to transform different spatiotemporal sequences of incoming inputs into different output patterns (5, 6). Therefore, knowing whether synaptic inputs are clustered or dispersed over dendrites at a given time (fig. S1) is critical for determining the dendritic computational power (7, 8); however, these dynamics are still poorly understood.

We monitored spontaneous synaptic inputs using dual patch-clamp recordings under confocal visualization from different apical dendritic branches of individual CA3 pyramidal cells in rat hippocampal slices that were cultured for 12 to 19 days (fig. S2A) (9). Large postsynaptic potentials (that is, putative synchronized inputs) often occurred in only one branch (fig. S2B). The Euclidean distance in membrane potentials between two branches was distributed with a long tail (fig. S2C), suggesting that dendrites received spatially biased synchronous inputs.

Next, we intracellularly injected CA3 pyramidal cells with Fluo-5F through somatic patch-clamp pipettes (Fig. 1A) (10). We voltage-clamped a neuron at –30 mV and imaged the dendrites in an area of approximately 100 × 100 μm2 that contained an average of 98.5 ± 16.7 spines (mean ± SD, ranging from 52 to 235 spines) (Fig. 1B). Transient calcium elevations occurred spontaneously in the spines (Fig. 1, C and D, and movie S1). These activities were spatially restricted within the spines (Fig. 1C) and were time-locked to the occurrence of spontaneous excitatory postsynaptic currents (EPSCs) in the patched neurons (fig. S3A). The local application of 1 mM l,d-2-amino-5-phosphonopentanoic acid (AP5), an N-methyl-d-aspartate (NMDA) receptor antagonist, abolished spine calcium events without affecting the overall frequency of the spontaneous EPSCs (fig. S3, B to G). The mean frequency of spine activities was 1.5 events/min, which was approximately half the mean firing rate of the presynaptic neuron population (fig. S4). This difference may result from stochastic synaptic transmission, with a probability of about 50% between CA3 pyramidal cells ex vivo (11), as well as from false-negative detection of a fraction of spine calcium events.

Fig. 1

Imaging of spontaneous synaptic inputs. (A) A stack image of a CA3 pyramidal cell filled with Fluo-5F. (B) The locations of 137 spines (red) monitored from the boxed region shown in (A). (C) Spontaneous calcium events of spines in the region boxed in (B). ΔF/F, percent increase in fluorescence intensity. (D) The spatiotemporal pattern of the calcium events of the 137 spines. Each dot indicates a single event of a single spine. (E) The event frequencies are plotted versus the path distance from the soma. Each black dot represents a single spine (n = 1084 spines in 11 videos from 11 slices). Red dots and bars indicate the means ± SDs. (F) A Lorenz curve representing the proportion of the total inputs that is assumed by the proportion of spines with the lowest input frequencies. (G) Distributions of the input frequencies and the head sizes of the spines.

The locations of spines were three-dimensionally determined post hoc to measure the path length from the soma along the dendrites. The spines did not differ with respect to activity levels between the basal and radial oblique dendrites or between the proximal and distal dendrites (Fig. 1E). Therefore, all data were pooled in the following analyses. The Gini coefficient of the spine activity was 0.78; approximately 20% of the spines exhibited 80% of the calcium activity (Fig. 1F). The activity frequency and the spine head size, each of which approximated a log-normal distribution, correlated only weakly with each other (Fig. 1G, Spearman's rank r = 0.15, P = 5.9 × 10−7).

We calculated the spatial correlations of spontaneous spine activities; for a given activity in a “focused” spine, the probability of observing activity in other spines in a time window of 100 ms was plotted as a function of the path length from the focused spine (Fig. 2A, top; n = 11 videos from 11 neurons in 11 slices). The probability was compensated with the spine density observed at a given distance to avoid sampling bias caused by the limited lengths of monitored dendritic segments. Spine coactivation was significantly more frequent within interspine intervals of 8 μm as compared to the chance level, which is defined here as the mean probability of observing spine activity at distances greater than 10 μm (|Z| ≥ 5.77, P ≤ 8.0 × 10−9, Z-test for a population mean). The spatially clustered spine activation was also significant as compared to randomized surrogates (fig. S5A, P < 0.01). We did not observe clustered synaptic inputs in fast-spiking parvalbumin-positive interneurons in the CA3 stratum pyramidale (fig. S6, |Z| ≥ 0.63, P ≤ 0.53; fig. S5B; n = 11 videos from eight neurons).

Fig. 2

Functionally clustered spine activity ex vivo. (A) (Top) The probability of observing spines coactivated within 100 ms as a function of the distance from a given spine (n = 11 videos from 11 neurons). The chance level and its 95% confidence intervals (purple) were estimated from the distribution of distances of more than 10 μm. (Bottom) Spine coactivation in response to electrical stimulation of the CA3 stratum radiatum (n = 16 videos from 10 neurons). (B) Assemblets (thick dots) in a raster plot of 235 spines. Assemblets that appear in hot zones 1, 2, and 3 in (E) are colored in yellow, red, and blue; otherwise, they are colored in green. (C and D) Distribution of the number of spines participating in single assemblets (C) and the repetition numbers of each assemblet (D). Chance (purple) was estimated by the random shuffling of inter-event intervals and is represented by the mean value of 1000 surrogates and their 95% confidence intervals. The data from 11 neurons are pooled. (E) A heat map of the frequency of assemblets in dendrites. (F) Representative assemblet dynamics in hot zones 1, 2, and 3.

To examine the synaptic activation patterns in vivo, we conducted somatic whole-cell patch-clamp recordings (12) and two-photon calcium imaging from spines of layer 2/3 pyramidal cells in the barrel cortex of anesthetized young adult mice. Calcium activities were simultaneously monitored from 31.6 ± 13.7 spines (ranging from 16 to 48 spines) from 10 apical or basal dendrites of four cells (Fig. 3, A and B). Spontaneous activities frequently occurred in neighboring spines; the probability of observing the coactive spines significantly increased within 6 μm along the dendrites (Fig. 3C, |Z| ≥ 5.61, P ≤ 2.1 × 10−8; fig. S5C).

Fig. 3

Functionally clustered spine activity in vivo. (A) A stack image of dendrites of an Alexa 594–loaded layer 2/3 pyramidal cell in the mouse somatosensory cortex in vivo. (B) Typical traces of spontaneous calcium activity from eight spines shown in (A). (C) The probability of observing coactivated spines as a function of the inter-spine path distance (n = 10 dendritic segments in four cells). The chance level and its 95% confidence intervals (purple) were estimated from the distribution of distances of more than 10 μm.

The synchronization of adjacent spines can be explained by five possible mechanisms: (i) convergent afferents from a population of spontaneously synchronized presynaptic neurons (cell assembly) (fig. S7A); (ii) multiple innervations of a single presynaptic axon (fig. S7B); (iii) spillover of diffusible molecules (such as glutamate) to neighboring synapses (fig. S7C);(iv) spatial segregation of spine excitation by local dendritic inhibition (fig. S7D); or (v) local depolarization-induced increase in a chance of calcium influx in neighboring spines. We can rule out mechanisms ii to v on the basis of the results of the three following experiments ex vivo. First, we applied electrical field stimulation to the CA3 stratum radiatum and produced synchronized network activity. The stimulation intensity was set to evoke a compound EPSC with an amplitude greater than 400 pA in the patched neurons. This cell assembly–irrelevant artificial synchronization did not generate spatially clustered spine activation (Fig. 2A, bottom, |Z| ≥ 0.45, P ≤ 0.65; fig. S5D; n = 16 videos from 10 neurons). This result is inconsistent with ii, iii, and v. Second, we can also rule out mechanism ii because biocytin reconstructions of synaptically connected neurons revealed that 51 of 55 (92.7%) putative synapses arising from 12 presynaptic neurons contacted single spines (fig. S8). Finally, we can exclude mechanism iv because spine activation remained clustered in dendrites that were disinhibited by the local application of 1 mM picrotoxin (fig. S9, |Z| ≥ 3.97, P ≤ 7.2 × 10−5; fig. S5E; n = 8 videos from eight neurons). Therefore, mechanism i appeared to be the most plausible mechanism (fig. S1A).

To intuitively depict the observed spine coactivation, we defined an “assemblet” as a cluster of synchronized spine activities in which the distance from any spine in the cluster to the next nearest spine in the cluster was less than 10 μm (Fig. 2B). In the entire sample of 11 videos ex vivo, 31.5% of the spines participated in at least one assemblet, and assemblet activity accounted for 29.5% of the total spine activity. In single assemblet events, an average of 3.6 ± 0.7 spines, ranging from 2 to 12 spines, were activated during a period of 59 ± 33 ms and within an area of 4.7 ± 3.3 μm (Fig. 2C). Twenty-eight percent of the assemblets appeared more than once, and a portion of them repeated up to 30 times (Fig. 2D), whereas 58.9% of the spines that participated in one assemblet participated in other assemblets. The assemblet sizes and the numbers of repetitions were greater than the chance values expected by event-interval shuffling (Fig. 2, C and D, purple; P < 0.001), in which the time intervals between successive calcium events were randomly exchanged within each spine to collapse the time correlations between spines (13).

In regard to the population dynamics, 85.0% of the assemblets occurred sporadically, whereas the remaining 15.0% occurred in synchrony with other assemblets (fig. S10A). When synchronized, the assemblets tended to appear more than 80 μm apart from one other (fig. S10B).

Dendrites were spatially heterogeneous in emitting assemblets (Fig. 2, E and F). Therefore, we defined a hot zone as a continuous dendritic segment where a single assemblet or multiple assemblets that shared at least one spine occurred. The hot zones had an average area of 7.7 ± 6.7 μm, ranging from 0.4 to 28.2 μm (fig. S11A), and exhibited an average of 2.7 ± 4.8 assemblets/min, ranging from 0.04 to 25.0 assemblets/min (fig. S11B). Hot zones were dispersed at a density of 1.8 ± 1.2, ranging from 0 to 6, per 50 μm of dendritic length (fig. S11C). They are candidate sites for the initiation of dendritic spikes (14), but we rarely observed calcium sparks in the dendritic shafts, which may possibly be due to a lowered spine density in slice cultures (fig. S13B) as compared to cortical neurons in vivo.

Spines that participated in assemblets were larger in head size than nonparticipants (fig. S12), which suggests that assemblets are shaped by long-term synaptic plasticity (15). Indeed, adjacent spines were less synchronized in slices cultivated in the presence of 100 μM AP5 for 12 to 19 days (Fig. 4A, |Z| ≥ 1.05, P ≤ 0.30; fig. S5F; n = 8 videos from 8 neurons) and in immature slices cultivated for 3 to 4 days (Fig. 4B, |Z| ≥ 1.51, P ≤ 0.13; fig. S5G; n = 14 videos from 14 neurons) as compared to control mature slices (P < 0.01 each; Kolmogorov-Smirnov test). Neither the spine density nor the spontaneous or miniature EPSC levels differed between control and AP5-treated cultures (fig. S13). Thus, functional synaptic clustering is likely to develop through NMDA receptor–dependent circuit remodeling.

Fig. 4

NMDA receptor–dependent emergence of clustered synaptic inputs. (A) The probability of observing coactivated spines as a function of the path distance in slices cultivated in the chronic presence of 100 μM AP5 (n = 8 videos from eight neurons). (B) The probability of observing coactivated spines in immature networks of slices for 3 to 4 days in vitro (n = 14 videos from 14 neurons). Early calcium sparks in immature dendritic shafts were excluded from data analysis. The control is the same as in Fig. 2A.

Consistent with this idea, we found that glutamate receptors were preferentially inserted into neighboring spines after behavioral exploration in vivo, using adult transgenic mice in which GFP-GluR1 is expressed under control of the c-fos promoter (figs. S5H and S14) (16). Thus, the loci of synaptic plasticity are spatially clustered over dendrites. The clustered plasticity may result from interspine interactions that heterosynaptically modulate the threshold for long-term potentiation, such as local depolarization-induced Mg2+ unblock of nearby NMDA receptors or intracellular diffusion of plasticity-associated molecules (1720).

We found that synaptic inputs were frequently synchronized within a group of spines in the immediate vicinity of one another. Given that ex vivo networks are subject to massive axon reorganization during cultivation without external inputs, our data indicate that the locally convergent connectivity that generates assemblets emerges through self-organization (fig. S15). Thus, the default principle for designing circuit topology is biased to facilitate dendritic compartmentalization (21). The resultant clustered synchrony may offer opportunities for associative learning, because vicinal spines encode different information (22, 23).

Because the video frame rate of our spine imaging was limited to a maximum of 20 Hz to maintain the signal-to-noise ratio, we could not determine the internal structure of assemblets; however, given that the ex vivo hippocampal network synchrony accompanies sharp waves and ripples (24), assemblets are expected to coordinate temporal activity sequences (25). Such sequential activation would facilitate nonlinear synaptic integration and enhance the computational power of a single neuron (5).

Note added in proof: Two recent studies have reported phenomena partly related to those described here, demonstrating activity-dependent clustering of synaptic inputs to developing dendrites of hippocampal slice cultures (26) and clustered synaptic plasticity in the developing somatosensory cortex (27).

Supporting Online Material

www.sciencemag.org/cgi/content/full/335/6066/353/DC1

Materials and Methods

Figs. S1 to S15

References

Movie S1

References and Notes

  1. Materials and methods are available as supporting material on Science Online.
  2. Acknowledgments: The authors are grateful to K. Morita (University of Tokyo) for his comments on the manuscript. This work was partly supported by Grants-in-Aid for Science Research from the Ministry of Education, Culture, Sports, Science and Technology of Japan (nos. 21220006, 22115003, 23115504, and 23800019); the Strategic Research Program for Brain Sciences (development of biomarker candidates for social behavior); and by the Funding Program for Next Generation World-Leading Researchers (grant LS023).
View Abstract

Navigate This Article