Synaptic Clustering and Memory Formation

Abstract

In the study of memory engrams, synaptic memory allocation is a newly emerged theme that focuses on how specific synapses are engaged in the storage of a given memory. Cumulating evidence from imaging and molecular experiments indicates that the recruitment of synapses that participate in the encoding and expression of memory is neither random nor uniform. A hallmark observation is the emergence of groups of synapses that share similar response properties and/or similar input properties and are located within a stretch of a dendritic branch. This grouping of synapses has been termed "synapse clustering" and has been shown to emerge in many different memory-related paradigms, as well as in in vitro studies. The clustering of synapses may emerge from synapses receiving similar input, or via many processes which allow for cross-talk between nearby synapses within a dendritic branch, leading to cooperative plasticity. Clustered synapses can act in concert to maximally exploit the nonlinear integration potential of the dendritic branches in which they reside. Their main contribution is to facilitate the induction of dendritic spikes and dendritic plateau potentials, which provide advanced computational and memory-related capabilities to dendrites and single neurons. This review focuses on recent evidence which investigates the role of synapse clustering in dendritic integration, sensory perception, learning, and memory as well as brain dysfunction. We also discuss recent theoretical work which explores the computational advantages provided by synapse clustering, leading to novel and revised theories of memory. As an eminent phenomenon during memory allocation, synapse clustering both shapes memory engrams and is also shaped by the parallel plasticity mechanisms upon which it relies.

Keywords: dendrites; engram; in-branch localization; memory; synaptic clustering.

Figure 1

Synaptic and cellular engram formation. (A) Dendritic spikes can be induced either from localized activation of a co-active group of synapses (clustered synapse allocation) or from more dispersed activation of synapses within the same branch segment (In-branch allocation). (B) In both cases, the elicitation of dendritic spikes is integrated with the somatic compartment, controlling the action potential generation or bursting behavior of the neuron and enabling local or global plasticity to occur. (C) Neuronal populations activated for each memory are selected by mechanisms such as excitability and CREB activation leading to the selection of a population memory engram. Mechanisms for the generation of synaptic clusters: (D) cooperative sharing of plasticity-related resources such as proteins facilitates cooperative LTP in nearby synapses after LTP induction. The spreading of activation of plasticity-related proteins, enzymes, and mRNAs may prime nearby synapses for subsequent plasticity. (E) Co-active axons in nearby synapses can drive cooperative plasticity by initiating similar resource-sharing mechanisms as a result of coincident activation (F) a single axon may make multiple contacts in a short segment of a dendrite, thereby driving cooperative plasticity via synchronized synapse activation.

Figure 2

Computational basis of 2-layer neuronal integration. (A) Nonlinear integration in CA1 neurons can results in a sigmoidal-like response of the dendrite to synaptic input. y-axis: observed depolarization (excitatory post-synaptic potentials, EPSP), x-axis: expected depolarization (EPSP) if the dendrite had a completely linear response. (B) The simplified model which predicts the firing rate output of a CA1 neuron is a 2-layer neural network with sigmoidal activation functions [adapted from Poirazi et al. (2003a,b), with permission]. (C) Hippocampal fast-spiking interneurons (the morphology shown) also have nonlinear dendrites. (D) Dendrites with supralinear responses axes as in (A). (E) Dendrites with sublinear responses. Adapted with permission from Tzilivaki et al. (2019).

Figure 3

Clustering in memory. Computational modeling suggests that two memories that overlap temporally and thus have the ability to share proteins and resources via cooperative plasticity and synaptic tagging, are co-allocated in overlapping neuronal populations. (A) Increase in the size of memory engram (neuronal populations active during recall) of the first memory when paired with a second one as a function of the interval between the two memories (x-axis). LTP cooperativity enables enhanced allocation of the memory for the first 2 h. Panel (B) as in (A), but for the second encoded memory. The increased excitability of the neurons encoding the first memory enables enhanced allocation of the second memory for up to 5 h. (C) The two memories being encoded are allocated in overlapping populations of neurons, and the overlap varies both as a function of the interval between memories (x-axis) and dependent on whether protein synthesis is limited to a single branch, or the entire neuron, or both (color coding). (D) In addition to overlapping populations, memories are also encoded in overlapping dendrites, i.e., they create clusters together, dependent on time and locus of protein synthesis similarly to (C). (E) Imaging of retrosplenial cortex dendrites shows that pre-training synaptic turnover in a stretch of dendrite correlates with the performance of contextual fear memory after a learning task. (F) The learning task resulted in a higher incidence of clustered synapses post-training compared to controls. **p < 0.01. (G) Increased clustering of synapses gained post-training was found to correlate with behavioral performance. Panels (A–D) adapted from Kastellakis et al. (2016), with permission. Panels (E–G) adapted from Frank et al. (2018), with permission.