The Role of Intrinsic Plasticity in Engram Physiology and Temporal Memory Linking

The neuroplastic mechanisms that endow the brain with the capacity to learn and form memories remain a major topic of research in neuroscience and in the development of brain-inspired computational algorithms. Although research into the formation of engrams, the neural assemblies, and codes by which memories are created and stored is primarily focused on synaptic plasticity, a growing body of research is investigating the connection between other forms of plasticity, including intrinsic plasticity, and memory storage.

Intrinsic plasticity, in contrast to synaptic plasticity, is a modification of neuronal excitability. Evidence to support the existence of a relationship between memory formation and intrinsic excitability has existed for nearly half a century (Zhang and Linden, 2003). An early study, conducted by Brons and Woody, of pavlovian conditioning in cats’ sensorimotor cortices found that learned motor performance was reflected in sustained increases of cortical neural excitability recorded via surface electrodes (reviewed in Zhang and Linden, 2003), a landmark observation that motivated thousands of studies. One notable experiment, performed by Moyer, investigated the transience of learning-related changes in rabbits after the animals were conditioned to blink in response to a tone in order to avoid a puff of air to the eye (reviewed in Zhang and Linden, 2003). In electrophysiological recordings of the rabbits’ hippocampal pyramidal neurons, Moyer et al. observed persistent increases in intrinsic excitability for up to 7 d post-training. After which, the conditioned eyeblink response remained strong, despite intrinsic excitability levels returning to their baseline values. Based on this finding, the authors speculated that increases in excitability prime neurons for learning and memory consolidation over a short time period. This led to the hypothesis that neuron priming via intrinsic plasticity makes the network more amenable to changes in synaptic efficacy.

Further efforts to unravel the mechanisms of intrinsic plasticity and its relation to engram formation identified a contribution of cAMP response element–binding protein (CREB) to memory allocation in mice (Zhou et al., 2009). This work revealed that neurons with higher CREB levels were preferentially recruited into the memory trace, because (1) they were more excitable than their neighbors and (2) their synapses underwent larger weight changes after conditioning. Rashid et al. (2016) built upon Zhou's findings by showing that CREB-activated excitability dynamics helped link memories of two events presented to the network in temporal proximity. Increased excitability of the neuronal ensemble formed in response to the first context persisted for a short period; if a second event was presented during this time, the first population of excited neurons was poised to participate in the engram representing the second context. This created a structural overlap between the two engrams, thus linking the memories.

It is important to note that synaptic plasticity was not evident in the findings of Rashid et al. (2016), implying that excitability changes were sufficient to enable engram formation and memory linking. However, it is possible that synaptic changes went undetected or that other neural mechanisms not explored by the authors contributed to memory formation and linkage. A good way to test the sufficiency of identified neural mechanisms to produce the observed effect is through computational models. Therefore, in a recent paper in the Journal of Neuroscience, Delamare et al. (2024) used a recurrent neural network (RNN) to reproduce the experimental results of Rashid et al. (2016) and Cai et al. (2016) regarding the role of CREB-governed inherent neuronal excitability on engram formation and overlap.

Although most computational studies that investigate engram formation focus on synaptic plasticity. There are plenty of studies that consider memory storage from an intrinsic plasticity context. Some of these studies presage the approach of Delamare et al. (2024). An early work used a coupled plasticity mechanism, which includes intrinsic and synaptic learning rules on a full bidirectional, biophysically simplified network, and conveyed how changes in excitability act as a memory trace for learning (Janowitz and Van Rossum, 2006). Despite successfully showing that intrinsic plasticity boosted Hebbian learning, excitability was represented as a simple two-state variable, which doesn't capture the complexities of neuronal excitability dynamics as seen in the brain. Later work used a more biophysically accurate model, which accounted for CREB-activated excitability dynamics and a bioinspired intrinsic connectivity network (ICN), to demonstrate that both intrinsic excitability and competitive synaptic interactions may bias neuron allocation during engram formation (Kim et al., 2013). However, Kim et al. used an ICN model instead of an RNN. While both of these architectures have the ability to store memory traces, ICNs are a theoretical framework that is used to analyze brain organization and physiology and does not contain recurrent synaptic connections. Moreover, Kim et al. (2013) only evaluated the engram response of a single event, and not the engram overlap of temporally proximal events. More recent computational studies, such as Kastellakis et al. (2016), evaluated neuronal assembly overlap using a CREB-induced excitability model in conjunction with multiple synaptic plasticity mechanisms on a simplistic integrate-and-fire feedback network. But their architecture differed from that of Delamare's in a couple of ways. First, they used inhibitory feedback connections instead of recurrent ones, which typically transmit contextual signals from one neuronal population to another and are not specifically designed for retaining state information. Additionally, Kastellakis et al. focused on the excitability of a single neuronal compartment, the dendrite, instead of the entire neuron. Delamare et al. (2024) built off these works by exploring the role of CREB-governed excitability in memory linking using an RNN outfitted with neuron-wide excitability and a Hebbian learning rule.

Delamare's simulations involved testing the recurrent model's ability to form individual memory engrams, overlapping engrams in response to two temporally contiguous contexts, and links between two events presented in quick temporal succession. Additional trials were performed to determine the degree to which excitability biases memory allocation, as well as to evaluate the extent of inhibition-induced neuronal competition in memory allocation.

To capture the temporal dynamics of the stimuli more effectively, Delamare used a more biologically plausible variant of a traditional RNN, termed the rate-based RNN. In this framework, the model's rate dynamics are governed by continuous-time differential equations, as opposed to classic RNNs which use discrete equations. When a neuron's firing rate reached a predefined threshold, the excitability of the neuron instantaneously jumped up to a higher value before decaying over time back to its initial value, mimicking what occurs in biological neurons when excitability changes are induced by an increase in CREB levels. These network units were partitioned into three populations, each representing one of three presented contexts. During training, the firing rates of one of the three neuronal assemblies were manually increased to some predefined value, analogous to “activating” the ensemble, for a short duration over a number of epochs. A simulated shock was also applied to a subpopulation's synaptic connections, as a virtual unconditioned stimulus, synchronously with the context during the training phase. The shock acts as a potent stimulus that enhances the excitability of neurons, making them more likely to participate in forming a memory trace, or engram. This helped the researchers understand how increased excitability influences the likelihood of a neuron being included in an engram. To measure the memory recall performance of the model during the recall phase of the experiment, a behavioral readout variable was introduced that summates the firing rates of neurons in the ensemble over a short temporal window. Several network attributes, such as the recurrent weight matrix and engram overlap, were evaluated in addition to the memory strength metric.

To get intuition for the results, we will discuss one example that simulated a fear conditioning experiment inspired by Cai et al. (2016). Delamare et al. sought to determine whether engram overlap creates a relationship between temporally proximal memories. Three contexts, red, yellow, and blue, were presented to the network. The red context came first, followed by the yellow 7 d later, and the blue was presented after another 5 h. The purpose of the differing temporal disparities was to verify that engram overlap only occurred in their model when two contexts were presented in quick succession. To measure memory linking, the last context was presented a second time with shock, after which the recall performance for each context was measured. The memory strength when recalling either of the temporally proximal contexts (5 h delay) was high, but it was low for the temporally distal (7 d delay) context. Moreover, the ensembles created by the presentation of the yellow and blue contexts overlapped structurally as a result of their temporal proximity, whereas the ensemble generated by the red context was distinct. These results confirm that engram overlap evokes memory linking, thereby reproducing the experimental findings of Cai et al. (2016). Delamare et al. also conducted an ablation study involving a single context to further strengthen their deductions regarding the impact of increased intrinsic excitability on memory allocation. During the presentation, the excitability of a subpopulation of engram neurons that responded preferentially was enhanced. During recall, the same neuronal subgroup was blocked. They found that inhibiting the enhanced subpopulation resulted in worse memory recall performance compared with a control case in which the neurons’ excitability was not manipulated. This ablation study experimentally confirmed the link between intrinsic excitability and engram formation, confirming their earlier findings.

Engram evolution over time was not considered in Delamare et al. (2024), but it is an important topic to address, given that engrams have been found to be dynamic in the brain. The same group recently published a paper investigating this very question using a recurrent spiking neural network (Tomé et al., 2024). They found that postencoding memories undergo a refinement process in which neurons drop in and out of the engram and inhibitory synaptic plasticity drives this process. However, the model behavior was slightly different in that Tomé et al. didn't consider CREB-governed excitability, leaving an open question to be addressed.

Another area of inquiry suggested by Delamare et al. is the precise type of information transferred between overlapping engrams. This line of research has been explored with respect to conceptual memory linking. Gastaldi et al. (2021) proposed a computational framework capable of performing a chain-like recall of closely associated concepts, similar to how the recall of the first word in a sentence sparks the sequential activation of the rest of the sentence. However, the information transfer mechanism used when recollecting chains of temporally contiguous events is still unknown.

By demonstrating the rate-based RNN's capacity to synthesize and sustain engrams, link memories via structural assembly overlaps, and allocate memory through a neuronal competition mechanism, Delamare et al. successfully reproduced the experimental findings of Rashid et al. (2016) and Cai et al. (2016). Furthermore, Delamare's simplified computational approach helped clarify the precise role of intrinsic excitability in memory formation, setting the stage for future investigations into the plasticity mechanisms that underpin engram physiology.