TL;DR
This paper presents a computational model demonstrating how the "readiness" of a neuron to fire—its intrinsic excitability—determines which neurons are recruited to store a new memory. By simulating a recurrent neural network, the authors show that events occurring close in time share overlapping neural populations (engrams) because the high excitability triggered by the first event biases the allocation of the second, effectively "linking" the two memories.
What Problem Does This Paper Address?
In neuroscience, the "engram" is the physical ensemble of neurons that represents a specific memory. A long-standing question is how the brain decides which specific neurons among millions should be "allocated" to a new memory. While traditional theories focus on synaptic plasticity (the strengthening of connections between neurons), recent experimental evidence suggests that a neuron’s internal state—specifically its intrinsic excitability—plays a crucial role.
Furthermore, we know that memories of events happening close in time (e.g., within a few hours) often become linked, so that recalling one brings up the other. This paper seeks to provide a mechanistic theoretical framework to explain how time-varying excitability drives both the initial allocation of neurons to an engram and the subsequent overlap between engrams that enables memory linking.
Methods
The authors developed a rate-based recurrent neural network (RNN). Unlike standard models that focus solely on Hebbian "fire together, wire together" synaptic rules, this model incorporates two interacting systems:
- Synaptic Plasticity: A Hebbian learning rule that modifies the weights between neurons based on their activity.
- Intrinsic Excitability: Modeled as a time-varying threshold ($\epsilon_i$). When a neuron becomes part of an engram, its excitability increases (lowering its firing threshold) for a transient period before decaying back to baseline over approximately 24 hours.
The researchers used this model to simulate classic behavioral experiments, such as fear conditioning, where they could measure "memory strength" (a proxy for freezing behavior in rodents) by calculating the total activity of the engram during a recall phase.
Datasets
As a theoretical computational study, the paper does not use external biological datasets. Instead, it generates synthetic data through simulations of:
- Contextual Receptive Fields: Simulated inputs representing different environments.
- Experimental Benchmarks: The model's outputs (engram overlap and memory strength) were directly compared against published experimental results from the lateral amygdala (Rashid et al., 2016) and the hippocampal CA1 (Cai et al., 2016).
Key Insights and Ideas
- The "Winner-Take-All" Competition: The model reveals that memory allocation is a competitive process driven by global inhibition. When a stimulus arrives, neurons with higher initial excitability fire first and more strongly; via inhibitory interneurons, they suppress their neighbors, effectively "winning" the right to represent that memory.
- The Temporal Window of Linking: The 24-hour decay constant of excitability (linked to the transcription factor CREB) creates a "temporal window." If a second event occurs within ~6 hours, the neurons from the first engram are still highly excitable and are "re-recruited" into the second engram. If the delay is 24 hours, the excitability has reset, leading to distinct, non-overlapping engrams.
- Excitability vs. Input Strength: The authors found that even if a neuron receives weaker sensory input, high intrinsic excitability can compensate, allowing it to be recruited into an engram over "fitter" neighbors.
- Functional Overlap: The structural overlap (shared neurons) leads to functional linking. Stimulating a subset of neurons from the first memory can trigger the "pattern completion" of the second memory, explaining why temporally close events are recalled together.
Key Results
- Overlap Dynamics: The model successfully reproduced the observation that engram overlap is significantly higher for events separated by 6 hours compared to those separated by 24 hours.
- Biasing Allocation: By artificially increasing the excitability of a subset of neurons in the simulation, the authors could force those neurons to be recruited into the engram, matching "optogenetic" manipulations seen in real-world experiments.
- Inhibition as a Regulator: Reducing the network's global inhibition during the second event allowed for the formation of a memory even when the "preferred" (highly excitable) neurons were blocked, confirming that inhibition is the primary driver of the competition for allocation.
- Predictive Constraints: The model predicts that memory linking only occurs within a specific range of feed-forward input weights; if inputs are too strong, they override the excitability bias; if too weak, the second memory fails to form.
Limitations and Future Work
The authors acknowledge that their rate-based model is a simplification. Real neurons communicate via discrete electrical spikes, and future work could involve spiking neural networks to capture finer temporal dynamics.
Additionally, the model uses a homogeneous global inhibition mechanism. In the actual brain, there are diverse types of inhibitory interneurons (like PV+ or SOM+ cells) that likely play more nuanced roles in shaping engrams. Finally, the study focuses on the formation of overlaps but does not address systems consolidation—how these linked memories might change or move from the hippocampus to the cortex over weeks or months.
Summary updated