Distributed subthreshold representation of sharp wave-ripples by hilar mossy cells

Ayako Ouchi, Taro Toyoizumi, Nobuyoshi Matsumoto, Genki MinamisawaeLife 2025 · DOI

TL;DR

This study reveals how a small population of hippocampal mossy cells (MCs) collectively encode complex information from sharp wave-ripples (SWRs) originating in the CA3 region. Using a novel machine learning approach, the researchers demonstrate that even a single MC can predict SWR waveforms, and that this predictive power scales sublinearly with more MCs, suggesting a distributed and pseudo-orthogonal representation of SWR information.

What Problem Does This Paper Address?

The hippocampus is crucial for memory formation and retrieval, and its intricate circuitry involves layers of neurons that process information. Some of these layers, like the hilar mossy cells (MCs) in the dentate gyrus, are remarkably sparse compared to other hippocampal regions. This raises a fundamental question: how can such a small population of neurons efficiently encode and transmit complex information, especially during critical events like sharp wave-ripples (SWRs)? SWRs are synchronous neural events thought to be involved in memory replay and consolidation, originating in the CA3 region and propagating through the hippocampus. Understanding how MCs, acting as a bottleneck, process this SWR information is key to deciphering hippocampal function.

Methods

The researchers employed a multi-pronged approach combining electrophysiology and machine learning.

  • Simultaneous Patch-Clamp Recordings: They performed in vitro and in vivo whole-cell patch-clamp recordings from multiple MCs (up to five in vitro) while simultaneously recording local field potentials (LFPs) from the CA3 region to capture SWRs. This allowed them to observe the subthreshold membrane potential (Vm) dynamics of MCs during SWR events.
  • Machine Learning Model: A novel neural network, inspired by autoencoder architectures, was developed to predict SWR waveforms based on the Vm traces of MCs. This model featured an encoder-decoder structure designed for efficient data compression and feature extraction.
  • Cross-Validation and Surrogate Data: To rigorously assess the model's predictive power, they used a five-fold cross-validation strategy. Crucially, they compared predictions made using real MC Vm data against predictions made using shuffled Vm data (surrogate data) to distinguish genuine information encoding from chance correlations.
  • Statistical Analysis: Various statistical methods, including paired t-tests, ANOVA, and non-parametric tests, were used to quantify differences in prediction accuracy and assess the significance of their findings.

Datasets

The study utilized data collected from in vitro and in vivo electrophysiological recordings in mice.

  • In vitro recordings: Data were obtained from multiple MCs simultaneously recorded in acute hippocampal slices.
  • In vivo recordings: Data were collected from single MCs in urethane-anesthetized mice.

The paper does not introduce new benchmark datasets but rather uses the collected electrophysiological data for training and evaluating their machine learning model.

Key Insights and Ideas

  • Subthreshold Information Encoding: The study strongly suggests that crucial information about SWR waveforms is encoded in the subthreshold membrane potential dynamics of MCs, rather than solely relying on spiking activity.
  • Distributed Representation: SWR information is not localized to a single MC but is distributed across a population of MCs. While individual MCs show some specificity in the SWRs they can predict, there's also a degree of overlap.
  • Pseudo-Orthogonal Coding: The finding that the predictable SWRs for different MCs are largely independent (low correlation coefficients) points towards a "pseudo-orthogonal" coding scheme. This means that each MC contributes unique information, minimizing redundancy and maximizing the capacity of the MC population to represent diverse SWRs.
  • Bottleneck Efficiency: The sublinear increase in predictive power with more MCs highlights the efficiency of this sparse layer. It suggests that adding more MCs provides diminishing returns, indicating an optimal population size for compressing and relaying information without significant loss.

Key Results

  • Predictable SWRs from Single MCs: Even a single MC's Vm activity could predict SWR waveforms significantly better than chance (about 9% of SWRs predictable by a single MC).
  • Improved Prediction with More MCs: The accuracy of SWR prediction increased sublinearly as the number of simultaneously recorded MCs increased. With five MCs, approximately 30% of total SWRs could be predicted.
  • Model Performance: The neural network achieved significantly lower RMSE (root mean square error) when predicting SWR waveforms from real MC Vm data compared to shuffled data, indicating successful information extraction.
  • Spatial Bias: MCs located in the lower blade of the dentate gyrus showed higher predictability for SWRs, suggesting a potential anatomical organization of information processing.
  • No Correlation with Intrinsic Properties: The predictability of SWRs by an MC was not correlated with its intrinsic electrophysiological properties like membrane capacitance or resistance.

Limitations and Future Work

  • In Vivo vs. In Vitro Differences: The authors acknowledge that in vivo recordings might capture more complete synaptic input compared to in vitro slices, and that anesthesia can affect neural activity. Direct comparison of in vivo CA3 LFP with in vivo MC Vm would be ideal but is technically challenging.
  • Subthreshold Focus: While powerful, the study primarily focuses on subthreshold activity. Future work could explore how this subthreshold encoding translates to spiking activity and its downstream effects.
  • Anatomical Specificity: The observed spatial bias in predictability warrants further investigation to understand the underlying circuit mechanisms.
  • Future Experiments: The authors suggest future experiments in awake animals to verify these findings and explore how memory information transmitted via feedforward pathways differs from this backprojected SWR information. They also propose investigating the role of MCs in encoding specific, experience-related memories.

Summary updated

Abstract

In neural information processing, the nervous system transmits neuronal activity between layers of neural circuits, occasionally passing through small layers composed only of sparse neurons. Hippocampal hilar mossy cells (MCs) constitute such a typical bottleneck layer. However, how efficient information encoding is achieved within such constrained layers remains poorly understood. To address this, we focused on sharp wave-ripples (SWRs) - synchronous neural events originating in the CA3 region - and investigated functional diversity within MC populations using in vivo/in vitro patch-clamp recordings in mice. By combining machine learning algorithms, we developed a model to predict CA3 SWR waveforms based on the synaptic response waveforms of MCs, suggesting that SWR-related information is indeed encoded in their subthreshold activity. While individual MCs were generally associated with specific SWR clusters, partial overlap across some MCs was also observed, indicating that CA3 activity is distributed across the MC population. Our findings suggest that CA3 SWR activity is represented in a pseudo-orthogonal manner across MC populations, allowing the small MC layer to effectively compress and relay hippocampal information.
BibTeX
@inproceedings{ouchi2025distributed,
  author = {Ayako Ouchi and Taro Toyoizumi and Nobuyoshi Matsumoto and Genki Minamisawa},
  title = {{Distributed subthreshold representation of sharp wave-ripples by hilar mossy cells}},
  booktitle = {eLife},
  year = {2025},
  doi = {10.7554/elife.97270},
}
Figure