Multimodal Memory Components and Their Long-Term Dynamics Identified in Cortical Layers II/III but Not Layer V

Dong Li, Guangyu Wang, Hong Xie, Yi Hu, Ji‐Song Guan, Claus C. HilgetagFrontiers in Integrative Neuroscience 2019 · DOI

TL;DR

This study reveals a striking layer-specific organization of long-term memory representation in the mouse cortex. Superficial layers (II/III) exhibit consistent, distributed memory dynamics across different brain areas, suggesting a unified multimodal memory system. In contrast, deep layer V shows more localized and variable dynamics, implying a distributed storage of past experiences across distinct cortical regions.

What Problem Does This Paper Address?

The brain's ability to integrate information from multiple senses (multimodal integration) and form lasting memories is fundamental to cognition. While we know that different cortical areas are involved in learning and memory, how this information is represented and maintained over time, particularly across different layers within the cortex, remains poorly understood. Specifically, the paper aims to uncover whether the long-term dynamics of neural activity patterns associated with multimodal memories are consistent across different cortical areas and layers, or if they exhibit specific layer- and area-dependent characteristics. Understanding this could shed light on how the brain balances plasticity (learning new things) with stability (retaining old information).

Methods

The researchers employed a sophisticated approach combining in vivo two-photon calcium imaging of immediate early gene (IEG) expression in mice with a machine learning framework.

  • IEG Expression Imaging: Mice were exposed to various multimodal environments. After exploration, IEG expression (a marker of recent neural activity) was imaged in specific cortical layers (II/III and V) using a genetically engineered mouse line (BAC-EGR-1-EGFP). This allowed them to visualize which neurons were active in response to different experiences.
  • Machine Learning for Prediction: A gradient boosting decision tree algorithm (LightGBM) was used to predict neural activity patterns on a given day based on activity from previous days. This "prediction power" served as a measure of how well past experiences could predict current neural states, reflecting memory-dependent dynamics.
  • Cross-Location and Intra-Location Prediction: The study performed two types of predictions:
  • Intra-location: Predicting activity within the same cortical location and layer.
  • Cross-location: Predicting activity in one cortical location/layer based on data from a different location/layer. This was crucial for assessing the generality of memory representations.
  • Repeat Environment Prediction: This method tested how well past experiences in a repeated environment could predict current activity, examining the decay of memory signals over time and the influence of environment complexity.

Datasets

The study utilized data from four mice. The specific environments included:

  • Home Cage (baseline, weak multimodal inputs)
  • Context A, B, C (stronger, specific visual, somatosensory, and olfactory inputs)
  • Training A, B, C (contexts paired with foot shocks)
  • Retrieval A, B, C (contexts without foot shocks)
  • Enriched Environment and Tunnel (strong visual and somatosensory inputs)
  • Horizontal and Vertical Stimulus (visual-only environments)

The data collected consisted of IEG expression levels from thousands of neurons across multiple cortical locations and layers over several weeks. No new benchmark datasets were introduced; rather, existing imaging data was analyzed using novel predictive modeling techniques.

Key Insights and Ideas

  • Layered Memory Systems: The most significant finding is the clear distinction in memory representation dynamics between superficial (II/III) and deep (V) cortical layers. This suggests that different layers might serve fundamentally different roles in memory processing.
  • Distributed vs. Localized Memory: Layers II/III exhibit a remarkably consistent memory dynamic across diverse cortical areas and hemispheres. This implies a highly distributed and unified system for representing multimodal memories. Conversely, layer V shows more variable dynamics that are sensitive to the specific cortical location, suggesting that information from past experiences is stored in a more localized and compartmentalized manner.
  • Temporal Sensitivity: Layers II/III appear to be more sensitive to recent temporal information (within days), indicating a dynamic updating of memory representations. Layer V's dynamics are more influenced by the specific history of interactions with different environments, suggesting a more complex interplay of past experiences.
  • Environment Complexity Matters for Layer V: The study found that layer V's memory dynamics are more influenced by the complexity of remembered contexts (i.e., the number and type of sensory modalities involved) compared to layers II/III.

Key Results

  • Cross-location Prediction: Models trained on layers II/III could predict activity in other cortical locations with high accuracy (prediction powers often between 0.6 and 0.8). In contrast, cross-location prediction in layer V performed significantly worse, with many more instances of poorer performance compared to layers II/III.
  • Intra-location Prediction: While still showing a bias towards better performance in layers II/III, the difference between layers was less pronounced for intra-location predictions, suggesting that the layer-specific differences observed in cross-location prediction are not solely due to technical imaging depth issues.
  • Repeat Environment Prediction: Layers II/III showed a stronger sensitivity to the environment-repeat interval (time since the last exposure to the same environment), indicating a faster decay or updating of memory signals compared to layer V. This supports the idea that II/III layers are more attuned to recent temporal information.

Limitations and Future Work

  • Layer Vb Dominance: The authors acknowledge that the properties observed for "layer V" might be heavily influenced by layer Vb, which contains more neurons than layer Va. Future work could aim to disentangle the dynamics of these sub-layers.
  • Technical Limitations: Discriminating between layers II and III, and between layers Va and IV/Vb, proved challenging with the current imaging and annotation methods, leading to combined analysis of these layers.
  • Feature Localization: The current approach can predict memory dynamics but cannot pinpoint which specific cortical areas or features within those areas are responsible for encoding particular pieces of information. Future research should focus on localizing these memory traces.
  • Anatomical Mechanisms: The underlying anatomical mechanisms driving these layer-specific long-term dynamics remain an open question for future investigation.
  • Functional Roles: The precise functional roles of these distinct layer-specific dynamics in learning and memory processes require further experimental exploration.

Summary updated

Abstract

Activity patterns of cerebral cortical regions represent the current environment in which animals receive multi-modal inputs. These patterns are also shaped by the history of activity that reflects learned information on past multimodal exposures. We studied the long-term dynamics of cortical activity patterns during the formation of multimodal memories by analyzing <i>in vivo</i> high-resolution 2-photon mouse brain imaging data of Immediate Early Gene (IEG) expression, resolved by cortical layers. Strikingly, in superficial layers II/III, the patterns showed similar dynamics across structurally and functionally distinct cortical areas and the consistency of dynamic patterns lasted for one to several days. By contrast, in deep layer V, the activity dynamics varied across different areas, and the current activities were sensitive to the previous activities at different time points, depending on the cortical locations, indicating that the information stored in the cortex at different time points was distributed across different cortical areas. These results suggest different roles of superficial and deep layer neurons in the long-term multimodal representation of the environment.
BibTeX
@inproceedings{li2019multimodal,
  author = {Dong Li and Guangyu Wang and Hong Xie and Yi Hu and Ji‐Song Guan and Claus C. Hilgetag},
  title = {{Multimodal Memory Components and Their Long-Term Dynamics Identified in Cortical Layers II/III but Not Layer V}},
  booktitle = {Frontiers in Integrative Neuroscience},
  year = {2019},
  doi = {10.3389/fnint.2019.00054},
}

Loading related content...

Figure