TL;DR
This paper introduces a novel framework for identifying "cell assemblies" – groups of neurons that work together – directly from neural spike data. By combining a biophysically-inspired measure of neuron interaction with a powerful graph-based community detection method, they can reveal these assemblies at various levels of detail, from fine-grained to coarse-grained structures. This approach offers a more comprehensive understanding of how neural populations encode information compared to existing methods.
What Problem Does This Paper Address?
As neuroscience tools advance, we can now record the electrical activity (spikes) of hundreds of neurons simultaneously. A fundamental question is how these individual neurons coordinate to form functional groups, known as "cell assemblies," which are thought to underlie complex cognitive functions. Identifying these assemblies is challenging because their size, number, and the precise nature of their interactions are often unknown beforehand. Existing methods often struggle to capture the dynamic, directed, and potentially hierarchical nature of these interactions, or they require prior assumptions about the data. This paper aims to bridge this gap by providing a method that can uncover these assemblies without such constraints.
Methods
The authors propose a two-step approach:
- Biophysically-Inspired Causal Similarity Measure: They developed a new way to quantify the relationship between pairs of neurons based on their spiking history. This measure is "causal" because it considers the temporal order of spikes (i.e., if neuron A fires before neuron B). It also incorporates a simplified model of how excitatory and inhibitory postsynaptic potentials (EPSPs and IPSPs) influence a neuron's firing probability, reflecting their distinct biological roles. This results in a directed "functional connectivity matrix" (FCM).
- Markov Stability for Multiscale Community Detection: The FCM is then treated as a directed network. They apply the Markov Stability framework, a graph-theoretical method that uses a simulated diffusion process on the network to identify communities (groups of nodes). The key advantage here is that Markov Stability can scan across different "Markov times," which correspond to different levels of granularity or resolution, revealing partitions at multiple scales without needing to pre-specify the number of groups.
Datasets
The authors evaluated their method on several datasets:
- Synthetic Spiking Data:
- Data with embedded and hierarchical cell assemblies.
- Data with feedforward-like firing patterns.
- Simulated data from Leaky-Integrate-and-Fire (LIF) networks with various architectures (E-E clustered, E-E hierarchical, E-I clustered). These LIF network simulations serve as a controlled environment to test the method's ability to recover known structural connectivity from spiking dynamics.
- Experimental Data:
- Retinal Ganglion Cell (RGC) recordings from mouse and salamander.
- Hippocampal CA1 and CA3 recordings from rats navigating a linear track.
Key Insights and Ideas
- Uncovering Hidden Hierarchies: The Markov Stability framework's ability to scan across multiple "Markov times" allows it to reveal hierarchical structures in neural activity that simpler clustering methods might miss.
- Biophysical Realism Matters: Incorporating simplified biophysical principles (like the difference between excitatory and inhibitory influences) into the similarity measure leads to more biologically relevant functional connections.
- Directionality is Key: Treating the functional connectivity as a directed graph is crucial for accurately identifying feedforward relationships and understanding the flow of information in neural circuits.
- No A Priori Assumptions Needed: The method's strength lies in its ability to discover cell assemblies without requiring prior knowledge of their number, size, or specific types of interactions.
Key Results
- Synthetic Data: The method successfully identified embedded and hierarchical cell assemblies, as well as feedforward patterns, with high classification accuracy (often 100% for synthetic data). It demonstrated superior performance compared to hierarchical clustering and Modularity optimization in recovering known structures, especially hierarchical ones.
- LIF Networks: The algorithm accurately recovered the underlying structural connectivity in simulated LIF networks, identifying clustered and hierarchical assemblies with high fidelity (e.g., 99.5% accuracy for E-E clustered networks). It also successfully detected mixed excitatory-inhibitory cell assemblies.
- Experimental Data:
- Mouse RGCs: The method correctly grouped spike trains from three distinct RGCs, achieving 98.7% accuracy. It also revealed a coarser partition grouping two OFF cells together.
- Salamander RGCs: Two distinct groups of RGCs were identified, corresponding to transient and sustained response types, with statistically significant differences in their temporal characteristics.
- Rat Hippocampal Data: The analysis identified inhibitory neurons and four groups containing place cells, accounting for 100% of the identified place cells.
Limitations and Future Work
- Parameter Sensitivity: While the method is robust to the choice of the synaptic time constant ($\tau$), it's acknowledged that more specialized measures could outperform the simple biophysical similarity for specific cases.
- Interpretation of Multiple Partitions: In noisy real-world data, multiple candidate partitions might emerge. The authors suggest that further biological knowledge can complement the method's output to select the most relevant partition.
- Future Work: The authors suggest several avenues for future research, including:
- More explicit use of neuron firing statistics (e.g., burst firing).
- Developing more specialized biophysical coupling measures.
- Enforcing sparsity constraints on couplings.
- Investigating how cell assemblies change over time and relate to learning.
- Studying the relationship between neuronal time scales and Markov times.
- Exploring the use of this method for more specialized applications, such as place cell detection.
Summary updated