TL;DR
This paper bridges the gap between detailed, single-neuron models and simplified neural mass models by creating a "construct-valid" neural mass model. This allows for more accurate analysis of laminar recordings from the brain, revealing how neuromodulators like acetylcholine influence activity in different cortical layers.
What Problem Does This Paper Address?
Understanding how the brain works requires models that operate at different scales. We have highly detailed "compartmental" models that simulate individual neurons and their intricate structures, and more abstract "neural mass" models that represent larger populations of neurons. However, directly linking these scales has been challenging. Specifically, analyzing data from laminar microelectrode recordings, which capture activity across different cortical layers, requires models that can resolve this fine-grained spatial information. Traditional neural mass models, while computationally efficient, often treat brain regions as homogeneous and struggle to differentiate layer-specific activity. This paper aims to bridge this gap by developing a neural mass model that inherits biological plausibility from detailed compartmental models, enabling it to analyze laminar recordings and investigate the effects of neuromodulation on cortical circuits.
Methods
The core innovation lies in creating a neural mass model that is "construct-valid" with respect to more detailed compartmental models. This is achieved through a multi-step process:
- Compartmental Model Simulation: The authors start with a well-established compartmental model of a cortical column (Jones et al., 2007) that captures detailed neuronal morphology and synaptic interactions. They simulate responses from this model.
- Neural Mass Model Derivation: They then derive a corresponding neural mass model (based on the Jansen and Rit model, but with distinct superficial and deep populations) that aims to replicate the dynamics of the compartmental model.
- Parameter Prior Estimation: Crucially, they fit the neural mass model to the simulated data from the compartmental model. The resulting "posterior" parameter estimates from this fit are then used as "prior" constraints for the neural mass model when analyzing real empirical data. This ensures the neural mass model's parameters are biologically informed by the more detailed model.
- Dynamic Causal Modelling (DCM): The derived neural mass model is then used within a DCM framework to analyze empirical laminar recordings. DCM allows for Bayesian inference on model parameters and structure, enabling the estimation of connectivity and the comparison of different model hypotheses.
This approach ensures that the neural mass model is not just a black box but has a foundation in the biophysical realism of compartmental models.
Datasets
The paper uses two primary datasets:
- Simulated Data: Data generated from a validated compartmental model of a cortical column (Jones et al., 2007). This simulated data is used to derive the priors for the neural mass model and to establish its construct validity.
- Empirical Data: Local Field Potential (LFP) data recorded from the primary visual cortex (V1) of awake mice using 32-channel laminar probes. This data was collected during a visual perception paradigm that included optogenetic stimulation of basal forebrain cholinergic neurons. This dataset allows for the testing of the model's ability to distinguish between cortical layers and to assess the effects of neuromodulation.
Key Insights and Ideas
- Bridging Scales: The paper demonstrates a practical method for creating neural mass models that are "construct-valid" with respect to detailed compartmental models. This is achieved by using the latter to inform the priors of the former.
- Laminar Resolution: The developed neural mass model, when integrated with DCM, can effectively distinguish between neural activity in superficial and deep cortical layers, a capability often lacking in standard neural mass models.
- Neuromodulation Effects: The model successfully captures how cholinergic input (via optogenetic stimulation) influences cortical activity, specifically showing an enhancement of superficial layer activity relative to deep layers.
- Predictive Coding Link: The observed disinhibition of superficial layers by cholinergic input aligns with predictive coding theories, where neuromodulators might boost prediction errors that ascend the cortical hierarchy.
- Bayesian Model Comparison: The use of Bayesian model comparison (specifically, relative log-evidence) is shown to be a powerful tool for validating the spatial arrangement of model populations (i.e., correctly assigning superficial/deep activity to supragranular/infragranular layers).
Key Results
- The symmetric compartmental model used for simulations produced evoked responses highly correlated (r=0.9343) with the original Jones et al. (2007) model, demonstrating functional equivalence.
- The neural mass model, informed by compartmental model simulations, could accurately reproduce simulated evoked responses and spectral densities.
- Bayesian model comparison strongly favored the "plausible" model where superficial pyramidal cells were mapped to supragranular layers and deep to infragranular layers, with a relative log-evidence of >3 for both laser ON (1498 vs. 1394) and laser OFF (2334 vs. 2201) conditions. This indicates the model can reliably distinguish laminar activity.
- Analysis of empirical data suggested that cholinergic stimulation led to a significant increase in superficial layer activity (e.g., excitatory activity from superficial pyramidal cells increased by 400%) and a disinhibition of superficial pyramidal cells (self-inhibition decreased by 32%).
Limitations and Future Work
- Model Simplification: While the neural mass model inherits construct validity, it is still a simplification of the detailed compartmental model. The paper acknowledges that it cannot capture all biophysical details, such as back-propagation of action potentials.
- Generalizability: The current study focuses on a specific cortical region (V1) and a specific neuromodulatory system (basal forebrain cholinergic input). Future work could extend this approach to other brain regions and neuromodulators.
- Predictive Coding Hypotheses: The authors suggest that their model can be used to test specific hypotheses related to predictive coding, such as the role of different cortical layers in processing predictions and prediction errors. This is a promising avenue for future research.
- Network Dynamics: Future work will involve extending the model to investigate the laminar topography of current source density and explore causal influences on network dynamics more broadly.
- Technological Advancements: The authors note that further technological advances in laminar probes and compartmental modeling will be crucial for future progress.
Summary updated