Accurate Identification of Communication Between Multiple Interacting Neural Populations

Area of Science:

• Neuroscience
• Computational Neuroscience
• Systems Neuroscience

Background:

• Simultaneous neural recordings across multiple brain regions are now feasible.
• Existing models struggle to accurately differentiate communication sources influencing neural populations.
• This limitation hinders a clear understanding of inter-regional neural communication.

Purpose of the Study:

• To introduce a novel computational framework, Multi-Region Latent Factor Analysis via Dynamical Systems (MR-LFADS).
• To develop a model capable of disentangling inter-regional communication, external inputs, and local neural dynamics.
• To improve the accuracy of modeling brain-wide information processing.

Main Methods:

• MR-LFADS is a sequential variational autoencoder designed for analyzing multi-region neural data.
• The model employs dynamical systems to capture temporal dependencies in neural activity.
• It is validated using simulations of task-trained multi-region networks and large-scale electrophysiology data.

Main Results:

• MR-LFADS demonstrated superior performance in identifying communication across simulated neural networks compared to existing methods.
• The model successfully predicted brain-wide effects of circuit perturbations in real electrophysiology data, even for perturbations not used during training.
• MR-LFADS effectively disentangles distinct sources of neural activity, including inter-regional communication and local dynamics.

Conclusions:

• MR-LFADS offers a significant advancement in modeling neural communication across multiple brain regions.
• The model provides a more accurate representation of brain-wide information processing.
• MR-LFADS is a valuable tool for uncovering fundamental principles governing neural interactions and information flow in the brain.