Rapid, interpretable data-driven models of neural dynamics using recurrent mechanistic models

Area of Science:

• Computational Neuroscience
• Systems Neuroscience
• Machine Learning

Background:

• Modeling neural systems is complex, with trade-offs between detailed, intractable models and simplified, less predictive ones.
• Existing methods struggle with model complexity, fitting efficiency, and interpretability.

Purpose of the Study:

• To present a novel modeling paradigm for creating predictive, mechanistic models of neurons and small neural circuits.
• To address the challenges of model complexity, efficiency, and interpretability in neural modeling.

Main Methods:

• Utilized systems theory, combining linear state-space models and nonlinear artificial neural networks.
• Developed two types of membrane current models: flexible lumped currents and interpretable data-driven conductance-based currents.
• Introduced recurrent mechanistic models (RMMs) for efficient training on intracellular recordings.

Main Results:

• RMMs can be trained in seconds to minutes, a significant improvement over previous methods.
• Successfully applied RMMs to model the dynamics and synaptic connections of neurons in the Stomatogastric Ganglion.
• Demonstrated the reliability, efficiency, and interpretability of the RMM approach.

Conclusions:

• RMMs offer a powerful new approach for estimating predictive neural models.
• The efficiency and interpretability of RMMs enable new experimental possibilities in closed-loop neurophysiology.
• RMMs facilitate online estimation of neural properties in living preparations.