Physically constrained neural networks for inferring physiological system models.

Area of Science:

• Biomedical Sciences
• Computational Neuroscience
• Systems Biology

Background:

• Systems biology and neurophysiology models are vital but computationally demanding.
• Traditional models often require complex, multiscale, and multiphysics strategies.
• Deep neural networks show promise for approximating complex nonlinear problems efficiently.

Purpose of the Study:

• To apply deep neural networks for solving complex biophysical models.
• To utilize physically constrained neural networks (PINNs) for neurophysiology modeling.
• To infer parameters and time-courses from real neurophysiological data.

Main Methods:

• Validated deep neural networks using synthetic data.
• Employed PINNs to solve the Hodgkin-Huxley model.
• Inferred model parameters and hidden time-courses from real electrophysiology data under varied stimulation.

Main Results:

• Achieved accurate signal reconstruction with low variability across neural spikes.
• Obtained parameter ranges consistent with existing biological knowledge.
• Demonstrated the ability of neural networks to fit complex dynamics using integrated biological knowledge.

Conclusions:

• PINNs provide a computationally efficient and accurate method for complex biophysical modeling.
• Integrating biological knowledge into neural networks enhances their fitting capabilities for real-world data.
• This approach advances systems neurophysiology and biomedical applications.