Abstract
Recent advances in connectomics and biophysics demand detailed neuronal models that incorporate both network interactions and cellular dynamics. Creating these "ElectroPhysiomes" requires estimating hundreds of parameters for individual neurons from electrophysiological recordings—a task traditionally slow and computationally expensive. We present ElectroPhysiomeGAN (EP-GAN), a deep generative model that combines a Recurrent Neural Network (RNN) encoder with a Generative Adversarial Network (GAN) to instantly generate Hodgkin-Huxley (HH) model parameters. Validated on both synthetic and experimental C. elegans data, EP-GAN outperforms traditional optimization methods in both accuracy and speed, enabling whole-organism biophysical simulations.
Introduction: The Challenge of Parameter Estimation
To understand the nervous system, researchers must bridge the gap between structural connectivity (the connectome) and functional behavior. This requires biophysical models, typically based on Hodgkin-Huxley equations, which describe the dynamics of various ion channels. However, a single detailed neuron model can involve over 170 parameters. Traditionally, these are found through manual "hand-tuning" or evolutionary algorithms that can take days to converge for a single cell. As we move toward simulating entire nervous systems, such as the somatic neurons of C. elegans, these methods become a significant bottleneck.
Methods: The EP-GAN Architecture
ElectroPhysiomeGAN (EP-GAN) shifts the paradigm from optimization to generation. Instead of searching for parameters for each new cell, the model learns the underlying mapping from electrophysiological responses to biophysical parameters.
The architecture consists of three primary components:
- RNN Encoder: Utilizes Gated Recurrent Units (GRUs) to process time-series membrane potential recordings and steady-state current profiles, compressing them into a latent representation.
- Generative Adversarial Network (GAN): A Wasserstein GAN with Gradient Penalty (WGAN-GP) where the Generator produces parameter sets and the Discriminator ensures they match the distribution of valid biological parameters.
- Physics-Informed Loss: The model incorporates a reconstruction loss by simulating the neuron's response using the generated parameters and comparing it to the original input, ensuring the output is biophysically grounded.
Key Results and Validation
The model was trained on 32,000 synthetic HH neurons and validated against both existing benchmarks (RIM, AFD, and AIY neurons) and six newly recorded C. elegans neurons (AWB, AWC, URX, RIS, DVC, and HSN).
Accuracy and Robustness
EP-GAN achieved a Root Mean Square Error (RMSE) of 6.7mV, significantly better than traditional evolutionary algorithms like NSGA2 (11.2mV). Notably, the model remains robust even with incomplete data, maintaining performance when up to 25% of the voltage trace or 75% of the current profile is missing.
Inference Speed
The most dramatic improvement is in computational efficiency. While traditional methods would require approximately 67 days to model the somatic nervous system of C. elegans, EP-GAN can complete the task in less than a day. Once trained, the inference for a single neuron takes only milliseconds.
Discussion and Future Directions
While EP-GAN represents a major step forward, it currently focuses on graded potential neurons. Future iterations will aim to capture the complex dynamics of spiking neurons and improve the modeling of fine-grained temporal features like voltage "overshoots." The ultimate goal is the integration of these individual models into a full "ElectroPhysiome," providing a functional, biophysically detailed simulation of the entire C. elegans nervous system.
Summary updated