Основні поняття
A deep generative model called ElectroPhysiomeGAN (EP-GAN) can efficiently estimate the biophysical parameters of Hodgkin-Huxley neuron models from recorded electrophysiological responses of neurons.
Анотація
The paper introduces a novel deep generative model called ElectroPhysiomeGAN (EP-GAN) that can estimate the biophysical parameters of Hodgkin-Huxley (HH) neuron models from recorded electrophysiological data such as membrane potential responses and steady-state current profiles.
Key highlights:
- EP-GAN combines a Generative Adversarial Network (GAN) architecture with a Recurrent Neural Network (RNN) Encoder to generate HH-model parameters given the input electrophysiological recordings.
- The method can generate a large number of HH-model parameters (over 170) from the neuron's membrane potential responses and steady-state current profiles.
- EP-GAN outperforms existing optimization-based methods like Differential Evolution and Genetic Algorithms in terms of accuracy and inference speed when modeling non-spiking neurons in the C. elegans nervous system.
- EP-GAN preserves its performance even when provided with incomplete input data, such as up to 25% missing membrane potential responses or 75% missing steady-state current profiles.
- Ablation studies show the importance of the reconstruction losses (for membrane potential and current) in aligning the Generator to produce accurate HH-model parameters.
- EP-GAN can generalize to model novel neurons in C. elegans that were not previously characterized, achieving lower errors compared to other methods.
- The method is currently limited to non-spiking neurons but could potentially be extended to model neurons with spiking activity in the future.
Статистика
The HH-model used in this study has 15 ion channels and a total of 216 parameters, of which 176 are trainable.
EP-GAN was trained on 32,000 simulated HH-model neurons.
For the validation scenarios, EP-GAN achieved an average RMSE error of 5.84 mV and 5.81 mV for neurons sampled from the training and test data, respectively.
For the 3 experimentally recorded C. elegans neurons (RIM, AFD, AIY), EP-GAN achieved an overall RMSE error of 6.7 mV, which is 40% lower than the next best method (NSGA2 at 11.2 mV).
For the 6 additional C. elegans neurons (AWB, AWC, URX, RIS, DVC, HSN), EP-GAN achieved an even lower overall RMSE error of 5.35 mV.
Цитати
"EP-GAN can learn a translation from electrophysiologically recorded responses and propose projections of them to parameter space."
"EP-GAN is currently limited to non-spiking neurons in C. elegans as it was designed and trained with HH-model describing the ion channels of these neurons."