Neural Network based ion-channel model fitting

Implementation of PINNS (physically inspired neural networks) to model ion-channels dynamics. A combination of experimental data and surrogate data (artificially generated using Hodgkin-Huxley ODEs) are used to train the PINNs-model.

Repository structure

• "replication/" contains the original code (Tensorflow) from Raissi's paper and the replication in PyTorch

  • Raissi_original/Burgers.py is the original code
  • Raissi_replication/burgers_rep3.py is the replication. It learns the PDE parameters both for noiseless and noisy data.
  • the working directory should be replication/

• "ion_channel_fit/" contains the data and code related to the modeling of ion-channels with PINNS.

  • fit_all_act.py contains the class for the model fitting on data with multiple activation voltages
  • fit_one_act.py contains the class for the model fitting on data with one selected activation voltage
  • fit_one_notebook.ipynb is used to train the single-activation-voltage model, plot and store the results
  • fit_one_reps.ipynb is used to train the single-activation-voltage model on several experimental repetitions, plot and store the results
  • fit_all_notebook.ipynb is used to train the multiple-activation-voltage model, plot and store the results
  • fit_all_reps.ipynb is used to train the multiple-activation-voltage model on several experimental repetitions, plot and store the results
  • "surrogate/" contains a notebook that can generate artificial data with selected levels of noise
  • "plots/" is the directory where the results' plots are stored
  • the working directory must be ion_channel_fit/

PARAMETERS

In the replication part, the following parameters can be tuned

• PLOT: whether to plot results
• MAX_ITER_LBFGS: maximal number of iterations per LBFGS-optimization step
• NOISE: noise percentage for the fitting on noisy data

In the ion-channel fitting part, the following parameters can be tuned:

• PLOT: whether to plot results
• MAX_ITER_LBFGS: maximal number of iterations per LBFGS-optimization step
• LINESEARCH: line search function (either "strong_wolfe" or None)
• NN_DEPTH: depth of the neural nets (number of layers)
• NN_BREADTH: breadth of the neural nets (number of neurons per layer)
• E_K: channel reversial potential (here, potassium)
• G_K: channel ref conductance (here, potassium)
• ADAM_IT: number of ADAM iteration to perform
• WHOLE_DATA: whether or not to use the whole data as training set
• DEVICE: "cpu" or "cuda". Whether to train on cpu or gpu
• EXP_REP: repetition (activation voltage) to consider (only for fit_one_act)