Automatic Cell Tuner (act)

act provides tools for optimization-based parameter selection for biologically realistic cell models developed in NEURON. The project is inspired by the ASCT library.

act relies on a simulation-based optimization, i.e., for a pipeline

Parameters -> Black-box simulator -> Simulated data

it tries to obtain parameter estimates indirectly by working with simulated data.

Installation

Currently, act can be installed from GitHub using pip or locally with the standard pip installation process.

pip install act-neuron

git clone https://github.com/V-Marco/ACT.git
cd ACT
pip install .

Usage

Prerequisites

Conceptually, act requires three components.

1. A .hoc file which declares the cell's properties.
2. Modfiles for this .hoc file.
3. Target voltage data of shape (num_cur_inj, ...) to predict on OR parameters to simulate target data with.

Pipeline

act operates in original and segregated modes. Original mode runs in the following steps:

1. Generate a parameter set uniformly randomly from a (lower; upper) interval for each current injection.
2. Simulate a voltage trace for each current injection and respective parameter set.
3. Extract key summary features (e.g., inter-spike time), and keep parameter sets for those voltage traces which match the target voltage trace in these summary features.
4. Repeat steps 1-3 until the specified number of current injections is matched.
5. Train a neural network model to predict conductance values from a voltage trace using saved sets as targets.
6. Predict conductance values by applying the trained model to the target voltage data. Take the maximum of each predicted value across all current injections.

Segregated mode changes step 5 so that the model is trained on regions of a voltage trace. The regions can be specified in terms of time (X-axis) or voltage (Y-axis) bounds.

Setting up a simulation

Simulations' parameters are defined as python classes in simulation/simulation_constants.py.

• Names of parameters to optimize for are defined in the params property. The names must match the hoc file. Lower and upper bounds are specified in lows and highs properties.
• Segregated parameters and respective time/voltage bounds are specified as lists-of-lists in the respective segr_... properties.

Running a simulation

simulation/simulation_configs.py is an example of the simulation configs that can be read by ACT

simulation/run_simulation.py is an example script of running act on Pospichil's cells.

simulation/analyze_res.py is an example script which gives a summary of the model's quality.

ACT can also generate traces over a large parameter space (#amps * #parameter-splits-between-low-high ^ #parameters) To do this you must generate a BMTK network, and run using MPI. Eg:

pip install -e .
module load mpich-x86_64-nopy
pip install mpi4py bmtk

cd simulation
python generate_traces.py build
sbatch batch_generate_traces.sh
#    This will create: 
#        parameter_values.json which contains every permutation of parameters
#        output/v_report.h5 which contains the resulting trace


python run_simulation.py
python analyze_res.py

Examples (Jupyter Notebook)

examples/Pospischil_sPYr/main.ipynb example of running act on Pospichil's cells

On Google Colab:

• Pospischil_sPYr
• LA Type A