asct 0.0.5

Automated Single Cell Tuner

Automated Single Cell Tuner (ASCT)

The aim of this tool is the automate the parameter selection process for biologically realistic neuronal models using SBI.

Installation:

The installation process differs depending on the platfrom being used, currently only linux and windows have been tested with ASCT.

Linux Installation:

Installation on linux should be relatively straight-forward. There are two major options

1. Install via pip pip install asct
2. Install manually.
   • Download the source code from this github repository git clone https://github.com/pbcanfield/ASCT
   • Change directory to where the repositort was downloaded cd ASCT
   • Run the setup.py script to install the package locally python setup.py install (to install normally) or python setup.py develop (to install in developer mode)

Windows installation:

The installation for windows users should be identical to that of the linux users with several additional step.

1. First you must manually install NEURON on your local machine.
2. Install anaconda.
3. Follow the instructions in the Linux installation section in an anaconda terminal.

Running parameter inference.

The CLI interface.

ASCT is a CLI based tool which takes a configuration file and command line arguments as inputs. For basic usage, run the command python optimize_cell.py -h. Which will produce the following output:

usage: asct [-h] [-g] [-c] [-l] [-n N] config_dir [save_dir]

Uses SBI to find optimal parameter sets for biologically realistic neuron simulations.

positional arguments:
  config_dir  the optimization config file directory
  save_dir    [optional] the directory to save figures to

optional arguments:
  -h, --help  show this help message and exit
  -g          displays graphics
  -c          compiles modfiles automatically (currently only available on linux systems)
  -l          store log files
  -n N        the number of found parameters to show (must be less than the threshold sample size in the optimization
              config file)

From here the user has several options which they can specify depending on their individual needs.

Compiling modfiles on windows:

It should be noted that currently ASCT does not support automatic modfile compilation on windows. To run parameter inference on these systems, the user will have to compile their own modfiles manually using the mkrnrndll utility. This will generate a file named nrnmech.dll which must then be copied into the directory where ASCT is being run.

Configuration file setup.

The only required argument for ASCT is a user defined json configuration file. This is a file which specifies all relevant information which is needed to infer the parameters a user is interested in. There are five subsections required in every configuration file for ASCT: manifest, optimzation_settings,run,conditions, and optimzation_parameters. Each subsection has a variety of options which the user can set for a general run, some of which are required, and some are conditional based on the user's job. Here is the general breakdown of a config file by section.

"manifest"

• Required Settings:
  • "job_type": Specifies if a user wants to optimize parameters from provided current response information or if they would like to validate against a ground truth model. Can be set to either "ground_truth" or "from_data" respectively.
  • "architecture": Specifies what optimization architecture the user wishes to use with this model. There are three options: "convolution","hybrid", and "summary". "convolution" uses a Convolutional Neural Network (CNN) to learn a user defined number of features to encode voltage response data, "summary" uses a user defined function for the same purpose, and "hybrid" uses a combination of both.
  • "template_name": Specifies the name of the NEURON HOC template to use for optimization.
  • "template_dir": The directory where the above template HOC file is stored on the local machine.
• Conditional Settings:
  • If "job_type" is set to "ground_truth":
    1. "target_template_name": Specifies the name of the target NEURON HOC template to use for optimization.
    2. "target_template_dir": The directory where the above target template HOC file is stored on the local machine.
  • If "job_type" is set to "from_data":
    1. "input_data": The CSV file storing the current injection responses for the cell you wish to tune. The file should be comma-seperated where each column corresponds to a seperate current injection value. There should be exactly 1024 data points in each column and each data point should be a voltage (mV) corresponding to the membrane potential at each time step of the experiment.
  • "modfiles_dir": The directory where the mod files for the above neuron template are stored on the local machine. (This is used for automatic compilation on linux which is not supported on windows).

"summary" (Only required if "architecture" is set to "summary" in "manifest")

• Required Settings:
  • "summary_file": This is the file which specifies user defined summary statistics functions. Each function must be implemented by the user and must two numpy arrays of length 1024 for the first two positional arguments. The first array stores the membrane voltage of a given cell and the second is the time vector.
  • "function_name": This is the name of the function to use in the provided file.
• Conditional Settings: All other parameters in this section are conditional based on the function definition. They will be passed into the function as kwargs.

"Run" (all time units are in ms)

• Required Settings:
  • "tstop": The duration of each simulation.
  • "delay": The time before current injection is provided.
  • "duration": The length of the current injection

"Conditions" (all voltage measurements are in mV)

• Required Settings:
  • "v_init": The initial voltage of the simulation.

"optimization_settings"

• Required Settings:
  • "num_simulations": The number of samples to take within the prior distribution.
  • "num_rounds": The number of rounds if using multi-round inference.
  • "workers": The number of workers to use for inference.
• Conditional Settings:
  • If "architecture" is set to "convolution" or "hybrid":
    • "features": The number of features the CNN will learn.

"optimization_paraemeters"

• Required Settings:
  • "current_injections": A list of current injection values to optimize at (in nA).
  • "parameters": The list of parameters to optimize, this should match the exact name of the respective HOC mechanism.
  • "lows": A list of the lower bound for each of the above parameters.
  • "highs": A list of the upper bounds for each of the above parameters.
• Conditional Settings:
  • If "job_type" is set to "ground_truth":
    • "ground_truth": A list of the model ground truth parameters.