All code uses the mackelab sbi package (https://github.com/sbi-dev/sbi) in a python 3.9 environment.
This directory contains:
- The parameters (Simulations_modelparameters.csv) and resulting MEA features (Simulations_MEAfeatures.csv) of the 300,000 simulations used to train the NDE for the paper. Every row represents one simulation. In "Simulations_modelparameters", the columns in ascending order represent the parameters: 'noise', '$g_{Na}$', '$g_{K}$', '$g_{AHP}$', '$g_{AMPA}$', '$g_{NMDA}$', 'Conn%', r'$\tau_{D}$', 'U (STD)', 'U asyn'. In "Simulations_MEAfeatures", the columns in ascending order represent the MEA features: 'MFR', 'NBR', 'NBD', 'PSIB', '#FBs', 'CVIBI', 'mean CC', 'sd CC', 'mean ISI CC', 'sd ISI CC', 'ISI dist', 'mean ISI', 'sd ISI temp', 'sd isi elec', 'MAC'
- The "TrainedNDE": the neural density estimator trained using the 300,000 simulations in the mackelab sbi package. This can be loaded in python to evaluate and obtain the posterior distribution as described in 'FindPosterior.py".
- The "Simulator.py" contains the functions to perform simulations with the biophysical computational model of hiPSC-derived neuronal networks on MEA, and the function to extract the MEA features from either simulations or experimental data. For the simulations, the "elecranges.dat" is necessary.
- "MakeFigures.py": contains some functions to construct images as used in:
- "FindPosterior.py": contains the code to construct images like in the paper, and to asses your own experimental data using the method described in the paper. Also three example experimental observations are provided example_observations: "SCN_WTC_2410", "SCN_DS_2410" and "SCN_GEFS_2410" to construct some images and try the code if your own data is not available.
- The parameters (Simulations_modelparameters.csv) and resulting MEA features (Simulations_MEAfeatures.csv) of the 300,000 simulations used to train the NDE for the paper. Every row represents one simulation. In "Simulations_modelparameters", the columns in ascending order represent the parameters: 'noise', '$g_{Na}$', '$g_{K}$', '$g_{AHP}$', '$g_{AMPA}$', '$g_{NMDA}$', 'Conn%', r'$\tau_{D}$', 'U (STD)', 'U asyn'. In "Simulations_MEAfeatures", the columns in ascending order represent the MEA features: 'MFR', 'NBR', 'NBD', 'PSIB', '#FBs', 'CVIBI', 'mean CC', 'sd CC', 'mean ISI CC', 'sd ISI CC', 'ISI dist', 'mean ISI', 'sd ISI temp', 'sd isi elec', 'MAC'
- The "TrainedNDE": the neural density estimator trained using the 300,000 simulations in the mackelab sbi package. This can be loaded in python to evaluate and obtain the posterior distribution as described in 'FindPosterior.py".
- The "Simulator.py" contains the functions to perform simulations with the biophysical computational model of hiPSC-derived neuronal networks on MEA. The "elecranges.dat" is necessary for this.
- "FeatureExtraction.py" has the function to extract the MEA features from spike trains of either either simulations or experimental data. It also has a function to compute spike rates per electrode that can be used as input to an embedding network.
- "MakeFigures.py": contains some functions to construct images as used in functions below
- "FindPosterior.py": contains the code to construct images like in the paper, and to asses your own experimental data using the method described in the paper. Also three example experimental observations are provided example_observations: "SCN_WTC_2410", "SCN_DS_2410" and "SCN_GEFS_2410" to construct some images and try the code if your own data is not available.
- "TrainNDE.py" shows how you can train an NDE like done for the paper, on provided data or own data. NDE (Masked Autoregressive Flow) can be trained with MEA features, or can be trained jointly with an 2D convolutional embedding network that takes firing rates per electrode as input (doesn't perform so well).
- "PosteriorPredictiveChecks.py": Is used to perform posterior predictive checks, and calculate parameter recovery error of trained NDEs. Can be done on "TrainedNDE" or on the 2 trained NDEs in the "Simulations_Training_embedding_network" directory where there are also NDEs trained with embedding network. Can also be applied to own trained NDE.
- "DetModelMisspecification.py": can detect model misspecification based on a method proposed by Schmitt et al. 2024, this requires "AllExperiments_MEAfeatures.csv" and "Simulations_MEAfeatures.csv" for our model simulations and experimental data used in the manuscript.
All experimental data can be requested from the original authors:
- Dravet Syndrome and GEFS+ data: https://academic.oup.com/brain/article/146/12/5153/7226643
