neurofe 0.1.0

Fit neural mass models to empirical TMS-EEG data with gradient-based optimization.

neurofe

Fit neural mass models to empirical TMS-EEG data using gradient-based optimization.

Based on: Momi et al. (2022) Stimulation mapping and whole-brain modeling reveal gradients of excitability and recurrence in cortical networks.

Features

• Jansen-Rit neural mass model — full PyTorch implementation with conduction delays, structural connectivity, and leadfield projection
• Gradient-based fitting — fit model parameters to empirical TMS-EEG data using Adam optimizer with Bayesian priors
• Virtual dissection — lesion brain regions and observe the effect on simulated EEG
• Wong-Wang-Deco model — parameter framework defined (forward model contributions welcome)
• Standalone NumPy simulator — lightweight Jansen-Rit simulator for parameter exploration

Installation

pip install neurofe

From source:

git clone https://github.com/neurofe/neurofe.git
cd neurofe
pip install -e .

With optional dependencies:

pip install -e ".[examples]"   # matplotlib, pandas for running examples
pip install -e ".[viz]"        # mne, nibabel, etc. for brain surface visualization
pip install -e ".[dev]"        # pytest for development

Quick start

import numpy as np
import neurofe

# Load your data as numpy arrays
sc = np.load("sc.npy")           # (n_regions, n_regions) structural connectivity
lm = np.load("leadfield.npy")   # (n_channels, n_regions) leadfield matrix
dist = np.load("dist.npy")      # (n_regions, n_regions) distance matrix
eeg = np.load("eeg.npy")        # (n_channels, n_timepoints) empirical EEG

# Create and fit the model
model = neurofe.JansenRit(sc=sc, leadfield=lm, dist=dist)
result = model.fit(eeg, epochs=120)
result = model.predict(eeg)

# Access results
result.eeg_test    # simulated EEG
result.E_test      # excitatory activity
result.I_test      # inhibitory activity
result.loss        # training loss history

Virtual dissection

Lesion specific brain regions and observe the effect on simulated EEG:

# Lesion nodes 5 and 10 — returns a new model, original is unchanged
lesioned_model = model.lesion(nodes=[5, 10])
result_lesioned = lesioned_model.predict(eeg)

Documentation

• Getting Started — installation, input data, and quick start
• Models — model API, parameters, and loss functions
• Virtual Dissection — lesion API, usage patterns, and interpretation
• Examples — walkthrough of all example scripts

Examples

See the examples/ directory:

• example_synthetic_data.py — full pipeline with randomly generated data (no real data needed)
• example_train_jr.py — training on empirical TMS-EEG data
• example_param_exploration.py — parameter sweeps with the standalone NumPy simulator
• example_goodness_of_fit.py — evaluating model fit quality

Citation

If you use neurofe in your research, please cite:

@article{momi2022stimulation,
  title={Stimulation mapping and whole-brain modeling reveal gradients of excitability and recurrence in cortical networks},
  author={Momi, Davide and others},
  journal={Nature Communications},
  year={2022}
}

License

MIT. See LICENSE.