Pretrained Models and Feature Datasets for ncpi

This dataset contains derived electrophysiological features, trained machine-learning inverse models, test-set predictions, and evaluation metrics generated from large-scale neural circuit simulations.

The resources were produced using the ncpi Python package to support neural circuit parameter inference from simulated electrophysiological signals.

The original raw simulation signals are not included in this deposit. Instead, this record contains derived feature matrices, reusable trained inference models, test-set outputs, and model-evaluation summaries.

Important note

The included features and machine-learning models have not yet been experimentally tested or independently validated. They are provided as research resources for reproducibility, benchmarking, and further methodological development.

Simulation datasets

Results are provided for three simulation datasets.

Hagen_v1

Current-dipole-moment signals generated using the first version of the Hagen neural circuit model. This model uses current-based synapses.

The Hagen_v1 dataset comprises approximately 2 million simulations generated by sampling neural circuit parameters within biologically plausible ranges. These ranges were constrained a priori according to biological principles. Hagen_v1 is processed in streaming batch mode using six simulation batches.

Model description: https://doi.org/10.1371/journal.pcbi.1010353

Hagen_v2

Current-dipole-moment signals generated using the second version of the Hagen neural circuit model. As in Hagen_v1, this model uses current-based synapses.

The Hagen_v2 dataset is a more constrained simulation set, comprising approximately 100,000 simulations. In addition to tighter parameter-range constraints, simulations were further filtered post hoc based on firing rates, spike-based metrics, and power-spectrum properties. Hagen_v2 is loaded by concatenating the first-search and second-search simulation sets.

Model description: https://doi.org/10.1371/journal.pcbi.1010353

Cavallari

Proxy electrophysiological signals generated using the Cavallari neural circuit model. In contrast to the Hagen datasets, the Cavallari model uses conductance-based synapses.

Model description: https://doi.org/10.3389/fncir.2014.00012

All signals were processed assuming a sampling frequency of 1600 Hz.

Inferred neural circuit parameters

Separate inverse models were trained for the Hagen and Cavallari simulation datasets. All inverse models estimate two target variables:

• E_I

• J_ext

For the Cavallari dataset, J_ext corresponds to the external-input scaling parameter ext_input_scale.

Hagen model parameters

Each Hagen simulation directly varies seven neural circuit parameters:

The four recurrent coupling strengths are combined into the inferred excitatory-to-inhibitory target:

E_I = (J_EE / J_EI) / (J_IE / J_II)

Therefore, each Hagen inverse model estimates:

• E_I

• J_ext

The individual recurrent coupling strengths and synaptic time constants are varied during simulation but are not estimated separately by the inverse models.

Cavallari model parameters

Each Cavallari simulation modifies seven parameters relative to the default Cavallari configuration available in the ncpi repository:

The default AMPA decay constants are 2.0 ms for excitatory neurons and 1.0 ms for inhibitory neurons. The default GABA-A decay constant is 5.0 ms for both populations. The default values of the thalamic external-input conductances are 0.234 and 0.317, respectively. Cortico-cortical external inputs are not modified.

The four recurrent conductances are combined into the inferred excitatory-to-inhibitory target:

E_I = (g_EE / g_EI) / (g_IE / g_II)

Therefore, each Cavallari inverse model estimates:

• E_I

• J_ext, corresponding to ext_input_scale

The individual recurrent conductances, synaptic time-constant scaling factors, and external-input components are varied during simulation but are not estimated separately by the inverse models.

Feature configurations

Three feature configurations are included:

Spectral parameterization was performed over 5–200 Hz, with a goodness-of-fit threshold of R² ≥ 0.9. Invalid or non-finite feature rows were excluded before model training.

Trained inverse models

Four inverse-model types were trained for each dataset and feature configuration:

This produces up to 36 trained dataset–feature–model combinations:

3 datasets × 3 feature configurations × 4 model types

Models were trained using a reproducible random 85% training / 15% testing split with random seed 0.

For ridge regression, multilayer perceptron regression, and random forest regression, test-set predictions and mean squared error metrics were computed on the held-out test set.

For neural posterior estimation, the posterior estimator and related inference assets were trained and saved, but held-out test-set prediction and evaluation were skipped by design to avoid expensive posterior sampling.

Included files

The archive contains two main directories:

new_features
new_models

new_features

This directory includes:

• Computed feature matrices in NumPy format

• Valid-row masks

• Reproducible training and testing indices

• Cached feature matrices generated by the training workflow

• Precomputed per-batch Hagen_v1 feature and target files produced in streaming mode

new_models

This directory includes:

• Serialized trained models and feature scalers

• NPE inference, density-estimator, and posterior assets

• Model configuration files

• Test target arrays, y_test.npy, for evaluated scikit-learn models

• Model predictions, y_pred.npy, for evaluated scikit-learn models

• Per-parameter and mean squared error metrics, metrics.json

• NPE metrics files indicating that test evaluation was skipped

• Dataset-level result summaries, summary.json

Software

The processing and training workflow was implemented using ncpi and the script:

run_massive_training_v3.py

The ncpi source code is available at:

https://github.com/necolab-ugr/ncpi

The script used to generate these results is also included in this deposit.

Recommended use

This dataset is intended for:

• Reproducibility of the ncpi inverse-model training workflow

• Benchmarking feature-based and simulation-based inference methods

• Reuse of trained inverse models for methodological development

• Evaluation of derived electrophysiological feature representations

• Further validation against experimental EEG, MEG, or related electrophysiological datasets

Limitations

The raw simulation signals are not included in this record. The deposit contains derived features, trained models, predictions, metrics, and workflow assets.

The included inverse models should be considered research tools. They have not yet been experimentally validated and should not be interpreted as clinically or biologically validated estimators without further independent testing.