i-do-ai/fci-neuron-simulations

FCI neuron simulations

The simulated input-output data of the 24 detailed compartmental neuron models compared in

Dendritic morphology and synaptic nonlinearities enhance functional complexity in human cortical neurons Ido Aizenbud, Daniela Yoeli, David Beniaguev, Christiaan P. J. de Kock, Michael London, Idan Segev. PNAS 123(28), e2533168123 (2026).

Code, neuron models and instructions: https://github.com/ido4848/FCI. The trained networks that the paper's Functional Complexity Index (FCI) values come from are in the companion model repository i-do-ai/fci-neuron-tcns.

Every model was simulated with NEURON under in-vivo-like random synaptic input (Poisson trains on ~1000 excitatory and ~1000 inhibitory "super synapses" spread over the dendrites, at input rates calibrated per model so that the cell fires at ~1 Hz), and its somatic voltage and output spikes were recorded. A deep network (TCN) trained on these simulations to predict the output spikes from the input spikes gives an AUC, and the FCI is a fixed function of that AUC.

240,000 training and 60,000 test simulations of 10.5 s each (200.5 GB in total, 1-23 GB per model). 12 human models (human synaptic parameters) and 12 rat models (rat synaptic parameters), pyramidal cells of layers 2-6, morphologies and biophysics from Eyal et al., the Allen Cell Types Database, Hay et al. and the Blue Brain Project.

Models

Model	Species	Layer	Source	Input rate exc / inh (Hz)	FCI (Fig. 2A)	Train	Test
Rat_L2_TPC_BBP_Mandge_diams_fixed_passive_dends_simple_soma	rat	L2	bbp	16.9 / 14.4	0.1877	3.9 GB	1.0 GB
Rat_L23_PC_cADpyr229_1_BBP_passive_dends_simple_soma	rat	L2/3	bbp	18.2 / 15.7	0.2260	4.1 GB	1.0 GB
Rat_L23_PC_cADpyr229_5_BBP_passive_dends_simple_soma	rat	L2/3	bbp	20.2 / 16.7	0.2156	4.4 GB	1.1 GB
Rat_L4_TPC_BBP_Mandge_passive_dends_simple_soma	rat	L4	bbp	15.0 / 12.5	0.2032	3.6 GB	0.9 GB
Rat_L4_PC_cADpyr230_1_BBP_passive_dends_simple_soma	rat	L4	bbp	28.1 / 23.6	0.2362	5.5 GB	1.4 GB
Rat_L4_PC_cADpyr230_2_BBP_passive_dends_simple_soma	rat	L4	bbp	19.0 / 16.1	0.2452	4.2 GB	1.0 GB
Rat_L5_TPC_BBP_Mandge_passive_dends_simple_soma	rat	L5	bbp	47.1 / 35.2	0.2510	7.8 GB	1.9 GB
Rat_L5_TTPC1_cADpyr232_1_BBP_diams_fixed_passive_dends_simple_soma	rat	L5	bbp	54.9 / 43.7	0.2509	9.0 GB	2.2 GB
Rat_L5b_PC_2_Hay_passive_dends_simple_soma	rat	L5b	hay	39.8 / 33.0	0.2342	7.1 GB	1.8 GB
Rat_L6_UPC_BBP_Mandge_passive_dends_simple_soma	rat	L6	bbp	10.1 / 9.0	0.2399	2.8 GB	0.7 GB
Rat_L6_TPC_BBP_Mandge_passive_dends_simple_soma	rat	L6	bbp	15.5 / 13.6	0.2052	3.7 GB	0.9 GB
Rat_L6_IPC_BBP_Mandge_diams_fixed_passive_dends_simple_soma	rat	L6	bbp	10.8 / 8.8	0.1972	2.9 GB	0.7 GB
Human_L23_PC_0603_11_937_Eyal_passive_dends_simple_soma	human	L2/3	eyal	50.5 / 102.7	0.4294	12.6 GB	3.2 GB
Human_L23_PC_1303_03_448_Eyal_passive_dends_simple_soma	human	L2/3	eyal	49.8 / 102.4	0.4165	12.7 GB	3.2 GB
Human_L3_PC_0_BBP_passive_dends_simple_soma	human	L3	bbp	31.6 / 67.3	0.4190	8.8 GB	2.2 GB
Human_L4_PC_BBP_Mandge_diams_fixed_passive_dends_simple_soma	human	L4	bbp	21.9 / 53.3	0.3957	7.2 GB	1.8 GB
Human_L4_PC_539661667_Allen_passive_dends_simple_soma	human	L4	allen	28.0 / 65.2	0.3626	8.5 GB	2.1 GB
Human_L4_PC_569818704_Allen_passive_dends_simple_soma	human	L4	allen	15.2 / 36.2	0.3757	5.4 GB	1.4 GB
Human_L5_PC_BBP_Mandge_passive_dends_simple_soma	human	L5	bbp	27.3 / 59.6	0.3618	8.1 GB	2.0 GB
Human_L5_PC_0_BBP_passive_dends_simple_soma	human	L5	bbp	68.8 / 150.0	0.3672	16.9 GB	4.2 GB
Human_L5_PC_790872626_Allen_passive_dends_simple_soma	human	L5	allen	23.6 / 52.9	0.3934	7.3 GB	1.8 GB
Human_L6_PC_558211203_Allen_passive_dends_simple_soma	human	L6	allen	14.9 / 35.3	0.4004	5.3 GB	1.3 GB
Human_L6_PC_548494556_Allen_passive_dends_simple_soma	human	L6	allen	12.5 / 29.1	0.3146	4.7 GB	1.2 GB
Human_L6_PC_528614014_Allen_passive_dends_simple_soma	human	L6	allen	9.5 / 21.4	0.3274	3.8 GB	1.0 GB

Input rates are the mean over the simulations of the average firing rate per super synapse. The FCI column is the value printed above each neuron in Fig. 2A of the paper (see the model repository for how it follows from the three trained networks). models.json holds these numbers, the sizes and the AUCs per seed in machine readable form.

Layout

models/<model>/dataset_summary.pkl        summary of the whole dataset (calibrated input rates per simulation, ...)
models/<model>/train/summary.pkl          summary of the split, read by the trainer
models/<model>/train/shard_000.h5 ...     500 simulations per shard, 20 shards
models/<model>/test/summary.pkl
models/<model>/test/shard_000.h5 ...      5 shards
models.json

<model> is the model's folder name under simulating_neurons/neuron_models/ in the code repository, e.g. Rat_L5b_PC_2_Hay_passive_dends_simple_soma.

Format of a shard

HDF5, one group per simulation:

Dataset / attribute	dtype	shape	Content
sim_<n>/exc_indptr	uint32	(segments+1,)	CSR row pointer of the excitatory input matrix (segments × ms)
sim_<n>/exc_indices_delta	uint16	(nnz,)	column (ms) of every nonzero, delta coded within a row: the first nonzero of a row stores its column, every later one the distance from the previous nonzero
sim_<n>/exc_data	uint8	(nnz,)	number of synaptic activations of that segment in that ms
sim_<n>/inh_indptr, inh_indices_delta, inh_data			the same for the inhibitory input
sim_<n>/somatic_voltage	float32	(ms,)	somatic membrane potential, mV, 1 kHz
sim_<n>/output_spike_times	int64	(spikes,)	ms
sim_<n>.attrs			exc_shape, inh_shape, output firing rate, input rates, average voltage, ... and args_json (the simulation's arguments)
file attrs			model, split, shard, simulation_indices, format_version

The first 500 ms of every simulation are initialization and are skipped by the trainer. The delta coding is lossless; the voltage was recorded in float64 and stored in float32.

import h5py, numpy as np
from scipy import sparse

def read_csr(g, tag):
    indptr = g[f'{tag}_indptr'][:].astype(np.int64)
    delta = g[f'{tag}_indices_delta'][:].astype(np.int64)
    data = g[f'{tag}_data'][:]
    cumulative = np.cumsum(delta)
    starts = indptr[:-1]
    base = np.where(starts > 0, cumulative[np.maximum(starts - 1, 0)], 0)
    indices = cumulative - np.repeat(base, np.diff(indptr))
    return sparse.csr_matrix((data.astype(np.float64), indices, indptr), shape=tuple(g.attrs[f'{tag}_shape']))

with h5py.File('models/Rat_L5b_PC_2_Hay_passive_dends_simple_soma/test/shard_000.h5') as h:
    g = h['sim_0']
    exc = read_csr(g, 'exc').toarray()          # (1041, 10500)
    inh = read_csr(g, 'inh').toarray()
    v = g['somatic_voltage'][:]
    spikes = g['output_spike_times'][:]

utils/simulation_shards.py in the code repository has this reader (and the writer), and training_nets/train_neuron_tcn.py trains on a downloaded model folder directly:

pip install huggingface_hub
python data_release/download.py --model Rat_L5b_PC_2_Hay_passive_dends_simple_soma --splits train test
python training_nets/train_neuron_tcn.py --simulation_dataset_folder paper_data/simulations/models/Rat_L5b_PC_2_Hay_passive_dends_simple_soma ...

Citation

@article{aizenbud2026fci,
    title   = {Dendritic morphology and synaptic nonlinearities enhance functional complexity in human cortical neurons},
    author  = {Aizenbud, Ido and Yoeli, Daniela and Beniaguev, David and de Kock, Christiaan P. J. and London, Michael and Segev, Idan},
    journal = {Proceedings of the National Academy of Sciences},
    volume  = {123},
    number  = {28},
    pages   = {e2533168123},
    year    = {2026},
    doi     = {10.1073/pnas.2533168123}
}