fci-neuron-tcns

FCI neuron TCNs

The trained networks behind the Functional Complexity Index values reported 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: https://github.com/ido4848/FCI. Training data: i-do-ai/fci-neuron-simulations.

For each of the 24 neuron models: the three seeds of the depth-3, width-128 temporal convolutional network of the paper (neuron_tcn_d3_w128_fk54_lk12_sd0_5_bs180_lr0_01_wd1e_08_vlw0_02_t{0,1,2} internally), trained to predict the neuron's output spikes and somatic voltage from its synaptic input. FCI = log10(1000 (1 - AUC)) / log10(1000 (1 - 0.9)), from the AUC on the normalized test set (a subset of the test simulations whose mean output rate is exactly 1 Hz).

Layout

models/<model>/results.csv, final_results.pkl                        final AUCs of all 12 networks trained on the model (depths 1, 2, 3, 7 x 3 seeds)
models/<model>/d3_w128_t<seed>/args.pkl, final_results.pkl           the network's training arguments and final results
models/<model>/d3_w128_t<seed>/models/model_<epoch>_<steps>          the best checkpoint (torch.save)
models/<model>/d3_w128_t<seed>/models/model_*_light.pkl              every evaluation along training (best_save + eval_history, plain pickle)

<model> is the model's folder name under simulating_neurons/neuron_models/ in the code repository. A d3_w128_t<seed> folder has the layout train_neuron_tcn.py writes, so

python calculate_fci.py --neuron_tcn_folder models/<model>/d3_w128_t0 --use_normalized

prints the FCI of that seed's checkpoint. The checkpoint is a dict with model_state_dict, args, in_chans, best_save, eval_history, epoch, step, ...; load it with Tcn.get_tcn_from_file(path, 'cpu') from training_nets/train_neuron_tcn.py. Compared with the checkpoints written during training, the optimizer state was removed and the weights were cast from float64 to float32 (355 MB → 59 MB per network; the network runs in float64 and its outputs change by ~1e-7). One checkpoint (Rat_L6_IPC_BBP_Mandge_diams_fixed_passive_dends_simple_soma, seed 1) was pruned before the release and only its light checkpoints, i.e. its AUC, survive.

Which number is in the paper

Every network was evaluated on the normalized test set every ~50,000 steps and its model_*_light.pkl files keep that history (eval_history['normalized_test'][step]['auc']). The checkpoint shipped for a seed is best_save: the evaluation with the highest normalized AUC over the whole run. The paper (Fig. 2A) took, per seed, the highest normalized AUC among the evaluations made within the first 500,000 training steps, and printed above each neuron the minimum FCI over its three seeds - the "Fig. 2A" column below, which reproduces the paper's 24 values exactly. Where a run kept improving after step 500,000 the shipped checkpoint is a little better than the paper's number for that seed.

import pickle, glob, numpy as np
from calculate_fci import auc_to_fci
light = max(glob.glob('models/<model>/d3_w128_t0/models/model_*_light.pkl'), key=lambda f: int(f.split('_')[-3]))
history = pickle.load(open(light, 'rb'))['eval_history']['normalized_test']
paper_auc = max(v['auc'] for step, v in history.items() if step <= 500000)
print(auc_to_fci(paper_auc))

Networks

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}
}