SBI Framework for Mechanistic Modeling of Cortical Dysfunction in Dementia

This repository contains the code associated with the paper:

Cárdenas Sánchez, M. et al. (2026)
A Framework Integrating Spiking Cortical Circuit Modeling and Simulation-Based Inference to Probe Biomarkers of Cortical Dysfunction in Alzheimer’s Disease
Interdisciplinary Sciences: Computational Life Sciences (Springer Nature)
DOI: https://doi.org/10.1007/s12539-026-00817-8

---

Overview

This work introduces a computational framework that combines the causal interpretability of mechanistic modeling with the predictive power of simulation-based inference (SBI) to uncover cortical circuit dysfunction underlying dementia-related neurophysiological changes.

The framework integrates:

• Biophysically grounded spiking cortical circuit simulations
• Generation of current dipole moments (CDM) and EEG signals
• Feature extraction from simulated and empirical signals
• Training of neural posterior estimation (NPE) models
• Posterior inference on empirical EEG data
• Statistical analysis and figure generation

The repository contains the full pipeline used in the study, including benchmarking, model training, posterior inference, and generation of publication figures.

---

Repository Structure

For clarity and reproducibility, the code has been organized into three clearly separated blocks, corresponding to the main stages of the article:

SBI_mechanistic_modeling_AD/
├── sbibm/
│   ├── 1_benchmark_parallel.py
│   ├── 2_merge_results.py
│   ├── 3_plot_fig2.py
│   └── all_results/
│
├── training_models/
│   ├── 4_train_CDM_catch22.py
│   ├── 5_train_CDM_power.py
│   ├── 6_train_CDM_wavelets.py
│   ├── 7_plot_fig3.py
│   └── 11_train_EEG_catch22.py
│
└── empirical_dataset/
    ├── 8_create_empirical_dataset.py
    ├── 9_plot_fig4.py
    ├── 10_plot_fig5.py
    ├── 12_compute_posteriors_emp_dataset.py
    ├── 13_plot_fig6_y_fig7.py
    ├── Analysis.py
    └── tools.py

Each block represents a different conceptual stage of the framework.

---

What Each Block Does

1️. sbibm/ — SBI Benchmarking and Configuration Selection

This block performs benchmarking of different SBI configurations using synthetic tasks from SBIBM.

It includes:

• Parallel benchmarking runs
• Aggregation and merging of benchmark results
• Generation of the benchmarking figure (Figure 2 in the paper)

This stage is used to validate and select the SBI configuration employed in the rest of the study.

---

2️. training_models/ — Training on Simulated Data

This block contains scripts for training SBI models using features extracted from simulated CDM and EEG signals.

It includes:

• Training using different feature sets (e.g., catch22, power-based features, wavelets)
• Saving trained models, scalers, and inference objects
• Diagnostic and performance plots (Figure 3)

The models trained in this stage are later used for posterior inference on empirical EEG data.

---

Pre-trained SBI models availability

If you want to skip the SBI training stage, the trained models used in this work are publicly available on Zenodo and can be downloaded directly:

https://zenodo.org/records/20375410

The repository includes the trained posterior models (model.pkl), inference objects (inference.pkl), and corresponding scalers (scaler.pkl) for the different experiments described in the paper.

---

3️. empirical_dataset/ — Empirical EEG Inference and Final Figures

This block handles the empirical data pipeline:

• Creation of the empirical EEG feature dataset
• Posterior inference on real subjects using trained SBI models
• Statistical analysis of inferred mechanistic parameters
• Generation of final figures (Figures 4–7)

This stage connects the mechanistic modeling framework to real neurophysiological data from dementia cohorts.

---

Execution Order

To reproduce the full pipeline described in the paper:

1. Run the scripts in sbibm/ (benchmarking and model selection).
2. Run the scripts in training_models/ (train SBI models on simulated data).
3. Run the scripts in empirical_dataset/ (apply trained models to empirical EEG data and generate final figures).

The number at the beginning of each script (e.g., 1_, 4_, 8_, etc.) indicates the recommended execution order to follow the pipeline presented in the article.

---

Simulation data availability (Zenodo)

The simulation outputs used by the pipeline (e.g., feature matrices and parameter vectors θ) are available on Zenodo:

• Zenodo API record: https://zenodo.org/api/records/15351118

Depending on your usage, you may download the archive manually from Zenodo (recommended for most users) or programmatically via the Zenodo API.

---

Computing environment

The experiments in this repository were executed on Albaicín, the High-Performance Computing (HPC) cluster of the University of Granada (UGR, Spain).

• The scripts are written to be runnable locally, but are HPC-oriented:
  • large datasets (simulation banks, empirical EEG feature matrices)
  • long-running training and posterior sampling
  • output paths designed for scratch/parallel environments

Note: This repository does not include SLURM .sh submission files.

---

Citation

If you use this repository, please cite:

Cárdenas Sánchez, M. et al. (2026).
A Framework Integrating Spiking Cortical Circuit Modeling and Simulation-Based Inference to Probe Biomarkers of Cortical Dysfunction in Alzheimer’s Disease.
Interdisciplinary Sciences: Computational Life Sciences.
https://doi.org/10.1007/s12539-026-00817-8