$\texttt{NOBLE}$: Neural Operator for Biological Learning and Electrophysiology

$\texttt{NOBLE}$ is a neural operator-based framework for modeling and predicting electrophysiological responses in biological neurons. The project leverages Fourier Neural Operators (FNO) to learn complex mappings between input stimuli and neuronal responses, enabling efficient simulation and analysis of neural dynamics. This repository accompanies our NeurIPS 2025 paper.

Installation

Prerequisites

• Python 3.10 or higher
• CUDA-compatible GPU (recommended)

Setup

1. Clone the noble repository:

git clone https://github.com/neuraloperator/noble.git
cd noble

2. Run the installation script to install the nerualoperator library (outside the noble directory) and the noble codebase:

bash install_noble.sh

Project Structure

noble/
├── src/                          # Main source code
│   └── training/                 # Training pipeline and models
│       ├── engine/               # Training engines
│       ├── data/                 # Data loading utilities
│       ├── neuro/                # Neuroscience-specific modules
│       └── configs/              # Configuration files
│
├── inference/                    # Example scripts and utilities
│   ├── ensemble_generation.py
│   ├── arbitrary_ensemble_generation.py
│   ├── compute_ephys_features.py
│   ├── compare_ephys_features_experiments.py
│   ├── generate_FI_curve.py
│   ├── noble_models/             # Pre-trained models for inference
│   └── utils/                    # Utility functions for inference scripts
│
├── data/                         # Data storage
│   └── e-features/               # Electrophysiological features
│
└── training_scripts/             # SLURM job submission scripts

Features

• Fourier Neural Operator (FNO) Architecture: Advanced neural operator for learning complex temporal dynamics
• Electrophysiological Feature Extraction: Automated computation of spike features using eFEL
• Multi-scale Modeling: Support for different temporal resolutions and downsampling factors
• Embedding Systems: Flexible embedding of amplitude, frequency, and electrophysiological features
• Biophysical Neuron Models: Integration with detailed biophysical simulations
• Training and Fine-tuning: Comprehensive training pipeline with configurable hyperparameters
• Visualization Tools: Built-in plotting and analysis utilities

Test Examples

We provide five different example scripts in the noble/inference/ directory to help you explore and test the trained $\texttt{NOBLE}$ model on:

• Generating ensemble predictions
• Computing electrophysiological features
• Generating frequency-current curves

cd noble/inference

Citation

If you find this work useful, please cite:

@inproceedings{ghafourpour2025noble,
  title     = {NOBLE: Neural Operator with Biologically-informed Latent Embeddings to Capture Experimental Variability in Biological Neuron Models},
  author    = {Ghafourpour, Luca and Duruisseaux, Valentin and Tolooshams, Bahareh and Wong, Philip H and Anastassiou, Costas A and Anandkumar, Anima},
  booktitle = {Advances in Neural Information Processing Systems},
  volume    = {39},
  year      = {2025},
  doi       = {arXiv:2506.04536}
}

Contact

For questions and support, please contact Luca Ghafourpour ([email protected]).