💡 About

This repository contains the code that can be used to reproduce the results presented in the paper 'Modelling chemical reaction networks using neural ordinary differential equations'.

📦 Contents

src/
    data/                   - Vector field implementations, preprocessing
    models/                 - (neural) ODE implementations, neural networks, model saving & loading utilities
    utils/                  - Path utilities, facade patterns
    visualization/          - Visualization functionalities
        
    
train_files/            
    oscillations/           - Training scripts for experimental oscillating data
    single_pulse/           - Training scripts for experimental single-pulse data
    synthetic/              - Training scripts for synthetic datasets
    
trained_models/
    oscillations/           - Saved models for experimental oscillating data.
    single_pulse/           - Saved models for experimental single-pulse data.

experimental_paper_plots.py - Script that can be used to generate the experimental figures used in the thesis.
plot_data.py                - Script that can be used to generate the appendix figures for the data points. 
training_with_synthetic_data.ipynb  - An example of generating a synthetic dataset and training a nODE. 

*.csv, *.txt                - Data file, retrieved from https://www.nature.com/articles/s41586-023-06310-2

⚙️ Installation

To run the project, install the dependencies listed in requirements.txt. For example, using Anaconda,

(base) C:\path\to\ChemicalReactionNetworks> conda activate your_condaenv
(your_condaenv) C:\path\to\ChemicalReactionNetworks> pip install -r requirements.txt

or venv

(base) C:\path\to\ChemicalReactionNetworks> source your_virtualenv/bin/activavate
(your_virtualenv) C:\path\to\ChemicalReactionNetworks> pip install -r requirements.txt

🐎 Getting Started

For more details on how to get started, please find the file getting_started.md.

🚧 Training neural ODEs on your own dataset

If you want to train a neural ODE on your own dataset, you can use the train_loop(...) functionality that is provided in scr/models/train_loop.py. In case you would like to combine a neural network with a theoretical vector field, you can initialize the theoretical vector field by implementing the ReactionWrapper interface (src/data/ReactionWrapper). Have a look at the already implemented vector fields (src/data/vector_fields/) for an example.

Create the neural ODE with the UDE implementation (scr/models/ODEs.py) where you can provide your own ReactionWrapper and the neural network specifications.

model = UDE(width, depth, reaction, network, ...)

An example of this is also provided in the training_with_synthetic_data.ipynb notebook, and in the getting_started.md file in the train_files/synthetic/ folder.

♻️ Reproducibility

Currently, the repository is aimed at making the results and figures presented in the thesis reproducible. The results found on the synthetic datasets can be reproduced by running the scripts train_files/synthetic/. These scripts train the models and plot the results. As it is time-consuming to train the models on the experimental data, trained models have been provided in the trained_models/ folder. One can reproduce the experimental figures and results by running:

(your virtualenv) C:\path\to\ChemicalReactionNetworks> python experimental_paper_plots.py

In case you would like to retrain the models, you can run the training scripts that are included in train_files/.