Optimizing ODE-derived Synthetic Data for Transfer Learning in Dynamical Biological Systems

This code repository provides the necessary tools to reproduce the findings from Zabbarov & Witzke et al., 2024, on optimizing synthetic dataset characteristics for a simulation-based transfer learning approach to predicting dynamical biological systems.

Installation

To run the code, please set up a new environment with Python 3.10 or later and install the required dependencies, i.e. using pip:

pip install -r dev_requirements.txt

Please call each script from the root directory of this repository.

Structure of Repository

• Datasets: The datasets for the rotifers-algae, lynx-hares and COVID-19 experiments are provided in /src/data. Here, we also provide the scripts for generating synthetic datasets from the calibrated ODE models. For further details, see the README.mds in the respective folders.
• Transfer Learning and Deep Learning Experiments: The code to run and evaluate the transfer learning and deep learning experiments for each biological system is contained in /src/experiments/julian. The folder has the following structure:
  • algae-rotifers (coherent) w/ small DL model architecture: /algae-rotifers/run-1
  • algae-rotifers (coherent) w/ large DL model architecture: /algae-rotifers/run-4
  • algae-rotifers (incoherent) w/ small DL model architecture: /algae-rotifers/run-5
  • algae-rotifers (incoherent) w/ large DL model architecture: /algae-rotifers/run-8
  • COVID-19 w/ small DL model architecture: /algae-rotifers/run-1
  • COVID-19 w/ large DL model architecture: /algae-rotifers/run-6
  • lynx-hares w/ small DL model architecture: /algae-rotifers/run-10
  • lynx-hares w/ large DL model architecture: /algae-rotifers/run-18
  • evaluation.ipynb notebook provides the code used to evaluate the experiments. The results from the experiments are loaded from the /results folder.
  • You can use the preprocessing_results... notebooks to preprocess the results after running the experiments again.
• DL models: The small and large architecture DL models are defined as plugins for SimbaML in /src/models.
• Custom Metrics: You can define custom evaluation metrics in the /src/metrics folder.
• ODE calibration: The code used to calibrate the ODEs for each each biological system is contained in /src/experiments/simon.
• Plotting Function: The scripts used to generate the plots in the paper are provided in /plotting.

Walkthrough

In the following, we provide details on how to run the scripts in this repo. Call each script from the root directory of this repository.

1. Generation of ODE-derived Synthetic Datasets

For each of the experiments, we provide a Python script to generate synthetic datasets from the respective ODE models. You can start the data generation process by calling the following files in the /data-folder:

Data Generation with SIR model for COVID-19 experiments:

python src/data/covid/sir.py

Data Generation with SAR model for rotifers-algae experiments:

python src/data/algae-rotifers/rosenbaum.py

Data Generation with LV model for lynx-hares experiments:

python src/data/lynx-hares/lotka_volterra.py

Details on the sampling intervals for initial conditions and kinetic parameters used to generate the data are also listed in the mentioned scripts.

2. Time Series Prediction

2.1 Setup Experiments

To set up the time series prediction experiments, call the setup_experiment.sh scripts in each of the experiments/julian folders. In the following, we follow as an example the experiments on the COVID-19 experiments using the small DL model architectures.

src/experiments/julian/covid/run-1/setup_experiment.sh

Transfer Learning

Running the shell script will generate the required toml-files for configuring the transfer learning pipeline of SimbaML. You will find a folder and toml-file for each synthetic dataset (listed in src/data/<experiment>/synthetic) in a runs_transfer_learning-folder in the same location as the setup_experiment.sh. You can adjust the time series prediction in the generate_tomls_transfer_learning.pyfile. Additionally, a shell script (run_multivariate.sh) will be generated with which you can schedule the transfer learning run for each synthetic dataset at once.

DL Baseline

Running the setup_experiment.sh will also generate toml-files to start the DL baseline. You will find five experiment-folder (one for each seed) in the /runs_baseline-folder in the same location as the setup_experiment.sh.

2.2 Run Experiments

Before running experiments, add the absolute path of the src-folder of this repository to PYTHONPATH:

export PYTHONPATH="{PYTHONPATH}:<path_to_repository>/opt-synthdata-4tl/src"

You can start the transfer learning and DL experiments like so:

sh src/experiments/julian/covid/run-1/run_baseline.sh
sh src/experiments/julian/covid/run-1/run_multivariate.sh

You find the prediction results for the deep learning and transfer learning experiments in the /runs_baseline and run_transfer_learning folder, respectively. Note that you might have to adjust PyTorch Lightning's internal num_worker depending on the architecture on which the experiments are executed to speed up the training process.

Citation

If you use our pipeline from this repository and its code, please cite our preprint Zabbarov & Witzke et al., 2024:

Julian Zabbarov, Simon Witzke, Maximilian Kleissl, Pascal Iversen, Bernhard Y. Renard, Katharina Baum Optimizing ODE-derived Synthetic Data for Transfer Learning in Dynamical Biological Systems bioRxiv 2024.03.25.586390; doi: https://doi.org/10.1101/2024.03.25.586390