This repo provides a collection of benchmark datasets used in a sequence of papers targeting multimarginal trajectory and SDE inference from snapshot data Shen et al. 2025, Berlinghieri et al. 2025.
There are in general three levels of realisticity: 1) fully synthetic, where we define a model and generate the data; 2) semi-synthetic, where we use real velocity fields but simulate particles; and 3) real, where we do not know the model and did not simulate the data.
We provide the data, the method to generate the data, citations, and license information. Please consider citing the original work where the data was initially collected, generated, and/or introduced.
The repository generally requires torchsde, and the data are in the shape [snapshots, batch, dimensions]. Consider changing the code if a different sampling frequency and/or sample size is desired. The data does have the same batch size for each snapshot and the snapshots are equally spaced, but this is not a required behavior.
This is a classic dynamical system describing predator-prey dynamics. To generate the data, follow
cd model
python make_data.py LV
Consider citing Shen et al. 2025. The data is licensed with this repository.
Data will be saved in asset as LV_data.npz. There are three keys: N_steps, the number of snapshots, and Xs, the data of size [snapshots, batch, dimensions].
The Repressilator system is a minimal biological clock. For more biological information, we recommend reading this tutorial. Note that this model is called a protein model in the tutorial because they assumed measurements were done with optics and thus measured protein. Usually, the mRNA-only model will look similar. To generate the data with the mRNA-only model:
cd model
python make_data.py Repressilatormrna
Consider citing Shen et al. 2025 and Elowitz and Leibler 2000. The data is licensed with this repository.
This model also includes protein in the reaction network, see this. One can derive the mRNA-only model by assuming the protein quickly reaches equilibrium.
To generate the data with the mRNA and protein model:
cd model
python make_data.py Repressilatormrnaprotein
Consider citing Berlinghieri et al. 2025 and Elowitz and Leibler 2000. The data is licensed with this repository.
The data is from high-resolution (1 km) bathymetry data from a HYbrid Coordinate Ocean Model (HYCOM) reanalysis Panagiotis 2014. This dataset was released by the US Department of Defense and is thus in the public domain. The dataset provides hourly ocean current velocity fields for the region extending from 98E to 77E in longitude and from 18N to 32N in latitude, covering every day since January 1st, 2001. We focus on a specific time point, June 1st, 2024, at 5 PM at surface level, and a particular spatial region where a vortex is observed.
To generate the data:
cd model
python make_data.py GoM
The code will read HYCOM data from asset and simulate particles.
Consider citing Panagiotis 2014 and Shen et al. 2025. HYCOM data is funded by the US federal government and thus in the public domain.
The data is a real single-cell RNA-sequencing dataset that tracks T cell–mediated immune activation in peripheral blood mononuclear cells (PBMCs) Jiang et al. 2024. They recorded gene-expression profiles every 30 minutes for 30 hours. We use the 41 snapshots collected between 0 h and 20 h—prior to the onset of steady state; we take 20 alternating snapshots (at integer hours) for training and the remaining 21 for validation.
We use a 30-dimensional projection ("gene program") of the original measurements.
The original program data is in asset/pbmc_timecourse.csv. The first column is cell ID, the second to third-from-last are cell programs, while the last two are time collected and cell type annotations. A slightly cleaned and subsampled version is given in asset/processed_pbmc_data_sub500_every_2_until20.npz.
Consider citing Jiang et al. 2024 and Berlinghieri et al. 2025. The data is under CC BY-NC 4.0 with Jiang et al. 2024.
We welcome contributions. Please start a PR and include a description of the data, original citation, and license.