Our paper appeared on the cover of Cell Systems, Volume 17, Issue 6
CAROT applies the conditional Monge Gap to single-cell RNA sequencing data of Chimeric Antigen Receptor T cells. It extends the Conditional Monge Gap, introduced in the Nature Machine Intelligence paper Conditional Monge Gap enables generalizable single-cell perturbation modelling, with CAR-specific dataloaders, embeddings, and trainers. The notebooks directory contains notebooks for generating the paper figures and additional analyses.
We use poetry as package manager and tested the code in Python 3.10.
pip install poetry # into your base env
git clone [email protected]:AI4SCR/carot.git
cd carot
poetry install -vIf the installation was successful, activate the environment interactively via poetry shell.
You can find example configs in tests/configs/ for the unconditional and conditional settings.
To train a conditional Monge model:
from pathlib import Path
from carot.datasets.conditional_loader import ConditionalDataModule
from carot.trainers.conditional_monge_trainer import ConditionalMongeTrainer
from cmonge.utils import load_config
config_path = Path("tests/configs/conditional_synthetic.yml")
config = load_config(config_path)
logger_path = Path(config.logger_path)
datamodule = ConditionalDataModule(config.data, config.condition)
trainer = ConditionalMongeTrainer(jobid=1, logger_path=logger_path, config=config.model, datamodule=datamodule)
trainer.train(datamodule)
trainer.evaluate(datamodule)If you find this work useful, please cite:
@article{driessen2026modeling,
title={Modeling chimeric antigen receptor response at the single-cell level with conditional optimal transport},
author={Driessen, Alice and Born, Jannis and Rueda, Roc{\'\i}o Castellanos and Reddy, Sai T and Rapsomaniki, Marianna},
journal={Cell Systems},
volume={17},
number={6},
year={2026},
pages={101591},
doi={10.1016/j.cels.2026.101591},
url={https://www.cell.com/cell-systems/fulltext/S2405-4712(26)00073-6}
}