Recent techniques enable functional characterization of single-cells, which allows for the study of cellular and molecular mechanisms in complex biological processes, including cell development. Many methods have been developed to utilize single-cell datasets to reveal cell developmental trajectories, such as dimensionality reduction and pseduotime. However, these methods generally produce static data snapshots, challenging a deeper understanding of the mechanistic dynamics underlying cell development. To address this, we have developed CellTRIP, a multi-agent reinforcement learning model to recapitulate the dynamic progression and interaction of cells during development or progression. CellTRIP takes single or bulk cell data, single or multimodality, and trains a collaborative reinforcement learning model that governs the dynamic interactions between cells that drive development. In particular, it models single cells as individual agents that coordinate progression in a latent space by interacting with neighboring cells. The trained model can further prioritize and impute cellular features and in-silico predict the dependencies of cell development from feature perturbations (e.g., gene knockdown). We apply CellTRIP to both simulation and real-word single-cell multiomics datasets including brain development and spatial-omics, revealing potential novel mechanistic insights on gene expression and regulation in complex developmental processes.

CellTRIP confers several unique advantages over comparable methods:

• Take any number or type of modalities as input
• Variable cell input quantity,
• Interactive latent space,
• Great generalizability.

Application	Preview	Description	Performance
Multimodal integration		Policy trained on 300 simulated single-cells with multimodal input applied to an integration environment
Cross-modal imputation		Imputation of spatial data from gene expression, simulated until convergence using a CellTRIP trained model
Development simulation		Cell differentiation simulation on single-cell brain data, with CellTRIP agents controlling cell movement in the previously unseed environment
Trajectory recovery, inference, and prediction		Trajectory estimation on single-cell multimodal human brain data across several age groups
Perturbation analysis		Estimated effect size calculation of randomly selected genes from a CellTRIP imputation model on spatial data

To install CellTRIP, first clone and navigate to the repository,

git clone https://github.com/Oafish1/CellTRIP
cd CellTRIP

# CellTRIP may also be installed directly from GitHub without cloning, but does not have version controlled dependencies, and is therefore not recommended
pip install celltrip@git+https://[email protected]/Oafish1/CellTRIP

Create and activate a conda virtual environment using Python 3.10,

conda create -n celltrip python=3.10
conda activate celltrip

Install dependencies using pip

# Development install (Recommended)
# For full development capabilities, also install ffmpeg, poppler-utils, boto3, cupy, docker, and poppler-utils
pip install -r requirements.txt
pip install -e .

# Base install
pip install -e .

CellTRIP can be used either from the command line,

python train.py ...
# TODO

or as a python package,

# TODO

After training the model, analysis can be performed using the examples/analysis.ipynb notebook,

python analysis.py ...
# TODO

this script will generate...