Prediction beyond observed states
Virtual cells must forecast cellular change under biological context and intervention.
Virtual cells must do more than represent observed states — they must predict what happens next. CellCompass determines when existing dynamics are sufficient and when evolving measurements require a new model to be designed, tested, and retained.
A hand-drawn sketch of the idea — a cell on the bench, an atlas streaming into the machine, and the design cycle that draws new dynamics as new questions arrive.
Grounded in benchmark and real-data evidence
The premise
New measurements reveal biology that existing dynamics cannot express. CellCompass turns those measurements into new models—and retains them for the next study.
Virtual cells must forecast cellular change under biological context and intervention.
Each new experiment can expose a relation that existing dynamics cannot express.
Scientific capacity should grow across studies instead of restarting from one-off solutions.
The modeling principle
Sampled marginals constrain how populations move between observed time points.
transportClone barcodes link early progenitors to their descendants and constrain future fate probabilities.
fate massChanges in observed abundance motivate models that allow population mass to vary over time.
unbalancedRNA and protein readouts require trajectories that carry molecular programs across modalities.
RNA → proteinTransport, bridge and flow-matching models share a modular structure. CellCompass changes the components needed to represent a new experimental constraint while reusing the remaining training and simulation procedures.
Which cells can be connected across observation times? RNA state, metabolic directions, clone barcodes or paired protein measurements can inform these connections.
How do connected cells change between observations? Paths can be deterministic or stochastic and can incorporate measured directions of molecular change.
What changes along the path? Models can carry effective population mass, represent fate branches or predict surface-protein states along RNA trajectories.
The engine includes balanced and unbalanced formulations, with flow-matching methods for learning velocity, growth and score-based diffusion fields. Spatial interactions and history-dependent processes are directions for future extension.
DynBench100 uses controlled ground truth to test dynamical reasoning across velocity, growth-driver recovery, held-out distributions, fate probabilities, perturbation response, and gene-regulatory recovery. CellCompass achieved the highest aggregate performance among the tested coding and biomedical agents.
CellCompass scored 0.712 with GPT-5.5 high and 0.667 with MiMo-v2.5-pro. In the repeated-trial comparison, GPT-5.5 high reached a mean score of 0.672.
Across all 40 medium-complexity scenarios, with two runs per condition, the mean six-task score was 0.736 with the engine and 0.672 without it. The complete system scored higher in 35 of 40 scenarios, with the largest gains in fate and regulatory recovery.
When an experiment reveals a relation that existing models cannot express, CellCompass proposes and implements a new formulation, tests it against the measurements and retains the validated method.
Experimental measurements reveal a biological relationship that existing dynamics do not represent.
Three experiments. Three new models. Follow the design choices and the measured results.
Terrain and positions illustrate design choices, not a measured loss surface.
Across four cancer cell lines and three stimuli, CellCompass reused velocity–growth dynamics to reconstruct context-dependent EMT routes.
Launch the web interface, configure a supported model provider and describe the question you want to investigate. CellCompass prepares the data, proposes a dynamical model and evaluates the generated trajectories.
For algorithm-design studies, enable Planner stop hook under Advanced settings → Review and stop hook before initializing the agent. Select LLM prompt evaluator to continue unfinished model development.
cellcompass --llm-model gpt-5.5 --llm-thinking-level high \
web --host 127.0.0.1 --port 8000 --device cpuOpen http://localhost:8000, configure the provider and click Initialize Agent.
My time-course scRNA-seq data are in /path/to/data.h5ad. How does this cell population change over time? Check the data and sampling times, then propose a dynamical model and explain how you would evaluate it.
The initial CellCompass team brings together single-cell dynamics, mathematical modeling, and AI-assisted model design.
Peking University
Peking University
University of California, Irvine
Peking University
Reuse an existing model when it is adequate, or begin a measurement-grounded design campaign.
Requirements: Conda, Node.js 22.12 or later, and a supported provider’s API key or Codex sign-in. Small examples can run on CPU. An NVIDIA GPU is recommended for biological model training.
git clone https://github.com/zhenyiizhang/CellCompass-review.git
cd CellCompass-review
conda create -n cellcompass python=3.10 -y
conda activate cellcompass
python -m pip install -c requirements-runtime-constraints.txt -r requirements-agent.txt -e ./CytoBridge-main -e .
npm --prefix web ci
npm --prefix web run build