Building virtual-cell dynamics
that evolve with biological measurements

As experimental measurements evolve, the models used to understand cells must evolve with them. CellCompass translates evolving measurements into executable virtual-cell dynamics, using an AI model-design system to reason, search, code, validate, and revise the model.

Get started Explore the science
  • 100DynBench scenarios
  • 4measurement constraints
  • 4real single-cell studies
  • ∞long-horizon research
In motion

From a living cell to a designed dynamics.

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.

a moving sketch · sound on for the score

Grounded in benchmark and real-data evidence

DynBench100 controlled tasks HCT116 pulse-labeled drug response LARRY clone-informed hematopoiesis CITE-seq RNA-to-protein dynamics CytoBridge dynamics engine

The premise

As measurements evolve, the model must evolve with them.

CellCompass translates new experimental evidence into executable virtual-cell dynamics: not a static analysis, but a model that changes as the biology being measured becomes richer.

Figure 1 system overview showing biological questions, evolving measurements, the CellCompass model-design loop, and reusable dynamics families.
Figure 1 · system overview Experimental evidence reshapes the dynamics base.
01

Evidence defines the target

Time, lineage, growth, and multi-omic measurements specify what the dynamics must preserve.

02

The model family changes

Transport, bridges, mass change, and regulatory coupling are selected or designed as evidence expands.

03

The agent executes it

CellCompass reasons, searches, codes, validates, and revises until the dynamics match the measurement.

The modeling principle

Every measurement changes what the model has to become.

CellCompass starts from the biological question, reads the measured structure, and converts it into an explicit dynamics constraint. As evidence accumulates, the model family evolves from transport to mass-changing, lineage-conditioned, stochastic, or multi-omic dynamics.
01
Time course

Velocity through state space

Sampled marginals constrain how populations move between observed time points.

transport
02
Clone record

Branching future fate

Lineage barcodes force the dynamics to preserve hidden fate divergence.

fate mass
03
Growth / death

Changing population mass

Proliferation and apoptosis reshape balanced transport into mass-changing dynamics.

unbalanced
04
Multi-omics

Coupled regulatory state

RNA and protein readouts require trajectories that carry molecular programs across modalities.

RNA → protein
What makes it different

An AI model-design system for evolving dynamics.

The scientific object is the evolving virtual-cell model. The agent is how CellCompass makes that evolution operational: reasoning, searching, coding, validating, and revising on top of the open CytoBridge dynamics engine.

Autonomous algorithm design

Proposes, implements, and tunes brand-new dynamical-modeling algorithms inside isolated, version-controlled workspaces.

Theory-grounded reasoning

Literature & theory-book RAG over optimal transport and flow matching informs every modeling decision with citations.

Open, extensible engine

New methods register as first-class models on the open CytoBridge engine — reusable for simulation and downstream analysis across datasets.

Benchmark-driven campaigns

Staged trial budgets with archived trials and automatic promote / reject against baseline gates.

Reproducible by design

A registry tracks every proposal, snapshot, decision, and locked release as the single source of truth.

Long-horizon autonomy

Guarded subprocesses, timeout & OOM protection, and durable checkpoints let a campaign run for days and recover itself — the agent never crashes.

The dynamics framework

Four forces of life, one transport equation.

CellCompass and the CytoBridge engine model cellular processes as a spatiotemporal dynamical generative model. Add forces, and the optimal-transport formulation grows with them.

RUOT continuity equation learned from snapshots
∂tρt + ∇·(ρt vt) = ½σ²Δρt + gt ρt
velocity field v growth rate g score / diffusion σ

A single neural field f(x, t) → (v, g, s) is trained so the population density ρt transports between observed time-points — simulation-free, via flow matching, or by solving the PDE directly.

Model family
Dynamical OT Unbalanced OT RUOT CRUFM CytoBridge
Spatiotemporal dynamical generative model: cells differentiate, grow, die, and interact across time, modeled as transport between an initial and target distribution.

The spatiotemporal dynamical generative model — velocity, growth, score, and interaction transport an initial distribution to later time-points.

DynBench benchmark

Evaluated on controlled dynamical modeling tasks.

DynBench turns simulated single-cell dynamical scenarios into hidden-ground-truth tasks. Agents receive time-course data and a task specification, then are scored on velocity, growth drivers, held-out distributions, fate probabilities, perturbation response, and GRN recovery.

100scenarios across easy, medium, hard, and adversarial settings
0.702top CellCompass total score on DynBench100
6dynamical objectives evaluated per agent
***selected20 repeat advantage after Holm correction
DynBench evaluates CellCompass across controlled dynamical modeling tasks, showing task design, difficulty levels, baseline ranking, task profiles, difficulty scaling, BRGU-SFM component performance, and repeat stability.
DynBench100 probes six biological objectives and increasing difficulty; CellCompass leads baseline agents and selects BRGU-SFM as a stronger overall trade-off.
Closed-loop algorithm design

A new dynamical algorithm, discovered end‑to‑end.

From a real biological gap to a locked, validated method — proposed, reviewed, implemented, and hardened through staged gates. Then the loop begins again.

01 · 07 Motivation

A real biological gap

A dataset where growth, stochasticity, or cell–cell interaction is poorly captured by existing dynamical models.

    Reviewer-gated. A proposal clears review before any code is written; baseline gates then promote or reject every stage automatically.
    Long-horizon. Isolated execution, durable checkpoints, and self-recovery let a campaign run for days — continuously debugging itself.
    Registry-first. Stages, gates, archived trials, and locked releases live in a registry — the source of truth, not the chat.
    Honest. Metrics use model-generated trajectories from t₀, never future observed targets — no metric hacks.
    Real campaign replay

    CellCompass searches with evidence, not a single prompt.

    These are three completed research campaigns. The animation replays the actual trial history: what biological problem was posed, how the model design changed, and where the metric moved before the final method was locked. The statistics report the full CellCompass search span; the animated curve zooms into the scored trials for the selected method.

    GSE305370 · CITE-seq time course

    RNA-only dynamics became a protein-state generator.

    CellCompass turned a CITE-seq time course into CITE-PCFM, a model that starts from RNA and generates time-resolved ADT surface-protein states.

    What changed during search
      Biological problem Input question and measurement constraints
      CITE-seq panel showing the biological request and CellCompass RNA-to-protein formulation.
      Output analysis Downstream biological result
      CITE-seq panel showing CellCompass-designed RNA-to-protein dynamics and downstream protein-fate programs.
      Optimization history

      Surface-protein distribution error

      trial 1 / 70 0.3233
      high low real-data trial order
      trial result best retained so far lower is better
      Cross-design search landscape

      CellCompass moved between algorithm basins before locking CITE-PCFM.

      The upper chart zooms into CITE-PCFM tuning. This panel replays the broader search across candidate algorithm families.

      Model reuse, not just invention

      Sometimes the right model already exists.

      Asked to build EMT dynamics for four cancer cell lines under three stimuli, CellCompass judged the existing VGFM velocity–growth formulation adequate — and spent its effort on fitting, evaluation and interpretation instead of inventing a new method.

      • 12 trajectoriesseparate velocity–growth dynamics for every cell-line × stimulus pair — 34,994 cells, no new algorithm required.
      • Context-gated routesthe same stimulus yields mesenchymal conversion, partial/hybrid EMT, or epithelial retention depending on the initial cell-line state.
      • Three-layer mechanismstimulus-proximal pathways, cell-state gating, and growth/mass coupling explain why the routes separate.
      Generated endpoint route atlas across twelve cell-line and stimulus conditions, showing mesenchymal conversion, partial or hybrid EMT, epithelial-retaining and mixed routes.
      Generated endpoint EMT-core shift vs. partial-EMT/hybrid shift across all twelve cell-line × stimulus conditions.
      How a run unfolds

      A guided path from raw data to reported biology.

      Intake & inspect

      Load an .h5ad, sanity-check AnnData, and confirm the time & latent contracts.

      Preprocess

      Normalize, reduce dimensions, and build the canonical temporal backbone for training.

      Select theory

      Choose the transport formulation — balancing velocity, growth, stochasticity, and interaction.

      Train & evaluate

      Fit the dynamical model, preview the runtime call-chain, and measure against held-out marginals.

      Downstream analysis

      Velocity & score streams, fate branches, driver genes, landscapes, and terminal states.

      Report

      Compile logs, figures, and parameters into a polished HTML or Markdown report.

      Meet it where you work

      Terminal, TUI, or a full Web UI.

      One unified runtime behind every entry point. Start interactive, fire a one-shot turn, or open the timeline-replay Web UI — sessions resume seamlessly across all of them.

      • LangGraph checkpointing & event-based timeline restore
      • Provider-aware auth: API key, Gemini & Codex OAuth
      • Boundary-based context compaction for long runs
      # Reconstruct dynamics, then report — in one turn
$ cellcompass run data.h5ad \
    --question "Analyze the main fate branches and driver genes" \
    --output ./run_demo \
    --device cuda

✓ preprocessing · AnnData contracts verified
✓ theory selected · Regularized Unbalanced OT
✓ training · velocity + growth + score
✓ downstream · fate branches, driver genes
✓ report · ./run_demo/report.html
      # Converse with the agent, resume any time
$ cellcompass

CellCompass › Check the dataset first, then propose a workflow.
  ↳ /status   /model   /compact   /resume   /jobs

› Design a new growth-aware flow-matching algorithm
✓ proposal drafted · review & campaign started
✓ trial 3/8 promoted · beats baseline gate
      # Full Web UI with live timeline replay
$ cellcompass web --host 0.0.0.0 --port 8000

✓ runtime ready · FastAPI lifespan hooks installed
→ open http://localhost:8000

# timeline · proposal review · skill toggles · compaction
      ∞long-horizon campaigns — resumable, and self-recovering for days
      1registry of record for every proposal, gate & locked release
      0metric hacks — inference always starts from t₀
      The research behind it

      Built on a line of published theory.

      npj Syst. Biol. Appl. · 2025 Deciphering cell-fate trajectories using spatiotemporal single-cell transcriptomic data
      ICLR · 2025 · Oral Learning stochastic dynamics from snapshots through regularized unbalanced optimal transport
      NeurIPS · 2025 Modeling cell dynamics and interactions with the Unbalanced Mean-Field Schrödinger Bridge
      NeurIPS · 2025 Joint Velocity-Growth Flow Matching for single-cell dynamics modeling
      NeurIPS · 2025 Variational Regularized Unbalanced Optimal Transport: single network, least action
      Nature Methods stVCR: reconstructing spatio-temporal dynamics of cell development using optimal transport
      Core team

      The core team behind CellCompass.

      The initial CellCompass team brings together single-cell dynamics, mathematical modeling, and AI-assisted model design.

      Peking University
      University of California, Irvine
      Zhenyi Zhang

      Zhenyi Zhang

      Peking University

      Zihan Wang

      Zihan Wang

      Peking University

      Qing Nie

      Qing Nie

      University of California, Irvine

      Peijie Zhou

      Peijie Zhou

      Peking University

      Point it at your data.
      Get dynamics back.

      Install the agent and the CytoBridge engine in editable mode, then start your first run.

      install
      # 1 · install the agent + core engine
$ pip install -e . && pip install -r requirements-agent.txt
$ cd CytoBridge-main && pip install -e .

# 2 · start exploring
$ cellcompass "Analyze the available dataset and propose the next workflow"
      View CytoBridge on GitHub Read the docs