A Claude-backed copilot for reasoning over a whole-cell E. coli simulation — grounded, and honest about its own limits.
Whole-cell models compute the dynamic, regulatory, single-cell behaviour of a living cell from first principles — the regime that steady-state flux-balance analysis and human intuition can't reach. But they are locked behind deep expertise, a heavy compute stack, and long run times. Cellarium is a small agentic workbench that lets a scientist ask a question in plain English and get back a grounded answer — every number tied to real simulation output, never fabricated — with two guardrails that make the answer trustworthy:
- Feasibility / validation-envelope check — refuses experiments the model was never built or validated to simulate (e.g. a mid-run carbon-source switch), and says why.
- Output QC — inspects each simulated generation and withholds any degenerate / non-viable result instead of laundering it into a clean-looking number.
Built for Built with Claude: Life Sciences (Builder track). This is a fresh, from-scratch implementation. It runs on the Covert-lab whole-cell E. coli model as an external backend, which you obtain separately under its Stanford Academic Software License (non-commercial; redistribution requires Stanford's written permission). Cellarium bundles none of that model — it is the new agent + guardrail + interface layer, and points at your own licensed checkout. See docs/DECISIONS.md D3 for the licensing/data-distribution constraints.
The differentiator is not "the AI answers the question" — a naive tool can print a number. It is that the agent imposes the scientific rigor (validated envelope, replication, grounding) that a mechanistic model demands, and catches the failure modes a scientist would otherwise trust. That rigor layer is a specialization of the reviewer-agent pattern, applied to systems-biology modelling — a domain a scientific AI workbench does not yet cover.
src/cellarium/
envelope.py validated-perturbation envelope check (feasibility guardrail)
qc.py per-generation output QC (integrity guardrail)
model.py thin backend adapter (cached + live hook to the public wcEcoli model)
tools.py grounded read tools exposed to the agent
agent.py Claude (Anthropic Messages API) tool-using loop
cli.py run a question end-to-end
ui/index.html Claude-Science-style demo interface
data/cache/ cached demo simulation results (reproducible, fast)
docs/DEMO.md the demo script
python -m venv .venv && . .venv/Scripts/activate # (or bin/activate on unix)
pip install -e .
cp .env.example .env # add your ANTHROPIC_API_KEY
python -m cellarium.cli "Do genetically identical E. coli cells behave differently, and why?"Open ui/index.html in a browser for the interface mockup.
Cellarium's users are hypothesis generators, not decision-makers: the model's predictions are hypotheses, and the tool prioritizes and explains experiments rather than certifying outcomes. It removes computational expertise, not scientific judgement. Organism: E. coli K-12 MG1655 (a lab strain).
Cellarium's own code is MIT — see LICENSE. The whole-cell model it depends on is not MIT: it is under Stanford's academic (non-commercial) license and is obtained/run separately by the user. Cellarium redistributes no model code or model-derived data. See docs/DECISIONS.md D3.