Cellarium
Cellarium is a tool for asking questions of the whole-cell model of E.
Core trick
A council commits to a falsifiable prediction before seeing data, then an agent with read-only simulation tools tests it and cites the numbers it read.
Description
Cellarium is a tool for asking questions of the whole-cell model of E. coli from Macklin et al. (Science, 2020). That model tracks about 16,000 molecular species in a single cell, updated every second across a full cell cycle. It is very detailed, and checking whether it behaves correctly for a given perturbation normally takes an expert days.
Cellarium divides the work between two agents. A Socratic Council (a proposer, a skeptic, and a judge) reads the question and writes down a specific prediction, including the result that would count as the model being wrong. It does this without seeing any of the simulation data. A second agent, Cellwright, then tests the prediction. It has 38 tools for reading real simulation output and searching PubMed, it reports the numbers it actually read and cites its sources, and it cannot start a new simulation without a person approving it.
Three findings from real runs:
- We asked whether knocking out a tRNA synthetase raises ppGpp, which is what the textbook stringent response predicts. The Council committed to "2 to 4 times higher" in advance. The simulations showed ppGpp falling about 90 percent (t = −27.85), so the prediction failed. Writing it down first is what makes this a real disagreement instead of an explanation added afterward.
- Asked whether nitrate turns on the nitrate-reductase genes, the first answer was yes. Once we controlled for the fact that adding nitrate also removes oxygen, those genes looked like a general anaerobic response rather than a nitrate-specific one.
- Deleting ribosomal RNA operons lowered ribosome content and growth rate together. That is the opposite of what Scott's efficiency law describes. A literature search during the run found work on ribosome-limited antibiotic susceptibility (Greulich and Scott, 2015), a regime that has not been reproduced in a whole-cell model.
Why it matters: models like this are usually trusted or ignored as a whole. Cellarium exposes the prediction, the failure condition, and the source of each number, so a scientist can check the model against known biology and see where it disagrees. The same approach would apply to other large mechanistic simulations, and the runs we collect become a dataset others can reuse.
README excerpt
# Cellarium **A glass box over whole-cell reasoning.** Ask a question about *E. coli*; a **blind Socratic Council** frames it into a **falsifiable hypothesis — without seeing the data**; then a grounded agent, **Cellwright**, tests it against **real whole-cell simulations** and the **published literature**, and closes the loop by proposing experiments for your approval. Every number rides with its provenance. The agent never launches a run on its own. > Cellarium's own code is MIT; the whole-cell model it runs on is obtained separately under Stanford's > academic license (see [License](#license)). **Jump to:** [The problem](#the-problem) · [The two halves](#the-two-halves) · [Major results](#major-results) · [Install & run](#install--run-the-three-tiers) · [Architecture](#architecture) · [License](#license) --- ## The problem 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, and their output — a molecular movie of tens of thousands of species at every timestep — is effectively a black box. Two things are hard: **framing a question the model can actually answer**, and **trusting the answer** (grounding every number, and catching where the model is wrong). Today's "AI scientist" systems optimize for novelty and rarely ground what they claim