CellCanon: A Verifiable World-State Agent for Single-Cell Prediction and Reasoning

Single-cell perturbations are condition-wide, yet language agents can fragment one condition across thousands of gene-wise calls. We introduce CellCanon, the first single-cell perturbation agent to unify transcriptome-wide prediction, evidence-grounded explanation, and selective action through a shared, typed Cellular World State. A learned Cellular World Model writes the quantitative field once per drug–cell condition; specialized roles use the shared state to connect evidence, explanations, decisions, and observations without repeated per-gene reasoning. On 55,598 held-out drug-disjoint SciPlex3 rows, CellCanon raises DE Macro-F1 from 0.550 to 0.709 and conditional-DIR Macro-F1 from 0.524 to 0.793 over VCWorld, attains DE AUPRC 0.776 versus 0.406 for the strongest learned predictor, and cuts full-run language-model calls 230.7-fold and tokens 513.6-fold. Response-blind tests recover the exact top-3 mechanism in 11/12 cases versus 0.375 at random; typed verification converts 0/32 packed outputs into 32/32 executable writes; and reliability gating lowers harmful optional actions from 0.779 to 0.239 while raising positive yield from 0.221 to 0.325. CellCanon thus couples high predictive accuracy and low language use with inspectable explanation and selective action in one shared state.