CellCanon: A Verifiable World-State Agent for Single-Cell Drug-Perturbation Prediction, Explanation, and Selective Action

Drug discovery requires agents that combine accurate perturbation-response prediction with interpretable evidence and reliable follow-up. We introduce CellCanon, a world-state agent that unifies prediction, explanation, and selective action through a shared, typed Cellular World State. A numerical world model constructs the transcriptome-wide predictive field once per drug–cell condition; specialized roles reuse this state to link claims to evidence, prioritize candidate mechanisms, verify revisions, and evaluate follow-up. Explicit provenance and write permissions make interpretations and state changes inspectable. On the sampled held-out-drug SciPlex3 track, CellCanon improves DE AUPRC by 91.3% over the strongest evaluated learned predictor, and DE and three-state Macro-F1 by 26.7% and 50.7%, respectively, over their strongest evaluated hard-decision baselines. Compared with a gene-wise reasoning baseline, it requires 513.6-fold fewer language-model tokens and 230.7-fold fewer logical generations. In retrospective cross-cell updating, reliability gating reduces the per-drug rate of performance-degrading updates by 69.3% relative to an always-execute policy. CellCanon makes strong prediction and low language cost compatible with evidence-linked explanation and controlled follow-up through a reusable scientific state.