CellCanon: An Agentic Cellular World Model for Prediction, Explanation, and Observation-Driven Updating

Zhaokang Liang, Shuyang Zhuang, Xiaoran Jiao, Hao ChenarXiv 2026 · arXiv

TL;DR

CellCanon is the first "agentic" cellular world model that unifies single-cell perturbation prediction, evidence-grounded explanation, and real-time updating from new experimental observations. By integrating molecular knowledge into a shared world state, it achieves a massive 63% improvement in differential expression prediction while reducing computational costs (LLM tokens) by over 99% compared to gene-wise reasoning approaches.

What Problem Does This Paper Address?

Understanding how cells respond to drugs is central to biological discovery, but the combinatorial space of drugs, doses, and cell types is too vast to measure experimentally. While numerical models can predict these responses, they are often "black boxes" that can't explain why a prediction was made. Conversely, Large Language Models (LLMs) can reason about biology but are computationally expensive when applied to thousands of genes individually.

Furthermore, most models are static; they cannot easily incorporate new experimental data to refine their predictions for unobserved conditions without expensive retraining. CellCanon addresses this by creating a unified agent that can predict, explain its reasoning using established biological literature, and update its internal "world state" as new data arrives.

Methods

CellCanon introduces an agentic architecture centered around a Shared World State. This state links numerical predictions, acquired measurements, and source-linked biological claims.

  1. Knowledge-Conditioned Retrieval: Instead of reasoning gene-by-gene, CellCanon uses an LLM (like DeepSeek-V4-Flash) to compile a "condition-level" evidence packet. This includes drug-target interactions, pathways, and baseline cell expression from databases like ChEMBL and Reactome.
  2. Predictive Core: This evidence is encoded into a shared representation that guides gene-specific numerical readouts (MLPs). These readouts predict continuous response, differential expression (DE) probability, and direction (up/down).
  3. Mechanism Explainer: The agent evaluates regulatory paths to provide proofs for its claims or identify cases where evidence is insufficient.
  4. Observation-Driven Update: When a new measurement is provided (e.g., a drug's effect on a new cell line), the agent uses a kernel-based updater to revise predictions for that same drug in other, still-unmeasured contexts—all without changing the underlying model weights.
Figure 1: An agent for prediction, explanation, and updating. Evidence and a training-response atlas condition prediction. The world state links responses, source-linked claims, and measurements. Evidence review supports explanations; Update transfers observed responses across cell contexts with fixed predictor weights.

Datasets

  • SciPlex3: The primary evaluation dataset, containing transcriptional responses for thousands of drugs across three cell lines (A549, K562, MCF7).
  • Knowledge Bases: Integrated data from ChEMBL, Reactome, OmniPath, DoRothEA, PROGENy, and MSigDB.
  • Source-Grounded Benchmark: A newly introduced benchmark of 966 questions across 72 conditions designed to test the model's ability to provide evidence-linked explanations.

Key Insights and Ideas

  • Context Sharing is Key: The authors realized that genes in a single drug-cell condition share the same biological background. By compiling this evidence once per condition rather than once per gene, they achieved a 99.77% reduction in token usage for predictions.
  • Hybrid Numerical-Symbolic Reasoning: CellCanon bridges the gap between "sub-symbolic" neural networks (which handle messy expression data) and "symbolic" biological knowledge (which handles discrete facts like "Drug A inhibits Protein B").
  • In-Context Biological Learning: The "Update" feature acts as a form of in-context learning for biology. By seeing how a drug behaves in one cell type, the model can immediately improve its "belief" about how that drug will behave in another, similar cell type.
Figure 2: Shared evidence for single-cell prediction. a, Evidence compilation is shared across genes; continuous responses and DE/DIR probabilities guide generation from matched DMSO cells and doses. b, Text and structured features guide atlas retrieval and task readouts. Testing generated populations against DMSO gives final DE/DIR (Section B.4).

Key Results

  • State-of-the-Art Prediction: On 339 held-out conditions, CellCanon improved differential-expression AUPRC by 63.3% and Macro-F1 by 17.3% over the best existing baselines.
  • Superior Correlation: It reached a logFC Pearson correlation of 0.6944, significantly higher than the 0.4306 achieved by chemCPA.
  • Perfect Grounding: In the explanation benchmark, the model correctly answered all answerable questions with supporting proofs and correctly identified every case where evidence was insufficient.
  • Efficient Updating: In simulated feedback loops, using just one measured context per drug allowed CellCanon to exceed the performance of baselines that were fully fine-tuned on the new data.
CellCanon significantly outperforms baselines in logFC association and response ranking across gene panels.

Limitations and Future Work

  • Knowledge Base Dependency: The model's explanatory power is limited by the completeness of the underlying knowledge bases (e.g., if a drug-target interaction isn't in ChEMBL, the model may struggle).
  • Scope of Perturbations: The current study focuses on chemical perturbations; extending this to genetic perturbations (like CRISPR screens) or multi-drug combinations is a logical next step.
  • Real-time Interaction: While agentic, the current implementation is evaluated in a simulated feedback loop. Deploying this in an active lab environment to guide real-time experimental design remains future work.

Summary updated

Abstract

Biological discovery requires using molecular knowledge and new measurements to infer responses beyond the conditions already observed. We introduce CellCanon, the first agentic cellular world model to unify single-cell expression prediction, evidence-grounded explanation, and observation-driven updating. Its predictive core reuses condition-level biological evidence across gene-specific numerical readouts to guide single-cell expression generation. A shared world state links predictions, acquired measurements, and source-linked claims, supporting explanation and revision. On 339 held-out SciPlex3 conditions across three cell lines, CellCanon improves differential-expression AUPRC and Macro-F1 over the best baseline for each metric by 63.3% and 17.3%, respectively, and direction Macro-F1 by 7.2%. Relative to gene-wise reasoning on matched drug-cell cohorts, measured language-model token use falls by 99.77% for Predict and 98.41% for Explain. On the source-grounded benchmark, Explain correctly answers all answerable questions with supporting proofs and identifies all insufficient-evidence cases. In simulated experimental feedback, Update uses one measured source context per drug to exceed the best fine-tuned baselines in differential-expression AUPRC and Macro-F1 without changing model weights. Together, these capabilities support a cumulative account of cellular responses, with predictions and biological judgments that can be revised as evidence accumulates.
BibTeX
@misc{liang2026cellcanon,
  author = {Zhaokang Liang and Shuyang Zhuang and Xiaoran Jiao and Hao Chen},
  title = {{CellCanon: An Agentic Cellular World Model for Prediction, Explanation, and Observation-Driven Updating}},
  year = {2026},
  eprint = {2611.05890},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2611.05890},
}
Figure