CellCanon: An Agentic Cellular World Model for Prediction, Explanation, and Observation-Driven Updating
Listen
Radio episode about this paper
Transcript
Introduction to the show: ident: Genomics Radio. Generated commentary on the latest computational biology and genomics papers.
Ines: We're looking at the paper: "CellCanon: An Agentic Cellular World Model for Prediction, Explanation, and Observation-Driven Updating".
Marcus: Here's the gist of the paper. Biological discovery requires using molecular knowledge and new measurements to infer responses beyond the conditions already observed CellCanon, the first agentic cellular world model to unify single-cell expression prediction, evidence-grounded explanation, and observation-driven updating.
Ines: Core Model Architecture CellCanon introduces a shared world state that links predictions, acquired measurements, and source-linked claims to support explanation and revision.
Marcus: This state connects three roles: language models interpret evidence, tools check citations and paths, and numerical models predict and update responses.
Ines: The shared world state contains the continuous response ybq, DE probabilities pDEq, pupq, Eq C(t)q containing claims and their dependencies, and Ot recording acquired measurements.
Paper summary: Ines: So we're talking about CellCanon today, this paper is about this new agentic cellular world model that tries to unify prediction, explanation, and updating of single-cell expression Marcus exactly. It claims it uses a shared world state to link predictions with measurements and source linked claims so you can actually explain and revise the findings Ines right?
Marcus: Right. The main thesis here is that biological discovery needs molecular knowledge and new measurements to go beyond what we already see, and CellCanon is presented as this first agentic cellular world model doing that Yuki by reusing condition level biological evidence across different gene specific numerical readouts to guide the expression generation Ines
Ines: It sets up a shared world state that connects three roles: language models interpreting the evidence, tools checking citations and paths, and numerical models predicting and updating responses Marcus This state holds continuous responses, differential expression probabilities, and claims along with their dependencies Yuki
Marcus: And the predictive core reuses condition level biological evidence across gene specific numerical readouts to guide the single cell expression generation Ines So it uses that shared representation of features to map it to a continuous gene panel that includes things like MLPRNAc Yuki
Ines: The explanation and updating parts are pretty interesting because Explain assesses source support for target actions, control cell expression, regulatory paths, and source statements returning supported or contradicted judgments with natural language rationales Marcus And Update uses a newly measured drug response to revise predictions for the same drug in unmeasured cell contexts using a support trained kernel ridge model Yuki
Marcus: The performance metrics they report are pretty strong too. On three hundred thirty-nine held out SciPlex three conditions across three cell lines, CellCanon improves differential expression AUPRC and Macro F1 over the best baseline for each metric by sixty-three point thirty three percent and seventeen point three percent respectively Ines <ref:2611.05890#pg1,conditions across three cell lines, CellCanon improves differential expression AUPRC and Macro>
Paper summary: Ines: That improvement on direction Macro F1 is seventy point two percent too Marcus And relative to gene wise reasoning on matched drug cell cohorts, the measured language model token use falls by ninety-nine point seven seven percent for predict and ninety eight point four one percent for explain Yuki
Marcus: On the source grounded benchmark, Explain correctly answers all answerable questions with supporting proofs and identifies all insufficient evidence cases Ines And 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 Yuki
Ines: This supports a cumulative account of cellular responses where predictions and biological judgments can be revised as evidence accumulates Marcus And the sensitivity analysis shows that removing biological annotations lowers differential expression Macro F1 from zero point eight zero seven seven to zero point seven one one seven Yuki
Marcus: It also shows that removing measured DMSO expression and detection lowers the delta Pearson from zero point three zero seven four to zero point two eight one zero Ines The raw source knowledge view gives differential AP of zero point eight six six one and direction mAP of zero point nine zero six five Yuki
Ines: The observation driven update effect is pretty clear too when Update uses one source context per drug it reaches differential AUPRC and Macro F1 of zero point nine zero seven seven and zero point eight two six one compared to zero point six two seven nine and zero point seven six three for the strongest fine tuned comparator Marcus <ref:2611.05890#pg2>
Marcus: And at forty eight observations, CellCanon’s differential Macro F1 rises from zero point eight one nine nine to zero point eight three nine three Yuki This shows that observation driven updating supports a cumulative account of cellular responses where predictions and biological judgments can be revised as evidence accumulates Ines
Ines: The evidence grounded reasoning part is also pretty robust because Explain correctly answers all answerable questions with supporting proofs and identifies all insufficient evidence cases on the source grounded benchmark Marcus It uses the deepseek flash api for source analysis, question answering, and localized review Yuki
Marcus: And I’m interested in how the gene panel prediction works. They have an additional panel containing nine three one measured lincs l one thousand landmark genes outside of two thousand highly variable genes Ines With support differential expression direction label supervision for these genes, the independent classification heads achieve AUPRC of zero point seven six seven four and auprc of nine two one for differential expression queries and seven nine zero eight for direction queries Yuki <ref:2611.05890#pg2>
Paper summary: Ines: And the combined panel has eighty thousand two hundred twenty-eight sampled queries and twelve thousand four hundred twenty eight differential positives Marcus This prediction results use generated expression testing, but observation feedback results use the shared condition heads Yuki
Marcus: So what does this mean for us as a field? The paper CellCanon: An Agentic Cellular World Model for Prediction, Explanation, and Observation-Driven Updating suggests we can build a system that doesn't just predict expression but also explains why it predicts it and keeps learning from new measurements Ines
Ines: It means we move toward systems where biological judgments are not static; they can be revised as evidence accumulates through this feedback loop Marcus It’s about making the inference process more robust by explicitly modeling how we use external data to correct our internal models Yuki
Marcus: The implication for large cohort studies is that if you can reduce language model token use by almost ninety-nine percent for prediction and explanation, you can make these complex biological models much more efficient to run Ines It addresses the computational cost of reasoning over vast amounts of genomic data Yuki
Ines: But there’s a limitation they point out. The context removed model retains the response atlas and cell specific modules but excludes measured control expression and detection information from evidence compilation and encoding Marcus So it still relies on having that initial biological annotation to be fully effective Yuki
Marcus: Exactly. They show that without those initial annotations, the performance drops significantly, which tells us how much reliance we have on pre-existing biological context for these models to work well Ines
Ines: So the core idea of CellCanon: An Agentic Cellular World Model for Prediction, Explanation, and Observation-Driven Updating is creating a system that unifies these three things—prediction, explanation, and updating—into one cohesive agentic structure Marcus It's about moving from static prediction to dynamic inference guided by real-time observation Yuki
Conclusion: Ines: So, to wrap up this whole CellCanon thing, we’re talking about this agentic cellular world model that manages prediction, explanation, and updating all at once Marcus
Marcus: Right. The authors are pushing this unified system of single cell expression prediction and using new measurements to actually improve the biology beyond what we already know Yuki
Ines: It's a shared world state linking predictions with measurements and claims so you can explain things and revise them as new data comes in Marcus
Marcus: And the big numbers show that this approach improves differential expression accuracy by sixty point three percent and macro F1 by seventeen point three percent over the best previous methods Yuki
Ines: That level of improvement is significant, especially when you look at how it uses language model tokens—it drops them almost to zero for prediction and explanation Marcus
Marcus: It means we can build these complex biological models much more efficiently for large cohort studies because the reasoning part doesn't need as many tokens Yuki
Ines: But there are still caveats, like how the model relies on initial biological annotations to work well, which is something the authors point out in their sensitivity analysis Marcus
Marcus: Exactly. The system isn't magic; it still needs that starting biological context to be fully effective Yuki
Ines: So what does this mean for biology right now? It means we can move toward systems where biological judgments aren't static; they can be revised as evidence accumulates Marcus
Marcus: It’s about making the inference process more robust by explicitly modeling how we use external data to correct our internal models Yuki
Ines: Next up, we're going to talk about what this means for the actual experimental setup and how this updating mechanism works in practice Marcus
Zhaokang Liang, Shuyang Zhuang, Xiaoran Jiao Hao Chen
Zhejiang University
q-bio.QM, cs.AI
Submitted: 2026-11-04
Updated: 2026-11-04
Comments: 31 pages, 5 figures. Code: https://github.com/ZhaokangLiang123/CellCanon
Code: https://github.com/ZhaokangLiang123/CellCanon
License: http://creativecommons.org/licenses/by/4.0/
Importance score: 100/100
The gist: Biological discovery requires using molecular knowledge and new measurements to infer responses beyond the conditions already observed CellCanon…
Key concepts
- Shared World State
- This central component links three elements: predictions, acquired measurements from experiments, and source-linked claims about biological facts. It acts as a unified memory that allows the model to explain its reasoning and revise its predictions when new evidence is gathered.
- Predictive Core
- The core mechanism generates single-cell expression by reusing condition-level biological evidence across various gene-specific numerical readouts. A language model selects relevant source evidence, which is then mapped through a shared representation to predict the continuous gene panel for a specific cell.
- Update Mechanism
- This component uses newly measured drug responses to revise previous predictions for the same drug in unmeasured cell contexts. It employs a support-trained kernel ridge model that combines similarity between measured sources and chemical/biological similarities to infer the target response without retraining the main model weights.
Terminology
Summary
Biological discovery requires using molecular knowledge and new measurements to infer responses beyond the conditions already observed CellCanon, the first agentic cellular world model to unify single-cell expression prediction, evidence-grounded explanation, and observation-driven updating. p. 0 CellCanon: An Agentic Cellular
World Model for Prediction, Explanation,
and Observation-Driven Updating
Zhaokang Liang∗ Shuyang Zhuang∗ Xiaoran Jiao Hao Chen†
Zhejiang University
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.Where the paper says it
Core Model Architecture
CellCanon introduces a shared world state that links predictions, acquired measurements, and source-linked claims to support explanation and revision. This state connects three roles: language models interpret evidence, tools check citations and paths, and numerical models predict and update responses. p. 2 3 Method
3. 1 Cellular world model and evidence state
The shared world state connects three roles: language models interpret evidence, tools check citations and paths, and numerical models predict and update responses. Training drugs Dtr supply the
support atlas. Each query state contains a continuous response ybq, DE probabilities p
DE
q
, and
conditional up-regulation probabilities p
up
q
:
S
(t)
q =
yb
(t)
q
, p
DE,(t)
q
, p
up,(t)
q
, Eq, C
(t)
q
, Ot
, (1)
where Eq contains source-linked evidence, C
(t)
q contains claims and their dependencies, and Ot
records acquired measurements.Where the paper says it
Where the paper says it
Predictive Mechanism
The predictive core reuses condition-level biological evidence across gene-specific numerical readouts to guide single-cell expression generation. Biological evidence compilation involves the language model selecting source-linked target evidence and qualifying its relevance and assay scope. p. 2 2 Predict: language-selected evidence meets an observation atlas
Biological evidence compilation. For each condition, the language model selects source-linked
target evidence and qualifies its relevance and assay scope. Frozen text embeddings encode selected evidence and weighted target descriptions; structured features preserve evidence categories and quantitative context. Their concatenation xq undergoes
training-fitted projection and normalization, zq = Ptr(xq) (Equation (6) and Section B). Continuous and classification readouts. A cell-specific MLP maps the shared representation to
the continuous gene panel:
ybdc = µc + sy MLPRNA
c
(udc), (4)
where µc
and sy are training response normalization statistics.Where the paper says it
Where the paper says it
Where the paper says it
Explanation and Updating
Explain assesses source support for target action, control-cell expression, regulatory paths, and source statements, returning supported, contradicted, or insufficient evidence judgments with citations and natural-language rationales. Update uses a newly measured drug response to revise predictions for the same drug in unmeasured cell contexts. p. 3 Section F specifies the source-question and observation-state tasks. 4 Update: incorporating newly measured responses
Update uses a newly measured drug response to revise predictions for the same drug in unmeasured
cell contexts. A support-trained kernel ridge model combines measured source-response similarity with chemical
and biological similarities to infer the target response (Equations (12) and (13)). The parent predictor
and update maps remain fixed.Where the paper says it
Where the paper says it
Performance and Efficiency Gains
CellCanon improves differential-expression AUPRC and Macro-F1 over the best baseline for each metric by 63.3% and 17.3%, respectively, on 339 held-out SciPlex3 conditions. p. 0 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%, respectively, and direction MacroF1 by 7. 3%, respectively, and direction MacroF1 by 7. Relative to gene-wise reasoning on matched drug–cell cohorts, measured language-model token use falls by 99. Relative to gene-wise reasoning on matched drug–cell cohorts, measured language-model token use falls by 99. 77% for Predict and 98. 77% for Predict and 98. 41% for Explain. Target conditions remain unobserved
(Sections G and H. 2 Predicting expression and differential responses
CellCanon achieves DE AUPRC of 0. 2 Predicting expression and differential responses
CellCanon achieves DE AUPRC of 0. 8067 and Macro-F1 of 0. 8067 and Macro-F1 of 0. 8164, versus 0. 8164, versus 0. 4938 and 0. 4938 and 0. 6961 for
the strongest baselines (Table 1).Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Observation-Driven Improvement
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. At 48 observations, CellCanon’s DE Macro-F1 rises from 0. p. 6 7431 versus 0. At 48 observations, CellCanon’s DE Macro-F1 rises from 0. 7431 versus 0. At 48 observations, CellCanon’s DE Macro-F1 rises from 0. 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.Where the paper says it
Where the paper says it
Where the paper says it
Evidence-Grounded Reasoning
Explain correctly answers all answerable questions with supporting proofs and identifies all insufficient-evidence cases on the source-grounded benchmark. The explanation procedure uses the DeepSeek Flash API (deepseek-flash) for source analysis, question answering, and localized review. p. 4 1 per condition (Section A). We use the DeepSeek-V4-Flash API for
evidence compilation and frozen Qwen3-Embedding-0. On the
source-grounded benchmark, Explain correctly answers all answerable questions with complete proofs,
abstains on all insufficient-evidence cases, and retrieves valid paths among the top three for all eligible
conditions (Table 4). Observation-state and withdrawal tests track revisions as measurements arrive
or sources are removed (Section F).Where the paper says it
Where the paper says it
Sensitivity Analysis
Removing biological annotations lowers DE Macro-F1 from 0.8077 to 0.7117, while removing measured DMSO expression and detection lowers ∆-Pearson from 0. p. 7 Removing biological annotations lowers
DE Macro-F1 from 0. Removing measured DMSO expression and
detection lowers ∆-Pearson from 0. (a) Shared evidence for prediction
Per-gene LLM prediction
Condition-shared prediction
Gene query
g
qg= (c, d, g)
Retrieved context LLM task Gene-level prediction
DE DIR
Condition query
q = (c, d)
Knowledge
conditioning
LLM evidence compiler
Encoders
Training
support atlas
World model
Shared response
profiles
r …
pDE …
pUP …
DMSO cells
Dose
Probability-guided
cell generator
Single-cell
expression
Cell population
Response · DE · DIR
(b) Predictive core
Knowledge conditioning
Drug + cell context
LLM evidence
compiler
LLM encoder
Feature encoder
Projector
Support keys World Model Query
Support atlas
Training keys +
observed responses
Training
observations
Retrieval & gated fusion
Response predictor
MLP
Response regressor DE classifier DIR classifier
Dose
DMSO cells
Baseline statistics
Probability-guided cell generator
Single-cell expression
Response DE DIR
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. Observation-state and withdrawal tests track revisions as measurements arrive
or sources are removed (Section F). 5 Biological context and independent readouts
Three ablations assess biological information and task-specific supervision (Table 5). Section I gives token totals and cohort definitions. 5 Discussion and Conclusion
CellCanon connects accurate single-cell response prediction, source-grounded explanation, and
observation-driven updating in an agentic cellular world model. Retrieval combines learned compatibility
with chemical, target, and mechanism-of-action similarities sr(d, i):
ℓdi,c =
q
⊤
dckic
√
H
+
X
3
r=1
softplus(wr)sr(d, i), αdi,c = softmaxi∈Ad
(ℓdi,c), (2)
hdc = LayerNorm
qdc + sigmoid(Wgqdc + bg) ⊙ Wo
X
i∈Ad
αdi,cvic! (3)
A cell-specific map udc = MLPc(hdc) produces the shared readout representation. (a) Shared evidence for prediction
Per-gene LLM prediction
Condition-shared prediction
Gene query
g
qg= (c, d, g)
Retrieved context LLM task Gene-level prediction
DE DIR
Condition query
q = (c, d)
Knowledge
conditioning
LLM evidence compiler
Encoders
Training
support atlas
World model
Shared response
profiles
r …
pDE …
pUP …
DMSO cells
Dose
Probability-guided
cell generator
Single-cell
expression
Cell population
Response · DE · DIR
(b) Predictive core
Knowledge conditioning
Drug + cell context
LLM evidence
compiler
LLM encoder
Feature encoder
Projector
Support keys World Model Query
Support atlas
Training keys +
observed responses
Training
observations
Retrieval & gated fusion
Response predictor
MLP
Response regressor DE classifier DIR classifier
Dose
DMSO cells
Baseline statistics
Probability-guided cell generator
Single-cell expression
Response DE DIR
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. 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%, respectively, and direction MacroF1 by 7. 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%, respectively, and direction MacroF1 by 7. 4
Δ-Spearman c Response ranking
CellCanon
CMonge
BioLORD
Squidiff
chemCPA
PerturbNet
PRnet
STATE
Figure 3: Response-focused prediction. a, Per-condition logFC Pearson on 2,000 genes: boxes
show quartiles, gray dashed lines medians, whiskers the furthest points within 1.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Update Effectiveness
With one source context per drug, CellCanon reaches DE AUPRC/Macro-F1 of 0.9077/0.8261, compared with 0. p. 6 With one source context per drug, CellCanon reaches DE AUPRC/Macro-F1 of 0. 8261,
compared with 0. 8261,
compared with 0. 7630 for the strongest fine-tuned comparator (Table 3). We introduce CellCanon, the first agentic cellular world model to unify single-cell expression prediction, evidence-grounded explanation, and observation-driven updating. Encoded molecular evidence and a training-response atlas feed continuous multilayer perceptron (MLP) and classification
readouts, whose gene-specific outputs guide single-cell generation (Figure 2).Where the paper says it
Where the paper says it
Where the paper says it
Gene Panel Prediction
The additional panel contains the 931 measured LINCS L1000 landmark genes outside the 2,000 highly variable genes. With support DE/DIR label supervision for these genes, the independent classification heads achieve AUPRC of 0. p. 7 Replacing independent classification heads with continuous-response
scores lowers DE Macro-F1 to 0. 5278 with
validation-selected thresholds, supporting
distinct supervision for response magnitude
and DE occurrence. DE and DIR use AUPRC, AUROC, Macro-F1, accuracy, precision, and recall. DIR conditions on reference-DE queries with upregulation as positive. Table 1: Differential-response prediction on the 2,000-gene SciPlex3 panel. DE/joint use sampled test queries; DIR uses reference-DE queries with up-regulation as positive. BaiZe predicts expression and supports few-shot cross-species adaptation; BaiZeAgent retrieves and explains saved results [Zeng et al. CellCanon links prediction, sourcegrounded explanation, and measurement-driven revision through a shared world state.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.
Improvements for AI systems
-
Agentic Cellular World Model for Prediction, Explanation, and Observation-Driven Updating: This system can unify single-cell expression prediction, evidence-grounded explanation, and observation-driven updating by linking predictions to source-linked claims and acquired measurements through a shared world state.
-
Evidence-Grounding Biological Explanation: The system can evaluate target actions and regulatory paths by returning
supported, contradicted, or insufficient evidence judgments with citations and natural-language rationales,
distinguishing between binding, signed regulation, and solvent-control RNA detection. -
Observation-Driven Updating: The system can revise predictions for the same drug in unmeasured cell contexts using a measured source context per drug to exceed best fine-tuned baselines without changing model weights.
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.
Sources
- MapPFN: Learning Causal Perturbation Maps in Context
- LLM-Guided Retrieval for Prediction of Molecular Perturbation Responses
- Qwen3 Embedding: Advancing Text Embedding and Reranking Through Foundation Models
Related papers
- ViroGym: Realistic Large-Scale Benchmarks for Evaluating Viral Proteins
- Probing and steering biology across Boltz-1s trunk-diffusion boundary
- Trajectory Inference of Human Aging from Cross-Sectional DNA Methylation Data
- Bridging the Modality Bottleneck in Pathology MIL through Molecularly Guided Feature Adaptation
- Molecular notation can act as a shortcut in AI benchmark evaluation
- Towards Efficient Auto-Interpretability of Boltz-1 Sparse Autoencoders