acceptodds
Under review as a conference paper at ICLR 2027

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

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.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
single-cell perturbation modeling, virtual cells, scientific agents, cellular world models, drug response prediction, uncertainty-aware decision making

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_69D3OpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper presents CellCanon, a framework combining cellular perturbation prediction, source-grounded explanation, and observation-driven updating. Its predictive component combines condition-level biological knowledge with retrieval from a training-response atlas, followed by continuous-response and separate DE/direction readouts. These outputs guide single-cell expression generation. An explanation component checks biological claims against source evidence, while a fixed numerical updater incorporates newly observed responses to improve predictions for the same drug in other cell contexts. Evaluation covers held-out SciPlex3 drug–cell conditions, source-grounded explanation tasks, and retrospective observation-feedback experiments.

Critical strengths
  1. CellCanon reports substantial improvements in differential-response prediction across a broad set of comparators. On the main evaluation, DE AUPRC improves from the strongest baseline’s 0.4938 to 0.8067, and joint unchanged/up/down Macro-F1 improves from 0.5400 to 0.7420. The component studies also provide useful evidence for biological annotations and separately supervised response and classification readouts. These results make the predictive approach worth considering beyond its broader agent framing.

  2. Sharing condition-level evidence across genes avoids repeated language-model analysis, while the numerical updater provides a concrete way to use newly measured source-cell responses without test-time parameter updates. The explicit distinction between source evidence, predictions, and measurements is also useful for scientific applications. Although the individual techniques are largely established, their integration addresses a meaningful practical problem.

Critical weaknesses
  1. The contribution of the emphasized LLM and agent components is not sufficiently isolated. The ablations establish the value of biological information and independent classification heads, but provide weaker evidence for LLM-based evidence compilation itself. In Table 15, raw-source encoding slightly outperforms LLM-compiled knowledge on DE AP and F1, while compilation improves direction metrics. Furthermore, comparisons against expression predictors do not fully isolate the effects of additional classification supervision and decoding. Matched controls using the same information, supervision, and readout would better establish which proposed mechanisms drive the gains.

  2. The evidence for a tightly integrated “agentic cellular world model” is limited. The shared state tracks predictions, evidence, claims, and observations, but the demonstrated numerical dependencies mainly involve source measurements updating target predictions. The paper does not establish that explanation revision or claim-dependency management improves prediction or acquisition decisions; observation schedules are fixed, and explanation and numerical transfer are evaluated separately. The explanation benchmark primarily tests source-supported judgments rather than the faithfulness of explanations to the numerical predictor. Appendix F.1 also states that evaluation-panel error analysis informed refinement of the correction procedure, making an independent evaluation of the final pipeline important.

Questions for authors
  1. Which outputs of Explain or the claim-dependency state are consumed by Predict or the numerical Update algorithm? With identical compiled knowledge, atlas, weights, and observation schedules, would removing Explain and claim management change any reported numerical result?

  2. How do the predictive gains compare with a simpler knowledge-conditioned retrieval model equipped with the same DE/direction supervision and count decoder? Are any existing comparisons sufficient to separate these effects from the proposed LLM compilation and integration?

  3. Which parts of the correction procedure were modified following evaluation-panel error analysis? Is there an untouched evaluation set, or another assessment performed after the final procedure was frozen, that confirms the reported explanation performance?

7 comments
AuthorsOpenReview ↗

Response to Reviewer 69D3 (Part 1/6):

Thank you for highlighting the gains in differential-response prediction, the efficiency of sharing evidence across genes, and the practical value of keeping sources, predictions, and measurements distinct while updating with fixed model weights.

To address your concerns, we added the following comparisons and experiments.

  • LLM compilation compared with raw text, shuffled judgments, and deterministic rules, plus a replication with an open-weight model.
  • Matched kNN, Linear, and training-mean baselines before and after single-cell generation, a single-seed comparison, and a chemical-scaffold holdout.
  • Tests of classification heads, DE count decoding, and probability-guided generation with the other components held fixed where possible.
  • Updates from the same 48 measurements on SciPlex3 and from 1, 3, or 5 measured cell lines on Tahoe-100M, including comparisons with simpler ways to transfer measured responses.
  • Removing Explain from the shared state, testing claim revision after observations or sources change, and checking prediction confidence and available target evidence.
  • Three evaluations constructed after freezing Explain: unseen-drug questions, real source withdrawal and recombination, and multi-step reasoning on new graph templates.

1. The contribution of the emphasized LLM and agent components is not sufficiently isolated.

On pre-cleaned sources, rules and LLM compilation gave similar discrete accuracy with the same structured representation. The LLM improved continuous prediction over raw text and improved one post-generation ranking metric over the rules. To separate the effects of compilation and source volume, all variants used the same upstream source records across 558 conditions. We compared raw text, shuffled judgments, a deterministic parser that fills the same evidence fields, and an open-weight end-to-end replication.

The deterministic parser and LLM compilation reached 0.7575 and 0.7532 sampled Joint F1. Across 20 discrete comparisons, all Holm-adjusted p-values were at least 0.4215. Both exceeded matched kNN at 0.7218 and Linear at 0.6749. The structured evidence, supervision, and readout are shared by both variants. The open-weight Qwen3.5-35B-A3B replication was within 0.0065 on the main prediction endpoints, including 0.7384 versus 0.7420 Joint F1. Separately, it reached 95.90% answerable Explain accuracy versus 100% for the submitted pipeline, and Update RMSE was 4.4475 versus 4.4969.

With the source records and text budgets matched, LLM judgments improved pre-generation Pearson correlation from 0.6240 for raw text to 0.6340 and lowered RMSE from 5.0056 to 4.9278, with Holm-adjusted p = 0.0014. Shuffling judgments within splits or across drugs gave Pearson correlations of 0.6252 and 0.6257. Relative to these shuffles, LLM judgments reduced RMSE by 0.0729 and 0.0714, with Holm-adjusted p = 0.0008. With the same balanced validation criterion, post-generation Top 100 Δ-Spearman was 0.3753 for the LLM and 0.3597 for the rules. The gain was 0.0156, with a 95% interval of 0.0036 to 0.0275 and unadjusted two-sided p = 0.0129. Compiled evidence links each claim to its sources. Reusing condition-level evidence across genes cut prediction token use by 99.77% relative to VCWorld.

Compiler comparisons on 339 SciPlex3 test conditions

Evidence compilationSampled DE APSampled DE F1Sampled direction F1Sampled Joint F1Top 50 R²Δ-PearsonTop 100 Δ-Spearman
Submitted CellCanon, DeepSeek-V4-Flash0.80670.81640.84720.74200.91450.32540.3809
Open-weight Qwen3.5-35B-A3B0.80650.81640.84380.73840.91420.31910.3761
LLM compilation, balanced validation0.80860.82530.85020.75320.90210.30560.3753
Deterministic rules, balanced validation0.81790.83040.84790.75750.90370.29800.3597

For the DE AP comparison quoted in the review, 0.4938 is VCWorld's two-point trapezoidal PR area. Its step-wise AP is 0.2196. Under matched inputs and generation, the strongest baseline is kNN at 0.7574, and CellCanon reaches 0.8067, a gain of 4.93 percentage points.

See the revised manuscript, Tables 1–2 and 4, Appendix E, Tables 15–16, and Appendix I, Table 27.

AuthorsOpenReview ↗

Response to Reviewer 69D3 (Part 2/6):

2. How do the predictive gains compare with a simpler knowledge-conditioned retrieval model equipped with the same DE/direction supervision and count decoder?

CellCanon improved on the matched kNN model by 4.93 percentage points in sampled DE AP and 2.02 points in sampled Joint F1 after single-cell generation. We compared cosine-weighted kNN over the 20 nearest support examples, ridge regression with per-gene L2-regularized logistic heads, and a drug-agnostic prior using cell-line mean response and gene-wise prevalence. They share the same 96-dimensional biological and chemical features, 50-candidate validation budget, DE and direction supervision, count decoding, and frozen generator. We generated 341,202 cells per model and applied the same Wilcoxon and Benjamini–Hochberg tests.

Across 112 test drugs, the 95% interval for the sampled DE AP gain over kNN was 2.92 to 7.31 points, with p = 0.0001. The Joint F1 comparison had p = 0.0247. The dense gains were 6.12 points in DE AP and 5.13 points in Joint F1. The Systema correlation after removing systematic cell-line shifts was +0.0526 versus +0.0186, p = 0.0046. A single-seed CellCanon reached 0.8102 DE AP and 0.7582 Joint F1, exceeding deterministic kNN by 5.28 and 3.64 points, with p = 0.0001. CellCanon led matched Linear by 7.60 points in DE AP and 6.71 in Joint F1, with p = 0.0001 and p = 0.0004.

When we retrained on the SciPlex3 split with zero shared Bemis–Murcko scaffolds, cross-split Tanimoto similarity below 0.30, and 339 conditions, CellCanon reached 0.7414 sampled Joint F1 before generation. The gains were 13.42 points over Linear at 0.6072, with Holm-adjusted p = 0.0038, and 18.03 points over kNN at 0.5611.

SciPlex3 results after single-cell generation

MethodSampled DE APSampled DE F1Sampled direction F1Sampled Joint F1Top 50 R²Δ-PearsonSystema correlation
CellCanon0.80670.81640.84720.74200.91450.3254+0.0526
kNN0.75740.80400.82190.72180.88870.2910+0.0186
Linear0.73070.77450.77810.67490.89530.2945+0.0056
Drug-agnostic prior0.66770.70520.74600.59400.89690.2768+0.0012
PRnet0.40570.69610.70060.54000.90390.1528+0.0001
BioLORD0.26950.61840.79030.49000.91970.2538−0.0004

See the revised manuscript, Tables 1–2, Appendix C, and Appendix L, Table 36.

AuthorsOpenReview ↗

Response to Reviewer 69D3 (Part 3/6):

3. Are any existing comparisons sufficient to separate these effects from the proposed LLM compilation and integration?

Yes. The matched comparisons separate the prediction model from the shared compiled inputs. The additional tests then show what changes when we remove the classification heads, change the count decoder, or turn off probability guidance during generation.

  • Prediction before generation. With matched features, supervision, and count decoding, CellCanon gained 3.84 percentage points in sampled DE AP over kNN, 0.8588 versus 0.8204, with raw p = 0.0058 and Holm-adjusted p = 0.0232. The Joint F1 gain was 3.90 points, 0.7242 versus 0.6852, with raw p = 0.0180 and adjusted p = 0.0350. Against Linear, the gains were 5.19 and 6.86 points, with adjusted p = 0.0350 and p = 0.0270.
  • Classification supervision. Removing the heads from CellCanon reduced sampled DE AP to 0.1464, with Holm-adjusted p = 0.0036. Adding the same heads and count decoder to external models raised chemCPA from 0.1337 to 0.7675 DE AP. With those components matched, CellCanon reached 0.7242 Joint F1 versus 0.6359 for BioLORD, the strongest adapted deep model, a gain of 8.83 points. Adapted CMonge, PRnet, and chemCPA reached 0.5420 to 0.6326 Joint F1.
  • Count decoding. On the dense panel of 678,000 queries with 1.28% true DE, count decoding called 15,480 genes as DE across queries, versus 53,022 to 82,707 calls using sampled probability thresholds. Count MAE improved from 132.8 to 33.5, and dense Joint F1 from 0.453 to 0.557. With per-model dense thresholds selected on validation data, CellCanon led Linear by 5.94 and kNN by 4.94 points, with 95% intervals above zero.
  • Probability-guided generation. Holding the condition head fixed, guidance raised DE Macro-F1 from 0.6873 to 0.8315, a gain of 14.42 points with Holm-adjusted p = 0.0005. Continuous RMSE was 4.9090 without guidance and 4.9078 with guidance. The corresponding comparison for the submitted model showed a gain of 16.58 points.

SciPlex3 predictions before single-cell generation

ModelSampled DE APSampled DE F1Sampled direction F1Sampled Joint F1Dense DE APDense Joint F1Systema correlation
CellCanon0.85880.80770.83930.72420.33710.5472+0.0545
kNN0.82040.77940.82270.68520.29830.5276+0.0194
Linear0.80690.76700.79280.65560.22260.5007+0.0184
Drug-agnostic prior0.76850.70990.79200.56990.17620.45760.0000
BioLORD with matched heads and decoder0.83950.73160.81210.63590.26500.5368−0.0004
CellCanon without classification heads0.14640.45400.82390.30400.01210.3298+0.0336

See the revised manuscript, Table 6 and Appendices D–E, including Table 17.

Reviewer_Ut25OpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes CellCanon, an agentic cellular model that unifies prediction, evidence-grounded explanation, and updating for single-cell prediction. The cellular state is shared across agentic steps, which supports explanation and revision. This shows the predictive performance improvement in DE and DIR prediction.

Critical strengths

None

Critical weaknesses
  • The “first agentic cellular world model” claim is too strong: Related work enumerates prior agents without stating what they lack or how the proposed method improves on them. Recent systems that combine verifiable LLM reasoning with database-grounded perturbation prediction, such as VCR-Agent [1] and RBio1 [2], are not cited, even though they connect prediction and reasoning for perturbations in an agentic system. As prior works already handle the connection among prediction, agents, and evidence-grounded explanation, I cannot find any significant difference to call this work “The first agentic cellular world model”. The update is the only point that was missing in prior work, but it is still marginal to call it so.
  • Strong statistical baselines are cited but never run: The paper cites [3], which reports that deep perturbation models do not outperform simple statistical baselines. However, the paper never shows the performance of the linear or nearest-neighbor baseline that appears in Table 1 and only shows weak Bernoulli(0.5).

[References]

[1] Jang, Y., et al. Towards Autonomous Mechanistic Reasoning in Virtual Cells. EMNLP Findings 2026.

[2] Istrate, A. M.,et al. RBio1-training scientific reasoning LLMs with biological world models as soft verifiers. bioRxiv 2025.

[3] Ahlmann-Eltze, C., et al. Deep-learning-based gene perturbation effect prediction does not yet outperform simple linear baselines. Nature Methods 2025.

Questions for authors
  • For the hard-decision metrics, would a prevalence-matched random reference change the reported Macro-F1 gap?
  • How do the authors define “agentic cellular world model”?
  • How do strong statistical baselines, including the linear model and k-nearest neighbor model, work?
7 comments
AuthorsOpenReview ↗

Response to Reviewer Ut25 (Part 1/6):

Thank you for noting how the shared cellular state supports explanation and revision, and for highlighting the reported gains in differential expression and direction prediction.

To address your concerns, we added the following comparisons and experiments.

  • A comparison with prior reasoning agents, a test of probability-guided single-cell generation, and a comparison of language-model token use.
  • Updates from the same 48 measurements on SciPlex3 and from 1, 3, or 5 measured cell lines on Tahoe-100M, plus a test of how claims are revised as observations and sources change.
  • Matched training-mean, linear, and nearest-neighbor baselines before and after single-cell generation, and deep predictors with the same classification heads and count decoder.
  • A single-seed comparison, a drug-specific correlation analysis, chemical-scaffold and unseen-cell-line holdouts, and single-cell prediction across 30 Tahoe-100M cell lines.
  • A random reference matched to class frequencies in the support set.

1. The “first agentic cellular world model” claim is too strong … How do the authors define “agentic cellular world model”?

We define CellCanon by what its shared state lets it do: predict single-cell responses, check biological claims against sources, and revise predictions when measurements arrive. For each drug and cell line, the state records predicted expression changes, calibrated probabilities of differential expression and upregulation, molecular evidence, claims and their dependencies, and new measurements.

VCWorld, VCR-Agent [1], and RBio1 [2] already connect perturbation prediction with mechanistic reasoning. We removed the priority claim and expanded the comparison to cover these methods and PerTurboAgent. Here, agentic refers to tool-assisted evidence compilation, local review of claims with mandatory abstention when sources are insufficient, and automatic revision or invalidation of affected claims. Observations follow a fixed schedule. New measurements update numerical predictions with fixed weights, and explanations check and revise claims using the same state.

Capabilities of perturbation models and reasoning agents

MethodSingle-cell expressionDifferential expression and directionExplanationUpdate
PRnet, PerturbNet, STATE, BioLORD, chemCPA, CMonge, SquidiffContinuous cell-level profilesPost hoc calls without explicit DE count calibrationNot supportedGradient fine-tuning
VCWorldNot supportedGene-wise binary callsFree-text rationalesStatic inference outputs
VCR-Agent [1]Not supportedDirection calls via verifier-filtered VC-TRACES and graph distillationVerifier filtering during synthesis on Tahoe-100MOffline synthesis and distillation
RBio1 [2]Not supportedCalibrated direction calls on PerturbQATraining rewards from biological world models and Gene OntologyFixed weights at inference
CellCanonProbability-guided single-cell profilesContinuous effects, calibrated DE and direction heads, and DE count decodingSource and effect-sign verification with required abstentionMeasured responses update predictions across cell lines with fixed weights, and affected claims are invalidated

[1] Jang et al., Towards Autonomous Mechanistic Reasoning in Virtual Cells, arXiv:2604.11661. The EMNLP Findings 2026 proceedings title is Towards Structured Mechanistic Reasoning in Single-cell Prediction.

[2] Istrate et al., RBio1: Training scientific reasoning LLMs with biological world models as soft verifiers, bioRxiv 2025.

See the revised manuscript, Abstract, Sections 1–2 and 3.1–3.4, and Appendix L, Table 35.

AuthorsOpenReview ↗

Response to Reviewer Ut25 (Part 2/6):

2. The update is the only point that was missing in prior work, but it is still marginal to call it so.

CellCanon also differs in what it predicts: it generates continuous single-cell profiles from which both effect sizes and differential expression and direction calls can be recovered. VCWorld makes gene-wise binary calls, VCR-Agent uses verifier-filtered trajectories and graph distillation for direction calls, and RBio1 answers directional PerturbQA questions using training-time verification. In our matched test, probability-guided generation improved post-generation DE Macro-F1 by 14.42 percentage points with Holm-adjusted p = 0.0005. Continuous RMSE was 4.9090 without guidance and 4.9078 with guidance.

The same drug and cell line share upstream targets, pathways, and the overall sparsity of differential expression. Sharing evidence across genes gives their predictions a common basis. VCWorld used about 420 million tokens on SciPlex3. CellCanon used 99.77% fewer tokens for prediction. With the same DeepSeek-V4-Flash backbone, VCWorld achieved 0.2481 DE precision and 0.2196 DE AP.

With the same 48 observations on SciPlex3, Update reached 0.7645 Joint F1 with fixed weights, a gain of 15.77 percentage points over fine-tuned PerturbNet at 0.6068. On Tahoe-100M, we used 36 unseen drugs and 10 fixed measurement schedules across 30 cell lines from 13 organs. Responses from 1, 3, or 5 measured lines improved sampled Joint F1 on the remaining 29, 27, or 25 lines by 5.91, 8.88, or 10.47 points over Predict before any observations. Holm-adjusted p-values were at most 0.0112. The gains over uniform residual shift from the same starting predictions were 3.96, 4.58, and 5.78 points. Uniform residual shift led by 0.68 to 0.86 points on direction F1 among truly differentially expressed genes. DMSO kernel transfer reached 0.5158 to 0.5924 Joint F1, with p = 0.0024 for the comparison. Tracking claim dependencies also reduced stale active claims from 148 to zero and token use by 50.6% as observations or sources changed.

Results by number of measured Tahoe-100M cell lines

Measured linesMethodSampled DE APSampled DE F1True-DE direction F1Sampled Joint F1Dense DE APDense Joint F1
0Predict0.95060.73440.79920.61170.52140.5863
1CellCanon Update0.95370.78890.82840.67080.53870.6266
1Uniform residual shift0.95210.73960.83520.63120.52910.6006
3CellCanon Update0.95820.80140.86080.70050.56130.6508
3Uniform residual shift0.95410.75090.86940.65470.53940.6158
5CellCanon Update0.96170.80890.87890.71750.57890.6647
5Uniform residual shift0.95450.75170.88620.65970.54100.6199

The zero-observation row is evaluated on the same 29 target lines as the one-measurement comparison. For the 27- and 25-line comparisons, the paired sampled Joint F1 baselines were 0.6117 and 0.6128, and the dense baselines were 0.5861 and 0.5869.

See the revised manuscript, Sections 3.2–3.4 and 4.3–4.4, Table 6, and Appendices F, H, I, and K, including Tables 26, 27, and 34.

AuthorsOpenReview ↗

Response to Reviewer Ut25 (Part 3/6):

3. How do strong statistical baselines, including the linear model and k-nearest neighbor model, work?

Both Linear and kNN outperform all seven external deep models on DE AP and Joint F1, and CellCanon improves on both under the same inputs, supervision, and generation procedure. On sampled queries, the gains were 4.93 percentage points in DE AP and 2.02 in Joint F1 over kNN, and 7.60 and 6.71 points over Linear.

The drug-agnostic prior uses the mean response of the support drugs in each cell line and each gene's observed DE and direction frequencies. It corresponds to the training-mean baseline discussed by Ahlmann-Eltze et al. [3]. Linear uses ridge regression for continuous log fold change and per-gene L2-regularized logistic heads for DE and direction. The kNN model computes a cosine-similarity-weighted average of the 20 nearest support examples in the same 96-dimensional knowledge and chemical features. The models share the 50-candidate validation budget, explicit DE and direction supervision, the DE count decoder, and the frozen single-cell generator. We generated 341,202 cells per matched model and used Wilcoxon testing with Benjamini–Hochberg correction. The Systema correlation removes the mean perturbation response of the support drugs in the same cell line [4].

SciPlex3 results after single-cell generation

MethodSampled DE APSampled DE F1Sampled direction F1Sampled Joint F1Top 50 R²Δ-PearsonTop 100 Δ-SpearmanSystema correlation
CellCanon0.80670.81640.84720.74200.91450.32540.3809+0.0526
kNN0.75740.80400.82190.72180.88870.29100.3553+0.0186
Linear0.73070.77450.77810.67490.89530.29450.3602+0.0056
Drug-agnostic prior0.66770.70520.74600.59400.89690.27680.3495+0.0012
PRnet0.40570.69610.70060.54000.90390.15280.2089+0.0001
PerturbNet0.48710.67320.64250.49100.44610.15000.2062+0.0074
BioLORD0.26950.61840.79030.49000.91970.25380.3199−0.0004
STATE0.42550.60400.74180.48850.90630.24010.2809−0.0004
CMonge0.34750.64790.66990.46860.73590.14840.1822−0.0002
chemCPA0.13370.44220.67530.38210.90260.25430.3632+0.0018
Squidiff0.08800.10950.58460.09460.85900.07030.2396−0.0015
VCWorld0.21960.55000.52410.3767N/AN/AN/AN/A
Prevalence-matched random0.15560.49800.49620.3317N/AN/AN/AN/A
Uniform random0.15560.43240.49820.2881N/AN/AN/AN/A

The main evaluation covers 339 conditions. PerturbNet and the Squidiff row use the matched 111-drug subset. On all 339 conditions, Squidiff reaches 0.0875 DE AP and 0.0944 Joint F1. Excluding the one PerturbNet training-overlap compound gives 0.5021 DE AP and 0.4894 Joint F1. The table reports step-wise AP. The two-point trapezoidal areas for VCWorld and uniform random are 0.4938 and 0.3666, respectively.

[3] Ahlmann-Eltze et al., Deep-learning-based gene perturbation effect prediction does not yet outperform simple linear baselines, Nature Methods 2025.

[4] Viñas Torné et al., Systema: a framework for evaluating genetic perturbation response prediction beyond systematic variation, Nature Biotechnology 2025.

See the revised manuscript, Section 4.1, Tables 1–2, and Appendices C–D.

Reviewer_YCKJOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

CellCanon is introduced as an agentic model for single-cell expression prediction, evidence-grounded explanation, and observation-driven updating. Molecular evidence encodings and support atlas feed retrieval and gates fusion as well as response MLP predictors. The readouts guide probabilistic single-cell generation. These predictions are linked mechanism explainer and observation-driven update, through a world state. CellCanon was tested on a three cell-line SciPlex3 dataset and showed competitive performances against competing methods.

Critical strengths
  • The target is strong in achieving single-cell expression prediction, evidence-grounded explanation, and observation-driven updating.
  • The compared methods include SOTAs and recent methods; the performance improvement is robust across multiple evaluation metrics.
Critical weaknesses
  • The approach's innovation is not presented in a convincing way. Related methods are each briefly described but not in a direct and insightful way to contrast CellCanon; the impression of Related Works was limited to that those methods are all different compared to CellCanon. But the mentioned methods do include those that at least both predict and explain cell states upon perturbations and do include agents and world models. To what extent is CellCanon conceptually new remains unclear, in contrast to the claim of the first agentic world model to unify the three tasks. To what extent is CellCanon making a method contribution also remains unclear, despite the emphasis of a shared world state unifying the three tasks.

  • The SciPlex3 test set of 3 cell lines is relatively small and limited in scope. The splits are focused on drug-level splits but drugs can be similar and targeting the same or similar genes; splits based on drug similarity, target genes, and cell lines would further provide the stress test for incremental generalizability.

  • It is possible that the knowledge bases and the evidence curation present data leakage.

Questions for authors

Please see details in Weaknesses and kindly clarify the conceptual and algorithmic contributions, provide more evidence of the performance gains (bigger data or more critical splits), and address the data-leakage concern.

8 comments
AuthorsOpenReview ↗

Response to Reviewer YCKJ (Part 1/6):

Thank you for recognizing the value of bringing together single-cell expression prediction, evidence-grounded explanation, and observation-driven updating, and for highlighting the robustness of the gains across evaluation metrics.

To address your concerns, we added the following comparisons and experiments.

  • A comparison with prior reasoning agents and their token use, matched statistical and deep prediction baselines, and a test of probability-guided single-cell generation.
  • Blind tests of source and pathway verification against generic self-reflection, and a test of claim revision when observations or sources change.
  • Chemical and target-overlap checks for the original SciPlex3 split, a 30-cell-line Tahoe-100M benchmark, and a separate 20-drug blind confirmation.
  • Retrained holdout tests for new chemical scaffolds, disjoint drug targets, and unseen drugs and cell lines.
  • Fixed-weight updates with matched observations on SciPlex3 and with 1, 3, or 5 measured cell lines on Tahoe-100M.
  • Checks of source records and test-file access, retraining with verified sources, shuffled-response and shuffled-judgment controls, and replication with deterministic rules and an open-weight model.

1. Related methods are each briefly described but not in a direct and insightful way to contrast CellCanon … To what extent is CellCanon conceptually new remains unclear, in contrast to the claim of the first agentic world model to unify the three tasks.

CellCanon's contribution is an inspectable state that connects prediction, source verification, and updating for each drug and cell line. CellCanon generates single-cell expression profiles, checks biological claims against specified sources and abstains when evidence is insufficient, and uses new measurements to revise predictions without retraining. The state records predicted expression changes, the probabilities of differential expression and upregulation, source evidence, claims and their dependencies, and new measurements.

We removed the priority claim. VCWorld, VCR-Agent [1], and RBio1 [2] already connect perturbation prediction with mechanistic reasoning. We expanded the comparison to cover these methods and PerTurboAgent. We also clarified that observations follow a fixed schedule and that uncertainty refers to calibrated probabilities and the strength of supporting evidence.

Capabilities of perturbation models and reasoning agents

MethodSingle-cell expressionDifferential expression and directionExplanationUpdate
PRnet, PerturbNet, STATE, BioLORD, chemCPA, CMonge, SquidiffContinuous cell-level profilesPost hoc calls without explicit state or count calibration, with high false positives or low recallNot supportedGradient fine-tuning
VCWorldNot supportedGene-wise binary callsGenerated rationales without deterministic source verification or required abstentionStatic inference outputs
VCR-Agent [1]Not supportedDirection calls via verifier-filtered VC-TRACES and graph distillationVerifier filtering during synthesis on Tahoe-100MOffline synthesis and distillation
RBio1 [2]Not supportedCalibrated direction calls on PerturbQATraining rewards from biological world models and Gene OntologyFixed weights at inference
CellCanonProbability-guided single-cell profilesContinuous effects, calibrated DE and direction heads, and DE count decodingDeterministic verification of source support and effect signs, with mandatory abstentionCross-cell transfer of measured residuals without retraining, with invalidation of affected claims

[1] Jang et al., Towards Autonomous Mechanistic Reasoning in Virtual Cells, arXiv:2604.11661. The EMNLP Findings 2026 proceedings title is Towards Structured Mechanistic Reasoning in Single-cell Prediction.

[2] Istrate et al., RBio1: Training scientific reasoning LLMs with biological world models as soft verifiers, bioRxiv 2025.

See the revised manuscript, Abstract, Sections 1–2, 3.1 and 3.4, Figures 1 and 5, and Appendix L, Table 35.

AuthorsOpenReview ↗

Response to Reviewer YCKJ (Part 2/6):

2. To what extent is CellCanon making a method contribution also remains unclear, despite the emphasis of a shared world state unifying the three tasks.

CellCanon shares evidence across genes, verifies claims against their sources, and uses new measurements to revise both predictions and affected claims. We evaluated each of these operations directly.

Predict. CellCanon shares biological evidence across all genes for a given drug and cell line. Genes can be evaluated using common upstream targets and signaling pathways, with calibration that accounts for the sparsity of differential expression across the condition. Statistical tests on the generated cells recover differential expression and direction calls alongside continuous log fold changes and genome-wide perturbation rankings. VCWorld makes separate binary calls for each gene and used about 420 million tokens on SciPlex3. On the same DeepSeek-V4-Flash backbone, VCWorld reached 0.2481 DE precision and 0.2196 DE AP. CellCanon used 99.77% fewer tokens for prediction.

With the condition head held fixed, probability-guided generation improved post-generation DE Macro-F1 by 14.42 percentage points at essentially the same continuous RMSE, with Holm-adjusted p = 0.0005. For the submitted model, the guidance comparison raised DE Macro-F1 from 0.6506 to 0.8164. Dense DE calibration error was 0.0079 for CellCanon and 0.0106 to 0.0134 for the statistical baselines.

With identical 96-dimensional features, DE and direction supervision, count decoding, and single-cell generation, CellCanon reached 0.8067 DE AP and 0.7420 Joint F1. Its gains in those two metrics were 13.90 and 14.80 percentage points over the drug-agnostic prior, 7.60 and 6.71 over Linear, and 4.93 and 2.02 over knowledge-space kNN. The p-values were 0.0001 for the prior, at most 0.0004 for Linear, and 0.0001 for DE AP and 0.0247 for Joint F1 for kNN. Before generation, CellCanon reached 0.7242 Joint F1 versus 0.6359 for adapted BioLORD, the strongest adapted deep baseline.

Explain. Explain checks whether the cited source records and signed multi-step pathways support a claim. It returns “insufficient evidence” when incomplete or unresolved evidence prevents a conclusion. It also checks whether multiple unknown effect signs can be assigned consistently along a pathway, as in the TSA and HDAC1 example. On untouched held-out drugs, source and pathway verification reduced the rate of falsely supported claims from 8.02% to 0.00%, with Holm-adjusted p = 0.0036. Generic LLM self-reflection at the same budget reduced accuracy from 89.42% to 83.01%.

Update. Update uses new measurements to revise predictions for other cell lines with fixed weights. The shared state also records which claims depend on a prediction or source, so changes can invalidate the affected claims. Disabling Explain left the numerical Predict and Update metrics identical. On SciPlex3, CellCanon Update reached 0.7645 Joint F1 versus 0.6068 for PerturbNet fine-tuned on 48 observations. On Tahoe-100M, measurements from 1, 3, or 5 cell lines improved Joint F1 on the remaining 29, 27, or 25 lines by 5.91 to 10.47 percentage points over Predict and by 3.96 to 5.78 points over uniform residual shift from the same starting predictions. Across seven updates, dependency tracking reduced stale active claims from 148 to zero, LLM calls by 49.4%, and token use by 50.6%.

See the revised manuscript, Sections 3.2–3.4 and 4.2–4.4, Tables 1 and 6, and Appendices E, F, H, I, and K, including Tables 17, 19, 26, 27, and 34.

AuthorsOpenReview ↗

Response to Reviewer YCKJ (Part 3/6):

3. The SciPlex3 test set of 3 cell lines is relatively small and limited in scope.

We extended the evaluation to 30 human cancer cell lines and checked performance by chemical similarity and target overlap in the original SciPlex3 split. Tahoe-100M includes 180 compounds across 30 lines from 13 organs. We used 108 support drugs, 36 validation drugs, and 36 held-out test drugs, yielding 1,080 test conditions. All three models used the same features, validation budgets, DE and direction supervision, and single-cell generation procedure, with 823,038 cells generated per model. CellCanon led both baselines on sampled and dense DE Macro-F1 and Joint F1. Regularized Linear scored comparably on DE AP and continuous RMSE. Before generation, sampled DE AP was 0.9506 for CellCanon, 0.9481 for Linear, and 0.9305 for kNN.

Tahoe-100M single-cell Wilcoxon results after generation

ModelSampled DE APSampled DE F1Sampled direction F1Sampled Joint F1Dense DE F1Dense Joint F1Top 50 R²RMSE
CellCanon0.93400.74700.79730.65540.71640.59090.86735.4309
Linear0.93000.68880.80710.58840.70590.57490.87665.3974
kNN0.91430.66730.78890.57110.68640.55730.86885.5507

In a separate blind cohort of 20 drugs and 600 conditions, the frozen model reached 0.9379 DE AP, 0.9097 DE Macro-F1, and 0.7823 Joint Macro-F1 using the recall-calibrated single-cell Wilcoxon readout.

The original SciPlex3 split uses 54 support, 19 validation, and 112 test drugs. Against 10,000 random splits of the same size, the average Morgan Tanimoto similarity to the closest support drug was 0.2464, versus 0.2590 for random splits. The original split was at the 7.71st percentile for chemical similarity and at the 2.15th percentile for target overlap among annotation-matched random splits.

Among the 112 test drugs, 88 had a maximum similarity of 0.30 or less to the support set and covered 267 conditions. On this subset, CellCanon reached 0.7423 DE AP and 0.6799 Joint F1 after generation. It exceeded matched Linear by 8.44 percentage points in DE AP and 4.88 points in Joint F1, and exceeded matched kNN by 5.43 points in DE AP. The 95% confidence intervals were 4.92 to 11.57, 0.32 to 8.77, and 3.02 to 8.19 points. The corresponding p-values were 0.0001, 0.0146, and 0.0001.

On the 32 test drugs with no target overlap with the support set, covering 99 conditions, CellCanon led Linear and kNN by 7.25 and 7.58 points in DE AP. The 95% confidence intervals were 1.22 to 13.32 and 4.36 to 13.39 points, with p = 0.0177 and p = 0.0269. Joint F1 was comparable (−0.02 points versus Linear and −0.49 points versus kNN).

See the revised manuscript, Appendix A for the SciPlex3 split, and Appendix K, Table 33, for the Tahoe-100M evaluation and blind confirmation.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.