CellCanon: An Agentic Cellular World Model for Prediction, Explanation, and Observation-Driven Updating
Organizations: 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. 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.
Figures & tables
| Metric | BioLORD | chemCPA | CMonge | PRnet | Squidiff | PerturbNet | STATE | VCWorld | Random | CellCanon |
|---|---|---|---|---|---|---|---|---|---|---|
| DE | ||||||||||
| AUPRC | 0.2694 | 0.1345 | 0.3474 | 0.4058 | 0.0875 | 0.4870 | 0.4255 | 0.4938 | 0.3666 | 0.8067 |
| AUROC | 0.7734 | 0.4536 | 0.8140 | 0.8572 | 0.1279 | 0.8807 | 0.8371 | 0.6527 | 0.4998 | 0.9449 |
| F1 | 0.6184 | 0.4422 | 0.6479 | 0.6961 | 0.1086 | 0.6732 | 0.6040 | 0.5500 | 0.4324 | 0.8164 |
| Acc. | 0.6886 | 0.4616 | 0.7402 | 0.7698 | 0.1144 | 0.8453 | 0.8275 | 0.6259 | 0.4996 | 0.9163 |
| Precision | 0.3125 | 0.2092 | 0.3432 | 0.3941 | 0.1056 | 0.5060 | 0.4093 | 0.2481 | 0.1555 | 0.8370 |
| Metric | BioLORD | chemCPA | CMonge | PRnet | Squidiff | PerturbNet | STATE | DMSO reference | CellCanon |
|---|---|---|---|---|---|---|---|---|---|
| Top50 R 2 | 0.9197 | 0.9026 | 0.7359 | 0.9039 | 0.8590 | 0.4461 | 0.9063 | 0.8880 | 0.9145 |
| logFC Pearson | 0.2935 | 0.4306 | 0.2392 | 0.2061 | 0.2034 | 0.4268 | 0.3046 | – | 0.6944 |
| -Pearson | 0.2538 | 0.2543 | 0.1484 | 0.1528 | 0.0703 | 0.1500 | 0.2401 | – | 0.3254 |
| Top100 -Spearman | 0.3199 | 0.3632 | 0.1822 | 0.2089 | 0.2396 | 0.2062 | 0.2809 | – | 0.3809 |
| Metric | BioLORD | chemCPA | CMonge | PRnet | PerturbNet | STATE | CellCanon |
|---|---|---|---|---|---|---|---|
| DE | |||||||
| AUPRC | 0.1337 | 0.3188 | 0.1306 | 0.0948 | 0.6279 | 0.5535 | 0.9077 |
| AUROC | 0.3925 | 0.8069 | 0.3350 | 0.1467 | 0.9135 | 0.8663 | 0.9792 |
| F1 | 0.2627 | 0.5786 | 0.1965 | 0.1241 | 0.7630 | 0.7003 | 0.8261 |
| Acc. | 0.2680 | 0.7199 | 0.1976 | 0.1338 | 0.8637 | 0.8555 | 0.9185 |
| Precision | 0.1236 | 0.2767 | 0.1023 | 0.1271 | 0.5846 | 0.5925 | 0.9021 |
| Endpoint | Correct / total |
|---|---|
| Label + complete proof | 561 / 561 |
| Correct abstention | 405 / 405 |
| Valid source path@3 | 63 / 63 |
| Biological annotations | DE F1 | DIR F1 |
|---|---|---|
| Full predictor | 0.8077 | 0.8393 |
| Without annotations | 0.7117 | 0.7900 |
| Measured context | -Pearson | -Spearman |
| With DMSO features | 0.3074 | 0.2383 |
| Without DMSO features | 0.2810 | 0.2292 |
| Task readouts | DE F1 | DIR F1 |
Appendix figures & tables23 assets
Supplementary material from the paper’s appendix.
Appendix
| Output | RMSE | Pearson | DE AUPRC | DE F1 | DIR F1 |
|---|---|---|---|---|---|
| Shared condition heads | 4.4762 | 0.6943 | 0.8588 | 0.8077 | 0.8393 |
| Control-conditioned expression | 4.4770 | 0.6942 | 0.7982 | 0.6506 | 0.8203 |
| Probability-guided expression | 4.4755 | 0.6944 | 0.8067 | 0.8164 | 0.8472 |
| Task | AUPRC | AUROC | Macro-F1 |
|---|---|---|---|
| DE | |||
| DIR |
| Method | LogFC RMSE |
|---|---|
| BioLORD | 11.7617 |
| chemCPA | 9.1093 |
| CMonge | 15.1668 |
| PRnet | 15.6431 |
| Squidiff | 18.5063 |
| PerturbNet | 6.7180 |
| Method | All-gene -Pearson | All-gene -Spearman | Top100 -Pearson | Top100 -Spearman |
|---|---|---|---|---|
| BioLORD | 0.2538 | 0.1298 | 0.3618 | 0.3199 |
| chemCPA | 0.2543 | 0.1534 | 0.3888 | 0.3632 |
| CMonge | 0.1484 | 0.0900 | 0.1916 | 0.1822 |
| PRnet | 0.1528 | 0.0859 | 0.2357 | 0.2089 |
| Squidiff | 0.0703 | 0.0328 | 0.2719 | 0.2396 |
| PerturbNet | 0.1500 | 0.1085 | 0.2405 | 0.2062 |
| Variant | RMSE | DE AP | DE F1 | DIR mAP | DIR F1 | Joint F1 |
|---|---|---|---|---|---|---|
| Condition heads | 4.4762 | 0.8588 | 0.8077 | 0.9135 | 0.8393 | 0.7242 |
| Without biological knowledge | 4.4699 | 0.7922 | 0.7117 | 0.8593 | 0.7900 | 0.5776 |
| Without text embeddings | 4.4688 | 0.8640 | 0.8066 | 0.9102 | 0.8370 | 0.7228 |
| Without measured context | 4.4781 | 0.8611 | 0.8118 | 0.9093 | 0.8360 | 0.7281 |
| Variant | Task | AP | PR area | AUROC | F1 | Accuracy | Precision | Recall |
|---|---|---|---|---|---|---|---|---|
| Condition heads | DE | 0.8588 | 0.8588 | 0.9676 | 0.8077 | 0.9163 | 0.8876 | 0.5294 |
| DIR | 0.9292 | 0.9291 | 0.9156 | 0.8393 | 0.8398 | 0.8737 | 0.8233 | |
| Without biological knowledge | DE | 0.7922 | 0.7921 | 0.9546 | 0.7117 | 0.8875 | 0.8396 | 0.3425 |
| DIR | 0.8905 | 0.8904 | 0.8695 | 0.7900 | 0.7900 | 0.8616 | 0.7295 | |
| Without text embeddings | DE | 0.8640 | 0.8640 | 0.9676 | 0.8066 | 0.9162 | 0.8920 | 0.5251 |
| DIR | 0.9278 | 0.9278 | 0.9131 | 0.8370 | 0.8375 | 0.8714 | 0.8212 |
| Gene panel | Metric | Condition heads | No context | Difference [95% interval] |
|---|---|---|---|---|
| All 2,000 | -Pearson | 0.3074 | 0.2810 | -0.0265 [-0.0410, -0.0130] |
| All 2,000 | -Spearman | 0.2383 | 0.2292 | -0.0091 [-0.0144, -0.0039] |
| Top20 | -Pearson | 0.4264 | 0.4030 | -0.0235 [-0.0423, -0.0062] |
| Top20 | -Spearman | 0.3894 | 0.3628 | -0.0265 [-0.0471, -0.0081] |
| Top50 | -Pearson | 0.3961 | 0.3733 | -0.0228 [-0.0370, -0.0093] |
| Top50 | -Spearman | 0.3700 | 0.3420 | -0.0280 [-0.0433, -0.0133] |
| Readout and DE decision rule | DE AUPRC | DE F1 | DIR mAP | DIR F1 | Joint F1 |
|---|---|---|---|---|---|
| Independent heads + count decoder | 0.8588 | 0.8077 | 0.9135 | 0.8393 | 0.7242 |
| Continuous scores + count decoder | 0.1429 | 0.4538 | 0.8901 | 0.8134 | 0.3029 |
| Continuous scores + fixed thresholds | 0.1429 | 0.5278 | 0.8901 | 0.8134 | 0.4361 |
| Comparison | Metric | Difference | Lower | Upper |
|---|---|---|---|---|
| Without biological knowledge | DE AP | -0.066687 | -0.102294 | -0.011355 |
| Without biological knowledge | DE F1 | -0.096053 | -0.152370 | -0.019405 |
| Without biological knowledge | DIR mAP | -0.054199 | -0.067440 | -0.036245 |
| Without biological knowledge | DIR F1 | -0.049380 | -0.068552 | -0.030069 |
| Without biological knowledge | Joint F1 | -0.146592 | -0.214128 | -0.053563 |
| Without text embeddings | DIR mAP | -0.003370 | -0.006790 | -0.000359 |
| Variant | RMSE | DE AP | DE F1 | DIR mAP | DIR F1 | Joint F1 |
|---|---|---|---|---|---|---|
| Raw-source knowledge | 4.4723 | 0.8661 | 0.8136 | 0.9065 | 0.8319 | 0.7339 |
| LLM-compiled knowledge | 4.4762 | 0.8588 | 0.8077 | 0.9135 | 0.8393 | 0.7242 |
| Method | Label accuracy | Label + proof | Correct insufficiency |
|---|---|---|---|
| TF–IDF retrieval + rules | 0.5989 | 0.5989 | 405 / 405 |
| Complete-source symbolic reference | 1.0000 | 1.0000 | 405 / 405 |
| CellCanon | 1.0000 | 1.0000 | 405 / 405 |
| Observation budget | MSE | DE AP | DE AUROC | DE F1 | DIR F1 |
|---|---|---|---|---|---|
| No update | 20.4360 | 0.8806 | 0.9710 | 0.8199 | 0.8402 |
| 12 observations | 20.3889 | 0.8833 | 0.9715 | 0.8210 | 0.8419 |
| 24 observations | 20.2662 | 0.8893 | 0.9723 | 0.8242 | 0.8410 |
| 48 observations | 20.1772 | 0.8924 | 0.9726 | 0.8393 | 0.8425 |
| 112 observations | 20.0427 | 0.9077 | 0.9792 | 0.8261 | 0.8453 |
| 224 observations | 20.0447 | 0.9030 | 0.9777 | 0.8583 | 0.8445 |
| Budget | Task | AP | AUROC | F1 | Accuracy | Precision | Recall |
|---|---|---|---|---|---|---|---|
| 0 | DE | 0.8806 | 0.9710 | 0.8199 | 0.9162 | 0.9013 | 0.5570 |
| 0 | DIR | 0.9323 | 0.9179 | 0.8402 | 0.8408 | 0.8845 | 0.8167 |
| 48 | DE | 0.8924 | 0.9726 | 0.8393 | 0.9230 | 0.8975 | 0.6064 |
| 48 | DIR | 0.9305 | 0.9157 | 0.8425 | 0.8432 | 0.8838 | 0.8224 |
| 112 | DE | 0.9077 | 0.9792 | 0.8261 | 0.9185 | 0.9021 | 0.5715 |
| 112 | DIR | 0.9313 | 0.9167 | 0.8453 | 0.8460 | 0.8840 | 0.8283 |
| Method | Learning rate | Updates |
|---|---|---|
| STATE | 400 | |
| PRnet | 400 | |
| chemCPA | 400 | |
| PerturbNet | 400 | |
| CMonge | 400 | |
| BioLORD | 100 |
| Method | DE AUPRC | DE F1 | DIR F1 | DE Acc. |
|---|---|---|---|---|
| Fine-tuning: generated-expression readouts | ||||
| STATE | +0.0872 | +0.0808 | +0.0716 | +0.0297 |
| PRnet | +0.0001 | +0.0065 | +0.0552 | +0.0076 |
| chemCPA | -0.0014 | +0.0865 | -0.0785 | -0.0346 |
| PerturbNet | +0.1023 | +0.0839 | +0.0557 | +0.0229 |
| CMonge | -0.0031 | -0.0064 | -0.0042 | -0.0056 |
| Method | Sources | DE AUPRC | DE F1 | DE Acc. | DIR F1 | DIR Acc. |
|---|---|---|---|---|---|---|
| Fine-tuning: generated-expression readouts | ||||||
| STATE | None | 0.4663 | 0.6195 | 0.8259 | 0.7606 | 0.7613 |
| 48 | 0.5314 | 0.6702 | 0.8434 | 0.7912 | 0.7930 | |
| One source/drug | 0.5535 | 0.7003 | 0.8555 | 0.8323 | 0.8329 | |
| PRnet | None | 0.0947 | 0.1176 | 0.1262 | 0.5936 | 0.6566 |
| 48 | 0.0945 | 0.1216 | 0.1310 | 0.5441 | 0.6282 | |
| Method | Sources | Top50 | logFC Pearson | All 2,000 -Pearson | Top100 -Spearman |
|---|---|---|---|---|---|
| STATE | None | 0.8998 | 0.3045 | 0.2417 | 0.2800 |
| 48 | 0.9146 | 0.3144 | 0.2758 | 0.3115 | |
| One source/drug | 0.9230 | 0.3214 | 0.3017 | 0.3384 | |
| PRnet | None | 0.8475 | 0.2386 | 0.0977 | 0.2149 |
| 48 | 0.8535 | 0.2394 | 0.0941 | 0.2045 | |
| One source/drug | 0.1368 | 0.2411 | 0.0724 | 0.1183 |
| Method / stage | Input | Output | Total | Saved |
|---|---|---|---|---|
| Predict: 339 conditions; 55,598 reference gene queries | ||||
| VCWorld | 310,614,134 | 109,667,313 | 420,281,447 | – |
| CellCanon Predict | 882,462 | 99,131 | 981,593 | 99.77% |
| Explain: 72 conditions; 12,399 reference gene queries | ||||
| VCWorld | 69,060,037 | 24,334,813 | 93,394,850 | – |
| CellCanon Explain | 1,347,957 | 141,533 | 1,489,490 | 98.41% |
| Stage | Conditions | Input | Output | Total | Per condition |
|---|---|---|---|---|---|
| Predict: test | 339 | 882,462 | 99,131 | 981,593 | 2,896 |
| Predict: all splits | 558 | 1,403,786 | 155,684 | 1,559,470 | 2,795 |
| Explain: evaluated cohort | 72 | 1,347,957 | 141,533 | 1,489,490 | 20,687 |
| Panel | Task | AUPRC | AUROC | F1 | Accuracy | Precision | Recall |
|---|---|---|---|---|---|---|---|
| Additional 931 | DE | 0.6842 | 0.8777 | 0.7466 | 0.8957 | 0.8082 | 0.4194 |
| Additional 931 | DIR | 0.8879 | 0.9051 | 0.8030 | 0.8144 | 0.8714 | 0.6671 |
| All 2,931 | DE | 0.7710 | 0.9280 | 0.7965 | 0.9100 | 0.8298 | 0.5270 |
| All 2,931 | DIR | 0.9254 | 0.9186 | 0.8377 | 0.8378 | 0.8692 | 0.8012 |
| Budget | Task | AP | AUROC | F1 | Accuracy | Precision | Recall |
|---|---|---|---|---|---|---|---|
| 0 | DE | 0.8179 | 0.9357 | 0.8136 | 0.9051 | 0.8581 | 0.5683 |
| 0 | DIR | 0.9157 | 0.9488 | 0.8599 | 0.8676 | 0.8194 | 0.8376 |
| 48 | DE | 0.8420 | 0.9392 | 0.8430 | 0.9166 | 0.8641 | 0.6421 |
| 48 | DIR | 0.9163 | 0.9490 | 0.8651 | 0.8724 | 0.8235 | 0.8472 |
| 112 | DE | 0.8618 | 0.9553 | 0.8182 | 0.9074 | 0.8685 | 0.5748 |
| 112 | DIR | 0.9160 | 0.9490 | 0.8644 | 0.8714 | 0.8210 | 0.8488 |
| Budget | Task | AP | AUROC | F1 | Accuracy | Precision | Recall |
|---|---|---|---|---|---|---|---|
| 0 | DE | 0.8618 | 0.9615 | 0.8179 | 0.9127 | 0.8867 | 0.5606 |
| 0 | DIR | 0.9274 | 0.9284 | 0.8491 | 0.8495 | 0.8665 | 0.8219 |
| 48 | DE | 0.8774 | 0.9635 | 0.8405 | 0.9209 | 0.8859 | 0.6176 |
| 48 | DIR | 0.9264 | 0.9273 | 0.8523 | 0.8527 | 0.8671 | 0.8286 |
| 112 | DE | 0.8944 | 0.9729 | 0.8235 | 0.9150 | 0.8910 | 0.5723 |
| 112 | DIR | 0.9271 | 0.9282 | 0.8540 | 0.8543 | 0.8665 | 0.8335 |
| Variant | RMSE | DE AP | DE F1 | DIR mAP | DIR F1 | Joint F1 |
|---|---|---|---|---|---|---|
| Complete model | 4.4762 | 0.8588 | 0.8077 | 0.9135 | 0.8393 | 0.7242 |
| Full atlas, core-only heads | 4.4762 | 0.8588 | 0.8077 | 0.9135 | 0.8390 | 0.7242 |
| Core-only atlas and heads | 4.4715 | 0.8584 | 0.8056 | 0.9102 | 0.8326 | 0.7216 |
Single-Cell Perturbation Modeling
Single-cell perturbation modeling predicts how genetic or chemical interventions change cellular states across cells and contexts.
See all 70 papers