CellCanon links cell-response predictions to evidence—and revises them as observations arrive

A prediction that can be revised

Single-cell experiments measure how individual cells change their gene activity after a drug treatment. But even large screens leave many combinations of drugs and cell types untested. Computational models try to fill those gaps; the harder challenge is keeping their predictions connected to evidence that researchers can inspect and revise.

In a November 2026 preprint, Zhaokang Liang, Shuyang Zhuang, Xiaoran Jiao and Hao Chen introduce CellCanon, a system combining response prediction, evidence-grounded explanation and updating from newly available observations. Their “cellular world model” is not a complete simulation of a living cell. It models gene-expression responses and maintains records linking predictions, biological claims and measurements.

Share the evidence, specialize the predictions

CellCanon separates interpreting biological evidence from calculating each gene’s response. A language model compiles drug and cell-context information once for a treatment condition. Numerical models then reuse that information across genes, alongside an atlas of measured training responses.

Separate prediction components estimate response magnitude, whether a gene changes significantly, and whether that change is upward or downward. Those outputs guide the generation of modeled single-cell expression profiles using matched solvent-control cells. The generated profiles are predictions, not additional laboratory measurements.

This division also reduces repeated language-model work. On matched evaluation cohorts, the authors measured 99.77% lower token use for prediction and 98.41% lower token use for explanation than VCWorld’s gene-by-gene approach. These are token-use comparisons, not measurements of total computing cost, energy consumption or laboratory savings.

Stronger scores, with a sensitivity tradeoff

The authors evaluated prediction on 339 held-out SciPlex3 conditions across A549, K562 and MCF7 cell lines, withholding test drugs from training. Responses were assessed at 24 hours, pooling information across doses. Continuous-expression evaluation covered 2,000 genes.

For identifying differentially expressed genes—genes whose activity changes significantly relative to controls—CellCanon achieved a precision–recall area score of 0.8067 and a Macro-F1 score of 0.8164. The strongest comparator scores for those respective metrics were 0.4938 and 0.6961. These correspond to the authors’ reported relative improvements of 63.3% and 17.3%, not percentage-point increases.

The gains do not mean the system detected every response. Its reported precision was 0.8370, but recall was 0.5739: predictions were comparatively selective, and the authors note lower sensitivity than PRnet. CellCanon also improved response-correlation measures, although it did not lead every expression metric. Classification scores used all reference-positive genes plus sampled negatives, rather than an unrestricted whole-transcriptome evaluation.

Explanations that distinguish support from uncertainty

CellCanon’s explanation component checks whether supplied molecular records support, contradict or leave insufficient evidence for a claim. It distinguishes drug binding from functional action, and RNA detection in untreated controls from direct evidence that a drug engages its target in treated cells.

On a source-grounded benchmark of 966 questions across 72 conditions, the authors report correct judgments with complete supporting proofs for all 561 answerable questions, plus correct abstention on all 405 insufficient-evidence questions.

That perfect score has a narrow meaning. The questions used reusable templates and source-defined rules; some question strings overlapped with validation, and evaluation-panel error analysis informed refinement of the correction procedure. A complete-source symbolic reference also scored perfectly. These results demonstrate performance on structured evidence checking, not universal reliability in explaining biology. In a separate source-withdrawal challenge, CellCanon correctly revised 14 of 16 claims—not all of them.

Updating without a new laboratory campaign

To test updating, the researchers progressively revealed existing SciPlex3 measurements, simulating experimental feedback. A measured response for one cell context was used to revise predictions for the same drug in another, still-hidden context. The predictor and update-model weights remained fixed.

With one observed source context per drug, the authors report differential-expression precision–recall area and Macro-F1 scores of 0.9077 and 0.8261, versus 0.6279 and 0.7630 for the strongest fine-tuned comparators. However, CellCanon’s update scores came from condition-level prediction components, while comparators were evaluated through statistical testing of generated expression profiles—a methodological distinction when interpreting the comparison.

What remains unproven

The study supports a workflow for combining predictions, evidence checks and revisions within this dataset. It does not establish prospective laboratory discovery, clinical usefulness or generalization beyond the tested cell lines, timing and perturbation setting. Source-supported regulatory paths remain candidate explanations, not experimentally confirmed causal mechanisms. The paper is a preprint; these findings should be read as the authors’ reported results, not as established validation.

The supplied PDF extraction disrupts tables and formulae and does not preserve the figures reliably. This account therefore relies on explicit written results and labeled descriptions rather than reconstructing missing visual evidence.