CellCanon links perturbation prediction, sourced explanation and observation-driven updating in one world state, letting a single new measurement revise unmeasured cell lines
Synopsis
Single-cell perturbation screens still leave many compounds and cell contexts unmeasured, and CellCanon puts expression prediction, source-grounded explanation and observation-driven updating into one shared world state. A language model compiles drug and target evidence, and numerical readouts then guide single-cell generation; on 339 held-out SciPlex3 conditions it reaches differential-expression AUPRC 0.8067 and Macro-F1 0.8164, against 0.4938 and 0.6961 for the best baseline. In simulated experimental feedback, one measured source context per drug lifts differential-expression AUPRC to 0.9077 for unmeasured cell lines with a fixed-weight updater.
Interpretation
CellCanon lets one condition-level body of biological evidence serve continuous response, differential expression and direction readouts at once, and records predictions, source-linked claims and new measurements in a single world state. Earlier numerical perturbation models output expression profiles only, while gene-wise language reasoning infers response labels separately, so the two do not share an inspectable evidence chain. On 339 held-out SciPlex3 conditions across A549, K562 and MCF7, CellCanon reaches differential-expression AUPRC 0.8067 and Macro-F1 0.8164 versus 0.4938 and 0.6961 for the strongest baselines, and direction Macro-F1 0.8472 versus 0.7903.
The prediction path has a language model select source-linked target evidence per condition, retrieves similar drugs from a training response atlas, and uses cell-specific readouts to produce gene-level outputs that guide single-cell generation. Evidence compilation happens once per condition and is reused across genes rather than triggering a separate language-model call for each gene. Against VCWorld's gene-wise prediction and chain-of-thought explanation on matched drug–cell cohorts, language-model token use falls by 99.77% for Predict across 339 conditions and 98.41% for Explain across 72 conditions.
Update uses one measured source context's response to revise predictions for the same drug in an unmeasured cell line, while the parent predictor and update maps keep their weights fixed. The comparison is six response models fine-tuned on the revealed source-cell populations, which need gradient updates to absorb new observations. With one source context per drug, CellCanon reaches differential-expression AUPRC/Macro-F1 of 0.9077/0.8261 versus 0.6279/0.7630 for the strongest fine-tuned comparator, PerturbNet, and joint three-state Macro-F1 of 0.7431 versus 0.6116.
Perspective
The evaluation covers 339 held-out SciPlex3 conditions in A549, K562 and MCF7 at 24 hours with pooled doses; continuous metrics span 2,000 genes and the classification panel adds 931 LINCS landmark genes. Explanation is assessed on a source-grounded benchmark of 966 questions across 72 conditions, and updating is tested by revealing existing SciPlex3 measurements under fixed acquisition schedules to simulate experimental feedback.
The update experiment is a retrospective simulation: target conditions stay unobserved and source selections are fixed before adaptation, so it reflects performance as existing measurements are progressively revealed. Differential-expression Macro-F1 rises from 0.8199 to 0.8393 at 48 observations, yet differential-expression AP at 224 observations is below the 112-observation value, so more feedback does not improve every metric monotonically. Reference judgments in the explanation benchmark are defined by molecular source records, and runtime correction was not given the reference answers.