CELLULAR PHENOTYPES · MODEL DEVELOPMENT

CellScientist.

Model Revision by
Diagnostic Routing
for Morphological Perturbation Prediction

Turn execution and validation feedback into a decision about what to change next.

Read the paper Code Data

The BBBC047 trajectory includes improvements and regressions, retaining Ref-RGAF at validation PCC 0.4368. — **Progress has a history.** The retained model follows both improvements and regressions. This case reports search-used validation PCC.

Mengran Li^1,2,*Bo Li^3,*Jiaying Wang² Wenbin Xing² Chengyang Zhang⁴ Jinlin Wu¹ Zhen Lei¹ Jiebo Luo¹ Stan Z. Li⁵ Zelin Zang^1,†

* Equal contribution · † Corresponding authorAffiliations +

Centre for Artificial Intelligence and Robotics, Hong Kong Institute of Science & Innovation, Chinese Academy of Sciences
Sun Yat-sen University
University of Macau
Sichuan University
Westlake University

01 / THE IDEA

A score tells you how well.
Diagnostics guide what comes next.

Predicting how a chemical perturbation changes cellular morphology connects molecular interventions to measurable phenotypes. A useful model must coordinate chemical representation, dose, cellular context, and response modeling.

A disappointing score leaves a design decision unresolved. CellScientist connects diagnostics and earlier attempts to local model revisions, keeping the prediction task and evaluation protocol consistent throughout the search.

CellScientist connects a structured coupled design space to local revision and a complete recorded trajectory. — Every edit belongs to a design state. Every outcome becomes evidence for the next decision.

HRT

Keep the evidence.

History-Aware Revision Tracking connects design states to earlier attempts, including failures and regressions.

LCA

Make the edit executable.

Limited-Change Application checks candidate eligibility and applies bounded repair before evaluation.

PDR

Choose where to revise.

Performance-Discrepancy Refinement maps observed discrepancies to model components for local revision.

Full CellScientist framework linking the task contract, LLM proposals, eligibility checks, validation diagnostics, revision targets and recorded outcomes. — From a model-design state to an executable candidate, and from its outcome to the next edit.

02 / CONTROLLED EVIDENCE

Better revisions.
Fewer candidate evaluations.

A matched comparison isolates the revision policy within a shared 768-candidate language. Four task/split settings, five paired seeds, and a reporting fold held out from search.

Candidate-evaluation budget

CELLSCIENTIST − RANDOM ROUTING

+0.0373

mean held-out PCC advantage

95% confidence interval[0.0292, 0.0454]

A larger mean advantage appears at small evaluation budgets, where each revision has more influence on the search.

Paired-setting t-interval across four settings · Table 3

Held-out prediction quality

Global PCC ↑

CellScientistRandom routing

BBBC036 Plate split

0.2674

0.2277

BBBC036 SMILES split

0.2346

0.2049

BBBC047 Plate split

0.3457

0.3066

BBBC047 SMILES split

0.3614

0.3207

0.000.100.200.300.40

Bar length: five-seed mean. Whiskers: ±1 sample standard deviation.

The budget counts evaluated candidates, including the shared initial candidate. At ten evaluations, the mean advantage narrows and its confidence interval includes zero.

View all five policies and exact values

Held-out global PCC, mean ± sample standard deviation over five paired seeds. Budget: 5 evaluations.
Task / split	Fixed initial model	Flat retry	Random routing	AIDE-style	CellScientist
BBBC036 / Plate	0.1664 ± 0.0005	0.2073 ± 0.0056	0.2277 ± 0.0502	0.2477 ± 0.0294	0.2674 ± 0.0003
BBBC036 / SMILES	0.1547 ± 0.0005	0.1882 ± 0.0129	0.2049 ± 0.0417	0.2231 ± 0.0169	0.2346 ± 0.0009
BBBC047 / Plate	0.2442 ± 0.0008	0.2630 ± 0.0130	0.3066 ± 0.0514	0.3153 ± 0.0399	0.3457 ± 0.0004
BBBC047 / SMILES	0.2504 ± 0.0004	0.2916 ± 0.0003	0.3207 ± 0.0582	0.3371 ± 0.0417	0.3614 ± 0.0003

Download chart data (JSON) ↗

03 / MODELS & TRAJECTORIES

See the predictions.
Follow the decisions.

Selected-model benchmarks and open search trajectories show how the workflow develops task-specific predictors. BBBC036, BBBC047, and CPG0016 form the morphology benchmark suite; transcriptomic and temporal tasks extend the evaluation.

BBBC047 five-fold R-squared distributions comparing CS-model, RealMLP, RF-TD, and TabR. — **Prediction quality across folds.** R² distributions compare CS-model with fixed reference predictors; annotations follow the paper.

BBBC047 candidate validation PCC and best-so-far scores across ten model-revision iterations. — **Improvements, regressions, and retention.** The best-so-far curve preserves a useful candidate even when a later experiment underperforms.

The BBBC047 trajectory reports validation scores used during search. The controlled comparison above uses a separate held-out reporting fold.

04 / INSIDE THE WORKFLOW

Inspect the components,
as well as the final score.

Registered component audits examine what history, candidate checks, and localized repair contribute to the workflow.

Repeated proposals with HRT

170 unique proposals in 170 evaluated steps. Without persistent history: 82 unique proposals in 180 steps.

20 matched controlled runs

40/40

Registered violations handled

LCA repairs 15 cases and blocks 25 others. All 20 valid controls are accepted.

60 registered contract cases

15/15

Held-out faults repaired

Structured PDR repairs all registered held-out faults in one round on average, with no unrelated edits.

15 faults across five component addresses

05 / CODE, DATA & PAPER

Reproduce the workflow.
Build on the revision process.

START FROM THE REPOSITORY

git clone https://github.com/limengran98/CellScientist.git
cd CellScientist
python -m venv .venv
source .venv/bin/activate
python -m pip install -e ".[data]"
python -m cellscientist --help

Full setup, including Windows instructions ↗

GITHUBCode & reproducibility guidesConfigurations, audits, recorded outputs, and extension points.HUGGING FACECellScientist dataPreprocessed morphology benchmarks and documented inputs.MANUSCRIPTRead the complete paperMethod, experimental protocols, full results, and appendices.

The paper uses Gemini 3 Pro, with Qwen2.5-0.5B-Instruct for the open-weight controlled audit. Other LLM APIs can be connected through the OpenAI-compatible interface. LLM configuration guide ↗

CITATION

Reference this work

BIBTEX

@misc{li2026cellscientist,
  title = {CellScientist: Model Revision by Diagnostic Routing
           for Morphological Perturbation Prediction},
  author = {Li, Mengran and Li, Bo and Wang, Jiaying and
            Xing, Wenbin and Zhang, Chengyang and Wu, Jinlin and
            Lei, Zhen and Luo, Jiebo and Li, Stan Z. and Zang, Zelin},
  year = {2026},
  howpublished = {Preprint},
  url = {https://github.com/limengran98/CellScientist}
}