CellScientist: Dual-Space Hierarchical Orchestration for Closed-Loop Refinement of Virtual Cell Models

Mengran Li, Bo Li, Jiaying Wang, Wenbin Xing, Yixuan Dong, Chengyang Zhang, Hongliang Zhang, Yuzhong Peng, Jinlin Wu, Bob Zhang, Bingo Wing-Kuen Ling, Fuji Yang, Zhen Lei, Jiebo Luo, Zelin ZangarXiv 2026 · DOI

TL;DR

This paper introduces CellScientist, a novel framework for refining virtual cell models (VCMs). It tackles the challenge of effectively updating complex biological models by creating a closed loop between high-level hypotheses and low-level executable code, enabling more robust and auditable model development.

What Problem Does This Paper Address?

Virtual Cell Modeling (VCM) aims to predict how cells respond to various interventions, like drug treatments. However, building these models is tricky because predictions often fail, and it's hard to pinpoint why. The failure could stem from a flawed assumption about cell biology, a poor design choice in how the model represents data, an implementation bug, or even incorrect task constraints. Existing AI-driven workflows often follow a linear "build-test-fix" approach, where feedback from failed tests is too general to guide specific revisions. This leads to a "refinement-routing problem": how do we effectively route these prediction discrepancies back to the right level of the model for correction? CellScientist addresses this by introducing a structured, hierarchical approach that connects the abstract reasoning about biological hypotheses with the concrete implementation of those hypotheses.

Methods

CellScientist employs a "dual-space hierarchical orchestration" framework. It operates in two interconnected spaces:

  • Hypothesis Space (H): This high-level space represents biological and modeling assumptions as a structured, dynamic typed hypergraph. Nodes in this graph represent specific modeling decisions (e.g., how to represent perturbations, what architecture to use), and hyperedges capture dependencies between these decisions. This hierarchical structure helps in localizing failures.
  • Implementation Space (I): This low-level space contains the actual executable code that realizes the hypotheses. The framework ensures that proposed implementations are "admissible" – meaning they satisfy task, interface, and domain constraints.

The core of CellScientist is a closed "Hypothesis → Implementation → Hypothesis" loop:

  1. Hypothesis to Implementation: A hypothesis (h) guides the generation of an executable candidate (c) through a process called Logic-Constrained Alignment (LCA). LCA checks for structural validity and repairs local violations.
  2. Execution and Feedback: The executable candidate (c) is run on validation data.
  3. Feedback to Hypothesis: Performance discrepancies (m) are collected and then routed by the Performance-Discrepancy Refinement (PDR) module to a specific address (at) within the hypothesis topology. This feedback then triggers a local edit (dt) to update the hypothesis (h) and topology (G).

This structured feedback loop allows for targeted revisions, moving beyond simple code repair to actual hypothesis refinement.

Datasets

CellScientist was evaluated on several benchmark datasets:

  • Cell Painting Morphology Prediction: BBBC021, BBBC036, BBBC047, and CPG0016. These datasets represent cellular phenotypes under various chemical perturbations.
  • LINCS2020 Transcriptomic Perturbation Prediction: This dataset captures gene-expression responses to perturbations. Two variants were used: "Seven Cell Lines" (dense, chemical-blind split) and "Full Data" (sparse, long-tail distribution).
  • Single-cell Perturbation Prediction: Norman (CRISPR knockouts), Schiebinger (cytokine perturbations), and Papalexi RNA/Protein (joint RNA and protein responses).

Key Insights and Ideas

  • Structured Refinement is Crucial: Simply fixing code when predictions fail isn't enough for complex VCM. CellScientist shows that routing feedback to specific modeling assumptions is key to robust improvement.
  • Hypothesis-Implementation Alignment: Keeping abstract hypotheses and their concrete code implementations tightly coupled prevents the model from drifting away from its intended biological reasoning.
  • Auditable Design Trajectories: CellScientist doesn't just find a good model; it records the entire process—including failed attempts, regressions, and successful revisions. This creates a traceable history, making the model development process transparent and understandable.
  • Hierarchical Reasoning for Localization: The hypergraph structure of the hypothesis space allows the system to pinpoint where a failure likely originated, enabling more efficient and targeted revisions.

Key Results

  • Improved Predictive Performance: The final executable models selected by CellScientist consistently outperformed baseline methods across morphology, transcriptomic, and single-cell perturbation prediction benchmarks. For example, on the BBBC047 dataset, the CS-model achieved a PCC of 0.4368, outperforming previous methods.
  • Enhanced Workflow Efficiency: CellScientist demonstrated better execution success (SR) and rule-compliant success (R-SR) compared to simpler baselines, indicating more reliable and compliant model generation.
  • Robustness: The framework showed robustness to different LLM backbones, prompt variations, and run-level stochasticity, suggesting its reliability in practical applications.

Limitations and Future Work

  • Scope of Interpretation: The authors emphasize that the traces should not be interpreted as proof of perfect failure attribution or that the final candidate is uniquely optimal. The system documents how diagnostics and validation outcomes were associated with revisions, not necessarily the absolute true cause of every failure.
  • Future Work: The authors suggest extending quantitative audits of routing decisions, broadening budget-frontier comparisons across different agentic workflows, and incorporating stronger biological validation of generated design rationales.

Summary updated

Abstract

Virtual Cell Modeling (VCM) requires models that not only predict perturbation responses, but also support targeted revision when predictions fail. Current LLM-assisted modeling workflows face a refinement-routing problem: prediction discrepancies are observed through executable implementations, but the relevant revision may involve the modeling assumption, representation design, implementation, or task constraint. Without structured feedback propagation across these levels, iterative refinement may repair code while failing to revise the assumption responsible for the discrepancy. We propose CellScientist, a dual-space hierarchical framework that couples a high-level hypothesis space with a low-level executable implementation space. CellScientist represents modeling decisions as structured states, realizes them as admissible programs under task and interface constraints, and routes execution discrepancies back to targeted hypothesis or implementation updates. This enables a closed Hypothesis -> Implementation -> Hypothesis loop where failures become structured signals for model refinement rather than debugging events. Across morphology and transcriptomic benchmarks, with additional single-cell perturbation evaluations, the final executable models selected by CellScientist improve over reference baselines under fixed split and evaluation protocols, while the workflow produces auditable refinement traces.
BibTeX
@misc{li2026cellscientist,
  author = {Mengran Li and Bo Li and Jiaying Wang and Wenbin Xing and Yixuan Dong and Chengyang Zhang and Hongliang Zhang and Yuzhong Peng and Jinlin Wu and Bob Zhang and Bingo Wing-Kuen Ling and Fuji Yang and Zhen Lei and Jiebo Luo and Zelin Zang},
  title = {{CellScientist: Dual-Space Hierarchical Orchestration for Closed-Loop Refinement of Virtual Cell Models}},
  year = {2026},
  eprint = {2605.07335},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2605.07335},
}
Figure