HarmonyCell: Automating Single-Cell Perturbation Modeling under Semantic and Distribution Shifts

Wenxuan Huang; Mingyu Tsoi; Yanhao Huang; Xinjie Mao; Xue Xia; Hao Wu; Jiaqi Wei; Yuejin Yang

cs.AIMar 2, 2026

HarmonyCell: Automating Single-Cell Perturbation Modeling under Semantic and Distribution Shifts

Authors: Wenxuan Huang, Mingyu Tsoi, Yanhao Huang, Xinjie Mao, Xue Xia, Hao Wu, Jiaqi Wei, Yuejin Yang,

Organizations: Fudan University · Shanghai Artificial Intelligence Laboratory · The Hong Kong University of Science and Technology · Xi’an Jiaotong-Liverpool University ·

Abstract

Single-cell perturbation studies face dual heterogeneity bottlenecks: (i) semantic heterogeneity--identical biological concepts encoded under incompatible metadata schemas across datasets; and (ii) statistical heterogeneity--distribution shifts from biological variation demanding dataset-specific inductive biases. We propose HarmonyCell, an end-to-end agent framework resolving each challenge through a dedicated mechanism: an LLM-driven Semantic Unifier autonomously maps disparate metadata into a canonical interface without manual intervention; and an adaptive Monte Carlo Tree Search engine operates over a hierarchical action space to synthesize architectures with optimal statistical inductive biases for distribution shifts. Evaluated across diverse perturbation tasks under both semantic and distribution shifts, HarmonyCell achieves a 95% valid execution rate on heterogeneous input datasets (versus 0% for general agents) while matching or even exceeding expert-designed baselines in rigorous out-of-distribution evaluations. This dual-track orchestration enables scalable automatic virtual cell modeling without dataset-specific engineering.

Figures & tables

Figure 1: Uniqueness of HarmonyCell . Existing task-specific LLM agents (CellForge etc.) but require rigid data input format, while general-purpose agents often lack biological knowledge. HarmonyCell can enhance its capabilities through biological priors, and also handle the problem of data heterogeneity.

Figure 2: Architecture of HarmonyCell. The system integrates: (1) an LLM-driven Semantic Unifier to canonicalize heterogeneous h5ad inputs; (2) a Retrieval-Augmented Agent for meta-initialization using historical strategies; and (3) an Executor Agent guided by MCTS over a hierarchical action space (bottom panel). This hierarchical decomposition significantly bolsters search stability by proactively pruning branches that lead to erroneous ‘bug’ nodes before full execution. While runtime failures trigger a ReAct-style debugging loop, successfully validated pipelines are stored in a persistent knowledge base for future reuse.

Abilities	Heterogeneity Data Unification	Biological Prior	Model Exploration	Collaborative Coding
General-Purpose Agents	$-$	$-$	$-$	✓
Specialized Cell Scientists	$-$	✓	✓	✓
HarmonyCell	✓	✓	✓	✓

Table 1: Comparison of capabilities. General-purpose agents ( Yang et al., 2025 ; Ou et al., 2025 ) support collaborative coding but lack biological priors, model exploration, and heterogeneity data handling. Specialized cell scientists ( Tang et al., 2025 ) use biological priors and model exploration, but are limited to standardized data. HarmonyCell unifies all four functions, enabling end-to-end modeling across heterogeneous datasets.

Agents	Error Type			Valid Execution Rate
Agents	Preprocess Error	Model Error	Hallucinated Success	Valid Execution Rate
AIDE ( Jiang et al., 2025 )	35%	50%	15%	0%
R&D Agent ( Yang et al., 2025 )	45%	30%	25%	0%
HarmonyCell	0%	5%	0%	95%

Table 2: Superiority in Handling Semantic Heterogeneity. While general coding agents (AIDE, R&D Agent) failed all 20 virtual cell modeling trials despite detailed manual guidance, HarmonyCell achieved a 95% success rate. This underscores its robust capability in resolving data heterogeneity and automating preprocessing.

Figure 3: HarmonyCell successfully handles semantic heterogeneity, and enables synergistic dataset scaling. We compared the generalization performance of models trained on single-source Adamson ( Adamson et al., 2016 ) and Replogle ( Replogle et al., 2022 ) datasets against the joint dataset harmonized by HarmonyCell’s Semantic Unifier. Models were evaluated on held-out test sets under a unified separate cross-validation protocol.

	CosLogFC $\uparrow$	RMSE $\downarrow$	DeltaPCC $\uparrow$	CosLogFC $\uparrow$	RMSE $\downarrow$	DeltaPCC $\uparrow$
Gene Perturbation	$\circ$ Norman ( Norman et al., 2019 )			$\circ$ Adamson ( Adamson et al., 2016 )
Biolord ( Piran et al., 2024 )	0.57	0.05	0.42	0.52	0.02	0.54
Sams VAE ( Bereket and Karaletsos, 2023 )	0.58	0.05	0.44	0.51	0.02	0.55
CPA ( Lotfollahi et al., 2021 )	0.55	0.05	0.39	0.41	0.03	0.32
HarmonyCell	0.61	0.09	0.62	0.32	0.09	0.49
Drug Perturbation	$\circ$ Srivatsan-Sciplex2 ( Srivatsan et al., 2020 )			$\star$ Srivatsan-Sciplex3 ( Srivatsan et al., 2020 )

Table 3: Performance on disjoint datasets under statistical heterogeneity. The model is evaluated on datasets with distinct statistical heterogeneities (Unseen Perturbation and Unseen Cell). HarmonyCell consistently matches or exceeds expert-level baseline performance, particularly in maintaining correlation (DeltaPCC) across distribution shifts. Underlined results indicate the best performance among baselines, while bold results indicate HarmonyCell’s performance. The uparrow

\uparrow

stands for higher value, while the downarrow

\downarrow

stands for lower value for better performance of a metric.

Figure 4: Ablation Study: Necessity of Semantic Unifier. During execution, HarmonyCell containing the Semantic Unifier exhibits a more stable and error-free execution workflow compared to agents without the module.

Figure 5: Ablation Study: Necessity of Hierarchical Action Space. HarmonyCell achieves superior convergence speed and accuracy compared to the non-hierarchical ablated agents, effectively surpasses the state-of-the-art specialized baseline models.

Figure 6: Case Study: MCTS Exploration for Norman dataset. The number inside each node represents the DeltaPCC validation value of the model at that node.

Appendix figures & tables2 assets

Supplementary material from the paper’s appendix.

Appendix

Component	Symbol	Value	Description
Retrieval (RAG)	$\tau_{filter}$	0.3	Threshold for filtering irrelevant historical tasks
Retrieval (RAG)	$m$	3	Number of top neighbors retrieved
MCTS Selection	$\alpha$	0.7	Weight for Optimistic Value ( $Q_{mix}=\alpha Q_{max}+(1-\alpha)\bar{Q}$ )
	$C$	1.0	Exploration constant in UCT formula
	$N_{sim}$	32	Number of MCTS simulation iterations
Reward Function	$w_{p}$	0.8	Weight for validation performance ( $\Delta$ PCC)

Appendix

Table 4: Hyperparameters for HarmonyCell MCTS Engine.

\mathcal{M}=\text{LLM}(\mathcal{P})\in\mathbb{M}

Appendix

Algorithm 1 Formal Semantic Unification via LLM-Guided Schema Alignment

Single-Cell Perturbation Modeling

Single-cell perturbation modeling predicts how genetic or chemical interventions change cellular states across cells and contexts.

See all 61 papers

AI Scientists for Building Virtual-Cell Models
Aug 4, 2025Xiangru Tang, Jiapeng Chen, Daniel Shao +12AI Agents for Scientific Discovery Autonomous Research Agents
Yale University · University of Pennsylvania · Helmholtz Zentrum München +3
PerturbDiff: Functional Diffusion for Single-Cell Perturbation Modeling
Feb 23, 2026Xinyu Yuan, Xixian Liu, Ya Shi Zhang +3Single-Cell Perturbation Modeling Single-Cell Transcriptomics
Mila - Québec AI Institute · University of Montréal · University of Ottawa +3
CellRFT: Reinforcement Fine-Tuning for Single-Cell Perturbation Modeling
Sep 17, 2026Jie Yan, Li Liu, Hanze Guo +7Multi-Objective Reinforcement Learning Single-Cell Perturbation Modeling
Chinese Academy of Sciences · Peking University · Renmin University of China +2

Aug 4, 2025J/K move · Enter open · S save

AI Scientists for Building Virtual-Cell Models

Xiangru Tang, Jiapeng Chen, Daniel Shao, Zhuoyun Yu+11

Yale University · University of Pennsylvania · Helmholtz Zentrum München+3

AI Agents for Scientific Discovery Autonomous Research Agents Single-Cell Perturbation Modeling LLM Agent Orchestration

Open paper AskPDF arXiv

Autonomous agents, from virtual laboratories to biological data-analysis agents and co-scientists, have begun to turn scientific goals into hypotheses. Whether they can go further and convert real, noisy and complex biological data into executable model hypotheses and architectures that encode testable assumptions remains unclear. Predicting how cells respond to genetic and chemical perturbations is a natural test of this ability: it is central to decoding disease mechanisms and therapeutic discovery, yet it demands cross-disciplinary expertise rarely concentrated in a single research group. Here, we introduce CellForge, an organized team of autonomous agents for single-cell perturbation modeling. Given a dataset and a task, it surveys the literature, parses the data, identifies relevant baselines, and formulates a model hypothesis through critic-mediated discussion among agents specialized in data modeling, single-cell biology, and deep learning. It then writes the training code, runs the experiments, and returns a trained predictor including architecture, code, and validation results. The architecture itself is the hypothesis: CellForge produces it end-to-end, with no human specifying the design, loss, or training protocol. Across six perturbation datasets spanning transcriptomic, epigenomic, and proteomic contexts, CellForge yields models that are competitive with hand-crafted state-of-the-art results on established tasks and extends the same workflow to paired multi-omic readouts. Blinded human evaluators rate CellForge's research plans as more scientifically sound and innovative than those from single LLMs or deep-research systems. In summary, we show autonomous agents can complete the full hypothesis-to-validation cycle in single-cell biology, with an AI effectively building another AI, in a complex setting that currently requires a team of human experts.

Abstract

Single-Cell Perturbation Modeling

Explore similar work

AI Scientists for Building Virtual-Cell Models

PerturbDiff: Functional Diffusion for Single-Cell Perturbation Modeling

CellRFT: Reinforcement Fine-Tuning for Single-Cell Perturbation Modeling