Skip to content

CellAgent: LLM-Driven Multi-Agent Framework for Natural Language-Based Single-Cell Analysis

🕒 Published (v1): 2026-01-01 · Source: ICLR · Venue: ICLR 2026 · link

Why this paper was selected

LLM-driven multi-agent framework for natural language single-cell analysis [ICLR 2026]

Ask a follow-up

Open an assistant pre-loaded with this paper's context.

💬 Ask ChatGPT✦ Ask Claude

TL;DR

CellAgent is a hierarchical multi-agent framework that automates end-to-end scRNA-seq and spatial transcriptomics analysis via natural language. It orchestrates a Planner, Executor, and Evaluator in a self-reflective optimization loop, grounded in a curated domain toolkit (sc-Omni). It achieves >96% execution success across 60+ datasets and matches or exceeds human-expert performance on five bioinformatics benchmarks.

Problem

Single-cell and spatial transcriptomics analysis requires dual expertise in computational programming and biology: researchers must manually select tools, tune hyperparameters, and write complex pipelines. Existing LLM-based agents (AutoBA, BioMANIA) improve execution success rates but lack automated mechanisms to assess biological relevance of results, preventing autonomous algorithm selection and hyperparameter optimization across heterogeneous datasets.

Method

CellAgent uses a three-tier hierarchical decision-making architecture:

  1. Planner (\(A^{LLM}_p\)): receives user task description \(u_{task}\), dataset summary \(\psi(D)\), and optional preferences \(u_{req}\), then produces an ordered subtask sequence \(t_1, t_2, \ldots, t_n\) using expert-codified scRNA-seq best-practices baked into its system prompt.

  2. Executor (Tool Selector \(A^{LLM}_t\) + Code Programmer \(A^{LLM}_c\)): for each subtask \(t_i\), selects tools \(T_{t_i}\) from sc-Omni, then generates Python code \(c_i\) using tool docstrings and a memory context \(M\) containing finalized code from prior steps. Autonomous self-correction handles runtime errors.

  3. Evaluator (\(A^{LLM}_e\)): scores \(K\) candidate solutions per subtask using task-specific quantitative metrics (e.g., iLISI for batch correction, RANKPCC for imputation) and qualitative visual assessments, then selects \(\bar{c}_i = \arg\max_j A^{LLM}_e(\{c^j_i\})\). Fairness is enforced by providing only anonymized outputs and diagnostic plots—never prompt text or tool names—to the Evaluator.

Memory architecture: global memory stores only final validated code \(M \leftarrow \{\bar{c}_1, \bar{c}_2, \ldots\}\) (high information density, low token cost); local memory captures full error/correction traces within a subtask, then is discarded after success.

Code Sandbox: Jupyter Notebook via nbconvert isolates execution from the core framework, ensuring reproducibility and security.

Key Contributions

  • CellAgent framework: hierarchical Planner–Executor–Evaluator multi-agent pipeline for fully automated, natural-language-driven scRNA-seq and ST analysis.
  • sc-Omni toolkit: expert-curated, >15-task Python/R toolset with standardized docstring-driven APIs covering quality control through trajectory inference and spatial imputation.
  • Self-Reflective Optimization Mechanism: automated multi-run evaluation using domain-appropriate quantitative and qualitative criteria, replacing manual assessment.
  • Dual-layer memory control: global code-only memory for inter-step coherence; ephemeral local memory for intra-step self-correction, keeping context footprint minimal.
  • Secure code sandbox: nbconvert-based Jupyter execution environment decoupling analysis from the agent core.
  • Online interactive platform enabling community access.

Results

  • Execution success rate: >96% averaged across 60+ datasets.
  • Cell type annotation: average consistency score 0.85 vs. scGPT 0.77, ScType 0.59, Celltypist 0.75, SCSA 0.47, CellMaker 0.58 (6 datasets).
  • Batch correction (overall score): CellAgent 0.67 vs. scVI 0.66, Liger 0.65, Harmony 0.60, Combat 0.57, scGPT 0.57 (5 datasets).
  • Trajectory inference (overall score): CellAgent 0.50 vs. Slingshot 0.47, Scorpius 0.44, Page tree 0.40 (8 datasets).
  • Spatial domain identification (ARI): CellAgent 0.47 vs. BayesSpace 0.46, DeepST 0.47, SEDR 0.44, spaGCN 0.38, stLearn 0.38 (12 DLPFC slices).
  • Spatial imputation (Accuracy Score): CellAgent 0.88 vs. Tangram 0.75, SpaGE 0.73, gimVI 0.71, nanoSpaRc 0.41, stPlus 0.40 (7 datasets).

Limitations

  • Evaluator is powered specifically by GPT-4o, creating a dependency on a proprietary external model and potential cost/availability constraints.
  • Evaluation criteria for qualitative tasks (trajectory continuity, spatial coherence) rely on LLM visual judgment, which may not fully substitute for domain-expert review.
  • Benchmark datasets are limited to specific tissues/platforms; generalization to novel sequencing technologies or non-standard organisms is not assessed.
  • The ablation study on memory and base model variants is deferred to an appendix; the paper does not fully characterize failure modes or cases where self-reflective optimization does not converge.
  • The paper is truncated before the full spatial imputation results section, so completeness of that benchmark cannot be confirmed from the provided text.

Relevance to Harnesses / Meta-Harnesses

CellAgent is a domain-specific meta-harness: it does not perform any single analysis directly but instead orchestrates a pipeline of LLM agents, a curated tool library, a sandboxed code executor, and an automated evaluation loop to produce end-to-end scientific workflows from natural language. The self-reflective optimization mechanism—where the Evaluator triggers re-runs of the Executor with different algorithm choices and selects the best outcome—is a concrete harness-level feedback loop analogous to the retry-and-verify patterns central to general meta-harness design. The dual-layer memory architecture (global validated state vs. ephemeral local scratchpad) is a reusable pattern for any multi-step agentic harness where intermediate errors should not pollute downstream context. For researchers tracking harnesses, CellAgent demonstrates how domain-specific evaluation oracles can be integrated as first-class harness components to replace human-in-the-loop quality gates.