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 automates complex single-cell bioinformatics workflows end-to-end

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 single-cell RNA sequencing (scRNA-seq) and spatial transcriptomics (ST) analysis through natural language. It comprises three specialized LLM agents—Planner, Executor, and Evaluator—coupled with a curated bioinformatics toolkit (sc-Omni) and a self-reflective optimization loop. Benchmarks across 60+ datasets show >96% execution success and state-of-the-art performance on five downstream tasks.

Problem

Existing scRNA-seq and ST analysis pipelines require dual expertise in computational programming and domain biology to manually select tools, tune hyperparameters, and interpret results. Prior LLM-based attempts (AutoBA, BioMANIA) prioritize execution success over biological quality, and lack automated mechanisms to evaluate biological relevance and select optimal algorithms across heterogeneous datasets.

Method

CellAgent implements a three-tier hierarchical decision process over the space of analysis pipelines:

  1. Planner decomposes a user's natural-language request into ordered subtasks \(t_1, t_2, \ldots, t_n\) by inspecting a dataset summary \(\psi(D)\) and applying codified best-practice knowledge: \(t_1,\ldots,t_n \leftarrow A_p^{\text{LLM}}(p_p^{\text{sys}}, u_{\text{task}}, u_{\text{req}}, u_D, \psi(D))\).
  2. Executor (Tool Selector + Code Programmer) selects tools \(T_{t_i}\) from sc-Omni, generates Python code \(c_i\), and performs autonomous self-correction on execution errors. Code is context-aware via a memory module \(M\) that carries only finalized code from prior steps.
  3. Evaluator (powered by GPT-4o) scores multiple candidate solutions using task-specific quantitative metrics (e.g., iLISI for batch correction, Accuracy Score for imputation), qualitative visual assessments, and anonymized outputs to prevent self-circularity: \(\bar{c}_i = A_e^{\text{LLM}}(p_e^{\text{sys}}, u_{\text{req}}, u_D, t_i, \{c_i^j\})\).

Memory control uses a dual architecture: global memory retains only final validated code (high information entropy, minimal tokens); local memory captures the full error/correction trace within a subtask and is discarded upon completion. sc-Omni consolidates tools for 15+ analytical tasks as self-documenting Python classes. A Jupyter-based code sandbox (via nbconvert) isolates execution for security and reproducibility.

Key Contributions

  • Hierarchical Planner→Executor→Evaluator multi-agent architecture for autonomous bioinformatics pipeline construction
  • Self-Reflective Optimization Mechanism: automated multi-candidate evaluation with fairness safeguards (anonymized tool names/prompts) replacing manual assessment
  • sc-Omni: expert-curated, self-documenting toolkit covering 15+ scRNA-seq and ST tasks
  • Dual global/local memory architecture that retains only successful code paths, minimizing token footprint
  • Jupyter sandbox for secure, reproducible code execution
  • Interactive online platform enabling community access

Results

  • Execution success rate: >96% average across 60+ datasets
  • Cell type annotation: average consistency score 0.85 vs. scGPT 0.77, Celltypist 0.75, ScType 0.59, SCSA 0.47, CellMaker 0.58
  • Batch correction (overall score): CellAgent 0.67 vs. scVI 0.66, Liger 0.65, Harmony 0.60, Combat 0.57, scGPT 0.57
  • Trajectory inference (overall score): CellAgent 0.50 vs. Slingshot 0.47, Scorpius 0.44, PageTree 0.40, Page 0.07, RaceID/StemID 0.06
  • Spatial domain identification (ARI): CellAgent 0.47 vs. BayesSpace 0.46, DeepST 0.47, SEDR 0.44, spaGCN 0.38, stLearn 0.38
  • Spatial imputation (Accuracy Score): CellAgent 0.88 vs. Tangram 0.75, SpaGE 0.73, gimVI 0.71, nanoSpaRc 0.41, stPlus 0.40

Limitations

  • The Evaluator is powered by GPT-4o, introducing dependency on a closed, paid external model and potential evaluation bias despite anonymization safeguards
  • Text is truncated before full ablation results are reported; the extent to which performance depends on the base LLM backbone is not fully detailed in the provided excerpt
  • Biological correctness of Evaluator judgments relies on visual and metric proxies; truly novel biology without established metrics may evade quality control
  • Benchmarks are confined to five task types; generalizability to less-studied modalities (e.g., multi-omics, ATAC-seq) is untested
  • Self-corrective loops and multi-candidate evaluation increase computational cost versus single-shot approaches

Relevance to Agentic AI / LLM Agents

CellAgent is a concrete instantiation of the hierarchical planning + tool-use + self-reflection agent pattern applied to a high-stakes scientific domain, demonstrating that role-differentiated LLM agents (distinct Planner, Executor, Evaluator) outperform monolithic LLM baselines in both task success and output quality. The Self-Reflective Optimization Mechanism—where an Evaluator agent scores multiple parallel Executor candidates using domain-specific metrics rather than LLM self-rating—is a practically important technique for grounding agent quality assessment in objective signals. The dual global/local memory design addresses the stateless LLM context problem with a principled discard-after-use strategy, contributing a reusable pattern for long-horizon agentic workflows. This work exemplifies how domain-specific tool curation (sc-Omni) and structured agent roles can bridge the gap between general-purpose LLM agent frameworks (AutoGen, LangChain) and the precision requirements of scientific computing.