ATLAS: A Multi-Agent LLM Framework for Accurate and Interpretable Single-Cell Annotation

ATLAS is an open-source multi-agent LLM framework reproduced based on the CASSIA paper, designed specifically for automated cell type annotation of single-cell RNA sequencing (scRNA-seq) data. Through the division of labor and collaboration among 7 cooperative agents, it achieves high-quality, interpretable annotation results with controllable costs, demonstrating the strong potential of AI agent collaboration in biomedical research.

Single-cell RNA sequencing technology enables single-cell resolution analysis of gene expression, but the massive amount of data also poses challenges for cell cluster annotation: traditional manual annotation is time-consuming and subjective. Large Language Models (LLMs) have the ability to understand biological literature and gene functions, providing new possibilities for automated annotation. ATLAS is exactly a multi-agent LLM framework addressing this issue.

ATLAS adopts a "divide and conquer" strategy, decomposing the annotation task into sub-tasks handled by 7 agents:

• Core Agents: Annotator (infers cell types), Validator (verifies consistency and iteratively corrects), Formatter (converts to JSON format), Scoring (quality scoring), Reporter (generates HTML reports);
• Optional Enhancement Modules: Annotation Boost (corrects low-confidence annotations, e.g., identifying gold standard errors), RAG (retrieves authoritative databases to enhance annotations).

Cost Optimization: Controls costs through intelligent model selection—for example, low-cost models (DeepSeek v3, Gemini Flash) are used for scoring/formatting, while strong models (Claude Sonnet4.5) are used for annotation/validation. The default pipeline costs approximately $0.04 per run; Knowledge Base Integration: Built-in with authoritative data from CellMarker2.0 and Cell Ontology, and unifies access to multiple LLMs (GPT-5, Claude, etc.) via OpenRouter, requiring only one API key.

Results from ATLAS's test suite demonstrate its reliability:

• Clear CD8+ T cell case: Correct (score 92);
• Plasma cell case: Identified through noise (score 78);
• Key breakthrough: Identified gold standard errors (corrected incorrectly labeled monocytes to enteric glial cells); These results prove that multi-agent collaboration can discover issues missed by human experts.

Application Value: ATLAS provides interpretable reasoning chains, auditable report outputs, low-cost automated annotation, and workflows that integrate scattered knowledge; Relationship to CASSIA: Faithfully reproduces the CASSIA methodology, simplifies LLM provider support (only OpenRouter), has a more readable codebase, and is suitable as a learning example.

Limitations: Only focuses on cell type annotation and does not extend to processes like batch effect correction; relies on predefined knowledge bases, with insufficient coverage of new cell types; Future Directions: Integrate real-time literature retrieval, expand spatial transcriptome annotation, develop visualization interfaces, and establish a community feedback loop.

ATLAS is a successful case of AI agent application in vertical domains, achieving performance beyond a single model through task decomposition and agent collaboration. This model is not only applicable to single-cell biology but also provides a reference architecture for other scientific fields that require professional knowledge integration and complex reasoning.