ATLAS: A Multi-Agent Large Language Model Framework for Accurate Single-Cell Annotation

ATLAS is a bioinformatics framework that combines multi-agent collaboration and large language models, aiming to solve the challenge of cell type annotation for single-cell RNA sequencing (scRNA-seq) data. Its core innovation lies in using a multi-agent system for division of labor and collaboration, comprehensively judging cell types from dimensions such as gene expression, pathway enrichment, and literature knowledge, achieving dual improvements in annotation accuracy and interpretability, and providing a new tool for biomedical research.

Single-cell RNA sequencing technology has completely transformed the resolution of biological tissue research, but faces a core challenge: how to accurately classify tens of thousands of cell types? Traditional methods rely on manual labeling or automatic classification using reference datasets, which have problems such as insufficient accuracy, poor interpretability, and difficulty in handling rare cell types. The ATLAS project innovatively combines large language models and multi-agent architecture to provide a new solution to this problem.

ATLAS (Accurate and Interpretable Single-Cell Annotation via Multi-Agent LLM Framework) is an open-source tool whose core is a multi-agent system:

• Gene Expression Analysis Agent: Interprets gene expression profiles and identifies signature gene features;
• Pathway Enrichment Analysis Agent: Understands cell functions from the perspective of biological pathways;
• Literature Knowledge Retrieval Agent: Integrates biomedical literature to ensure annotations are consistent with domain knowledge;
• Consensus Decision Agent: Integrates results from various agents through a consensus mechanism to improve accuracy and transparency. The design draws on the model of human expert consultation, making comprehensive judgments on cell types from multiple dimensions.

The ATLAS workflow includes:

1. Data Preprocessing: Quality control, normalization, dimensionality reduction;
2. Agent Initialization: Launch analysis tasks based on the number of cell clusters;
3. Parallel Analysis: Each agent independently analyzes the assigned cell clusters;
4. Knowledge Integration: The literature agent provides external knowledge support;
5. Consensus Formation: Integrate opinions through voting or weighted mechanisms;
6. Result Output: Generate annotations with confidence scores.

Accuracy Improvement:

• Multi-angle verification: Complementary evaluation of cell clusters by different agents;
• Error self-check: Disagreements among agents mark cases requiring manual review;
• Knowledge fusion: Combining data-driven analysis and knowledge-driven reasoning. Interpretability Enhancement:
• Decision traceability: Annotations are accompanied by an evidence chain;
• Confidence quantification: Provides confidence scores;
• Literature citations: Links to relevant research literature;
• Disagreement visualization: Shows the consistency of agents' opinions.

ATLAS can be applied to:

• Tumor microenvironment research: Precisely identify subtypes of tumor-infiltrating immune cells, providing a basis for immunotherapy;
• Developmental biology: Capture the dynamic differentiation process of embryonic development and map developmental trajectories;
• Drug response research: Annotate changes in cell types before and after drug treatment, identifying targets and drug resistance mechanisms.

Current Limitations:

• High computational cost: Large overhead from multi-agent and LLM calls;
• Knowledge timeliness: Literature knowledge bases need regular updates;
• Rare cell types: Annotation accuracy is limited when literature is scarce. Future Directions:
• Optimize agent collaboration protocols to reduce redundant computation;
• Integrate real-time literature update mechanisms;
• Develop active learning modules to continuously improve from expert feedback.

ATLAS represents a new paradigm for AI-empowered scientific research—an augmented intelligence system for human-machine collaboration. By combining the knowledge processing capabilities of LLMs, the advantages of multi-agent collaboration, and biomedical expertise, it opens up new possibilities for single-cell data analysis. In the era of precision medicine, such tools will promote the translation of basic research on disease cells to clinical applications.