Cell type annotation remains one of the most critical yet time-consuming steps in single-cell RNA sequencing analysis. Researchers from the University of Wisconsin-Madison have developed CASSIA (Collaborative Agent System for Single-cell Interpretable Annotation), a novel multi-agent LLM framework that automates this process while providing interpretable, high-quality annotations without requiring reference datasets.

The multi-agent LLM system underlying CASSIA

(a) A user interacts directly with the Onboarding platform by specifying species, tissue type, and a collection of markers associated with cell subtypes within that tissue, if known. Any information associated with experimental conditions, interventions, or other sample specific information may be provided. (b) Together, this input is used to create the user prompt given to the Annotator agent. The Annotator agent performs a comprehensive annotation of the single-cell data using a zero-shot chain-of-thought approach that mimics the standard workflow that a computational biologist would typically follow for cell annotation. Results are then passed to the Validator agent to check marker and cell type consistency; results failing validation are passed back to the Annotator and this iterative process continues until results pass validation (or the maximum number of iterations is reached). Results are then moved to the Formatting agent which summarizes each cell annotation; this summary along with the full conversation history is provided to the Scoring agent for quality scoring. The Reporter agent then generates a comprehensive report documenting the complete annotation process, including agent conversations, quality evaluation reasoning, and validation decisions with supporting evidence to facilitate transparent interpretation of results. The output from default CASSIA is shown in (c). Optional agents include those shown in (d).

Multi-Agent Architecture

Unlike single-prompt LLM approaches, CASSIA employs five specialized AI agents working in sequence to ensure annotation quality:

1. Annotation Agent – Analyzes marker genes using a structured 6-step reasoning process, identifying both general cell types and specific subtypes

2. Validation Agent – Performs quality assurance by checking marker consistency and detecting potential mixed cell populations

3. Formatting Agent – Converts the analysis into structured, standardized output

4. Scoring Agent – Evaluates annotation confidence on a 0-100 scale using cell ontology-aware metrics

5. Reporting Agent – Generates interactive HTML reports with detailed reasoning traces

This collaborative approach mimics expert biologist workflows, with agents providing feedback to each other through iterative validation loops. When initial annotations score below threshold, the system can invoke an optional Annotation Boost agent for deeper analysis of challenging clusters.

Reference-Free Annotation

A key advantage of CASSIA is its reference-free design. The framework requires only marker gene expression data from standard tools like Seurat’s FindAllMarkers() or Scanpy’s rank_genes_groups(), eliminating the need for curated reference datasets or atlas matching. This makes CASSIA applicable across diverse tissues, species, and experimental conditions.

Flexible Model Support

CASSIA supports virtually any LLM through a unified provider interface:

– Commercial APIs: OpenAI GPT-5.2, Anthropic Claude Sonnet 4.5, Google Gemini 2.5 Flash

– OpenRouter: Access 100+ models through a single API key

– Custom endpoints: DeepSeek, GLM, Kimi, and other OpenAI-compatible APIs

– Local LLMs: Ollama and LM Studio for complete data privacy and zero API costs

Advanced Capabilities

Beyond core annotation, CASSIA provides specialized agents for complex analyses:

– Annotation Boost Agent: When clusters receive low confidence scores, this agent performs iterative hypothesis-driven refinement. It summarizes prior conversation history, generates new hypotheses about cell identity, and systematically tests them against additional markers, which significantly improves accuracy on challenging or ambiguous clusters.

– RAG (Retrieval-Augmented Generation) Agent: Automatically retrieves relevant information from marker databases, cell ontology trees, and scientific literature. Using a PCA-based approach, it identifies the most informative references for each annotation task, enabling accurate identification of rare cell types and novel populations without manual literature searches.

– Uncertainty Quantification: Runs multiple annotation iterations with majority voting to boost accuracy .

– Symphony Compare: Multi-model consensus building where different LLMs debate annotations for ambiguous clusters

– Subclustering Agent: Hierarchical analysis of heterogeneous cell populations

– Merging Agent: Consolidates granular annotations into biologically meaningful groupings at multiple detail levels

Accessibility

CASSIA prioritizes accessibility with multiple deployment options. The free web interface at cassia.bio allows researchers to try core features without installation. For full functionality, native R and Python packages provide seamless integration with existing single-cell workflows. Batch processing with configurable parallel workers enables efficient annotation of large datasets.

The framework generates comprehensive outputs including CSV files with predictions and confidence scores, JSON conversation histories preserving full reasoning chains, and interactive HTML reports for visual exploration of results.

Performance and Validation

Published in Nature Communications (2025), CASSIA has been validated across multiple tissues and species, consistently outperforming reference-free annotation methods in benchmark comparisons. The associated LLM annotation benchmark platform (sc-llm-benchmark.com) provides ongoing evaluation of model performance and cost-effectiveness.

Availability

– Web Interface: https://cassia.bio

– Documentation: https://docs.cassia.bio

– GitHub: https://github.com/ElliotXie/CASSIA

– Benchmark: https://sc-llm-benchmark.com