### mLLMCelltype: A Multi-Model Framework for Cell Type Annotation in Single-Cell RNA-seq Data

We present mLLMCelltype, a computational framework that employs multiple large language models (LLMs) to perform automated cell type annotation in single-cell RNA sequencing (scRNA-seq) data. The tool is available as both a web application (https://www.mllmcelltype.com) and a Python package.

#### Methodology

mLLMCelltype implements an iterative consensus framework where multiple LLMs collaborate to annotate cell types based on marker gene expression profiles. The framework incorporates:

– **Multi-model integration**: Supports latest LLMs including GPT-4.1, o3/o4, Claude 4 (Opus/Sonnet), Gemini 2.5, DeepSeek R1, Grok 3, Qwen Max 2025, and others

– **Uncertainty quantification**: Calculates Consensus Proportion (CP) and Shannon Entropy to assess annotation confidence

– **Iterative refinement**: Models engage in structured deliberation to converge on annotations

– **Cross-species compatibility**: Applicable to human, mouse, and other organisms

#### Implementation

The tool offers three deployment options:

1. **Web Interface**: Browser-based application requiring no local installation

– Accepts standard formats (CSV, TSV, Excel)

– Provides real-time processing feedback

– Exports results in multiple formats

2. **Python Package**: For programmatic access and pipeline integration

– Installation via pip: `pip install mllmcelltype`

– Compatible with Scanpy/AnnData workflows

– Batch processing capabilities

3. **R Package**: For R users and Seurat integration

– Currently available on GitHub: `devtools::install_github(“cafferychen777/mLLMCelltype”)`

– CRAN submission in progress

– Compatible with Seurat workflows

#### Technical Architecture

– Cloud deployment on Google Cloud Run with auto-scaling

– Persistent storage using Google Cloud Storage

– Global content delivery via Cloudflare CDN

– RESTful API for programmatic access

#### Performance and Validation

The multi-model consensus approach addresses common challenges in LLM-based analysis:

– Reduces individual model biases

– Mitigates hallucination through cross-validation

– Maintains annotation consistency across hierarchical cell type relationships

#### Publication

Full methodological details are available in:

Yang C, Zhang X, Chen J. mLLMCelltype: A Unified Cross-Species Foundation Model for Cell Type Annotation Using Multiple Large Language Models. bioRxiv. 2025. doi: 10.1101/2025.04.10.647852

#### Availability

– **Web Application**: https://www.mllmcelltype.com

– **Source Code**: https://github.com/cafferychen777/mllmcelltype

– **Documentation**: Available at the GitHub repository

– **License**: MIT License

This tool provides researchers with an accessible solution for cell type annotation in scRNA-seq analysis, particularly beneficial for laboratories without extensive computational infrastructure or bioinformatics expertise. The platform supports over 30 state-of-the-art language models and includes free model options through OpenRouter for budget-conscious researchers.