cafferychen777/mLLMCelltype

What it solves

mLLMCelltype addresses the challenge of automated cell type annotation in single-cell RNA sequencing (scRNA-seq) data. Traditionally, identifying cell types from gene expression data can be error-prone or dependent on specific reference datasets. This framework reduces single-model biases and errors by using a consensus-based approach across multiple large language models (LLMs).

How it works

The tool takes marker genes (differentially expressed genes) for each cell cluster and uses a variety of LLMs (such as GPT, Claude, Gemini, and others) to predict the cell type. Instead of relying on a single model, it implements a consensus framework where multiple models analyze the data and their predictions are combined. The process includes iterative discussion rounds where models evaluate evidence and refine their annotations, and it provides uncertainty metrics like Consensus Proportion and Shannon Entropy to flag unreliable predictions.

Who it’s for

It is designed for researchers in single-cell transcriptomics and bioinformatics who use platforms like Scanpy (Python) or Seurat (R) and need a reference-free method to annotate cell populations.

Highlights

• Multi-LLM Consensus: Combines predictions from 10+ LLM providers to improve accuracy over single-model baselines.
• Reference-Free: Annotates cell types without requiring pre-training or external reference datasets.
• Iterative Refinement: Models engage in multiple rounds of deliberation to refine results.
• Uncertainty Quantification: Provides metrics to identify and validate uncertain annotations.
• Cross-Platform Support: Fully integrated with Scanpy and Seurat workflows via dedicated Python and R packages.
• Broad Model Compatibility: Supports a wide range of providers including OpenAI, Anthropic, Google, DeepSeek, and OpenRouter.