Benchmarking large language models improves cell type annotation in transcriptomics, suggesting enhanced biological insights.
Key Points
The aim is to improve cell type annotation by leveraging large language models for better adaptability and performance across various biological contexts.
Benchmarking 79 large language models across 1130 datasets
Using an evaluation framework that incorporates ontology structure and semantic reasoning
Developing an open-source R package and web platform for user-friendly access
Claude 3.5 Sonnet demonstrated the best overall performance with a weighted accuracy of 76%
AICellType enables flexible deployment via OpenRouter or custom APIs
The platform supports diverse species, tissues, and integrates with Seurat workflows