scHilda: Hierarchical Integration of LLM with KG database for single cell type annotation

Area of Science:

• Computational Biology
• Genomics
• Artificial Intelligence

Background:

• Single-cell RNA sequencing (scRNA-seq) cell type annotation is crucial but faces accuracy and generalization challenges.
• Existing automated methods struggle with novel cell types and LLM hallucinations.
• Integrating external knowledge bases with LLMs is difficult due to data imperfections.

Purpose of the Study:

• To develop a novel framework, scHilda, for robust and accurate cell type annotation in scRNA-seq data.
• To address LLM hallucinations and knowledge base deficiencies in biological data analysis.
• To improve the interpretability and generalization of automated cell annotation.

Main Methods:

• scHilda integrates a Knowledge Graph (KG) with Large Language Models (LLMs).
• A hierarchical arbitration strategy identifies major cell lineages using a global KG, then refines subtypes with focused subgraph information.
• Dynamic, knowledge-enhanced reasoning constrains LLM decision-making, mitigating hallucinations and knowledge base errors.

Main Results:

• scHilda achieves state-of-the-art performance on multiple benchmark datasets, outperforming existing methods.
• The framework demonstrates robustness in complex mixed samples and enables lightweight LLMs to approach top-tier performance.
• Statistical evaluations and interpretability case studies confirm scHilda's efficiency and transparent decision-making.

Conclusions:

• scHilda offers a new paradigm for trustworthy biological AI by integrating LLMs with structured biological knowledge.
• The framework significantly improves accuracy, interpretability, and robustness in scRNA-seq cell annotation.
• scHilda effectively constrains LLM reasoning, reducing hallucinations and enhancing biological insights.