Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

ReCellTy: Domain-specific knowledge graph retrieval-augmented LLMs workflow for single-cell annotation

Dezheng Han, Yibin Jia, Ruxiao Chen, Wenjie Han, Shuaishuai Guo, Jianbo Wang

TL;DR

This work addresses precise cell type annotation in single-cell RNA-seq by leveraging a domain-specific knowledge graph to augment LLM reasoning. It builds a graph-structured feature-marker database from CellMarker2.0 and a multi-task retrieval framework that guides LLMs to infer cell types via broad-type and feature-function evidence. Across 11 tissues, ReCellTy yields higher human evaluation scores by up to 0.21 and semantic similarity gains of up to 6.1% compared with general-purpose LLMs and CellMarker2.0 baselines. The approach improves interpretability and compatibility with standard pipelines like Seurat, enabling scalable, diverse, knowledge-guided cell type reconstruction with a dynamic graph-augmentation workflow.

Abstract

To enable precise and fully automated cell type annotation with large language models (LLMs), we developed a graph structured feature marker database to retrieve entities linked to differential genes for cell reconstruction. We further designed a multi task workflow to optimize the annotation process. Compared to general purpose LLMs, our method improves human evaluation scores by up to 0.21 and semantic similarity by 6.1% across 11 tissue types, while more closely aligning with the cognitive logic of manual annotation.

ReCellTy: Domain-specific knowledge graph retrieval-augmented LLMs workflow for single-cell annotation

TL;DR

This work addresses precise cell type annotation in single-cell RNA-seq by leveraging a domain-specific knowledge graph to augment LLM reasoning. It builds a graph-structured feature-marker database from CellMarker2.0 and a multi-task retrieval framework that guides LLMs to infer cell types via broad-type and feature-function evidence. Across 11 tissues, ReCellTy yields higher human evaluation scores by up to 0.21 and semantic similarity gains of up to 6.1% compared with general-purpose LLMs and CellMarker2.0 baselines. The approach improves interpretability and compatibility with standard pipelines like Seurat, enabling scalable, diverse, knowledge-guided cell type reconstruction with a dynamic graph-augmentation workflow.

Abstract

To enable precise and fully automated cell type annotation with large language models (LLMs), we developed a graph structured feature marker database to retrieve entities linked to differential genes for cell reconstruction. We further designed a multi task workflow to optimize the annotation process. Compared to general purpose LLMs, our method improves human evaluation scores by up to 0.21 and semantic similarity by 6.1% across 11 tissue types, while more closely aligning with the cognitive logic of manual annotation.
Paper Structure (9 sections, 3 figures)

This paper contains 9 sections, 3 figures.

Table of Contents

  1. Methods
  2. Data processing
  3. Feature-marker GraphRAG
  4. Multi-Task framework
  5. Comparative methods
  6. Evaluations
  7. Application
  8. Data availability
  9. Code availability

Figures (3)

  • Figure 1: |Overview of cell type annotation methods.a, Knowledge graph-driven LLM for automated annotation. b, Semi-automated annotation with expert-LLM collaboration. c, Traditional annotation methods prior to the advent of LLM.
  • Figure 2: |Structure and method of the knowledge graph-based cell type annotation framework.a, Visual representation of a subset of the data within the graph database. b, Seven node types and their relationship types within the graph database. c, Data processing pipeline and cell annotation question-answering workflow.
  • Figure 3: |Performance evaluation.a, Human evaluation scores of four large models under different methods. b, Human evaluation scores of two large models across various tissues. c, Overall human evaluation scores for each method. d, Semantic evaluation scores of four large models under different methods. e, Semantic evaluation scores of two large models across various tissues. f, Intra-group semantic variance of annotation results for each model and method. g, Human evaluation scores of two models and CellMarker 2.0 across various tissues. h, Semantic evaluation scores of two models and CellMarker 2.0 across various tissues.