GPTAnno: Ontology-tree-guided hierarchical cell type annotation based on GPT models for single-cell data

Abstract

Cell type annotation is critical for interpreting single-cell transcriptomic data but remains challenging due to uncertain cellular clustering granularity and inconsistent labeling across studies. Here we present GPTAnno, an automated, ontology-tree-guided, uncertainty-aware, hierarchical cell type annotation method based on GPT models. GPTAnno directly handles gene expression matrices, integrates multi-resolution clustering with large language model reasoning constrained by the cell ontology to produce standardized, ontology-aware, and reproducible annotations with automatic resolution selection. GPTAnno selects optimal clustering resolutions based on the annotation distance on the ontology tree and quantifies annotation uncertainty to flag ambiguous clusters for expert review. Benchmarking across twelve large-scale datasets demonstrates GPTAnnos superior accuracy on annotating cell types across various species, tissues, and disease contexts against existing methods. Implemented in R and Python with Seurat and Scanpy compatibility, GPTAnno allows simple inputs to streamline the reproducible annotation, considerably reducing human efforts in repeated reclustering, assigning and examining the labels.