scCAP: An ontology-aware large language model framework for hierarchical and standardized single-cell type annotation

Background: Accurate cell type annotation remains challenging in single-cell RNA sequencing analysis. Existing tools (CellTypist, SCimilarity, SingleR, Geneformer) exhibit distinct biases and lack standardized nomenclature, hindering cross-study comparability. Furthermore, current methods fail to provide hierarchical annotations reflecting cell type ontologies. Methods: We developed scCAP, a meta-learning framework integrating multiple annotation tools using an ontology-aware large language model (LLM) architecture. scCAP employs HybridAttentionNet combining: (1) a knowledge stream using self-attention on semantic embeddings, and (2) a confidence stream learning tool-specific reliability patterns. A key innovation is our LLM-based label interpretation: for each prediction, the LLM generates structured descriptions including full names, marker genes, and pathways, enabling automatic abbreviation expansion and cluster annotation interpretation. These descriptions are embedded for standardized semantic comparison across heterogeneous nomenclatures. The LLM also classifies labels into hierarchical ontology levels (L1: lineages; L2: cell classes; L3: subtypes), validated against Cell Ontology. Tool dropout regularization during training ensures robust performance even when individual tools are unavailable. Additionally, iterative pseudo-labeling progressively refines predictions by incorporating high-confidence annotations as training signals. The modular architecture allows seamless tool addition/replacement without retraining. Results: We evaluated scCAP on Cancer Cell Census Atlas and Human Lung Cell Atlas data against four annotation tools. At Level 3, scCAP achieved 0.898 average semantic similarity, outperforming SingleR (0.865), CellTypist (0.861), SCimilarity (0.857), and Geneformer (0.805). At ≥80% similarity threshold, scCAP achieved 87.7% accuracy versus 73.3% for CellTypist. scCAP demonstrated consistent hierarchical performance (L1:0.858; L2:0.874; L3:0.898) with higher median similarity and tighter variance than individual tools. Conclusions: scCAP leverages LLMs for ontology-aware label standardization and hierarchical classification. Its plug-and-play architecture enables incorporating emerging tools while maintaining consistent outputs, facilitating meta-analyses of tumor microenvironment heterogeneity.