PROTOTYPE-BASED CONTINUAL LEARNING FOR SINGLE-CELL ANNOTATION
Large-scale single-cell atlases provide an increasingly comprehensive view of cellular diversity, but their continued expansion across studies poses a fundamental challenge: preserving consistent cell identities while capturing biological variation in cellular states. Most existing annotation frameworks are built around static references, making it difficult to incorporate newly generated datasets into established cellular representations without retraining on historical data. When updated sequentially, these methods are further constrained by catastrophic forgetting and batch-specific biases, limiting scalability and the continuity of knowledge integration. Here we introduce scEvolver, a continual learning framework for single-cell annotation that incrementally accumulates knowledge through memory-guided refinement of cell-type prototypes without revisiting historical data. Across sequencing platforms, tissue contexts and molecular modalities, scEvolver supports robust annotation and external query mapping with substantially fewer labelled reference cells. By preserving consistent cell-type semantics across datasets while capturing biologically meaningful within-class heterogeneity, scEvolver enables the identification of epithelial cell-state transitions in inflammatory gut disease. External mapping to the healthy Human Lung Cell Atlas further reveals shared cell-state deviations across multiple diseases, including an FCGR3A+ inflammatory monocyte programme in sarcoidosis, chronic obstructive pulmonary disease and idiopathic pulmonary fibrosis, highlighting scEvolvers potential to characterize context-specific cellular dynamics in complex disease settings.