Overview
CellMaster uses an LLM’s biological knowledge to annotate single-cell RNA-seq data without task-specific pretraining or a fixed marker database. It produces explanations that experts can inspect and refine, supporting collaboration on rare, ambiguous, and novel cell populations.
Across nine datasets covering eight tissues, the paper reports a 7.1% accuracy improvement over the strongest baselines in automatic mode. Human feedback increases the reported advantage to 18.6%, with a 22.1% gain for subtype annotation.