Highly accurate reference and method selection for universal cross–data set cell type annotation with CAMUS

Cell type annotation is a critical and essential task in single-cell data analysis. Various reference-based methods have provided rapid annotation for diverse single-cell data. However, selection of the optimal references and methods is often overlooked. To this end, we present a cross-data set cell type annotation methodology with a universal reference data and method selection strategy (CAMUS) to achieve highly accurate and efficient annotations. We demonstrate the advantages of CAMUS by conducting comprehensive analyses on 672 pairs of cross-species scRNA-seq data sets. The annotation results with references selected by CAMUS achieves substantial accuracy gains (25.0%-124.7%) over random selection strategies across five reference-based methods. CAMUS achieves high accuracy in choosing the best reference-method pair among 3360 pairs (49.1%). Moreover, CAMUS shows high accuracy in selecting the best methods on the 80 scST data sets (82.5%) and five scATAC-seq data sets (100.0%), illustrating its universal applicability. In addition, we utilize the CAMUS score with other metrics to predict the annotation accuracy, providing direct guidance on whether to accept current annotation results.