Abstract
Single-cell RNA sequencing (scRNA-seq) technology has garnered considerable attention as it enables the exploration of cellular heterogeneity from a single-cell perspective. Various unsupervised methods, such as biclustering and clustering methods, offer a theoretical foundation for understanding the structure and function of cells. However, accurately identifying cell subtypes within complex scRNA-seq data remains challenging. To evaluate the current development status; summarize the strengths, weaknesses, and improvement strategies of unsupervised methods; and provide guidelines for future research, we surveyed five biclustering and 21 clustering methods applied to different types of scRNA-seq datasets. We employed three external and two internal metrics to determine clustering performance on 10 publicly available real datasets. Dataset properties are quantified from six perspectives to discover the most suitable biclustering or clustering methods. The results of this survey indicate that biclustering methods are effective for identifying local consistency or for deeply mining partially annotated datasets. Conversely, clustering methods are more suitable for dealing with unknown datasets. This survey aids in identifying cellular heterogeneity by recommending appropriate methods based on different dataset characteristics.