Cell type annotation is a fundamental task in single-cell genomics. Although various methods have been developed for automatic cell type annotation, they often function as black-box models, making predictions without explaining their reasoning and lacking proper uncertainty estimation for their predictions. Furthermore, they often struggle to annotate rare cell types. We introduce ProtoCloud, a self-explaining deep generative model trained end-to-end to embed cells into a structured, low-dimensional space organized around cell type-specific prototypes. Coupled with a specifically-designed data augmentation strategy, it matches or outperforms existing methods in cell type annotation across 11 large-scale datasets, particularly for rare cell types. Moreover, ProtoCloud improves data annotation quality by identifying and re-annotating misannotated training cells through a built-in certainty quantification mechanism based on cell-prototype similarity. Finally, ProtoCloud provides interpretable predictions by identifying key genes that drive its classifications, facilitating the discovery of both known and novel cell type marker genes. Applied to a time-course dataset of post-injury retinal neurons, ProtoCloud successfully annotates previously unassigned cells; on the esophageal cell atlas, it identifies rare but potentially important cell populations and their marker genes relevant to esophageal inflammation.