The repetitive analysis of similar scRNA-seq samples for cell identification is both time-consuming and labor-intensive. Transferring annotations from a reference to a query dataset offers innovative strategies for single-cell annotation, with deep learning’s prowess in feature extraction serving as a robust foundation for these approaches. In this paper, we present the Multi-level Feature Extractor-attention model (MLFE-Att), a deep learning framework that leverages the multi-head attention mechanism to learn intricate cellular correlation features, thereby enhancing the identification of cell categories with minor differences. Deeper and non-linear features of cells are learned by the module that contains multiple fully connected layers and activation layers. Experimental evaluations on the bone marrow scRNA-seq dataset demonstrate that MLFE-Att outperforms the other two baseline models, MLFE-CNN and MLFE-LSTM, with a 99.84% identification accuracy. The results on the new subset of the bone marrow dataset indicate that MLFE-Att offers dependable support for identifying cells in the new sample, achieving a 92.76% accuracy, and supplies insightful feedback for manual cellular labeling.