scPerformer-Based Deep Clustering of Single Cell RNA Sequencing Data

Single-cell RNA sequencing (scRNA-seq) has revolutionized our understanding of cellular diversity, yet clustering scRNA-seq data remains challenging due to high dimensionality, sparsity, and batch effects. In this paper, we introduce scPerformer, a deep learning framework designed for efficient clustering of scRNA-seq data. The model leverages Performer, an optimized and scalable variant of the Transformer, integrating Principal Component Analysis (PCA) for dimensionality reduction and Convolutional Neural Networks (CNNs)for feature extraction, thereby achieving both high accuracy and computational efficiency. We assessed scPerformer on multiple scRNA-seq datasets, including Mouse Pancreas (Baron), Human Pancreas (Baron), Cell-bench Lung, PBMC (Zheng), Brain (MouseALM & HumanMTG), and Tabula Muris (TM) Mouse. Our results show that scPerformer significantly improves Mutual Information (MI) and Adjusted Rand Index (ARI) scores over leading methods, such as TransCluster and scDeepCluster. Furthermore, the linear attention mechanism in scPerformer drastically reduces computational time, enabling the model to efficiently process large datasets. Importantly, scPerformer demonstrated strong generalization to unseen data, maintaining high performance on external validation datasets without extensive preprocessing. These findings highlight scPerformer's potential as a powerful tool for large-scale biological data analysis, advancing both the accuracy and scalability of scRNA-seq clustering.