scLTF: A self-training clustering framework for large-scale single-cell RNA-Seq data

Abstract

Single-cell RNA sequencing (scRNA-seq) technology has rapidly advanced in recent years, driving significant breakthroughs in developmental biology, cancer research, immunology, and other related fields. However, existing clustering methods still face performance bottlenecks when handling large-scale scRNA-seq data. To address this challenge, we propose a self-training clustering method for large-scale single-cell RNA-seq data, termed scLTF (Single-cell Louvain-Transformer Framework for Large-scale Clustering). This method first generates preliminary clusters using a fast Louvain algorithm and then selects key cells based on a custom “Cluster Representativeness Index”. Subsequently, a Transformer model is employed for iterative representation learning and optimization to optimize the final clustering results. Experiments on multiple real-world datasets demonstrate that scLTF achieves overall superior or comparable performance in clustering accuracy compared with several state-of-the-art deep clustering methods, with evaluation metrics (ARI, NMI, and ACC) improving on average by approximately 3.9%–24.1%; meanwhile, its runtime is only about 6.7%–43.9% of existing deep learning-based methods. scLTF integrates the advantages of traditional graph-based clustering and deep learning models, achieving both high clustering performance and computational efficiency, thus providing a practical and reliable tool for biomedical research and cellular analysis.