scDAEC: deep clustering of single-cell data based on dual autoencoder

Abstract

Single-cell RNA sequencing (scRNA-seq) provides high-resolution transcriptomic profiles for dissecting cellular heterogeneity, but accurate clustering remains difficult because gene-expression matrices are sparse, noisy, and affected by dropout events. Existing deep clustering models often emphasize either gene-expression reconstruction or cell–cell structural learning, which may limit their ability to jointly account for gene-expression sparsity and preserve cellular topology. Here, we propose scDAEC, a dual-autoencoder framework for unsupervised clustering of scRNA-seq data. scDAEC applies an adaptive zero-inflated negative binomial (ZINB) based denoising autoencoder (DAE) to model gene-expression uncertainty, combined with a graph autoencoder (GAE) to learn topology-aware cell representations. A dynamic combined multi-head attention (DCMHA) module then integrates denoised gene-expression embeddings and graph-based embeddings for clustering optimization. Experiments on 13 real scRNA-seq datasets showed that scDAEC achieved improved clustering accuracy and stable performance compared with other representative methods. Based on the dataset-level mean results, scDAEC achieved overall mean Adjusted Rand Index (ARI) and Normalized Mutual Information (NMI) values of 0.864 and 0.859, respectively. Compared with the strongest competing methods, scDAEC improved the average ARI and NMI by 10.79% and 10.96%, respectively. The source code is available at https://github.com/chanDaniel-cc/scDAEC .