AAE-SC: A scRNA-Seq Clustering Framework Based on Adversarial Autoencoder

Abstract

Single-cell RNA sequencing (scRNA-seq) provides the expression profiles of individual cells, and it is expected to provide higher cellular differential resolution than traditional bulk RNA sequencing. In scRNA-seq analysis, clustering is crucial for identifying cell types, and can be potentially exploited to understand high-level biological processes. Recently, autoencoder has been successfully applied in scRNA-seq clustering problem and achieved promising results. Most existing works focus on characterizing the sparsity of data, and directly utilize the bottleneck feature of the autoencoder for clustering might not be optimal. In this paper, a novel framework named Adversarial AutoEncoder ScRNA-seq Clustering (AAE-SC) is proposed to bring an additional constraint on the bottleneck feature. Specifically, AAE-SC adds an AAE module on top of the bottleneck layer, and constrains the bottleneck feature distribution to be aligned with a consistent distribution. Also, the AAE and the reconstructed modules are jointly optimized to generate a highly discriminative and consistent feature, which is further proceeded for clustering. We find that by using AAE-SC to impose certain constraints on the features of the hidden layer, the performance of clustering can be improved. Experimental results on three real-world datasets demonstrate that the proposed AAE-SC framework outperformed state-of-the-art methods by 2% at least and 5% at most. And AAE-SC shows more robustness than the baseline model for downsampled and unbalanced cluster size datasets.

Keywords

• Single-cell RNA-seq data
• adversarial autoencoder
• clustering analysis
• unsupervised learning