scDeepC3: scRNA-seq Deep Clustering by A Skip AutoEncoder Network with Clustering Consistency

Abstract

Single-cell RNA sequencing (scRNA-seq) reveals the heterogeneity and diversity among individual cells and allows researchers conduct cell-wise analysis. Clustering analysis is a fundamental step in analyzing scRNA-seq data which is needed in many downstream tasks. Recently, some deep clustering based methods exhibit very good performance by combining the AutoEncoder reconstruction-based pre-training and the fine-tune clustering. Their common idea is to cluster the samples by the learned features from the bottleneck layer of the pre-trained model. However, these reconstruction-based pre-training cannot guarantee that the learned features are beneficial to the clustering. To alleviate these issues, we propose an improved scRNA-seq Deep Clustering method by a skip AutoEncoder network with Clustering Consistency (i.e., named scDeepC3) from two aspects, an efficient network structure and a stable loss function. In particular, we introduce an adaptive shortcut connection layer to directly add the shallow-layer (encoder) features to deep-layer (decoder). This will increase the flow of forward information and back-forward gradients, and make the network training more stable. Considering the complementarity between the features of different layers, which can be seen as different views of the original samples, we introduce a clustering consistency loss to make the clustering results of different views consistent. Experimental results demonstrate that our proposed scDeepC3 achieves better performance than state-of-the-arts and the detailed ablation studies are conducted to help us understand how these parts make sense.