Wasserstein Graph Convolutional Network with Attention for Imbalanced scRNA-seq Data Knowledge Discovery

Discovering complex molecular patterns in imbalanced scRNA-seq data remains a challenge, despite numerous efforts from different perspectives. In this study, we propose a novel deep learning model: the Wasserstein Graph Convolutional Network with an attention mechanism (twGCN), designed for semi-supervised learning to address this challenge. The proposed model overcomes the weaknesses of traditional Graph Convolutional Networks by capturing more data intricacies and geometry. This is achieved by integrating a Wasserstein distance-based loss function optimization along with an attention mechanism. Unlike traditional scRNA-seq data preprocessing, we employ a robust scaling approach to normalize scRNA-seq data, which generally contains a large number of outliers. Our methods demonstrate significant advantages over peer methods in discovering single-cell patterns in benchmark data. More importantly, the proposed twGCN can handle both low-dimensional and high-dimensional scRNA-seq data obtained after feature selection. To our knowledge, this study will positively impact both deep learning and bioinformatics, inspiring future research.