deepMNN: Deep Learning-Based Single-Cell RNA Sequencing Data Batch Correction Using Mutual Nearest Neighbors

• DOI: 10.3389/fgene.2021.708981
• Publication Date: 2021-08-10T08:11:37Z
• Authors: Bin Zou; Tongda Zhang; Ruilong Zhou; Xiaosen Jiang; Huanming Yang; Xin Jin; Yong Bai
• Keywords: Batch processing; Regularization; Overfitting

It is well recognized that batch effect in single-cell RNA sequencing (scRNA-seq) data remains a big challenge when integrating different datasets. Here, we proposed deepMNN, novel deep learning-based method to correct scRNA-seq data. We first searched mutual nearest neighbor (MNN) pairs across batches principal component analysis (PCA) subspace. Subsequently, correction network was constructed by stacking two residual blocks and further applied for the removal of effects. The loss function deepMNN defined as sum weighted regularization loss. used compute distance between cells MNN PCA subspace, while make output similar input. experiment results showed can successfully remove effects datasets with identical cell types, non-identical multiple batches, large-scale well. compared performance state-of-the-art methods, including widely methods Harmony, Scanorama, Seurat V4 recently developed MMD-ResNet scGen. demonstrated achieved better or comparable terms both qualitative using uniform manifold approximation projection (UMAP) plots quantitative metrics such entropies, ARI F1 score, ASW score under various scenarios. Additionally, allowed one step. Furthermore, ran much faster than other These characteristics made it have potential be new choice gene expression analysis.