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

Bin Zou; Tongda Zhang; Ruilong Zhou; Xiaosen Jiang; Huanming Yang; Xin Jin; Yong Bai

doi:10.3389/fgene.2021.708981

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

Front Genet. 2021 Aug 10:12:708981. doi: 10.3389/fgene.2021.708981. eCollection 2021.

Authors

Bin Zou¹, Tongda Zhang¹, Ruilong Zhou^{1

2}, Xiaosen Jiang^{1

2}, Huanming Yang^{1

3}, Xin Jin^{1

4

5}, Yong Bai¹

Affiliations

¹ BGI-Shenzhen, Shenzhen, China.
² College of Life Science, University of Chinese Academy of Sciences, Beijing, China.
³ James D. Watson Institute of Genome Sciences, Hangzhou, China.
⁴ School of Medicine, South China University of Technology, Guangzhou, China.
⁵ Guangdong Provincial Key Laboratory of Human Disease Genomics, Shenzhen Key Laboratory of Genomics, BGI-Shenzhen, Shenzhen, China.

Abstract

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

Keywords: batch effect correction; deep learning; mutual nearest neighbor; residual network; scRNA-seq data integration.