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

Area of Science:

• Computational Biology
• Genomics
• Bioinformatics

Background:

• Batch effects are a significant challenge in integrating single-cell RNA sequencing (scRNA-seq) datasets, hindering downstream analysis.
• Accurate integration of scRNA-seq data is crucial for robust biological insights.

Purpose of the Study:

• To develop and evaluate deepMNN, a novel deep learning-based method for effective batch effect correction in scRNA-seq data.
• To demonstrate deepMNN's capability in handling diverse and large-scale scRNA-seq datasets.

Main Methods:

• Mutual nearest neighbor (MNN) pairs were identified in a principal component analysis (PCA) subspace.
• A batch correction network, comprising stacked residual blocks, was designed for batch effect removal.
• A loss function combining batch loss and regularization loss was utilized to guide the network training.

Main Results:

• deepMNN successfully corrected batch effects across datasets with identical and non-identical cell types, multiple batches, and large scales.
• Performance evaluation showed deepMNN achieved superior or comparable results against state-of-the-art methods (Harmony, Scanorama, Seurat V4, MMD-ResNet, scGen) using qualitative (UMAP) and quantitative metrics.
• deepMNN demonstrated efficient integration of multi-batch scRNA-seq data in a single step and significantly faster processing for large datasets.

Conclusions:

• deepMNN is a powerful and efficient deep learning tool for batch effect correction in scRNA-seq data.
• Its ability to handle complex datasets and its speed make it a promising new option for large-scale single-cell gene expression analysis.