Integration of scRNA-seq data by disentangled representation learning with condition domain adaptation

Area of Science:

• Genomics
• Bioinformatics
• Computational Biology

Background:

• Integrating single-cell RNA sequencing (scRNA-seq) data from multiple batches and conditions is crucial for studying cellular heterogeneity.
• Existing methods face challenges in effectively handling diverse datasets and disentangling biological variation from technical artifacts.

Purpose of the Study:

• To develop an advanced scRNA-seq data integration method named scDisco.
• To effectively reduce batch effects while preserving and disentangling biological and condition-specific variations.

Main Methods:

• scDisco employs a domain-adaptive decoupling representation learning strategy using variational autoencoders.
• It constructs condition-specific domain-adaptive networks with Domain-Specific Batch Normalization layers.

Main Results:

• scDisco successfully reduces batch effects and disentangles biological from condition-specific effects.
• The method enhances condition-specific representations, enabling the identification of condition-specific genes.
• Evaluations on simulated and real datasets demonstrate scDisco's effectiveness in visualization, cell clustering, and gene identification.

Conclusions:

• scDisco is an effective variational autoencoder-based integration method.
• It significantly improves scRNA-seq data analysis tasks, including batch effect reduction, cell clustering, and identification of condition-specific genes.