FADVI: disentangled representation learning for robust integration of single-cell and spatial omics data

Area of Science:

• Computational biology
• Genomics
• Bioinformatics

Background:

• Integrating diverse omics datasets (scRNA-seq, scATAC-seq, spatial transcriptomics) is crucial but hindered by technical batch effects.
• Current methods struggle to distinguish technical variation from genuine biological signals, limiting downstream analysis.
• Accurate data integration is essential for understanding complex biological systems.

Purpose of the Study:

• To develop a robust framework, FADVI, for accurate integration of single-cell and spatial omics data.
• To disentangle technical batch effects from biological variation.
• To provide an interpretable method for large-scale omics data analysis.

Main Methods:

• FADVI utilizes a variational autoencoder framework with a partitioned latent space.
• It combines supervised classification, adversarial training, and cross-covariance penalty for independent representation learning.
• The method is benchmarked across single-cell RNA sequencing (scRNA-seq), single-cell ATAC sequencing (scATAC-seq), and spatial transcriptomics datasets.

Main Results:

• FADVI demonstrated superior performance compared to state-of-the-art integration methods across multiple omics modalities.
• The framework successfully preserved biological variation while effectively correcting for batch effects.
• FADVI enabled feature attribution, identifying genes linked to cell type identity and batch-specific variations.

Conclusions:

• FADVI offers a powerful and interpretable solution for integrating large-scale single-cell and spatial omics data.
• The method overcomes limitations of existing approaches by effectively separating technical and biological variation.
• FADVI provides a robust framework for advancing downstream analyses and biological discoveries.