FactVAE: a factorized variational autoencoder for single-cell multi-omics data integration analysis.

Area of Science:

• Computational biology
• Genomics
• Bioinformatics

Background:

• Single-cell multi-omics technologies enable simultaneous profiling of multiple molecular layers within individual cells, advancing the study of cell states and functions.
• Current data integration methods often fail to preserve critical feature information and leverage existing regulatory knowledge, limiting comprehensive cellular insights.

Purpose of the Study:

• To develop an innovative factorized variational autoencoder (FactVAE) for robust and accurate integration and analysis of single-cell multi-omics data.
• To enhance the preservation of feature information and incorporate known regulatory knowledge for improved understanding of cell functions.

Main Methods:

• FactVAE integrates a factorization principle into a variational autoencoder framework to preserve feature information and capture non-linear sample information.
• Incorporation of known regulatory knowledge during model training and utilization of a knowledge transfer strategy for cell embedding optimization and data augmentation.

Main Results:

• FactVAE demonstrated superior clustering performance compared to benchmark methods on diverse single-cell multi-omics datasets, including spatial multi-omics data.
• The method generated augmented data revealing clear cell-type-specific motif expression and enabled the inference of reliable gene regulatory relationships.

Conclusions:

• FactVAE offers a promising solution for single-cell multi-omics data analysis, providing superior performance, strong scalability, and enhanced biological insights.
• The model's ability to preserve feature information and leverage regulatory knowledge facilitates more accurate cell-type identification and gene regulatory network inference.