TarDis: Achieving robust and structured disentanglement of multiple covariates

Area of Science:

• Computational Biology
• Genomics
• Machine Learning

Background:

• Single-cell genomics data exhibit significant heterogeneity due to technical artifacts and biological variations.
• Integrating multi-source datasets in single-cell genomics presents challenges in maintaining biological signal integrity.
• Existing methods struggle to effectively disentangle technical covariates from true biological differences.

Purpose of the Study:

• To develop a deep generative model for disentangling covariate structures in diverse single-cell genomics datasets.
• To distinguish technical artifacts from biological variations for improved data integration.
• To create interpretable latent representations for hypothesis-driven research.

Main Methods:

• Introduction of targeted disentanglement (TarDis), an end-to-end deep generative model.
• Utilizing covariate-specific loss components and a self-supervised approach.
• Generating multiple latent-space representations for continuous and categorical covariates, plus unexplained variation.

Main Results:

• TarDis outperforms existing methods in data integration and covariate disentanglement.
• The model demonstrates robust out-of-distribution prediction capabilities.
• TarDis generates interpretable and structured latent spaces, including ordered representations for continuous covariates.

Conclusions:

• TarDis provides a powerful analytical platform for single-cell genomics research.
• The model enhances the ability to gain insights into cellular dynamics.
• TarDis facilitates targeted therapeutic interventions by clarifying biological variations.