Learning interpretable representations of single-cell multi-omics data with multi-output Gaussian processes.

Area of Science:

• Computational Biology
• Genomics
• Machine Learning

Background:

• Single-cell genomics data presents challenges in representation learning due to its nonlinear and multi-modal nature.
• Existing methods often face a trade-off between the expressive power of black-box models and the interpretability of simpler methods.

Purpose of the Study:

• To develop a unified framework that balances expressive power and interpretability for single-cell genomics data.
• To learn distinct, interpretable representations for both cells and genes from multi-modal single-cell data.

Main Methods:

• A novel approach combining an embedding layer for representation learning with multi-output Gaussian processes for interpretability.
• Learning separate latent representations for samples (cells) and features (genes).
• Utilizing a gene relevance map to connect cell and gene clusters.

Main Results:

• Demonstrated that a few interpretable latent dimensions can effectively capture the underlying data structure.
• Successfully learned distinct representations for cells and genes.
• Established interpretable relationships between cell clusters and their associated marker genes.

Conclusions:

• The proposed model offers a powerful yet interpretable method for analyzing single-cell genomics data.
• The approach facilitates the discovery of biological insights by linking cell types to specific gene expression patterns.