Interpretable generative deep learning: an illustration with single cell gene expression data.

Area of Science:

• Computational Biology
• Bioinformatics
• Genomics

Background:

• Deep generative models (DGMs) excel at learning complex patterns in high-dimensional omics data.
• Techniques like variational auto-encoders (VAEs) provide low-dimensional latent representations of gene expression.
• Interpreting these latent spaces and their relation to observed biological data remains a challenge.

Purpose of the Study:

• To introduce and overview deep generative models for omics data analysis.
• To illustrate the application of DGMs, particularly VAEs, with single-cell gene expression data.
• To present methods for enhancing the interpretability of DGMs in biological contexts.

Main Methods:

• Overview of deep generative models, including variational auto-encoders.
• Generation of synthetic omics data to assess representation uncertainty.
• Development and illustration of methods to infer relationships between latent variables and observed data/phenotypes.

Main Results:

• DGMs effectively capture non-linear dependencies in high-dimensional omics data.
• Latent representations aid in understanding relationships between gene expression and experimental factors.
• Novel approaches successfully improve the interpretability of DGMs for biological data.

Conclusions:

• Deep generative models are powerful tools for uncovering structure in omics data.
• Enhanced interpretability methods make DGMs more accessible for biological discovery.
• The presented techniques demonstrate utility in single-cell gene expression analysis.