siVAE: interpretable deep generative models for single-cell transcriptomes.

Area of Science:

• Genomics
• Bioinformatics
• Machine Learning

Background:

• Variational autoencoders (VAEs) are used for genomic data dimensionality reduction.
• Current VAEs lack interpretability, obscuring the biological meaning of embedding dimensions.
• This limits their utility in downstream genomic analysis and discovery.

Purpose of the Study:

• To present siVAE, a novel interpretable VAE designed for genomic data analysis.
• To enhance the interpretability of dimensionality reduction techniques for biological data.
• To enable the identification of gene modules and hubs directly from interpretable embeddings.

Main Methods:

• Development of siVAE, a VAE architecture incorporating interpretability by design.
• Application of siVAE to genomic datasets for dimensionality reduction and feature representation analysis.
• Utilizing siVAE's interpretable embeddings to identify gene modules associated with specific phenotypes.

Main Results:

• siVAE provides interpretable dimensions, clarifying the biological features captured by the model.
• The model successfully identified gene modules and hubs without requiring explicit gene network inference.
• Connectivity within identified gene modules was significantly associated with phenotypes such as neuronal differentiation efficiency and dementia.

Conclusions:

• siVAE offers a powerful, interpretable alternative to standard VAEs for genomic data analysis.
• Interpretable generative models can facilitate the discovery of biologically relevant gene modules and hubs.
• siVAE demonstrates broad applicability in linking genetic patterns to diverse human phenotypes.