Similarity-assisted variational autoencoder for nonlinear dimension reduction with application to single-cell RNA...

Area of Science:

• Computational biology
• Machine learning
• Data visualization

Background:

• Deep generative models, like variational autoencoders (VAEs), are used for nonlinear dimension reduction and visualizing large datasets.
• VAEs can map samples to a latent space and generate data but often struggle to clearly display grouping patterns without annotations.
• Similarity-based methods (e.g., t-SNE, UMAP) show clear groupings but lack encoder/decoder structures.

Purpose of the Study:

• To enhance variational autoencoders (VAEs) for clearer visualization of complex datasets.
• To integrate similarity information into the VAE framework to improve latent space representation.
• To develop a conditional VAE for biological applications, accounting for covariate effects.

Main Methods:

• Proposed a novel approach by incorporating similarity information into the variational autoencoder (VAE) framework.
• Extended the VAE approach to a conditional VAE to manage covariate effects during dimension reduction.
• Utilized the UMAP loss function to encode similarity information within the VAE.

Main Results:

• The proposed method demonstrated superior performance in producing clear grouping structures compared to existing methods.
• Achieved effective dimension reduction for single-cell RNA sequencing datasets, revealing latent patterns.
• Successfully adjusted for covariate effects, leading to more interpretable and useful dimension reduction results.

Conclusions:

• The novel VAE approach effectively produces clearer biological grouping patterns by leveraging data similarity.
• The integration of similarity information, particularly via the UMAP loss function, enhances VAE performance.
• The method offers improved dimension reduction for large-scale biological data analysis.