siVAE

siVAE provides interpretable deep generative modeling of single-cell transcriptomic data to produce low-dimensional embeddings and identify gene modules and hubs linked to phenotypes.

Key Features:

• Variational Autoencoder (VAE) backbone: Implements a deep generative model based on variational autoencoders for modeling single-cell transcriptome distributions.
• Interpretable latent space: Exposes which data features are represented by each latent dimension to improve interpretability of embeddings.
• Dimensionality reduction and visualization: Transforms high-dimensional single-cell transcriptomic data into a lower-dimensional representation suitable for visualization and analysis.
• Gene module and hub identification: Identifies gene modules and hub genes without requiring explicit gene network inference.
• Phenotype connectivity association: Links gene connectivity patterns inferred from the model to phenotypic outcomes.

Scientific Applications:

• Single-cell transcriptome interpretation: Interprets latent dimensions to reveal biological processes captured in single-cell RNA-seq data.
• Gene module discovery: Detects gene modules and hub genes that reflect coordinated expression programs.
• Phenotype association analysis: Associates gene module connectivity with phenotypic outcomes such as induced pluripotent stem cell (iPSC) neuronal differentiation efficiency and dementia.

Methodology:

Uses a variational autoencoder-based deep generative model with an interpretable latent space to perform dimensionality reduction into low-dimensional embeddings, identify gene modules and hubs without explicit gene network inference, and associate inferred gene connectivity patterns with phenotypes.