scDHA

scDHA performs decomposition and dimensionality reduction of single-cell RNA sequencing (scRNA-seq) data to extract cell-level signals from large, noisy datasets.

Key Features:

• Hierarchical autoencoder framework: Implements a single-cell Decomposition using Hierarchical Autoencoder architecture composed of two principal modules.
• Non-negative kernel autoencoder: Filters genes or components that contribute insignificantly to the data's part-based representation.
• Feature filtering: Retains relevant features for downstream analysis by removing low-contribution genes/components.
• Stacked Bayesian autoencoder: Projects refined data into a low-dimensional space for complexity reduction.
• Compressed-space perturbation: Repeatedly perturbs the compressed latent space to mitigate overfitting and learn more generalized representations.
• Scalability and noise handling: Tailored to manage large scRNA-seq datasets and high noise levels.
• Unsupervised cell segregation: Performs effective unsupervised segregation of cells.
• Transcriptome visualization: Provides detailed visualization of transcriptome landscapes.
• Cell classification: Enables precise cell classification from single-cell profiles.
• Pseudo-time inference: Supports accurate pseudo-time trajectory inference.

Scientific Applications:

• Unsupervised clustering: Segregates cell populations using unsupervised learning on scRNA-seq data.
• Cell type classification: Classifies cells into distinct types based on learned low-dimensional representations.
• Trajectory and pseudo-time analysis: Infers developmental or temporal trajectories via pseudo-time ordering.
• Transcriptome landscape visualization: Visualizes global and local transcriptomic structures from single-cell datasets.
• Large-scale noisy data analysis: Analyzes large, noisy scRNA-seq datasets while reducing dimensionality and noise impact.

Methodology:

scDHA applies a non-negative kernel autoencoder for gene/component filtering, a stacked Bayesian autoencoder for low-dimensional projection, and repeated perturbation of the compressed latent space to reduce overfitting.