scAAnet

scAAnet performs non-linear archetypal analysis of single-cell RNA sequencing (scRNA-seq) data to identify gene expression programs (GEPs) and infer their relative activities across cells.

Key Features:

• Non-Linear Archetypal Analysis: Identifies continuous spectra of cellular states and shared gene expression programs (GEPs) across cell types rather than discrete clusters.
• Autoencoder Architecture: Uses deep autoencoders to capture non-linear relationships in scRNA-seq data in an unsupervised manner.
• Count Distribution-Based Loss Term: Incorporates a loss term that accounts for sparsity and overdispersion characteristic of raw count scRNA-seq data.
• Archetypal Constraint: Adds an archetypal constraint to the loss function to guide identification of representative GEPs (archetypes) within the dataset.

Scientific Applications:

• Extraction of biologically meaningful GEPs: Applied to publicly available scRNA-seq datasets to recover interpretable gene expression programs and their activities across cells.
• Pancreatic islet dataset: Revealed insights into cellular states and functions within pancreatic tissues.
• Lung idiopathic pulmonary fibrosis dataset: Characterized gene expression dynamics relevant to lung disease.
• Prefrontal cortex dataset: Provided information on neuronal diversity and function in the prefrontal cortex.

Methodology:

Data preprocessing of raw count scRNA-seq with consideration for sparsity and overdispersion. Model training of deep autoencoders using a custom loss function that includes a count distribution-based term and an archetypal constraint. Performance evaluation via simulations demonstrating improved accuracy and biological relevance compared to existing methods.