scLDS2

scLDS2 models rare cell types in single-cell RNA-sequencing (scRNA-seq) data using a deep generative framework to identify and classify rare cell populations.

Key Features:

• Adversarial Learning for Discrimination: Employs adversarial learning to estimate cell-type distributions and distinguish rare from non-rare cells in scRNA-seq data.
• Sparse Sample Generation with ℓ1-norm: Generates sparse synthetic samples via ℓ1-norm regularization to highlight essential features and improve interpretability.
• Block Structure Learning with Nuclear-norm Optimization: Learns block structures through nuclear-norm optimization to group similar cells and directly identify cell types from generated samples.
• Unified Generative Framework: Integrates sample generation, classification of generated and true samples, and feature extraction, transforming rare cell type detection into a classification problem for joint learning.

Scientific Applications:

• Biomarker discovery: Enables identification of cell-type–specific biomarkers from scRNA-seq data by improving discrimination of rare populations.
• Cancer diagnostics and therapy development: Facilitates detection and characterization of rare malignant or tumor-infiltrating cell populations relevant to cancer research.
• Rare-cell analysis in genomics: Applies to studies requiring precise analysis of rare cell types across diverse scRNA-seq datasets.

Methodology:

Combines deep generative modeling with adversarial learning, ℓ1-norm regularization for sparse sample generation, nuclear-norm optimization for block-structure learning, and a unified framework that jointly performs sample generation, classification of generated and true samples, and feature extraction.