scIGANs

scIGANs performs imputation and denoising of single-cell RNA sequencing (scRNA-seq) data using generative adversarial networks (GANs) to recover dropout (false zero) values while preserving cell-to-cell variability.

Key Features:

• Imputation of Dropouts: Uses GANs to generate synthetic cells that impute dropout (false zero) values and recover biologically meaningful gene expression levels.
• Avoidance of Oversmoothing: Performs imputation using generated rather than observed cells to preserve inherent stochasticity and avoid oversmoothing of cell-to-cell variability.
• Balanced Performance Across Cell Populations: Balances imputation performance between major and rare cell populations to maintain representation of rare cell types.
• Robustness to Dataset Variability: Operates robustly on small datasets with low expression or minimal cell-to-cell variance, is compatible with multiple scRNA-seq protocols, and scales to datasets with over 100,000 cells.
• Enhancement of Downstream Analysis: Improves the reliability of downstream analyses by providing higher-quality imputed expression data for transcriptomic profiling and cellular state inference.
• Implementation: Implemented in PyTorch with support for GPU acceleration and CPU computing.

Scientific Applications:

• Transcriptomic Profiling: Facilitates characterization of transcriptomic profiles at single-cell resolution to study cellular heterogeneity.
• Comparative Studies: Provides a competitive alternative for benchmarking and comparing imputation methods in scRNA-seq analyses.

Methodology:

Employs generative adversarial networks (GANs) to generate synthetic cells used for imputation of scRNA-seq dropouts; implemented in PyTorch with GPU and CPU support.