DISC

DISC infers gene structure and reconstructs gene expression profiles obscured by dropout events in single-cell RNA sequencing (scRNA-seq) data to improve cell type identification and analysis of cellular heterogeneity.

Key Features:

• Dropout inference: Reconstructs expression signals lost to dropout events in scRNA-seq datasets.
• Deep learning network: Employs a novel deep learning network to model relationships between gene structure and expression.
• Semi-supervised learning: Uses semi-supervised learning to leverage both labeled and unlabeled data.
• Gene structure and expression inference: Infers gene structural features and expression levels obscured by dropouts.
• Reconstruction/imputation: Reconstructs gene expression profiles from sparse scRNA-seq data as an alternative to traditional imputation.
• Benchmarking: Demonstrated superior performance over seven state-of-the-art imputation methods across ten real-world datasets.
• Reduced false signals: Limits introduction of false signals compared with traditional imputation techniques.
• Improved downstream analysis: Enhances reliability of downstream analyses such as cell type identification and structural characterization of genes.
• Generalization: Semi-supervised approach improves generalization across different datasets.

Scientific Applications:

• Single-cell transcriptomics: Recovering missing expression data and improving analysis of scRNA-seq studies.
• Cell type identification: Improving accuracy of cell type classification by reconstructing expression profiles.
• Analysis of cellular heterogeneity: Providing a more accurate representation of cellular heterogeneity by reducing dropout-induced artifacts.
• Gene structure characterization: Inferring gene structural features obscured by sparse sequencing data.
• Imputation benchmarking: Serving as an alternative method and benchmark in comparative studies of imputation techniques.

Methodology:

DISC employs a novel deep learning network and semi-supervised learning to reconstruct gene expression profiles from sparse scRNA-seq data, leveraging both labeled and unlabeled samples.