DCA

DCA denoises single-cell RNA sequencing (scRNA-seq) count data using a deep count autoencoder that models negative binomial noise with optional zero-inflation to capture overdispersion and dropout for improved downstream analysis.

Key Features:

• Deep count autoencoder: Uses a deep learning-based autoencoder network tailored to scRNA-seq count data to learn nonlinear representations of gene expression.
• Negative binomial noise model: Employs a negative binomial likelihood to model count distribution and overdispersion in scRNA-seq data.
• Zero-inflation option: Supports an adjustable zero-inflation model to account for dropout events in sparse single-cell counts.
• Nonlinear gene-gene dependency capture: Learns nonlinear relationships among genes to improve reconstruction of expression profiles.
• Scalability: Computational complexity scales linearly with the number of cells, enabling application to datasets containing millions of cells.
• Performance: Demonstrates improved denoising accuracy and faster runtime compared with existing imputation approaches.
• Evaluation: Validated on both simulated and real scRNA-seq datasets.

Scientific Applications:

• Denoising for downstream analyses: Improves input data quality for typical scRNA-seq analyses by reducing technical noise and dropout effects.
• Gene expression reconstruction: Provides more accurate reconstructed gene expression profiles to support biological interpretation and discovery.

Methodology:

Implements a deep autoencoder trained on raw scRNA-seq counts using a negative binomial noise model with an optional zero-inflation term; complexity scales linearly with the number of cells; evaluated on simulated and real datasets.