A deep neural network to de-noise single-cell RNA sequencing data.

Area of Science:

• Genomics
• Bioinformatics
• Computational Biology

Background:

• Single-cell RNA sequencing (scRNA-seq) is crucial for understanding cellular heterogeneity.
• scRNA-seq data analysis faces challenges due to high rates of measurement dropouts (zero gene expression).

Purpose of the Study:

• To introduce ZiPo, a deep neural network designed to accurately estimate rates and predict library sizes in scRNA-seq data.
• To address the issue of zero inflation and dropouts in scRNA-seq datasets.

Main Methods:

• Developed ZiPo, a deep artificial neural network incorporating adjustable zero inflation.
• Utilized deep autoencoders, Poisson, and negative binomial distributions.
• Implemented novel strategies like library size prediction and residual connections.

Main Results:

• ZiPo effectively captures dropouts by incorporating adjustable zero inflation.
• Introduced a scale-invariant loss term for improved model interpretability and sparsity.
• Demonstrated ZiPo's efficiency in handling large, singular, and mixed scRNA-seq datasets.

Conclusions:

• ZiPo offers an advanced solution for analyzing scRNA-seq data, particularly in handling zero inflation and dropouts.
• The model provides enhanced biological interpretability and efficient processing of large datasets.
• ZiPo shows advantages over existing techniques for scRNA-seq data analysis.