Model-based autoencoders for imputing discrete single-cell RNA-seq data.

Area of Science:

• Bioinformatics
• Genomics
• Computational Biology

Background:

• Discrete data imputation is under-explored, particularly for single-cell RNA sequencing (scRNA-seq) data, which often contains prevalent 'false' zero counts due to dropout events.
• Accurate imputation of missing values in scRNA-seq data is crucial for effective downstream analyses, including cell type identification and differential expression analysis.

Purpose of the Study:

• To develop a novel imputation method for discrete scRNA-seq data using a Zero-Inflated Negative Binomial (ZINB) model-based autoencoder.
• To explicitly model dropout events using the Gumbel-Softmax distribution and optimize zero-inflated reconstruction against the raw count matrix.

Main Methods:

• Proposed a novel Zero-Inflated Negative Binomial (ZINB) model-based autoencoder for discrete data imputation.
• Incorporated the Gumbel-Softmax distribution to explicitly model dropout events causing missing values.
• Optimized the zero-inflated reconstruction with respect to the raw count matrix.

Main Results:

• Extensive experiments on simulation datasets demonstrated significant improvements in imputation accuracy.
• Real data experiments showed enhanced separation of distinct cell types.
• The proposed imputation method improved the accuracy of differential expression analysis.

Conclusions:

• The novel ZINB autoencoder effectively imputes discrete scRNA-seq data by addressing zero-inflation and dropout events.
• The imputation method enhances the biological interpretability of scRNA-seq data by improving cell type classification and differential expression analysis.