Imputation of single-cell gene expression with an autoencoder neural network

Area of Science:

• Genomics
• Bioinformatics
• Computational Biology

Background:

• Single-cell RNA sequencing (scRNA-seq) offers high-resolution gene expression analysis but suffers from high dropout rates, leading to zero read counts and complicating downstream analyses.
• Dropout events in scRNA-seq data represent a significant challenge for accurate interpretation of gene expression profiles.

Purpose of the Study:

• To develop advanced imputation methods for addressing dropout events in scRNA-seq data.
• To enhance the accuracy and scalability of scRNA-seq data analysis through effective imputation.

Main Methods:

• Development of nonparametric deep learning imputation methods: LATE (Learning with AuToEncoder) and TRANSLATE (TRANSfer learning with LATE).
• LATE utilizes autoencoders trained with random initial parameters.
• TRANSLATE leverages reference gene expression datasets for improved initial parameter estimation in LATE.

Main Results:

• LATE and TRANSLATE demonstrate superior performance compared to existing scRNA-seq imputation methods on both simulated and real datasets.
• The proposed methods achieve lower mean squared error, accurately recover nonlinear gene-gene relationships, and improve cell type separation.
• Both methods exhibit high scalability, capable of processing over one million cells within hours on a GPU.

Conclusions:

• The nonparametric autoencoder-based imputation approach is highly effective and efficient for scRNA-seq data.
• LATE and TRANSLATE offer a powerful solution for overcoming dropout limitations in single-cell genomics.