Exploring the optimization of autoencoder design for imputing single-cell RNA sequencing data.

Area of Science:

• Computational Biology
• Bioinformatics
• Genomics

Background:

• Single-cell RNA sequencing (scRNA-seq) data often suffers from sparsity, necessitating imputation methods.
• Autoencoders are widely used for scRNA-seq data imputation, but optimal design choices are not well-established.
• Effective imputation is vital for accurate downstream analyses like cell clustering and differential gene expression.

Purpose of the Study:

• To empirically investigate the impact of autoencoder architecture and hyperparameters on scRNA-seq data imputation.
• To provide practical guidance for optimizing autoencoder design in single-cell bioinformatics.
• To compare findings with common practices in other machine learning fields.

Main Methods:

• Utilized diverse real and simulated scRNA-seq datasets.
• Systematically evaluated various neural network architectures (depth, width).
• Assessed different activation functions (e.g., sigmoid, tanh, ReLU) and regularization strategies.

Main Results:

• Deeper and narrower autoencoder architectures generally improved imputation performance.
• Sigmoid and tanh activation functions consistently outperformed ReLU and other common functions.
• Regularization enhanced imputation accuracy and the reliability of downstream cell clustering and differential expression analyses.

Conclusions:

• The study provides evidence-based recommendations for autoencoder design in scRNA-seq imputation.
• Optimal autoencoder configurations differ from those commonly used in computer vision.
• These findings facilitate more accurate and robust analysis of scRNA-seq data.