Denoising single-cell RNA-seq data with a deep learning-embedded statistical framework.

Area of Science:

• Computational biology
• Genomics
• Bioinformatics

Background:

• Single-cell RNA sequencing (scRNA-seq) data is crucial for understanding cellular heterogeneity.
• Technical noise and zero counts in scRNA-seq present significant analytical challenges.
• Existing imputation methods have limitations in capturing complex gene expression patterns or lack interpretability.

Purpose of the Study:

• To develop a novel computational framework for robust imputation of scRNA-seq data.
• To effectively decompose technical variability while preserving biological heterogeneity.
• To improve downstream analyses such as cell type classification and differential expression.

Main Methods:

• Introduced ZILLNB (Zero-Inflated Latent factors Learning-based Negative Binomial), a framework combining zero-inflated negative binomial (ZINB) regression with deep generative modeling.
• Employed an ensemble architecture of Information Variational Autoencoder (InfoVAE) and Generative Adversarial Network (GAN) to learn latent representations.
• Utilized an Expectation-Maximization algorithm for iterative optimization of ZINB regression parameters.

Main Results:

• ZILLNB outperformed existing methods in cell type classification tasks (mouse cortex, human PBMC) with higher Adjusted Rand index (ARI) and Adjusted Mutual Information (AMI).
• Demonstrated superior performance in differential expression analysis, showing improved area under the ROC (AUC-ROC) and Precision-Recall (AUC-PR) curves with lower false discovery rates.
• Identified distinct fibroblast subpopulations in idiopathic pulmonary fibrosis (IPF) datasets, validated by marker gene expression and pathway analysis.

Conclusions:

• ZILLNB offers a principled approach to handle technical artifacts in scRNA-seq data, preserving biological variation.
• The integration of statistical and deep learning methods ensures robust performance across various single-cell analysis tasks.
• ZILLNB is effective for cell type identification, differential expression analysis, and discovering rare cell populations.