scVGAMF: a novel imputation method for scRNA-seq data by integrating linear and non-linear features.

Area of Science:

• Genomics
• Bioinformatics
• Computational Biology

Background:

• Single-cell RNA sequencing (scRNA-seq) reveals gene expression dynamics and cellular heterogeneity.
• Dropout events in scRNA-seq data present a significant challenge for accurate analysis.
• Existing imputation methods often rely on linear assumptions, neglecting complex regulatory relationships.

Purpose of the Study:

• To develop a novel imputation method, scVGAMF, that addresses dropout events in scRNA-seq data.
• To integrate both linear and non-linear features for improved imputation performance.
• To enhance downstream scRNA-seq data analyses, including clustering and differential gene identification.

Main Methods:

• scVGAMF utilizes a hybrid approach combining variational graph autoencoders and non-negative matrix factorization.
• It identifies highly variable genes, clusters cells, and constructs gene/cell similarity matrices.
• A neural network integrates linear and non-linear features for missing value prediction.

Main Results:

• scVGAMF demonstrates superior performance in gene expression recovery compared to existing methods.
• The method improves accuracy in cell clustering, differential gene identification, and pseudo-trajectory analysis.
• Ablation studies confirm the benefit of integrating both linear and non-linear features.

Conclusions:

• scVGAMF offers a robust solution for imputing dropout events in scRNA-seq data.
• The integration of diverse features significantly enhances the performance of scRNA-seq data analysis.
• This method advances the understanding of transcriptional regulation by improving data quality.