Self-supervised deep learning of gene-gene interactions for improved gene expression recovery

Area of Science:

• Genomics
• Bioinformatics
• Computational Biology

Background:

• Single-cell RNA sequencing (scRNA-seq) provides cellular-level biological insights but suffers from technical limitations causing inaccurate gene counts due to omitted low gene expression values.
• Current imputation methods, including deep learning, often fail to reliably impute gene expressions as they lack mechanisms to incorporate biological knowledge like gene-gene interactions.

Purpose of the Study:

• To develop a novel genomic data analysis framework that leverages gene-gene interactions for more reliable imputation of scRNA-seq data.
• To improve the identification of distinctive cellular patterns by extracting and integrating intricate biological characteristics.

Main Methods:

• Genes are repositioned into a 2D grid, spatially reflecting their interactive relationships.
• A self-supervised 2D convolutional neural network is utilized to extract contextual features from spatially configured genes, imputing omitted values without requiring labeled datasets.

Main Results:

• The proposed strategy demonstrates superior performance in gene expression imputation compared to existing methods.
• Extensive experiments on both simulated and real scRNA-seq datasets validate the effectiveness of the approach.

Conclusions:

• Leveraging gene-gene interactions through spatial configuration and self-supervised learning significantly enhances gene expression imputation in scRNA-seq data.
• This method offers a more accurate and biologically informed approach for analyzing single-cell genomic data, improving the discovery of cellular signatures.