SpaGene: A Deep Adversarial Framework for Spatial Gene Imputation

Area of Science:

• Computational Biology
• Genomics
• Bioinformatics

Background:

• Single-cell RNA sequencing (scRNA-seq) offers high-resolution gene expression but lacks spatial context.
• Spatial transcriptomics provides spatial resolution but has limited transcriptomics coverage.
• Integrating both data types is crucial for a comprehensive understanding of tissue architecture and function.

Purpose of the Study:

• To introduce SpaGene, a novel deep learning framework for integrating scRNA-seq and spatial transcriptomics data.
• To impute missing gene expression data in spatial transcriptomics datasets using scRNA-seq information.
• To enhance the understanding of tissue biology, cellular interactions, and disease progression.

Main Methods:

• SpaGene utilizes a deep learning architecture comprising two encoder-decoder pairs, two translators, and two discriminators.
• The framework is designed to effectively impute gene expression within spatial transcriptomics datasets.
• Performance was benchmarked against six representative methods using a controlled gene-holdout evaluation protocol across diverse datasets.

Main Results:

• SpaGene demonstrated superior performance compared to baseline methods, improving average Pearson correlation coefficient and cell-wise structural similarity index.
• The model significantly reduced root mean squared error, indicating more accurate recovery of held-out spatial gene expression.
• Application to lung tumor tissue revealed spatial immune cell enrichment at tumor boundaries and improved detection of microenvironment-driven pathways.

Conclusions:

• SpaGene effectively integrates scRNA-seq and spatial transcriptomics data, providing accurate gene expression imputation.
• The framework offers valuable insights into tissue spatial patterns, immune cell distribution, and tumor-immune interactions.
• These findings support further biological validation and advance the study of complex biological systems.