Gene spatial integration: enhancing spatial transcriptomics analysis via deep learning and batch effect mitigation.

Area of Science:

• Genomics
• Bioinformatics
• Computational Biology

Background:

• Spatial transcriptomics (ST) is crucial for linking tissue organization to cellular functions.
• Current ST methods primarily focus on proximity, neglecting other spatial information like distribution.
• Batch effects from diverse sample origins and technologies hinder ST data analysis.

Purpose of the Study:

• To develop a deep learning method for integrated analysis of multiple ST datasets.
• To focus on the spatial distribution aspect of gene expression data.
• To leverage single-cell analysis tools for enhanced ST data interpretation.

Main Methods:

• Introduced Gene Spatial Integration (GSI), a data integration pipeline using representation learning.
• Employed an Autoencoder network to extract spatial embeddings and integrate them into gene expression feature space.
• Developed a method to handle multiple ST samples with minimal data loss and batch effect removal.

Main Results:

• GSI successfully integrates multiple ST samples, enhancing analysis tool performance.
• Significant improvements in clustering accuracy were observed when using GSI with Seurat and GraphST.
• For instance, GSI almost doubled the ARI score for Seurat clustering in sample 151673 (0.225 to 0.405).
• GSI improved GraphST clustering ARI score in sample 151672 from 0.614 to 0.795.

Conclusions:

• Gene distribution is a vital spatial aspect for ST data analysis.
• Integration and batch effect removal are critical for refined tissue characteristic analysis.
• The GSI pipeline offers a powerful approach for advanced ST data interpretation.