Predictive and robust gene selection for spatial transcriptomics.

Area of Science:

• Genomics
• Computational Biology
• Neuroscience

Background:

• Single-cell transcriptomics aims to integrate spatial information with cellular molecular states.
• Selecting optimal gene panels for spatial transcriptomics is a critical challenge.
• Current methods often target a small fraction of genes, potentially missing key biological insights.

Purpose of the Study:

• To introduce PERSIST, a deep learning framework for identifying informative gene targets for spatial transcriptomics.
• To leverage reference single-cell RNA sequencing (scRNA-seq) data for gene panel selection.
• To improve the efficiency and accuracy of spatial transcriptomics by optimizing gene panel composition.

Main Methods:

• Developed a flexible deep learning framework named PERSIST.
• Utilized reference scRNA-seq datasets from diverse brain regions, species, and technologies.
• Employed a binarization strategy for gene expression levels to facilitate model generalization.

Main Results:

• PERSIST reliably identifies gene panels that enable more accurate prediction of genome-wide expression profiles.
• The identified panels capture more biological information using significantly fewer genes.
• Models trained on scRNA-seq data generalized effectively to spatial transcriptomics data using PERSIST's approach.

Conclusions:

• PERSIST offers a powerful and adaptable method for selecting optimal gene panels in spatial transcriptomics.
• The framework enhances the ability to capture comprehensive cellular information with reduced gene sets.
• PERSIST facilitates the integration of scRNA-seq data with spatial transcriptomics, overcoming technological shifts.