CoSTA: unsupervised convolutional neural network learning for spatial transcriptomics analysis.

Area of Science:

• Genomics
• Computational Biology
• Bioinformatics

Background:

• Spatial transcriptomics technologies offer new insights into gene regulation within tissues.
• Identifying genes with similar spatial expression patterns can elucidate gene regulatory networks.
• Current methods often overlook spatial organization, treating data points independently.

Purpose of the Study:

• To introduce CoSTA, a novel approach for analyzing spatial transcriptomics data.
• To leverage convolutional neural networks (ConvNets) for clustering gene expression matrices based on spatial patterns.
• To reveal gene regulatory relationships by learning spatial similarities.

Main Methods:

• CoSTA utilizes convolutional neural network (ConvNet) clustering to analyze gene expression matrices.
• The method focuses on learning spatial similarities between genes.
• It quantifies expression pattern similarity, moving beyond simple categorization.

Main Results:

• CoSTA emphasizes broader spatial patterns over pixel-level correlations in gene expression.
• The approach provides a quantitative measure of expression pattern similarity between gene pairs.
• CoSTA identifies narrower, biologically relevant sets of significantly related genes compared to existing methods.

Conclusions:

• CoSTA offers a novel deep learning-based perspective on spatial transcriptomics analysis.
• The method effectively utilizes spatial information from neighboring pixels to understand expression patterns.
• CoSTA is applicable to various matrix-formatted spatial transcriptomics data and potentially histology images.