CoSTA

CoSTA applies unsupervised convolutional neural network (CNN) learning to quantify and cluster spatial gene expression patterns in spatial transcriptomics gene expression matrices.

Key Features:

• Convolutional neural network clustering: Employs unsupervised CNNs to cluster gene expression matrices and learn spatial similarities between genes.
• Quantitative pairwise pattern similarity: Computes a quantitative measure of similarity for each pair of gene expression patterns.
• Pattern-focused spatial analysis: Emphasizes broader spatial expression patterns over pixel-level correlations when comparing genes.
• Validation on diverse datasets: Analyzes simulated and previously published spatial transcriptomics data to discern spatial relationships across tissues and cell types.
• Applicability to multiple data types: Operates on any spatial transcriptomics dataset represented in matrix form and can be extended to histology images from similar but not identical biological sections.
• Narrow gene-set identification: Identifies narrower sets of significantly related genes based on spatial expression-pattern similarity.

Scientific Applications:

• Gene regulatory network inference: Identifies spatially related gene sets to inform reconstruction of spatial gene regulatory networks.
• Cell-type spatial regulatory analysis: Discerns spatial relationships of gene expression across different cell types within tissues.
• Developmental biology: Elucidates spatial gene regulation patterns relevant to developmental processes.
• Cancer research: Maps spatially organized expression patterns within tumors to inform tumor biology studies.
• Regenerative medicine: Provides spatial expression context relevant to tissue repair and regeneration studies.

Methodology:

Unsupervised convolutional neural network (CNN) clustering of gene expression matrices and computation of quantitative pairwise expression-pattern similarity, applied to simulated and previously published spatial transcriptomics datasets.