ZipAEr: A compressive convolutional autoencoder for high-dimensional spatial omics data at subcellular resolution.

Area of Science:

• Genomics
• Bioinformatics
• Computational Biology

Background:

• Spatial transcriptomics generates high-throughput data, but analytical complexity hinders biological interpretation.
• Existing methods often reduce data at the cell level, losing fine-grained spatial context.
• Conventional autoencoders are unsuitable for high-channel spatial omics data.

Purpose of the Study:

• To develop a novel computational approach, ZipAEr, for extracting informative latent features from spatial omics data.
• To address the limitations of traditional autoencoders in handling the high dimensionality of spatial omics datasets.
• To enable more effective downstream analysis of complex spatial transcriptomics data.

Main Methods:

• ZipAEr, a convolutional autoencoder, operates at the transcript level to preserve subcellular and extracellular spatial context.
• It reduces both spatial dimensions and channel count using convolutional layers.
• A channel weighting mechanism in the loss function ensures balanced representation of lowly expressed genes.

Main Results:

• ZipAEr achieves data compression of spatial omics data by two to three orders of magnitude.
• Key spatial and molecular features are effectively preserved during compression.
• The method enables computationally feasible downstream analyses like classification and clustering.

Conclusions:

• ZipAEr provides an effective method for analyzing complex spatial transcriptomics data.
• The transcript-level approach preserves essential spatial and molecular information.
• This facilitates advanced downstream analyses previously limited by computational constraints.