Abstract
Recent advances in spatial transcriptomics have produced rich, high-throughput datasets, but their biological interpretation remains challenging due to analytical complexity. We present ZipAEr, a convolutional autoencoder tailored to extract informative latent features from spatial omics data. Unlike traditional methods that reduce data at the cell level, ZipAEr operates at the transcript level, preserving both subcellular and extracellular spatial context. Conventional autoencoders, built for images with three channels (red, green, blue), cannot handle spatial omics data with thousands of input channels representing genes and proteins. ZipAEr addresses this by reducing both spatial dimensions and channel count through its convolutional layers. It also introduces channel weighting in the loss function to ensure balanced representation of lowly expressed genes. ZipAEr effectively compresses spatial omics data by two to three orders of magnitude while preserving key spatial and molecular features. The resulting latent representation enables downstream analyses, such as classification and clustering, which would otherwise be computationally infeasible with raw data.