Learning context-aware, distributed gene representations in spatial transcriptomics with SpaCEX

Distributed gene representations are pivotal in data-driven genomic research, offering a structured way to understand the complexities of genomic data and providing foundation for various data analysis tasks. Current gene representation learning methods demand costly pretraining on heterogeneous transcriptomic corpora, making them less approachable and prone to over-generalization. For spatial transcriptomics (ST), there is a plethora of methods for learning spot embeddings but serious lacking method for generating gene embeddings from spatial gene profiles. In response, we present SpaCEX, a pioneer cost-effective self-supervised learning model that generates gene embeddings from ST data through exploiting spatial genomic "context" identified as spatially co-expressed gene groups. SpaCEX-generated gene embeddings (SGE) feature in context-awareness, rich semantics, and robustness to cross-sample technical artifacts. Extensive real data analyses reveal biological relevance of SpaCEX-identified genomic contexts and validate functional and relational semantics of SGEs. We further develop a suite of SGE-based computational methods for a range of key downstream objectives: identifying disease-associated genes and gene-gene interactions, pinpointing genes with designated spatial expression patterns, enhancing transcriptomic coverage of FISH-based ST, detecting spatially variable genes, and improving spatial clustering. Extensive real data results demonstrate these methods superior performance, thereby affirming the potential of SGEs in facilitating various analytical task. Significance StatementSpatial transcriptomics enables the identification of spatial gene relationships within tissues, providing semantically rich genomic "contexts" for understanding functional interconnections among genes. SpaCEX marks the first endeavor to effectively harnesses these contexts to yield biologically relevant distributed gene representations. These representations serve as a powerful tool to greatly facilitate the exploration of the genetic mechanisms behind phenotypes and diseases, as exemplified by their utility in key downstream analytical tasks in biomedical research, including identifying disease-associated genes and gene interactions, in silico expanding the transcriptomic coverage of low-throughput, high-resolution ST technologies, pinpointing diverse spatial gene expression patterns (co-expression, spatially variable pattern, and patterns with specific expression levels across tissue domains), and enhancing tissue domain discovery.