MARVEL: Microenvironment Annotation by Supervised Graph Contrastive Learning

Recent advancements in in situ molecular profiling technologies, including spatial proteomics and transcriptomics, have enabled detailed characterization of the microenvironment at cellular and subcellular levels. While these techniques provide rich information about individual cells’ spatial coordinates and expression profiles, extracting biologically meaningful spatial structures from the data remains a significant challenge. Current methodologies often rely on unsupervised clustering followed by cell type annotation based on differentially expressed genes within each cluster and most of the time will require other information as the reference (e.g., HE-stained images). This is labor-intensive and demands extensive domain knowledge. To address these challenges, we propose a supervised graph contrastive learning framework, MARVEL. MARVEL is a supervised graph contrastive learning method that can effectively embed local microenvironments represented by cell neighbor graphs into a continuous representation space, facilitating various downstream microenvironment annotation scenarios. By leveraging partially annotated examples as strong positives, our approach mitigates the common issues of false positives encountered in conventional graph contrastive learning. Using real-world annotated data, we demonstrate that MARVEL outperforms existing methods in three key microenvironment-related tasks: transductive microenvironment annotation, inductive microenvironment querying, and the identification of novel microenvironments across different slices.