MaskGraphene: Advancing joint embedding, clustering, and batch correction for spatial transcriptomics using graph-based self-supervised learning

Abstract

Recent advancements in spatial transcriptomics (ST) have underscored the importance of integrating data from multiple ST slices for joint analysis. A major challenge remains generating interpretable joint embeddings that preserve geometric information for downstream analyses. Here we introduce MaskGraphene, a graph neural network that combines self-supervised and self-contrastive training to integrate gene expression and spatial location into joint embeddings. By employing clusterwise alignment and a graph attention autoencoder with masked self-supervised and triplet loss optimizations, MaskGraphene effectively preserves geometric structures while achieving batch correction. In benchmarks against seven state-of-the-art methods, MaskGraphene consistently demonstrated superior alignment accuracy and geometric fidelity across diverse ST datasets. Its interpretable embeddings significantly enhanced downstream applications, including domain identification, spatial trajectory reconstruction, biomarker discovery, and the creation of topographical maps of brain slices. Notably, MaskGraphene successfully recovered layer-wise brain structures with near-perfect accuracy. MaskGraphene provides a powerful and versatile framework for advancing ST data integration and analysis, unlocking valuable biological insights.