AdPrST:An Adversarial Graph Deep Learning Pre-Clustering Framework for Deciphering Spatiotemporal Structures in...

Spatially Resolved Transcriptomics (SRT) has revolutionized our understanding of gene expression within tissue microenvironments, yet accurately deciphering spatiotemporal structures-encompassing spatial domain identification, trajectory inference, and pseudo-spatiotemporal map construction-in complex tissues remains a formidable challenge. AdPrST begins with a pre-clustering process on gene expression data to establish initial domain groupings. It then constructs dual-view graph structures using K-Nearest Neighbors (KNN) for local similarities and r-radius for broader spatial contexts. Through adversarial self-supervised contrast, leveraging Wasserstein distance-based GANs and contrastive learning, AdPrST generates robust low-dimensional embeddings for each view. These embeddings are fused via a dot-product attention mechanism, guided by pre-clustering labels, to achieve accurate spatial domain identification. Benchmarking across multiple datasets demonstrated AdPrST's superior performance over state-of-the-art methods, highlighting its potential to advance spatial transcriptomics research by elucidating spatial functional patterns and developmental trajectories. In particular, AdPrST excels in inferencing spatiotemporal structures, reconstructing developmental sequences and temporal features in complex tissues.