PAST

PAST integrates prior information with a variational graph convolutional autoencoder and self-attention to characterize spatial domains in spatial transcriptomics (ST) data.

Key Features:

• Variational Graph Convolutional Autoencoder: Employs a variational graph convolutional autoencoder to integrate ST data and capture complex spatial patterns.
• Prior Information Integration (Bayesian neural network): Incorporates prior information via a Bayesian neural network framework to leverage reference datasets for context-aware analysis.
• Self-Attention Mechanism: Uses a self-attention mechanism to model spatial dependencies and emphasize relevant features across spatial domains.
• Ripple Walk Sampler Strategy: Applies a ripple walk sampler strategy to enable scalable sampling and processing of large ST datasets.
• Multislice Joint Embedding and Automatic Annotation: Supports multislice joint embedding and automatic annotation of spatial domains across multiple slices or datasets.
• Spatial Trajectory and Pseudotime Analysis: Enables inference of spatial trajectories and pseudotime to examine cellular dynamics and developmental processes.
• Visualization of Spatial Heterogeneity: Produces representations that aid visualization of spatial heterogeneity within biological samples.

Scientific Applications:

• Spatial Domain Characterization: Characterizes spatial domains in ST data to delineate tissue architecture and region-specific expression patterns.
• Visualization: Facilitates visualization of spatial heterogeneity for interpretation of spatial expression landscapes.
• Spatial Trajectory Inference and Pseudotime Analysis: Supports inference of spatial trajectories and pseudotime analyses to investigate cellular dynamics and developmental processes.
• Multislice Joint Embedding and Automatic Annotation: Enables joint embedding and automatic annotation across multiple slices or datasets to harmonize and compare spatial domains.

Methodology:

Computational methods explicitly include a variational graph convolutional autoencoder, a Bayesian neural network for prior/reference-data integration, a self-attention mechanism, and a ripple walk sampler strategy.