Identifying the best integration strategy for SRT

Determine which integration strategy—physical-space alignment of tissue sections, latent-space integration with spatially informed embeddings, or pseudobulking-based integration—performs best for integrating spatially resolved transcriptomics datasets.

Background

The paper delineates three major strategies for integrating multi-sample SRT: (i) physical-space alignment (e.g., landmark-based registration, thin-plate splines, Gaussian processes, diffeomorphic mapping, optimal transport), (ii) latent-space integration using spatially aware embeddings (e.g., PRECAST, GraphST, BayesSpace), and (iii) pseudobulking approaches that aggregate expression within spatial domains.

Each approach offers distinct trade-offs in scalability, sensitivity to local spatial signals, and dependence on image registration or clustering. A systematic determination of which approach is optimal under what circumstances remains unresolved.