NexuST: A Hierarchical Foundation Model for Spatial Transcriptomics

Spatial transcriptomics offers a way to study the molecular states of individual cells within tissue organization. However, existing foundation models struggle to integrate fine-grained gene data with spatial context on a large scale. Researchers have now introduced NexuST, a hierarchical foundation model designed to tackle this challenge.

The NexuST model operates on a two-level approach, repeatedly combining gene-level molecular modeling with cell-level spatial modeling. These two levels can refine one another during the end-to-end pretraining process. To develop the pretraining data, the team curated a dataset called HumanST-46M, which includes 45.7 million human cells from 72 datasets covering 11 organs and three imaging-based platforms.

When evaluated on four held-out datasets totaling about 2.6 million cells, NexuST demonstrated state-of-the-art or competitive performance in various tasks. These tasks include cell-type annotation, region prediction, gene recovery, and neighborhood-composition prediction. The study also revealed that cell-intrinsic expression data remains valuable even for spatial tasks, as evidenced by an expression-only PCA baseline.

Notably, NexuST showed particularly strong improvements in scenarios where spatial context is crucial. By establishing a hierarchical framework, NexuST sets the stage for future spatial transcriptomics foundation models to build upon.