From descriptive to generative: foundation-model approaches for spatial transcriptomics

Spatial transcriptomics has advanced the study of biological systems by facilitating the mapping of gene expression within native tissue environments. Nevertheless, the intrinsic high dimensionality, pronounced data sparsity, and intricate nonlinear relationships between tissue morphology and the transcriptome pose challenges for conventional analytical frameworks in capturing long-range biological dependencies. This manuscript introduces a unified conceptual framework, termed "cells as words, tissues as sentences," for examining how the contextual learning and generative capabilities of large language models (LLMs) and other foundation-model approaches can be applied to characterize the spatial organization of tissues. We synthesize evidence from published studies showing that Transformer-based LLMs, Mamba/state-space models, and diffusion-based models have been applied to tasks including spatial domain identification, cell-cell interaction inference, and cross-modal alignment. This shift from descriptive analysis toward generative modeling supports reconstruction and prediction tasks, while highlighting the need for rigorous validation of perturbational and spatiotemporal inference. Ultimately, this framework provides a systematic perspective for integrating spatial information and evaluating the demonstrated capabilities and limitations of foundation models, while identifying opportunities for generating testable hypotheses about tissue organization and disease progression.