Foundation models (FMs), which are deep learning models pretrained on large-scale data and applied to diverse downstream tasks, have transformed natural language processing and multimodal AI. However, in spatial transcriptomics (ST), no FM has yet demonstrated the capacity to generate novel, validated biological discoveries. The authors argue that this gap exists because ST data lack an explicit sequence-like structure, are noisy, and are more costly to collect than single-cell RNA sequencing data, making them unsuitable for simply reusing existing single-cell FMs. Therefore, how to leverage ST data to construct better foundation models is a highly promising research direction that warrants further exploration.
The paper distinguishes two types of FMs for ST analysis. Seq-based FMs are pretrained directly on large-scale ST sequencing data using self-supervised learning, with examples including NicheCompass, Nicheformer, STFormer, and CellPLM. Knowledge-based FMs instead leverage existing LLMs or large multimodal models pretrained on biological text or pathology images, such as QuST-LLM and Geneverse, to transfer general knowledge into spatial analysis. The authors also highlight an emerging hybrid approach combining both paradigms, as seen in spEMO and scGPT-spatial. Details are summarized in Figure 1.
The authors argue that FMs should tackle substantive, high-impact problems rather than simple tasks such as basic clustering. Specifically, FMs should help automate and standardize preprocessing pipelines — including quality control, normalization, and annotation — to reduce subjectivity and improve reproducibility across studies. They should also enhance performance on key downstream tasks such as cell-type annotation, spatial niche clustering, gene expression imputation, and spatial deconvolution.
A major opportunity for spatial FMs lies in accelerating biological discovery, reducing the need for costly wet-lab experiments. By analogy with tools like ChemCrow in chemistry, an FM-powered AI agent for ST data could identify novel cell types, predict perturbation effects, and explore spatially induced biological patterns. These areas remain largely unexplored. The paper stresses that the value of an FM must be weighed against the cost of human resources and model training.
In the future, several challenges must be addressed for ST FMs to reach their potential. These include collecting high-quality and diverse training data, designing pretraining objectives appropriate for non-sequential transcriptomic data, and building rigorous benchmarking frameworks that go beyond low-level tasks. Computational costs must also be carefully managed, and the authors advocate for open-sourcing models at multiple scales and providing online demos to make these tools broadly accessible to the research community.

DOI
10.1002/qub2.70010