STPath: a generative foundation model for integrating spatial transcriptomics and whole-slide images

Spatial transcriptomics (ST) offers insights into gene expression patterns and their spatial context within the tumor microenvironment, but remains limited by the scalability of current sequencing technologies. Existing approaches infer ST from whole-slide images (WSIs) using pretrained encoders, yet are restricted by narrow gene coverage, organ-specific training, and dataset-specific fine-tuning. In light of this, we present STPath, a generative foundation model pretrained on large-scale WSIs paired with ST profiles. This extensive pretraining enables STPath to directly predict gene expression across 38,984 genes and 17 organs without downstream fine-tuning. STPath integrates histology images, gene expression, organ type, and sequencing technology modality within a geometry-aware Transformer, trained via masked gene expression prediction with tailored noise schedules to capture gene-gene dependencies and enable high-quality inference. Evaluated on six tasks spanning 23 datasets and 14 biomarkers, including expression prediction, spot imputation, spatial clustering, biomarker prediction, mutation prediction, and survival prediction, STPath demonstrates strong applicability for scalable ST-based pathology applications.