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

Area of Science:

• Computational Biology
• Genomics
• Pathology

Background:

• Spatial transcriptomics (ST) provides crucial spatial context of gene expression within the tumor microenvironment.
• Current ST methods face scalability limitations and require organ-specific training and dataset fine-tuning.
• Existing approaches infer ST from whole-slide images (WSIs) with restricted gene coverage.

Purpose of the Study:

• To develop a scalable and versatile foundation model for spatial transcriptomics analysis.
• To overcome limitations of existing ST inference methods, including narrow gene coverage and organ specificity.
• To enable direct prediction of gene expression across a wide range of genes and organs without downstream fine-tuning.

Main Methods:

• Introduced STPath, a generative foundation model pretrained on large-scale WSIs and ST profiles.
• Integrated histology images, gene expression, organ type, and sequencing technology into a geometry-aware Transformer.
• Employed masked gene expression prediction with tailored noise schedules for training.

Main Results:

• STPath directly predicts gene expression for 38,984 genes across 17 organs without fine-tuning.
• Demonstrated strong performance on six diverse tasks, including expression prediction, imputation, clustering, and biomarker, mutation, and survival prediction.
• Validated across 23 datasets and 14 biomarkers, showcasing broad applicability.

Conclusions:

• STPath offers a scalable solution for spatial transcriptomics-based pathology.
• The model's pretraining enables high-quality inference and broad applicability across different organs and datasets.
• STPath advances the integration of histology and gene expression data for comprehensive tumor microenvironment analysis.