SAGE-FM: A lightweight and interpretable spatial transcriptomics foundation model.

Area of Science:

• Computational biology
• Genomics
• Bioinformatics

Background:

• Spatial transcriptomics provides gene expression data with spatial context.
• Developing computational models to interpret these spatial relationships is crucial.
• Existing methods may not fully capture the nuances of spatial gene regulation.

Purpose of the Study:

• Introduce SAGE-FM, a lightweight foundation model for spatial transcriptomics.
• Leverage graph convolutional networks (GCNs) for spatially aware gene expression analysis.
• Demonstrate the model's ability to learn coherent embeddings and capture regulatory relationships.

Main Methods:

• Developed SAGE-FM using graph convolutional networks (GCNs).
• Trained the model on 416 human Visium samples across 15 organs.
• Employed a masked central spot prediction objective for training.

Main Results:

• SAGE-FM learned spatially coherent embeddings, recovering 91% of masked genes with significant correlations.
• Embeddings outperformed existing methods in unsupervised clustering and preserving biological heterogeneity.
• Achieved 81% accuracy in oropharyngeal squamous cell carcinoma spot annotation and improved glioblastoma subtype prediction.

Conclusions:

• Simple, parameter-efficient GCNs can function as effective foundation models for spatial transcriptomics.
• SAGE-FM provides biologically interpretable and spatially aware insights.
• The model generalizes well to downstream tasks, enhancing spatial gene expression analysis.