SpatialFusion: A lightweight multimodal foundation model for pathway-informed spatial niche mapping.

Area of Science:

• Spatial biology
• Computational pathology
• Multimodal deep learning

Background:

• Foundation models advance knowledge transfer across data types but are underdeveloped for spatial biology.
• Existing models often focus on single-cell data and spatial context, lacking integration of transcriptomic and morphological information.
• Delineating functional niches requires a comprehensive approach beyond spatial proximity.

Purpose of the Study:

• To introduce SpatialFusion, a lightweight multimodal foundation model for spatial biology.
• To integrate histopathology, gene expression, and pathway activity for unified representation.
• To identify biologically coherent microenvironments based on pathway activation patterns.

Main Methods:

• Developed SpatialFusion, a multimodal foundation model integrating histopathology, gene expression, and inferred pathway activity.
• Unified these diverse data types into a single representation for analysis.
• Compared SpatialFusion against existing niche-detection methods and spatial foundation models.

Main Results:

• SpatialFusion achieved competitive performance, consistently resolving fine-grained spatial niches with unique pathway signatures.
• Identified a pre-malignant niche in morphologically normal mucosa near colorectal tumors.
• Revealed distinct malignant microenvironments in non-small cell lung cancer predictive of tumor stage.

Conclusions:

• SpatialFusion provides a versatile framework for multimodal spatial analysis in biology.
• Enables the discovery of novel morpho-molecular niches with significant biological and clinical relevance.
• Advances the field of spatial biology by integrating diverse data modalities for deeper insights.