Spatium: A Protein Language Foundation Model for Spatial Proteomics

Area of Science:

• Biochemistry
• Computational Biology
• Proteomics

Background:

• Spatial proteomics offers single-cell protein insights but faces challenges with limited, overlapping measurement spaces.
• Current analyses often use statistical or task-specific models, leaving scalable representation learning underexplored.

Purpose of the Study:

• To develop a model that learns stable cellular identity representations from constrained spatial protein measurements.
• Introduce Spatium, a protein language foundation model designed for spatial proteomics data.

Main Methods:

• Trained Spatium on over 51 million cells from multiple spatial proteomics platforms.
• Developed a foundation model to learn intrinsic co-expression hierarchies capturing cell identity.
• Evaluated Spatium's performance on downstream tasks including cell identity recovery and microenvironment characterization.

Main Results:

• Spatium learns generalizable cell state representations robust to panel composition and measurement scale.
• The model accurately recovers cell identities and reveals distinct spatial microenvironments via marker enrichment.
• Spatium enables reconstruction of missing protein measurements while preserving biological patterns.

Conclusions:

• Spatium provides stable, interpretable cell representations for spatial proteomics data.
• The model demonstrates robust performance across diverse biological and experimental settings with lightweight adaptation.
• This approach advances the analysis of complex spatial proteomic landscapes.