The Virtual Tissues foundation model resolves spatial proteomics across scales

Area of Science:

• Computational biology
• Proteomics
• Cancer research

Background:

• Spatial proteomics advances cancer tissue analysis but faces computational challenges.
• Current methods lack generalizability, hindering biomarker discovery and knowledge transfer.

Purpose of the Study:

• Introduce Virtual Tissues (VirTues), a foundation model for spatial proteomics.
• Enable marker-aware, multi-scale analysis of proteins, cells, niches, and tissues from imaging data.

Main Methods:

• Developed VirTues, a general-purpose foundation model for spatial proteomics.
• Utilized multiplex imaging data for training a single pretrained backbone.
• Enabled zero-shot annotation across diverse datasets and marker panels.

Main Results:

• VirTues supports marker reconstruction, cell segmentation/typing, niche annotation, and biomarker discovery.
• In triple-negative breast cancer, VirTues biomarkers predicted immunotherapy response.
• VirTues biomarkers stratified disease-free survival in an independent cohort.

Conclusions:

• VirTues offers a versatile platform for spatial proteomics analysis.
• The model facilitates robust biomarker discovery and patient stratification in cancer.
• VirTues outperforms existing methods in predicting treatment response and survival.