Generalized cell phenotyping for spatial proteomics with language-informed vision models

Area of Science:

• Proteomics
• Bioinformatics
• Computational Biology

Background:

• Spatial proteomics enables cell-level analysis but struggles with dataset generalization.
• Existing methods face challenges with varying marker panels and data heterogeneity.

Purpose of the Study:

• To develop a novel, generalizable cell phenotyping approach for spatial proteomics.
• To create a language-informed vision model that adapts to diverse datasets and marker panels.

Main Methods:

• Utilized a transformer with channel-wise attention for a language-informed vision model.
• Trained the model on a diverse dataset with cell type labels from literature and the NIH Human BioMolecular Atlas Program (HuBMAP).

Main Results:

• Demonstrated robust performance across various cell types, tissues, and imaging modalities.
• Achieved superior accuracy and generalizability compared to existing methods through comprehensive benchmarking.

Conclusions:

• The novel approach significantly advances automated spatial proteomics analysis.
• Offers a generalizable and scalable solution for cell phenotyping in multiplexed imaging data.