MAPS: pathologist-level cell type annotation from tissue images through machine learning.

Area of Science:

• Spatial biology
• Proteomics
• Machine learning

Background:

• Highly multiplexed protein imaging enables analysis of protein distribution in native cellular and tissue contexts.
• Current cell annotation methods for high-plex spatial proteomics data are resource-intensive and require expert input, limiting scalability.
• This hinders the analysis of large-scale spatial proteomics datasets.

Purpose of the Study:

• To introduce MAPS (Machine learning for Analysis of Proteomics in Spatial biology), a novel machine learning approach.
• To enable rapid and precise cell type identification from spatial proteomics data.
• To overcome the scalability and practicality constraints of existing annotation methods.

Main Methods:

• MAPS utilizes machine learning algorithms for automated cell type identification.
• The approach was validated on multiple MIBI and CODEX datasets (in-house and public).
• Performance was compared against existing cell annotation techniques.

Main Results:

• MAPS demonstrates superior speed and accuracy compared to current annotation methods.
• The tool achieves pathologist-level precision, including for challenging cell types like tumor cells of immune origin.
• MAPS provides human-level accuracy in cell type identification.

Conclusions:

• MAPS offers a scalable and rapid machine learning-based solution for spatial proteomics data annotation.
• The tool democratizes cell type identification, accelerating research in tissue biology.
• MAPS has the potential to significantly advance the understanding of disease mechanisms.