Morphology-Aware Profiling of Highly Multiplexed Tissue Images using Variational Autoencoders.

Area of Science:

• Biomedical Imaging
• Computational Biology
• Proteomics

Background:

• Spatial proteomics offers deep insights into cellular organization within tissues.
• Current segmentation methods struggle with signal spillover and lack morphological detail.
• Accurate single-cell analysis is crucial for understanding tissue microenvironments and disease.

Purpose of the Study:

• To develop an improved method for single-cell analysis in spatial proteomics.
• To overcome limitations of conventional segmentation, including signal spillover and loss of morphological information.
• To generate more accurate and nuanced cell type and state characterizations.

Main Methods:

• Combined unsupervised, pixel-level machine learning (autoencoders) with traditional segmentation.
• Developed a novel approach for analyzing high-plex spatial proteomics images.
• Generated single-cell data capturing protein abundance, morphology, and local neighborhood context.

Main Results:

• The new method accurately captures protein abundance, cell morphology, and spatial relationships.
• Successfully overcame the issue of signal spillover between adjacent cells.
• Achieved a more nuanced characterization of cell types and states compared to segmentation alone.

Conclusions:

• The developed method enhances the accuracy and detail of single-cell analysis in spatial proteomics.
• This approach mimics the nuanced analysis of human experts while overcoming technical limitations.
• Offers a powerful new tool for histopathology, disease diagnosis, and biological research.