QuantCell: machine learning based cell annotation integrating qualitative and quantitative imaging profiles.

Area of Science:

• Spatial omics
• Bioinformatics
• Computational biology

Background:

• Spatial omics technologies provide high-resolution imaging of gene and protein expression.
• Accurate cell annotation is crucial for analyzing complex tissue data but remains challenging.
• Limited markers, overlapping expression, and rare cell types hinder traditional annotation methods.

Purpose of the Study:

• To present QuantCell, a machine learning (ML) framework for improving cell annotation in spatial omics data.
• To leverage quantitative imaging data for more reliable cell identification.
• To offer a robust and automated solution for spatial omics analysis.

Main Methods:

• QuantCell employs multiple ML models to identify the optimal approach for cell annotation.
• A user-defined false discovery rate (FDR) is applied to ensure annotation reliability.
• The framework was validated using PhenoCycler imaging data from mouse bone marrow.

Main Results:

• QuantCell increased cell annotation from 33.1% to 90.2% at a 5% FDR.
• Achieved 96.5% accuracy compared to conventional annotation methods.
• Outperformed existing methods in predictive accuracy and minimized bias towards abundant cell types, enabling rare cell detection.

Conclusions:

• QuantCell offers an effective ML-based solution for spatial omics cell annotation.
• The framework enhances annotation accuracy, reliability, and efficiency, particularly for rare cell populations.
• QuantCell is compatible with various imaging platforms and adaptable to specific research needs, reducing reliance on manual annotation.