Application and characterization of the multiple instance learning framework in flow cytometry.

Area of Science:

• Immunology
• Computational Biology
• Biomedical Data Science

Background:

• Flow cytometry enables single-cell profiling but relies on labor-intensive, bias-prone manual gating.
• Manual gating involves sequential biomarker selection to isolate cell populations, limiting throughput and reproducibility.

Purpose of the Study:

• To introduce and evaluate multi-instance learning frameworks for automated flow cytometry data analysis.
• To provide an interpretable, data-driven alternative to manual gating for identifying cell phenotypes.

Main Methods:

• Application of multi-instance learning frameworks to flow cytometry datasets.
• Examination of network architecture's impact on predictive performance and rare cell detection.
• Utilization of attention mechanisms for direct identification of phenotype-associated cell subsets.

Main Results:

• Multi-instance learning models demonstrated strong performance across various biomedical applications.
• Successful application in cancer subtyping, HIV stratification, AML MRD prediction, and COVID-19 severity assessment.
• Attention mechanisms enabled interpretable identification of clinically significant cell subsets.

Conclusions:

• Multi-instance learning presents a scalable and interpretable framework for automated flow cytometry analysis.
• This approach has broad potential in precision medicine and translational immunology.
• Automated analysis overcomes limitations of manual gating, improving efficiency and reducing bias.