scMILD: Single-cell multiple instance learning for sample classification and associated subpopulation discovery

Area of Science:

• Computational Biology
• Immunology
• Genomics

Background:

• Linking cellular states to clinical phenotypes is crucial but challenging in single-cell analysis.
• Existing methods often require detailed cell-level labels, limiting their application.

Purpose of the Study:

• To develop a weakly supervised framework for identifying condition-associated cells using only sample-level labels.
• To bridge the gap between single-cell observations and high-level clinical phenotypes.

Main Methods:

• Introduced single-cell multiple instance learning for sample classification and associated subpopulation discovery (scMILD).
• Validated scMILD's accuracy through controlled simulations and application to diverse disease datasets.
• Applied scMILD to analyze monocytes in COVID-19 and stratify Lupus patients.

Main Results:

• scMILD robustly identifies condition-associated cells with sample-level labels.
• Revealed a temporal transition in COVID-19 monocytes from antiviral to stress-response states.
• Successfully stratified Lupus patients and distinguished shared vs. disease-specific inflammatory states.

Conclusions:

• scMILD provides a validated and versatile strategy for dissecting cellular heterogeneity.
• Enables robust linking of cellular states to clinical phenotypes in various diseases.
• Facilitates discovery of shared and disease-specific cellular signatures across conditions.