A self-supervised embedding of cell migration features for behavior discovery over cell populations

Area of Science:

• Immunology
• Computational Biology
• Systems Biology

Background:

• Cell population dynamics are crucial for understanding immunological processes like inflammation and disease.
• Spatial omics and single-cell analysis are key tools in modern immunology research.
• Hierarchical organization of immunological scenarios reveals distinct cell behaviors across different groups.

Purpose of the Study:

• To develop a self-supervised learning approach for discovering cell dynamics and behaviors in immunology.
• To analyze neutrophil migration in an infarct inflammation model.
• To model hierarchical structures and temporal consistency in cell behavior analysis.

Main Methods:

• Utilized a recurrent neural network with hand-crafted spatio-temporal features.
• Employed a novel multi-task contrastive loss for training.
• Focused on modeling hierarchical organization (groups-behaviors-samples) and temporal consistency.

Main Results:

• Generated embeddings that improve cell behavior separability and therapy log-likelihood.
• Outperformed traditional feature extraction and state-of-the-art methods.
• Achieved better results with reduced dimensionality (16 features vs. 21).

Conclusions:

• The method enables population-level single-cell analysis by automatically discovering shared cell behaviors.
• It facilitates prediction of therapy effectiveness based on the proportions of discovered behaviors.
• This approach advances understanding of cell dynamics in complex biological systems.