Deep learning representations of human Immune Health for precision immunology

Area of Science:

• Immunology
• Computational Biology
• Artificial Intelligence

Background:

• The human immune system comprises diverse cell types and states, crucial for sensing and responding to various health conditions.
• Immune memory and network rewiring store information about exposures, aging, and disease, shaping immune history and future potential.
• Genetic information alone cannot capture the dynamic remodeling of the immune system.

Purpose of the Study:

• To develop a deep learning framework for transforming high-dimensional immune profiles into representations of immune health.
• To create a quantitative method for comparing immune states across individuals and over time.
• To establish a foundation for precision immunology using clinically actionable embeddings.

Main Methods:

• Developed MAESTRO (Masked Encoding Set TRansformer with self-distillation), a self-supervised deep learning framework.
• Trained MAESTRO on 1,792 peripheral blood samples (over 418 million immune cells) across 13 clinical diagnoses.
• Utilized attention-based modeling to capture deep network architecture of immune states from cytometry data.

Main Results:

• MAESTRO generates immune fingerprints that are stable within individuals but diverse across health, disease, and treatment states.
• These fingerprints capture immune architecture beyond cell proportions, enabling efficient clinical prediction.
• Model embeddings retain temporal information, reflecting past exposures and predicting future immune responses.

Conclusions:

• MAESTRO provides a reusable foundation for precision immunology by converting immune complexity into actionable embeddings.
• Demonstrated translational application in metastatic Pancreatic Ductal Adenocarcinoma (PDAC) for patient stratification and predicting immunotherapy response.
• The framework enables diagnosis, monitoring, and therapy selection by analyzing immune landscape circuitry.