Cytometry masked autoencoder: An accurate and interpretable automated immunophenotyper.

Area of Science:

• Immunology
• Computational Biology
• Bioinformatics

Background:

• Single-cell cytometry is vital for immune system research.
• Current cell annotation methods struggle with scalability, robustness, and accuracy.

Purpose of the Study:

• To automate immunophenotyping and cell type annotation using a novel approach.
• To improve the interpretability and cross-study comparability of cytometry data analysis.

Main Methods:

• Developed a cytometry masked autoencoder (cyMAE) model.
• Employed a two-phase training: self-supervised learning on unlabeled data, followed by fine-tuning on annotated data.
• Validated the model across multiple studies using the same cytometry panel.

Main Results:

• cyMAE accurately and interpretably annotates cell types in cytometry data.
• The model improves the prediction of subject-level metadata.
• Training cost is amortized over repeated inferences on new datasets.

Conclusions:

• cyMAE offers a scalable and robust solution for automated immunophenotyping.
• This approach represents a significant advancement for large-scale immunology research.
• The model's adherence to user-defined cell types enhances interpretability.