A biology-driven deep generative model for cell-type annotation in cytometry.

Area of Science:

• Biotechnology
• Computational Biology
• Immunology

Background:

• Cytometry is crucial for single-cell phenotyping in complex biological samples.
• Manual gating, the traditional method for cell annotation, suffers from poor reproducibility and batch effect sensitivity.
• High-dimensional data from advanced cytometers (spectral, mass) challenge manual analysis.

Purpose of the Study:

• To introduce Scyan, an automated cell type annotation network for cytometry data.
• To address the limitations of manual gating in terms of reproducibility, speed, and high-dimensional data handling.
• To provide an interpretable and efficient deep learning solution for cytometry analysis.

Main Methods:

• Development of Scyan, a Single-cell Cytometry Annotation Network utilizing deep generative models (normalizing flows).
• Mapping of high-dimensional protein expression data into a biologically relevant latent space.
• Leveraging prior expert knowledge of cytometry panels for automated annotation.

Main Results:

• Scyan significantly outperforms existing state-of-the-art models on multiple public cytometry datasets.
• The model demonstrates enhanced speed and interpretability compared to traditional methods.
• Scyan effectively addresses complementary tasks including batch-effect correction, debarcoding, and cell population discovery.

Conclusions:

• Scyan accelerates and simplifies cell population characterization, quantification, and discovery in cytometry.
• The network offers a robust and efficient alternative to manual gating for complex, high-dimensional cytometry data.
• Automated annotation with Scyan enhances reproducibility and reduces batch effects in single-cell analysis.