Skip to main content
FOCIS 2025 Main banner
Final Program


Final Program

Systems Immunology

Immunology-Driven Semi-Automatic Annotation of Single-cell Data Using Machine Learning Approaches

W214 - Immunology-driven Semi-automatic Annotation of Single-cell Data Using Machine Learning Approaches

Wednesday, June 25, 2025
7:30am - 7:30pm East Coast USA Time
Abstract Text: The need for a unified system of cellular nomenclature has driven the development of single-cell atlases. To be truly comprehensive these must encompass millions of cells and incorporate cellular diversity induced by a range of disease perturbations. However, challenges such as non-biological variation and inconsistent sample metadata hinder their construction, requiring optimized data wrangling and integration workflows. As the scale of these datasets grows, cell-type annotation becomes increasingly labor-intensive, with automated tools often struggling to balance accuracy and adaptability. To address these challenges, we assembled a dataset of 25 million cells spanning 40 diseases, integrating data from 20 public studies and in-house efforts, with contributions from over 1000 donors. This dataset represents a wide range of disease states, including autoimmune conditions (e.g., systemic lupus erythematosus, rheumatoid arthritis, IgG4-related disease), infections (e.g., Zika, viral hepatitis, CMV, EBV), and iatrogenic causes (e.g., immune-checkpoint inhibitor toxicities). We benchmarked data harmonization and integration methods to retain biological signals. To support cellular annotation we developed a semi-automated framework employing a decision tree approach with XGBoost, combining gene and protein expression based on finely-annotated training data. This achieved a mean balanced accuracy of 0.85 across 22 cell types, with a run-time of under five minutes. This strategy significantly reduces annotation time while remaining versatile and adaptable to a wide range of contexts. This study offers a structured framework for constructing single-cell reference atlases, paired with an innovative methodology to expedite cell-type identification and a resource for understanding disease-associated cellular diversity.