How to Simplify the Integration of Human Cell Types

Harmonizing cell types across datasets is a critical step in building a standardized and unified Human Cell Atlas (HCA). A team led by Sarah A. Teichmann at the Wellcome Sanger Institute has introduced CellHint, a powerful tree-based predictive clustering tool designed to address differences in annotation resolution and technical biases in single-cell datasets.

CellHint quantifies transcriptome similarities between cells with high accuracy, organizing cell types into a hierarchical relationship graph. This approach defines shared and unique subtypes across datasets, providing a powerful framework for harmonizing cell annotations. When applied to multiple immune cell datasets, CellHint successfully replicated expert-curated annotations, demonstrating its accuracy and reliability. In addition, the tool revealed previously underexplored relationships between healthy and diseased lung cell states in eight diseases. This insight highlights its utility in identifying subtle cellular changes associated with disease. The team also presents a rapid cross-dataset integration workflow guided by the harmonized cell types and hierarchical structure. The workflow identified underappreciated cell types in the adult hippocampus, demonstrating its potential to reveal new biological insights. To further validate its versatility, the team applied CellHint to 12 tissues, covering 38 datasets, and collated a comprehensive cross-tissue database of approximately 3.7 million cells. This database, combined with a machine learning model developed for automatic cell annotation, provides an important resource for researchers studying human tissues.

This study, published in the journal Cell, provides a key tool for the single-cell community, facilitating the coordination and integration of datasets to build a standardized and deeply annotated human cell atlas. By improving cross-dataset compatibility and revealing new cell type relationships, CellHint is shaping the future of cell biology and biomedical research.

Reference

[1] Chuan Xu et al., Cell 2023 (DOI: 10.1016/j.cell.2023.11.026)