A supervised ontology-aware cell annotation method for single-cell transcriptomic data.

Area of Science:

• Computational biology
• Genomics
• Bioinformatics

Background:

• Current single-cell RNA sequencing (scRNA-seq) annotation methods often overlook the inherent hierarchical structure of cell types.
• This limitation can lead to inaccurate or inconsistent cell type classifications.

Purpose of the Study:

• To develop a novel method for scRNA-seq annotation that incorporates ontological consistency.
• To improve the performance of existing annotation models without requiring retraining.
• To introduce a fast, interpretable, and scalable classifier for large scRNA-seq datasets.

Main Methods:

• A probability propagation strategy was developed to enforce ontological consistency in cell type classification.
• This strategy was integrated with a lightweight logistic regression model trained on a large dataset of 42 million human cells, resulting in the SOCAM classifier.
• A novel hop-based F1 score was introduced for ontology-aware evaluation of classification performance.

Main Results:

• The probability propagation strategy improved performance when applied to existing models without retraining.
• SOCAM demonstrated to be a fast and interpretable classifier.
• The hop-based F1 score provided a more accurate evaluation for ontology-aware classification.

Conclusions:

• The proposed probability propagation strategy effectively enforces ontological consistency in cell type classification for scRNA-seq data.
• SOCAM offers a significant advancement in scRNA-seq annotation by providing a scalable, fast, and interpretable solution.
• The developed ontology-aware evaluation metric aids in assessing the biological relevance of cell type classifications.