Improving atlas-scale single-cell annotation models with hierarchical cross-entropy loss

Area of Science:

• Computational biology
• Genomics
• Bioinformatics

Background:

• Accurate cell type annotation is crucial for single-cell RNA sequencing (scRNA-seq) data analysis.
• Existing computational models often fail to leverage the inherent hierarchical structure of cell type ontologies.
• This limitation hinders the generalizability of predictive models.

Purpose of the Study:

• To introduce a novel loss function that incorporates biological hierarchy into machine learning models for cell type annotation.
• To improve the performance of computational models on out-of-distribution scRNA-seq datasets.
• To guide future research towards data generation strategies that enhance algorithm generalizability.

Main Methods:

• Developed a hierarchical cross-entropy loss function.
• Applied the loss function to various machine learning architectures, including linear models and transformers.
• Evaluated model performance on out-of-distribution datasets to assess generalizability.

Main Results:

• The hierarchical cross-entropy loss improved out-of-distribution performance by 12-15% across different model architectures.
• The modification introduced no additional computational cost.
• Performance gains were achieved without increasing model complexity.

Conclusions:

• Incorporating biological hierarchy into model training objectives is an effective strategy for improving cell type annotation.
• Focusing on generating new data that strengthens connections between annotated cell types is critical for developing more generalizable algorithms.
• This approach offers a more promising avenue for advancing scRNA-seq analysis than solely increasing model complexity.