Scaling cross-tissue single-cell annotation models.

Area of Science:

• Genomics
• Computational Biology
• Machine Learning

Background:

• Accurate cell type identification is crucial for single-cell transcriptomics.
• Existing methods struggle to scale and generalize across diverse biological contexts.

Approach:

• Introduced scTab, a feature-attention-based model for tabular data.
• Trained scTab on 22.2 million human cells using a novel data augmentation scheme.
• Incorporated deep ensembles for uncertainty quantification and accounted for label ontology in evaluation.

Key Points:

• Cross-tissue cell type annotation necessitates nonlinear models.
• scTab's performance scales with training data and model size, outperforming linear models.
• Data augmentation significantly enhances model generalization.

Conclusions:

• scTab offers a scalable and generalizable solution for automated cell type prediction.
• Deep learning approaches provide significant advantages for large-scale single-cell RNA sequencing analysis.
• The scTab codebase and data are publicly available for benchmarking foundation models.