scTab: Scaling cross-tissue single-cell annotation models.

Area of Science:

• Computational Biology
• Single-Cell Transcriptomics
• Machine Learning in Genomics

Background:

• Accurate cell identification is crucial for single-cell transcriptomics.
• Existing machine learning models struggle to scale and generalize across diverse tissues.
• Automating cell type prediction remains a significant challenge in the field.

Purpose of the Study:

• To introduce scTab, a novel deep learning model for automated cell type prediction.
• To address the limitations of scaling neural networks and cross-tissue generalization.
• To demonstrate the efficacy of a new data augmentation scheme for improved model performance.

Main Methods:

• Developed scTab, a deep learning model tailored for tabular single-cell RNA-seq data.
• Trained scTab on a large corpus of 22.2 million single-cell RNA-seq observations.
• Implemented a novel data augmentation strategy to enhance model generalization across tissues.

Main Results:

• scTab demonstrates effective cross-tissue cell annotation, requiring nonlinear models.
• Model performance scales with both training dataset size and model complexity.
• The proposed data augmentation significantly improves model generalization capabilities.

Conclusions:

• scTab represents a de novo cell type prediction model for single-cell RNA-seq data.
• Deep learning methods, combined with large-scale curated datasets and data augmentation, offer significant benefits for cell type prediction.
• The developed model shows promise for accurate and scalable cell annotation across diverse biological samples.