scEMB: Learning context representation of genes based on large-scale single-cell transcriptomics.

Area of Science:

• Computational Biology
• Genomics
• Bioinformatics

Background:

• Single-cell transcriptomics generates vast cellular data, revealing heterogeneity but posing computational challenges.
• Analyzing complex gene-gene relationships requires advanced computational tools.
• Existing methods struggle with integrating diverse single-cell data and predicting gene functions.

Purpose of the Study:

• To develop a novel deep learning model, scEMB, for analyzing large-scale single-cell transcriptomics data.
• To capture context-aware gene embeddings and improve downstream analysis tasks.
• To enable accurate prediction of gene perturbation effects and facilitate therapeutic target discovery.

Main Methods:

• Developed scEMB, a transformer-based deep learning model.
• Trained scEMB on over 30 million single-cell transcriptomes using a novel binning strategy.
• Integrated data across multiple platforms to preserve gene expression hierarchies and cell-type specificity.

Main Results:

• scEMB demonstrated superior performance in batch integration, clustering, and cell type annotation compared to scGPT and Geneformer.
• The model accurately predicted in silico gene perturbation effects in CRISPR-edited datasets and microglia state transitions.
• scEMB identified potential Alzheimer's disease risk genes and showed robust fine-tuning capabilities for domain-specific applications.

Conclusions:

• scEMB is a powerful tool for extracting biological insights from complex gene expression data.
• The model's ability to simulate in silico perturbation effects accelerates therapeutic development.
• scEMB has the potential to advance precision medicine through enhanced gene correlation analysis and target discovery.