CpG Transformer for imputation of single-cell methylomes

Area of Science:

• Genomics
• Bioinformatics
• Computational Biology

Background:

• Single-cell DNA methylation sequencing (scBS-seq, scRRBS-seq) is crucial for understanding cellular heterogeneity.
• Current protocols suffer from incomplete data coverage, necessitating robust imputation methods.
• Accurate imputation requires models that capture underlying biological processes in methylation data.

Purpose of the Study:

• To develop an effective imputation technique for single-cell DNA methylation data.
• To adapt advanced neural network architectures for methylation matrix analysis.
• To create a model that addresses the limitations of existing sequencing protocols.

Main Methods:

• Adapted the transformer neural network architecture for methylation matrices.
• Combined axial attention with sliding window self-attention mechanisms.
• Developed the CpG Transformer model for imputation tasks.

Main Results:

• CpG Transformer achieved state-of-the-art performance on scBS-seq and scRRBS-seq datasets.
• Demonstrated the interpretability of the CpG Transformer model.
• Showcased rapid transfer learning capabilities for new datasets with minimal resources.

Conclusions:

• CpG Transformer provides a powerful and interpretable solution for single-cell DNA methylation imputation.
• The model's transfer learning ability reduces computational and time costs for new applications.
• This advancement facilitates more accurate analysis of single-cell epigenomic data.