CpG Transformer

CpG Transformer performs imputation of missing values in single-cell DNA methylation matrices to enable more complete and accurate single-cell methylome analyses.

Key Features:

• Transformer Neural Network Architecture: Processes methylation matrices using a transformer-based neural network adapted to single-cell methylome data.
• Axial attention with sliding-window self-attention: Combines axial attention and sliding-window self-attention to capture local and global dependencies in methylation matrices.
• Performance on scBS-seq and scRRBS-seq: Demonstrated superior imputation performance across datasets derived from single-cell bisulfite sequencing (scBS-seq) and reduced representation bisulfite sequencing (scRRBS-seq).
• Interpretability: Provides interpretability features that allow examination of the model's predictions and inferred biological patterns.
• Rapid transfer learning: Supports rapid transfer learning to adapt models to new single-cell methylome datasets with minimal additional training.

Scientific Applications:

• Single-cell methylome recovery: Enables recovery and completion of sparse single-cell DNA methylation profiles for downstream analysis.
• Epigenetic regulation studies: Facilitates analysis of DNA methylation patterns to investigate epigenetic regulation.
• Cancer epigenomics and disease progression: Supports analysis of methylation changes relevant to cancer research and disease progression.
• Developmental biology and cellular differentiation: Assists studies of methylation dynamics during development and cellular differentiation.
• Personalized medicine: Provides more complete methylome information that can inform research relevant to personalized medicine.

Methodology:

Processes methylation matrices with a transformer-based neural network that incorporates axial attention combined with sliding-window self-attention to learn from local and global contexts for imputation.