High-fidelity bidirectional translation between single-cell transcriptomes and DNA methylomes with scBOND

Area of Science:

• Computational Biology
• Genomics
• Epigenetics
• Single-cell Multiomics

Background:

• Single-cell multiomic sequencing provides deep insights into cellular heterogeneity by profiling gene expression and epigenetic landscapes simultaneously.
• Technical challenges and high costs limit the widespread application of these advanced sequencing technologies.
• Existing computational methods for cross-modality translation between scRNA-seq and scDNAm data have limitations, including unidirectionality and inadequate modeling of regulatory associations.

Purpose of the Study:

• To introduce scBOND, a bidirectional cross-modal translation framework for scRNA-seq and scDNAm data.
• To address limitations of existing methods by improving context-specific association modeling and biological relevance in evaluation.
• To enhance model generalization in data-scarce scenarios using a biologically informed augmentation strategy (scBOND-Aug).

Main Methods:

• Developed scBOND, a bidirectional framework utilizing a mixture-of-experts block for context-dependent patterns and self-attention for signal fidelity.
• Implemented scBOND-Aug, a variant incorporating biologically informed data augmentation to improve performance with limited paired data.
• Evaluated scBOND and scBOND-Aug against baseline methods using extensive experimental datasets.

Main Results:

• scBOND consistently outperformed baseline methods in both translation directions, achieving high accuracy while preserving cellular structure.
• The framework successfully identified subtle, functionally significant differences between closely related cell types in mouse embryonic data.
• scBOND reconstructed scDNAm profiles from RNA-only data, revealing cell-type and stage-specific regulatory mechanisms in human neurons and oligodendrocyte lineage.

Conclusions:

• scBOND offers a robust and accurate solution for bidirectional cross-modal translation between scRNA-seq and scDNAm data.
• The framework effectively recovers biological signals and enhances the discovery of regulatory mechanisms, even with limited paired data.
• scBOND and scBOND-Aug significantly advance the computational analysis of single-cell multiomic data, broadening accessibility and insight generation.