A multi-modal diffusion model with dual-cross-attention for multi-omics data generation and translation.

Area of Science:

• Computational Biology
• Genomics
• Bioinformatics

Background:

• Single-cell multi-omics technologies provide deep insights into cellular functions.
• Integrating diverse omics data is challenging due to experimental limitations.
• Advanced computational methods are crucial for high-fidelity data generation and analysis.

Purpose of the Study:

• Introduce scDiffusion-X, a latent diffusion model for multi-omics data integration, generation, and translation.
• Develop a Dual-Cross-Attention (DCA) module for adaptive and interpretable cross-modal relationships.
• Enable inference of cell-type-specific gene regulatory networks (GRNs) for biological discovery.

Main Methods:

• Developed scDiffusion-X, a latent diffusion model incorporating a Dual-Cross-Attention (DCA) module.
• Employed extensive benchmarking experiments to evaluate model performance.
• Designed a gradient-based interpretation framework for GRN inference.

Main Results:

• scDiffusion-X demonstrated superior performance in generating realistic multi-omics data, preserving cellular heterogeneity and data structure.
• The model achieved excellent scalability for large datasets.
• Accurate modality translation with uncertainty quantification was achieved.
• Inferred comprehensive cell-type-specific heterogeneous GRNs.

Conclusions:

• scDiffusion-X offers a flexible and interpretable approach to multi-omics data integration and generation.
• The model facilitates accurate cross-modal predictions and biological network inference.
• scDiffusion-X accelerates discovery in single-cell multi-omics research by dissecting regulatory relationships and predicting responses.