NeuroMDAVIS: Visualization of Single-Cell Multi-Omics Data under Deep Learning Framework.

Area of Science:

• Computational Biology
• Bioinformatics
• Genomics

Background:

• Single-cell technologies generate multi-omics data, offering multiple views of cellular information.
• Analyzing high-dimensional multi-omics data for biological insights is challenging due to data complexity.
• Existing dimensionality reduction and visualization methods struggle to preserve both local and global data structures.

Purpose of the Study:

• To introduce NeuroMDAVIS, a novel unsupervised deep neural network for joint visualization of multi-modal biological datasets.
• To develop a method that integrates different omics layers into a shared latent space for enhanced biological interpretation.
• To provide a robust visualization tool that competes with state-of-the-art methods.

Main Methods:

• Developed NeuroMDAVIS, an unsupervised deep neural network model.
• Implemented joint visualization by transforming and integrating feature spaces of multiple omics modalities.
• Captured modality-specific and common information across omics layers into a shared latent space.

Main Results:

• NeuroMDAVIS effectively learns local and global data relationships, producing meaningful low-dimensional embeddings.
• The model demonstrated strong performance in downstream classification and clustering tasks (accuracy, precision, recall, F1 score).
• NeuroMDAVIS competes effectively against established visualization techniques like t-SNE, UMAP, Fit-SNE, IVIS, and MultiMAP.

Conclusions:

• NeuroMDAVIS is the first model offering joint visualization for multi-modal biological datasets.
• The model provides a robust and efficient approach for analyzing complex multi-omics data.
• NeuroMDAVIS facilitates deeper understanding of biological systems through integrated data visualization.