MultiGEOmics: Graph-Based Integration of Multi-Omics via Biological Information Flows.

Area of Science:

• Computational Biology
• Bioinformatics
• Machine Learning in Medicine

Background:

• Multi-omics datasets offer comprehensive biological insights but are challenging to integrate.
• Current graph-based methods often ignore crucial cross-omics regulatory signals and struggle with missing data.
• Existing approaches often fail to model interdependencies between different omics layers effectively.

Purpose of the Study:

• To introduce MultiGEOmics, a novel graph integration framework for multi-omics data.
• To explicitly model cross-omics regulatory signals and dependencies for improved biological insights.
• To develop a method robust to missing omics data for reliable machine learning applications.

Main Methods:

• Developed MultiGEOmics, an intermediate-level graph integration framework.
• Incorporated explicit cross-omics regulatory signals into graph representation learning.
• Modeled both omics-specific and cross-omics dependencies using biologically inspired approaches.

Main Results:

• MultiGEOmics learns robust cross-omics embeddings, performing well even with partially missing data.
• Evaluated on eleven datasets across cancer and Alzheimer's disease, showing consistent strong predictive performance under various missing-data scenarios.
• Demonstrated interpretability by identifying key omics types and features driving predictions.

Conclusions:

• MultiGEOmics effectively integrates multi-omics data, overcoming limitations of existing methods.
• The framework provides reliable and interpretable predictions, even with incomplete datasets.
• Enables advanced machine learning applications in biology and medicine by leveraging integrated omics information.