A multi-omics integration framework using multi-label guided learning and multi-scale fusion.

Area of Science:

• Computational Biology
• Bioinformatics
• Genomics

Background:

• High-throughput sequencing generates vast omics data, necessitating multi-omics integration for complex disease research.
• Current multi-omics integration methods face limitations including manual feature selection, narrow applicability, and insufficient capture of inter-sample and cross-omics interactions.

Purpose of the Study:

• To introduce mmMOI, an end-to-end framework for multi-omics integration designed to overcome existing limitations.
• To enhance model interpretability and reduce bias by directly processing raw omics data without manual feature selection.

Main Methods:

• Developed a multi-label guided multi-view graph neural network for adaptive omics data representation learning across datasets.
• Designed a multi-scale attention fusion network integrating global and local attention for accurate multi-omics data integration.
• Employed an end-to-end framework processing raw high-dimensional omics data.

Main Results:

• mmMOI significantly outperforms state-of-the-art methods in classification tasks.
• Demonstrated high stability and adaptability of mmMOI across diverse biological contexts and sequencing technologies.
• Successfully identified key disease-associated biomarkers, enhancing biological interpretability.

Conclusions:

• mmMOI offers a robust and interpretable solution for multi-omics integration.
• The framework improves classification performance and biomarker discovery for complex diseases.
• mmMOI provides a valuable tool for advancing biological research and clinical applications.