A Graph-Based Deep Learning Framework with Gating and Omics-Linked Attention for Multi-Omics Integration and...

Area of Science:

• Computational Biology
• Bioinformatics
• Genomics

Background:

• Multi-omics data integration offers a holistic view of biological systems, crucial for advancing disease classification and biomarker discovery.
• The inherent heterogeneity and high dimensionality of omics data pose significant analytical challenges, hindering effective integration.
• Existing methods often struggle with interpretability and adaptively handling diverse omics data types.

Purpose of the Study:

• To introduce MOGOLA (Multi-Omics integration by Gating and Omics-Linked Attention), a novel deep learning framework for robust and interpretable multi-omics data integration.
• To address the analytical challenges posed by heterogeneous and high-dimensional omics data.
• To improve disease classification and biomarker discovery through advanced computational methods.

Main Methods:

• MOGOLA employs a hybrid graph learning module using Graph Convolutional Networks and Graph Attention Networks for intra-omics feature extraction.
• A gating and confidence mechanism adaptively weighs the importance of features across different omics types.
• A cross-omics attention-based fusion module captures complex inter-omics relationships for integrated analysis.

Main Results:

• MOGOLA consistently outperformed eleven state-of-the-art methods across four benchmark datasets (BRCA, KIPAN, ROSMAP, LGG).
• Ablation studies confirmed the significant contribution of each MOGOLA module to its overall performance.
• Biomarker identification using MOGOLA demonstrated its potential for clinical applications.

Conclusions:

• MOGOLA provides a robust and interpretable framework for multi-omics data integration.
• The proposed deep learning approach advances computational biology and precision medicine.
• MOGOLA facilitates more effective disease classification and biomarker discovery from complex biological data.