GREMI: An Explainable Multi-Omics Integration Framework for Enhanced Disease Prediction and Module Identification.

Area of Science:

• Computational Biology
• Bioinformatics
• Systems Biology

Background:

• Multi-omics integration shows promise for complex disease prediction.
• Current methods often prioritize accuracy over biomarker discovery.
• Biomolecular interactions are crucial for understanding disease.

Purpose of the Study:

• To propose GREMI, a two-phase framework for multi-omics classification and explanation.
• To improve disease prediction by incorporating biomolecular interaction information.
• To discover meaningful biomarkers and provide biomedical rationale for disease outcomes.

Main Methods:

• Graph attention architecture on co-functional networks for feature representation.
• Joint-late mixed strategy and true-class-probability block for classification confidence.
• Multi-view approach with Monte Carlo Tree Search (MCTS) for biomarker module identification.

Main Results:

• GREMI outperforms state-of-the-art methods in seven classification tasks.
• The framework effectively handles data mutual interference with increasing omics types.
• Identified modules show functional and disease relevance, validated on an independent cohort.

Conclusions:

• GREMI offers a robust approach for multi-omics classification and biomarker discovery.
• The framework provides enhanced prediction performance and interpretable insights.
• GREMI advances the understanding of complex diseases through integrated omics analysis.