MGDMCL: A multi-omics integration framework based on masked graph dynamic learning and multi-granularity feature...

Area of Science:

• Computational Biology and Bioinformatics
• Genomics and Multi-omics Data Integration
• Precision Medicine and Disease Etiology

Background:

• Integrating multi-omics data is crucial for understanding complex diseases and advancing precision medicine.
• Graph-based approaches offer robust capabilities for multi-omics data analysis but face limitations in fixed sample similarity graphs and exploring inter-omic feature relationships.
• Existing methods struggle to fully leverage the interconnectedness of features across different omics datasets.

Purpose of the Study:

• To propose MGDMCL (Masked Graph Dynamic Learning and Multi-granularity Feature Contrastive Learning), an innovative framework for integrating multiple omics data.
• To overcome limitations of existing graph-based methods by adaptively adjusting sample similarity graphs and exploring inter-omic feature interactions.
• To enhance the accuracy and robustness of biomedical classification tasks through advanced data integration techniques.

Main Methods:

• Employs masked graph dynamic learning to adaptively adjust sample similarity graphs (SSG) for each omics type, generating multi-layer feature representations via graph convolutional networks (GCN).
• Integrates multi-layer features from different omics at the layer-level and applies multi-granularity feature contrastive learning to derive consensus feature representations.
• Incorporates true class probability to evaluate classification confidence from consensus features across different layers, enhancing classification robustness.

Main Results:

• MGDMCL demonstrates significant performance improvements over state-of-the-art methods in biomedical classification tasks.
• Experiments conducted on five diverse public datasets (LGG, ROSMAP, LUSC, BRCA, KIPAN) validate the effectiveness of the proposed framework.
• The approach successfully addresses limitations in current multi-omics integration methods, particularly in feature representation and inter-omic relationship exploration.

Conclusions:

• MGDMCL offers a more effective approach for integrative multi-omics data analysis, significantly advancing the field.
• The framework shows substantial potential for improving biomedical classification applications, contributing to better disease understanding and diagnosis.
• Open-source implementation is available, facilitating further research and application in the scientific community.