Effects of similarity networks in graph-based multi-omics classification

Area of Science:

• Bioinformatics
• Computational Biology
• Genomics

Background:

• Precision medicine requires accurate disease subtype classification, especially for complex conditions like breast cancer and Alzheimer's disease.
• Graph-based deep learning shows promise for multi-omics integration by modeling inter-sample relationships via similarity networks.
• Optimal construction strategies for these similarity networks remain underexplored.

Purpose of the Study:

• To systematically evaluate six distinct similarity network construction strategies for multi-omics disease classification.
• To assess the impact of different similarity metrics on classification performance, variance, and statistical robustness.
• To identify the most effective similarity measure for graph-based deep learning in precision medicine.

Main Methods:

• A graph convolutional network (GCN) integrated with a view correlation discovery network (VCDN) framework was utilized.
• Six similarity network construction strategies were evaluated: Cosine Similarity, Cosine Distance, RBF-based measures, and two hybrid combinations.
• Performance was assessed using two benchmark datasets (BRCA and ROSMAP) with cross-validation.

Main Results:

• Cosine Similarity consistently outperformed all other evaluated metrics in classification accuracy, F1-score, and AUC.
• Cosine Similarity demonstrated the lowest standard deviation across cross-validation splits, indicating high robustness.
• Simple angular similarity (Cosine Similarity) proved effective in capturing biologically meaningful structure in high-dimensional omics data.

Conclusions:

• Simple similarity measures, such as Cosine Similarity, can outperform more complex techniques in multi-omics data analysis for disease classification.
• Cosine Similarity offers a robust and effective approach for constructing similarity networks in graph-based deep learning models.
• These findings support the development of more effective and interpretable graph-based models for advancing precision medicine.