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

Accurate classification of disease subtypes is a fundamental requirement of precision medicine especially for complex and heterogeneous conditions such as breast cancer and Alzheimer's disease. Recent advances in graph-based deep learning have shown strong potential in multi-omics integration by modeling inter-sample relationships through similarity networks. Yet, the question of how best to construct these networks remains an open and underexplored challenge. In this work, we present a systematic evaluation of six distinct similarity network construction strategies including Cosine Similarity, Cosine Distance, RBF-based measures, and two hybrid combinations leveraging a graph convolutional network (GCN) integrated with a view correlation discovery network (VCDN) framework for multi-omics disease classification. Using two benchmark datasets (BRCA and ROSMAP), we assessed the impact of each method on classification performance, variance across runs, and statistical robustness. Surprisingly, our results demonstrate that Cosine Similarity outperforms all other metrics, consistently achieving the highest accuracy, F1-score, and AUC, while also showing the lowest standard deviation across cross-validation splits. Despite the growing popularity of kernel-based and hybrid similarity designs, our findings highlight the unique effectiveness of simple angular similarity in capturing biologically meaningful structure in high-dimensional omics data. In our study, we showed that simple yet biologically meaningful similarity measures like Cosine Similarity can outperform more complex techniques in accuracy, consistency, and clarity. This insight sets the stage for building more effective and interpretable graph-based models to support precision medicine.