Benchmarking graph representation learning algorithms for detecting modules in molecular networks

Background: A common task in molecular network analysis is the detection of community structures or modules. Such modules are frequently associated with shared biological functions and often disrupted disease. Detection structure entails clustering nodes graph, many algorithms apply a algorithm on an input node embedding. Graph representation learning offers powerful framework to learn embeddings perform various downstream tasks such as clustering. Deep embedding methods based graph neural networks can have substantially better performance machine graphs, including module detection; however, existing studies focused social citation networks. It currently unclear if deep offer any advantage over shallow for detecting networks. Methods: Here, we investigated synthetic real cell-type specific gene interaction detect identify pathways affected by sequence nucleotide polymorphisms. We used multiple criteria assess quality clusters connectivity well overrepresentation processes. Results: On networks, variational autoencoder had superior measured modularity metrics, followed closely methods, node2vec Laplacian However, worsens when overall between increases. On did not clear depended upon properties metrics. Conclusions: detection-based be beneficial some but depends metrics properties. Across different types, best performing algorithms.