A graph clustering algorithm with hypergraph learning and a core-attachment strategy for protein complex identification.

Protein complexes play a crucial role in cellular biological processes. Identifying these complexes is essential for understanding cellular functions and biological mechanisms. Graph clustering approaches to identify protein complexes in protein-protein interaction (PPI) networks have become a significant research hotspot in data mining and bioinformatics. Many graph clustering methods have been developed for protein complex identification. However, most existing methods only utilize original networks to discover dense subgraphs and ignore higher-order topological characteristics. Considering the prevalent multi-relational and complex interactions in biological networks, a graph clustering algorithm based on hypergraph learning and a core-attachment strategy is proposed for protein complex identification, called HLCA. Hypergraph networks are employed to directly model multi-relational interactions. Based on this method, a multi-level hypergraph is used as higher-order topology and a core-attachment strategy are adopted to identify protein complexes. Firstly, the original PPI network is transformed into a hypergraph network. Secondly, a hierarchical compression strategy is applied to recursively compress the hypergraph into smaller hypergraphs at various levels, forming a multi-level analytical framework. Thirdly, hypergraph convolution is performed across different hierarchical levels to obtain node representations at each level. These node representations are then combined to produce complete node embeddings. Based on these node embeddings, a weighted PPI network is constructed by cosine similarity from the original PPI network. Core clusters are obtained in this weighted network by cluster density. Finally, remaining protein nodes are added to the core clusters using a core-attachment strategy combining hyperedge density and overlap. The effectiveness of HLCA is evaluated by comparing it with other protein complex identification methods on multiple datasets. Experimental results show that the proposed method outperforms comparison methods regarding F-measure and Accuracy.