Abstract: Protein complex is the basis of cell structure and biochemical mechanism. How to recognize protein complex accurately has become a popular research direction in recent years. Traditional algorithms has low sensitivity and F-measure in searching protein complexes based on structural information, and the artificial construction features can not reflect the real information of the graph when the existing supervised learning algorithms use machine learning algorithms to identify protein complexes. In order to solve the aforementioned problems, a graph2vec SVM recognition algorithm is proposed. In this algorithm, the protein complex is regarded as a dense subgraph, and the modularity of the subgraph is considered. graph2vec technology is used to transform the graph information into vectors, and SVM classifier is used to recognize the protein complex, which improves the sensitivity of protein complex re- cognition and F-measure. Compared with four popular unsupervised learning algorithms (ClusterONE, CMC,HC-PIN and Coach) and three supervised learning algorithms (SCI-BN, SCI-SVM and RM), the algorithm shows good performance in terms of accuracy, sensitivity and F-measure.

Key words: protein complex, gragh2vec, support vector machine, protein-protein interaction network