Biologically, protein complexes are the key molecular entities to perform many essential biological functions, such as transcription and replication of DNA, catalyzing biological reactions, signal transduction, cell cycle and so on. Understanding principles of cellular organization and function can be enhanced if we detect known and predict undiscovered protein complexes within the cells. While there are a number of ways to detect protein complexes experimentally, but there are several limitations to this methods. The remarkable thing is that thousands of different proteins exist, which have the ability to interact with other proteins. Thus, the major problem is to identify the groups of millions of protein that is proposed. Due to these limitations, alternative computational approaches for detecting the complexes are thus useful complements to the experimental methods for detecting protein complexes. Over the last decade, high-throughput experimental techniques have allowed us to collect a large amount of protein-protein interaction (PPI) data for many species. A popular ion for representing this data is the PPI network, which makes it possible to predict protein complexes from the network. Such predictions may be used as an inexpensive tool to direct biological experiments. This networks can be represented as undirected graphs, in which nodes represent proteins and edges represent interactions between pairs of proteins. This networks allow us to tackle the problem of complex prediction with the aid of clustering techniques. But because of too much noise which include waste edges and incomplete data in such graphs, available graph clustering algorithms have not achieved to appropriate results. Also unfortunately, many of clustering algorithms have several limitations that are not suitable for using of PPI networks. Such as, some of them are designed only for unweighted graphs and some of them assign proteins to the only one group. While PPI networks are modeled in a weighted graph and many evidences have demonstrated that many proteins belong to more than one main group and protein complexes overlap with each other. In this research we propose a novel four-part method based on removing hubs to reduce and considerable amount of noise in the network. This new algorithm utilizes both given edge weights and can find overlapping clusters. So by using of our proposed method complexes can be more accurately distinguished on different data sets than ClusterONE algorithm. Keywords : Complex detection, PPI networks, Hubs removal, Graph clustering