In the previous section, we hypothesized that identifying essential proteins in the dynamic network can optimize the
performance of topology-based methods. To prove the validity of this hypothesis, we take three topology-based methods (DC, SoECC and LAC) as examples. These methods are selected because they can be used to predict essential proteins from different angles. DC bases on the number of neighbors of each current node, SoECC focuses on the importance of the edges connected with current node and LAC considers the significance of neighbors of current node. We compare the performance of DC, SoECC and LAC based on the static PPI network with that applied in the dynamic network. When implementing the three methods in the dynamic PPI network, for each protein, we first calculate the centrality scores related to each of those temporal networks that contain the protein. Then, we obtain the mean of these centrality scores and regard it as the final centrality score of this protein. The final score of a protein that is not included in any temporal network is set to 0. The proteins that are not included in the dynamic PPI network and the interactions related to those proteins are deleted from the static PPI network to achieve a fair comparison. Regardless of which method and which branch of the six types of top ranked proteins is considered, more essential proteins are correctly predicted from the dynamic network than from the original static network and the remaining static network (Fig. 3). Therefore, we can conclude that predicting essential proteins based on the dynamic network topology can achieve better performance. In addition, deleting the proteins excluded from the dynamic network only has minimal effect on the performance of the three methods. That is to say, it is the usage of dynamic PPI network that results in the improvement of prediction precision rather than the deletion of proteins.