Selecting Near-native Protein Structures From Ab Initio Models Using Ensemble Clustering

Determining the three-dimensional structure of protein molecules is a cornerstone for many aspects of modern biological research.Based on the comparison of protein structures,people can classify the protein into different family and understand its evolution.Besides,it is also critical to reveal the protein function.Nowadays,considerable efforts have been made in determining the protein structure experimentally.But due to the heavy costs and difficulties in specimen preparation,the number of available protein structures still lags far behind the number of available protein sequences.It is urgent to develop the protein structure prediction method.Before this problem is solved fundamentally,it remains the most successful way to predict the protein structure that doing by combining the energy function with statistics.But there is another question: how to select the near-native structure from so many possible decoys?A popular way is clustering.Here,we proposed an ensemble clustering method based on k-medoids to solve this problem.We first run k-medoids many times to generate clustering ensembles,then using a voting method to combine the clustering results.And defined a confidence score based on the cluster quality,considering the cluster size and the internal similarity.The experiments show that,compared to the SPICKER method which is used in I-TASSER server,the proposed method can usually result in better solution in terms of the similarity between the selected near-native structure and the native structure.