our findings show the best partitioning solution used cluster count less than 100. Our first set of experiments was performed with a number of clusters range from 2 to 1,000. However, we decided to decrease this range for two reasons: (i) the time consuming taken for performing practical virtual screening of large database of ligands in an ensemble with 1,000 representative MD conformations; and (ii) the high level of accuracy achieved by using a representative ensemble with 200 MD conformations. To support the second reason above described, we analyzed and compared all clustering solutions taking into account the level of coverage reached by them in terms of dispersion and MD trajectory representativeness. The dispersions among the partitions generated from 10 to 200 clusters were analyzed by assessing the SQD values (Eq 10). The resulting SQD values by clustering method for Attribute, Cavity RMSD and MK886 PubMed ID:http://www.ncbi.nlm.nih.gov/pubmed/19668191 Protein RMSD data sets are in S1, S2 and S3 Tables, respectively. While Figs 4 and 5 show the SQD values as a function of the cluster PLOS ONE | DOI:10.1371/journal.pone.0133172 July 28, 2015 15 / 25 An Approach for Clustering MD Trajectory Using Cavity-Based Features Fig 4. Comparative performance of partitioning clustering methods for the three data sets under study. Variations in the SQD values as a function of the number of clusters for PubMed ID:http://www.ncbi.nlm.nih.gov/pubmed/19667298 k-means and k-medoids are showed in the graphs (a) and (b), respectively. The black points identify the optimal partitioning solutions. doi:10.1371/journal.pone.0133172.g004 PLOS ONE | DOI:10.1371/journal.pone.0133172 July 28, 2015 16 / 25 An Approach for Clustering MD Trajectory Using Cavity-Based Features Fig 5. Comparative performance of hierarchical agglomerative clustering methods