Clustering of chemical data sets for drug discovery

Chemoinformatics clustering algorithms are important issues for drug discovery process. So, there are many clustering algorithms that are available for analyzing large chemical data sets of medium and high dimensionality. The quality of these algorithms depends on the nature of data sets and the accuracy needed by the application. The applications of clustering algorithms in the drug discovery process are compound selection, virtual library generation, High-Throughput Screening (HTS), Quantitative Structure-Activity Relationship (QSAR) analysis and Absorption, Distribution, Metabolism, Elimination and Toxicity (ADMET) prediction. Based on Structure-Activity Relationship (SAR) model, compounds with similar structure have similar biological activities. So, clustering algorithms must group more similar compounds in one cluster. K-Means, bisecting K-Means and Ward clustering algorithms are the most popular clustering algorithms that have a wide range of applications in chemoinformatics. In this paper, a comparative study between these algorithms is presented. These algorithms are applied over homogeneous and heterogeneous chemical data sets. The results are compared to determine which algorithms are more suitable depending on the nature of data sets, computation time and accuracy of produced clusters. Accuracy is evaluated using standard deviation metric. Experimental results show that K-Means algorithm is preferable for small number of clusters for homogeneous and heterogeneous data sets in terms of time and standard deviation. Bisecting K-Means and Ward algorithms are preferable for large number of clusters for homogeneous and heterogeneous data sets in term of standard deviation, but bisecting K-Means algorithm is preferable in term of time.