Improving Jarvis-Patrick algorithm for drug discovery

Clustering algorithms play an important role in chemoinformatics and especially in the drug discovery process. Clustering methods may be hierarchical or non-hierarchical. Non-hierarchical algorithms have fast processing for clustering chemical data sets than hierarchical algorithms. One of the most popular non-hierarchical clustering algorithms that are used in many applications in the drug discovery process is Jarvis-Patrick algorithm. The applications of Jarvis-Patrick in the drug discovery process are compound selection, compound acquisition, low-throughput screening and Quantitative Structure-Activity Relationship (QSAR) analysis. Jarvis-Patrick groups compounds in a cluster based on a three neighborhood conditions. These three conditions groups compounds, which are not similar enough, in the same cluster. Adding dissimilar compounds in the same cluster will lead to poor compound selection, compound acquisition and QSAR analysis. In this paper, standard Jarvis-Patrick is modified by adding a fourth condition which computed only if the three standard conditions are true. This condition computes the increasing in the value of Square Error (SE) of the cluster by adding a compound and compares it with expected increasing in SE to determine whether to add a compound to the cluster or not. The result shows that our modification produces clusters with less SE values in the produces clusters.