Optimized subtractive clustering for cluster-based compound selection

Abstract

Compound selection algorithm has become a need to pharmaceutical industry due to the increasing number of chemical compounds to be screened. One of the widely used methods in compound selection is cluster-based selection where the compound datasets are grouped into clusters and representative compounds are selected from each cluster. This paper proposes the use subtractive clustering in compound clustering by finding the optimal data points to be defined as a cluster centers based on the density of surrounding data points. The technique resolves the problem of determining the suitable number of clusters for the data. Different values of cluster radii and inter-cluster squash factor have been evaluated. For subtractive clustering, good values of squash factor are between 0.375 and 0.45 and the cluster radii from 0.35 to 0.45 because they always give the highest proportion of active structures in active cluster datasets. The results obtained from subtractive clustering has also been used in fuzzy c-mean (FMC) and k-means. We found that the proportion of actives in active cluster subsets are better when fcm and k-means are based on the results produced by subtractive clustering compared to results from subtractive clustering. K-means produced the best results among the three clustering methods.