Random or rational design? Evaluation of diverse compound subsets from chemical structure databases.

The performance of rational design to maximize the structural diversity of databases for lead finding and lead refinement was investigated. Rational methods such as maximum dissimilarity methods or hierarchical cluster analysis for designing compound subsets were compared to a random approach to study their efficiency for an enhancement of the diversity of three different databases. All investigations were done based on 2D fingerprints as a validated molecular descriptor. To compare the performance of the rational selection methods to a random approach, we additionally used probability calculations. When using maximum dissimilarity-based selections, a single compound can be a member of different neighborhoods as defined by the similarity threshold value, while in hierarchical clustering each compound is assigned to only a single cluster. Therefore the relationship between the similarity threshold of the maximum diversity selection method and a 2D similarity search threshold was studied. In contrast to hierarchical clustering analysis, maximum dissimilarity selections allow to use a similarity threshold for adding a new compound to an already selected compound list. Reasonable values for this similarity threshold are presented here. More diverse subsets were designed using maximum dissimilarity selections, which cover more biological classes than using random selections. An optimally diverse subset without redundant structures containing only 38% of one original dataset was generated, where no structure is more similar than 0.85 to its nearest neighbor, but all biological classes were represented. When it is acceptable to cover only 90% of all biological targets, 3.5-3.7 times more compounds need to be selected using a random approach than in a rational design approach. Such coverage rate shows the highest efficiency of design techniques compared to a random approach. In those subsets no compound is closer than 0.70 to its nearest neighbor. Furthermore a comparative molecular field analysis (CoMFA) is used to evaluate designed and randomly chosen subsets for a database consisting of inhibitors of the angiotensin-converting enzyme. It was shown that designed subsets using maximum dissimilarity methods lead to more stable quantitative structure-activity relationship (QSAR) models with higher predictive power compared to randomly chosen compounds. This predictive power is especially high when there is no compound in the test dataset with a similarity coefficient less than 0.7 to its nearest neighbor in the training set.