Molecular diversity and representativity in chemical databases

It is now common practice in the pharmaceutical industry to use molecular diversity selection methods. With the advent of high throughput screening and combinatorial chemistry, compounds must be rationally selected from databases of hundreds of thousands of compounds to be tested for several biological activities. We explore the differences between diversity and representativity. Validation runs were made for different diversity selection methods (such as the MaxMin function), several representativity techniques (selection of compounds closest to centroids of clusters, Kohonen neural networks, nonlinear scaling of descriptor values), and various types of descriptors (topological and 3D fingerprints) including some validated whole-molecule numerical descriptors that were chosen for their correlation with biological activities. We find that only clustering based on fingerprints or on whole-molecule descriptors gives results consistently superior to random selection in extracting a diverse set of activities from a file with potential drug molecules. The results further indicate that clustering selection from fingerprints is biased toward small molecules, a behavior that might partly explain its success over other types of methods. Using numerical descriptors instead of fingerprints removes this bias without penalising performance too much.