Scaffold‐Hopping Potential of Ligand‐Based Similarity Concepts

ion from chemical structure. Consequently, such methods have been employed for the design of screening libraries, relying on their scaffold-hopping potential. 3, 8–12] In this study we compared the scaffold-hopping efficiency of topological, three-dimensional and molecular-surface-based pharmacophore pair descriptors with a popular substructure fingerprint method. Two molecules are considered to have different scaffolds if they have different topologies. This idea is based on the concept that druglike molecules are built up from a scaffold and side chains. There are several reasons for seeking a set of diverse scaffolds. Different chemotypes offer a choice in terms of chemical accessibility and prospects for lead optimization. Multiple lead structures (“backup” compounds) lower the chance of attrition in drug development through undesirable ADMET (absorption, distribution, metabolism, excretion, and toxicity) properties. Scaffold hopping can also be applied to move from natural substrates to more druglike chemotypes. 5, 14] Furthermore, the creation of intellectual property is facilitated when multiple novel bioactive agents are available. Different virtual-screening concepts have been proposed for scaffold hopping. These include three-dimensional pharmacophore models, 15] pseudoreceptors, protein-structure-based de novo design, 17] and ligand-based similarity searching. Typically, rapid similarity searching is based on the comparison of descriptor vectors rather than on the explicit alignment of molecules to a reference and can thus be efficiently applied to screening large datasets. Herein, we concentrated on such methods. Similarity searching is founded on the similarity principle, which states that similar molecules exhibit similar biological effects. A straightforward similarity-searching approach is to compare the connection tables to assess the similarity between two molecules. Such methods include substructure fingerprints like the MACCS keys, which are based on exact chemical substructures. Substructure matching approaches were reported to be among the most successful for virtual screening. 7] The classification of intermolecular interactions into general pharmacophore types provides a way to obtain a more general description of the underlying chemotypes of molecules. 9] Three such descriptors were employed in the work reported herein: the topological CATS descriptor, 21] the three-dimensional CATS3D descriptor, and the molecular-surface-based SURFCATS descriptor (Figure 1). Molecular representations that are based on three-dimensional conformations like molecular surface-based descriptors are independent from the molecular connectivity and should have a favorable scaffold-hopping potential. 23] The three CATS descriptors describe a molecule in the form of a histogram that contains the normalized frequencies of all pairs of potential pharmacophore points (PPP) in a molecule. In our study, PPP pairs were further subdivided into PPP–PPP distan[a] S. Renner, Prof. Dr. G. Schneider Beilstein Endowed Chair for Cheminformatics Institute of Organic Chemistry & Chemical Biology Johann Wolfgang Goethe University Siesmayerstraße 70, 60323 Frankfurt (Germany) URL: www.modlab.de Fax: (+49)69-798-29826 E-mail : g.schneider@chemie.uni-frankfurt.de Supporting information for this article is available on the WWW under http://www.chemmedchem.org or from the author. Figure 1. The CATS family of descriptors: CATS, CATS3D, and SURFCATS. All descriptors are based on a PPP (potential pharmacophore point)-type description of the underlying molecule. For each descriptor, pairs of PPPs are transformed into a correlation vector. CATS is calculated from the topological distances of atom-based PPP pairs. For CATS3D the spatial distances between atom-based PPPs are used instead. SURFACTS uses the spatial distances between PPPs on the contact surface of a molecule. Here the PPPs represent the atom types of the nearest atom to each surface point. ChemMedChem 2006, 1, 181 – 185 A 2006 Wiley-VCH Verlag GmbH & Co. KGaA, Weinheim 181 ces and different pharmacophore types. For CATS, pairs of PPPs with shortest topological distances of up to ten bonds were counted, matching at least one of the pharmacophore types: anion, cation, hydrogen-bond donor, hydrogen-bond acceptor, or hydrophobic. For CATS3D and SURFCATS, pairs of PPPs were considered to fall into one of 20 equal-distance bins from 0–20 E. For the latter two methods, one of the pharmacophore types anion, cation, hydrogen-bond donor, hydrogenbond acceptor, polar (hydrogen-bond acceptor and hydrogenbond donor), or hydrophobic were assigned with the ph4_ aType function in the software suite MOE. For SURFCATS, surface points were calculated with the Gauss–Connolly function in MOE with a spacing of 2 E which were subsequently assigned to the pharmacophore type of the nearest atom. Finally, each bin of the three descriptors was scaled by the added occurrence of the respective PPPs. For comparison with a conceptually different descriptor, the MACCS keys were used as implemented in MOE. To assess the degree of scaffold hopping, one must define the term “scaffold”. Herein, we followed the concept of Xu and Johnson employing the software suite Meqi, which has recently been devised for the analysis of chemical libraries. 27] Two different definitions of a scaffold were applied: cyclic system (“Scaffold”, Sc) and reduced cyclic system (“Reduced Scaffold”, ReSc) (Figure 2). In Meqi, each molecular topology is specified by a particular molecular equivalence index (meqi) which is used to distinguish between different scaffolds and reduced scaffolds. Ligands from ten different target classes from the COBRA database of annotated ligands (version 2.1, 4705 molecules) were used as reference for retrospective virtual screening: angiotensin-converting enzyme (ACE, 44 compounds, 28 scaffolds, 17 reduced scaffolds), cyclooxygenase 2 (COX2, 94, 27, 14), corticotropin-releasing factor (CRF antagonists, 63, 33, 23), dipeptidyl peptidase IV (DPP, 25, 13, 7), human immunodeficiency virus protease (HIVP, 58, 46, 31), matrix metalloproteinase (MMP, 77, 47, 19), neurokinin receptors (NK, 118, 65, 49), peroxisome proliferator-activated receptor (PPAR, 35, 29, 17), b-amyloid-converting enzyme (BACE, 44, 13, 12), and thrombin (THR, 188, 100, 36). According to the number of scaffolds and reduced scaffolds in relation to the number of molecules, the datasets range from sets with a low scaffold diversity (for example, COX2) to sets with a large relative scaffold diversity (such as PPAR and HIVP). The complete COBRA database contained 1628 different scaffolds and 637 distinct reduced scaffolds. For retrospective screening, each molecule from each target class was taken iteratively as the reference molecule for a virtual screening experiment, in which all other molecules were ranked according to their similarity to the reference molecule. For quantification of “similarity” three similarity indices were employed: Manhattan distance [Eq. (1)] , Euclidean distance [Eq. (2)] , and Tanimoto similarity [Eq. (3)]: