A comparison of ligand based virtual screening methods and application to corticotropin releasing factor 1 receptor Gary Tresadern a,*, Daniele Bemporad b, Trevor Howe b a Johnson & Johnson, Pharmaceutical Research & Development, Janssen-Cilag S.A., Calle Jarama, 75, Poligono Industrial, 45007 Toledo, Spain b Johnson & Johnson, Pharmaceutical Research & Development, Janssen Pharmaceutica N.V., Turnhoutseweg 30, 2340 Beerse, Belgium 1. Introduction The corticotropin releasing factor 1 (CRF1) receptor is a target of interest in neuroscience [1]. The endogenous 41-amino acid peptide ligand, CRF (also known as corticotropin releasing hormone), regulates the body���s response to stress through the release of adrenocorticotropic hormone (ACTH). CRF has been shown to mediate stress-induced changes in the autonomic system and to cause neuroendocrine and behavioural effects [2]. Clinical data has shown that patients with depression and post-traumatic stress disorder show significantly elevated concentrations of CRF in cerebrospinal fluid and may have down regulated CRF receptors [3���5]. Two receptor subtypes, CRF1 and CRF2, have been identified and shown to be widely distributed throughout the central nervous system (CNS) and periphery [6]. Compound R121919, Fig. 1, was the first CRF1 antagonist to be evaluated in a phase IIa clinical trial for depression and showed positive effects [7]. A selective CRF1 receptor antagonist would therefore represent a novel class of compounds for the treatment of anxiety, depression, and stress-related diseases. The CRF1 receptor is a secretin-like class B GPCR [8]. A variety of NMR techniques have been used to solve the structure of the isolated extracellular N-terminal domain [9]. The small molecule antagonist binding domain, however, is believed to be located in the 7-transmembrane region [10]. Since the first reported non- peptidic small molecule CRF antagonists in 1991 the number of known compounds has increased significantly and have been reviewed elsewhere [11���13]. There are in excess of one hundred CRF composition of matter patents and although there are many chemically distinct series they often share the same pharmaco- phoric features. The reported CRF antagonists generally show high affinity and selectivity for the CRF1 receptor. Representative CRF1 compounds [14���19] are shown in Fig. 1. Given the expanding prior art it is increasingly difficult to find novel intellectual property space for CRF1 chemistry. For this reason we considered a virtual screening approach to identify new active compounds for this target. High throughput screening (HTS) against a discrete biological target is a preferred technique for finding active starting points in early phase drug discovery. Confirmation and follow-up assays generally provide compounds of sufficient interest for a medicinal chemistry program to begin. Complementary to this, virtual screening can be used as a process for ranking molecules by their potential for activity against the target [20]. Therefore assuming the virtual screening method is of sufficient accuracy the potential benefits are substantial in terms of screening time, compound provision, library depletion and protein supply. As such, virtual screening of compound collections prior to assay or virtual libraries prior to chemical synthesis is widespread [21,22]. Available methods for virtual screening are often classified as structure or ligand based and the choice of which to apply is dependent on available information or desired outcome [23���25]. Journal of Molecular Graphics and Modelling 27 (2009) 860���870 A R T I C L E I N F O Article history: Received 13 December 2008 Received in revised form 12 January 2009 Accepted 14 January 2009 Available online 23 January 2009 Keywords: Corticotropin releasing factor CRF1 CRH1 Scaffold-hopping Ligand based virtual screening A B S T R A C T Ligand based virtual screening approaches were applied to the CRF1 receptor. We compared ECFP6 fingerprints, FTrees, Topomers, Cresset FieldScreen, ROCS OpenEye shape Tanimoto, OpenEye combo- score and OpenEye electrostatics. The 3D methods OpenEye Shape Tanimoto, combo-score and Topomers performed the best at separating actives from inactives in retrospective experiments. By virtue of their higher enrichment the same methods identified more active scaffolds. However, amongst a given number of active compounds the Cresset and OpenEye electrostatic methods contained more scaffolds and returned ranked compounds with greater diversity. A selection of the methods were employed to recommend compounds for screening in a prospective experiment. New CRF1 actives antagonists were found. The new actives contained different underlying chemical architecture to the query molecules, results indicative of successful scaffold-hopping. �� 2009 Elsevier Inc. All rights reserved. * Corresponding author. Tel.: +34 925 24 5782. E-mail address: gtresade@its.jnj.com (G. Tresadern). Contents lists available at ScienceDirect Journal of Molecular Graphics and Modelling journal homepage: www.elsevier.com/locate/JMGM 1093-3263/$ ��� see front matter �� 2009 Elsevier Inc. All rights reserved. doi:10.1016/j.jmgm.2009.01.003

Structure based virtual screening utilizes a 3D representation of the biological target whereas ligand based approaches have no such requirement. Ligand based virtual screening ranks com- pounds by their similarity towards known active ligands. In this work we compare a selection of ligand-based methods. There are thousands of 2D molecular property descriptors which could potentially be used for the calculation of molecular similarity [26,27]. Atom based 2D fingerprints describe the connectivity in molecules but tend to be limited to identifying analogues of close chemical structure, one example is extended connectivity finger- prints (ECFP) from Scitegic [28]. More abstract 2D molecular descriptors such as Feature Trees (FTrees) are based on graph representations of molecules and have less dependence on the underlying 2D structure [29]. Where molecules are represented and compared by their 3D properties the challenge of conforma- tional flexibility and alignment arises. One such method which seeks to overcome the alignment problem is Topomers. Individual 3D fragments of the input molecules are aligned by their valence bonds [30]. The subsequent molecular comparison is performed using fragment steric fields with the inclusion of pharmacophoric features. Methods such as ROCS (Rapid Overlay of Chemical Structures) perform 3D shape based similarity on any number of user-supplied conformations which are often the output of a conformational search [31,32]. The whole molecules are aligned and the Tanimoto difference between their 3D steric fields is calculated. Throughout this work we refer to this method using the acronym OEST (OpenEye Shape Tanimoto). The ROCS colour score includes the option to measure 3D similarity with a feature-based definition of atom-centres and there is also the choice to combine both steric and colour-score, the so-called combo-score denoted as OECS (OpenEye Combo-Score). Comparison of the electrostatic potential of aligned molecules can be performed by means of the program EON which we refer to as OEET (OpenEye Electrostatic Tanimoto) [33]. The Cresset-Fieldscreen approach also performs complete 3D conformational analysis of compounds using a custom forcefield [34]. Subsequent molecular comparison uses four different 3D fields, positive and negative charge, steric shape and hydrophobicity. To render the comparison computationally tract- able for virtual screening of large databases the fields are simplified to their maxima and minima [35,36]. The 3D shape and electrostatic field descriptors encode molecular characteristics likely to be important for biological recognition. However, given their more abstract nature they are less dependent on the underlying atom connectivity. As such they are well suited for scaffold-hopping which ideally should be ignorant of covalent structural frameworks. This is a topic receiving increasing attention [32,37���40]. Given the lack of complete receptor 3D structure and abundance of small molecule ligands the CRF1 target is well suited to a ligand based virtual screening approach. We were particularly interested in finding new chemical series and as such employed 3D field based methods with scaffold-hopping potential. In this work we describe the comparison of a variety of commercially available techniques, FTrees, Topomers, ROCS (OEST and OECS), EON (OEET) and Cresset-Fieldscreen. The scitegic 2D fingerprint method, ECFP6, and simple descriptors such as MW, ALogP and element counts are used for comparison. We performed both retrospective and prospective virtual screening analyses. For the retrospective analysis we searched over compounds previously screened for CRF1 antagonism. We analysed the ability of the different methods to identify known active and reference compounds seeded into CRF1 inactive compounds from an in- house HTS. We also investigated the ability of the methods to successfully scaffold-hop. To achieve this we report the retrieval rate of different active scaffolds from amongst the active set. For the prospective work we used a selection of the same methods to recommend new compounds for screening versus CRF1. We demonstrate that 3D methods such as OEST and OECS had the best enrichment at identifying known active compounds. In addition the same methods retrieved more new active scaffolds amongst the proportion of ranked compounds studied. Methods such as OEET and Cresset appear to offer the most diversity when compared with fingerprints. The latter two methods delivered more new scaffold diversity amongst a fixed number of confirmed actives however, it was at the cost of a lower active retrieval rate. 2. Computational and experimental details 2.1. Data set and query molecules The dataset used for the retrospective analysis consisted of 1261 active and 175,196 inactive HTS molecules. The active set was a combination of 899 reported reference compounds extracted from CRF1 patents and 362 in-house CRF1 compounds with antagonistic pIC50 activity 6. The structures for the 899 reference actives are provided in the supplementary information accom- panying this work. The confirmed inactive compounds were taken from the output of a previous in-house CRF1 antagonist HTS. For the prospective part of this work we searched over available compounds from the Johnson & Johnson corporate collection. The compounds in the datasets were considered in their neutral form. The datasets were converted into the required searchable database format for each software package. The case-by-case details are given in the following sections. To shed further light on this proprietary dataset we compared histograms of simple descriptors ALogP, MW, surface area and volume, Table 1. The plots help to understand the potential elementary differences between actives and inactives. All plots Fig. 1. Representative CRF1 compounds CP-154526 [14], antalarmin [15], R121919 [16], R278995 [17], CP-316311 [18] and CP-376395 [19]. G. Tresadern et al. / Journal of Molecular Graphics and Modelling 27 (2009) 860���870 861