De Novo Fragment Design: A Medicinal Chemistry Approach to Fragment-Based Lead Generation Francisco X. Talamas,* Gloria Ao-Ieong, Ken A. Brameld, Elbert Chin, Javier de Vicente, James P. Dunn, Manjiri Ghate, Anthony M. Giannetti, Seth F. Harris, Sharada S. Labadie, Vincent Leveque, Jim Li, Alfred S-T. Lui, Kristen L. McCaleb, Isabel Na ́ jera, Ryan C. Schoenfeld, Beihan Wang, and April Wong Hoffmann-La Roche Inc., Pharma Research & Early Development, 340 Kingsland Street, Nutley, New Jersey 07110, United States * S Supporting Information ABSTRACT: The use of fragments with low binding affinity for their targets as starting points has received much attention recently. Screening of fragment libraries has been the most common method to find attractive starting points. Herein, we describe a unique, alternative approach to generating fragment leads. A binding model was developed and a set of guidelines were then selected to use this model to design fragments, enabling our discovery of a novel fragment with high LE. ■ INTRODUCTION In the last two decades, fragment screening has been embraced by the pharmaceutical and biotechnology industry, becoming an integral part of drug discovery. Much of this can be attributed to the early success of several companies that developed fragments into clinical candidates 1,2 or a marketed drug. 3 Fragments as starting points for lead identification efforts are typically found by using two complementary approaches: (1) screening a library of fragments or (2) by de novo fragment design using computer-assisted methods. The use of fragment libraries is the most common approach. A summary of the process and successes of using libraries for fragment-based drug discovery has been documented by Erlanson. 4 Two decades ago, computer-assisted de novo design methods such as LEGEND 5 attracted the attention of researchers to use computational tools to create novel structures that could become drugs. Since then, a large number of new programs have been described in the literature that addresses the approach of de novo design. Two excellent reviews by Loving 6 and Rognan 7 describe the methods and approaches used by the different programs. Even with the success obtained by using software to do de novo fragment design, some drawbacks have been identified and are difficult to overcome. 7 Herein, we describe our approach to de novo fragment design targeting the hepatitis C virus (HCV) nonstructural 5B (NS5B) polymerase that plays an important role in the life cycle of the virus by replicating the viral genome. 8 We approached de novo fragment design by implementing some of the same principles utilized by computer programs. Key receptor interactions were identified, and design constraints were defined. The generation of the novel compounds was done differently. The fragments were created by medicinal chemists that applied their own experience to design novel molecules that addressed all the recognized disadvantages. ■ RESULTS AND DISCUSSION A library of ∼2700 fragments was screened against the HCV NS5B polymerase using surface plasmon resonance (SPR) and confirmed 163 hits. Using different parameters, 29 fragments were selected for crystallography but only fragment 1 cocrystallized with NS5B (NS5B BK K D = 78 μM and IC 50 = 130 μM). Attempts to directly improve potency, and some of the physicochemical and ADME properties of the scaffold provided by 1 were unsuccessful. At the same time that the screening activities were ongoing, a de novo fragment design approach was also taken to come up with leads that could then be moved into optimization as an alternative to library screening. Of the four allosteric sites in NS5B known, 9 the palm I site was selected for targeting using this approach. Our first step was to build a model containing key interactions in the palm I allosteric site. We used the crystallographic data available from our internal efforts (2 and 3 10,11 ) and the relevant structures in the public domain (4- 6 12-14 ) available at the time (early 2007). As mentioned above, we were not able to use 1 as an explicit starting point for our medicinal chemistry efforts, however, 1 played an important role in helping us develop our model. Careful analysis of the crystallographic data revealed that the pyrazolopyrimidine N-2 makes a hydrogen bond accepting interaction with the backbone NH of Tyr448, and the NH hydrogen on the N-1 was within hydrogen bond distance (2.4 Å) to the backbone carbonyl of Gln446 (Figure 1). This interaction with Gln446 was unexpected and to our knowledge the first time that a small molecule binding in the palm I site exhibited it. Once we had selected the relevant molecules (1-6) from which to develop the model, it was decided to focus mainly on the hydrophobic area of the allosteric site (for additional descriptions of all the interactions in this site see ref 10). First, Received: February 19, 2013 Brief Article pubs.acs.org/jmc © XXXX American Chemical Society A dx.doi.org/10.1021/jm4002605 | J. Med. Chem. XXXX, XXX, XXX-XXX