On the Art of Compiling and Using 'Drug‐Like' Chemical Fragment Spaces

Ever since the first rational approaches to the discovery of promising lead candidate structures were applied, it has been a challenge for both medicinal and computational chemists to assess, generate, and combine promising structural motifs to form new and potent chemical entities for biological screening against potential drug targets. Many scientists have committed themselves to the analysis and identification of valuable chemical building blocks and have also developed strategies on how to best recombine them. In this context, the retrosynthetic fragmentation and recombination of chemical motifs derived from known inhibitors is a common and well-known procedure. Meanwhile, fragment-based approaches have become established and valuable processes in pharmaceutical lead discovery and validation. Several application studies have yielded promising lead candidates. Chemical space is huge. Corporate as well as public databases are in the millions and are still increasing in size in order to cover a larger part of the chemical universe. For several good reasons, there is the common trend to standardize experimental and computational protocols in pharmaceutical research. This trend demands systematic and consistent approaches, although they can hardly match the creativity and intuition of medicinal chemists. Consequently, they can and should not substitute, but rather assist, the expert in this task. The most prominent automated example for fragment generation is the retrosynthetic combinatorial analysis procedure (RECAP). It was the first of its kind to apply 11 distinct rules that were supposed to model chemical motifs that could easily be formed by combinatorial chemistry. In this context, the “fragment space” concept was introduced. In contrast to a fragment library, such a space consists not only of a set of fragments, but also of a set of rules that specifies how to recombine fragments by fusing the respective chemical motifs. RECAP is widely used and often referred to, yet even though authors frequently state to have used modified improved versions of the original, actual publications that communicate the extensions that were carried out are rare. An extension of the fragment space concept was recently published, but with a focus on obtaining scaffolds and not on retaining supposedly ’drug-like’ substituents or functional groups. Apart from that, the question remains what a ’drug-like’ fragment space actually is, and whether or not ’drug-likeness’ depends on the origin of the fragments: that is, if they necessarily have to be derived from drugs. In this context, it is highly interesting and important to measure the extent and accuracy with which current models and methods are able to represent the available chemical space. In an attempt to improve existing approaches for the automatic decomposition of molecules into fragments, we compiled a new and more elaborate set of rules for the breaking of retrosynthetically interesting chemical substructures (BRICS) and used this for obtaining fragments from biologically active compounds and vendor catalogue sources. Based on this, we compiled corresponding fragment spaces by specifying a complementary set of rules for the recombination of the corresponding chemical motifs. Furthermore, we put considerable effort into compiling a set of high-quality, high-performance, and, in contrast to all other approaches, publicly available fragments that are meant to serve as a possible basis for various molecular design objectives and techniques. We incorporated more elaborate medicinal chemistry concepts and, for example, modeled explicit isosteric replacements for cyclic and acyclic cases and further distinguished activated from inactivated heterocyclic ring systems and their corresponding substituents. Overall, this work led us to more comprehensive sets of fragments, and the corresponding fragment spaces show a significant increase in performance over existing methods. Moreover, by incorporating fragments from vendor catalogue sources, the performance can be increased even further. The shredding procedure we used for BRICS applies all possible retrosynthetic cuts simultaneously, which avoids the generation of overlapping (redundant) fragments. This is in accordance with RECAP and simplifies the calculation later on. Scheme 1 shows a simple fragmentation example and highlights the key steps. In addition to splitting retrosynthetically relevant bonds, we directly included substructure filters into the shredding procedure to avoid the generation of unwanted chemical motifs as well as small terminal fragments such as single hydrogen and halogen atoms, hydroxy, nitro, carboxylate, methoxy, methyl, ethyl, and isopropyl groups. These motifs are therefore discarded or left uncleaved, respectively. The BRICS model consists of 16 chemical environments indicated by link atoms of different types. The corresponding fragment prototypes are depicted in Scheme 2 and show only the direct chemical environment of the cleavage sites for reasons of simplicity. Therefore, the diversity of the fragments is within the R groups that can also contain further links. Note that the carbonyl and alkyl fragments are shown twice (L1/L6, L4/L8). This is because we wanted to keep track of their origin for medicinal chemistry and modeling reasons, that is, whether they appeared as cyclic or acyclic substituents or linkers. The corresponding fragment space results from the definition of the [a] J. Degen, Dr. A. Zaliani, Prof. Dr. M. Rarey Center for Bioinformatics, University of Hamburg Bundesstrasse 43, 20146 Hamburg (Germany) Fax: (+49)40-42838-7352 E-mail : rarey@zbh.uni-hamburg.de [b] Dr. C. Wegscheid-Gerlach Medicinal Chemistry VII Computational Chemistry Bayer Schering Pharma AG, M:llerstrasse 178, 13342 Berlin (Germany) Supporting information for this article is available on the WWW under http://dx.doi.org/10.1002/cmdc.200800178 and from the authors’ website .