Structural and Potency Relationships between Scaffolds of Compounds Active against Human Targets
Ye Hu, Jürgen Bajorath · ChemMedChem · 2010
The analysis of molecular scaffolds as core structures for druglike compounds has traditionally played an important role in medicinal chemistry.1, 2 Scaffolds have been defined, for example, on the basis of synthetic and retrosynthetic criteria,3 by focusing on ring systems,4 or by applying a molecular hierarchy of R-groups, ring structures, and linkers between rings,5 which represents the currently most widely applied definition. A variety of statistical analyses have been carried out to characterize scaffold distributions in drugs and pharmaceutically relevant active compounds5–10 or screening libraries.11 For example, we have previously analyzed the distribution of scaffolds in compounds at different pharmaceutical development stages.10 In this study, we identified sets of scaffolds that preferentially occur in bioactive molecules, clinical trials compounds, or drugs. However, structural relationships between scaffolds found in these compounds were not explored. Of particular interest in scaffold analysis is the ability to identify different molecular scaffolds that have the same specific activity, a task often referred to as scaffold hopping.12–15 The exploration of scaffold hopping ability is a major focal point in both medicinal chemistry12, 13 and computational design.14, 15 For computational compound screening methods, the assessment of scaffold hopping potential has become one of the most important criteria.15 The search for different scaffolds sharing the same activity, through chemical and/or computational means, is based on assumed scaffold diversity among specifically active compounds. However, the degree of scaffold diversity among currently available active compounds has not yet been investigated in a systematic manner. Therefore, we asked the question as to what currently available compound data might tell us about scaffold diversity. To these ends, a large-scale analysis of scaffolds in compounds that are active against currently available human drug targets was carried out. Structural relationships between scaffolds were systematically explored at different levels and related to compound potency distributions. For our analysis, ChEMBL db (CDB)16 represented the most relevant public domain compound source. CDB is a well-curated database that contains >500 000 compounds with more than two million activity annotations and broad target coverage. The majority of CDB entries represent compound optimization data, that is, high-confidence activity annotations,16 which we considered an important criterion for a global target-based assessment of scaffold diversity. From CDB, 31 158 compounds active against human targets were selected that had the highest target confidence level (i.e., CDB target confidence score 9) for direct interactions (target relationship type "D"). These compounds represented 577 different target sets. From these sets, a total of 12 047 scaffolds were extracted according to Bemis and Murcko.5 Hence, all R-groups were removed from ring systems, but linkers between rings were retained. Figure 1 a shows the distribution of Bemis and Murcko scaffolds over all target sets. These sets contained from one to 615 scaffolds, but the majority of sets contained fewer than 50 scaffolds. Only 44 target sets consisted of compounds representing only a single scaffold. Figure 1 b shows the distribution of compound potency ranges over target sets. Approximately 75 % of the target sets contained active compounds with a potency spread of more than two orders of magnitude (the extreme case being renin inhibitors, with potency up to the attomolar (10−18 M) level). We compared the potency of compounds representing each scaffold. The median potency of all compounds representing a unique scaffold was calculated as the so-called "scaffold potency". Target set statistics. For all 577 target sets, distributions are reported for a) the number (#) of scaffolds, b) compound potency range, and c) percentage of scaffolds that are either involved in substructural relationships (black) or that have topologically equivalent scaffolds (light grey); that is, [# Scaffolds with Structural Relationships]/[Total # Scaffolds]. Scaffolds were also transformed into carbon skeletons (CSKs) by converting all heteroatoms to carbon atoms and all non-single bonds to single bonds. Thus, scaffolds producing the same CSK are topologically equivalent. On the basis of scaffolds and CSKs, two types of structural relationships were explored for each target set: 1) a scaffold is a substructure of another scaffold; 2) different scaffolds yield the same CSK. Accordingly, scaffold diversity was assessed at two levels. We considered scaffolds to be structurally diverse if they were 1) not involved in substructure relationships with others and 2) yielded unique CSKs (i.e., were topologically distinct). Figure 1 c reports the ratio of scaffolds involved in two types of structural relationships over the total number of scaffolds for all target sets. For the analysis of substructural and CSK relationships between scaffolds, the benzene ring, the most generic scaffold, was not considered because benzene was a substructure of the majority of CDB scaffolds. For 465 of 533 target sets containing multiple scaffolds (∼87 %), structural relationships were detected, which was a rather unexpected finding. Moreover, for 107 target sets, all scaffolds were found to be involved in substructural and/or CSK relationships. For those target sets containing at least two scaffolds and one structural relationship, we first determined the number of scaffold pairs that were involved in substructure relationships. Figure 2 a shows that one or more substructural scaffold relationships were observed in ∼85 % of the target sets. A total of 261 sets contained more than five substructural relationships. Among these, vascular endothelial growth factor receptor (VEGFR) 2 antagonists contained 475 substructure pairs, the overall largest number. Distribution of scaffold relationships. a) For 465 target sets containing at least two scaffolds and one structural relationship, the number (#) of scaffold pairs with substructural relationships is reported. b) Three overlapping sequential substructural scaffold paths of length 4 are shown for the VEGFR2 antagonist set; the median compound potency value is reported for each scaffold. c) Distribution of the number of scaffold pairs yielding the same CSKs. In 397 target sets with one or more substructure relationships, a total of 9020 unique substructure relationships were detected. The corresponding scaffold pairs were analyzed for sequential substructure relationships on a per-target basis (i.e., A is a substructure of B, B of C, and C of D, etc.). It was found that ∼80 % of the substructure relationships had a path length of 1 (i.e., A is a substructure of B, but B is not a substructure of another scaffold), ∼18 % had a path length of 2 (i.e., A is a substructure of B and B of C), and 1.5 % had of length of 3. Only five substructure paths of length 4 were identified. Three of these paths were overlapping and are shown in Figure 2 b. Hence, the majority of scaffolds were found to be involved in substructure relationships, but sequential relationships involving more than three scaffolds were rare. Next we identified scaffold pairs yielding the same CSKs. Figure 2 c shows their distribution. In nearly 90 % of all target sets, scaffold pairs yielding the same CSKs were detected. For 270 of these sets, more than five scaffold pairs were topologically equivalent (the maximum being 662 pairs for ligands of the melanin-concentrating hormone receptor 1). Thus, most of the target sets contained a substantial number of topologically equivalent scaffolds. For 348 of 533 target sets consisting of multiple scaffolds, both substructural relationships and CSK equivalences were observed. In 49 target sets, only substructural relationships were found, and in 68 sets, only CSK equivalences. In only 68 other target sets, no structural relationships were detected. Figure 3 a reports the distribution of scaffolds involved in different types of relationships over all target sets containing multiple scaffolds. As can be seen, sets with both substructure and CSK relationships typically consisted of many scaffolds. In contrast, targets with no scaffold relationships mostly contained only very few scaffolds. Figure 3 b shows an example of a target set with both substructural scaffold relationships and CSK equivalences. Figure 3 c and 3 d show representative examples of the only 14 target sets with more than five scaffolds but no structural relationships. In these cases, scaffolds were either completely unrelated (Figure 3 c) or related by symmetry and/or polymer character (Figure 3 d). Hence, the limited number of target sets with no substructural or CSK relationships included rather unusual active compounds. However, the majority of target sets were characterized by well-defined structural scaffold relationships involving most, if not all scaffolds. Target sets with scaffolds having defined structural relationships. a) The number of scaffolds involved in different types of relationships over all target sets containing multiple scaffolds; no structural relationship ("None"; white), only substructural relationships ("Substructure"; light grey), only CSK equivalence ("CSKeq"; dark grey), or substructure and CSK relationships ("Both"; black). b) Scaffolds from an exemplary target set (protein tyrosine kinase erbB-4 inhibitors) are shown that contain both types of structural relationships. Arrows denote substructure relationships, and grey shading indicates CSK equivalence. Median compound potency values are reported for all scaffolds. c), d) Scaffolds are shown for two representative target sets without structural relationships including inhibitors of c) tyrosine protein kinase BTK and d) glutathione S-transferase A1. We also carried out an analysis of structural relationships between scaffolds isolated from a set of 1586 clinical trials compounds extracted from the MDL Drug Data Report (MDDR),17 a set of 2980 registered or launched drugs extracted from DrugBank18 and the MDDR, as described previously,10 and a set of 50 000 synthetic compounds randomly collected from ZINC.19 These calculations were carried out without target set constraints because these compounds could not be systematically organized into defined target sets based on activity annotations. Thus, as a control, we also calculated scaffold relationships for the entire collection of bioactive CDB compounds without applying the target set organization. For all of these sets of medicinal chemistry relevant compounds, >90 % of the scaffolds were found to be involved in structural relationships. Hence, scaffold diversity, as defined herein, among these compounds was generally much lower than we anticipated, and bioactive compounds mirrored this low level of structural diversity. When target set constraints were applied, structural relationships among scaffolds were decreased; importantly, however, still >80 % of scaffolds extracted from specifically active CDB compounds displayed defined structural relationships. Therefore, these structurally related scaffolds corresponded to compounds with well-defined activity against currently available targets. We previously showed that different scaffolds have different propensities to form activity cliffs20 (an activity cliff is formed by two or more structurally similar compounds with marked difference(s) in potency21). Therefore, we also investigated activity cliff formation among topologically equivalent scaffolds. For each scaffold pair yielding the same CSK, all possible compound pairs were selected that represented one or the other scaffold. If the potency of compounds in a pair differed by at least two orders of magnitude, it was considered to form an activity cliff (and termed "compound cliff pair"). Figure 4 a reports the number of topologically equivalent scaffold pairs that were represented by multiple compound cliff pairs. In 258 target sets, no such scaffold pairs were found. However, 79 target sets contained more than five scaffold pairs represented by multiple compound cliff pairs. Distribution of compound cliff pairs. a) The target set distribution of scaffold pairs with varying numbers of compound cliff pairs is shown. b) The number of target sets that contain only scaffold pairs with unidirectional potency, bidirectional potency, or both. c) The distribution of uni- and bidirectional scaffold pairs represented by multiple compound cliff pairs. We then analyzed the potency "direction" among compound cliff pairs; that is, we examined whether compounds representing one of two topologically equivalent scaffolds were always more potent than compounds representing the other ("unidirectional") or not ("bidirectional"). Figure 4 b reports the potency direction among these pairs on a target set basis. A total of 122 target sets were exclusively unidirectional in compound cliff pair potency distribution, and only five sets were exclusively bidirectional; the remaining 80 targets contained both uni- and bidirectional scaffold pairs. Thus, a notable tendency for unidirectional potency among topologically equivalent scaffold pairs with activity cliff potential was observed. Figure 4 c shows the distribution of compound cliff pairs over topologically equivalent scaffold pairs, which further corroborates this trend. A total of 1564 unique scaffold pairs with at least two compound cliff pairs were found in 202 target sets, and 1398 of these scaffold pairs displayed unidirectional potency; 1218 unidirectional pairs occurred in single target sets, and the remaining 180 pairs in multiple sets. In contrast, only 166 bidirectional scaffold pairs were found in 85 target sets, 145 of which only occurred in single sets. The uni- and bidirectional scaffold pair sets shared 46 pairs. Thus, cliff-forming compounds representing these 46 pairs showed target-specific differences in activity. The prevalence of topologically equivalent unidirectional scaffolds indicates that the choice of these scaffolds plays an essential role for achieving high compound potency. In Figure 5, exemplary scaffolds with defined structural relationships are shown that correspond to compounds forming significant activity cliffs. These scaffolds display substructural relationships (Figure 5 a) or are topologically equivalent (Figure 5 b). Activity-cliff-forming scaffolds with defined structural relationships. a) Ten scaffolds from inhibitors of the β-2 adrenergic receptor are shown that are involved in substructural relationships. The root scaffold is shown on a grey background, and nine scaffolds containing the root as a substructure are arranged according to increasing scaffold potency values (i.e., the median potency of compounds corresponding to each scaffold). Structural regions outside the root substructure are circled. b) Four scaffolds extracted from thymidylate synthase inhibitors that yield the same carbon skeleton are shown. Differences in heteroatom positions and bond orders are highlighted, and scaffold potency values are reported. In summary, we have shown that the majority of currently available molecular scaffolds representing compounds active against human targets display substructure relationships and/or are topologically equivalent. Of the 12 047 scaffolds analyzed herein, a total of 9993 scaffolds were involved in these well-defined structural relationships. Thus, true scaffold diversity among active compounds was limited. However, we have also shown that this applies, in the absence of target set constraints, to randomly selected medicinal chemistry relevant and druglike compounds or drugs. Hence, bioactive compounds essentially mirror general scaffold relationships in currently available small molecules. These findings might suggest that biologically relevant chemical space is small, perhaps even smaller than previously thought, and that the identification of distinctly different compound structures sharing a specific target activity might often be difficult. However, the high frequency of substructural and topological relationships between currently available medicinal chemistry relevant scaffolds might also suggest that many medicinal chemistry efforts build upon prior knowledge and modify known structural motifs. Regardless, among topologically equivalent scaffolds with the potential to form activity cliffs, many scaffolds display clear preferences over others to yield highly potent compounds. These observations indicate that structurally related scaffolds might yield rather different target-dependent compound activity. Thus, minor differences in scaffold structures might substantially affect compound optimization efforts. All calculations were carried out with in-house written Perl or Scientific Vector Language (SVL)22 scripts and Pipeline Pilot23 programs. For CDB compounds with multiple potency measurements reported for the same target (either Ki or IC50 values), the geometric mean was calculated to yield the final potency value. Compounds, scaffolds, and CSKs were represented in SMILES24 format for processing.