Figures
Abstract
The ability of target proteins to bind structurally diverse compounds and compounds with different degrees of promiscuity (multi-target activity) was systematically assessed on the basis of currently available activity data and target annotations. Intuitive first- and second-order target promiscuity indices were introduced to quantify these binding characteristics and relate them to each other. For compounds and targets, opposite promiscuity trends were observed. Furthermore, the analysis detected many targets that interacted with compounds representing a similar degree of structural diversity but displayed strong tendencies to recognize either promiscuous or selective compounds. Moreover, target families were identified that preferentially interacted with promiscuous compounds. Taken together, these findings further extend our understanding of the molecular basis of polypharmacology.
Citation: Hu Y, Bajorath J (2015) Quantifying the Tendency of Therapeutic Target Proteins to Bind Promiscuous or Selective Compounds. PLoS ONE 10(5): e0126838. https://doi.org/10.1371/journal.pone.0126838
Academic Editor: Andrea Cavalli, University of Bologna & Italian Institute of Technology, ITALY
Received: March 6, 2015; Accepted: April 8, 2015; Published: May 22, 2015
Copyright: © 2015 Hu, Bajorath. This is an open access article distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited
Data Availability: All data used are freely available in the ChEMBL database (version 20) and can be obtained applying the protocol reported in the manuscript.
Funding: The authors received no specific funding for this work.
Competing interests: The authors have declared that no competing interests exist.
Introduction
Polypharmacology is an emerging theme in pharmaceutical research and chemical biology based upon the premise that compounds frequently act on multiple targets [1–5], thereby triggering complex functional responses and pharmacological effects. Compound promiscuity, defined as the ability of small molecules to specifically interact with multiple targets, provides the molecular basis of polypharmacology [6,7]. On the other hand, since there are many more active compounds than targets available, polypharmacology also requires the ability of targets to specifically bind multiple (and structurally distinct) ligands. In other words, many pharmaceutically relevant proteins must be “good” small molecule targets. Otherwise, polypharmacology on a larger scale would be difficult to rationalize. An analysis of compounds active against the current spectrum of pharmaceutical targets has revealed that many targets recognize large numbers of structurally diverse compounds [8], which is well in accord with assumed ligand-target interaction characteristics underlying polypharmacology, as discussed above.
While compound/drug promiscuity has been the topic of a number of investigations and reviews [5–7], promiscuity at the target level has thus far only been little explored in a systematic manner. Compound promiscuity can be quantified by collecting available high-confidence activity/target annotations, thereby providing a conservative estimate of the degree of promiscuity [5,6]. Analogously, one might estimate target promiscuity by counting the number of known structurally distinct active compounds for a given target for which well-defined activity measurements are available. Such simple measures are sufficient to assign different promiscuity levels to active compounds and targets on the basis of currently available data or aid in the generation of compound-based target or drug-target networks. However, they do not provide any information about the potential interplay of promiscuity at the ligand and target levels.
Having studied compound promiscuity from different viewpoints [6,7], we have been interested in exploring target promiscuity taking compound promiscuity information into account. Specifically, we have asked the questions if there might be detectable tendencies for targets to either recognize promiscuous or selective compounds and how such tendencies might relate to the ability of targets to interact with increasing amounts of structurally diverse compounds. The analysis presented herein was designed to address these and related questions and has yielded in part surprising results, as detailed in the following.
Material and Methods
Data collection
From the latest version of ChEMBL (release 20) [9], compounds were extracted for which direct interactions (i.e., assay relationship type “D”) with human targets at the highest level of confidence (i.e., assay confidence score 9) were reported. Only “single protein” targets were considered. Two different types of potency measurements, including (assay-independent) equilibrium constants (Ki) and (assay-dependent) IC50 values, were separately collected (because these types of measurements should not be directly compared). To ensure high data confidence, only explicitly defined potency values were retained. All approximate measurements such as “>”, “<”, or “∼” were discarded. Compounds with multiple Ki or IC50 values for the same target were selected if all values fell within the same order of magnitude. Then, the geometric mean of all values was calculated as the final potency annotation. In addition, only compounds with at least 1 μM potency (i.e., pKi or pIC50 ≥ 6) were considered. Furthermore, all targets with active compounds were organized into target families following the protein classification hierarchy of ChEMBL and UniProt family annotations [10].
On the basis of these selection criteria, two activity measurement-dependent data sets were generated, including a Ki and an IC50 value-based set. If a compound was annotated with both Ki and IC50 values, it was assigned to both sets. In addition, from all qualifying compounds, molecular scaffolds were extracted by removing all side chains and retaining ring systems and linkers between them [11]. Scaffolds were isolated to represent structurally distinct compound series. In addition, scaffolds were further reduced to cyclic skeletons (CSKs) by converting all heteroatoms to carbon and all bond orders to one [12]. Hence, each CSK represented a set of topologically equivalent scaffolds.
Assessment of target promiscuity
To assess the degree of target promiscuity, different indices were defined, as illustrated in Fig 1. On the basis of high-confidence compound activity data assembled from ChEMBL, the activity profile of a compound was generated by collecting all available target annotations. Accordingly, for each compound, the number of its known targets was counted to yield the compound promiscuity index (CPI). In the example in Fig 1, compound 1 is active against three targets, yielding a CPI value of 3. Furthermore, compounds active against the same target were grouped. For example, in Fig 1, target TA interacts with four compounds (1, 6, 9, 10) and target TC with a distinct set of three compounds (2, 4, 5). For each target, the number of unique scaffolds representing active compounds was determined as the first-order target promiscuity index (TPI_1). Furthermore, CPI values of all compounds known to interact with a given target were summed and the average CPI value was calculated to yield the second-order target promiscuity index (TPI_2). For example, in Fig 1, the four compounds active against target TA contain three unique scaffolds, resulting in a TPI_1 value of 3. In addition, these four compounds have a total of nine target annotations, yielding a TPI_2 value of 2.3 for TA. By contrast, compounds 2, 4, and 5 are exclusively active against TC, resulting in a TPI_2 value of 1 for TC.
Shown is a workflow that illustrates how first- and second-order target promiscuity indices are calculated. On the basis of compound activity data, the activity profile of a compound is generated by collecting all available target annotations (top). Accordingly, for each compound, the number of targets it is active against is counted to yield the compound promiscuity index (CPI). Then, all compounds active against the same target are grouped (bottom). For each target, the number of unique scaffolds contained in its ligands is determined as the first-order target promiscuity index (TPI_1). Furthermore, CPI values of all compounds interacting with a given target are summed and the average CPI value is calculated as the second-order target promiscuity index (TPI_2).
Results and Discussion
Activity data and compound sets
Initially, we briefly summarize the results of data selection and curation and the assembly of the data sets upon which our subsequent promiscuity analysis was based.
Organization of compound data sets.
On the basis of the data selection and curation criteria detailed above, two sets of compounds were assembled for which high-confidence activity data for human targets were available by separately considering Ki and IC50 measurements, as reported in Table 1. In this context, it is also noted that records of inactivity in target-based assays were not available for compounds selected for promiscuity analysis. The Ki value-based set consisted of 43,086 compounds active against 613 targets. These compounds formed a total of 67,049 compound-target interactions and were represented by 16,071 unique scaffolds and 7880 CSKs. The IC50 set was much larger than the Ki set, containing 75,244 compounds annotated with 1069 targets forming nearly 95,000 compound-target interactions. The IC50 set compounds yielded 28,875 scaffolds and 12,856 CSKs (Table 1).
Compound, scaffold, and CSK distributions.
Fig 2 reports the distribution of compounds, scaffolds, and CSKs over different target proteins. For ~35% (Ki set) and ~31% (IC50 set) of all targets, only one to five compounds were available, as reported in Fig 2A. For the majority of the targets, 10 or more active compounds were available. Moreover, 32 targets (i.e., ~5%; Ki) and 36 targets (~3%; IC50) with more than 500 active compounds were identified. Fig 2B and 2C reveal comparable distributions for scaffolds and CSKs for the Ki and IC50 sets. For large numbers of target proteins, active compounds were found to contain one to five scaffolds or CSKs. In particular, for ~20% (Ki) and ~16% (IC50) of the targets, only one scaffold or CSK was available. On average, compounds active against each target yielded 45 and 38 scaffolds and 29 and 25 CSKs for the Ki and IC50 value-based sets, respectively, reflecting the average degree of scaffold diversity across current pharmaceutical targets. Compared to the IC50 set, targets in the Ki set were generally associated with more compounds, scaffolds, and CSKs. Targets for which fewer than 10 active compounds were available were not further considered (given their low degree of exploration). The final Ki and IC50 data sets assembled for promiscuity analysis comprised 354 and 649 targets, respectively.
The percentage of targets with increasing numbers of (A) compounds, (B) scaffolds, and (C) CSKs is reported for the Ki (red) and IC50 (blue) data sets.
Promiscuity indices
Concept.
Different promiscuity indices were defined for our analysis, as illustrated in Fig 1. Counting the number of target annotations for a given compound yielded the compound promiscuity index (CPI), a standard measure for assessing the degree of compound promiscuity that is often applied [6]. Furthermore, to assess target promiscuity, two indices were defined. For each target, the number of unique molecular scaffolds from all active compounds was determined, yielding the first-order target promiscuity index (TPI_1). This index accounted for the ability of a target to interact with structurally diverse compounds. We note that this index did-by design- not consider the number of compounds represented by each scaffold, which would often bias the statistics. For example, if a scaffold represented 10 related active analogs, it was considered equivalent to a scaffold representing two actives. Hence, the total number of different core structures recognized by a given target was accounted for by TPI_1 (not the absolute number of compounds represented by them). In addition, CPI values of all compounds active against a target were summed and the average CPI value was calculated to yield the second-order target promiscuity index (TPI_2). Thus, TPI_2 accounted for the degree of promiscuity among all compounds active against the target. Accordingly, different from TPI_1, the total number of active compounds was taken into consideration in the calculation of TPI_2. The minimal value of TPI_2 was 1, indicating that all compounds active against a given target were exclusively active against this target. By contrast, a TPI_2 value of 5 would indicate that compounds active against the target were on average active against five targets. Therefore, comparison of TPI_1 and TPI_2 revealed if a target that interacted with a certain amount of structurally distinct compounds might preferentially bind promiscuous compounds (with multi-target activities) or more selective compounds. These comparison can be extended to multiple targets, for example, targets with the same or similar TPI_1 values (i.e., targets binding compounds with a comparable level of scaffold diversity) or entire target families. For example, in Fig 1, targets C and D interact with compounds represented by a single scaffold, thus yielding the same TPI_1, but different TPI_2 values (i.e., 1.0 vs. 2.7) because these active compounds have different promiscuity.
We also note that the conventional CPI definition applied here does not take into account if targets of promiscuous compounds are related to each other or not. However, it has recently been shown that only ~2% of bioactive compounds are promiscuous across different unique target families on the basis of high-confidence activity data (as used herein) [7]. Thus, most promiscuous compounds act on related targets, as quantified by CPI calculations. This also has implications for the consideration of other possible compound promiscuity measures. For example, one could envision introducing a CPI variant to account for activity against unique target families, rather than individual targets. However, given the very low promiscuity rate across different families, most values of this CPI variant would be one (and hence not suitable for TPI_2 calculations).
Distribution of promiscuity indices.
For 354 (Ki) and 649 targets (IC50) with at least 10 active compounds, the distribution of TPI_1 and TPI_2 values is reported in Fig 3. The value distributions were comparable for the Ki and IC50 sets. The TPI_1 value distribution in Fig 3A shows that the majority of targets had active compounds yielding more than 10 distinct scaffolds. The average TPI_1 value was 77 and 61 for the Ki and IC50 sets, respectively, indicating that many targets bound structurally diverse compounds (i.e., active compounds had many different core structures). Fig 3B shows the TPI_2 value distribution. Similar to previous studies reporting that ~35% of active database compounds had multi-target activity [1,7], our CPI calculations revealed that ~33% of compounds in the Ki but only 17% in the IC50 set were active against more than one target. The average CPI values were 1.6 (Ki) and 1.3 (IC50).
Shown is the distribution of (A) TPI_1 and (B) TPI_2 values for 354 targets from the Ki (red) and for 649 targets from the IC50 set (blue), respectively. For each of these targets, at least 10 active compounds were available.
In light of these findings, one might also anticipate obtaining comparably low TPI_2 values. Surprisingly, however, only ~18% of all targets interacted with compounds having exclusive single-target activity (i.e., producing a TPI_2 value of 1). By contrast, more than 80% of the targets interacted with one or more compounds having multi-target activity. For ~36% (Ki) and ~30% (IC50) of the targets, TPI_2 values larger than 2 were obtained (with average TPI_2 values of 2.1 and 2.0 for the Ki and IC50 sets, respectively). Hence, essentially opposite promiscuity trends were observed for compounds and targets. Whereas the majority of compounds was only active against a single target, most targets bound varying numbers of promiscuous compounds.
Comparison of TPI_1 and TPI_2 values
Relationships between TPI_1 and TPI_2 values were further analyzed. As shown in Figs 4A and 5A for the Ki and IC50 sets, respectively, there was no apparent correlation between these two target promiscuity indices. Targets with TPI_1 values of less than 200 had a much broader distribution of TPI_2 values than targets with largest TPI_1 values (> 200). Furthermore, heat map representations of promiscuity index combinations were generated for targets from the Ki and IC50 sets, shown in Figs 4B and 5B, respectively. In these heat maps, rows represent seven ranges of TPI_2 values and columns six ranges of TPI_1 values. Each cell indicates the number of targets having corresponding TPI_1 and TPI_2 values. In addition, each row reflects the distribution of TPI_1 values for targets having comparable TPI_2 values and each column the distribution of TPI_2 values for targets having similar TPI_1 values.
(A) For 354 targets from the Ki set, their TPI_1 and TPI_2 values are compared. Each dot in the scatter plot represents a target. The correlation coefficient (R2) for TPI_1 and TPI_2 values is provided. (B) Relationships between TPI_1 and TPI_2 values are captured in a heat map in which cells are colored according to the population density of targets. In addition, the number of targets is reported for cells that were populated with more than 20 targets using white numbers.
For 649 targets from the IC50 set, their TPI_1 and TPI_2 values are compared. The representation is according to Fig 4.
For the Ki set (Fig 4B), targets interacting with compounds containing up to 200 distinct scaffolds displayed many different TPI_2 values covering five or six value ranges. The overall largest TPI_2 value (11.1) was observed for the histamine H2 receptor with a set of 26 antagonists (represented by 24 scaffolds). Hence, these antagonists were highly promiscuous. The majority of the targets produced low to intermediate TPI_2 values ranging from 1 to 3. The five most populated cells in the heat map contained 211 targets (i.e., ~60%). A subset of 117 targets with active compounds containing at most 20 scaffolds yielded TPI_2 values between 1 and 2.
Table 2 lists 10 exemplary targets from the Ki set that yielded the same or very similar TPI_1 values of varying magnitude but significantly different TPI_2 values. For example, compounds active against dihydroorotate dehydrogenase and NADPH oxidase 5 contained the same number scaffolds. However, inhibitors of dihydroorotate dehydrogenase had no other reported activities (TPI_2 value of 1.0), whereas all inhibitors of NADPH oxidase 5 had multi-target activity, resulting in a TPI_2 value of 3.4. In addition, for two related G protein coupled receptors (GPCRs; purinergic receptor P2Y12 and alpha-2c adrenergic receptor), known antagonists contained comparably large numbers of scaffolds (142 vs. 149), but their TPI_2 values differed significantly (1.0 vs. 5.9). Thus, purinergic receptor P2Y12 antagonists were exclusively active against this target, whereas 87.5% of the alpha-2c adrenergic receptor antagonists had multi-target activity.
For the IC50 set (Fig 5B), observations similar to the Ki set were made. Four targets were identified that produced TPI_2 values greater than 10 including alpha-1d, -2b, and -2c adrenergic receptors and fibroblast growth factor receptor 3. Compounds active against these targets contained nine to 43 scaffolds. As reported in Table 3, a variety of targets were identified having the same or very similar TPI_1 but significantly different TPI_2 values.
Taken together, these results revealed that many different targets that recognized ligands with comparable degrees of structural diversity displayed markedly different tendencies to preferentially interact with selective or promiscuous compounds; a rather unexpected finding.
Target family promiscuity
In light of these observations, the distribution of TPI_2 values was analyzed for 10 target families from the Ki set and 14 families from the IC50 set, which contained at least 10 targets each, as reported in Table 4. As discussed in the following, target families displayed very different promiscuity patterns.
Fig 6 reports the intra-family distribution of TPI_2 values in a pie chart format. In Fig 6A (Ki set), three target families (ID 59, 64, and 208) that contained 11 or 12 targets yielded distinct intra-family TPI_2 distributions. For the chemokine receptor family (ID 64), ~1/3 of the targets only interacted with selective compounds (TPI_2 value of 1) while ~2/3 yielded TPI_2 values between 1 and 2, due to ligands with multi-target activity. For the nuclear hormone receptor family (208), no target was found to only interact with selective compounds. Fig 6B (IC50 set) reveals comparable results for four families (ID 64, 143, 222, and 234) with 10 or 11 targets including chemokine receptors, which displayed varying preferences for selective or multi-target compounds.
The distribution of targets with varying TPI_2 values is reported in pie charts for (A) 10 target families from the Ki set and (B) 14 families from the IC50 set that contain at least 10 individual targets. Each color-coded pie chart segment reports the proportion of targets with TPI_2 values falling into a specific range. Seven value ranges are defined and colored-coded, as indicated on the right. For each family, an ID (bold) and the number of targets are provided. For example, “59: 12” means that family 59 contains 12 targets (Ki set). Target families are listed in Table 4.
Several of the target families in Table 4 were closely related to each other including different GPCR, kinase, or protease families. For related families, different promiscuity patterns also emerged. For example, four GPCR families (ID 64, 165, 183, and 281) were associated with both the Ki and IC50 sets and showed different distributions of TPI_2 values. The degree of target promiscuity increased from the chemokine (64) over the lipid-like ligand (165) and short peptide (281) to the monoamine (183) receptor family. Hence, targets in these families showed an increasing tendency to bind promiscuous ligands. Furthermore, the serine/threonine (275) and tyrosine (319) kinase families displayed similar distributions of TPI_2 values for the IC50 set (Fig 6B) that notably differed from the PI3/PI4-kinase family (222).
Finally, the analysis of TPI_2 value distributions also identified target families with an overall strong preference to interact with promiscuous compounds including, for example, the carbonic anhydrase (Fig 6A; ID 59), histone deacetylase (Fig 6B; 143), or monoamine receptor family (Fig 6A and 6B; 183). In particular, histone deacetylases and monoamine receptors continue to be high-profile therapeutic targets and medicinal chemistry efforts are often heralded to identify new active compound classes for them. However, targets in these families are shown to display a strong tendency to recognize promiscuous compounds and are likely to be involved in many polypharmacological effects. The characteristics should be considered in the context of drug development.
Conclusions
An intuitive methodological framework has been introduced to systematically explore target promiscuity. Although the exploration of polypharmacology has thus far mostly focused on compound promiscuity, differences in the ability of targets to interact with small molecules inevitably also make important contributions to the formation of polypharmacological networks. For our analysis of target promiscuity, simple first- and second-order target promiscuity indices were designed to quantify the tendency of targets to recognize structurally diverse and promiscuous compounds and relate these characteristics to each other. Care was taken to select high confidence activity data and target annotations as a basis for the analysis. Because assay-independent Ki and assay-dependent IC50 values cannot be directly compared, Ki- and IC50-based data sets were separately generated and yielded similar results in promiscuity analysis. However, for compounds and targets, opposite promiscuity trends were detected. The majority of compounds were only active against a single target, whereas most targets bound varying numbers of promiscuous compounds. On the basis of TPI_1 calculations, many targets interacted with compounds representing different levels of scaffold diversity. TPI_2 calculations then revealed that many targets preferentially bound either selective or promiscuous compounds. Importantly, a variety of targets with ligands of comparable structural diversity displayed markedly different preferences to interact with compounds having single- or multi-target activity. This was also observed for targets capable of binding structurally highly diverse compounds. Furthermore, preferences for binding of selective vs. promiscuous compounds emerged at the level of target families that mostly interacted with promiscuous compounds. Structural features of targets or families that correlate with their propensity to interact with promiscuous vs. selective compounds are currently unknown, which provides opportunities for future research.
Taken together, the findings reported herein further improve our understanding of promiscuity at the level of targets and refine our view of the molecular basis of polypharmacology. In addition, through calculation and comparison of target promiscuity indices, as introduced herein, it can easily be estimated how likely it might be to identify selective compounds for a target of interest on the basis of available compound activity data for this and closely related targets. Furthermore, targets that are most likely to contribute to polypharmacology networks can also be identified via the same route. These practical applications should be of considerable interest in pharmaceutical research.
Acknowledgments
The authors are grateful to Gerald Maggiora for many discussions concerning polypharmacology and ongoing collaborative efforts.
Author Contributions
Conceived and designed the experiments: YH JB. Performed the experiments: YH. Analyzed the data: YH JB. Wrote the paper: YH JB.
References
- 1. Paolini GV, Shapland RH, van Hoorn WP, Mason JS, Hopkins AL. Global mapping of pharmacological space. Nat. Biotechnol. 2006;24: 805–815. pmid:16841068
- 2. Hopkins AL. Network pharmacology: the next paradigm in drug discovery. Nat. Chem. Biol. 2008;4: 682–690. pmid:18936753
- 3. Peters JU. Polypharmacology—foe or friend? J. Med. Chem. 2013;56: 8955–8971. pmid:23919353
- 4. Anighoro A, Bajorath J, Rastelli G. Polypharmacology: challenges and opportunities in drug discovery. J. Med. Chem. 2014;57: 7874–7887. pmid:24946140
- 5. Jalencas X, Mestres J. On the origins of drug polypharmacology. Med. Chem. Comm. 2013;4: 80–87.
- 6. Hu Y, Bajorath J. Compound promiscuity: What can we learn from current data? Drug Discov. Today 2013;18: 644–650. pmid:23524195
- 7. Hu Y, Bajorath J. How promiscuous are pharmaceutically relevant compounds? A data-driven assessment. AAPS J. 2013;15: 104–111. pmid:23090085
- 8. Hu Y, Bajorath J. Global assessment of scaffold hopping potential for current pharmaceutical targets. Med. Chem. Comm. 2010;1: 339–344.
- 9. Bento AP, Gaulton A, Hersey A, Bellis LJ, Chambers J, Davies M, et al. The ChEMBL bioactivity database: an update. Nucleic Acids Res. 2014;42: D1083–D1090. pmid:24214965
- 10. UniProtConsortium. Reorganizing the protein space at the Universal Protein Resource (UniProt). Nucleic Acids Res. 2012;40: D142–D148.
- 11. Bemis GW, Murcko MA. The properties of known drugs. 1. Molecular frameworks. J. Med. Chem. 1996;39: 2887–2893. pmid:8709122
- 12. Xu YJ, Johnson M. Using molecular equivalence numbers to visually explore structural features that distinguish chemical libraries. J. Chem. Inf. Comput. Sci. 2002;42: 912–926. pmid:12132893