Exploring Structural Relationships between Bioactive and Commercial Chemical Space and Developing Target Hypotheses for Compound Acquisition.

Analog series were systematically extracted from more than 650 000 bioactive compounds originating from medicinal chemistry and screening sources and more than 3.6 million commercial compounds that were not biologically annotated. Then, analog series-based (ASB) scaffolds were generated. For each scaffold from a bioactive series, a target profile was derived and ASB scaffolds shared by bioactive and commercial compounds were determined. On the basis of our analysis, large segments of commercial chemical space were not yet explored biologically. Shared ASB scaffolds established structural relationships between bioactive and commercial chemical space, and the target profiles of these scaffolds were transferred to commercially available analogs of active compounds. This made it possible to derive target hypotheses for more than 37 000 compounds without biological annotations covering more than 1000 different targets. For many molecules, alternative target assignments were available. Target hypotheses for these compounds should be of interest, for example, for hit expansion, acquisition of compounds to design or further extend focused libraries for drug discovery, or testing of expanded analog series on different targets. They can also be used to search for analogs and complement compound series during target-directed optimization. Therefore, all of the commercial molecules with new target hypotheses as well as key scaffolds identified in our analysis and their target profiles are made freely available.

Introduction

Volumes of publicly available bioactive compounds and other small molecules increase rapidly, as exemplified by three major databases that collect bioactive compounds from different sources and small molecules from chemistry vendors. For example, release 22 of ChEMBL,[1] the major public repository of active compounds from the medicinal chemistry literature and patents contains nearly 1.7 million compounds with activity against more than 11 000 targets. In addition, PubChem,[2] the major public resource of screening data, lists more than 93 million compound entries and 1.25 million assays. Taken together, these two major databases contain more compounds and activity data than could have been imagined just a few years ago. Moreover, release 15 of ZINC,[3] which collects compounds from chemistry vendors worldwide, reports an astonishing number of more than 200 million entries. Thus, in terms of compounds, structures, and activity data, chemistry is without doubt entering the “big data” era.[4] ZINC compounds are not biologically annotated but represent a vast sample of chemical space that is relevant for drug discovery. Accordingly, the large numbers of bioactive compounds and vendor molecules that are available provide an unprecedented opportunity to systematically compare structural relationships between active compounds and other small molecules that are currently without biological annotations and attempt to derive target hypotheses for molecules that are available commercially. Therefore, we have carried out a large-scale comparison of compound subsets selected from ChEMBL and PubChem on the basis of well-defined activity criteria and target annotations and ZINC molecules that are most likely available from vendor sources. The analysis was facilitated using the recently introduced concept of analog series-based (ASB) scaffolds.[5] Different from other scaffold definitions,[6] the ASB scaffold formalism first systematically extracts analog series from compound collections and then isolates conserved core structures from series that distinguish between analogs on the basis of a single substitution site (if available).[5] Thus, ASB scaffolds carry series information and account for structural relationships. Furthermore, ASB scaffolds can be annotated with targets of compounds they represent. Biologically annotated ASB scaffolds can then be used to develop target hypotheses for other compounds containing the same scaffold. This approach was recently successfully applied to target deconvolution of chemical cancer cell line screens.[7]

Herein, we report our systematic comparison of bioactive and commercial compounds and derivation of target hypotheses for nonannotated molecules. The large number of target hypotheses that were obtained, frequently including multiple assignments for compounds, should be of interest for a variety of practical applications as discussed below.

Materials and Methods

Compounds and Activity Data

Commercially available compounds were collected from ZINC version 15 (accessed March 5, 2017).[3] To narrow down commercial chemical space and focus our analysis on compounds desirable for medicinal chemistry, the druglike subset was selected from ZINC using the “tranche browser”. Reactivity was set to “anodyne” (i.e., compounds that do not contain problematic functionalities, also excluding pan-assay interference compounds[2]) and availability to “in-stock” (i.e., ready for delivery). The resulting compound subset was downloaded in two dimensional SMILES format and standardized with inhouse Python scripts implemented in OpenEye chemistry toolkit.[9] These selection procedures resulted in set of 3 658 425 unique compounds from ZINC.

Bioactive compounds were assembled from ChEMBL[1] version 22 and the current release of PubChem BioAssays.[2] From ChEMBL, only compounds with high-confidence activity data were considered. Therefore, compounds involved in direct interactions (target relationship type “D”) with human targets at the highest confidence level (target confidence score 9) were selected. As potency measurements, only assay-independent equilibrium constants (Ki values) and assay-dependent IC50 values were considered. Approximate measurements associated with “>”, “<”, or “∼” were discarded. If multiple Ki or IC50 values were available for the same compound, the geometric mean of the values was calculated as the final potency annotation, provided all of the values fell within the same order of magnitude. Otherwise, the measurements were discarded. Given these selection criteria, a total of 224 532 unique compounds were obtained with activity against 1687 human targets.

From PubChem BioAssays, a subset of 426 294 unique compounds tested both in primary assays (percentage of inhibition from a single dose) and confirmatory assays (dose–response assays yielding IC50 values) was selected.[10] Only compounds classified as active or inactive were taken, whereas compounds with designations such as unspecified or inconclusive were discarded. From primary assays, RNA interference screens were removed, whereas all of the chemical screens were retained, including cell-based assays for which no individual target was specified. For confirmatory assays, a series of selection criteria was applied with the aim to collect chemical compounds tested against single target proteins,[10] yielding a total of 299 021 unique compounds with activity against 772 targets.

Scaffold Analysis

ASB scaffolds were identified following a previously described two-stage procedure.[5] First, the analog series was identified by applying the matched molecular pair (MMP) formalism.[8] An MMP is defined as a pair of compounds that differ only by a structural change (chemical transformation) at a single site.[11] Thus, an MMP consists of a shared core and a pair of exchanged substituents. Specifically, MMPs were generated using an adaptation of Hussain and Rea method,[12] which systematically fragments molecules at exocyclic single bonds. Instead of random fragmentation, we applied retrosynthetic combinatorial analysis procedure (RECAP) fragmentation rules,[13] according to which, exchanged fragments conformed with chemical reactions, yielding RECAP-MMPs.[14] Transformations in RECAP-MMPs were confined to limited size because previously established fragment size restrictions[15] were applied to limit RECAP-MMPs to pairs of compounds representing typical analogs. Specifically, the size of the conserved RECAP-MMP core was required to be at least twice the size of the larger exchanged fragment that was allowed to consist of at most 13 heavy atoms.[15] These size restrictions ensured that substituents were limited to maximally a condensed two-ring system with no more than three additional atoms conforming with largest substitutions typically seen in the analog series from medicinal chemistry.

Second, a network was generated in which the nodes represented compounds and edges pairwise RECAP-MMP relationships. In this network, each separate cluster represented a unique series of analogs[16] from which ASB scaffolds were extracted.[5] Hence, ASB scaffold generation combines analog search and scaffold extraction algorithms. Because analog series often contained multiple possible RECAP-MMP cores, a computational search was carried out for a core that captured all of the MMP relationships within the series. If more than one core met this criterion, the largest one was selected. The qualifying RECAP-MMP core represented the ASB scaffold of the series.[5] ASB scaffold analysis and compound statistics are summarized in Table . All of the calculations were carried out using inhouse Perl and Python scripts with the aid of the OpenEye chemistry toolkit and KNIME protocols.[9,17]

Table 1

Compound, Analog Series, and ASB Scaffold Statisticsa

	ZINC	ChEMBL	PubChem
# unique CPDs	3 658 425	224 532	426 294
# analog series	264 496	22 015	42 513
# CPDs in analog series	2 147 784 (58.7%)	133 441 (59.4%)	259 019 (60.7%)
# ASB scaffolds	208 004 (78.6%)	15 625 (71%)	31 602 (74.3%)
# CPDs in series with ASB scaffolds	604 382 (28.1%)	51 308 (38.4%)	92 794 (35.8%)
# CPDs per series	2–321	2–60	2–65

The distribution of compounds (CPDs) over analog series obtained from ZINC, ChEMBL, and PubChem subsets is reported. In addition, the ASB scaffold statistics are provided.

Target Annotations

A pool of unique 1687 targets originating from selected ChEMBL compounds and 772 targets from selected PubChem compounds was used to develop plausible target hypotheses for commercially available ZINC molecules. For each analog series, the union of target annotations from bioactive compounds comprising the series was determined and assigned to ZINC molecules sharing the corresponding ASB scaffold. To identify unique targets and assign targets to families, the UniProt classification scheme was applied.[18] Therefore, target GI numbers from PubChem and IDs from ChEMBL were mapped to UniProt IDs, hence ensuring consistency of target annotations.

Results and Discussion

ASB Scaffold Analysis

ASB scaffolds were systematically extracted from the analog series originating from 224 532 ChEMBL compounds for which high-confidence activity data and target annotations were available, 426 294 most extensively assayed PubChem compounds, and 3 658 425 ZINC molecules that had the highest probability of commercial availability. As reported in Table , for each database, around 60% of the compounds participated in the formation of analog series. ZINC, ChEMBL, and PubChem compounds yielded 264 496, 22 015, and 42 513 series, respectively, from which ASB scaffolds were isolated. Figure shows the size distribution of analog series. In each case, more than 60% percent of all of the series were compound pairs, an additional 23–27% of the series consisted of 3–5 compounds, and less than 10% of 6–10. Thus, whereas some very large series were extracted from each database (Table ), small series dominated the distribution. From more than 70% (ChEMBL, PubChem) to close to 80% (ZINC) of all analog series, ASB scaffolds were obtained. ZINC, ChEMBL, and PubChem series yielded 208 004, 15 625, and 31 602 ASB scaffolds, respectively, which represented in total about a 3rd of the initially selected compounds (Table ). ASB scaffold statistics from ChEMBL were consistent with the previous findings reported by Stumpfe et al.[19] It is noted that ASB scaffolds from ChEMBL exclusively represented active compounds, whereas scaffolds from PubChem represented both active and consistently inactive screening compounds. Scaffolds from ZINC represented compounds without activity annotations. Importantly, the comparison revealed that ZINC contained many more analog series and ASB scaffolds than were available for bioactive compounds. In addition, more than 40% of the ZINC compounds did not participate in series. Taken together, these observations indicate that ZINC covers large chemical space that is most likely not yet explored biologically.

Figure 1

Distribution of compounds over series. Shown is the distribution of compounds over analog series with ASB scaffolds. Percentages give the proportion of series with 2 (blue), 3–5 (purple), 6–10 (green), and 11 or more (orange) compounds. In each case, the largest proportion of series were compound pairs.

Scaffold Overlap

Next, the overlap between the independently derived ASB scaffold populations was determined. The results are shown in Figure . ChEMBL and ZINC shared 1216 ASB scaffolds, whereas PubChem and ZINC shared 11 270 scaffolds. For PubChem, only ASB scaffolds representing multiple active compounds were taken into consideration (inactive compounds were omitted). The intersection between all of the 3 scaffold populations contained 581 scaffolds. On the basis of shared ASB scaffolds, structural relationships between bioactive and ZINC compounds were established. As reported in Figure , 2608 and 32 962 ZINC compounds shared ASB scaffolds and analog relationships with 1705 ChEMBL and 17 001 PubChem compounds, respectively. The 581 ASB scaffolds that were shared by all of the 3 scaffold populations represented 292 ChEMBL, 1230 PubChem, and 2193 ZINC compounds.

Figure 2

Overlap between ASB scaffolds from different sources. The Venn diagram reports the overlap of ASB scaffolds derived from ChEMBL 22, PubChem, and ZINC. The number of shared ASB scaffolds and corresponding compounds is reported. ZINC compounds were associated with the union of unique target annotations from bioactive compounds represented by the shared ASB scaffold. For ASB scaffolds shared by ChEMBL and PubChem, unique targets were determined.

Target Transfer

Targets of a series of bioactive compounds sharing the same ASB scaffold were assembled and assigned to the scaffold. The union of unique targets then constituted the target profile of the scaffold, representing a form of metadata. ZINC molecules sharing the same scaffold were analogs of bioactive compounds from which the target profile was derived. Accordingly, a shared ASB scaffold established a direct structural link between compounds with and without target annotations, and its target profile provided the target hypotheses for molecules from ZINC. Underlying ideas included that structural analogs are likely to display similar or overlapping biological activities, and that ASB scaffolds can be used for activity assignments at a higher level of structural abstraction compared to individual analogs comprising a series. Activity assignments to scaffolds are often attempted in compound-based scaffold analysis.[6] However, ASB scaffolds are more suitable for metalevel activity assignments on the basis of analogs comprising a series than compound-based scaffolds.[5]Figure shows 3 exclusive subsets of 1216, 11 270, and 581 ASB scaffolds representing compounds from ChEMBL and ZINC, PubChem and ZINC, and all of the 3 databases, respectively. ChEMBL and PubChem compounds sharing ASB scaffolds with ZINC molecules were active against a total of 543 and 586 unique targets, respectively. Thus, despite the availability of many more PubChem than CHEMBL compounds sharing ASB scaffolds with ZINC molecules, the target coverage of these compound sets was comparable. In addition, compounds from ChEMBL and PubChem represented by the 581 ASB scaffolds shared by all of the 3 databases were active against a total of 600 targets. In total, by combining these exclusive sets, 588 and 610 targets from ChEMBL and PubChem were obtained, which could be assigned to 37 763 ZINC compounds. The distribution of these targets over different families is shown in Figure a for ChEMBL and Figure b for PubChem. Although the distributions were partly similar, there were notable differences. For example, compounds from ChEMBL covered larger proportions of kinases and proteases than PubChem compounds, whereas PubChem covered a larger proportion of targets that did not belong to major therapeutically relevant families (“other targets”). The 581 ASB scaffolds that were common to ChEMBL, PubChem, and ZINC represented the compounds active against a large number of targets, thus assigning a key role to these conserved ASB scaffolds.

Figure 3

Distribution of targets over different families. The distribution of targets over different families assigned to ZINC compounds on the basis (a) ChEMBL and (b) PubChem annotations is shown. UniProt-based target family assignments were available for 549 of 610 PubChem targets.

Promiscuity of Shared Scaffolds

On the basis of target profiles, the promiscuity of scaffolds was determined, also representing an analysis at the level of metadata (i.e., scaffold promiscuity was inferred from unique targets of compounds represented by a given ASB scaffold). Table reports the distribution of the 1216 and 11 270 ASB scaffolds and the corresponding compounds from ChEMBL and PubChem over increasing numbers of targets (i.e., increasing promiscuity levels). The majority of ASB scaffolds shared by ChEMBL and ZINC compounds (i.e., 863) were only annotated with a single target. Hence, these scaffolds were classified as nonpromiscuous. In addition, there were more than 300 scaffolds with 2–4 targets available and smaller numbers with up to 25 targets (promiscuous scaffolds). In addition, for 3091 ASB scaffolds shared by PubChem and ZINC, only a single target was available. However, in this case, there were 4494 scaffolds with 2–4, 754 with 5, and 1822 scaffolds with 6–10 targets. Furthermore, more than 1100 ASB scaffolds with 11–25 or more targets were available. Thus, ASB scaffolds shared by PubChem and ZINC compounds were overall much more promiscuous than those shared by ChEMBL and ZINC.

Table 2

Promiscuity of Shared ASB Scaffoldsa

# targets	# ASB scaffoldsb	# ChEMBL CPDs	# ZINC CPDs	# ASB scaffoldsc	# PubChem CPDs	# ZINC CPDs
1	863	1154	1851	3091	3200	7832
2	193	285	417	2065	2473	5450
3	66	90	138	1420	1903	4010
4	52	94	139	1009	1518	2891
5	14	31	18	753	1224	2144
6–10	26	49	43	1823	3563	6467
11–15	1	1	1	605	1440	2248
16–20	−	−	−	240	738	889
21–25	1	1	1	117	360	426
>25	−	−	−	147	582	605

The table reports the number of shared ASB scaffolds associated with single- and increasing multitarget activities. For each ASB scaffold, the total number of bioactive compounds (from ChEMBL and PubChem) and ZINC compounds forming the analog series is reported.

ASB scaffolds shared by ChEMBL and ZINC.

PubChem and ZINC compounds.

Target Hypotheses

For a total of 37 763 commercial ZINC molecules, target hypotheses were derived on the basis of shared ASB scaffolds, covering a total of 1051 unique targets. Figure shows the representative examples. Figure a shows a series comprising eight analogs from ChEMBL that were active against varying numbers of seven different targets. On the basis of the shared ASB scaffold, this series was combined with a corresponding series from ZINC containing six analogs, hence yielding a total of seven target hypotheses for these ZINC molecules. In addition, in Figure b, a series of six PubChem analogs is shown. Each of these compounds was active against one of three targets. Five analogs from ZINC were found to contain the same ASB scaffold, thus providing three target hypotheses for these molecules. Furthermore, Figure c shows 5 PubChem and 4 ChEMBL analogs containing 1 of the 581 conserved ASB scaffolds found in all of the 3 databases. ChEMBL and PubChem analogs were active against a total of five different targets. The conserved ASB scaffold was also present in six ZINC molecules. Thus, for these analogs from ZINC, five plausible target hypotheses were available.

Figure 4

Analog series and shared ASB scaffolds. Shown is exemplary analog series with shared ASB scaffolds for (a) ChEMBL (blue box), (b) PubChem (green), (c) ChEMBL and PubChem, and ZINC compounds (orange) in an R-group table format. Substituents (R1) in analogs are in red. For each series, the union of targets (assigned targets) from bioactive compounds is provided that can be potentially assigned to ZINC compounds containing the same scaffold. In (a) eight analogs from ChEMBL that share an ASB scaffold with six analogs from ZINC are shown that are active against a total of seven targets including protein-tyrosine phosphatase LC-PTP (T1), induced myeloid leukemia cell differentiation protein Mcl-1 (T2), carboxy-terminal domain RNA polymerase II polypeptide A small phosphatase 1 (T3), nuclear receptor subfamily 4 group A member 1 (T4), protein tyrosine kinase 2 β (T5), estrogen receptor β (T6), and apoptotic protease-activating factor 1 (T7). In (b), six analogs from PubChem that share an ASB scaffold with five analogs from ZINC are shown that are active against a total of three targets including TDP1 protein (T1), thioredoxin glutathione reductase (T2), and dopamine receptor D3 (T3). In (c), 5 PubChem, 4 ChEMBL, and 6 ZINC analogs containing 1 of the 581 conserved ASB scaffolds are shown. PubChem and ChEMBL analogs were active against a total of five unique targets including dopamine D1 receptor (T1), TDP1 protein (T2), adenosine 5′-triphosphate-dependent Clp protease proteolytic subunit (T3), serotonin 6 (5-HT6) receptor (T4), and urea transporter 1 (T5).

Because more than 37 000 biologically unannotated compounds were linked to activities against more than 1000 targets via the ASB scaffold approach, often including multiple target hypotheses ASB scaffolds and resulting compounds, several practical applications can be considered. For example, these compounds provide a pool for target-directed hit expansion (“analog-by-catalog”) and for the design or extension of focused libraries used in drug discovery. To these ends, it is often attempted to acquire new compounds from vendor sources. Taken together, the results of your analysis assign priorities to commercial compounds for many targets. Furthermore, analog series are often evaluated on closely related targets in the search for selective compounds. For this purpose, analog series can also be further extended with commercial compounds linked to activities of interest. On the other hand, ZINC compounds containing promiscuous ASB scaffolds, as described above, can also be selected as candidates for exploring multitarget activities, another current topic of interest in drug discovery research. Thus, the pool of commercial compounds with analog relationships to known bioactives should have utility for a variety of applications.

Conclusions

Our study was designed to systematically compare bioactive and commercial chemical space on the basis of analog series, establish structural links via ASB scaffolds, and develop target hypotheses for commercial molecules. To these ends, a total of more than 4.2 million carefully selected compounds were analyzed. For ASB scaffolds, target profiles were generated on the basis of active compounds they represented, and scaffolds shared by bioactive compounds were determined. A subset of 581 conserved ASB scaffolds was identified whose profiles covered 600 different targets. The scaffolds played a central role in target exploration and transfer across different databases. Shared scaffolds and their target profiles made it possible to derive target hypotheses for 37 763 compounds from ZINC having a high probability of commercial availability. These compounds represent interesting candidates for acquisition and focused library design or other applications in compound optimization, as discussed above. Therefore, as a part of our study, key ASB scaffolds and ZINC compounds with new target annotations will be made freely available on an open access platform.[20]