In search of novel ligands using a structure-based approach: a case study on the adenosine A2A receptor
- 1.6k Downloads
In this work, we present a case study to explore the challenges associated with finding novel molecules for a receptor that has been studied in depth and has a wealth of chemical information available. Specifically, we apply a previously described protocol that incorporates explicit water molecules in the ligand binding site to prospectively screen over 2.5 million drug-like and lead-like compounds from the commercially available eMolecules database in search of novel binders to the adenosine A2A receptor (A2AAR). A total of seventy-one compounds were selected for purchase and biochemical assaying based on high ligand efficiency and high novelty (Tanimoto coefficient ≤0.25 to any A2AAR tested compound). These molecules were then tested for their affinity to the adenosine A2A receptor in a radioligand binding assay. We identified two hits that fulfilled the criterion of ~50 % radioligand displacement at a concentration of 10 μM. Next we selected an additional eight novel molecules that were predicted to make a bidentate interaction with Asn2536.55, a key interacting residue in the binding pocket of the A2AAR. None of these eight molecules were found to be active. Based on these results we discuss the advantages of structure-based methods and the challenges associated with finding chemically novel molecules for well-explored targets.
KeywordsAdenosine A2A receptor Virtual screening Explicit water Novel ligand Diverse chemical space
All structural modeling was performed using tools in the Schrödinger small-molecule discovery suite. Docking was done with Glide 6.3  using the SP scoring function with default settings. We used a previously generated ensemble of A2A receptor models prepared with the Protein Preparation Wizard , each containing different individual water molecules . The eMolecules database was prepared using LigPrep with default settings. Protonation and tautomeric states were assigned using Epik [18, 19]. This resulted in a fully expanded set of ~6.6 M stereoisomeric and tautomeric states from the 2.5 M ligands. The fully expanded set of molecules was docked into the upper node of the decision tree (DT) (i.e. the one that generated the highest enrichment in our previous study). Ligands proceeded through the decision tree according to previously defined rules (Fig. 2) based on Glide docking scores. For instance in order to end up in leaf 3, compounds should have a docking score better than -8.865 kcal/mol in the first node, a docking score better than −9.030 kcal/mol in the second and better than −9.115 in the third node of the decision tree. The compounds that ended up in leaf 1, leaf 3, and leaf 5 (5378 total) were considered to be “computational hits” and were subjected to subsequent filters (Fig. 2).
First we eliminated reactive compounds using the REOS filter (as implemented in KNIME) [20, 21]. Next, we rescored all poses using the MM-GBSA method , where we used the VSGB 2.0 implicit solvent model  and the OPLS2005 force field  to estimate a binding energy. For the MM-GBSA calculations, explicit water molecules were not considered. All compounds with a MM-GBSA binding energy worse than the mean were eliminated. Since our goal was to find novel scaffolds, we included an explicit similarity filter. This filter was based on all compounds ever tested on the A2AAR for activity from ChEMBL v17 (human, rat, mouse, bovine, guinea pig, sheep, rabbit, and pig), resulting in a total of 12,205 compounds. Tanimoto similarities between all computational hits and all tested compounds were calculated based on Molprint2D  fingerprints in Canvas  and computational hits were eliminated if the similarity to any of the compounds in ChEMBL was higher than a defined threshold (Tanimoto >0.25) . Next, we constructed two different sets: A) 71 compounds from the previous steps, ranked by solvent accessible ligand efficiency (LE2/3)  and filtered iteratively (based on the rank) by similarity within the selection of compounds (Tanimoto ≤0.25), and B) 8 compounds from the previous steps, prior to the filtering and ranking within the selection of computational hits (before step A), based on both visual inspection and further filtering to ensure a bidentate interaction with Asn2536.55 in the 6th transmembrane helix of the receptor was present.
Radioligand binding assays
[3H]ZM241385 (45.9 Ci/mmol) was purchased from ARC Scopus Research (Wageningen, The Netherlands). NECA was obtained from Sigma-Aldrich (St. Louis, MO, USA). The compounds selected by the VS were bought from Enamine, Vitas M, Chembridge, Chemdiv and Key Organics. All other materials were purchased from commercial sources.
HEK293 cells stably expressing human adenosine A2A receptor were cultured and harvested as has been described previously . Single point radioligand displacement assays were performed in order to determine radioligand displacement by the purchased molecules. 2.5 nM [3H]ZM241385 was incubated for 2 h. at 25 °C in a final volume of 100 μL with 15 µg A2AAR-WT membranes, in absence or presence of 10 μM of the purchased molecules. Total binding (TB) was determined in the absence of the purchased compounds and non-specific binding (NSB) was measured in the presence of 100 µM NECA. The amount of protein added ensured that TB was less than 10 % of the total radioactivity added to prevent radioligand depletion. Incubations were terminated by rapid vacuum filtration to separate the bound from free radioligand through 96-well GF/B filter plates using a FilterMate™ Harvester (PerkinElmer Life Sciences). Filters were subsequently washed ten times with ice-cold assay buffer and left to dry for at least 15 min at 55 °C. Thereafter, 25 µL of scintillation fluid (PE Microscint 20) was added and after at least 3 h the filter-bound radioactivity was measured by scintillation spectrometry using a PE 2450 MicroBeta2 Plate Counter (PerkinElmer Life Sciences). In order to determine the total counts and thus the exact concentration of radioligand, 3.5 mL of PE Emulsifier Safe was added to 25 µL of radioligand and the radioactivity was measured with the PE Liquid Scintillation Analyzer Tri-Carb 2900TR (PerkinElmer Life Sciences). The experiments were performed two times in triplicate. The experimental data was analyzed with GraphPad Prism 5.0 (GraphPad Software Inc., San Diego, CA, USA).
Aside from these two compounds, none of the other molecules showed more than 50 % radioligand displacement (SI Table 1). To make sure our results were not caused by the over rewarding of non-native interactions by Glide, we reevaluated the binding poses of the selected 1454 compounds, and compared these with the binding mode of adenosine A2A receptor ligands for which the binding mode was known. Indeed the poses of both the two actives and top scoring inactives are arguably overlapping, with all 7 compounds interacting with Asn2536.55 (SI Figures 1 and 2, poses of the 2 actives and 5 top scoring non-actives). However, we observed that known active compounds tended to form a bidentate H-bond with amino acid Asn2536.55, a residue crucial for ligand binding, whereas most of our top scoring virtual screening hits did not. Based on the 1454 compounds, a second selection of 8 compounds was made to ensure that these compounds form this critical interaction as well. These compounds were also tested in the same displacement assay; however none of them showed radioligand displacement above 50 % (SI Table 2).
Analysis of other compounds found in the Virtual Screen
Ten most frequent occurring ChEMBL ligands when compared with the compounds identified in our screen
T c b
Langmead et al. 
Novellino et al. 
Katritch et al. 
Katritch et al. 
Sanders et al. 
Sanders et al. 
Carlsson et al. 
Poulsen et al. 
van Muijlwijk-Koezen et al. 
Langmead et al. 
Also apparent in Table 2 is that certain substructures appear frequently; for example the succinimide moiety is found in 3 of the 10 entries shown. When we analyzed how frequently this substructure occurs in the raw results (the 5378 compounds prior to filtering) we found that it appeared 411 times (7.6 %), possibly because it is often used as a protective group in synthesis. Although in the docking the oxygen of the succinimide was often found to make an interaction with the backbone of Phe168EL2 (e.g. SI Figure 2), no evidence was found that it was a beneficial group. First of all, more than 60 % of the computational hits (prior to filtering) containing a succinimide were found to have an MM-GBSA score worse than mean (254 of the 411 of compounds). Moreover, when we queried compounds tested on the adenosine A2A receptor (ChEMBL v20) for this substructure mostly inactive compounds were found (37 compounds with pChEMBL < 5, compared to 0 highly active compounds (pChEMBL > 7), and 16 compounds with a moderate activity (pChEMBL value between 5 and 7)). To put this in context, the 1,3,5-triazine moiety found in reference antagonist ZM-241385 is only present in 11 of our potential hit compounds (0.2 %), versus 166 highly active ligands in ChEMBL. When we further analyzed the source of these compounds we found that, of the top 10 most frequently occurring ChEMBL compounds, 7 originated from computational studies. Additionally, although ChEMBL2151129 (Table 2, row 7) described by Carlsson et al.  was synthesized, the original scaffold was discovered using a structure-based VS . These results indicate that of the not-so-novel compounds discarded in our screen, many represent compounds found by computational methods that have been experimentally confirmed as A2AAR actives.
Retrospective analysis of the adenosine A2A receptor Ligand Discovery
The challenging nature of finding non-similar molecules in a virtual screen, for a thoroughly studied target, resulted in the identification of two hit compounds in this study. Therefore we compared our similarity threshold with the similarity of compounds tested on the adenosine A2A receptor. All compounds tested for A2A receptor binding in ChEMBL were sorted by year of publication, and a global maximum similarity to earlier reported compounds was calculated to categorize whether bioactive compounds were novel at the time of reporting.
Overall, the distribution of compounds across the different novelty bins seems to be skewed toward similarity over novelty, with a mean similarity of 0.6 and around 68 % (44.3 + 23.6 %, percentages shown in black, Fig. 3) of the compounds falling in the two bins above 0.5 similarity. The lowest similarity bin was the least populated, in which only 93 out of the 4501 active ligands were present (2.1 %). This is also demonstrated by the fact that in the majority of years, the bar charts do not cross the ≤0.25 similarity line. Some of the novel scaffolds that have been reported include the pyrazolo [4,3-e] 1,2,4-triazolo [1,5-c]pyrimidines in 1994, 1,2,4-triazolo [4,3-a]quinoxalin-1-ones in 2000, and 2-thioxothiazole derivatives in 2013 [1, 34, 35] (Fig. 3).
The 93 active ligands in the lowest similarity bin constitute, using the current definition, the total number of unique scaffolds that have been discovered over the years. Next we compared the number of actives as a fraction of the total number of reported compounds, across the different similarity bins (percentages shown in green). This confirms the expected result that the less similar a newly tested compound is, the lower the chance is that it will be active. However for active compounds, we did not find a significant difference between active (green, pChEMBL >7) and less highly active (yellow, pChEMBL 5–7) compounds, although this seems to vary strongly between years. Nevertheless, in the lowest similarity bin less than 1 in 5 compounds (18.5 %) was found to be active, thus demonstrating the difficulty in finding molecules that are both active (18.5 %) and novel (2.1 %) for a well-studied target with many actives.
Retrospective analysis of Virtual Screens
Because at least five structure-based VS have been conducted on the adenosine A2A receptor prior to the VS reported here, we assumed that similar trends (as in Fig. 3) could be observed for these earlier screens as well [6, 7, 8, 9, 10]. To test this hypothesis we retrieved all the compounds from the five known screens that were reportedly bought and tested (unfortunately, not all the papers reported full datasets of all tested compounds). The different screens yielded a combined total of 54 active compounds. When we compared these active molecules with the reported inactive compounds from those studies we learned that, similar to the retrospective analysis shown in Fig. 3, there is a trend toward lower similarity amongst the inactive compounds. For instance, almost 75 % of the reported inactives had a similarity ≤0.39 while only 55 % of the active compounds were found in these bins. There is a higher fraction of low similarity compounds reported in the five VS than in the retrospective data, e.g. more than 75 % of the actives from these VS had a similarity below 0.50 (Fig. 4) compared with 32 % for the retrospective data (Fig. 3; 30 + 2.1 % respectively). Nevertheless none of the actives found with VS were below the 0.25 similarity threshold, and using our current definition they cannot be considered novel scaffolds. The three most novel compounds originated from the VS by Heptares  where the authors explicitly selected non-similar molecules (hit 3, hit 4, and hit 9 in their publication) .
In this study we used a previously established structure-based virtual screening protocol  that incorporates explicit water molecules to find novel ligands for the adenosine A2A receptor. A more practical scenario would involve pre-filtering based on 2D-similarity and reactivity (REOS), increasing efficiency of the screen; however docking of all compounds allowed us to compare the non-similar/novel computational hits with the full spectrum of docked compounds. We defined ‘novel’ as compounds with a Tanimoto similarity smaller than or equal to 0.25 compared to any known compound tested for activity on the adenosine A2A receptor. Our hit rate was much lower (1.4 %) than in previously reported screens and we determined that the low hit rate could be most likely attributed to an overly stringent requirement to find novel compounds for a target that has a wealth of chemical matter. In retrospect a higher cutoff around the mean similarity (0.40) of adenosine A2AAR compounds (see “Retrospective analysis of the adenosine A2A receptor Ligand Discovery”) might have been more suitable. Because of the use of different fingerprints, similarity thresholds in other VS are only comparable qualitatively. Nonetheless, we can draw some conclusions about the results obtained here when compared with those from previous screens. In the study by Katritch et al. , who also included explicit water molecules in their VS, a hit rate of 41 % (23/56) at a test concentration of 10 µM was observed. Hits were compared with molecules in GLIDA , and their calculated similarities ranged from 0.44 to 0.68 . Another study by Carlsson et al. where modified partial charges were used on Asn2536.55 to favor hydrogen bonding to this residue yielded a comparable hit rate of 35 % (7/20) at a test concentration of 20 µM . Moreover, the authors employed filters such as molecular weight, formal charges, and the Similarity Ensemble Approach (SEA) to further narrow down their hit list. Compounds emerging from this screening approach were compared with adenosine receptor ligands from ChEMBL and Wombat, revealing similarities between 0.30 and 0.52. A subsequent study by Langmead et al.  was performed on a homology model of the adenosine A2A receptor yielding a hit rate of 9 % (20/230) at a test concentration of 55 µM. Compounds were filtered for CNS-like properties and ligands containing xanthine or furan moieties were removed. Subsequent hits were subjected to a biophysical mapping approach . The 10 hits that were reported all had relatively low similarities to known A2AAR antagonists, ranging between 0.19 and 0.31. However the authors noted that their most potent scaffold contained the same moiety as found in the previous two virtual screens described above [6, 7]. Another recent screen focused on finding fragments for the adenosine A2A receptor  resulted in a hit rate of 64 % (14/22) at a test concentration of 500 µM, while similarities to known compounds ranged between 0.28 and 0.41. Lastly, a VS was performed by Rodriguez et al. to identify agonists for the adenosine A2A receptor . Although none of the hits turned out to be agonists, a hit rate of 45 % (9/20) was achieved at a test concentration of 10 µM, with similarities ranging from 0.30 to 0.61.
In this study we attempted to determine the ability of docking-based VS to identify novel compounds. We explicitly selected compounds that bore no resemblance to any compound previously tested against the adenosine A2A receptor. Indeed, the adenosine A2A receptor is one of the best studied targets of all class A GPCRs, and the number of bioactivities (10,184) in ChEMBL (version 20) ranks it at number 7 of the 688 class A GPCRs . This represents the result of decades of screening and medicinal chemistry efforts (see also Figs. 1, 3) . In our screen we included not only active but also inactive compounds as similarity filters. This was done primarily to exclude inactive molecules identified in previous screens, but also to explore actual uncharted chemical space. In some VS [5, 8] not all the tested compounds, i.e. including inactives, have been disclosed. This arguably leads to a loss of useful information, in the form of validated ‘true negatives’ that can be used as decoys in a benchmark screen or to bias future screening efforts away from explored chemical space. As a consequence of this data incompleteness we may have tested compounds that had been found inactive in previous studies as well. Including this information in future VS reports would be helpful to prevent future screens from exploring the same (inactive) chemical space.
Despite the differences in data completeness, we attempted to compare previous VS results with our current similarity-filtered method. We found that ligands identified in a VS have a higher novelty than the ligands for the adenosine A2A receptor (Figs. 3, 4) in general. Inactives have an even higher novelty, which is the consequence of a tradeoff between similarity and activity. In fact it is a well-established paradigm that activity correlates with similarity [23, 24, 39]. Indeed, this is the basis for 2D virtual screening. Therefore depending on the goal of VS, and the extent of the mapped chemical space, one may wish to filter out molecules with a novelty above a certain threshold (e.g. ≥0.75) as the chances of success become increasingly small as similarity to known actives decreases. On this basis, virtual screening efforts often include a step of visual inspection, where ligands that interact in non-native manner (e.g. no hydrogen bond with key residue) are filtered out . Selecting compounds across a range of similarity, e.g. between 0.20 and 0.50 provides a good balance between potential activity and novelty (Fig. 3). Another possibility would be to test compounds that have a higher novelty at a higher concentration. In this study we considered hits as displaying approx. 50 % or higher radioligand displacement at 10 µM (equivalent to pChEMBL >5), but in previously performed VS concentrations between 10 and 500 µM were used as a test concentration. Additionally the number of selected compounds based on all the filters (Fig. 2) is on the low side due to the relatively high docking score cutoffs in the decision tree. These scores were determined based on known and potent ligands , and in hindsight these cutoffs could have been adjusted to select hits from a larger pool of compounds. Indeed, it is crucial to select ligands that end up in multiple leaves to capture more chemical space . Here we selected compounds that satisfied the docking score cutoff in at least two out of the three nodes (i.e. compounds ending up in leaf 1, 3, and 5, Fig. 2a). This could easily be extended to another ensemble strategy such as the “Z2 score”, which combines Z-scores from multiple screens and has been proven to outperform other ensemble methods .
Although some of the virtual screens (by us and others) have expanded A2A receptor chemical space with new chemotypes, they all draw from essentially the same pool of commercially available compounds (Table 2). In addition it has been shown that available chemical libraries only cover a limited part of the chemical space [42, 43]. Future studies using unbiased libraries that cover larger chemical spaces such as GDB  could potentially suggest new chemotypes for this important target.
Structure-based virtual screening is a widely used approach to identify hits for drug targets, and has been applied previously to the adenosine A2A receptor. We performed a virtual screen on this target while excluding all compounds similar to known actives and inactives in order to explore unchartered chemical space for this GPCR. Of the 79 compounds we tested, only 2 exhibited affinity for the target. We found several explanations for this relatively low hit rate: the overly stringent novelty threshold, the abundance of known ligands for the adenosine A2A receptor, and the overlap in results with previous virtual screening efforts. Finding novel compounds for a particular target of interest becomes increasingly difficult as the target is more extensively studied, as was the case with the adenosine A2A receptor studied here. Nonetheless, structure-based methods still seem to offer the greatest possibility to find novel actives for targets that have been well explored, since ligand-based methods are generally designed to find compounds similar to known actives. As such, researchers should choose their virtual screening method for a project based on their objectives and the known chemical matter for the target. Structure-based methods may not be necessary when sufficiently interesting compounds can be found with simpler/faster methods. However, when minimal biases from existing chemical matter is sought, structure-based methods will likely be preferred, even though there is still substantial room for these methods to improve based on treating the underlying physics of the system more accurately (protein flexibility, explicit waters, variable ionization states, etc.). With improvements of structure-based virtual screening methods coupled with rapidly increased computational capabilities to handle the complexities of sampling we expect to see the value of these methods continue to increase in the coming years.
Displacement results of the non-active molecules, docking poses of the two actives and five top scoring inactives, HPLC and 1H NMR spectra of the active molecules.
Adriaan P. IJzerman and Eelke B. Lenselink thank the Dutch Research Council (NWO) for financial support (NWO-TOP #714.011.001).
Eelke B. Lenselink and Thijs Beuming performed the calculations. Corine van Veen and Arnault Massink performed the biological experiments. Eelke B. Lenselink wrote the manuscript. All authors discussed, contributed to and revised the manuscript.
- 20.Berthold MR, Cebron N, Dill F, Gabriel TR, Kötter T, Meinl T, Ohl P, Sieb C, Thiel K, Wiswedel B (2008) KNIME: the Konstanz information miner. Springer, BerlinGoogle Scholar
- 22.Schrödinger Release 2013-3: Prime v, Schrödinger, LLC, New York, NY (2013)Google Scholar
- 25.Bender A, Mussa HY, Glen RC, Reiling S (2004) Similarity searching of chemical databases using atom environment descriptors (MOLPRINT 2D): evaluation of performance. J Chem Inf Comput Sci 44:1708–1718Google Scholar
- 36.Okuno Y, Tamon A, Yabuuchi H, Niijima S, Minowa Y, Tonomura K, Kunimoto R, Feng C (2008) Nucleic Acids Res 36(suppl 1):D907Google Scholar
- 40.Irwin JJ, Shoichet BK (2016) J Med Chem 12;59(9):4103Google Scholar
Open AccessThis article is distributed under the terms of the Creative Commons Attribution 4.0 International License (http://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution, and reproduction in any medium, provided you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license, and indicate if changes were made.