Computer-aided design of multi-target ligands at A1R, A2AR and PDE10A, key proteins in neurodegenerative diseases

Compounds designed to display polypharmacology may have utility in treating complex diseases, where activity at multiple targets is required to produce a clinical effect. In particular, suitable compounds may be useful in treating neurodegenerative diseases by promoting neuronal survival in a synergistic manner via their multi-target activity at the adenosine A1 and A2A receptors (A1R and A2AR) and phosphodiesterase 10A (PDE10A), which modulate intracellular cAMP levels. Hence, in this work we describe a computational method for the design of synthetically feasible ligands that bind to A1 and A2A receptors and inhibit phosphodiesterase 10A (PDE10A), involving a retrosynthetic approach employing in silico target prediction and docking, which may be generally applicable to multi-target compound design at several target classes. This approach has identified 2-aminopyridine-3-carbonitriles as the first multi-target ligands at A1R, A2AR and PDE10A, by showing agreement between the ligand and structure based predictions at these targets. The series were synthesized via an efficient one-pot scheme and validated pharmacologically as A1R/A2AR–PDE10A ligands, with IC50 values of 2.4–10.0 μM at PDE10A and Ki values of 34–294 nM at A1R and/or A2AR. Furthermore, selectivity profiling of the synthesized 2-amino-pyridin-3-carbonitriles against other subtypes of both protein families showed that the multi-target ligand 8 exhibited a minimum of twofold selectivity over all tested off-targets. In addition, both compounds 8 and 16 exhibited the desired multi-target profile, which could be considered for further functional efficacy assessment, analog modification for the improvement of selectivity towards A1R, A2AR and PDE10A collectively, and evaluation of their potential synergy in modulating cAMP levels. Electronic supplementary material The online version of this article (10.1186/s13321-017-0249-4) contains supplementary material, which is available to authorized users.


Background
Neurodegeneration involves the progressive loss of the structure and function of neurons, which is common in Parkinson's, Huntington's disease and schizophrenia [1]. Recently, there has been substantial interest in the search for alternative non-dopamine (non-DA) based approaches for the treatment of neurodegenerative diseases, as the classical DA-based approaches have long been associated with many undesirable side effects such as dyskinesia, hallucinations, and on/off effects [2]. Given that the adenosine neuromodulation system (via the adenosine A 1 and A 2A receptors) has been identified as a key target for the management of neurodegenerative diseases, this qualifies its targeting as a potential promising non-DA based treatment approach [3,4]. Indeed, modulation of cAMP levels has proven to have benefits in neuronal survival in an adenosine receptor-dependent manner [5]. In addition, recent findings suggest that phosphodiesterase 10A (PDE10A) also plays a role in neurodegenerative diseases such as Parkinson's, Huntington's disease, and schizophrenia [6][7][8]. Inhibition of PDE10A resulting in maintenance of elevated intracellular cAMP concentrations, has been suggested to be effective in the treatment of these diseases. Thus multi-target ligands that bind to different adenosine receptors subtypes (A 1 and A 2A receptors) while simultaneously inhibit PDE10A might be synergistic in modulating cAMP levels, which is of therapeutic potential for neurodegenerative diseases [9][10][11].
Conceptually, multi-target drugs work by creating a combination effect on multiple targets in the biological network simultaneously, which may (through e.g. synergistic effects) decrease the therapeutic dose required, thus increasing therapeutic efficacy, preventing drug resistance, and reducing target-related adverse effects [12][13][14]. Also, another advantage of multi-target drugs over other types of treatments such as combination therapies, is a reduced likelihood of drug-drug interactions [15,16].
However, it remains a challenging task for medicinal chemists to design drugs with a specific multi-target profile and to achieve selectivity for specific targets over offtarget effects with suitable pharmacokinetic properties [17,18]. In fact, the field of multi-target drug design has recently become an active field of research in the pharmaceutical industry, where around 20 designed multitarget drugs have either reached advanced development stages or are already approved [14,19,20].
In particular, for Central Nervous System (CNS) diseases, there has been growing interest in exploiting the multi-target profiles of existing compounds to investigate their potential applicability as drugs. For example, multitarget profiles of drugs and drug candidates affecting the dopaminergic system have been investigated. Examples include Aripiprazole, Amitriptyline, Chlorpromazine, and Clozapine [21]. In addition, various multi-target based virtual screening protocols for multi-target drug design have been developed [13,[22][23][24]. Examples of ligandbased protocols include in silico target prediction and Chemogenomic and pharmacophore-based approaches, which resulted in the discovery of CNS drugs with multitarget combinations such as MAO-A/MAO-B/AChE/ BuChE, AChE/BuChE, and H 3 -R/HMT/AChE/BuChE [21][22][23][24]. Structure-based approaches such as docking and molecular dynamics calculations have also been employed for the discovery of new multi-target ligands such as BuChE inhibitors/hCB2R and MAO-A/MAO-B/AChE/ BuChE ligands to treat neurodegenerative diseases [25].
In this work, we offer a computational strategy for designing synthetically feasible ligands that bind to A 1 R and A 2A R, and inhibit PDE10A-a novel multi-target combination of G protein-coupled receptors (GPCRs) and an enzyme, which has not, to our knowledge, been previously exploited. The designed ligands with this multi-target combination are intended as starting points for future development of multi-target drugs treating neurodegenerative diseases. It should be noted here that in the current study we only consider affinity of ligands to the above receptors, which we also experimentally validate as outlined below. However, for therapeutically relevant purposes also functional effects and optimization of selectivity towards A 1 R, A 2A R and PDE10A need to be considered, which will be the area of a future study.
The workflow of the current study is shown in Fig. 1. Starting with a focused chemical space consisting of known actives against A 1 R, A 2A R and PDE10A, new synthetically feasible compounds were established via RECAP (Retrosynthetic Combinatorial Analysis Procedure) [26,27], which fragments molecules at pre-defined bonds and recombines them in a combinatorial manner, and were then evaluated in silico, using target prediction and ligand/protein docking. Compounds with favorable assessments in both steps were carried forward for substructural analysis. This analysis identified compound series with the highest frequency of prediction as multitarget ligands against the desired set of targets, which is of advantage from the practical side, given their synthetic accessibility via a common synthetic route.
A series of 2-aminopyridine-3-carbonitriles were selected for prospective validation of the pipeline, a series which was synthetically accessible via a one pot synthetic scheme i.e. providing products with the desired properties: cost-effective, synthetically efficient and available in a timely fashion [28,29].
Subsequently the synthesized compounds were experimentally tested and confirmed as A 1 R/A 2A R-PDE10A multi-target ligands. Selectivity against other subtypes of both protein families confirmed the pharmacological profile of the compound series, and structure activity relationships (SAR) were also deduced. Hence, in this work we report a successful computational strategy, which allowed the discovery of the first A 1 R/ A 2A R-PDE10A multi-target ligands. The novel A 1 R/ A 2A R-PDE10A ligands are sought to display a combination effect in modulating the A 1 R, A 2A R, and PDE10A targets simultaneously similar to that of combination compounds of Adenosine receptors and PDEs, reported by Rickles et al., which were synergistic in modulating cAMP levels [10]. less than or equal to 1 µM revealed that the most frequently occurring common heterocycles among the actives against the three target classes were pyridine, pyrimidine, piperazine, and 1H-pyrazole (Additional file 1: Figure S1). Subsequently, A 1 R (2104), A 2A R (2489) and PDE10A inhibitors (679) containing those frequent heterocycles were subjected to RECAP analysis/synthesis in MOE (see Methods for details) [26]. As a result, 458,839 (potentially) synthetically accessible ligands were formed in silico. This list of candidates was filtered to those retaining the common heterocycles (listed above), in order to create a focused chemical space characteristic Fig. 1 The computational strategy for rational design of A 1 R/A 2A R-PDE10A multi-target ligands started with a focused chemical space consisting of known actives of A 1 R, A 2A R and PDE10A, and formed new synthetically feasible compounds which were subjected to target prediction and docking for synthesis and pharmacological evaluation of A 1 R, A 2A R and PDE10A (with the simultaneous tradeoff of reduced novelty), giving rise to 22,233 compounds.

Target prediction of the designed RECAP library
To assess the likelihood of active compounds against A 1 R, A 2A R and PDE10A, PIDGIN 1.0 (Prediction including Inactivity), a tool which uses ECFP 4 circular Morgan fingerprints and trained on ChEMBL actives and PubChem inactives, was used to perform in silico target prediction for the focused RECAP library (22,233 compounds) [24]. Subsequent enrichment analysis of the predictions was done using an estimation score, average ratio as developed by Liggi et al. [31] and via Chi square test [32]. For targets to be considered as enriched according to these methods, the estimation score and the Chi square test p value should be less than or equal to 0.01 and 0.05, respectively. Hence, upon analyzing the enrichment parameters for the A 1 R, A 2A R and PDE10A targets that were predicted for the focused RECAP library (Additional file 1: Figure S2), the three targets were predicted with an estimation score equal to 0 (enriched) as well as average ratios less than 0.1 (enriched) with Chi squared p values < 0.005. The percentage of RECAP compounds of the focused library that were predicted as actives against the A 1 R, A 2A R and PDE10A targets were 51.1, 52.8, and 24.5% respectively. These numbers are relatively high, which however is understandable given that the input to the RECAP analysis consisted of experimentally established known ligands of the above protein targets.

Docking of the compounds predicted as A 1 R/A 2A R-PDE10A multi-target ligands
In the next step docking and further substructure analysis were performed on compounds of the focused RECAP library, which were predicted as A 1 R/A 2A R-PDE10A multi-target ligands from the ligand-based side in the previous step. 2563 compounds were predicted as actives against the three desired targets, and they were subsequently docked against a high resolution (1.8 Å) A 2A R protein crystal structure (PDB ID: 4EIY) [33] its corresponding A 1 R homology model (see Methods for details), and PDE10A (PDB ID: 4DDL) [34].
Compounds which were carried forward to substructural analysis were selected when their docking score gave a value less than a pre-determined cut-off value computed from the docking scores. This cut-off value was evaluated as the docking score with the best F measure statistic obtained by docking a set of known actives and inactives against the protein crystal structures and the homology model (see Methods for details).
As a result, a distribution of RECAP compounds that were favorable as multi-target ligands by target prediction and docking was obtained, where 62.47% of the RECAP compounds that were predicted as A 1 R/ A 2A R-PDE10A multi-target ligands and docked against PDE10A exhibited docking scores lower than − 6.49 (the threshold of the best F measure discriminating between actives and inactives for known ligands). Out of the RECAP compounds which displayed docking scores lower than − 6.49 against PDE10A, 48.89 and 35.23% displayed docking scores lower than − 7.26 and − 8.49 against A 1 R and A 2A R (the thresholds of the best F measures).

Substructure analysis of the compounds predicted as A 1 R/ A 2A R-PDE10A multi-target ligands
Substructure analysis was performed on compounds having a favorable assessment by target prediction and docking (i.e. those compounds whose docking scores were below the threshold for all three targets). The analysis revealed frequently occurring series, which shared the same core structure and which are shown in Fig. 2.

Synthesis of novel 2-aminopyridine-3-carbonitriles
Due to both ease of the reaction and anticipated yield, a one-pot synthetic scheme was selected for synthesizing one promising series, 2-aminopyridine-3-carbonitriles. The design resulted in 25 compounds for synthesis of which 21 were novel compounds and four (1, 2, 5, and 17) have previously been reported in the literature [35][36][37][38]. Compounds 1-25 were screened against PAINs (PAN Assay Interference Compounds) [39] using FAFDrug3 [40], and none of the compounds exihibited potential PAINs liability. Subsequently, their synthesis was performed as shown in Scheme 1, and all products were obtained with good yields, ranging from 46 to 85% (see Methods for details).

Pharmacological evaluation of novel 2-aminopyridine-3-carbonitriles
Bioactivity testing was performed using A 1 and A 2A human adenosine receptors expressed in transfected CHO (A 1 ) and HeLa (A 2A ) cells, as well as AD293 cells that were transiently transfected with human PDE10A. Table 1 includes the list of synthesized 4,6-substituted 2-amino-pyridin-3-carbonitriles, along with their K i values against A 1 R, A 2A R, and IC 50 values against PDE10A. It can be seen that 15 compounds of the 25 synthesized 2-amino-pyridin-3-carbonitriles exhibited inhibitory activity against PDE10A below 10 μM. In addition, 13 compounds were adenosine receptor binders exhibiting selectivity towards A 1 R and A 2A R, which has not been the case in the previous work reported by Mantri et al., where 2-amino-pyridin-3-carbonitriles were promiscuous towards the four adenosine receptor subtypes [36].
Given that the objective of this work is to find compounds displaying specific multi-target activity, compounds 8, 16 It was previously reported that substituted pyridines exhibited PDE inhibitory activity [41,42], and 2-aminopyridin-3-carbonitriles are adenosine receptor ligands [36]. In this study we have now identified suitable compounds matching both criteria as A 1 R/A 2A R-PDE10A multi-target ligands, satisfying the original compound design objective.

(SAR) structure-activity relationship analysis
The purpose of the SAR analysis was to rationalize the variation in activity of the newly discovered A 1 R/A 2A R-PDE10A multi-target ligands against PDE10A, given that 2-amino-pyridin-3-carbonitriles have been discovered as a novel class of PDE10A inhibitors. Also due to the fact that compounds of this substructural class were documented as adenosine receptor ligands [36], computational SAR studies were focused on the PDE10A data, where the variation in potency was rationalized in relation to the physicochemical properties of the compounds (which were computed by FAFDrug3, Additional file 1: Table S1) [40].
A trend observed repeatedly in several cases was that when logP decreased, associated with an increase in tPSA, then this led to an improvement in the activity against PDE10A. Initial analysis concentrated on compounds 1-4, which have a phenyl substituent at position 4 of the pyridine ring. Compound 3 was the most potent PDE10A inhibitor with an IC 50   Hence, it could be deduced that in the majority of the series considered, where the substituents on a single position is varied, a decrease in computed lipophilicity associated with an increase in polarity generally improved the activity of compounds against PDE10A. This general trend can be attributed to the hydrophilic nature of the pocket, which favours the interactions between the ligand and the PDE10A protein by compounds exhibiting these properties.
where IC 50 is the concentration of compound that displaces the binding of the radioligand by 50%, C is the concentration of radioligand, and K D is the dissociation constant of each radioligand

Compound selectivity assessment
The selectivity of compounds 1-25 against the selected major off-targets A 2B R, A 3 R, PDE7A, PDE7B, and PDE9A, was predicted using PIDGIN at a threshold for binding greater than or equal to 0.8, and subsequently tested experimentally. It is noted here that the IC 50 values were determined for compounds with % inhibition at phosphodiesterases greater than 70%. As shown in Additional file 1: In general, the experimental results on off-target prediction for the synthesised 4,6-substituted 2-amino-pyridin-3-carbonitriles 1-25 agree with the predictions generated using PIDGIN utilised to bias the compound design towards selective compounds such as 8 (Additional file 1: Table S2). This compound would serve as a good starting point for analog modification to improve the selectivity of the synthesized ligands towards PDE10A.
It can be seen that compounds 8 and 25, with IC 50 values of 2.4 and 5.1 µM respectively, share similarities in predicted binding modes, since their pyridine rings display π-stacking with Phe 686 and Phe 719 of PDE10A (Fig. 3). These are the type of interactions predicted to be exhibited by the majority of the synthesized ligands from this work, as well as the only existing interactions between co-crystallised PDE10A inhibitors discovered by fragment screening (PDB ID: 5C2E, 5C1W, 5C29, 5C2A ligands with K i values of 2, 8, 700, 880, and 4.8 nM, respectively) [43]. It is noted that the ligand of 5C2A exhibits a considerable selectivity towards PDE10A over all the other PDEs (in the range of 100-1000 fold and greater over the majority of PDEs, with the least selectivity observed being in the range of 25-100 fold). This ligand exhibits only π-stacking interactions with Phe 686 and Phe 719 , similar to the mode of interactions of compound 8 with PDE10A, which is relatively selective over all tested PDEs, with the lowest selectivity being measured for PDE7B (of 55% inhibition at 10 µM ligand  [33,34] and the interactions with the A 1 R homology model reported in the literature [51,52] concentration) and compound 25, which is selective against all tested PDEs except PDE7B (Table 1 and Additional file 1: Table S2). Additional interactions were seen in analogs discovered by fragment screening, namely hydrogen bonding with Gln 716 and Tyr 683 in the PDE10A selectivity pocket (PDB ID: 5C28 and 5C2H with K i values of 2200 and 0.0082 nM respectively). [43] The ligand of 5C2H exhibits π-stacking with Phe 686 and Phe 719 and hydrogen bonding with Tyr 683 in the PDE10A selectivity pocket. The 5C2H ligand showed a very high selectivity towards PDE10A, greater than 5000 fold, which emphasizes the consideration of compound 8 for analog modification to target the selectivity pocket in order to improve the folds of selectivity towards PDE10A. In addition, hydrogen bonding with Tyr 683 in the PDE10A selectivity pocket is also seen in many other highly selective PDE10A inhibitors reported in the literature [44] (PDB ID: 5DH5, [45] 5B4L, [46] with K i = 0.23 nM, and IC 50 = 0.76 nM respectively), which further highlights the importance of analog modification to target the PDE10A selectivity pocket.
Moreover, it is noted that compounds 16 and 21 with IC 50 values of 3.2 and 10.0 µM respectively (which are selective against all tested PDEs except PDE7A, Table 1 and Additional file 1: Table S2) were predicted to exhibit an additional type of interaction, H-bonding with Gln 716 via their overlaid furan rings at position 6 of the pyridine ring (Additional file 1: Figure S3). In fact H-bonding with Gln 716 was the only interaction, besides π-stacking with Phe 686 and Phe 719 , which has been observed in many of the highly selective PDE10A ligands reported in the literature (PDB ID: 4DDL, [34] 3SN7, 3SNL, and 3SNI, [47] 5DH4 and 5DH5, [45] with IC 50 values of 4.9, 0.7, 0.7, 11 nM and K i = 0.23 nM respectively). As for other type of interactions generally exhibited by known PDE10A inhibitors such as hydrogen bonding with Gln 726 and π-stacking with Phe 729 (PDB ID: 5EDE) [48], none has been predicted for any of the compounds presented in this work.
Common predicted binding modes can also be observed for the synthesized compounds against the adenosine receptors A 2A R and A 1 R. Figure 3 displays the interactions of two representative compounds 18 and 25, which exhibit K i values of 948 and 55 nM respectively, and these are H-bonding of their pyridine rings with Asn 253 and π-stacking of their amino and carbonitrile groups with Phe 168 of A 2A R. As for A 1 R, the overlaid compounds 8 and 18, with K i values of 294 and 78 nM respectively, H-bond via their amino and carbonitrile groups with Asn 254 , and their pyridine rings are π-stacked against Phe 171 . It can be observed that the ligand/protein interactions predicted for the active compounds against the A 2A R are also those seen in the co-crystallised ligand/ protein crystal structures (PDB ID: 4EIY, [33] 3EML, [49] 5IU4, [50] with a K i value of 0.8 nM for ZM241385, which is the common ligand for the three PDB IDs). Similar was the case for the reported interactions with the A 1 R homology model in the literature (with IC 50 values of 2.9 and 6.2 nM for the reported ligands predicted to bind to the homology model of A 1 R) [51,52].
Generally the compounds exhibited good selectivity towards A 1 R and A 2A R (Table 1 and Additional file 1:  Table S2) with a nanomolar range of binding affinities. As for the selectivity towards PDE10A, it could be improved by analog modification of compound 8, which favors the hydrogen bonding with Tyr 683 in the PDE10A selectivity pocket. In addition, the potency of compounds against PDE10A could be optimized in itself, in order to achieve therapeutically relevant efficacy.

Computational assessment of CNS permeability
Compounds 8 and 16 exhibited the desired multi-target profile by inhibiting PDE10A and binding to A 2A R and/ or A 1 R. The physicochemical properties of these compounds were calculated by FAFDrug3 [40], and both compounds passed the Lipinski rule of 5 and the CNS filter, which takes into consideration the assessment of their ability to pass the blood brain barrier (Additional file 1: Figure S4) [53]. Hence, while further experimental work would be needed to establish the validity of those predictions, compounds 8 and 16 may serve as good starting points for further functional efficacy assessment and selectivity optimization towards PDE10A, A 2A R and/ or A 1 R for the subsequent consideration of multi-target drug development for the treatment of neurodegenerative diseases.

Conclusions
Here we report a successful computational strategy for designing the first A 1 R/A 2A R-PDE10A multi-target ligands as a therapeutic prospect for neurodegenerative diseases. A retrosynthetic approach was employed using MOE/RECAP, followed by target prediction and docking of the resulting library against the desired targets. We have identified 2-aminopyridine-3-carbonitriles as a series that showed agreement between both the ligandand structure-based predictions of activity against A 1 R, A 2A R and PDE10A. The synthesis of this series via a onepot synthetic scheme was pursued experimentally. As a result, compounds 8, 16, 21, and 25 were validated as A 1 R/A 2A R-PDE10A multi-target ligands with IC 50 values of 2.4, 3.2, 10.0, and 5.1 µM against PDE10A, and binding to A 1 R with K i values of 294 and 34 nM (8 and 16 respectively), and to A 2A R with K i values of 41, 95, and 55 nM (16, 21, and 25 respectively). Furthermore, selectivity profiling of the synthesized 4,6-substituted 2-amino-pyridin-3-carbonitriles against other subtypes of both protein families showed that the multi-target ligand 8 exhibited a minimum of twofold selectivity over all tested off-targets. In addition, compounds 8 and 16 exhibited the desired multi-target profile against A 1 R, A 2A R and PDE10A, which would serve as good starting points for further functional efficacy assessment and analog modification for the improvement of selectivity. In particular, this comprises investigating the signal transduction profiles of these compounds using techniques some of the authors have described before [51], as well as evaluating functional effects in cAMP assays to determine if these compounds do provide synergistic elevations in intracellular cAMP. One specific functional profile that would be of high interest and which is likely to elevate cAMP levels synergistically via the combination effect on multiple targets simultaneously, is the A 1 R antagonist/A 2A R agonist, and PDE10A inhibitor.
In summary we have investigated a computational approach for the design of multi-target ligands that was validated experimentally via synthesis and pharmacological evaluation of 2-aminopyridine-3-carbonitriles as A 1 R/A 2A R-PDE10A ligands. This approach is generally applicable to a wide range of multi-target ligand design problems, across disease areas and target families.

Selecting reference molecules for the design of multi-target ligands
Using SQL (script provided in Additional file 1), human A 1 R (2860), A 2A R (3566) ligands and PDE10A inhibitors (843) were extracted from the ChEMBL 20 database with K i and IC 50 values less than or equal to 1 μM respectively, and confidence scores of 8 or 9 [30]. Following extraction, the most frequent and common heterocycles between A 1 , A 2A receptor ligands and PDE10A inhibitors were found by performing substructure analysis on each structure using the "Chemistry-> Analyze scaffolds" function in DataWarrior 4.2.2 [54]. Analysis of A 1 R, A 2A R ligands and PDE10A inhibitors identified common and frequent heterocycles (pyridine, 1H-pyrazole, pyrimidine and 9H-purine for A 1 R and A 2A R), and these were extracted from each set using RDKit, 9.1, Python [55]. It should be noted that compounds containing 9H-purine were also extracted from the original set even though this substructure is characteristic of A 1 R and A 2A R only, since it is structurally similar to the common and frequent heterocycles identified (pyridine, 1H-pyrazole, and pyrimidine). Additional file 1: Figure S1 shows the most frequent heterocycles for the A 1 R, A 2A R ligands, and PDE10A inhibitors and their relative frequencies in each set. It was found that they are furan, pyridine, xanthine, 1H-pyrazole, pyrimidine, piperazine, and 9H-purine. All of these heterocycles ranked among the top 30 for A 1 R, A 2A R ligands and PDE10A inhibitors. This indicated their suitability for designing multi-target ligands at these protein targets, given the overlap in chemical (heterocyclic) space. In the case where no percentage is displayed for a particular target, this means that the heterocycle does not appear among the top 30 for the set of compounds involved.

Designing new multi-target ligands
A 1 R (2104), A 2A R (2489) and PDE10A inhibitors (679) consisting of the common and frequent heterocycles, were subjected to RECAP analysis/synthesis in MOE [26]. The RECAP function electronically fragments and recombines molecules based on chemical knowledge of 11 chemical bond types derived from common chemical reactions [27]. As a result, 458,839 novel RECAP-derived compounds were formed. Finally the designed RECAP library was filtered using RDKit, Python according to the common and frequent heterocycles identified, which narrowed the list down to 22,233 compounds.

Target prediction
The SMILES of the designed RECAP library were standardized using the ChemAxon Command-Line Standardizer where the following options were selected: "Remove Fragment" (keep largest), "Neutralize", "RemoveExplic-itH", "Clean2D", "Mesomerize" and "Tautomerize" [56]. The standardized canonical SMILES were exported to CSV files, and subjected to enriched target prediction using PIDGIN 1.0 implementing the method developed by Liggi et al. [24,31]. The target prediction for the designed RECAP library was performed using a recall probability threshold of 0.01 (which is a value consistent with greater confidence in the more positive predictions).
Enrichment calculations for the predicted targets of the designed RECAP library were performed to assess the likelihood of the active compounds against the targets of interest. In this procedure, the frequency of predicting A 1 R, A 2A R and PDE10A targets for the designed RECAP library was compared with a background distribution of a diverse library covering a large chemical space and was assessed by two parameters: the estimation score and the average ratio. The cutoff selected for considering a target as sufficiently enriched required an estimation score less than or equal to 0.01 [31]. The statistical relevance of the prediction was assessed via a Chi squared test with yates correction in Scipy [32], using the contingency table of the RECAP library and background of randomly sampled PubChem compounds (Additional file 1: Figure S2).

Receptor preparation
Docking with Glide [57] was performed against the human A 2A R protein crystal structure (PDB ID: 4EIY) bound to the antagonist ZM241385 and the PDE10A crystal structure (PDB ID: 4DDL) complexed with an inhibitor [33,34]. Protein structures were prepared using the protein preparation wizard of maestro 9.3 [58], following the default protocol which accounts for energy refinement, hydrogen addition, pKa assignment, and side-chain rotational isomer refinement. Resolved water molecules were discarded, and the structure was centered using the co-crystallized ligand as the center of the receptor grid generated for each protein structure. The A 1R homology model (Additional file 2) was constructed according to the method reported by Yaziji et al. [59][60][61], where the protein sequence of the human A 1 R (accession number P30542) was aligned with the A 2A R template of PDB ID: 4EIY.

Ligand preparation
The entire set of 2563 ligands was prepared for docking with LigPrep 2.5 [62] using the default settings and the Epik option which introduces energy penalties associated with ionization and tautomerization [63].

Cut-off generation for compound selection from docking models
In an attempt to validate the constructed A 2A R, A 1 R, and PDE10A docking models, a set of known actives and inactives were docked against each target to ensure that they enriched actives. 81 A 2A R receptor ligands reported in the literature were docked against the A 2A R model [64,65]. For consistency 81 ChEMBL actives were also selected (for each of the A 1 R and PDE10A proteins whose K i and IC 50 values are less than 10 µM), and these were docked against their respective target class. In addition, PubChem inactives (200 compounds) of each target class were docked.
A good separation was obtained for the medians of docking score distribution for actives versus inactives confirming that the actives are enriched. Additional file 1: Figure S5 shows the separation of the medians for the three docking models, − 6.93 (actives) versus − 5.64 (inactives) for the PDE10A docking model, − 7.66 (actives) versus − 6.01 (inactives) for the A 2A R docking model, and − 7.60 (actives) versus − 5.66 (inactives) for the A 1 R docking model. Statistical analysis was performed with R using a Mann-Whitney test [66] on the active and inactive docking score distributions of each target. The differences in medians were significant with p values < 0.05 (script provided in Additional file 1).
The F 1 score which is the harmonic mean of precision and recall, was computed (using a Python script, see Additional file 1) for all the docking scores of the ChEMBL actives and PubChem inactives for each model. A search was performed for a docking score threshold that gave the highest F 1 score, in order to perform substructure analysis on compounds that were predicted as A 1 R/A 2A R-PDE10A multi-target ligands by target prediction, and displayed docking scores that are lower than or equal to those with the highest F 1 score for each of the three docking models (A 1 R, A 2A R, and PDE10A, see Additional files 3, 4, and 5). Furthermore, the thresholds found are intended to serve as reference scores for any structure-based design problem at these target classes.

Docking
The RECAP compounds that were predicted as A 1 R/ A 2A R-PDE10A multi-target ligands were docked against the A 2A R protein crystal structure (PDB ID: 4EIY) [33], the A 1 R homology model and the PDE10A protein crystal structure (PDB ID: 4DDL) [34] to investigate the molecular interactions. The Glide docking parameters used here are given in Additional file 1: Table S3. The parameters were deduced from docking experiments using known actives and inactives against each protein model.

Synthesis of novel 4,6-substituted 2-amino-pyridin-3-carbonitriles
Due to both ease of the reaction and yield, a one-pot synthetic scheme was optimized for the purpose of synthesizing 2-aminopyridine-3-carbonitriles. For the other series, the synthetic routes were multi-step reactions, which due to synthetic complexity are not reported here.
Hence, given the fact that the 2-aminopyridine-3-carbonitriles were the only RECAP series that could be synthesized via a one-pot synthetic scheme [37,76,77], we have selected these for synthesis and subsequent validation as multi-target ligands. In particular, we selected compounds, which did not exihibit any potential PAINs liability upon screening with the FAFDrug3 ADME-Tox Filtering Tool [40].

Chemistry
Unless otherwise indicated, all starting materials, reagents and solvents were purchased and used without further purification. After extraction from aqueous phases, the organic solvents were dried over anhydrous sodium sulfate. The reactions were monitored by thin-layer chromatography (TLC) on 2.5 mm Merck silica gel GF 254 strips, and each of the purified compounds showed a single spot; unless stated otherwise, UV light and/or iodine vapor were used to detect compounds. The synthesis of the target compounds was performed in coated Kimble vials on a PLS (6 × 4) Organic Synthesizer with orbital stirring. Filtration and washing protocols for supported reagents were performed in a 12-channel vacuum manifold. The purity and identity of all tested compounds were established by a combination of HPLC, elemental analysis, mass spectrometry and NMR spectroscopy as described below. Purification of isolated products was carried out by column chromatography (Kieselgel 0.040-0.063 mm, E. Merck) or medium pressure liquid chromatography (MPLC) on a CombiFlash Companion (Teledyne ISCO) with RediSep pre-packed normal-phase silica gel (35-60 µm) columns followed by recrystallization. Melting points were determined on a Gallenkamp melting point apparatus and are uncorrected. The NMR spectra were recorded on Bruker AM300 and XM500 spectrometers. Chemical shifts are given as δ values against tetramethylsilane as internal standard and J values are given in Hz. Mass spectra were obtained on a Varian MAT-711 instrument. Analytical HPLC was performed on an Agilent 1100 system using an Agilent Zorbax SB-Phenyl, 2.1 mm × 150 mm, 5 µm column with gradient elution using the mobile phases (A) H 2 O containing 0.1% CF 3 COOH and (B) MeCN and a flow rate of 1 mL/min. The purity of all tested compounds was determined to be greater than or equal to 95%.
The synthesis of the 4,6-substituted 2-amino-pyridin-3-carbonitriles 1-25 was done via the one-pot synthetic route shown in Scheme 1. Varying both substituents on the ylidene malononitrile and the ketone reagents resulted in a variation of the substituents on positions 4 and 6 of the pyridine ring.

Synthetic procedure
Substituted ylidene malononitrile (1.0 mmol), ketone (1.0 mmol) and ammonium acetate (5.0 mmol) in a 1:1 toluene/EtOH mixture (7 mL) were stirred in a coated Kimble vial at 120 °C for 12-24 h. After reaction completion (TLC control), distilled water was added and the mixture was extracted with ethyl acetate (3 × 10 mL). The organic phase was dried (Na 2 SO 4 ) and evaporated under reduced pressure to afford an oily residue that was purified by column chromatography using n-hexaneethyl acetate in 2:1 mixture.
The activity measurements against the phosphodiesterases PDE7A, PDE7B, PDE9A and PDE10A were performed using AD293 cells that were transiently and separately transfected with human PDE7A, PDE7B, PDE9A, and PDE10A following the procedure described by Shipe et al. [43]. The IC 50 values were obtained by fitting the data with non-linear regression using Prism 2.1 software (GraphPad, San Diego, CA) [79], and the reported results are the mean of 3 experiments (n = 3) each performed in duplicate.