Is chemical synthetic accessibility computationally predictable for drug and lead-like molecules? A comparative assessment between medicinal and computational chemists

European Journal of Medicinal Chemistry 54 (2012) 679e689

Contents lists available at SciVerse ScienceDirect

European Journal of Medicinal Chemistry journal homepage: http://www.elsevier.com/locate/ejmech

Original article

Is chemical synthetic accessibility computationally predictable for drug and lead-like molecules? A comparative assessment between medicinal and computational chemists Pascal Bonnet* Janssen Research & Development, Division of Janssen Pharmaceutica N.V., Turnhoutseweg 30, 2340 Beerse, Belgium

a r t i c l e i n f o

a b s t r a c t

Article history: Received 24 February 2012 Received in revised form 4 June 2012 Accepted 12 June 2012 Available online 21 June 2012

The design of lead and drug-like molecules with expected desired properties and feasible chemical synthesis is one of the main objectives of computational and medicinal chemists. Prediction of synthetic feasibility of de novo molecules is often achieved by the use of in-silico tools or by advices received from medicinal and to a lesser extent from computational chemists. However, the validation of predictive tools is often performed on selection of compounds from external databases. In this study, we compare the synthetic accessibility (SA) score predicted by SYLVIA and the score estimated by medicinal chemists who synthesized the molecules. Therefore, we solicited 11 bench-based medicinal and computational chemists to score 119 lead-like molecules synthesized by same medicinal chemists. Their scores were compared with score calculated from SYLVIA software. Irrespective of the starting material database, we obtained a good agreement between average of medicinal and computational chemist scores for the ensemble of compounds; as well as between all chemists and SYLVIA SA scores with a correlation of 0.7. Furthermore, analysis of the marketed drugs since 1970 shows some consistency in average SYLVIA SA scores. Compounds entered in different phases of clinical trials show some large variation in synthetic accessibility scores due to natural-derived molecular scaffolds. Here, we proposed that the selection of compounds based on synthetically accessibility should not be done solely by one individual chemist to avoid personal gut-feeling appreciation from its experience but by a group of medicinal and computational chemists. By assessing synthetic accessibility of hundreds of compounds synthesized by medicinal chemists, we show that SYLVIA can be used efficiently to rank and prioritize virtual compound libraries in drug discovery processes. Ó 2012 Elsevier Masson SAS. All rights reserved.

Keywords: De novo design Drug discovery Synthetic accessibility Synthesis design Virtual screening

1. Introduction Computer-aided ligand design is heavily used by pharmaceutical companies in many drug design projects since it generates novel ideas to be exploited by medicinal chemists. The increase of computational power such as grid computing, processor performance and more recently cloud computing allows the generation of thousands in silico lead-like molecules in a reasonable time frames. The recent in silico tools developed by many software and pharmaceuticals companies can generate large number of de novo ligands with novel chemical structures such as VEHICLe [1], TIN [2] or GDB [3,4]. The virtual compounds generated for a specific project are then filtered by desired criteria such as Lipinski’s rules [5], affinity prediction, ADME-tox parameters or pharmacophoric features. However, the remaining prioritized compounds need to * Corresponding author. Tel.: þ32 14 605 961; fax: þ32 14 605 403. E-mail address: [email protected]. 0223-5234/$ e see front matter Ó 2012 Elsevier Masson SAS. All rights reserved. http://dx.doi.org/10.1016/j.ejmech.2012.06.024

be synthesized and validated in a relevant biological assay. In many cases, the ultimate prioritization step performed by medicinal chemists is assessed by the ease of synthesis of the compounds. Many de novo rationale design tools such as LUDI [6], LigBuilder [7,8], SkelGen [9], HOOK [10], BREED [11], SPROUT [12,13], FLexNovo [14] and PhDD [15] are aimed to generate large number of diverse ligands. Even though the generated molecules fulfill the expected requirements for binding affinity or drug-likeness, their chemical structure is often so complex that their synthesis cannot be executed in a fast and easy way. However, to validate in silico models, de novo molecules have to be synthesized and tested on biological systems. De novo methods generate large number of molecules which are often arduous to synthesize due to their unavailable starting materials, stereochemistry, ring complexity and substituents arrangement. Therefore, tools to compute synthetic accessibility are needed to further filter out de novo ligands. Synthetic accessibility corresponds to the ease of synthesis of organic compounds according to their synthetic complexity

680

P. Bonnet / European Journal of Medicinal Chemistry 54 (2012) 679e689

which combines starting materials information and structural complexity. To evaluate synthetic accessibility score, several methods have been developed which takes into account molecular complexity, information of starting materials or retrosynthetic analysis. The advantage of the methods that predict synthetic feasibility from structural complexity [16,17,18] is usually the speed of the calculations since they can easily process tens to hundreds of thousands of molecules in a reasonable time frame. Recently, several methods have been developed such as CAESA [19], RECAP [20], WODCA [21], LHASA [22], RASA [23], RSsvm [24] and SYLVIA [25] that perform retrosynthesis and/or synthetic accessibility prediction for compound libraries. To validate these methods, several attempts have been made where experienced chemists evaluate ease of synthesis of large data set of compounds. It has been shown in several examples that experienced medicinal chemists don’t score synthetic accessibility of compounds in a consensus manner [25e28]. Takaoka et al. [26] use a data set of 3980 diverse compounds and five chemists assigned two scores corresponding to ease of synthesis and compound drug-likeness. From these data the group developed predictive statistical models to rank novel compounds and to filter out hard-to-synthesize compounds. The models can be used to prioritize compounds acquire by external providers. In the project of Lajiness et al. [27] thirteen medicinal chemists reviewed 22,000 compounds divided into 11 lists of about 2000 compounds for their “attractiveness”. They have shown that the chemists are not consistent in rejecting undesirable compounds. Same conclusion was obtained when the chemists reviewed again a set of identical 2000 compounds. Podolyan et al. [24] presented two support vector machines-based models; RSsvm, a statistical model trained on a set of reactions and information of starting materials and DRsvm which takes into account synthetic information of nearest neighbors and is therefore not tied to a specific set of reactions or starting materials. To validate the SAscore, a score that estimates synthetic accessibility [28], Ertl et al. asked 9 experienced chemists to score 40 diverse molecules selected from the PubChem database. A very good enrichment (r2 ¼ 0.89) was obtained between consensus estimated score from medicinal chemists and calculated score from SAscore. The synthetic accessibility score (SAscore) is calculated from a combination of fragment contributions and a complexity penalty. In addition, Boda et al. [25] used a dataset of 100 diverse molecules extracted from the Journal of Medicinal Chemistry, which were estimated by 5 medicinal chemists. The weights of each individual component to calculate the total synthetic score of SYLVIA were estimated by linear regression analysis using the average scores provided by the medicinal chemists. The reliability of the method was estimated by comparing the average computational scores and chemist estimations, a good correlation of 0.89 was obtained from this analysis. Recently, a retrosynthesis-based scoring method called RASA [23] (Retrosynthesis-based Assessment of Synthetic Accessibility) was trained on 100 compounds extracted from the CMC (Comprehensive Medicinal Chemistry) database. Five chemists were selected to assess independently the synthetic accessibility of the compounds using publically available information. The weights of the three individual components contained in the scoring function were derived from linear regression analysis. To validate the scoring function, the former 5 chemists were asked to score 30 new compounds extracted from the CMC and 5 other chemists were asked to estimated synthetic accessibility of 25 additional compounds. Good correlation coefficients of 0.81 and 0.79 were obtained between the calculated RASA scores and the estimated scores respectively. Validation of synthetic accessibility score has been performed on compounds, not made by medicinal chemists involved into the assessment, but rather extracted from external libraries such as

MDL Drug Data Report (MDDR) [29], PubChem [30] or ZINC [31] databases. However, to correctly assess the ability of medicinal chemists to estimate synthetic accessibility of molecules, we validate their perception using a library of 119 compounds synthesized by the experienced bench-based medicinal chemists themselves and perform a cross-evaluation between all the medicinal chemists. At least one chemist knows about the number of synthetic steps, synthetic feasibility and starting material availability. Therefore the prediction of synthetic accessibility of the compounds is performed on a dataset where knowledge of synthesis is known by the chemists. Since in silico ligands are often proposed by computational chemists, we solicit 4 computational chemists to score the compounds having an average of 11 years of experience in the drug design field. Furthermore to check the consistency of scoring molecules by all the chemists, i.e. the medicinal and computational chemists, we randomly include many times same molecules in the library. In this study, SYLVIA software was used to calculate synthetic accessibility score. Synthetic accessibility is of high importance in early drug discovery stage but also in manufacturing processes of drug molecules. GMP (Good Manufacturing Practice) batch productions require being consistently reliable and reproducible in large scale chemical synthesis which usually prevent difficult synthetic routes to achieve low-cost manufacturing processes, high synthesis purity, quantity and quality. However, drugs with difficult synthesis steps were approved in the last few years such as Fuzeon (enfuvirtide) or several natural products [32]. The recent 2010 approval of oncology drug Halaven (eribulin) [33,34] by the FDA (Food and Drug Administration) [35] shows that highly synthetically complex drugs [36,37] can still be marketed despite economically manufacturing process challenges. In this study, we also analyzed the synthetic feasibility of marketed drugs using SYLVIA software as well as compounds in different clinical phases. The synthetic accessibility of large molecules, often derived from natural products is not covered due to their specific synthesis approach, and only small drug-like molecules are included in this study. In addition to others molecular informatics tools [38], the evaluation of synthetic accessibility of virtual ligand database could be very useful to distinguish difficult versus easy-to-make compounds. The synthetic accessibility score can be used to prioritize compounds in order to get compound synthesized and tested more rapidly. In this context, the global drug design cycle time could be accelerated substantially. 2. Methods 2.1. Datasets 11 chemists, including 7 medicinal chemists and 4 computational chemists with several years of experience in drug discovery agreed to score 119 lead-like molecules based on their synthetic accessibility (SA). In previous studies, the comparison of synthetic accessibility and feasibility of compounds by medicinal chemists was always performed on a dataset obtained from external libraries. In this study, each molecule included in the dataset was made by one of the selected medicinal chemists; therefore at least one medicinal chemist has experience of known chemical synthesis route and available building blocks for each compound. All selected compounds have a molecular weight greater than 300 Da and are not reaction intermediates bearing any protecting group. Compounds made by each chemist were extracted from the Johnson & Johnson corporate database. SYLVIA synthetic accessibility score (SSc) was calculated on all compounds and divided into 4 bins (SSc  3, 3 < SSc  4, 4 < SSc  5, SSc > 5) where high score indicates difficult compounds to synthesize. For each bin, the

P. Bonnet / European Journal of Medicinal Chemistry 54 (2012) 679e689

compounds were clustered using functional-class extendedconnectivity fingerprints (FCFP_4) computed in Pipeline Pilot version 7.5 [39] with an average of 10 compounds per cluster. Maximum dissimilarity was used for selecting new cluster center. An average of 17 compounds made by each chemist was selected (Table 1). This approach allows direct comparison between calculated SYLVIA synthetic accessibility scores and estimated chemist scores. All compounds were randomly sorted, so that chemists could not score consecutively compounds from same cluster. 10 compounds were presented each time to computational and medicinal chemists without the possibility to look back to their previous scores. Interestingly some enantiomers were randomly included into the dataset to assess consistency of the chemists in their scoring evaluation. To facilitate the comparison between predicted chemist scores and SYLVIA SA scores, the values were mean centered and scaled to unit variance for medicinal and computational chemist scores. This post-processing is necessary to avoid any bias due to overestimation of synthetic accessibility scores of compounds by chemists. We also analyzed a dataset of drug molecules extracted from the Prous Science Integrity database [40]. We selected drug molecules with molecular weight between 300 and 600 Da. Only molecules in clinical phases 1, 2 and 3 and at least once launched since 1970 were selected. Molecules having at least one inorganic atom (i.e. not like H, C, N, O, P, S, F, Cl, Br or I) were discarded as well as molecules having at least one lead-like violation criteria [41]. A total of 2420 and 799 molecules was obtained for molecules in phase 1, 2 or 3 and launched respectively. For each drug molecule, synthetic accessibility score and individual contributions were computed using SYLVIA. 2.2. Sylvia synthetic accessibility score The synthetic accessibility scores were calculated with SYLVIA version 1.2 (Molecular Networks Gmbh) [25]. SYLVIA evaluates the ease of synthesis of organic compounds by summing five weighted individual components. This latest version allows extraction of the five individual components contributing to the total synthetic accessibility score such as: Molecular Graph Complexity Score (MGCS), Ring Complexity Score (RCS), Stereochemical Complexity Score (SCS), Starting Material Similarity Score (SMSS), Reaction Center Substructure Score (RCSS). The first three components (MGCS, RCS and SCS) are based on chemical structures whereas SMSS and RCSS are based on data provided by external databases such as starting material database and product reaction center substructure database respectively. In the current study, we evaluate the effect of the variation in synthetic accessibility score by using the default starting material database (SMSS) and by providing the corporate starting material database (SMSS_rescore). More information regarding the calculation of each individual component is detailed by Boda et al. [25]. Briefly, MGCS uses the graph and information theory [42]. Molecular descriptors such as size, molecular symmetry, number of rings, number of bonds and atoms are included in the calculation of MGCS. Bridged ring Table 1 Number of clusters and molecules per cluster synthesized by each medicinal chemist. Chemist Chemist Chemist Chemist Chemist Chemist Chemist Chemist

1 2 3 4 5 6 7

Number of clusters

Number of molecules

1 7 3 3 3 2 2

19 15 15 12 18 20 20

681

systems and chiral centers will increase RCS and SCS values respectively and will therefore penalize the total synthetic accessibility score. It is considered that stereochemical centers increase the complexity of synthetic routes and compounds are more difficult to synthesize. SMSS is obtained by calculating the transformation-based similarity of target compounds to starting materials provided as an additional database. Transformationbased similarity depends on generalized reactions and on structural and topological molecular features. The transformed structures are compared to molecules in starting material database which combines unique molecules from Fluka, Acros and Maybridge catalogs [43]. It is expected that molecules with complex chemical features are easier to synthesize if starting materials are already available. To assess the effect of including additional starting materials, the corporate starting material database was also used to calculate the total synthetic accessibility score. SMSS values decrease for compounds having low transformation-based similarity to the standard starting material database. Finally, for calculating reaction center substructure score (RCSS), the target molecules are compared to known reaction center substructures which are provided in external databases. It is considered that RCSS is analogous to retrosynthetic reaction fitness in such a way that decomposed structural motifs of target molecules is compared to substructures derived from reaction database. The target molecule is dissected into decomposed retrosynthetic motifs using known retrosynthetic transformations and each motif is compared to reaction center substructures. The total synthetic accessibility score, which is the sum of the five individual contributing components added to a constant of 0.68, is scaled between 1 and 10 where larger values correspond to compounds which are more difficult to synthesize. 3. Results and discussion For all the 119 corporate compounds synthesized by medicinal chemists, the SYLVIA synthetic accessibility score was calculated using default parameters. Before being provided to the chemists, the compounds were randomly distributed for the scoring exercise. The synthetic accessibility of the compounds was estimated by 7 medicinal chemists and 4 computational chemists without the possibility to compare with their previous scores. Table 2 shows the correlation matrix between individual computational chemists (C) and medicinal chemists (M), as well as the average values. Despite poor correlation in general between medicinal chemist peers, the maximum correlation is of 0.54 between individual medicinal chemists (M4eM7 pair), 0.62 between individual computational chemists (C1eC4 pair) and 0.60 between computational chemists and medicinal chemists (M6eC3 pair). It is surprising to observe such poor correlations between experienced chemists, such as between M1 and M4, M1 and M7, M1 and M5 and finally M3 and M5 (correlation coefficient ¼ 0.04, 0.06, 0.13, 0.19 respectively). However, the correlation between average scores of computational chemists and average scores of medicinal chemists is reasonable (correlation coefficient ¼ 0.70), therefore synthetic accessibility of compounds is significantly different if it is estimated by individual chemists or by a group of chemists for which the average of all individual scores is higher. It shows that a group of experienced computational or medicinal chemists, instead of individuals has the greater ability to evaluate synthetic accessibility of molecules for compound prioritization. These results highlight the importance of selecting a group of chemists instead of individual chemist in assessing synthetic feasibility of compounds from any HTS (High-Throughput Screening) post-processing analysis or from virtual screening analysis. Fig. 1 highlights the most important components in the total synthetic accessibility score. Clearly, molecular graph complexity

682

P. Bonnet / European Journal of Medicinal Chemistry 54 (2012) 679e689

Table 2 Correlation coefficients for computational (C) and medicinal chemists (M) in assessing synthetic accessibility score of 119 compounds.

M1 M2 M3 M4 M5 M6 M7 C1 C2 C3 C4 Avg. C Avg. M Tot. Avg.

M1

M2

M3

M4

M5

M6

M7

C1

C2

C3

C4

Avg. C

Avg. M

Tot. Avg.

1.00 0.21 0.23 0.04 0.13 0.32 0.06 0.10 0.17 0.01 0.06 0.12 0.44 0.35

0.21 1.00 0.34 0.27 0.39 0.50 0.44 0.47 0.39 0.35 0.07 0.46 0.70 0.66

0.23 0.34 1.00 0.32 0.19 0.38 0.22 0.22 0.12 0.23 0.19 0.27 0.60 0.51

0.04 0.27 0.32 1.00 0.27 0.30 0.54 0.49 0.20 0.45 0.40 0.55 0.61 0.64

0.13 0.39 0.19 0.27 1.00 0.47 0.47 0.42 0.34 0.49 0.35 0.58 0.65 0.67

0.32 0.50 0.38 0.30 0.47 1.00 0.42 0.43 0.32 0.60 0.28 0.58 0.76 0.75

0.06 0.44 0.22 0.54 0.47 0.42 1.00 0.53 0.27 0.53 0.23 0.56 0.70 0.70

0.10 0.47 0.22 0.49 0.42 0.43 0.53 1.00 0.26 0.62 0.25 0.77 0.60 0.71

0.17 0.39 0.12 0.20 0.34 0.32 0.27 0.26 1.00 0.27 0.12 0.59 0.40 0.51

0.01 0.35 0.23 0.45 0.49 0.60 0.53 0.62 0.27 1.00 0.36 0.81 0.59 0.73

0.06 0.07 0.19 0.40 0.35 0.28 0.23 0.25 0.12 0.36 1.00 0.62 0.35 0.49

0.12 0.46 0.27 0.55 0.58 0.58 0.56 0.77 0.59 0.81 0.62 1.00 0.70 0.88

0.44 0.70 0.60 0.61 0.65 0.76 0.70 0.60 0.40 0.59 0.35 0.70 1.00 0.96

0.35 0.66 0.51 0.64 0.67 0.75 0.70 0.71 0.51 0.73 0.49 0.88 0.96 1.00

score (MGCS) is the strongest component to the total SYLVIA SA score. For many compounds the stereochemical complexity score (SCS) is null and SYLVIA SA score is relatively constant. An increase of SYLVIA synthetic accessibility score observed for some molecules (molecular number > 60 in Fig. 1) is mainly due to the additional contribution of stereochemical complexity score (SCS). The high synthetic accessibility score (SYLVIA SA score > 6) is due the contribution of molecular graph complexity score (MGCS) which increases because of the number of rings, molecular weight and number of bonds; descriptors used to calculate MGCS. When stereocenters are present in molecular structures, it largely increases the total SYLVIA SA score making these compounds in general more difficult to synthesize. The work of Lovering et al. [44] describes a simple method of measuring molecular complexity using carbon bond saturation. They show that compounds with greater complexity are more likely to succeed at each stage from discovery to drug and they observed an increase in the percentage of compounds that have one or more stereo centers. They argue that compounds with less aromatic rings would have better solubility, hence increasing the chance of clinical success. More complex molecules will explore greater chemical space. However, while

coupling reactions of sp2esp2 carbons are well described in the literature, sp3esp3 coupling reactions are still more challenging. As observed in the current study, compounds with high SYLVIA SA score have in general superior complexity score and medicinal chemists estimated these molecules difficult to synthesize. While it might be more challenging to synthesize complex molecules, the added value of such molecules in drug discovery could be greater. Most medicinal chemists would argue that the chemical synthesis of molecules containing stereocenters is often more challenging. However in some cases, asymmetric synthesis might become easier depending on the availability of enantiomerically pure starting materials, chiral auxiliaries or synthetic route accessibility [45] Correlation between SYLVIA synthetic accessibility score and combined medicinal and computational chemist scores is 0.66 (Table 3, Fig. 2) or 0.70 after removing one outlier (X chemist score ¼ 1.6, Y SYLVIA SA score ¼ 4.2). Interestingly, the correlation increases to 0.71 (or 0.74 after removing same outlier) when chemist score is estimated by only the medicinal chemist who synthesized the corresponding molecule instead of using the average of all medicinal chemist scores. This small improvement suggests that SYLVIA scoring function could be improved during the

7

SYLVIA Synthetic Accessibility Score

6

5

4

3

2

1

1 4 7 10 13 16 19 22 25 28 31 34 37 40 43 46 49 52 55 58 61 64 67 70 73 76 79 82 85 88 91 94 97 100 103 106 109 112 115 118

0

Molecule number MGCS

RCSS

RCS

SMSS

SCS

Cst

Fig. 1. Contribution of individual components from SYLVIA for all molecules in dataset ranked by synthetic accessibility score (MGCS: molecular graph complexity score, RCS: ring complexity score, SCS: stereochemical complexity score, SMSS: starting material similarity score, RCSS: reaction center substructure score, Cst ¼ 0.68).

P. Bonnet / European Journal of Medicinal Chemistry 54 (2012) 679e689 Table 3 Correlation matrix of synthetic accessibility score calculated by SYLVIA and computational (C) and medicinal chemists (M) scores. MGCS: molecular graph complexity score, RCS: ring complexity score, SCS: stereochemical complexity score, SMSS: starting material similarity score, RCSS: reaction center substructure score. In parenthesis are the scores obtained from chemists who synthesized the molecules. Property

Synthetic accessibility score

Synthetic accessibility rescore

Avg. C

Avg. M

Tot. Avg.

MGCS RCSS RCS SMSS SMSS rescore SCS Synthetic accessibility Score Synthetic accessibility rescore M1 M2 M3 M4 M5 M6 M7 C1 C2 C3 C4

0.79 0.39 0.04 0.68 0.54 0.81 1.00

0.77 0.38 0.05 0.70 0.59 0.82 1.00

0.61 0.29 0.17 0.36 0.28 0.58 0.71

0.49 0.31 0.18 0.24 0.19 0.43 0.55

0.58 0.33 0.19 0.31 0.24 0.53 0.66

1.00

1.00

0.71

0.55

0.66 (0.71)

0.16 0.30 0.18 0.51 0.37 0.43 0.49 0.61 0.41 0.58 0.39

0.18 0.32 0.17 0.50 0.37 0.43 0.50 0.62 0.42 0.57 0.38

0.12 0.46 0.27 0.55 0.58 0.58 0.56 0.77 0.59 0.81 0.62

0.44 0.70 0.60 0.61 0.65 0.76 0.70 0.60 0.40 0.59 0.35

0.35 0.66 0.51 0.64 0.67 0.75 0.70 0.71 0.51 0.73 0.49

(0.58) (0.28) (0.20) (0.39) (0.32) (0.60) (0.71)

validation process by soliciting the medicinal chemists who synthesized the compounds. All individual components used to calculate SYLVIA SA score are higher if the medicinal chemist score is taken from the chemist who synthesized the molecule than medicinal chemist score averaged over all medicinal chemists. As

683

expected, compounds which are estimated to be difficult to synthesize by the chemists have indeed a higher SYLVIA SA score. The default threshold of score 6 proposed in SYLVIA software confirms the reasonable upper limit to consider compounds difficult to access synthetically. Table 4 shows the correlations for groups of compounds synthesized by individual medicinal chemists. It also highlights the difference between SYLVIA SA scores estimated by all chemists and by only medicinal chemist who synthesized the corresponding compounds for each group. It appears that in almost all groups the correlation is better when the assessment is performed by the medicinal chemists who performed the synthesis of the compounds with the exception of chemist 4. To better characterize the correlation of the synthetic accessibility score calculated by SYLVIA and the score estimated by the chemists, we grouped the chemist scores into 3 categories: easy, middle and difficult and SYLVIA SA scores into 4 bins: low (SSc  3), medium-low (3 < SSc  4), medium-high (4 < SSc  5) and high (SSc > 5) (Fig. 3). With this data reduction approach, it is easier to appreciate the number of compounds that are predicted easy or difficult to synthesize by SYLVIA and by the chemists. Fig. 3 shows that the linear categories of chemist score/SYLVIA SA score such as easy/low, middle/medium and difficult/high are indeed mostly populated by the compounds. There is only one outlier in the category difficult/low that corresponds to compound predicted easy to synthesize by SYLVIA but perceived more difficult by the chemists. In addition, SYLVIA does not overestimate compounds difficult to synthesize since there is no compound in the category easy/high. As already mentioned above, Fig. 3 shows that there is a good agreement between categories obtained from chemist score and SYLVIA SA score. Table 3 shows the correlation values of SYLVIA synthetic accessibility score with individual chemists. Correlation between SYLVIA SA scores and scores averaged over all medicinal chemists is higher than with scores averaged over all computational chemists, 0.71 and 0.55 respectively. However the highest

Fig. 2. Average scores per compound from all medicinal and computational chemists compared to SYLVIA synthetic accessibility score.

684

P. Bonnet / European Journal of Medicinal Chemistry 54 (2012) 679e689

Table 4 Correlation matrix between synthetic accessibility scores and chemist scores using group of compounds synthesized by each medicinal chemist. G1 is the group of compounds synthesized by medicinal chemist M1. Group of compounds

Pearson correlation using average over all chemist scores

Pearson correlation using scores of chemist who synthesized the compounds

G1 G2 G3 G4 G5 G6 G7

0.69 0.29 0.67 0.64 0.62 0.80 0.80

0.73 0.45 0.73 0.47 0.63 0.81 0.85

correlation between individual medicinal chemists and computational chemists with SYLVIA SA score is 0.51 and 0.50 respectively. Overall, all individual components from SYLVIA score (MGCS, RCSS, RCS, SMSS and SCS) have a better correlation with medicinal chemists than with computational chemists. However, a closer analysis to the correlation obtained between individual medicinal or computational chemists score and SYLVIA SA score shows some striking differences. The low correlation of 0.12 and 0.27 for medicinal chemists M1 and M3 respectively compared to the average score of all medicinal chemists indicates that these two chemists have a different interpretation of assessing synthetic accessibility score of compounds compared to other medicinal chemists. These two medicinal chemists M1 and M3 have also the lowest correlation to synthetic accessibility score. These large deviations in synthetic feasibility prediction are potentially coming from the experience and background of each individual chemist. Table 3 shows the fluctuation in starting material similarity score using the corporate starting material database (SMSS rescore) instead of the default starting material database (SMSS). In comparison to the total SYLVIA SA score and rescore, the SMSS rescore is slightly lower than SMSS; 0.68 and 0.59 for score and

rescore respectively. By using the information of corporate starting materials, total SYLVIA SA rescore has lower values than SYLVIA SA score which includes default parameters. However there is a very good correlation between SMSS and SMSS rescore with a correlation coefficient of 0.91. Only for few compounds SMSS rescore is lower than SMSS which decreases the total SYLVIA SA score. Interestingly, the correlation between average chemist score and SMSS and SMSS rescore is 0.31 and 0.24 respectively. However the correlation between average chemist score and total SYLVIA SA score did not change with the use of corporate starting material; Pearson correlation is 0.66 for SYLVIA SA score and SYLVIA SA rescore. This result suggests that the starting material database provided in the software is correctly exemplified. To assess synthetic accessibility, SYLVIA uses multiple components which were weighted based on the linear regression analysis of 100 compounds scored by five chemists. Here we show that 11 new chemists ranking 119 different compounds gave similar trends suggesting some confidence in the application of SYLVIA to predict easy-to-synthesize compounds. This good correlation obtained from Fig. 2 suggests that SYLVIA can be efficiently used to predict synthetic accessibility of virtual compounds. We then look more closely to the prediction of synthetic accessibility given by chemists for some compounds and compare the results with SYLVIA SA score. Table 5 shows the synthetic accessibility score estimated by the chemists and calculated by SYLVIA for a diverse set of chemical structures that were synthesized by various medicinal chemists selected in this study. It is important to mention again that the compounds, and thus the enantiomers, were randomly presented to the chemists. With this approach, we can identify the variability of predicting synthetic accessibility of compounds during the experiment assuming that the enantiomers were made in same synthetic route. As example, compounds 6 and 7 are two enantiomers of trisubstituted 1,2,4triazole positive allosteric modulators of alpha-7 nicotinic acetylcholine receptor [56]. The average scores obtained from chemists

Fig. 3. Binned average chemist scores compared to SYLVIA synthetic accessibility scores. Chemist scores were grouped into 3 bins: easy (CSc ¼ 1), middle (CSc ¼ 2) and difficult (CSc ¼ 3) and SYLVIA SA scores were combined into 4 bins: binned score ¼ 3 if SYLVIA score < 4, binned score ¼ 4 if 4  SYLVIA score < 5, binned score ¼ 5 if 5  SYLVIA score < 5, binned score ¼ 6 if SYLVIA score  6. The number of compounds is proportional to the size of the circles.

P. Bonnet / European Journal of Medicinal Chemistry 54 (2012) 679e689

685

Table 5 Average SYLVIA SA scores and chemist scores obtained for some molecules used in the study. In parenthesis, SYLVIA SA score calculated using SMSS rescore and chemist score with values mean centered and scaled to unit variance. Molecule

Chemical structure

SYLVIA SA score

Chemist score

References

1

3.91 (3.91)

2.00 (0.48)

[53,54]

2

5.05 (5.04)

3.00 (0.44)

[55]

3

5.04 (5.02)

3.09 (0.57)

[49]

4

4.11 (4.09)

2.73 (0.18)

[49]

5

4.07 (4.05)

2.82 (0.26)

[49]

6

5.11 (5.11)

2.18 (0.21)

[56]

7

5.11 (5.11)

2.36 (0.07)

[50]

8

4.07 (4.06)

2.45 (0.05)

[57] (continued on next page)

686

P. Bonnet / European Journal of Medicinal Chemistry 54 (2012) 679e689

Table 5 (continued ) Molecule

Chemical structure

SYLVIA SA score

Chemist score

References

9

4.04 (4.03)

2.36 (0.04)

[58]

10

4.04 (3.97)

1.64 (0.81)

[59]

11

5.11 (5.05)

2.45 (0.01)

[60]

12

5.09 (5.03)

2.18 (0.28)

[60]

are very similar for the two enantiomers (2.2 and 2.4) but most interestingly the variability among individual chemists differs considerably. Amongst the 11 chemists, 2 medicinal chemists assigned a distinct synthetic accessibility score with a difference of 2 meaning that they contradict the synthetic accessibility for the two enantiomers. However for the remaining chemists, they all assign same score for the 2 enantiomers. Table 5 also shows some compounds with high degree of similarity such as compounds 2 and 3, compounds 4 and 5, compounds 10 and 11. Almost all the compounds with high chemical similarity were assigned with comparable averaged estimated synthetic accessibility scores; this shows a good consistency of the chemists in estimating synthetic accessibility of molecules. This is in contradiction with the work of Lajiness et al. [27] who observe that chemists reject the same compounds only about 50% of the time. However, Lajiness et al. have used a much larger number of compounds in their study to assess compound attractiveness. It is noteworthy to mention that most of the compounds presented in Table 5 have same range of synthetic accessibility and do not cover the chemical space of compounds used in this study. Table 5 shows that there is a good correlation between SYLVIA score and chemist score independently of the chemical structures. Fig. 4 represents the variation of SYLVIA SA score for drugs approved by the FDA (Food and Drug Administration) since 1970 until 2010 (see Methods). There is no major variation in SYLVIA SA score for the 798 drugs used in this study and approved since 1970; the average of synthetic accessibility score for all the compounds is 4.65 (þ/1.15). The maximum SYLVIA SA score reaches values of 8.5 because of the complexity of several approved drugs such as corticosteroid-containing natural compounds, macrocyclic compounds such as epothilones or mitosane-containing natural products. Surprisingly, minimum SYLVIA SA score reaches very low values which correspond to low molecular weight compounds. The average SYLVIA SA score was also calculated for NME (New

Molecular Entities) from Phase I until Phase III. Average SYLVIA SA scores are 4.79 (þ/1.11), 4.68 (þ/1.04) and 4.68 (þ/1.05) for NMEs in Phase I, II and III respectively. Although, the SYLVIA SA scores are similar to approved drugs, we notice a small decrease in the synthetic accessibility score from Phase I until marketed drugs. This trend is independent on the total number of compounds per clinical phase since the total number of compounds is 827, 1158 and 433 for compounds in Phase I, II, and III respectively. Fig. 5 indicates the variation of individual components included in SYLVIA SA score for each clinical phase. As already observed for compounds used in this study (Fig. 1), the main components driving the increase of SYLVIA SA score are SCS (Stereochemical Complexity Score) and in a lesser extend MGCS (Molecular Graph Complexity Score). It is interesting to observe similar “S” shape curve for all Phases I, II and III clinical trials. Roughly same maxima of 8.9 (PG-490-88), 8.8 (IPI926) and 8.3 (PEP-005) all derived from natural products, and a minima of 1.3 (Enzactin), 1.1 (Hexabid) and 2.7 (Elmustine) are reached for compounds in Phases I, II and III of clinical trials respectively. Here, we have shown that SYLVIA is an appropriate tool to estimate the synthetic accessibility of library of compounds [46] and it can be efficiently applied in filtering de novo ligands generated by structure-based and ligand-based virtual screening [47-50]. 4. Conclusions To prioritize thousands of compounds, in silico methods often used calculated molecular properties such as drug-likeness, ligand efficiency, fit quality, compound clustering, predicted affinity to a target, off-target prediction, intellectual property space and synthetic accessibility. Synthetic accessibility is becoming integrated in a more systematic way in the molecular design workflow and is commonly used in virtual screening to prioritize compounds. We have shown in this study that prediction of synthetic

P. Bonnet / European Journal of Medicinal Chemistry 54 (2012) 679e689

687

Fig. 4. Mean calculated SYLVIA synthetic accessibility score and standard deviation of approved drugs per year since 1970 until 2010 (Red line). Max. and Min. values of SYLVIA SA score are represented in blue and purple respectively. The total number of drugs approved by the FDA for each year in represented in secondary Y axis (Green line).(For interpretation of the references to color in this figure legend, the reader is referred to the web version of this article.)

accessibility by in silico tool correlates well with medicinal and computational chemist estimation. We found that despite poor correlation between certain individual chemists, there is a good correlation between scores obtained from medicinal chemists and those from computational chemists (correlation coefficient of 0.70). We have also shown the influence on the estimation of synthetic accessibility score for individuals versus a group of chemists. From this study, it is clear that the collective experience of chemists

working together will better capture the ease of synthesis of molecules than individual chemist. This is important for prioritizing compounds from a virtual study or for selecting compounds for library enrichment. The increase in SYLVIA synthetic accessibility score observed for some compounds is mainly due to the stereochemical complexity score which is one of the components used to calculate the total score. Depending on the availability of enantiomerically pure starting material, complex molecules with

Fig. 5. Contribution of individual components for all molecules in Phase I, II and III ranked by phase number and SYLVIA synthetic accessibility score (MGCS: molecular graph complexity score, RCS: ring complexity score, SCS: stereochemical complexity score, SMSS: starting material similarity score, RCSS: reaction center substructure score, Cst ¼ 0.68).

688

P. Bonnet / European Journal of Medicinal Chemistry 54 (2012) 679e689

stereocenters are in general more challenging to synthesize. In contradiction to other previous studies, the chemists involved in this work were mostly consistent in scoring enantiomers or structurally similar compounds. As observed with SYLVIA, chemists estimated chiral compounds to be more difficult to synthesize. Therefore SYLVIA can be used to prioritize compounds based on their ease of synthesis as well as starting material availability. The prediction of synthetic accessibility score of drugs approved by the FDA gives an average of 4.65 (1.15) with a maximum of 8.50 due to highly complex natural-like drugs. Similar values were obtained for compounds in Phase I, II and III clinical trials. Application of synthetic accessibility prediction in drug discovery is broad and SYLVIA can be applied as an additional filter in a virtual screening process together with Lipinski’s rule of five [5] or any unwanted chemistry filters [51]. Chemical synthetic accessibility score can be used in two different ways; to prioritize compounds with low synthetic accessibility score to rapidly validate hypothesis generated from de novo design approaches or to prioritize compounds with high score to explore new opportunities of novel chemical space since complex molecules are often discarded from medicinal chemists because of time and resource constraints. The next generation of synthetic accessibility prediction tool will probably combine retrosynthetic analysis, information of starting materials and empirically derived rules trained on a large and diverse data set and assessed by substantial number of chemists. A community volunteer-based project such a CDK [52] (Chemistry Development Kit) could be an interesting project to consider. A tool that could rapidly assess synthetic accessibility of a large virtual library would add value in the prioritization of de novo compounds. Acknowledgments The author would like to thank the medicinal and computational chemists who kindly accepted to score the compounds and Dr. Trevor Howe and Dr. Joannes T.M. Linders for their fruitful comments. The author also would like to thank Molecular Network for providing a trial version of SYLVIA in order to perform this study. Supplementary data Supplementary data associated with this article can be found in the online version, at http://dx.doi.org/10.1016/j.ejmech.2012.06. 024. These data include MOL files and InChiKeys of the most important compounds described in this article. References [1] W.R. Pitt, D.M. Parry, B.G. Perry, C.R. Groom, Heteroaromatic rings of the future, J. Med. Chem. 52 (2009) 2952e2963. [2] K.V. Dorschner, D. Toomey, M.P. Brennan, T. Heinemann, F.J. Duffy, K.B. Nolan, D. Cox, M.F.A. Adamo, A.J. Chubb, TIN e a combinatorial compound collection of synthetically feasible multicomponent synthesis products, J. Chem. Inf. Model. 51 (2011) 986e995. [3] T. Fink, J.-L. Reymond, Virtual exploration of the chemical universe up to 11 atoms of C, N, O, F: assembly of 26.4 million structures (110.9 million stereoisomers) and analysis for new ring systems, stereochemistry, physicochemical properties, compound classes, and drug discovery, J. Chem. Inf. Model. 47 (2007) 342e353. [4] L.C. Blum, J.-L. Reymond, 970 Million druglike small molecules for virtual screening in the chemical universe database GDB-13, J. Am. Chem. Soc. 131 (2009) 8732e8733. [5] C.A. Lipinski, F. Lombardo, B.W. Dominy, P.J. Feeney, Experimental and computational approaches to estimate solubility and permeability in drug discovery and development settings, Adv. Drug Deliv. Rev. 46 (2001) 3e26. [6] H. Bohm, The computer program LUDI: a new method for the de novo design of enzyme inhibitors, J. Comput. Aided Mol. Des. 6 (1992) 61e78. [7] R.-X. Wang, Y. Gao, L.-L. Lai, LigBuilder: a multi-purpose program for structure-based drug design, J. Mol. Model. 6 (2000) 498e516.

[8] Y. Yuan, J. Pei, L. Lai, LigBuilder 2: a practical de novo drug design approach, J. Chem. Inf. Model. 51 (2011) 1083e1091. [9] S. Firth-Clark, H.M.G. Willems, A. Williams, W. Harris, Generation and selection of novel estrogen receptor ligands using the de novo structure-based design tool, SkelGen, J. Chem. Inf. Model. 46 (2006) 642e647. [10] M. Eisen, D. Wiley, M. Karplus, R. Hubbard, HOOK: a program for finding novel molecular architectures that satisfy the chemical and steric requirements of a macromolecule binding site, Proteins: Struct., Funct., Bioinf. 19 (1994) 199e221. [11] A. Pierce, G. Rao, G. Bemis, BREED: generating novel inhibitors through hybridization of known ligands. Application to CDK2, P38, and HIV protease, J. Med. Chem. 47 (2004) 2768e2775. [12] V. Gillet, A. Johnson, P. Mata, S. Sike, P. Williams, SPROUT: a program for structure generation, J. Comput.-Aided Mol. Des. 7 (1993) 127e153. [13] K. Boda, A.P. Johnson, Molecular complexity analysis of de novo designed ligands, J. Med. Chem. 49 (2006) 5869e5879. [14] J. Degen, M. Rarey, FLEXNOVO: structure-based searching in large fragment spaces, ChemMedChem 1 (2006) 854e868. [15] Q. Huang, L.-L. Li, S.-Y. Yang, Q. Huang, PhDD: a new pharmacophore-based de novo design method of drug-like molecules combined with assessment of synthetic accessibility, J. Mol. Graphics Model. 28 (2010) 775e787. [16] R. Barone, M. Chanon, A new and simple approach to chemical complexity. Application to the synthesis of natural products, J. Chem. Inf. Comput. Sci. 41 (2001) 269e272. [17] C. Rücker, G. Rücker, S.H. Bertz, Organic synthesis e art or science? J. Chem. Inf. Comput. Sci. 44 (2004) 378e386. [18] T.K. Allu, T.I. Oprea, Rapid evaluation of synthetic and molecular complexity for in silico chemistry, J. Chem. Inf. Model. 45 (2005) 1237e1243. [19] V. Gillet, G. Myatt, Z. Zsoldos, A. Johnson, SPROUT, HIPPO and CAESA: tool for de novo structure generation and estimation of synthetic accessibility, Perspect. Drug Discovery Des. 3 (1995) 34e50. [20] X.Q. Lewell, D.B. Judd, S.P. Watson, M. Hann, RECAP-retrosynthetic combinatorial analysis procedure: a powerful new technique for identifying privileged molecular fragments with useful applications in combinatorial chemistry, J. Chem. Inf. Comput. Sci. 38 (1998) 511e522. [21] M. Pförtner, M. Sitzmann, in: J. Gasteiger (Ed.), Handbook of Cheminformatics, Wiley-VCH, Weinheim, 2003. pp. 1457e1507. [22] A. Johnson, C. Marshall, P. Judson, Starting material oriented retrosynthetic analysis in the LHASA program. 1. General description, J. Chem. Inf. Comput. Sci. 32 (1992) 411e417. [23] Q. Huang, L.-L. Li, S.-Y. Yang, RASA: a rapid retrosynthesis-based scoring method for the assessment of synthetic accessibility of drug-like molecules, J. Chem. Inf. Model. 51 (2011) 2768e2777. [24] Y. Podolyan, M.A. Walters, G. Karypis, Assessing synthetic accessibility of chemical compounds using machine learning methods, J. Chem. Inf. Model. 50 (2010) 979e991. [25] K. Boda, T. Seidel, J. Gasteiger, Structure and reaction based evaluation of synthetic accessibility, J. Comput.-Aided Mol. Des. 21 (2007) 311e325. [26] Y. Takaoka, Y. Endo, S. Yamanobe, H. Kakinuma, T. Okubo, Y. Shimazaki, T. Ota, S. Sumiya, K. Yoshikawa, Development of a method for evaluating drug-likeness and ease of synthesis using a data set in which compounds are assigned scores based on chemists’ intuition, J. Chem. Inf. Comput. Sci. 43 (2003) 1269e1275. [27] M.S. Lajiness, G.M. Maggiora, V. Shanmugasundaram, Assessment of the consistency of medicinal chemists in reviewing sets of compounds, J. Med. Chem. 47 (2004) 4891e4896. [28] P. Ertl, A. Schuffenhauer, Estimation of synthetic accessibility score of drug-like molecules based on molecular complexity and fragment contributions, J. Cheminform. 1 (2009) 8. [29] Web address: http://www.discoverygate.com. [30] Web address: http://pubchem.ncbi.nlm.nih.gov. [31] J.J. Irwin, B.K. Shoichet, ZINCea free database of commercially available compounds for virtual screening, J. Chem. Inf. Model. 45 (2005) 177e182. [32] T.D. Aicher, K.R. Buszek, F.G. Fang, C.J. Forsyth, S.H. Jung, Y. Kishi, M.C. Matelich, P.M. Scola, D.M. Spero, S.K. Yoon, Total synthesis of halichondrin B and norhalichondrin B, J. Am. Chem. Soc. 114 (1992) 3162e3164. [33] A. Mullard, 2010 FDA drug approvals, Nat. Rev. Drug Discovery 10 (2011) 82e85. [34] T.K. Huyck, W. Gradishar, F. Manuguid, P. Kirkpatrick, Eribulin mesylate, Nat. Rev. Drug Discov. 10 (2011) 173e174. [35] Web address: http://www.fda.gov/NewsEvents/Newsroom/PressAnnounc ements/ucm233863.htm. [36] K.L. Jackson, J.A. Henderson, A.J. Phillips, The halichondrins and E7389, Chem. Rev. 109 (2009) 3044e3079 (Washington, DC, U. S.). [37] M.J. Yu, W. Zheng, B.M. Seletsky, B.A. Littlefield, Y. Kishi, Case history: discovery of eribulin (HALAVENTM), a halichondrin B analogue that prolongs overall survival in patients with metastatic breast cancer, Annu. Rep. Med. Chem. 46 (2011) 227e241. [38] S.J. Lusher, R. McGuire, R. Azevedo, J.-W. Boiten, R.C. van Schaik, J.A. de Vlieg, A molecular informatics view on best practice in multi-parameter compound optimization, Drug Discov. Today 16 (2011) 555e568. [39] J.M. Stevenson, P.D. Mulready, Pipeline pilot 2.1, by Scitegic, J. Am. Chem. Soc. 125 (2003) 1437-1438. Pipeline Pilot 2.1 By Scitegic, 9665 Chesapeake Drive, Suite 401, San Diego, CA 92123-1365, http://www.scitegic.com. [40] Web address: http://www.prous.com/integrity (accessed 10.12.10).

P. Bonnet / European Journal of Medicinal Chemistry 54 (2012) 679e689 [41] T.I. Oprea, A.M. Davis, S.J. Teague, P.D. Leeson, Is there a difference between leads and drugs? a historical perspective, J. Chem. Inf. Comput. Sci. 41 (2001) 1308e1315. [42] G.W. Bemis, M.A. Murcko, The properties of known drugs. 1. Molecular frameworks, J. Med. Chem. 39 (1996) 2887e2893. [43] Web addresses: http://www.maybridge.com, https://www.sigmaaldrich.com, http://www.acros.be. [44] F. Lovering, J. Bikker, C. Humblet, Escape from flatland increasing saturation as an approach to improving clinical success, J. Med. Chem. 52 (2009) 6752e6759. [45] T. Katsuki, K.B. Sharpless, The first practical method for asymmetric epoxidation, J. Am. Chem. Soc. 102 (1980) 5974e5976. [46] M.D. Hack, D.N. Rassokhin, C. Buyck, M. Seierstad, A. Skalkin, P. ten Holte, T.K. Jones, T. Mirzadegan, D.K. Agrafiotis, Library enhancement through the wisdom of crowds, J. Chem. Inf. Model. 51 (2011) 3275e3286. [47] S. Jain, S.B. Bari, N. Rajnikant, De novo design and optimization of novel inhibitor molecules for HIV integrase, Int. J. ChemTech. Res. 3 (2011) 716e723. [48] T. Sundarapandian, J. Shalini, S. Sugunadevi, L.K. Woo, Docking-enabled pharmacophore model for histone deacetylase 8 inhibitors and its application in anti-cancer drug discovery, J. Mol. Graphics Modell. 29 (2010) 382e395. [49] C.H.T.d.P. Da Silva, V.B. da Silva, J. Resende, P.F. Rodrigues, F.C. Bononi, C.G. Benevenuto, C.A. Taft, Computer-aided drug design and ADMET predictions for identification and evaluation of novel potential farnesyltransferase inhibitors in cancer therapy, J. Mol. Graphics Modell. 28 (2010) 513e523. [50] A. Zaliani, K. Boda, T. Seidel, A. Herwig, C.H. Schwab, J. Gasteiger, H. Claußen, C. Lemmen, J. Degen, J. Pärn, M. Rarey, Second-generation de novo design: a view from a medicinal chemist perspective, J. Comput.-Aided Mol. Des. 23 (2009) 593e602. [51] J.B. Baell, G.A. Holloway, New substructure filters for removal of pan assay interference compounds (PAINS) from screening libraries and for their exclusion in bioassays, J. Med. Chem. 53 (2010) 2719e2740.

689

[52] C. Steinbeck, Y. Han, S. Kuhn, O. Horlacher, E. Luttmann, E.L. Willighagen, The chemistry development kit (CDK): an open-source java library for chemo- and bioinformatics, J. Chem. Inf. Comput. Sci. 43 (2003) 493e500. [53] D. Moffat, P. Davis, M. Hutchings, J. Davis, D. Berg, M. Batchelor, J. Johnson, J. O’Connell, R. Martin, T. Crabbe, J. Delgado, M. Perry, 4-Pyridin-5-yl-2-(3,4,5trimethoxyphenylamino)pyrimidines: potent and selective inhibitors of ZAP 70, Bioorg. Med. Chem. Lett. 9 (1999) 3351e3356. [54] P.D. Davis, D.F.C. Moffat, J.M. Davis, M.C. Hutchings, Preparation of substituted 2-anilinopyrimidines as protein kinase inhibitors, Int. Pat. Appl. (1997). WO 1997/19065. [55] J.-P.R.M.A. Bosmans, A.L.G. Meulemans, M.A.J. De Cleyn, H.J.M. Gijsen, Preparation of 4-(aminomethyl)piperidinebenzamides as gastrointestinal agents, Int. Pat. Appl. (2000). WO 2000/037461. [56] J.W.J.F. Thuring, T. Dinklo, A.S.J. Lesage, M.F.L. De Bruyn, W. Zhuang, Preparation of trisubstituted 1,2,4-triazoles and analogs as positive allosteric modulators of alpha 7 nicotinic acetylcholine receptors, Int. Pat. Appl. (2009). WO 2009/050186. [57] G.J. MacDonald, J.W.J.F. Thuring, P.C. Stanislawski, W. Zhuang, Y.E.M. Van Roosbroeck, F.A.M. Van Den Keybus, Trisubstituted 1,2,4-triazoles as nicotinic acetylcholine receptor modulators and their preparation, Int. Pat. Appl. (2009). WO 2009/115547. [58] E.J.E. Freyne, J.F.A. Lacrampe, T.P.S. Perera, P. Ten Holte, Y.A.E. Ligny, D.Y.R. Lardeau, T. Lavrijssen, Preparation of pyrazolopyrimidines as cell-cycle kinase inhibitors, Int. Pat. Appl. (2006). WO 2006/074985. [59] K. Van Emelen, M.G.C. Verdonck, S.F.A. Van Brandt, L.J.J. Backx, Preparation of carbonylamino derivatives as inhibitors of histone deacetylase, Int. Pat. Appl. (2003). WO 2003/076395. [60] K. Van Emelen, Preparation of N-(piperazinylmethyl-, piperidinylmethyl- and morpholinylmethyl) sulfonamides and amides as novel inhibitors of histone deacetylase, Int. Pat. Appl. (2003). WO 2003/076438, Bonnet 42.