Harnessing the potential of natural products in drug discovery from a cheminformatics vantage point

Tiago Rodrigues
Instituto de Medicina Molecular, Faculdade de Medicina da Universidade de Lisboa, Av. Prof. Egas Moniz, 1649-028 Lisboa, Portugal. E-mail: tiago.rodrigues@medicina.ulisboa.pt

Received 31st August 2017 , Accepted 24th October 2017

First published on 31st October 2017

Natural products (NPs) present a privileged source of inspiration for chemical probe and drug design. Despite the biological pre-validation of the underlying molecular architectures and their relevance in drug discovery, the poor accessibility to NPs, complexity of the synthetic routes and scarce knowledge of their macromolecular counterparts in phenotypic screens still hinder their broader exploration. Cheminformatics algorithms now provide a powerful means of circumventing the abovementioned challenges and unlocking the full potential of NPs in a drug discovery context. Herein, I discuss recent advances in the computer-assisted design of NP mimics and how artificial intelligence may accelerate future NP-inspired molecular medicine.

image file: c7ob02193c-p1.tif

Tiago Rodrigues

Tiago Rodrigues received a PharmD in 2006 and a PhD in Medicinal Chemistry in 2010 from the University of Lisbon, Portugal, under the supervision of Dr Francisca Lopes. He then completed a postdoctoral stay in the group of Prof. Gisbert Schneider at the ETH Zürich working on de novo design, microfluidics-assisted synthesis and macromolecular target identification for small molecules. He is currently a Marie Sklodowska-Curie Fellow with Dr Gonçalo Bernardes at the Instituto de Medicina Molecular, working on constructs for the targeted delivery of payloads and machine learning methods for natural product deorphaning.


Natural products (NPs) have a long-standing tradition of offering valuable starting points for development in chemical biology and early drug discovery programs, with particular increased incidence in cancerous and infectious diseases.1–8 They feature intricate molecular frameworks and pharmacophore arrangements that have been biologically pre-validated as protein binding motifs.9–13 Curiously, many naturally occurring chemotypes remain largely uncharted, suggesting that innovative chemical matter can be rationally designed in NP chemical space.5 Considering the poor accessibility to NPs and the general complexity of their synthetic routes, the manual design of synthetically tractable and focused NP-inspired libraries has risen as a viable alternative – a paradigm coined as biology-oriented synthesis (BIOS).14 On numerous occasions, the BIOS strategy has shown that building biologically-relevant chemical matter by grafting NP scaffolds into synthetic moieties can improve screening hit rates while unravelling hitherto unexploited chemical space.9 In this regard, scoring compound collections for their NP-likeness may find applicability in prioritizing chemical matter.15 Indeed, pharmaceutical companies have benefitted appreciably from such a strategy in development pipelines, taking into account the number of FDA-approved drugs incorporating meaningful NP fragments.5

Meanwhile, computer-assisted drug design has established itself as a hallmark for rational drug design, providing an array of different tools that build on the ever-increasing protein or ligand data.16 For example, molecular docking or similarity searches can be deployed to mine an enumerated fraction of chemical space in a time-efficient manner and increase the likelihood of success in follow up target- and/or cell-based assays.17–19 Above all, virtual screening methods offer an expeditious means of prioritizing chemical matter, and avoid full-deck screens to identify starting points for optimization. More recently, computer-assisted de novo design has re-entered the drug design toolbox as a consequence of improved algorithms, hardware and storage capacity.20,21 Most importantly, it does not rely on fully enumerated virtual compound libraries, as molecules are designed “on-the-fly”, i.e. as the search algorithm scavenges for activity islands.22–24 It is thus obvious that recent advances in methodology and concepts of computer-assisted drug design advocate for its applicability in guiding and leveraging NP-inspired drug discovery. Herein, I review modern early discovery approaches developed at the interface of computational and experimental medicinal chemistry. I discuss methods for visualizing chemical data in the NP realm, as well as algorithms for designing NP mimics. I also provide select examples of the use of machine learning methods for target identification that may serve as platform for future NP research.

Are natural products different from typical screening compounds?

Yes. It is widely recognized that NPs typically have a significantly larger fraction of sp3-hybridized bridgehead atoms10,25 and, consequently, stereogenic centres compared to purely synthetic small molecules. While synthetic pathways to NPs are commonly not trivial, biasing screening collections towards three-dimensionality and away from “flatland”26 contributes to the “chemical beauty”27 of the underlying molecular architectures and a generally lower promiscuous target binding behaviour.28 NPs do also favour aliphatic ring systems over aromatics, and feature higher oxygen/lower nitrogen atom contents compared to synthetic molecules.10 In fact, ca. 50% of NPs annotated in the Dictionary of Natural Products (DNP) do not have substructural counterparts among synthetic molecules.5

For example, 1000 random NPs and commercially available compounds collected from the TCM@Taiwan29 and ZINC1530 databases, respectively, display notorious structural dissimilarity, as attested from the projection of an underlying high dimensional space to the plane, while preserving the data points’ pairwise Euclidean distances (Fig. 1). ZINC15 compounds (in red) occupy a confined region of the chemical descriptor space, in line with the limited structural diversity. On the other hand, the NPs (in blue) cover a much larger region of the descriptor space, including that of ZINC15 compounds. For the purpose of in vitro screening, it is thus logical that even a small amount of appropriately selected NPs will sample numerous scaffolds and a larger slice of chemical space. Importantly, clustering and data visualization methods provide an intuitive means of selecting candidate molecules for screening.31,32 Moreover, a collection of resources comprising accessible NPs was recently made available with the aim of spurring nature-inspired research.33 How these resources can be used to motivate computer-based molecular design is discussed henceforth.

image file: c7ob02193c-f1.tif
Fig. 1 Projection (multidimensional scaling, MDS) of random NPs (n = 1000, blue) and commercially available compounds (n = 1000, red) from the TCM@Taiwan and ZINC15 databases, respectively. Exemplary NPs are highlighted. Descriptor set as implemented in RDKit (http://www.rdkit.org) and MOE v2016.08 (Chemical Computing Group, Canada): NumRotatableBonds, NumHeteroAtoms, NumRings, NumAromaticRings, NumSaturatedRings, NumAliphaticRings, NumAromaticHetercycles, NumSaturatedHeterocycles, NumAliphaticHeterocycles, NumAromaticCarbocycles, NumSaturatedCarbocycles, NumAliphaticCarbocycles, FractionCSP3 (RDKit), a_nN and a_nO (MOE). Plot computed with Python 2.7.13.

Deconvolution and visualization of natural product scaffolds

Cheminformatics tools offer a plethora of solutions to exploit both synthetic and naturally occurring scaffolds. Such an arsenal largely aids the rational design of potentially bioactive matter. Of marked relevance are tools that leverage visualization of scaffold distribution in a global context, facilitating interpretation of complex bioactivity traits and mining chemical space.31,34 For example, the automated deconvolution of intricate parent NPs into substructures, i.e. child entities, has proven to be a highly effective approach to access biologically meaningful compounds and prioritize molecular frameworks to inspire molecular design.35 More specifically, the Scaffold Hunter algorithm36,37 (Java-based open source tool downloadable at http://scaffoldhunter.sourceforge.net) organizes chemical diversity into scaffold trees by stepwise mapping of increasingly simpler frameworks – a process coined as “brachiation” (Fig. 2). In a particular application, the algorithm arranged the DNP from biology and synthetic tractability vantage points, in what is currently known as the Periodic Table of Natural Products.11 The method, lends an invaluable source of starting points for fragment-based drug discovery programs, but as in any other approaches, does not come without caveats; the gradual loss of bioactivity by smaller NP-derived fragments is an expected consequence of the recursive scaffold simplification, eventually leading to biological fingerprint decharacterization.11
image file: c7ob02193c-f2.tif
Fig. 2 Schematics of a dummy scaffold tree generated with Scaffold Hunter. Scaffold tree of compound 1 is given as example. Select Murcko scaffolds are highlighted.

In a proof-of-concept study, the approach was effective, with 19% of the designed and experimentally tested marine product dysidiolide-inspired molecules displaying the desired functional inhibition of 11β-hydroxysteroid dehydrogenase (11βHSD, Fig. 3a).11 Additionally, the in silico scaffold deconvolution offered a platform to the rational design of 11β-hydroxysteroid dehydrogenase subtype-selective entities across the developed chemical series. While a single example may not be enough for solid proof-of-concept, such paradigm appears to be primed for increasing screening hit rates. Waldmann and co-workers further elaborated on their hierarchical bioactivity-guided analysis technology to attain potent (IC50 = 3 ± 1 μM), selective and NP-inspired 5-lipoxygenase (5-LO) modulators (Fig. 3b).37 To that end, the automated brachiation process was applied to a 7-membered ring chemotype with a parent–child series annotated to 5-LO bioactivity. The fragment-based approach allowed the identification of minimal structural requirements for developing synthetically accessible and lignan-derived 5-LO inhibitors with suitable ligand efficiency.37 More recently, mapping of the NP-derived fragment space in the DNP afforded hundreds of thousands of possible starting points decorated with chemical handles to inspire medicinal chemistry efforts towards the design of new chemical entities.38 From those, a diverse array of fragment-like molecules were obtained and experimentally validated as p38α MAP kinase and phosphatase inhibitors, via functional assays and X-ray crystallography (Fig. 3c). Not surprisingly, only weak or moderate target inhibition was achieved with the NP-derived fragments, but at suitable ligand efficiencies these fragments may be further developed into appropriate chemical probes.38 But how can one accomplish it?

image file: c7ob02193c-f3.tif
Fig. 3 Cheminformatics-enabled biology-oriented synthesis of drug target modulators. (a) Development of a selective 11β-hydroxysteroid dehydrogenase 1 (11βHSD1) inhibitor inspired on minimal dysidiolide scaffold requirements. Retrosynthesis is depicted. (b) Identification of a lignan-inspired 5-lipoxygenase (5-LO) inhibitor. (c) Natural product derived fragments from the Dictionary of Natural Products as source of inspiration for drug discovery programs. Select examples of kinase and phosphatase inhibitors are shown.

De novo design in natural product space

As previously discussed, NP-derived fragments have served as seeds for developing fully-grown bioactive agents.39,40 However, harnessing their promise in chemical biology and drug discovery via the automated generation of NP mimics has been a far from trivial task up-to-date. The poor synthetic accessibility of the proposed constructs due to difficult-to-build atom connections, coupled to poor drug-likeness of the designed chemotypes has been a major bottleneck for the general deployment of computer-assisted de novo design in discovery programs, i.e. design of chemical entities from scratch.41 Recently, the evolution of the design paradigm to fragment- and reaction-driven algorithms has decisively contributed to mitigating the aforementioned tractability limitations.20 In 2015, Lanz and Riedl reported the receptor-based de novo design, synthesis and biological evaluation of uracil-inspired entities with tailored affinity for the matrix metalloproteinase-13 (MMP-13, Fig. 4) over the MMP-1, -2, -3, -7, -8, -9, -12 and -14 counterparts.42 The identification of hypothetical anchor points to MMP-13 in uracil, the multiple chemical handles for further fragment growth, together with careful planning of the optimal ligand–target recognition features were key for success.
image file: c7ob02193c-f4.tif
Fig. 4 Receptor-based de novo design of uracil-inspired MMP-13-selective inhibitor.

On the other hand, using a conceptually orthogonal approach, the ligand-based computational tool DOGS (Design of Genuine Structures, Insili.com)43 generates a library of molecules with identical pharmacophore features to the seed/template structure. In short, DOGS uses a fragment- and reaction-based approach leveraged by ca. 25[thin space (1/6-em)]000 curated and commercially available building blocks. These are virtually reacted in deterministic and stepwise fashion using up to 58 unique organic reaction rules to generate isofunctional molecules. The implemented chemistry toolbox is diverse and includes heterocyclic chemistry (e.g. Pictet–Spengler tetraisoquinoline, aminothiazole, benzoxazole synthesis) condensations (e.g. Claisen, Dieckmann), reductive amination, ether, ester and amide bond formations, cycloadditions (e.g. Diels Alder, Huisgen 1,3-cycloaddition) and cross couplings (e.g. Suzuki, Negishi, Heck).20 The de novo design protocol starts with a user-defined number of starting fragments for chemical space sampling, whereby pre-installed chemical handles work as attachment points for virtual combinatorial exploration using one reaction rule at a time. The resulting virtual compounds are scored via topological (2D) graph similarity to the template structure in terms of pharmacophore features (Fig. 5). More specifically, the graph kernel method may have different user-defined abstraction levels and compares hydrogen-bond donor/acceptor, positive/negative charge, aromatic, lipophilic, or no type features between molecules.44 The best scoring candidate is then selected for subsequent iterations. The recursive design is only completed when the virtual molecule reaches 100 ± 30% of the molecular weight of the template structure, or alternatively, with a user-defined number of virtual synthetic steps. Moreover, DOGS performs pseudo-retrosyntheses to suggest a synthetic pathway from the commercially available building blocks to the designed chemical entities.43 Potentially, the DOGS algorithm could be improved if an approach similar to the Reaxys® Synthesis Planner is implemented, which allows for streamlined reaction condition planning.

image file: c7ob02193c-f5.tif
Fig. 5 Application of DOGS. (a) Example of de novo designs mimicking the sesquiterpene anticancer agent (−)-englerin A. A molecular graph is built prior to the computational design. Each node in the graph corresponds to a pharmacophoric feature. The structures of two manually modified, synthesized and experimentally tested compounds are given. Bz = C6H5CH2. (b) Pharmacophore alignment between the template structure (dark grey) and the original designs (light grey), as computed from MOE v2016.08. Features: hydrophobic (green), hydrogen-bond acceptor (cyan) and aromatic/hydrophobic (orange).

With the DOGS technology in hand, Friedrich et al. designed and synthesized mimics of the anticancer sesquiterpene NP (−)-englerin A – a modulator of TRP channels (Fig. 5a).45,46 From a total of 903 designs, two of them were prioritized given the pharmacophoric similarity to the template, yet diverging molecular structures (Fig. 5b). Of note, the design algorithm does not take into account stereochemistry, which is a pending limitation of the method, considering its importance in ligand–target recognition. On the other hand the ‘fuzziness’ of the molecular graph descriptions allows for interchangeability of similar hydrophobic features without molecular structure adulteration from an algorithm's point of view. Thus, for the ease of synthesis, both designs were manually modified (Fig. 5a). The target compounds were readily obtained in moderate yields. In one case, acylation of an Evans’ chiral auxiliary followed by Suzuki cross coupling afforded the potential TRP modulator in two steps. In a second case, N-capping of 3-bromophenylalanine ester, followed by installation of a furanyl moiety via cross coupling yielded a racemic mixture of the designed compound.45

Both small molecules presented identical TRPM8 antagonism to the parent NP, and most importantly, provide initial proof-of-concept for the ligand-based design of synthetically accessible NP mimics. While other de novo design algorithms have been reported, to the best of my knowledge no direct application to NPs has been made. Nonetheless, a similar and commercially available tool – CoLibri (BioSolveIT GmbH) – can be expected to afford synthetically viable and bioactive mimics. Obviously, such design efforts are only reasonable when the drug targets have been characterized for the parent NP – a far from trivial ask in drug discovery and subject to several current studies.

Target identification and deorphanization of natural products

While phenotypic screens are a privileged means of identifying pathophysiology modulators with moderately acceptable physicochemical properties, their development into suitable chemical probes or drug leads is facilitated manifold if the engaged drug targets are known. However, the identification of drug targets for phenotypic screening hits is arguably one of the most demanding tasks in drug discovery. While in some cases, phenotype changes can be associated with certain pathways or drug targets, the ligand–target links are not always obvious.47 Chemoproteomic or “pull-down” assays of protein counterparts are an established method to identify drug targets.48 Though, they require chemical modification of the ligand at the risk of diminishing affinity for the relevant binding partners. Additionally, the technology is laborious and pull-down of membrane proteins is unlikely, at best, given the low protein content in some cases and the need of a membrane to stabilize the native, functional protein conformation.

Unfortunately, the rational deployment of NPs in drug discovery is commonly hampered by the lack of factual knowledge of which targets are modulated. Determinately, several in silico methods have surfaced with the aim of expediting drug target identification.49–56 As reference, a comprehensive list of webservers with potential application for NP target identification has recently been compiled.57 For example, 3D pharmacophore models in x,y,z coordinate space provide a means for interrogating NP targets and mapping the spatial and electrostatic requirements in putative ligand–target recognition.58 In an application, screening of secondary metabolites from Ruta graveolans with over 2000 pharmacophore models, followed by experimental testing provided insights into the binding partners of arborinine and rutamarin.59 Similarly, morphinans and natural isoquinolines were identified as moderate effectors of acetylcholinesterase.60 Moreover, physodic acid and perlatolic acid were unveiled as nanomolar-potent microsomal prostaglandin E synthease-1 (mPGES-1) inhibitors.61 Molecular docking can also find application in this particular task despite the higher computational cost and scoring functions’ caveats. In one example, inverse docking, i.e. docking a database of targets against a single query ligand, was effective in identifying cyclooxygenase-2 (COX-2) and the peroxisome proliferator-activated receptor gamma (PPARγ) as targets for meranzin.62

Dedicated software tools have also been developed to predict ligand–target relationships with minimal computational effort. One of such tools is PASS49,63 that predicts biological activities from topological fragment structure descriptors.64 Its application to >90 marine sponge alkaloids revealed potential anti-tumoral activity for a large proportion (80%) of the studied NPs.63 Naturally, said predictions require experimental validation and downstream target confirmation efforts. On the other hand, the Similarity Ensemble Approach (SEA)51 relates drug targets by ligand topology similarity. Moreover, it uses a statistical model to rank the significance of the calculated substructure similarity scores. While it has been most productively applied to synthetic small molecules, it was also prospectively used to identify the anti-plasmodial activity of physalins isolated from Physalis angulata.65

Target inference in these tools relies on chemical substructure similarity between the studied molecule and the reference ligands that are annotated to a macromolecular target, drug target sequence/structure or pathway information. As such, these algorithms have been explicitly tailored to analyse rather similar molecules, but not structurally intricate NPs. Given the divergent structures of NPs, the aforementioned computational tools generally fail to provide confident predictions and, therefore, macromolecular targets may remain elusive. The chemocentric Self-Organizing Maps (SOM)-based prediction of drug equivalence relationships (SPiDER) software addresses such limitations by using dual molecular representations: physicochemical properties and topological pharmacophores.66 The CATS2 descriptor used in SPiDER is ideally suited for target inference as it calculates pharmacophore feature-pair correlations from topological molecular graphs.67 While representing in an unbiased manner the relevant features for target engagement, the descriptor is sufficiently ‘fuzzy’ to allow for recognition of subtle patterns between molecules that are oblivious to commonly used substructural fingerprints.66 Hence, SPiDER was originally designed and tailored to confidently infer drug targets in the absence of structural similarity, such as in computationally de novo designed entities. In short, the unsupervised SOM algorithm tessellates the reference ligand space into clusters of molecules representing local neighborhoods. The query ligands are then assigned to one specific cluster of the map to offer a target counterpart. The operation is performed independently for each descriptor set and subsequently combined in a jury approach to yield a consensus score. By analysing the background distribution of consensus scores, a significance value (p value) for the qualitatively predicted targets is estimated (Fig. 6a).66,68 Building on prior successes, the SPiDER concept was further extended to structurally complex NPs. Albeit generally applicable, the method has limitations (e.g. small number of predictable targets) and is likely to provide null predictions in some cases. For example, only potassium channels are predicted for compound 1 at the threshold of significance (p = 0.048, Fig. 6b; http://www.cadd.ethz.ch/software/spider.html). The result may not be entirely surprising. Arguably, the calculated descriptors for complex NPs and smaller/drug-like molecules, such as those in the reference ligand database, will be distinct, taking into account the differences in atom content and connectivity. With that in mind, confident predictions could be obtained for archazolid A (Fig. 6b), using its derived fragments as activity blueprints for target prediction. The NP-derived fragment strategy allowed the discovery of new biology, e.g. activation of FXR, PPARγ and mPGES-1, that is potentially related to the already known anticancer activity of the macrocyclic NP.32 In similar fashion, the anticancer agent doliculide was deorphanized as an extremely potent prostanoid EP3 antagonist (IC50 = 16 nM), using synthetic intermediates for an ensemble prediction.69 The SPiDER method proved to be equally effective to discover modulation of calcium Cav1.2 channels as an off-target of (−)-englerin A. The approach included using a fragment-like NP as a proxy for predictions, whenever SPiDER does not afford confident ligand–target relationships for the parent compound.70 Unlike previous applications, the discovery of such an association was made using the publically available webserver. SPiDER was also efficiently employed to fragment-like NPs that provide biologically-motivated starting points for optimization. It was not only used to unveil weak modulation of the kappa opioid G-protein coupled receptor by sparteine,5 but also modulation of serotonin receptor 2B (5-HT2B) and platelet-derived growth factor receptor α (PDGFRα) by graveolinine and isomacroin, respectively.71 Significantly, the study provided a rationale to link 5-HT2B and cyclooxygenase-2 (COX-2) biology via graveolinine.

image file: c7ob02193c-f6.tif
Fig. 6 Target identification for natural products using SPiDER. (a) Schematics of the SPiDER workflow. (b) Examples of prospective applications of SPiDER.

More recently, inspired by the architecture of SPiDER, Schneider et al. reported the Target Inference Generator (TIGER) also using a consensus of two SOMs and topological pharmacophore descriptors as algorithm workhorses.72 With a capability of inferring up to 331 ligand–target relationships the software tool successfully identified the glucocorticoid, cholecystokinin 2 and orexin-1/2 receptors as target counterparts for the marine NP (±)-marinopyrrole A (Fig. 7). Moreover, TIGER successfully linked resveratrol (Fig. 7) to modulation of AMP-activated protein kinase (AMPK, IC50 = 10 μM), estrogen receptors α (ERα, Ki = 4 μM) and β (ERβ, Ki = 0.4 μM).73

image file: c7ob02193c-f7.tif
Fig. 7 Examples of natural products deorphanized with the TIGER software.


Nature provides a rich source of yet untapped chemical matter for drug discovery endeavours. On the other hand, the advent of advanced and innovative design algorithms, provides a unique platform to efficiently explore the vast NP chemical space, while cost-effectively tackling limitations in NP research. Significantly, storage and software limitations no longer hinder full-fledged and productive applications of such algorithms in molecular medicine. Hence, harnessing the potential of NPs by coalescing them with smart design tools appears an expected evolution of their exploitation process towards sustainable future drug discovery.

Chemocentric methods are incomparably faster than receptor-based approaches without sacrificing accuracy. As a feature, they also allow for swift and seamless data integration as it becomes available. In this regard, it is worth noting that the purpose of cheminformatics is to generate viable and motivated research hypotheses that require posterior experimental validation. Consequently, numerous failures are not only acceptable but also anticipated. Together with the obtained true hits, the generated negative data is in high demand to fuel the gradually popular machine learning algorithms. In particular, I expect that the conceptually opposing “big data”-driven deep learning and data-economical active learning variants will soon see novel uses in NP research. Most importantly, the impending potential of artificial intelligence in NP-inspired drug discovery is far from fulfilled, and advances in this field will shed light onto translational applications of many unexplored chemical entities.

Conflicts of interest

There are no conflicts to declare.


The author is a Marie Sklodowska-Curie Fellow (Grant Agreement 743640).

Notes and references

  1. G. Molinari, Adv. Exp. Med. Biol., 2009, 655, 13–27 CrossRef CAS PubMed .
  2. F. E. Koehn and G. T. Carter, Nat. Rev. Drug Discovery, 2005, 4, 206–220 CrossRef CAS PubMed .
  3. A. L. Harvey, Drug Discovery Today, 2008, 13, 894–901 CrossRef CAS PubMed .
  4. E. Patridge, P. Gareiss, M. S. Kinch and D. Hoyer, Drug Discovery Today, 2016, 21, 204–207 CrossRef CAS PubMed .
  5. T. Rodrigues, D. Reker, P. Schneider and G. Schneider, Nat. Chem., 2016, 8, 531–541 CrossRef CAS PubMed .
  6. D. J. Newman and G. M. Cragg, J. Nat. Prod., 2016, 79, 629–661 CrossRef CAS PubMed .
  7. A. L. Harvey, R. Edrada-Ebel and R. J. Quinn, Nat. Rev. Drug Discovery, 2015, 14, 111–129 CrossRef CAS PubMed .
  8. C. R. Pye, M. J. Bertin, R. S. Lokey, W. H. Gerwick and R. G. Linington, Proc. Natl. Acad. Sci. U. S. A., 2017, 114, 5601–5606 CrossRef CAS PubMed .
  9. H. van Hattum and H. Waldmann, J. Am. Chem. Soc., 2014, 136, 11853–11859 CrossRef CAS PubMed .
  10. M. L. Lee and G. Schneider, J. Comb. Chem., 2001, 3, 284–289 CrossRef CAS PubMed .
  11. M. A. Koch, A. Schuffenhauer, M. Scheck, S. Wetzel, M. Casaulta, A. Odermatt, P. Ertl and H. Waldmann, Proc. Natl. Acad. Sci. U. S. A., 2005, 102, 17272–17277 CrossRef CAS PubMed .
  12. L. Laraia, K. Ohsawa, G. Konstantinidis, L. Robke, Y. W. Wu, K. Kumar and H. Waldmann, Angew. Chem., Int. Ed., 2017, 56, 2145–2150 CrossRef CAS PubMed .
  13. M. Perni, C. Galvagnion, A. Maltsev, G. Meisl, M. B. Muller, P. K. Challa, J. B. Kirkegaard, P. Flagmeier, S. I. Cohen, R. Cascella, S. W. Chen, R. Limboker, P. Sormanni, G. T. Heller, F. A. Aprile, N. Cremades, C. Cecchi, F. Chiti, E. A. Nollen, T. P. Knowles, M. Vendruscolo, A. Bax, M. Zasloff and C. M. Dobson, Proc. Natl. Acad. Sci. U. S. A., 2017, 114, E1009–E1017 CrossRef CAS PubMed .
  14. S. Wetzel, R. S. Bon, K. Kumar and H. Waldmann, Angew. Chem., Int. Ed., 2011, 50, 10800–10826 CrossRef CAS PubMed .
  15. P. Ertl, S. Roggo and A. Schuffenhauer, J. Chem. Inf. Model., 2008, 48, 68–74 CrossRef CAS PubMed .
  16. G. Schneider, Nat. Rev. Drug Discovery, 2010, 9, 273–276 CrossRef CAS PubMed .
  17. D. B. Kitchen, H. Decornez, J. R. Furr and J. Bajorath, Nat. Rev. Drug Discovery, 2004, 3, 935–949 CrossRef CAS PubMed .
  18. J. J. Irwin and B. K. Shoichet, J. Med. Chem., 2016, 59, 4103–4120 CrossRef CAS PubMed .
  19. M. A. Miller, Nat. Rev. Drug Discovery, 2002, 1, 220–227 CrossRef CAS PubMed .
  20. T. Rodrigues and G. Schneider, Synlett, 2014, 170–178 CAS .
  21. P. Schneider and G. Schneider, J. Med. Chem., 2016, 59, 4077–4086 CrossRef CAS PubMed .
  22. T. Rodrigues, T. Kudoh, F. Roudnicky, Y. F. Lim, Y. C. Lin, C. P. Koch, M. Seno, M. Detmar and G. Schneider, Angew. Chem., Int. Ed., 2013, 52, 10006–10009 CrossRef CAS PubMed .
  23. T. Rodrigues, F. Roudnicky, C. P. Koch, T. Kudoh, D. Reker, M. Detmar and G. Schneider, Chem. Sci., 2013, 4, 1229–1233 RSC .
  24. T. Rodrigues, N. Hauser, D. Reker, M. Reutlinger, T. Wunderlin, J. Hamon, G. Koch and G. Schneider, Angew. Chem., Int. Ed., 2015, 54, 1551–1555 CrossRef CAS PubMed .
  25. T. Henkel, R. M. Brunne, H. Müller and F. Reichel, Angew. Chem., Int. Ed., 1999, 38, 643–646 CrossRef CAS .
  26. F. Lovering, J. Bikker and C. Humblet, J. Med. Chem., 2009, 52, 6752–6756 CrossRef CAS PubMed .
  27. G. R. Bickerton, G. V. Paolini, J. Besnard, S. Muresan and A. L. Hopkins, Nat. Chem., 2012, 4, 90–98 CrossRef CAS PubMed .
  28. P. A. Clemons, N. E. Bodycombe, H. A. Carrinski, J. A. Wilson, A. F. Shamji, B. K. Wagner, A. N. Koehler and S. L. Schreiber, Proc. Natl. Acad. Sci. U. S. A., 2010, 107, 18787–18792 CrossRef CAS PubMed .
  29. C. Y. Chen, PLoS One, 2011, 6, e15939 CAS .
  30. T. Sterling and J. J. Irwin, J. Chem. Inf. Model., 2015, 55, 2324–2337 CrossRef CAS PubMed .
  31. T. Miyao, D. Reker, P. Schneider, M. Funatsu and G. Schneider, Planta Med., 2015, 81, 429–435 CrossRef CAS PubMed .
  32. D. Reker, A. M. Perna, T. Rodrigues, P. Schneider, M. Reutlinger, B. Monch, A. Koeberle, C. Lamers, M. Gabler, H. Steinmetz, R. Muller, M. Schubert-Zsilavecz, O. Werz and G. Schneider, Nat. Chem., 2014, 6, 1072–1078 CrossRef CAS PubMed .
  33. Y. Chen, C. K. d. Bruyn and J. Kirchmair, J. Chem. Inf. Model., 2017, 57, 2099–2111 CrossRef CAS PubMed .
  34. J. Larsson, J. Gottfries, S. Muresan and A. Backlund, J. Nat. Prod., 2007, 70, 789–794 CrossRef CAS PubMed .
  35. H. Lachance, S. Wetzel, K. Kumar and H. Waldmann, J. Med. Chem., 2012, 55, 5989–6001 CrossRef CAS PubMed .
  36. S. Wetzel, K. Klein, S. Renner, D. Rauh, T. I. Oprea, P. Mutzel and H. Waldmann, Nat. Chem. Biol., 2009, 5, 581–583 CrossRef CAS PubMed .
  37. S. Renner, W. A. van Otterlo, M. Dominguez Seoane, S. Mocklinghoff, B. Hofmann, S. Wetzel, A. Schuffenhauer, P. Ertl, T. I. Oprea, D. Steinhilber, L. Brunsveld, D. Rauh and H. Waldmann, Nat. Chem. Biol., 2009, 5, 585–592 CrossRef CAS PubMed .
  38. B. Over, S. Wetzel, C. Grutter, Y. Nakai, S. Renner, D. Rauh and H. Waldmann, Nat. Chem., 2013, 5, 21–28 CrossRef CAS PubMed .
  39. E. A. Crane and K. Gademann, Angew. Chem., Int. Ed., 2016, 55, 3882–3902 CrossRef CAS PubMed .
  40. P. Burch, A. Chicca, J. Gertsch and K. Gademann, ACS Med. Chem. Lett., 2014, 5, 172–177 CrossRef CAS PubMed .
  41. G. Schneider and U. Fechner, Nat. Rev. Drug Discovery, 2005, 4, 649–663 CrossRef CAS PubMed .
  42. J. Lanz and R. Riedl, ChemMedChem, 2015, 10, 451–454 CrossRef CAS PubMed .
  43. M. Hartenfeller, H. Zettl, M. Walter, M. Rupp, F. Reisen, E. Proschak, S. Weggen, H. Stark and G. Schneider, PLoS Comput. Biol., 2012, 8, e1002380 CAS .
  44. M. Rupp, E. Proschak and G. Schneider, J. Chem. Inf. Model., 2007, 47, 2280–2286 CrossRef CAS PubMed .
  45. L. Friedrich, T. Rodrigues, C. S. Neuhaus, P. Schneider and G. Schneider, Angew. Chem., Int. Ed., 2016, 55, 6789–6792 CrossRef CAS PubMed .
  46. Y. Akbulut, H. J. Gaunt, K. Muraki, M. J. Ludlow, M. S. Amer, A. Bruns, N. S. Vasudev, L. Radtke, M. Willot, S. Hahn, T. Seitz, S. Ziegler, M. Christmann, D. J. Beech and H. Waldmann, Angew. Chem., Int. Ed., 2015, 54, 3787–3791 CrossRef CAS PubMed .
  47. L. Laraia and H. Waldmann, Drug Discovery Today: Technol., 2017, 23, 75–82 CrossRef PubMed .
  48. R. E. Moellering and B. F. Cravatt, Chem. Biol., 2012, 19, 11–22 CrossRef CAS PubMed .
  49. P. V. Pogodin, A. A. Lagunin, D. A. Filimonov and V. V. Poroikov, SAR QSAR Environ. Res., 2015, 26, 783–793 CrossRef CAS PubMed .
  50. J. Nickel, B. O. Gohlke, J. Erehman, P. Banerjee, W. W. Rong, A. Goede, M. Dunkel and R. Preissner, Nucleic Acids Res., 2014, 42, W26–W31 CrossRef CAS PubMed .
  51. M. J. Keiser, B. L. Roth, B. N. Armbruster, P. Ernsberger, J. J. Irwin and B. K. Shoichet, Nat. Biotechnol., 2007, 25, 197–206 CrossRef CAS PubMed .
  52. M. Reutlinger, T. Rodrigues, P. Schneider and G. Schneider, Angew. Chem., Int. Ed., 2014, 53, 582–585 CrossRef CAS PubMed .
  53. A. A. Antolin, X. Jalencas, J. Yelamos and J. Mestres, ACS Chem. Biol., 2012, 7, 1962–1967 CrossRef CAS PubMed .
  54. S. Z. Mohamad Zobir, F. Mohd Fauzi, S. Liggi, G. Drakakis, X. Fu, T. P. Fan and A. Bender, J. Evidence-Based Complementary Altern. Med., 2016, 2016, 2106465 Search PubMed .
  55. F. Mohd Fauzi, A. Koutsoukas, R. Lowe, K. Joshi, T. P. Fan, R. C. Glen and A. Bender, J. Chem. Inf. Model., 2013, 53, 661–673 CrossRef CAS PubMed .
  56. X. Fu, L. H. Mervin, X. Li, H. Yu, J. Li, S. Z. Mohamad Zobir, A. Zoufir, Y. Zhou, Y. Song, Z. Wang and A. Bender, J. Chem. Inf. Model., 2017, 57, 468–483 CrossRef CAS PubMed .
  57. E. Sam and P. Athri, Brief Bioinform., 2017 DOI:10.1093/bib/bbx1125 .
  58. J. M. Rollinger, Phytochem. Lett., 2009, 2, 53–58 CrossRef CAS .
  59. J. M. Rollinger, D. Schuster, B. Danzl, S. Schwaiger, P. Markt, M. Schmidtke, J. Gertsch, S. Raduner, G. Wolber, T. Langer and H. Stuppner, Planta Med., 2009, 75, 195–204 CrossRef CAS PubMed .
  60. D. Schuster, M. Spetea, M. Music, S. Rief, M. Fink, J. Kirchmair, J. Schutz, G. Wolber, T. Langer, H. Stuppner, H. Schmidhammer and J. M. Rollinger, Bioorg. Med. Chem., 2010, 18, 5071–5080 CrossRef CAS PubMed .
  61. J. Bauer, B. Waltenberger, S. M. Noha, D. Schuster, J. M. Rollinger, J. Boustie, M. Chollet, H. Stuppner and O. Werz, ChemMedChem, 2012, 7, 2077–2081 CrossRef CAS PubMed .
  62. Q. T. Do, C. Lamy, I. Renimel, N. Sauvan, P. Andre, F. Himbert, L. Morin-Allory and P. Bernard, Planta Med., 2007, 73, 1235–1240 CrossRef CAS PubMed .
  63. A. Lagunin, D. Filimonov and V. Poroikov, Curr. Pharm. Des., 2010, 16, 1703–1717 CrossRef CAS PubMed .
  64. D. Filimonov, V. Poroikov, Y. Borodina and T. Gloriozova, J. Chem. Inf. Comput. Sci., 1999, 39, 666–670 CrossRef CAS .
  65. M. S. Sa, M. N. de Menezes, A. U. Krettli, I. M. Ribeiro, T. C. Tomassini, R. Ribeiro dos Santos, W. F. de Azevedo Jr. and M. B. Soares, J. Nat. Prod., 2011, 74, 2269–2272 CrossRef CAS PubMed .
  66. D. Reker, T. Rodrigues, P. Schneider and G. Schneider, Proc. Natl. Acad. Sci. U. S. A., 2014, 111, 4067–4072 CrossRef CAS PubMed .
  67. M. Reutlinger, C. P. Koch, D. Reker, N. Todoroff, P. Schneider, T. Rodrigues and G. Schneider, Mol. Inf., 2013, 32, 133–138 CrossRef CAS PubMed .
  68. D. Reker, M. Seet, M. Pillong, C. P. Koch, P. Schneider, M. C. Witschel, M. Rottmann, C. Freymond, R. Brun, B. Schweizer, B. Illarionov, A. Bacher, M. Fischer, F. Diederich and G. Schneider, Angew. Chem., Int. Ed., 2014, 53, 7079–7084 CrossRef CAS PubMed .
  69. G. Schneider, D. Reker, T. Chen, K. Hauenstein, P. Schneider and K. H. Altmann, Angew. Chem., Int. Ed., 2016, 55, 12408–12411 CrossRef CAS PubMed .
  70. T. Rodrigues, F. Sieglitz, V. J. Somovilla, P. M. Cal, A. Galione, F. Corzana and G. J. Bernardes, Angew. Chem., Int. Ed., 2016, 55, 11077–11081 CrossRef CAS PubMed .
  71. T. Rodrigues, D. Reker, J. Kunze, P. Schneider and G. Schneider, Angew. Chem., Int. Ed., 2015, 54, 10516–10520 CrossRef CAS PubMed .
  72. P. Schneider and G. Schneider, Chem. Commun., 2017, 53, 2272–2274 RSC .
  73. P. Schneider and G. Schneider, Angew. Chem., Int. Ed., 2017, 56, 11520–11524 CrossRef CAS PubMed .

This journal is © The Royal Society of Chemistry 2017