5.1 Introduction

For the virtual screening-based discovery of bioactive compounds, pharmacophore models have been successfully used for three decades now. However, the concept of a pharmacophore, then also called toxicophore or haptophore, was first defined by Ehrlich [1] and later redefined by Schueler [2, 3]. It is defined by the IUPAC as “… the ensemble of steric and electronic features that is necessary to ensure the optimal supramolecular interactions with a specific biological target structure and to trigger (or to block) its biological response. A pharmacophore does not represent a real molecule or a real association of functional groups, but a purely abstract concept that accounts for the common molecular interaction capacities of a group of compounds toward their target structure.” [4]. Pharmacophore models consist of so-called chemical features that represent specific molecular interaction types such as hydrogen bonds, aromatic interactions, charged centers, metal interactions, or hydrophobic contacts. There are also steric constraints such as exclusion volumes or shape features that limit the size and extent of mapping compounds (Figure 5.1) [7].

Because of their universal representation of essential chemical functionalities for bioactivity, pharmacophore models are highly successful and widely applicable over all kinds of macromolecular targets. As in silico bioactivity predictions are being applied more and more in toxicology studies to reduce experimental efforts, especially animal testing, pharmacophore models are also being investigated for their applicability in this area. This chapter gives a brief introduction to the pharmacophore technology, present applications in toxicity prediction, and finally makes a statement on the future use of pharmacophore models for the so-called antitargets.

In general, pharmacophore models are used to investigate whether compounds have a set of chemical functionalities that are typical for a specific bioactivity. There are two types of pharmacophore models: structure-based and ligand-based models [7]. Structure-based models are directly retrieved from 3D structures of targets with bound active compounds, usually from X-ray co-crystal structures or NMR structures. Additionally, a docking pose in a target structure or even in a homology model may be used as a template. Nowadays, structures refined by molecular dynamics simulations are also employed for structure-based pharmacophore modeling [5]. If the 3D structure of the target alone is available, a grid may be used to identify hot spots for ligand binding in the structure cavities and define the locations of pharmacophore features [6, 8]. For generating structure-based pharmacophore models, (possible) protein-ligand interactions are directly translated into pharmacophore features forming a starting model for systematic refinement [7, 9]. If structural data are absent or not sufficient to sufficiently describe the properties of active compounds, ligand-based pharmacophore models can be developed based upon the 3D alignment of highly active, preferably rigid molecules. These models just describe the common chemical features of active ligands and may therefore not translate into direct interactions with the target. Owing to their focus on the presence of essential chemical functionalities at defined places in the 3D space and not on the entire structure of a compound, pharmacophore models are exceptionally well suited to discover novel chemotypes for a certain activity [10, 11].

In addition to the already described models, quantitative pharmacophores can be generated, which – similarly to a quantitative structure–activity relationship (QSAR) model – aim to predict the potency of the screened molecules. The collection of training molecules for such quantitative models should comprise at least 16 compounds spanning more than three orders of activity [12]. To assure the quality of the activity data, those molecules must have been experimentally investigated using the same biological assay, preferentially in the same lab, and tested by the same scientist. These quantitative models are not primarily preferred for virtual screening, but for ranking molecules according to their expected potency. For more detailed descriptions of the generation, validation, and applications of pharmacophore models, the interested reader is referred to standard literature in the field [13, 14] and to some extent excellent recent reviews, for example [7, 15–17].

One important aspect of selecting pharmacophore models for prospective screening is their predictive quality. It is essential to only use well-validated models to prospectively screen xenobiotic databases and select compounds for further in vitro investigation. Although there is no single path leading to high quality models, a few general recommendations can be given. First, the quality of the compound data for building and theoretically validating the model is crucial. A model can only be as precise as the underlying data. But how can one recognize high-quality data? First of all, it is definitely tempting to use structure-activity databases such as the ChEMBL [18] or PubChem [19]. These databases offer ready-to-download chemical structures along with activity data on many targets. Although tremendously useful as a starting point, these databases suffer from two drawbacks: First, they are not comprehensive and cover only a fraction of the published activity data of chemicals. Second, in the automated processing of thousands of structure and activity data, errors occur. This is a known problem and there are already efforts being made to improve the accuracy of the data [20]; however, all data coming from such databases should be verified with the original literature, just to be 100% sure.

During the data collection step, it is crucial to also collect a large set of inactive molecules to train the model to discard those compounds in virtual screening. In the best case, a dataset of inactive compounds is available in literature or one of the large structure-activity databases. PubChem offers several datasets from otherwise unpublished in vitro tests and is a helpful source when searching for inactive compounds. The aim is to gather about 40-fold more inactive compounds than actives to properly mimic the success chances of a random high-throughput screening. If this amount of data on inactive compounds is not available, so-called decoys can be employed for the study [21]. A decoy is a molecule with similar physicochemical properties (molecular weight, lipophilicity, numbers of hydrogen bond donors/acceptors, etc.) as the active compounds but with a different chemical structure. These decoys have a very low chance of coincidentally being active and can therefore be used as putative inactive compounds in the model validation. However, models can be trained more accurately when data of tested inactives and not only decoys are used [22].

For the decision on the compounds to be included in the model calculation, the in vitro assays used for evaluating these compounds must also meet certain criteria. Because pharmacophore models predict the direct interaction of the ligand with the target, the biological assay readout must also give this result. Cell-based or in vivo assays include many possible ways for the ligand to be intercepted before even reaching the target. Additionally, these systems offer many other potential targets for the ligand. Accordingly, no conclusions on direct ligand binding can be drawn from those assays. Cell-free assays allowing for a direct access to the binding site are the minimum requirement for measuring protein-ligand binding.

Once a suitable dataset has been assembled, structure- and ligand-based models can be calculated. For ligand-based models, it is advisable to preferentially use highly active compounds as the training set, because virtual hits are usually at least one order of magnitude less active than the training compounds of the screening model [23]. After the model's computation, it needs to be theoretically validated using the compounds from the dataset that have not been used for model generation. Some commonly evaluated metrics for pharmacophore model quality include the yield of actives, enrichment factor, accuracy, and receiver-operating characteristic curves/area under the curve, as reviewed by Braga et al. [24]. Individual models with promising performance can additionally be manually refined by adding or deleting exclusion volumes, changing the size and shifting the location of features, marking features as optional, adding shape constrictions, or using customized features [9]. After successful theoretical validation, the model's ability to recognize so far unknown active hits is experimentally tested. For cost and efficiency reasons, it is advisable to perform such experiments on in-house or commercial databases. The biological validation of the model's predictions usually gives true positive hit rates between 5% and – in exceptional cases – 50 % or more. On the basis of new experimental data, models can be refined and again experimentally validated to verify this improvement. Finally, such a well-validated model is ready for prospective use and a steady and reliable performance of the virtual screening predictions can be expected. On the other hand, the use of models that have not been experimentally validated cannot be recommended.

Nowadays, pharmacophore models are frequently combined with computationally more expensive screening methods such as docking or molecular dynamics simulations. Pharmacophore models are often used as fast pre-filters of large chemical databases. This strategy of virtual screening is very promising, because it covers different aspects that are important for the activity of a compound: The pharmacophore model accounts for the presence of physicochemical features important for changing a biological result and the docking or molecular dynamics part independently calculates the fitting of a compound into the binding site of the target. If a substance meets both criteria, the chance that a compound is active supposedly exceeds the success rates of the isolated virtual screening approaches.

5.2 Antitarget Screening

Regarding pharmacophore-based screening, there are only few differences in screening for a compound with a desired (target) or undesired (antitarget) effect. Actually, interference with a healthy organism can always be seen as unwanted and therefore, most known pharmacological targets are also antitargets in a toxicological sense. However, some targets are more susceptible to modulation by xenobiotics and more critical for health when modulated. Those targets have already attracted much attention and some pharmacophore-based screening studies have been conducted focusing on the identification of potentially harmful agents. In this chapter, some examples, grouped by toxicological area, are shown. Owing to space constraints, it is not possible to provide a comprehensive listing.

5.3 Prediction of Liver Toxicity

For the mechanism-based prediction of liver toxicity, especially nuclear receptors (NRs) involved in fatty acid and bile acid signaling have been investigated. Specifically, for the aryl hydrocarbon receptor (AhR), constitutive androstane receptor (CAR), estrogen receptor (ER), glucocorticoid receptor (GR), farnesoid X receptor (FXR), liver X receptor (LXR), peroxisome proliferator-activated receptor (PPAR), pregnane X receptor (PXR), and retinoic acid receptor (RAR), there is evidence of involvement in the development of hepatic steatosis [25].

One of the best-studied NR involved in liver steatosis is PPARγ, which is the pharmacological target of oral antidiabetic agents such as pioglitazone. This class of drugs, full PPARγ agonists, has been associated with weight gain and liver toxicity [26]. In terms of drug development, focus shifts toward PPARγ partial agonists in expectance of reduced side effects [27]. Owing to the wealth of structural and in vitro data for this receptor, the development and validation of pharmacophore models predicting PPARγ agonism is straightforward. There are even cocrystal structures with endocrine disrupters, for example, with the flame retardant tetrabromobisphenol A (Figure 5.2) [28]. Besides many pharmacophore models that were developed to find PPARγ agonists as novel therapeutic agents [29–32], some models have now also been specifically designed for toxicity prediction related to this target [33, 34]. However, until now the latter have not been used to prospectively identify potentially hepatotoxic compounds including in vitro experiments. Additionally, up to now, there is no sharp association of adipogenesis with full agonism, partial agonism, or antagonism on PPARγ. Pharmacophore modeling of (partial) NR agonism and antagonism may be challenging. Thus, although the modeling studies on PPARγ are quite advanced, it remains an interesting field for further and more in-depth research.

A few experimentally validated pharmacophore models have been reported for LXR [35]; however, they were designed for finding novel lead candidates for the treatment of high levels of blood lipids via LXRβ agonism. In contrast, hepatotoxicity is associated with LXRα activation [36]. Still, because the ligand binding sites of the two LXR subtypes are very similar, the models are also suitable to identify potentially hepatotoxic LXRα modulators.

Similar to LXR, no specific pharmacophore models designed for hepatotoxicity prediction have been reported for FXR and CAR. The available experimentally validated models from drug discovery [37–39] could and should be used in the near future to identify compounds involved in NR-induced liver steatosis. For PXR, AhR, and RAR, some models have been reported, but they are so far not experimentally validated, for example, see Ref. [40].

5.4 Prediction of Cardiovascular Toxicity

In the cardiovascular toxicity target group, one can find predominantly ion channels and receptors associated with the regulation of blood pressure and heart rate. Probably the best known cardiovascular antitarget is the human ether-a-go-go related gene (hERG) potassium channel involved in cardiac repolarization. Several experimental and even marketed drugs have been removed from the market due to hERG activity [41]. Additionally, some calcium and sodium channels also are important ion channel antitargets.

A whole chapter is dedicated to hERG-mediated toxicity in this book (Chapter 3), so in this part the focus lies on two hERG ligand pharmacophore modeling studies including experimental validation. The first one was reported by Ekins et al. and focused on antipsychotic agents blocking the hERG channel. They developed a quantitative pharmacophore model optimized for predicting the IC₅₀ values for the hERG block [42]. In comparison, Kratz et al. followed a parallel screening approach by developing several qualitative pharmacophore models for hERG blockers and using all of them for screening commercial databases (Figure 5.3). From the hit lists, 50 compounds were selected for testing, of which 20 significantly blocked the hERG channel with IC₅₀ values ranging from 0.13 to 2.77 μM [43]. These models were also used to find hERG blockers from widely consumed herbal remedies. Alkaloids from ipecac were thereby identified as micromolar inhibitors of this channel [44]. The compositions of the pharmacophore models developed by Ekins et al. and those by Kratz et al. are very similar. Common chemical features include four hydrophobic features and one positively ionizable group, preferably a tertiary amine. Compared to hERG, other cardiac ion channels such as the CaV1.2 or NaV1.5 channels are hardly investigated using pharmacophore models. To date, no experimentally validated pharmacophore model has been reported for these targets.

In their pioneering work, Klabunde et al. designed ligand-based pharmacophore models for antitarget G protein-coupled receptor (GPCR) ligands, among them the α_1A adrenergic receptor antagonists [45]. This receptor regulates blood pressure by modulating the relaxation of the vascular muscle tone. Its unwanted inhibition can lead to orthostatic hypotension, dizziness, and fainting spells [46]. Owing to different classes of α_1A receptor ligands, several models were built and theoretically validated against a set of 50 known active and nearly 1000 inactive compounds. A clear enrichment of active compounds was observed in this experiment. The models consisted of hydrophobic/aromatic outposts and a central positively ionizable feature. The class I model additionally contained a hydrogen bond acceptor. These pharmacophore models were later used in a prospective, multistep virtual screening campaign. Among the 80 tested virtual hits, 37 showed affinities less than 10 μM, three of them even less than 10 nM [47]. Similarly, α_2A receptor agonists can trigger hypotension [48]. However, no pharmacophore modeling studies have currently been conducted for this mechanism.

The serotonin receptors 5-HT_1A and 5-HT_2B have also been associated with cardiovascular toxicity. 5-HT_2B agonism can trigger valvular heart disease [49, 50]. While 5-HT_1A activity is already the subject of pharmacophore modeling studies (see the following), 5-HT_2B still needs attention in the near future to predict those serious effects.

5.5 Prediction of Central Nervous System (CNS) Toxicity

Central nervous system (CNS) toxicity studies are very complex, because many physiological processes in the brain involve multiple GPCRs, of which there are several subtypes with different functionalities and ligand selectivities. Moreover, many CNS-active compounds target several receptors and ion channels, thereby triggering different responses. In their GPCR antitarget pharmacophore modeling work, Klabunde and Evers modeled chemical functionalities of 5-HT_2A, dopamine D2, and adrenergic α_1A ligands and also compared them. All those receptors bind biogenic amines and therefore share structural features in the binding site that enable unspecific ligands to modulate several of them. The shared features comprised a positively ionizable nitrogen, several hydrophobic and/or aromatic features, and hydrogen bond acceptors (Figure 5.4) [45]. Compounds with a certain spatial arrangement of those chemical functionalities are susceptible to CNS toxicity, if they are able to cross the blood–brain-barrier.

Although it is also involved in blood pressure regulation, the 5-HT_1A receptor is primarily associated with central effects. Ngo et al. developed agonist and antagonist models for this receptor to use them for counter screening their test candidates. These were potential α₁ receptor antagonists for the treatment of benign prostatic hyperplasia (BPH). Some of the effects modulated by the 5-HT_1A receptor (deregulated sleep patterns, anxiety-like behavior, interrupted neural tone of the iris sphincter muscle, sexual dysfunction, inhibited bladder control) are especially unpleasant for BPH patients. Accordingly, they wanted to exclude potentially unselective compounds right at the beginning of their virtual screening campaign. Theoretical validation was followed by prospective virtual screening and biological testing of hits. Although the success of the selectivity prediction between both targets was moderate, a promising compound was identified in this study [54]. While the respective models may need refinement, the strategy followed in this work is a very rational and promising one and can be recommended for similar studies.

Most currently investigated CNS antitargets belong to the target group of GPCRs (e.g., histamine receptors, adrenoceptors, dopamine receptors, muscarinic receptors, opioid receptors, cannabinoid receptors) [55]. The promiscuity of many CNS-active ligands complicates the accurate understanding and prediction of CNS effects of compounds [51]. The development of pharmacophore models for this important target class has long been delayed owing to the lack of publicly available X-ray structures. Dai et al. developed a large pharmacophore model collection for GPCRs based on the available X-ray structures and many homology models for most of the known GPCRs [52]. However, those models have so far not been experimentally validated and the usefulness of such a large, automatically generated model library remains to be determined. If this system works, it will be tremendously useful for further drug development and toxicity assessments.

5.6 Prediction of Endocrine Disruption

Endocrine disruptors are xenobiotics (substances in food, consumer products, and the environment, e.g., drinking water), which interfere with human and/or wildlife hormone biosynthesis, metabolism, or action. Thereby they have effects on male and female reproduction, cancer forms, neuroendocrinology, thyroid, metabolism, obesity, and cardiovascular endocrinology. Nowadays, endocrine disruptors are considered a public health threat and a lot of research is performed to identify potentially harmful xenobiotics [53]. Of course, testing hundreds of thousands of compounds and mixtures is a challenging and costly task and needs sophisticated planning to be also effective in identifying the most hazardous substances first. In this endeavor, in silico virtual screening tools are used to prioritize compounds for in vitro experiments. While also other virtual screening methods are intensively applied in this field, for example, QSAR models [56], pharmacophore models have already contributed to the identification of substances interacting with targets associated with endocrine disruption. For this reason, these examples are explained in more detail to show feasible workflows and success stories in the area.

In the field of endocrine disruption, many studies focused on ligands directly binding to nuclear hormone receptors, which directly regulate sexual development, growth, fertility, and behavior. The effects of agonism and antagonism on the ERs, androgen receptor (AR), thyroid receptors (TRs), GRs, and progesterone receptors (PRs) are well understood, and those proteins are established drug targets. The association of unwanted modulation of these receptors with endocrine disruption is therefore obvious and those NRs are top-priority antitargets. In terms of prospectively identifying endocrine disruptors, there are currently no reports on successful studies. However, experimentally validated pharmacophore models for ERs [57–60], ARs [61–63], GRs [64], and TRs [65] could readily be used. Additionally, cofactor binding sites of the NRs are now recognized as druggable sites and must be considered in modulating NR activities [66]. For PRs, there are currently no suitable models available.

More recently, the enzymes catalyzing the biosynthesis of hormones are also considered in evaluating endocrine disruption. These include already approved drug targets such as aromatase (Aro, CYP19) and 5α-reductase (5αR), but also investigational targets like 3β-, 11β-, and 17β-hydroxysteroid dehydrogenases (HSDs) [67]. For some of these targets, prospective virtual screening studies aiming at the identification of previously unknown endocrine disruptors have been reported.

A series of papers investigated inhibitors for 11β-HSDs. The two isoforms 1 and 2 interconvert the active glucocorticoid cortisone and inactive cortisol, thereby regulating intracellular glucocorticoid concentrations in various tissues (Figure 5.5).

Because glucocorticoids have multiple effects in different tissues, the shifting of their concentrations can trigger advantageous or unwanted effects. In brief, selective 11β-HSD1 inhibition is evaluated as a strategy to treat obesity, Alzheimer's disease, depressive disorders, and the metabolic syndrome. 11β-HSD2 inactivates glucocorticoids. This function is especially important in tissues expressing the mineralocorticoid receptor, which can also be activated by cortisone, not only by its usual agonist aldosterone. 11β-HSD2 inhibition may therefore cause apparent mineralocorticoid excess, accelerate atherogenesis promoting cancer, decreasing testosterone levels in the testes, and causing fetal development disorders [68]. The in vivo effect of 11β-HSD2 inhibition in pregnant women is well known from Finland, where many people consume high amounts of licorice, which contains the potent inhibitor glycyrrhizin. It has been shown that high licorice consumption in pregnancy has an adverse impact on fetal development in utero and also later in life [69].

The following projects can be seen as pioneering works in the field of pharmacophore-based endocrine disruptor identification, because they cover all the steps from model development, validation, virtual screening, and the successful prospective identification of enzyme inhibitors from environmental chemicals. Schuster et al. first developed and experimentally validated pharmacophore models for 11β-HSD1 and unselective inhibitors [70]. This first validation was performed on commercially available substances. Because of the endocrine-disrupting effects of 11β-HSD2 inhibition, they later used the unselective model to screen a 3D database of putative endocrine-disrupting chemicals [71]. Out of the over 76,000 compounds in the virtual screening database, 29 fitted into the model and 5 of them were biologically tested. The two compounds lasalocid and AB110873, an antibiotic used in chicken farms and a silane rubber coupling agent, significantly inhibited 11β-HSD2 with IC₅₀ values in the low micromolar range. Both active hits were chemically very distinct from the currently known enzyme inhibitors, proving the scaffold-hopping potential of the pharmacophore model. It is noteworthy that the silane compound additionally directly activated the mineralocorticoid receptor in low micromolar concentrations, which would additionally increase the adverse effects triggered by 11β-HSD2 inhibition. The two models from reference [70] were later refined with new literature data and again used to virtually screen commercial, drug, and in-house natural product databases. Experimental validation of selected hits revealed several clinically used drugs as 11β-HSD1, 2, or nonselective inhibitors. The antihypertensive furosemide, the anti-inflammatory drug ibuprofen, and the natural products digitoxigenin, hecogenin, hispanolone, and marrubiin selectively inhibited 11β-HSD1. The fungicide ketoconazole, the calcium channel blocker lidoflazine, the vitamin B1 analog octotiamine, the antibiotic rifampicin, the food-flavoring agent monoolein, and the natural product gossypol were 11β-HSD2-selective inhibitors. The immunosuppressive rapamycin nonselectively inhibited both enzymes [9]. Finally, the Drugbank database consisting of 1543 FDA-approved drugs was virtually screened with one of the refined 11β-HSD inhibitor pharmacophore models. This led to the identification of several azole antifungals as 11β-HSD1 inhibitors. Further biological tests of additional fungicides from this class identified itraconazole and posaconazole as potent 11β-HSD2 inhibitors with submicromolar IC₅₀ values [72]. Although the pharmacophore-based virtual screening proved successful in these studies, the question raised was why the potent inhibitors itraconazole and posaconazole had not been identified by the model in the first place. First of all, one needs to check if the two compounds had been present in the Drugbank database – and they were. So theoretically, they could have been found in the initial virtual screening. Then, the azoles were fitted into all available models for 11β-HSD inhibitors. It turned out that the shape restriction that should prevent too large compounds from fitting the model was responsible for missing these potent but high-molecular-weight hits. Accordingly, in future studies, the shape should be deleted from the model before the virtual database screening. In general, going back to the pharmacophore models after the biological evaluation of virtual hits is a crucial step in the model development and refinement cycle [9, 72]. Only in this way the application domain of the models can be broadened and their predictive power optimized only in this way.

Whereas 11β-HSDs are catalyzing glucocorticoid metabolism, 17β-HSDs are key enzymes in the sex hormone metabolism network (Figure 5.6).

Some of them are evaluated as drug targets and, accordingly, experimentally validated pharmacophore models for the screening for endocrine disruptors are available (Table 5.1).

From this enzyme family, one study reported the prospective discovery of environmental chemical inhibitors: screening of an endocrine disruptor database for 17β-HSD3 ligands.

17β-HSD3 catalyzes the reduction of the 17-keto group to a hydroxyl group in the final step of testosterone synthesis (Figure 5.6). Its inhibition therefore decreases testosterone synthesis in the testis. The importance of this enzyme for normal sexual development is shown in patients suffering from a mutation in the 17β-HSD3 gene, so-called 17β-HSD3 deficiency or 46,XY disorder. These patients have female sexual characteristics at birth, but are genetically males. Because of the impaired testosterone synthesis in the fetal stage, they cannot develop male characteristics. However, at puberty, other testosterone sources than the one catalyzed by 17β-HSD3 become available and so the children develop secondary male features [78]. It is therefore crucial that this enzyme is not unintentionally inhibited in the critical phase of early life.

Nashev et al. employed a pharmacophore-based virtual screening of an endocrine disruptors database for searching 17β-HSD3 inhibitors among environmental chemicals [79]. In their virtual hit lists, some representative chemical UV filters were reported. Because humans are directly exposed to this class of compounds, and it has already been known that some UV filters are bioavailable via cutaneous application, they were investigated in vitro. Additionally, other chemical UV filters not included in the database were tested. Indeed, the study identified several benzophenones and camphor derivatives as micromolar inhibitors of 17β-HSD3. Some of the active compounds were also found to antagonize AR activation by testosterone, which synergistically impairs testosterone action in the organism. As mentioned above, it is critical to return to the pharmacophore model for refinement once the new biological test data become available. In this case, the authors focused on the benzophenone class of compounds because, interestingly, their activities ranged from very low micromolar activity to inactivity. So the authors developed a structure–activity-relationship model for the benzophenone class of chemical UV filters (Figure 5.7). The information from this study suggests that industry should shift toward the use of benzophenones with etherified hydroxyl groups in UV screens and plastics to avoid potential antiandrogenic effects of these products.

Besides these paradigm studies on 11β- and 17β-HSD inhibitors, other enzymes involved in sex steroid metabolism also need attention because their inhibition can cause endocrine disruption. These include 3β-HSDs, aldosterone synthase (CYP11B), Aro-synthesizing estrogen, and 5αR-synthesizing the potent androgen dihydrotestosterone. The in vivo impacts of the Aro and 5αR inhibition are well known. Aro inhibitors such as anastrozole and letrozole are approved drugs to treat postmenopausal, estrogen-dependent breast cancer. The 5α-reductase inhibitors finasteride and dutasteride are used for treating BPH and androgenic alopecia. Currently, no pharmacophore-based virtual screening studies aiming at identifying endocrine disruptors are reported for these enzymes. However, published models for Aro inhibitors [80–82] could readily be used for this purpose and speed up the discovery of so far unrecognized active xenobiotics.

Like the steroid receptors themselves, the substrates of the steroid-metabolizing enzymes are structurally very similar to each other. Accordingly, some of these enzymes also share substrates and inhibitors. Because substrates or products of the enzymes are often endogenous NR agonists, cross-activities of xenobiotics active on one of the enzymes or receptors are often observed, as exemplified on the silane AB110873 or benzophenone-1. Therefore, the activities of a chemical needs to be determined against all of these related targets to make an informed decision on its endocrine-disrupting potential [70, 71, 73–75, 79].

5.7 Prediction of ADME

As with hERG blocking, there is a chapter on absorption, distribution, metabolism, excretion (ADME) prediction in this book. The most important targets in this area are the xenobiotic-metabolizing CYP enzymes 1A2, 2C9, 2C19, 2D6, and 3A4 as well as efflux pumps such as P-glycoprotein. In this area, both the predictions of inhibitors and of substrates are of interest.

These enzymes and transporters have been studied for decades now, and there are already some pharmacophore models applied to prospectively identify potential ligands. For example, pharmacophore models for CYP1A2 [76] and CYP2D6 [77] have been employed to identify natural products inhibiting the enzymes.

5.8 General Remarks on the Limits and Future Perspectives for Employing Pharmacophore Models in Toxicological Studies

Although the pharmacophore approach has shown promising results in the case studies described above, antitarget screening compared to drug discovery screening has other requirements in terms of model quality. In drug discovery, the so-called cherry picking approach is followed, in which the virtual screening narrows down a large database to just a small fraction of hits, which have a high probability to be active. In contrast, toxicological screening must aim at the discovery of preferably all active compounds in a database. Therefore, the application domain of the pharmacophore models should be very large, which means that the models must recognize structurally diverse active hits from various activity ranges. This can be accomplished by generating very general pharmacophore models with few features (usually only three or four), creating partial query models with omitted features, or combining several more restrictive models in a parallel screening, as exemplified in the hERG study by Kratz et al. [43] Of course, such broad screening approaches will lead to a higher false positive rate; however, it enables the models to also find chemically distinct active compounds with unexpected activity, which would not be discovered by just looking at their structures.

The application of pharmacophore models in toxicology studies also has limits: Since pharmacophore models are developed for compounds interacting with specific targets, they can only be used for predicting mechanism-based toxicity. A general toxicity prediction such as “mutagenic” or “irritant” is not within the scope of pharmacophore-based virtual screening.

Another limitation is that much of the screening success depends on the composition of the database. For environmental chemical studies, it is a challenge to assemble databases of all chemicals an organism may be exposed to. For a good start, regulatory agencies publish lists of compounds approved for use in cosmetics, food additives, drugs, industrial chemicals, and more, for example, on the FDA homepage (www.fda.gov). These often contain the CAS numbers of the respective chemicals, which can then be used to extract the chemical structures of these compounds for the compilation of virtual screening databases. These compounds may be altered and metabolized before or when they enter the human organism and these metabolites can have different effects than the parent compounds. The same is true for natural products. Apart from this, not all natural compounds that we consume are known, therefore, they cannot be included in screening databases and suggested as active compounds in a virtual screening campaign. This problem is not restricted to pharmacophore-based studies but actually concerns all virtual screening endeavors that are independent of the screening method.

Finally, pharmacophore models work on a single compound – one target at a time . However, xenobiotics are usually taken in as mixtures, be it as food, nutraceuticals, cosmetic products, or other consumer products. In vitro and in vivo, such mixtures often have not one dominant, highly active ingredient but a mixture of compounds that leads to a synergistic effect [83]. The activity simulation of these mixtures is challenging and currently the focus of intensive investigation. It is to be determined how pharmacophore models can be best used for this task.

In recent years, in silico activity profiling has become popular. In these calculations, one compound is screened against a panel of models representing different targets. This experiment unifies therapeutic target fishing for a compound and the concomitant prediction of possible adverse effects. This concept has been developed for several virtual screening methods such as 2D similarity search using fingerprints [84], combined 2D-3D similarity [85] machine learning [86], docking [87, 88], and pharmacophore models [89]. Of course, for such a high number of models, complete experimental validation is hardly feasible and therefore most of these pharmacophore models are of uncertain quality. Furthermore, the approaches are limited to certain targets and can therefore not predict all pharmacologically relevant activities. However, as the field advances, more and more high quality models are becoming available. Pharmacophore-based target fishing has already led to the identification of novel targets for natural products [90, 91] and its importance in the field is growing. Therefore, in silico activity profiling, although challenging, is bound to be a very powerful tool in toxicity research.

References

1. 1 Ehrlich, P. (1909) Über den jetzigen Stand der Chemotherapie (Over the current state of the chemotherapy). Ber. Dtsch. Chem. Ges., 42, 17–47.
2. 2 Schueler, F.W. (1960) Cheobiodynamics and Drug Design, McGraw-Hill, New York.
3. 3 Güner OF, Bowen JP. Setting the record straight: the origin of the pharmacophore concept. J. Chem. Inf. Model. 54:1269–1283.
4. 4 Wermuth, C.G. (1998) Glossary of terms used in medicinal chemistry (IUPAC recommendations 1998). Pure Appl. Chem., 70, 1129–1143.
5. 5 Wieder, M., Perricone, U., Boresch, S. et al. (2016) Evaluating the stability of pharmacophore features using molecular dynamics simulations. Biochem. Biophys. Res. Commun., 470, 685–689.
6. 6 Ortuso, F., Langer, T., and Alcaro, S. (2006) GBPM: GRID-based pharmacophore model: concept and application studies to protein–protein recognition. Bioinformatics, 22, 1449–1455.
7. 7 Vuorinen, A. and Schuster, D. (2015) Methods for generating an applying pharmacophore models as virtual screening filters and for bioactivity profiling. Methods, 71, 113–134.
8. 8 Baroni, M., Cruciani, G., Sciabola, S. et al. (2007) A common reference framework for analyzing/comparing proteins and ligands. Fingerprints for ligands and proteins (FLAP): theory and application. J. Chem. Inf. Model., 47, 279–294.
9. 9 Vuorinen, A., Nashev, L.G., Odermatt, A. et al. (2014) Pharmacophore model refinement for 11β-hydroxysteroid dehydrogenase inhibitors: search for modulators of intracellular glucocorticoid concentrations. Mol. Inform., 33, 15–25.
10. 10 Hessler, G. and Baringhaus, K.H. (2010) The scaffold hopping potential of pharmacophores. Drug Discov. Today Technol., 7, e263–e269.
11. 11 Hu, Y., Stumpfe, D., and Bajorath, J. (2017) Recent advances in scaffold hopping. J. Med. Chem., 60, 1238–1246.
12. 12 Li, H., Sutter, J., and Hoffmann, R. (2000) HypoGen: an automated system for generating 3D predictive pharmacophore models, in Pharmacophore Perception, Development, and Use in Drug Design (ed. O.F. Güner), International University Line, La Jolla, CA, pp. 172–189.
13. 13 Güner, O. (2000) Pharmacophore Perception, Development, and Use in Drug Design, International University Line, La Jolla, CA.
14. 14 Langer, T. and Hoffmann, R.D. (2006) Pharmacophores and pharmacophore searches, in Methods and Principles in Medicinal Chemistry (eds R. Mannhold, H. Kubinyi, and G. Folkers), Wiley-VCH, Weinheim, p. 375.
15. 15 Akram, M., Kaserer, T., and Schuster, D. (2015) Pharmacophore modeling and pharmacophore-based virtual screening, in In Silico Drug Discovery and Design: Theory, Methods, Challenges and Applications (ed. C.N. Cavasotto), CRC Press, Boca Raton (FL), pp. 123–153.
16. 16 Qing, X., Lee, X.Y., De Raeymaeker, J. et al. (2014) Pharmacophore modeling: advances, limitation, and current utility in drug discovery. J. Receptor Ligand Channel Res., 7, 81–92.
17. 17 Van Drie, J.H. (2013) Generation of tree-dimensional pharmacophore models. WIREs Comput. Mol. Sci., 3, 449–464.
18. 18 Gaulton, A., Bellis, L.J., Bento, A.P. et al. (2012) ChEMBL: a large-scale bioactivity database for drug discovery. Nucleic Acid Res., 40, D1100–D1107.
19. 19 Wang, Y., Xiao, J., Suzek, T.O. et al. (2012) PubChem's bioassay database. Nucleic Acid Res., 40, D400–D412.
20. 20 Papadatos, G., Gaulton, A., Hersey, A., and Overington, J.P. (2015) Activity, assay and target data curation and quality in the ChEMBL database. J. Comput. Aided Mol. Des., 29, 885–896.
21. 21 Mysinger, M.M., Cachia, M., Irwin, J.J., and Shoichet, B.K. (2012) Directory of useful decoys, enhanced (DUD-E): better ligands and decoys for better benchmarking. J. Med. Chem., 55, 6582–6594.
22. 22 Heikamp, K. and Bajorath, J. (2013) Comparison of confirmed inactive and randomly selected compounds as negative training examples in support vector machine-based virtual screening. J. Chem. Inf. Model., 53, 1595–1601.
23. 23 Scior, T., Bender, A., Tresadern, G. et al. (2012) Recognizing pitfalls in virtual screening: a critical review. J. Chem. Inf. Model., 52, 867–881.
24. 24 Braga, R.C. and Andrade, C.H. (2013) Assessing the performance of 3D pharmacophore models in virtual screening: how good are they? Curr. Top. Med. Chem., 13, 1127–1138.
25. 25 Mellor, C.L., Steinmetz, F.P., and Cronin, M.T.D. (2016) The identification of nuclear receptors associated with hepatic steatosis to develop and extend adverse outcome pathways. Crit. Rev. Tox., 46, 138–152.
26. 26 Al Sharif, M., Alov, P., Vitcheva, V. et al. (2014) Modes-of-action related to repeated dose toxicity: tissue-specific biological roles of PPARγ ligand-dependent dysregulation in nonalcoholic fatty liver disease. PPAR Res., 2014, 432647; ID 432647.
27. 27 Agrawal, R., Jain, P., and Dikshit, S.N. (2012) Balaglitazone: a second generation peroxisome proliferator-activated receptor (PPAR) gamma (γ) agonist. Mini. Rev. Med. Chem., 12, 87–97.
28. 28 Riu, A., Grimaldi, M., le Maire, A. et al. (2011) Peroxisome proliferator-activated receptor γ is a target for halogenated analogs of bisphenol A. Environ. Health Persp., 119, 1227–1232.
29. 29 Kaserer, T., Obermoser, V., Weninger, A. et al. (2016) Evaluation of selected 3D virtual screening tools for the prospective identification of peroxisome proliferator-activated receptor (PPAR) γ partial agonists. Eur. J. Med. Chem., 124, 49–62.
30. 30 Kouskoumvekaki, I., Petersen, R.K., Fratev, F. et al. (2013) Discovery of a novel selective PPARγ ligand with partial agonist binding properties by integrated in silico/in vitro work flow. J. Chem. Inf. Model., 53, 923–937.
31. 31 Guasch, L., Sala, E., Castell-Auvi, A. et al. (2012) Identification of PPARgamma partial agonists of natural origin (I): development of a virtual screening procedure and in vitro validation. PLoS One, 7 http://journals.plos.org/plosone/article?id=10.1371/journal.pone.0050816 (accessed September 26, 2017).
32. 32 Tanrikulu, Y., Rau, O., Schwarz, O. et al. (2009) Structure-based pharmacophore screening for natural-Pproduct-derived PPARγ agonists. ChemBioChem, 10, 75–78.
33. 33 Tsakovska, I., Al Sharif, M., Alov, P. et al. (2014) Molecular modelling study of the PPARγ receptor in relation to the mode of action/adverse outcome pathway framework for liver steatosis. Int. J. Mol. Sci., 15, 7651–7666.
34. 34 Al Sharif M, Tsakovska I, Pajeva I, Alov P, Fioravanzo E, Bassan A, et al. The application of molecular modelling in the safety assessment of chemicals: a case study on ligand-dependent PPARγ dysregulation. Toxicology 2017; doi.org/10.1016/j.tox.2016.01.009.
35. 35 Temml, V., Voss, C.V., Dirsch, V.M., and Schuster, D. (2014) Discovery of new liver X receptor agonists by pharmacophore modeling and shape-based virtual screening. J. Chem. Inf. Model., 54, 367–371.
36. 36 Lehrke, M., Lebherz, C., Millington, S.C. et al. (2005) Diet-dependent cardiovascular lipid metabolism controlled by hepatic LXRα. Cell Metab., 1, 297–308.
37. 37 Schuster, D., Markt, P., Grienke, U. et al. (2011) Pharmacophore-based discovery of FXR agonists. Part I: model development and experimental validation. Bioorg. Med. Chem., 19, 7168–7180.
38. 38 Fu, J., Si, P., Zheng, M. et al. (2012) Discovery of new non-steroidal FXR ligands via a virtual screening workflow based on Phase shape and induced fit docking. Bioorg. Med. Chem. Lett., 22, 6848–6853.
39. 39 Lynch, C., Pan, Y., Li, L. et al. (2013) Identification of novel activators of constitutive androstane receptor from FDA-approved drugs by integrated computational and biological approaches. Pharm. Res., 30, 489–501.
40. 40 Ekins, S. and Erickson, J.A. (2002) A pharmacophore model for pregnane X receptor ligands. Drug Metab. Dispos., 30, 96–99.
41. 41 Schuster, D. and Langer, T. (2005) Why drugs fail – a study on side effects in new chemical entities. Curr. Pharm. Des., 11, 3545–3559.
42. 42 Ekins, S., Crumb, W.J., Sarazan, R.D. et al. (2002) Three-dimensional quantitative structure-activity relationship for inhibition of human ether-a-go-go-related gene potassium channel. J. Pharmacol. Exp. Ther., 301, 427–443.
43. 43 Kratz, J.M., Schuster, D., Edtbauer, M. et al. (2014) Experimentally validated hERG pharmacophore models as cardiotoxicity prediction tools. J. Chem. Inf. Model., 54, 2887–2901.
44. 44 Kratz, J.M., Mair, C.E., Oettl, S.K. et al. (2016) hERG channel blocking ipecac alkaloids identified by combined in silico - in vitro screening. Planta Med., 82, 1009–1015.
45. 45 Klabunde, T. and Evers, A. (2005) GPCR antitarget modeling: pharmacophore models for biogenic amine binding GPCRs to avoid GPCR-mediated side effects. ChemBioChem, 6, 876–889.
46. 46 Kehne, J.H., Baron, B.M., Carr, A.A. et al. (1996) Preclinical characterization of the potential of the putative atypical antipsychotic MDL 100,907 as a potent 5-HT_2A anaagonist with a favorable CNS safety profile. J. Pharmacol. Exp. Ther., 277, 968–981.
47. 47 Evers, A. and Klabunde, T. (2005) Structure-based drug discovery using GPCR homology modeling: successful virtual screening for antagonist of the alpha1A adrenergic receptor. J. Med. Chem., 48, 1088–1097.
48. 48 MacMillan, L.B., Hein, L., Smith, M.S. et al. (1996) Central hypotensive effects of the α_2a-adrenergic receptor subtype. Science, 273, 801–803.
49. 49 Huang, X.P., Setola, V., Yadav, P.N. et al. (2009) Parallel functional activity profiling reveals valvulopathogens are potent 5-hydroxytrypatmine_2B receptor agonists: implications for drug safety assessment. Mol. Pharmacol., 76, 710–722.
50. 50 Bhattacharyya, S., Schapira, A.H., Mikhailidis, D.P., and Davar, J. (2009) Drug-induced fibrotic valvular heart disease. Lancet, 347, 577–585.
51. 51 Kroeze, W.K. and Roth, B.L. (2012) Polypharmacological drugs: ‘magic shotguns’ for psychiatric diseases, in Polypharmacology in Drug Discovery (ed. J.U. Peters), John Wiley & Sons, Inc., Hoboken, NJ, pp. 135–148.
52. 52 Dai, S.X., Li, G.H., Gao, Y.D., and Huang, J.F. (2016) Pharmacophore-map-pick: a method to generate pharmacophore models for all human GPCRs. Mol. Inform., 35, 81–91.
53. 53 Diamanti-Kandrakis, E., Bourguignon, J.P., Giudice, L.C. et al. (2009) Endocrine-disrupting chemicals: an endocrine society scientific statement. Endocr. Rev., 30, 293–342.
54. 54 Ngo, T., Nicholas, T.J., Chen, J. et al. (2013) 5-HT_1A receptor pharmacophores to screen for off-target activity of α₁-adrenoceptor antagonists. J. Comput. Aided Mol. Des., 27, 305–319.
55. 55 Peters, J.U. (2013) Polypharmacology – foe or friend? J. Med. Chem., 56, 8955–8971.
56. 56 Dimitrov, S.D., Diderich, R., Sobanski, T. et al. (2016) QSAR toolbox – workflow and major functionalities. SAR QSAR Environ. Res., 27, 203–219.
57. 57 Brogi, S., Kladi, M., Vagias, C. et al. (2009) Pharmacophore modeling for qualitative prediction of antiestrogenic activity. J. Chem. Inf. Model., 49, 2489–2497.
58. 58 Chen, L., Wu, D., Bian, H.P. et al. (2014) Selective ligands of estrogen receptor β discovered using pharmacophore mapping and structure-based virtual screening. Acta Pharmacol. Sin., 35, 1333–1341.
59. 59 Huang, W., Wei, W., Yang, Y. et al. (2015) Discovery of novel selective ERα/ERβ ligands by multi-pharmacophore modeling and virtual screening. Chem. Pharm. Bull., 63, 780–791.
60. 60 Niinivehmas, S.P., Manivannan, E., Rauhamäki, S. et al. (2016) Identification of estrogen receptor α ligands with virtual screening techniques. J. Mol. Graph Model., 64, 30–39.
61. 61 Shen, H.C., Shanmugasundaram, K., Simon, N.I. et al. (2012) In silico discovery of androgen receptor antagonists with activity in castration resistant prostate cancer. Mol. Endocrinol., 26, 1836–1846.
62. 62 Liu, J., Liu, B., Guo, G. et al. (2015) Discovery of novel androgen receptor antagonists: a hybrid approach of pharmacophore-based and docking-based virtual screening. Anticancer Drugs, 26, 747–753.
63. 63 Voet, A., Helsen, C., Zhang, K.Y., and Claessens, F. (2013) The discovery of novel human androgen receptor antagonist chemotypes using a combined pharmacophore screening procedure. ChemMedChem, 8, 644–651.
64. 64 Onnis, V., Kinsella, G.K., Carta, G. et al. (2010) Virtual screening for the identification of novel nonsteroidal glucocorticoid modulators. J. Med. Chem., 53, 3065–3074.
65. 65 Greenidge, P.A., Carlsson, B., Bladh, L.G., and Gillner, M. (1998) Pharmacophoes incorporating numerous excluded volumes defined by X-ray crystallographic structure in three-dimensional database searching: application to the thyroid hormone receptor. J. Med. Chem., 41, 2503–2512.
66. 66 Lack, N.A., Axerio-Cilies, P., Tavassoli, P. et al. (2011) Targeting the binding function 3 (BF3) site of the human androgen receptor through virtual screening. J. Med. Chem., 54, 8563–8573.
67. 67 Vuorinen, A., Odermatt, A., and Schuster, D. (2013) In silico methods in the discovery of endocrine disrupting chemicals. J. Steroid Biochem. Mol. Biol., 137, 18–26.
68. 68 Vitku, J., Starka, L., Bicikova, M. et al. (2016) Endocrine disruptors and other inhibitors of 11β-hydroxysteroid dehydrogenase 1 and 2: tissue-specific consequences of enzyme inhibition. J. Steroid Biochem. Mol. Biol., 155, Part B, 207–216.
69. 69 Räikkönen, K., Seckl, J.R., Heinonen, K. et al. (2010) Maternal prenatal licorice consumption alters hypothalamic-pituitary-adrenocortical axis function in children. Psychoneuroendocrinology, 35, 1589–1593.
70. 70 Schuster, D., Maurer, E.M., Laggner, C. et al. (2006) The discovery of new 11β-hydroxysteroid dehydrogenase type 1 inhibitors by common feature pharmacophore modeling and virtual screening. J. Med. Chem., 48, 3454–3466.
71. 71 Nashev, L.G., Vuorinen, A., Praxmarer, L. et al. (2012) Virtual screening as a strategy for the identification of xenobiotics disrupting corticosteroid action. PLoS One, 7, e46958.
72. 72 Beck, K.R., Bächler, M., Vuorinen, A. et al. (2017) Inhibition of 11β-hydroxysteroid dehydrogenase 2 by the fungicides intraconazole and posaconazole. Biochem. Pharmacol., 130, 93–103.
73. 73 Schuster, D., Kowalik, D., Kirchmair, J. et al. (2011) Identification of chemically diverse, novel inhibitors of 17β-hydroxysteroid dehydrogenase type 3 and 5 by pharmacophore-based virtual screening. J. Steroid Biochem. Mol. Biol., 125, 148–161.
74. 74 Schuster, D., Nashev, L.G., Kirchmair, J. et al. (2008) Discovery of nonsteroidal 17β-hydroxysteroid dehydrogenase 1 inhibitors by pharmacophore-based screening of virtual compound libraries. J. Med. Chem., 51, 4188–4199.
75. 75 Vuorinen, A., Engeli, R., Meyer, A. et al. (2014) Ligand-based pharmacophore modeling and virtual screening for the discovery of novel 17β-hydroxysteroid dehydrogenase 2 inhibitors. J. Med. Chem., 57, 5995–6007.
76. 76 Zhu, R., Hu, L., Li, H. et al. (2011) Novel natural inhibitors of CYP1A2 identified by in silico and in vitro screening. Int J Mol Sci., 12, 3250–3262.
77. 77 Hochleitner, J., Akram, M., Gostner, J.M. et al. (2017) A novel combinatorial approach for the discovery of cytochrome P450 2D6 inhibitors from nature. Sci. Rep., 7, 8071.
78. 78 Faienza, M.F., Giordani, L., Delvecchio, M., and Cavallo, L. (2008) Clinical, endocrine, and molecular findings in 17beta-hydroxysteroid dehydrogenase type 3 deficiency. J. Endocrinol. Invest., 31, 85–91.
79. 79 Nashev, L.G., Schuster, D., Laggner, C. et al. (2010) The UV-filter benzophenone-1 inhibits 17β-hydroxysteroid dehydrogenase type 3: virtual screening as a strategy to identify potential endocrine disrupting chemicals. Biochem. Pharmacol., 79, 1189–1199.
80. 80 Schuster, D., Laggner, C., Steindl, T.M. et al. (2006) Pharmacophore modeling and in silico screening for new P450 19 (aromatase) inhibitors. J. Chem. Inf. Model., 46, 1301–1311.
81. 81 Muftuoglu, Y., Leimgruber, S.S., Sharlow, E.R. et al. (2012) An integrated strategy to identify new aromatase inhibitors. Curr. Trends Pharmacol., 16, 15–24.
82. 82 Neves, M.A.C., Dinis, T.C.P., Colombo, G., and Sá e Melo, M.L. (2009) An efficient steroid pharmacophore-based strategy to identify new aromatase inhibitors. Eur. J. Med. Chem., 44, 4121–4127.
83. 83 Cedergreen, N. (2014) Quantifying synergy: a systematic review of mixture toxicity studies within environmental toxicology. PLoS One, 9, e96580.
84. 84 Keiser, M.J., Roth, B.L., Armbruster, B.N. et al. (2007) Relating protein pharmacology by ligand chemistry. Nat. Biotechnol., 25, 197–206.
85. 85 Gfellerr, D., Grosdidier, A., Wirth, M. et al. (2014) SwissTargetPrediction: a web server for target prediction of bioactive small molecules. Nucleic Acid Res., 1, 1–7.
86. 86 Filimonov, D.A., Lagunin, A.A., Gloriozova, T.A. et al. (2014) Prediction of the biological activity spectra of organic compounds using the pass online web resource. Chem. Heterocycl. Compd., 50, 444–457.
87. 87 Li, H., Gao, Z., Kang, L. et al. (2006) TarFisDock: a web server for identifying drug targets with docking approach. Nucleic Acid Res., 34, W219–W224.
88. 88 Vedani, A., Dobler, M., Hu, Z., and Smiesko, M. (2015) OpenVirtualToxLab – a platform for generating and exchanging in silico toxicity data. Toxicol. Lett., 232, 519–532.
89. 89 Liu, X., Ouyang, S., Yu, B. et al. (2010) PharmMapper server: a web server for potential drug tarrget identification using pharmacophore mapping approach. Nucleic Acid Res., 38, W609–W614.
90. 90 Rollinger, J.M., Schuster, D., Danzl, B. et al. (2009) In silico target fishing for rationalized ligand discovery exemplified on constituents of Ruta graveolens. Planta Med., 75, 195–204.
91. 91 Duwensee, K., Schwaiger, S., Tancevski, I. et al. (2011) Leoligin, the major lignan fom Edelweiss, activates cholesteryl ester transfer protein. Atherosclerosis, 219, 109–115.