--~--------------------------

Virtual Test Kits for Predicting HarmfulEHects Triggered by Drugs andChemicals Mediated by Specific ProteinsAngelo Yedani, Max Dobler and Markus A. LillBiographics Laboratory 3R, CH-Basel

SummaryPoor pharmacokinetics and toxicity are not only frequent caus-es of late-stage failures in drug development but also a sourcefor unnecessary animal tests. In drug discovery and for theassessment of the toxic potential of chemicals, in silico tech-niques are nowadays considered as valuable alternatives to invivo approaches. Based on a receptor-modelling concept devel-oped at our laboratory (multidimensional QSAR), we havedeveloped and validated virtual test kits for the estrogen,androgen and aryl hydrocarbon receptor (~ endocrine disrup-tion), for cytochrome P450 3A4 (~ metabolie transformations)and most recently for the thyroid receptor. These surrogateshave been tested against a total of 430 compounds and areable to predict the binding affinity close to the experimental un-certainty. These results suggest that our approach is suited forthe in silico identification of adverse efJects triggered by drugsand chemieals. Consequently, we are prepared to ofJer a freetesting to selected academic institutions and non-profit orientedorganisations.

Zusammenfassung: In silico Tests zur Voraussage rezeptor-vermittelter Toxizität von Arzneistoffen und ChemikalienUngenügende pharmakokinetische Eigenschaften und Toxizitätsind in der pharmazeutischen Entwicklung nicht nur häufigeGründe für sogenannte .Late-stage failures ", sondern aucheine Quelle für unnötige Tierversuche. In der Entwicklungneuer Armeistoffe und bei der Abschätzung des toxischenPotenzials von Chemikalien werden computer-gestützte Ver-fahren heute als wertvolle Alternativen zu in vivo Verfahreneingestuft. Basierend auf einem in unserem Labor entwickeltenRezeptormodellierungskonzept (multidimensionales QSAR)haben wir virtuelle Tests für den Estrogen-, Androgen- undDioxinrezeptor (~ endokrine Störungen), für Cytochrom P450(~Metabolismus) sowie unlängst auch für den Thyroidre-zeptor entwickelt und validiert. Diese Surrogate wurden aninsgesamt 430 Verbindungen getestet und sind in der Lage,die Bindungsaffinität nahe der experimentellen Ungenauigkeitvorauszusagen. Diese Ergebnisse lassen den Schluss zu, dassunser Ansatz geeignet ist, von Arzneistoffen und Chemikalienausgehende unerwünschte Nebenwirkungen in silico zu erken-nen. In der Folge sind wir bereit, ausgewählten akademischenInstitutionen und nicht-profit orientierten Organisationenkostenfreie Tests anzubieten.

Keywords: virtual test kits, virtual experiments, in silico toxicology, multi-dimensional QSAR, reducing animal testing

1 Introduction

Some 40 years ago, Corwin Hansch be-gan quantifying relationships between acompound's physico-ehemical propertiesand its biological aetivity. His series ofpapers laid the foundations for a technol-ogy referred to as quantitative structure-activity relationships (QSAR). Sincethen, researchers have employed QSARto predict the activity of new drugs. Astrueture-aetivity relationship has firstbeen established in 1863 by Antoine F.

Cros (cf. Borman, 1990) who observedthat the toxieity of alcohols for mammalsincreased as the molecule's water solu-bility deereased. In 1899, Hans-HorstMeyer (1899) and Charles E. Overton(1901) identified a relationship betweenthe activity of 51 narcotiea and their ac-cumulation in the lipophilic phase. Be-tween 1935 and 1938 Louis Hammeteorrelated electronic properties of organ-ic acids and bases with their pKa value.In 1964 Spencer M. Free and James W.Wilson formulated their mathematieal

Received 22 June 2005; received in final form and accepted tor publication 4 August 2005

ALTEX 22, 3/05

model allowing to break down the ob-served biological aetivity to contribu-tions from individual funetional groupsand Corwin Hansch and Toshio Fujita(1964) published their seminal work.In his attempt to formulate a general

QSAR model, Hanseh went a step fur-ther by considering drug transport anddistribution as a two-step process (Han-sch and Clayton, 1973): after a "randornwalk" from the site of application to thereeeptor site, where the drug has to crossseveral aqueous and lipid barriers, thedrug interaets with the protein bindingsite. The underlying mathematieal for-mulation extended the applieability to in

123

VEDANI ET AL. ~-------------~-

vivo data. While this and related forrnu-lations could quantitatively model thebiological activities of congeneric seriesof neutral molecules, serious problemsarose for compounds with significantlydifferent pKa values which could only beappropriately addressed by including thepKa intq, the mathematical formulation(Lien, 1975). Shortly after the advent ofHansch analysis, there were only few pa-rameters (log P, TC, hydrophobicity,scales for e1ectronic properties and ster-ic parameters) but later hundreds of pa-rameters describing topology, connectiv-ity, electronic state and others weregenerated. Already in 1972, John Toplisspointed out that a large number of vari-ables in the models increased the risk ofrandom correlations. Many people ne-glected these warnings and, as a conse-quence, literature was deluged with aflood of meaningless models, containingunreasonable combinations of too manyparameters (Kubinyi, 2002). In 1973,Ungerer and Hansch defined the goodpractice in QSAR: "select independentvariables, justify the choice of the vari-ables by statistical procedures, apply theprinciple of parsimony, have a largenumber of objects as compared to thenumber of variables, try to find a quali-tative model of physicochemical or bio-chemical significance". Whereas suchrecommendations constitute asoundethical guideline, they are not helpful inthe derivation of quantitative models if,indeed, many different variables have tobe considered. Forward and backwardelimination procedures are worthless;even stepwise regression often ends upin a local minimum. The methods ofchoice for variable selection are evolu-tionary and genetic algorithms (cf. Ku-

Tab. 1. Dimensionality of QSAR approaches

binyi, 2002 and references cited therein).As with a sufficient number of variablesit is possible to "explain everything", itwould seem to be of utmost importanceto apply the QSAR to a set of test com-pounds not used to derive the model (seealso the "Setubal principles", for exam-ple under http://ecb.jrc.it/QSAR).It is an inherent weakness of classical

QSAR that it is restricted to propertiesand sub-structural features that do notconsider the three-dimensionality of re-al molecules. In 1979, a new approachwas proposed to describe molecularproperties by fields, calculated in a reg-ular grid (Cramer and Milne, 1979).Vectors representing the clustering ofmolecular preferences (with respect to ahypothetical binding site) were extract-ed from these fields by principal com-ponent analysis and correlated with bio-logical activities. It was only after 1988when partial least squares (PLS) analy-sis was used to correlate the field valueswith biological activity, the concept -now referred to as Comparative Molec-ular Field Analysis (CoMFA) - becamewidely used, particularly to derivequantitative models for protein bindingaffinities (Cramer et al., 1988; Kubinyi,1993; Kubinyi et al., 1998a, b). In CoM-FA, the first 3D-QSAR concept, thegroup of congeneric compounds used ina study should act via the same mecha-nism and, therefore, should fit into acommon pharmacophore. In contrast to2D-QSAR methods, however, they neednot to have the same molecular skele-ton. The crucial problem of CoMFAstudies - their limitation to a single 3Dconformation for each ligand - is appro-priately addressed by 4D-QSAR(Hopfinger et al., 1997; Ekins et al.,

1999; Vedani et al., 2000). Table 1 sum-marises the philosophy behind differentQSAR concepts.Unfortunately, it is frequently "over-

looked" in QSAR studies that the bind-ing affinity of a given compound doesnot simply depend on its physico-chemi-cal properties but mainly on its interac-tions with the target protein. Strong bind-ing to a bioregulator is only observed ifthese interactions overcompensate theprice for leaving, say, an aqueous envi-ronment. While the latter is often includ-ed in QSAR parametrisation, ligand-protein interactions are typically not.With now the 3D structures of over30,000 proteins available (see http:www.rcsb.org/pdb), it cannot be fath-omed why only few QSAR studies in-clude the binding pocket of the protein.Simply analysing the properties of thesmall molecule may not be sufficient toexplain the observed biochemical activi-ty, not to speak of selectivity, e.g. to-wards receptor subtypes. Why would na-ture come up with complex proteins ifthey were not to playa crucial role? Thisfact should be considered in structure-activity studies by including the proteincounterpart. In the following, such stud-ies will be referred to as protein-basedQSAR or receptor-modelling studies, re-spectively. The latter is particularly ap-plied when the 3D structure of the pro-tein is unknown but has been replaced byan atomistic or quasi-atomistic binding-site surrogate.Accommodation of a ligand molecule

in the binding pocket is often facilitatedby the adaptation of the protein to thevery ligand topology - a phenomenonreferred to as induced fit. It may includethe rearrangement of few side-chain

Dimension Method Protein

1D-QSAR Affinity correlates with pKa, logP, electronic properties, etc. no

2D-QSAR Affinity correlates with structural patterns (connectivity, 2D pharmacophore) no

3D-QSAR Affinity correlates with the three-dimensional structure of the ligands possible

4D-QSAR Substances are represented as an ensemble of conformers, orientations, typicalprotonation states, stereoisomers

5D-QSAR as 4D-QSAR + representation of different induced-fit models yes

6D-QSAR as 5D-QSAR + representation of different solvation models yes

124 ALTEX 22, 3/05

~ VEDANI ET AL.--~--------------------------

conformations but extend to wholeloops, flaps or even extended domains.Although clearly a key phenomenon, in-duced fit is mostly neglected in QSARstudies. In 1998, the impact of induced fitonto ligand binding in a QSAR contextwas recognised for the first time and,consequently, incorporated into theQuasar software (Vedani et al., 1998).Unfortunately, even with 4D-QSAR, amajor unknown persists: manifestation(and magnitude) of the induced fit. Toaddress this problem in an unbiased fash-ion, Vedani and co-workers extendedtheir concept by a further dimension(5D-QSAR; Vedani and Dobler, 2002;see also below). Since then, this technol-ogy was applied to several receptor andenzyme systems (e.g. Vedani et al.,2005a, b; Lill et al., in press; Lill et al.,submitted).Toxic agents, particularly those that

exert their actions with a great deal ofspecificity, sometimes act via receptorsto which they bind with high affinity.This phenomenon is referred to as recep-tor-mediated toxicity. Examples of solu-ble intracellular receptors, which are im-portant in mediating toxic responses,include the glucocorticoid receptorwhich is also involved in mediating toxi-city associated effects such as apoptosisof lymphocytes as well as neuronal de-generation as a response to stress, theperoxisome proliferator activated recep-tor, which is associated with hepatocar-cinogenesis in rodents, and the aryl hy-drocarbon receptor which is involved in awhole range of toxic effects (Gustaffson,1995). Harmful effects of drugs andchemieals can often be associated withtheir binding to other than their primarytarget - macromolecules involved inbiosynthesis, signal transduction, trans-port, storage, and metabolism (Rihova,1998; Fischer 2000; Lukasiuk and Pitkä-nen, 2000; Rymer and Good, 2001;Hampson and Grimaldi, 2002; Oliverand Roberts, 2002).Toxicity testing - mandatory by inter-

national regulations for drug develop-ment and chemical safety - is still asso-ciated with stressful animal tests. Whilemany in vitro approaches have been de-vised for targeting the various aspects oftoxicological phenomena, they require achemical or drug molecule to be physi-

ALTEX 22, 3/05

cally present (i.e. synthesised) beforetesting, are time consuming, and the re-sults are often associated with large stan-dard errors. In contrast to in vitro assays,computational approach es can be appliedto hypothetical substances as their 3Dstructure can readily be generated insilico. The nowadays-available computerpower permits to scan larger batches ofcompounds (e.g. parts of corporate orpublic databases) in a relatively shorttime. Toxicity-modelling algorithms aretypically based on quantitative structure-activity relationships, neuronal networks,artificial intelligence or rule-based expertsystems. In the following, we willdemonstrate the feasibility of protein-based QSAR for the quantitative predic-tion of receptor-mediated adverse effectstriggered by drugs and chemicals.

2 Methods

Quasar - a receptor-modelling conceptdeveloped at the Biographics Laboratory3R - is based on 6D-QSAR and explic-itly allows for the simulation of inducedfit (Vedani and Dobler, 2002; Vedani etal., 2005b). Quasar generates a familyof quasi-atomistic receptor surrogatesthat are optimised by means of a geneticalgorithm. The hypothetical receptor siteis characterised by a three-dimensionalsurface which surrounds the ligandmolecules at van der Waals distance andwhich is populated with atomistic prop-erties mapped onto it. The topology ofthis surface mimics the three-dimension-al shape of the binding site; the mappedproperties represent other information ofinterest, such as hydrophobicity, electro-static potential and hydrogen-bondingpropensity. The fourth dimension refersto the possibility of representing eachligand molecule as an ensemble of con-formations, orientations and protonationstates, thereby reducing the bias in iden-tifying the bioactive conformation andorientation (~4D-QSAR). Within thisensemble, the contribution of an individ-ual entity to the total energy is deter-mined by a normalised Boltzmannweight. As manifestation and magnitudeof induced fit may vary for differentmolecules binding to a target protein, thefifth dimension in Quasar allows for the

simultaneous evaluation of up to six dif-ferent induced-fit protocols (~5D-QSAR). The most recent extension ofthe Quasar concept to six dimensions(~ 6D-QSAR) allows for the simulta-neous consideration of different solva-tion models. This can either be achievedexplicitly where parts of the surface areaare mapped with solvent propertieswhereby position and size are optimisedby the genetic algorithm, or implicitly.Here, the solvation terms (ligand desol-vation and solvent stripping) are inde-pendently scaled for each different mod-el within the surrogate family, reflectingvarying solvent accessibility of the bind-ing pocket. The associated weights areevolved throughout the simulation. Likefor the fourth and fifth dimension, amodest "evolutionary pressure" is ap-plied to achieve convergence. Detailsmay be found in the program documen-tation (see http://www.biograf.chIPDFS/Quasar.pdf). In the Quasar concept, thebinding energy is calculated as folIows:

Ebinding= Eligand-receptor- Eliganddesolvation-T~S - Eligandstrain- Einducedfit

The contributions of the individual en-tities within an ensemble (conforma-tionlorientationlprotonation, induced fit,and solvation) are normalised to unityusing a Boltzmann criterion:

Et,inding, total = L. Ebinding, individual •

exp (-w, ·Ebinding.individual/E binding,individual,maximal)

The concept has been used to quantifythe interactions of various large systems,including GPCRs (Vedani et al., 2000;2005a) and nuclear receptors (Vedani etal., 2005b, 1999; Vedani and Dobler,2003, 2001; Dobler et al., 2003; Lill etal., in press; Lill et al., submitted), there-by most recently focusing on adverseeffects triggered by drugs and chemieals(see also: http://www.biograf.ch/pro-jects.html).Raptor, an alternative technology de-

veloped by our laboratory (Lill et al.,2004) explicitly and anisotropicallyallows for induced fit by a dual-shellrepresentation of the receptor surrogate,mapped with physico-chemical proper-ties (hydrophobic character and hydro-

125

VEDANI ET AL. ~----------------------------------------~---

gen-bonding propensity) onto it. InRaptor, induced fit is not limited tosteric aspects but includes the variationof the physico-chemical fields along withit. The underlying scoring function forevaluating ligand-receptor interactionsincludes directional terms for hydrogenbonding, -hydrophobiciry and therebytreats solvation effects implicitly. Thismakes the approach independent from apartial-eharge model and, as a conse-quence, allows to smoothly model ligandmolecules binding to the receptor withdifferent net charges. Details may befound in the program documentation (seehttp://www.biograf.chIPD FSlRaptor. pdf).In Raptor, the binding energy is deter-mined as follows:

tJ.G = tJ.Ghydrogen bonding + tJ.Ghydrophobic +TtJ.S + tJ.Ginduced fit

Albeit to a lesser extent than for 3D-QSAR, the identification of the bioactiveconformation is crucial for 6D-QSAR aswell- at least to the extent that is must bepresent in the 4D representation of theligand data. Depending upon whether ornot the 3D structure of the target protein

is available at atomic resolution, two fun-damentally different concepts are used toidentify this bioactive conformation:flexible docking for systems with known3D structure and pharmacophore hypoth-esis builder otherwise. For three of thesystems presented in this account (estro-gen and androgen receptor, cytochromeP450 3A4), the crystal structures areavailable, for the other (aryl hydrocarbonreceptor) it is not.Flexible docking aims at identifying

all potential binding modes (orientations,conformations) of a small molecule with-in the binding pocket of a protein. Theunderlying protocol should address twoaspects of ligand-protein binding whichwould seem to be important: 1. the simu-lation of induced fit, i.e. allowing theprotein to adapt its shape to the differentorientations and conformations of thesmall molecule during the search proce-dure and 2. the consideration of solventeffects (typically water). In our approach(software Yeti), the sampling is conduct-ed based on a Monte-Carlo Metropolisprotocol which allows to initially (i.e.before energy minimisation) considerapparently less favourable orientations or

conformations. Such calculations arequite computing intensive as for a "ex-hausting search", typically 5,000-10,000conformations/orientations are generatedand 500-1000 are fully minimised. Forour studies, we have used the Yeti concept(Vedani and Huhta 1990, 1991; see also:http://www.biograf.chIPDFSNeti.pdf).This protocol was applied for the estro-gen receptor (Vedani et al., 2005b) andandrogen receptor (Lill et al, in press)as well as for cytochrome P450 3A4 (Lillet al., submitted). For the Ah receptorwhere no experimental structure is avail-able, the alignment was performed basedon the molecular skeletons (Vedani et al.,1999; Vedani and Dobler, 2002, 2003).

2.1. The aryl hydrocarbon(Ah) receptor2,3,7,8-Tetrachlorodibenzo-p-dioxin(TCDD)and related compounds represent seriousenvironmental health hazards, whose ef-fects include tumor promotion, dermaltoxicity, immunotoxicity, developmentaland reproductive toxicity as well as in-duction and inhibition of various enzymeactivities. TCDD also induces differenti-ation changes affecting, for example, the

Fig. 1: Stereo view of the 40 representation (461 entities) of the 121 ligands used in the Ah receptor simulation.

126 ALTEX 22, 3/05

_____ ~------------------------------------------------------V~E~D~AN~I~E~T~A~L.

human epidermis - manifesting itself aschloracne. There is strong evidence thatthis effect is mediated by the aryl hydro-carbon receptor (AhR), a regulatory ele-ment involved in the mammalianmetabolism of xenobiotics. In targetcells, TCDD initially binds to the AhR,which accumalates in the nucleus as an"AhR-aryl hydrocarbon nuclear translo-cator" heterodimeric complex. The nu-clear AhR complex acts as a ligand-in-duced transcription factor, which bindsto transacting genornic dioxinlxenobi-otics responsive elements located in the5' -regulatory region upstream from theinitiation site. This interaction results intransactivation of gene transcription.Treatment of animals or cells withTCDD and related Ah-receptor agonistscan result in decreased enzyme activitiesand decreased gene expression (see, forexample, Safe, 1990; Landers andBunce, 1991; Rappe, 1993; Swanson andBradfield, 1993; Whitlock, 1993; Okey etal., 1994; Safe and Krishnan, 1995;Putzrath, 1997).To establish a QSAR for the Ah recep-

tor, we used a total of 121 dibenzodoxins,dibenzofurans, biphenyls and polyaromat-ic hydrocarbons (PARs): 91 representedthe training set, 30 the test set (Vedani etal., 1999; Vedani and Dobler, 2003;Dobler et al., 2003). The surrogate farnilywas generated with Quasar and included1000 receptor models, which wereevolved over 60,000 crossover cycles(corresponding to 60 generations) by al-lowing for transcription-error rate of 0.02.The simulation converged at a cross-vali-dated r2 of 0.816 and yielded a predictiver2 of 0.763. Both quantities represent val-ues averaged over the 1000 models. Athree-dimensional representation of the4D dataset is depicted in Figure 1, thequasi-atornistic model in Figure 2 and theexperimental and calculated ICso valuesare compared in Figure 3.The rrns deviation for the 91 ligand

molecules of the training set correspondsto an uncertainty factor of 1.9 in thebinding affinity; the maximal individualdeviation is to a factor 8.2 in the ICsoval-ue. A total of 30 compounds (not used formodel construction) were employed fortesting the predictive power of the recep-tor surrogate yielding a predictive r2 of0.773. On the average, the predicted

ALTEX 22. 3/05

Fig. 2: Stereo view of the Ah receptor surrogate with 2-CF3-3,6,7-tricholoro-dibenzo-p-dioxin bound. For elarity, the front seetion has been elipped. Areas eolored ingray/brown represent hydrophobie properties; areas in green H-bond donor funetions(H-bond aeceptors and salt bridges are not observed as the ligand data set laekseorresponding groups).

binding affinities of the test ligands devi-ate by a factor of 2.4 in ICsofrom the ex-periment; the maximal observed devia-tion eorresponds to a factor of 9.9 in ICso.This surrogate was then employed to testits ability to handle test compounds -both benign and harmful- that do not be-

long to any structural class from trainingor test set. The results suggest that thisreceptor model is suited to be used as avirtual test kit for identifying toxic ef-fects triggered by small molecules andmediated by the aryl hydrocarbon recep-tor (Vedani et al., in press).

7.0

2 Training set: n=91, q2=0.816

! Test set: n=30. p2=0.7638.0

"":''tJt!!S: 6.0oIl)

~CI 5.0oT

4.0

3.0

Quasar5.0

4.0 5.0 6.0 7.0-log ICso (exp.)

8.0 9.0

Fig. 3: Comparison of experimental and calculated ICso values for the Ah receptorsurrogate.

127

VEDANI ET AL. m.....---------------------------------------~---

2.2 The estrogen receptorNuclear receptors comprise a family ofligand-dependent transcription factorsthat transform extra- and intracellularsignals into cellular responses by trigger-ing the transcription of target genes. Inparticular, )bey mediate the effects ofhormones and other endogenous ligandsto regulate the expression of specificgenes. Among other members, this fami-

ly includes receptors for the varioussteroid hormones - e.g. the estrogen, an-drogen, progesterone, and glucocorticoidreceptor. Unbalanced production or cellinsensitivity to specific hormones mayresult in diseases associated with humanendocrine dysfunction (Zubay et al.,1995). The presence of hormonally ac-tive compounds - endocrine disruptors -in the biosphere has become a worldwide

Fig. 4: The 344 entities (40 data set) of the 106 ligands used in the estrogen receptorsimulation.

Fig. 5: Stereo view of the estrogen receptor surrogate with Coumestrol bound. Forclarity, the front section has been clipped. Areas colored in gray/brown representhydrophobie properties; areas in green H-bond donor functions (H-bond acceptors andsalt bridges are not observed as the ligand data set lacks corresponding groups).

128

environmental concem. It has been con-cluded that such compounds elicit a vari-ety of adverse effects in both humans andwildlife including promotion of hor-mone-dependent cancers, reproductivetract disorders, and a reduction in repro-ductive fitness. A number of receptor-mediated hormonal responses to toxicityare known, including xenobiotic effectson the thyroid hormone receptor, the epi-dermal growth factor receptor, the arylhydrocarbon receptor as well as effectsmediated by the androgen and the estro-gen receptor, respectively. A variety ofcompounds in the environment havebeen shown to display agonistic or antag-onistic activity towards the ER, includingboth natural products and synthetic com-pounds (Dibb, 1995; McLachlan andAmold, 1996; Guillette et al., 1995; Col-born, 1995; Feldmann and Krishnan,1995; Hoare et al., 1994; Korach et al.,1987). The concem over xenobioticsbinding to the ER has created a need toboth screen and monitor compoundswhich can modulate endocrine effects.This has been underscored by the US leg-islation in 1995/6, by mandating thatehernieals and formulations must bescreened for potential estrogenic activitybefore they are manufactured or used incertain processes (US Govemment,1996a, b). In Switzerland, the necessityfor a co-ordinated interdisciplinary ap-proach to the environmental and publichealth problems caused by endocrine dis-ruption has been recognised and a Na-tional Research Programme on En-docrine Disruption (NRP50) has beenlaunched in 2001 (Swiss National Sei-ence Foundation).Our 6D-QSAR study was based on

the X-ray crystal structure of the humanER a ligand-binding domain with bounddiethylstilbestrol (Shiau et al., 1998)which served as template for identifyingpotential binding modes of the 106 in-vestigated compounds. While for thesymmetric diethylstilbestrol moleculethe binding mode (the two phenolic hy-droxyl groups engage in hydrogen bondswith Glu 353 and His 524, respectively)is unambiguous, different orientationswithin the binding pocket are feasible formost of the asymmetric compounds usedin this study. Here, an only moderate in-duced fit - involving the rearrangement

ALTEX 22, 3/05

~ VEDANI ET AL.---~----------------------------------------

of protein side chains - tolerates differentligand orientations (Vedani et al. ,2005b). The three-dimensional structuresof all 106 compounds were generatedand optirnised in aqueous solution usingMacroModel (Mohamadi et al., 1990);atornic partial charges and solvation en-ergies were ealculated using AMSOL(Cramer and Truhlar, 1992). Subsequent-ly, we sampled all feasible arrangementswithin the binding pocket using a Monte-Carlo search protocol based on aMetropolis selection criterion (softwareYeti; Vedani and Huhta, 1990). For eachcompound, a total of 40,000 conforma-tions and orientations within the bindingsite were randomly generated; 4000thereof were fully rninirnised, includingan extended binding pocket of 12 Aaround the very ligand. For each ligandmolecule, up to four energetically favor-able arrangements (within 5.0 kcal/molof the lowest-energy conformation) wereretained and composed into a 4-D dataset (Fig. 4), totalling in 344 representa-tions for the 106 molecules. The ligandscomprise six different substance classesand cover a range in ICsoof seven ordersof magnitude (2.8 mM - 0.2 nM).Of the 106 agonist molecules used in

our study, 80 defined the training set, andthe remaining 26 the test set. In Quasar,the surrogate family comprised 200models that were evolved over 32,000crossover cycles, corresponding to 160generations. The simulation reached across-validated r2 of 0.895 and yielded apredictive r2 of 0.892. The receptor surro-gate is depicted in Figure 5; comparisonwith the binding site at the truebiological receptor shows that characteris-tic properties are well identified by the re-ceptor surrogate (Vedani et al., 2005b).Experimental and calculated ICso valuesare compared in Figure 6: the rms devia-tion for the 80 ligand molecules of thetraining set corresponds to a factor 2.0 inthe experimental ICsovalue; the maximalindividual deviation to a factor 8.6 in ICso.On the average, the 26 test compounds de-viate by a factor 2.7 in ICso from the ex-periment; the maximal observed deviationcorresponds to a factor 9.5.

2.3. The androgen receptorNuclear receptors represent the largestfamily of ligand-dependent eukaryotic

ALTEX 22. 3/05

9.0

Quasar 5.0

2 Training set (n=80) q2=0.89510.0 !Test set (n=26) p2:0.892

8.0.....•Uiä 7.0()-g 6.02~ 5.0T

4.0

3.0

2.0

2.0 3.0 4.0 5.0 6.0 7.0 8.0 9.0 10.0

-log ICSO (exp.)

Fig. 6: Comparison of experimental and calculated ICso values for the estrogenreceptor surrogate.

transcription factors transforrning extra-and intracellular signals into cellular re-sponses by triggering the transcriptionof target genes. In particular, they medi-ate the effects of hormones and otherendogenous ligands to regulate the ex-pression of specific genes, thereby regu-lating development and metabolism.Among other members, this farnily in-cludes receptors for the various steroidhormones - e.g. the androgen, estrogen,glucocorticoid and progesterone recep-tor. Unbalanced production or cell in-sensitivity to specific hormones may re-sult in diseases associated with humanendocrine dysfunction. Androgens andthe androgen receptor play an essentialrole in the growth of normal prostate.They are, however, also involved in thedevelopment of prostate cancer, repre-senting the most common male malig-nancy in the United States. Bothsteroidal and nonsteroidal derivativeshave shown clinical benefits aschemotherapeutic agents for prostatecancer. Still, several of these antiandro-gens, for example cyprosterone, showoverlapping effects with other hormonalsystems. Many environmental cherni-cals, e.g. flavones or kepone, displaysimilar structural properties: a hy-drophobic core and one or two terminalpolar groups. Consequently, they may

bind to a nuclear receptor influencingthe balance of the endocrine system.The presence of these so-called en-docrine disruptors in the biosphere hasbecome a worldwide environmentalconcern. It has been conc1uded that suchcompounds e1icit a variety of adverseeffects in both humans and wildlife in-c1uding promotion of hormone-depen-dent cancers, reproductive tract disor-ders, and areduction in reproductivefitness. A variety of compounds in theenvironment have been shown to dis-play agonistic or antagonistic activitytowards the androgen receptor, includ-ing both natural products and syntheticcompounds (Kavlock et al., 1996; Col-born et al. 1993; Carlsen et al. , 1992).The concern over xenobiotics binding tothe androgen receptor has created aneed to both screen and monitor com-pounds expected to modulate endocrineeffects. We therefore developed an insilico model to quantitatively predict thepotential of structurally diverse ligandsfor binding to the androgen receptor us-ing a multidimensional QSAR tech-nique (software Raptor, Lill et al. ,2004), specifically allowing for in-duced-fit. To identify the binding modeat the true biological receptor, a novelstepwise protocol consisting of flexibledocking, molecular dynamics (MD)

129

VEDANI ET AL. ~----------------------------------------~---

Fig. 7: Dual-shell representation of the androgen receptor surrogate with Miboleronebound. For clarity, the front section has been clipped. Areas colored in brown representhydrophobie properties; areas in red correspond to H-bond acceptors, blue to H-bonddonors and green to H-bond flip-flops.

8.0

2 Trainingset: 11=86,r 2=0.8589.0 !Test set: n=26, p2=0.792

-.-ci 7.0~Co--o 6.0an2~ 5.0T

4.0

3.0

Raptor2.0

4.0 5.0 6.0 7.0 8.0 9.0 10.0

-log ICSO(exp.)

Fig. 8: Comparison of experimental and calculated ICso values tor the androgenreceptor surrogate. (For clarity, seven threshold compounds have been omitted; cf. Lill etal., in press)

130

simulations and linear interaction-ener-gy analysis (LlE) was developed (Lill etal., in press).Our QSAR study was based on 119

ligands; 88 molecules thereof were as-signed to the training set and the remain-ing 31 used as test. Induced fit was ac-counted for by the dual-shell concept ofRaptor (Lill et a1., 2004). To allow fortopological and physico-chernical varia-tion at the true biological receptor withdifferent ligands bound, the Raptor re-sults were averaged over 10 individualmodels defining a surrogate family.The model converged at a cross-vali-

dated r2=0.858 for the training com-pounds and yielded a predictive r2=0.792for the test compounds. The model isshown in Figure 7; experimental and cal-culated ICsovalues are compared in Fig-ure 8. On the average, the predicted bind-ing affinity of the training ligandsdeviates by a factor 1.7 from the experi-ment; those of the test set deviate by afactor 1.6 in ICso.The maximal observeddeviation for individual molecules are7.8 and 13.9, respectively.

2.4. CytochromeP450 3A4During multi-drug medication, the indi-vidual entities may compete for the sameenzyme(s) to be metabolised. This canlead to undesired drug-drug interactions(Tanaka, 1998). In order to envision suchinteractions, a quantitative prediction ofassociated pharmacokinetic changeswould be beneficiary. A major mecha-nism is realised by inhibiting Cy-tochrome P450 3A4 (CYP3A4) mediatedbiotransformations, as this enzyme is re-sponsible for metabolising 50-60% oforally adrninistered marketed drugs. Forthis purpose, we developed a computa-tional model to accurately predict the in-hibitory potential of a diverse set of lig-and molecules using flexible docking andnovel multi-dimensional QSAR tech-niques (Lill et a1., submitted).The three-dimensional structures of

48 compounds were generated and opti-mised with MacroModel (Mohamadi,1990). After relaxation, the X-ray crys-tal structure of the human CYP3A4 en-zyme was used as template for the auto-matic docking of the individualcompounds. CYP3A4 can accommo-date a wealth of different substrates and

ALTEX 22, 3/05

~ VEDANI ET AL.--~--------------------------

Fig. 9: Dual-shell representation of the cytochrome P340 3A4 surrogate withamiodarone bound. For elarity, the front seetion has been elipped. Areas eolored inbrown represent hydrophobie properties; areas in red eorrespond to H-bond aeeeptors,blue to H-bond donors and green to H-bond flip-flops.

its binding pocket is rather wide andpredominantly hydrophobic in nature.As a consequence, different bin dingmodes and conformations had to beconsidered for small molecules bindingto it. Induced fit mayaIso facilitate lig-and binding, thereby hosting com-pounds varying in molecular volume byas much as a factor of seven. To identi-fy the possible binding modes, we sam-pled structurally and energetically feasi-ble arrangements within the bindingpocket using flexible docking combinedwith an extended Monte-Carlo searchprotocol based on a Metropolis selec-tion criterion (software Yeti; Vedani andHuhta, 1990) and allowing for an adap-tation of the pro tein side-chains to eachindividual ligand molecule.The simulation using the Raptor soft-

ware (Lill et al., 2004) reached a cross-validated r2 of 0.867 and yielded a pre-dictive r2 of 0.870. The binding-sitemodel is depicted in Figure 9. Compari-son with the binding site at the true bio-logical enzyme shows that both the hy-drophobic character and hydrogen-bondproperties of the binding packet are wellidentified by the model (Lill et al., sub-mitted). Experimental and calculatedICso values are compared in Figure 10.On the average, the predicted binding

ALTEX 22. 3/05

affinity of the training ligands deviatesby a factor 2.4 in ICso from the experi-ment; the maximal observed deviation is5.0. For the ligands of the test set, thecorresponding values are 2.0 and 5.2, re-spectively - an appreciable result, as the

10.0

9.0

8.0

.-..g 7.0fS:<:) 6.0It)

2CI 5.00T

4.0

2 Trainingset n=31,r2=0.867!Test set: n=9, p2=0.870

test set ineludes diverse molecules notsimilar in structure to any compound ofthe training set. The weak- and non-bind-ing molecules were predicted to bindwith affinities elose to or weaker than theexperimental threshold value (Lill et al.,submitted).

3 Discussion

3.1. Virtual test kitsThe four systems discussed in this ac-count - aryl hydrocarbon receptor, estro-gen receptor, androgen receptor, cy-tochrome P450 3A4 - for which aquantitative structure-activity relation-ship could be established and validatedfor a total of 430 compounds by meansof multi-dimensional QSAR are part ofthe envisioned Internet Laboratory forthe in silico identification of adverse ef-fects triggered by drugs and chemieals(Vedani and Dobler, 2001, 2003; Dobleret al., 2003; Vedani et al., in press). Theyare already functional as individual testkits and the Biographics Laboratory 3Ris prepared to offer a free testing to se-lected academic institutions and non-profit oriented organisations. The under-lying technology (software Yeti,

,,,,/,,,,,,,,,,,,,

.:,,,,,,

",,,,3.0

,,,,,,,o /,,/,,,,,,,,

Raptor2.02.0 ~-,--~~-1---,--L-L.---1---,,--1--,--..L---,--1---,-....J

2.0 3.0 4.0 5.0 6.0 7.0 8.0 9.0 10.0

-log 1CSO(exp.)

Fig. 10: Comparison of experimental and calculated ICso values for the cytochromeP450 3A4 surrogate. (For elarity, eight threshold eompounds have been omitted; cf. Lill etal., submitted)

131

VEDANI ET AL. m.....-------------~-

Symposar, Quasar, Raptor) runs onMacintosh and PC operating systems.We are presently extending our receptorlibrary by an additional member, the thy-roid receptor (Zumstein, 2005). Whilethe compound evaluation (softwareQuasar and Raptor) is very fast, the,Monte-Carlo search required to identifythe binding mode(s) is quite cpu inten-sive. Presently, up to four compounds/hour may be processed on a laptop com-puter; on a 32-node Linux cluster asmany as 100 compounds/hour (1500overnight) may be evaluated in silico.Any new extensions to the database,the technology or access procedures willbe posted at http://www.biograf.ch/projects.html.

3.2 3R relevanceThe envisioned Internet laboratory andthe already functional virtual test kitscan contribute to a significant reductionin animal testing. First, it allows for anearly - even before compound synthesis- recognition of potentially harmfulsubstances. By removing those candi-date substances from the evaluationpipeline, they will not be forwarded toany in vivo toxicity tests. These expec-tations are supported by the fact that ourvirtual experiments have so far not pro-duced any false-positive results. Ofcourse, with only a limited number ofenzyme/receptor systems known to me-diate adverse effects and even fewer ac-cessible in a QSAR context (due tolacking experimental affinity data),false-negative results will always bepresent. But this is by no means the aimof the approach - instead, it will selec-tively recognise potentially hazardouscompounds associated with majormechanisms (e.g. metabolic degrada-tion, endocrine disruption) and allow todiscard them early on. Second, a widelyused database of this kind would reducethe number of otherwise doubly-con-ducted toxicity tests at research labora-tories focussed on closely relatedbiomedical targets. The main advantageof the proposed virtuallaboratory is thatit can be applied to hypothetical sub-stances, is reproducible, fast and cheap.Another field of application is testing

of chemieals for toxicity - for exarnple

132

the 30,000 compounds that have to beretested by 2012 as defined in the Euro-pean Commission's well-documentedWhite paper on the strategy for a futurechemieals policy (2003) - and causing anestimated toll of 10 Million laboratoryanimals. Here, our system could prove tobe a useful in silico screening tool as newcompounds can be tested with only mod-erate "human" efforts. The importance ofQSARs has more recently been acknowl-edged by the OECD (2003) and the Dan-ish Environment Protection Agency hastaken the lead in use of structure-basedmethods to prioritise hazardous chemi-cals (Cronin, 2003).The complex task of maintaining a

virtuallaboratory on the Internet cannotbe accomplished by a single laboratory.It is therefore our intention to make itfreely available as soon as possible - ofcourse, by applying security measuresto avoid non-scientific use. The database will be managed as an "opensource mode", implying that all interest-ed, skilled parties may contribute to de-velopment and extension. For this pur-pose, an independent source ofcontinued funding is very important.The Foundation Biographics Laboratory3R is a non-profit organisation with on-ly limited financial means. We think thatfinancial support by an independentgovernmental agency would representan optimal solution. Use of the databasewould not be associated with any costsas we are prepared to provide to neces-sary software for free (universities, hos-pitals, regulatory bodies) or at primecosts for the pharmaceutical industry.

ReferencesBorman, S. (1990). Cros, A. F. A., PhDThesis, University of Strasbourg, 1863cited from "New QSAR techniqueseyed for environmental assessments",Chem. & Eng. News 19, 20-23.

Carlsen, E., Giwercman, A., Keiding, N.and Skakkebaek, N. E. (1992). Evi-dence for decreasing quality of semenduring past 50 years. Br. Med. J. 305,609-613.

Colborn, T. (1995). Environmental estro-gens: health implications for humansand wildlife. Environ. Health Perspect.

103 (supp!. 7), 135-136.Colborn, T., von Saal, F. S. and Soto, A.M. (1993). Developmental effects ofendocrine-disrupting chemieals inwildlife and humans (see comments).Environ. Health Perspect. 101, 378-384.

Cramer, C. J. and Truhlar, D. G. (1992).AMl-SM2 and PM3-SM3 para-metrized SCF solvation models forfree energies in aqueous solution.J. Comp.-Aided Molec. Design 6, 629-666.

Cramer, R. D. III, Patterson, D. E. andBunce, J. D. (1988). ComparativeMolecular Fields Analysis (CoMFA).1. Effect of shape on bin ding ofsteroids to carrier proteins. 1. Am.Chem. Soc. 110,5959-5967.

Crarner, R. D. III and Milne, M. (1979).The lattice model: a general paradigmfor shape-related structure-activitycorrelations. Abstracts 0/ papers 0/ theAm. Chem. Soc. Meeting, April 1979,COMP44.

Cronin, M. (2003). A toxic gamble.Chemistry and Industry 4, 13-14.

Dibb, S. (1995). Swimming in a sea ofestrogens, chemical hormone dis-rupters. The Ecologist 25,27-31.

Dobler, M., Lill, M. A. and Vedani, A.(2003). From crystal structures andtheir analysis to the in silico predictionof toxic phenomena. Helv. Chim. Acta86, 1554-1568.

Ekins, S., Bravi, G., Binkley, S. et al.(1999). Three- and four-dimensionalquantitative structure-activity relation-ship analyses of cytochrome P450 3A4inhibitors. J. Pharmaco!. Exp. Ther.290, 429-438.

Feldman, D. and Krishnan, A. (1995).Estrogens in unexpected places: Possi-ble implications for researchers andconsumers. Environ. Health Perspect.103 (supp!. 7), 129-133.

Fischer, B. (2000). Receptor-mediatedeffects of chlorinated hydrocarbons.Andrologia 32, 279-283.

Free, S. M. and Wilson, J. W. (1964). Amathematical contribution to structure-activity studies. 1. Med. Chem. 7, 395-399.

Guillette, L. J., Crain D. A., Rooney, A.A. and Pickford, D. B. (1995). Organi-zation versus activation: the role of en-

ALTEX 22, 3/05

~ VEDANI ET AL.-~-------------

docrine disrupting contaminants EDCsduring embryonie development inwildlife. Environ. Health Perspect. 103(suppl. 7), 157-164.

Gustaffson, J. A. (1995). Receptor-medi-ated toxicity. Toxicol. Lett. 135, 465-470.

Hammet, L. P,(1935). Reaction rates andindicator acidities. Chem. Rev. 17, 67-79.

Hampson, A. J. and Grimaldi, M. (2002).12-hydroxyeicosatetrenoate (12-HETE)attenuates AMPA receptor-mediatedneurotoxicity: Evidence for a G-pro tein coupled HETE receptor.1. Neurosci. 22, 257-264.

Hansch, C. and Fujita, T. (1964). p-cr-1tAnalysis. A method for the correlationof biological activity and chemicalstructure. J. Am. Chem. Soc. 86, 1616-1626.

Hansch, C. and Clayton, J. M. (1973).Lipophilic character and biologicalactivity of drugs 11:The parabolic case.J. Pharm. Sci. 62, 1-2l.

Hoare, S. A., Jobling, S., Parker, M. G. etal. (1994). Environmental persistentalkylphenolic compounds are estro-genic. Endocrinology 135, 175-182.

Hopfinger, A. J., Wang, S., Tokarski, J. S.et al. (1997). Construction of 3D-QSAR models using 4D-QSAR analy-sis forrnalism. J. Am. Chem. Soc. 119,10509-10524.

Kavlock, R. J., Daston, G. P., DeRosa, C.et al. (1996). Research needs for therisk assessment of health and environ-mental effects of endocrine disruptors:areport of the D.S. EPA-sponsoredworkshop. Environ. Health Perspect.104 (Suppl. 4),715-740.

Korach, K. S., Levy, L. A. and Sarver, P.J. (1987). Estrogen receptor stereo-chemistry: receptor binding and hor-monal responses. J. Steroid Biochem.27,281-290.

Kubinyi, H. (2002). From narcosis tohyperspace: The history of QSAR.Quant. Struct.-Act. Relat. 21, 348-356.

Kubinyi, H., Folkers, G. and Martin, Y.C. (1998a). 3D-QSAR in drug design.Volume 2. Ligand-protein complexesand molecular similarity, KluwerlESCOM, Dordrecht 1998; also pub-lished as Persp. Drug Disc. Design9-11, 1-416.

ALTEX 22, 3/05

Kubinyi, H., Folkers, G. and Martin, Y.C. (1998b). 3D-QSAR in drug design.Volume 3. Recent advances,Kluwer/ESCOM, Dordrecht 1998;also published as Persp. Drug Disc.Design 12-14,1-352.

Kubinyi, H. (ed.) (1993). 3D-QSAR indrug design. Theory, methods and ap-plications. Leiden: ESCOM SciencePublishers B.V.

Landers, J. P. and Bunce, N. J. (1991).The Ah receptor and the mechanism ofdioxin toxicity. Biochem. J. 276, 173-287.

Lien, E. J. (1975). Structure-absorption-distribution relationships. In E. J.Ariens (ed.), Drug Design (Volume V,81-132). NewYork: Academic Press.

Lill, M. A., Winiger, F., Vedani, A. andErnst, B. (in press). Impact of inducedfit on ligand binding to the androgenreceptor: A multidimensional QSARstudy to predict endocrine-disruptingeffects of environmental chemicals. J.Med. Chem.

Lill, M. A., Dobler, M. and Vedani, A.(submitted). Prediction of small-molecule binding to cytochrome P4503A4: Flexible docking combined withmultidimensional QSAR. Chem. Med.Chem.

Lill, M. A., Vedani, A. and Dobler, M.(2004). Raptor - combining dual-shellrepresentation, induced-fit simulationand hydrophobicity scoring in receptormodeling: Application towards thesimulation of structurally diverse lig-and sets. J. Med. Chem. 47, 6174-6186.

Lukasink, K. and Pitkanen, A. (2000).GABA(A)-mediated toxicity of hip-pocampal neurons in vitro. J. Neu-rochem. 74, 2445-2454.

McLachlan, J. A. and Arnold, S. F.(1996). Environmental estrogens. Am.Sei. 84, 452-46l.

Mohamadi, F., Richards, N. G. J., Guida,W. C. et al. (1990). MacroModel -An integrated software system formodeling organic and bioorganicmolecules using molecular mechanics.J. Comput. Chem. 11,440-467.

OECD (2003). http://webdominol.oecd.org/comnet/env/models.nsf.

Okey, A. B., Riddick, D. S. and Harper,P. A. (1994). The Ah receptor: Media-

tor of the toxicrty of 2,3,7,8- tetra-chlorodibenzo-p-dioxin (TCDD) andrelated compounds. Toxicol. Lett. 70,1-22.

Oliver, J. D. and Roberts, R. A. (2002).Receptor-mediated hepatocarcinogen-esis: role of hepatocyte proliferationand apoptosis. Pharmacol. Toxicol. 91,1-7.

Putzrath, R. M. (1997). Estimating re-lative potency for receptor-mediatedtoxicity: Re-evaluating the toxicityequivalence factor (TEF) model. Reg.Tox. Pharmacol. 25,68-78.

Rappe, C. (1993). Environmental con-centrations and ecotoxicological ef-fects of PCDDs, PCD Fs and relatedcompounds. Organohalogen Com-pounds, Dioxin 12, 163-170.

Rihova, B. (1998). Receptor-mediatedtargeted drug or toxin delivery. Adv.Drug Deliv. Rev. 29, 273-289.

Rymer, D. L. and Good, T. A. (2001).The role of protein activation in thetoxicity of amyloidogenic Aß (1- 40),Aß (25-35), and bovine calcitonin.J. Biol. Chem. 276, 2523-2530.

Safe, S. and Krishnan, K. (1995). Cellu-lar and molecular biology of Aryl hy-drocarbon (Ah) receptor-mediatedgene expression. In D. H. Degen, J. PSeiler and P. Bentley (eds.), Archivesof Toxicology (Suppl. 17, 116-124),Berlin: Springer Verlag.

Safe, S. (1990). Polychlorinatedbiphenyls (PCBs), dibenzo-p-dioxins(PCDDs), dibenzofurans (PCDFs) andrelated compounds: Environmentaland mechanistic considerations whichsupport the development of toxicequivalency factors (TEFs). CRC Crit.Rev. Toxicol. 21, 51-88.

Shiau, A. K., Barstad, D., Loria, P. M. etal. (1998). The structural basis of es-trogen receptor/coactivator recognitionand the antagonism of this interactionby tamoxifen endogenous estrogen17ß-estradiol (E2) and the syntheticnonsteroidal estrogen diethylstilbestrol(DES). Cell 95,927-937.

Swanson, H. I. and Bradfield, C. A.(1993). The Ah-Receptor: Genetics,structure and function. Pharmacoge-netics 3, 213-230.

Swiss National Science Foundation(2001). http://www.nrp50.ch.

133

VEDANI ET AL. m.....----------------------------------------~---

Tanaka, E. (1998). Clinically importantpharamcokinetic drug-drug interac-tions: Role of cytochrome P450 en-zymes. 1. Clin. Pharrn. Therap. 23,403-416.

Topliss, J. G. and Costello, R. J. (1972).Chance correlations in structure-activ-ity studies using multiple regressionanalysis. J. Med. Chem. 15, 1066-1068.

Ungerer, S. H. and Hansch, C. (1973).On model building in structure-activityrelationships. Areexamination ofadrenergic blocking of ß-halo-ß-arlyalkylarnines. 1. Med. Chern. 16,745-749.

US Govemment (1996a): Safe DrinkingWater Act Amendment. Public Law104-182 (Section 136), http://www.epa.gov.safewater/sdwaJindex.html.

US Govemment (1996b): Food QualityProtection Act. Public Law 104-170(Section 408), http://www.fda.gov/opacomllaws/foodqual/fqpatoc.htm.

Vedani, A., Dobler, M. and Lill, M. A. (inpress). In silico prediction of harrnfuleffects triggered by drugs and cherni-cals. Toxicol. Appl. Pharrnacol.

Vedani, A., Dobler, M. and Lill, M. A.(2005b). Combining protein modelingand 6D-QSAR - Simulating the bind-ing of structurally diverse ligands tothe estrogen receptor. 1. Med. Chern.48, 3700-3703.

Vedani, A., Dollinger, H., Hasselbach,K.-M. et al. (2005a). Novelligands forthe chemokine receptor-3 (CCR3): Areceptor modeling study based on 5D-

134

QSAR. J. Med. Chern. 48, 1515-1527.Vedani, A. and Dobler, M. (2003). Inter-net laboratory for predicting harrnfuleffects triggered by drugs and cherni-cals - A progress report. ALTEX 20,85-91.

Vedani, A. and Dobler, M. (2002). 5D-QSAR: The key for simulating in-duced fit? 1. Med. Chern. 45, 2139-2149.

Vedani, A., Briem, H., Dobler, M. et al.(2000). Multiple conformation andprotonation-state representation in 4D-QSAR: The neurokinin-l receptor sys-tem. J. Med. Chern. 43, 4416-4427.

Vedani, A., McMasters, D. R. undDobler, M. (1999). Genetische Algo-rithrnen im 3D-QSAR: Verwendungmultipler Wirkstofforientierungen zurverbesserten Voraussage der Toxizität.ALTEX 16,9-14; ALTEX 16,140-143.

Vedani, A. and Dobler, M. (2001). Inter-net laboratory for predicting harmfuleffects triggered by drugs and cherni-cals. ALTEX 18,110-114.

Vedani, A., Dobler, M. and Zbinden, P.(1998). Quasi-atornistic receptor sur-face models: A bridge between 3-DQSAR and receptor modeling. J. Arn.Chern. Soc. 120,4471-4477.

Vedani, A. and Huhta, D. W. (1991). Analgorithm for the systematic solvationof proteins based on the directionalityof hydrogen bonds. J. Arn. Chem. Soc.113,5860-5862.

Vedani, A. and Huhta, D. W. (1990). Anew force field for modeling metallo-proteins. J. Arn. Chern. Soc. 112,4759-

4767.White paper on the strategy for a futurechemieals policy (2003). Cf. http://eu-ropa.eu.int/comml environmentlchem-icals/whitepaper.htm.

Whitlockjr., J. P. (1993). Mechanistic as-pects of dioxin action. Chern. Res. Tox-icol. 6, 754-763.

Zubay, G. L., Parson, W. W. and Vance,D. E. (1995). Principles 0/ Biochern-istry. Dubuque, USA: Wm. C. BrownCommunications, Inc.

Zumstein, M. (2005). Establishing aquantitative structure-activity relation-ship for the a/ß selectivity of the thy-roid receptor. M.Sc. thesis, Universityof Basel.

AcknowledgementsThis research had been made possiblythrough grants by the Swiss NationalScience Foundation, the Margaret andFrancis-Fleitmann Foundation, Luceme/Switzerland and the Jacques en DollyGazan Foundation, Zug/Switzerlandwhich are all gratefully acknowledged.

Correspondence toProf. Dr. Angelo VedaniBiografik-Labor 3RFriedensgasse 35CH-4055 BaselSwitzerlandtel: +41-61-261 4256fax: +41-61-261 4258e-mail: ange1o@biograf.ch

ALTEX 22, 3/05