edinburgh research explorer · biosynthesis pathway' british journal of pharmacology, vol. 174, no....

Edinburgh Research Explorer

Is systems pharmacology ready to impact upon therapydevelopment?

Citation for published version:Benson, H, Watterson, S, Sharman, J, Mpamhanga, C, Parton, A, Southan, C, Harmar, A & Ghazal, P2017, 'Is systems pharmacology ready to impact upon therapy development? A study on the cholesterolbiosynthesis pathway' British Journal of Pharmacology, vol. 174, no. 23. DOI: 10.1111/bph.14037

Digital Object Identifier (DOI):10.1111/bph.14037

Link:Link to publication record in Edinburgh Research Explorer

Document Version:Publisher's PDF, also known as Version of record

Published In:British Journal of Pharmacology

Publisher Rights Statement:This is an open access article under the terms of the Creative Commons Attribution License, which permits use,distribution and reproduction in any medium, provided the original work is properly cited.

General rightsCopyright for the publications made accessible via the Edinburgh Research Explorer is retained by the author(s)and / or other copyright owners and it is a condition of accessing these publications that users recognise andabide by the legal requirements associated with these rights.

Take down policyThe University of Edinburgh has made every reasonable effort to ensure that Edinburgh Research Explorercontent complies with UK legislation. If you believe that the public display of this file breaches copyright pleasecontact [email protected] providing details, and we will remove access to the work immediately andinvestigate your claim.

Download date: 05. Apr. 2019

https://doi.org/10.1111/bph.14037https://www.research.ed.ac.uk/portal/en/publications/is-systems-pharmacology-ready-to-impact-upon-therapy-development(b046e10a-c6fe-46a9-8ad7-c576cc23f06d).html

RESEARCH PAPER

Is systems pharmacology ready to impactupon therapy development? A study on thecholesterol biosynthesis pathway

Correspondence Steven Watterson, Northern Ireland Centre for Stratified Medicine, University of Ulster, C-Tric, Altnagelvin HospitalCampus, Derry BT47 6SB, UK. E-mail: [email protected]

Received 23 September 2016; Revised 10 August 2017; Accepted 30 August 2017

Helen E Benson1,*,†, Steven Watterson2,* , Joanna L Sharman1 , Chido P Mpamhanga3,‡, Andrew Parton2,Christopher Southan1, Anthony J Harmar3 and Peter Ghazal4,5

1Centre for Integrative Physiology, University of Edinburgh, Edinburgh, UK, 2Northern Ireland Centre for Stratified Medicine, University of Ulster,

C-Tric, Derry, UK, 3Centre for Cardiovascular Science, University of Edinburgh, The Queen’s Medical Research Institute, Edinburgh, UK, 4Division

of Infection and Pathway Medicine, University of Edinburgh Medical School, Edinburgh, UK, and 5Centre for Synthetic and Systems Biology, CH

Waddington Building, King’s Buildings, Edinburgh, UK

*Joint first authors.†Current address: TB Section, Respiratory Disease Department, National Infection Service, Public Health England, 61 Colindale Avenue,London NW9 5EQ, UK.‡Current address: LifeArc, Accelerator Building, SBC Open Innovation Campus, Stevenage SG1 2FX, UK.

BACKGROUND AND PURPOSEAn ever-growing wealth of information on current drugs and their pharmacological effects is available from online databases.As our understanding of systems biology increases, we have the opportunity to predict, model and quantify how drugcombinations can be introduced that outperform conventional single-drug therapies. Here, we explore the feasibility of suchsystems pharmacology approaches with an analysis of the mevalonate branch of the cholesterol biosynthesis pathway.

EXPERIMENTAL APPROACHUsing open online resources, we assembled a computational model of the mevalonate pathway and compiled a set of inhibitorsdirected against targets in this pathway. We used computational optimization to identify combination and dose options thatshow not only maximal efficacy of inhibition on the cholesterol producing branch but also minimal impact on the geranylationbranch, known to mediate the side effects of pharmaceutical treatment.

KEY RESULTSWe describe serious impediments to systems pharmacology studies arising from limitations in the data, incomplete coverage andinconsistent reporting. By curating a more complete dataset, we demonstrate the utility of computational optimization foridentifying multi-drug treatments with high efficacy and minimal off-target effects.

CONCLUSION AND IMPLICATIONSWe suggest solutions that facilitate systems pharmacology studies, based on the introduction of standards for data capture thatincrease the power of experimental data. We propose a systems pharmacology workflow for the refinement of data and thegeneration of future therapeutic hypotheses.

AbbreviationsAPI, Application Programme Interface; BPS, British Pharmacological Society; BRENDA, Braunschweig Enzyme Database;CID, compound identifier; FDA, US Food and Drug Administration; FDFT1, farnesyl-diphosphate farnesyl transferase 1;

This is an open access article under the terms of the Creative Commons Attribution License, which permits use, distribution and reproduction in any medium,provided the original work is properly cited.

BJP British Journal ofPharmacologyBritish Journal of Pharmacology (2017) 174 4362–4382 4362

DOI:10.1111/bph.14037 © 2017 The Authors. British Journal of Pharmacologypublished by John Wiley & Sons Ltd on behalf of British Pharmacological Society.

http://orcid.org/0000-0002-2750-6410http://orcid.org/0000-0002-5275-6446http://creativecommons.org/licenses/by/4.0/

GtoPdb, Guide to Pharmacology Database; HMGCR, hydroxymethylglutaryl-coa reductase; HMGCS1,hydroxymethylglutaryl-CoA synthase; HPC, high-performance computing; KEGG, Kyoto Encyclopedia of Genes and Ge-nomes; IUBMB, International Union of Biochemistry and Molecular Biology; IUPHAR, International Union of Basic andClinical Pharmacology; n2s, name-to-structure; ODE, ordinary differential equation; SBGN, Systems Biology GraphicalNotation; SBGN-ML, Systems Biology Graphical Notation Markup Language; SBML, Systems Biology Markup Language

Introduction

The expansion of available genomic and proteomic data hasenhanced our understanding of biomolecular interactionnetworks. Consequently, the development of systems biologyapproaches has enabled us to better understand how cellularbehaviour emerges from these networks (Boran and Iyengar,2010a). Systems-level approaches have been used to predictthe on- and off-target impacts of an intervention (Boran andIyengar, 2010b) and to identify the most sensitive compo-nents in pathways that suggest candidate drug targets(Benson et al., 2013). They also have the untapped potentialto suggest therapies comprising combinations of drugschosen to strategically reprogram biomolecular interactionnetworks in order to drive the system from a diseased to ahealthy state (Zhao et al., 2013; van Hasselt and van derGraaf, 2015; Watterson and Ghazal, 2010). This approach,known as systems pharmacology (Boran and Iyengar, 2010b;Westerhoff et al., 2015), is underpinned by the expansion inpathway, pharmacology and medicinal chemistry databases.

For example, WikiPathways held 804 human pathways(http://www.wikipathways.org/index.php/WikiPathways:Statistics) with 253 added in 2015 (Kutmon et al., 2016).Kyoto Encyclopedia of Genes and Genomes (KEGG) PATHWAYholds 518 pathway maps (Kanehisa et al., 2017) (http://www.kegg.jp/kegg/docs/statistics.html). Reactome currently holds2148 human pathways involving 10684 proteins and isoforms(http://reactome.org/stats.html) (Croft et al., 2014; Fabregatet al., 2016). ChEMBL version 23 (Gaulton et al., 2016) includes14675320 bioactivities, and the International Union of Basicand Clinical Pharmacology/British Pharmacological Society(IUPHAR/BPS) Guide to Pharmacology (GtoPdb) contains15281 curated interactions in its 2017.5 release (Southan et al.,2016). In 2016, the Food and Drug Administration (FDA) newdrug approvals fell to 22, following 45 approvals in 2015 (USFood and Drug Administration, 2016a; US Food and DrugAdministration, 2016bb). According to DrugBank release 5.0,their distinctmolecular count of approved small-molecule drugsis 2037 (Law et al., 2013).

As this catalogue of pharmacological interactions growsand our understanding of pathway systems expands, it willbe advantageous to integrate these resources in order to de-vise new potential therapies. Drug combination-based inter-ventions represent an opportunity for therapy developmentthat can yield one-size-fits-all or personalized/stratified thera-pies, and they can target pathways precisely rather thanperturbing entire networks. Two US National Institute forHealth workshop white papers have made a strong case forsystems pharmacology (Sorger et al., 2011) as a way to reduceattrition in therapy, to stimulate drug development, to bridgethe gap between network biology and translational medicineand to enhance industrial–academic collaborations. Systemspharmacology is also likely to impact upon genomic

medicine (Westerhoff et al., 2015), Systems Pathology, Sys-tems Biology and Pharmacometrics (van der Greef andMcBurney, 2005; Vicini and van der Graaf, 2013) and thetools that could contribute to systems pharmacology havebeen described (Lehár et al., 2007; Berger and Iyengar, 2009;Kell and Goodacre, 2014).

Previous work under the domain of systems pharmacologyhas primarily focussed on pharmacokinetic–pharmacodynamicmodelling (Darwich et al., 2017). Industry has evaluated systemspharmacology as a tool to inform trial design in areas ofcardiovascular disease, endocrinology, neurogenerative disease,respiratory disease, oncology and infectious disease (Visser et al.,2014) and to inform regulatory development (Visser et al., 2014;Peterson andRiggs, 2015). There have a been a number of specificstudies of nerve growth factor (Benson et al., 2013), coagulation(Wajima et al., 2009), innate immunity (Madrasi et al., 2014), can-cer (Abaan et al., 2013) and atherosclerosis (Pichardo–Almarzaet al., 2015).

However, whilst there is much enthusiasm for systemspharmacology as a tool to improve the efficacy and safety ofthe drug development pipeline (van der Graaf and Benson,2011; Rostami-Hodjegan, 2012; Trame et al., 2016), the prac-tical challenges of systematically amalgamating pharmacol-ogy and pathway biology in a coherent framework have notbeen adequately addressed.

Here, we describe a systems pharmacology study of thecholesterol biosynthesis pathway, detailing the barriers toprogress that we encountered and suggesting solutions tothese impediments, before proposing amodel of how systemspharmacology studies could be conducted in future. In par-ticular, we build a dynamic ordinary differential equation(ODE) model of the pathway, which we parameterize as faras possible from the literature. We identify relevant pharma-cological agents that act on this pathway and parameterizethem as far as possible from the literature and online data-bases. We then use computational optimization techniquesto identify the drug combinations that are most effective atsuppressing the outputs of the pathway that lead to choles-terol production and that minimize off-target effects. In com-pleting our analysis, we identify many of the problems thatprevent this type of work being undertaken routinely, andwe suggest solutions that would enable systems pharmacol-ogy to make a regular contribution to therapy development.

As explored in previous studies (Mazein et al., 2011; 2013;Watterson et al., 2013; Bhattacharya et al., 2014; Caspi et al.,2016), the cholesterol biosynthesis pathway is critical to bothcardiovascular health (Lewington et al., 2007; Hendersonet al., 2016; Parton et al., 2016) and innate immunity (Blancet al., 2011; Lu et al., 2015; Robertson et al., 2016). As the tar-get of the statin class of drug, we would expect this pathwayto be amongst the most thoroughly characterized, and forthis reason, we have chosen it for our feasibility study of sys-tems pharmacology. For simplicity, we have focused on the

The feasibility of systems pharmacology BJP

British Journal of Pharmacology (2017) 174 4362–4382 4363

http://www.wikipathways.org/index.php/WikiPathways:Statisticshttp://www.wikipathways.org/index.php/WikiPathways:Statisticshttp://www.kegg.jp/kegg/docs/statistics.htmlhttp://www.kegg.jp/kegg/docs/statistics.htmlhttp://reactome.org/stats.htmlhttp://www.guidetopharmacology.org/GRAC/LigandDisplayForward?ligandId=2718

segment of the pathway that transforms acetyl-CoA tosqualene and that forks to produce geranylgeranyl-di-phosphate. As a precursor to cholesterol, we would expectsqualene synthesis to track cholesterol synthesis and so weuse this as a proxy. The branch of the pathway that producesgeranylgeranyl-diphosphate has been shown tomediate boththe innate immune response (Blanc et al., 2011) and themyopathy side-effects associated with statin treatment(Wagner et al., 2011). Any intervention that demonstrates aminimal impact on this branch will avoid one of thesignificant side effects associated with standard cholesterollowering treatments.

Methods

Pathway productionWe started from the representations available in KEGG(Kanehisa et al., 2014), MetaCyc (Caspi et al., 2016) and theGtoPdb (Southan et al., 2016) taking these resources to berepresentative of the community of online pathway data-bases. We reviewed the primary literature to establish thestructure of the mevalonate portion of the cholesterolbiosynthesis pathway, in particular the enzymes involvedin the pathway, the reactions they catalyse, their subcellularlocalization, the species in which they were identified andany known isoforms.

Diagrams of the pathway were created using the SystemsBiology Graphical Notation (SBGN) (Le Novère et al., 2009),the yEd diagram software (yWorks GmbH, http://www.yworks.com/products/yed) and the SBGN-ED add-on toVANTED (Czauderna et al., 2010). From these diagrams, webuilt kinetic models as systems of ODEs.

The ODE model of this pathway was built usingMichaelis–Menten kinetics to describe each step except theinteractions consuming isopentenyl diphosphate andproducing geranylgeranyl diphosphate and pre-squalenediphosphate. These steps were described using mass actionkinetics in order to simplify the process of calculating thesteady state of the model and hence the steady state behav-iour of the pathway. Mass action kinetics were justified bythe expectation that the pathway interactions would operatefar from substrate saturation making the dynamics robustagainst small fluctuations in enzyme concentration. Mass ac-tion rate constants were calculated from the Kcat, Km and Kiparameters as described elsewhere (Watterson et al., 2013)and enzyme concentrations were taken from experimentallymeasured values (Watterson et al., 2013).

The pathway map and the associated mathematicalmodel are available from the Supporting Information FilesS1 and S2 as Systems Biology Graphical Notation MarkupLanguage (SBGN-ML) (Van Iersel et al., 2012) and SystemsBiology Markup Language (SBML) files (Hucka et al., 2003)respectively.

Pathway parameterizationWe identified the kinetic parameters that quantify each reac-tion unambiguously (Km and Kcat) using the BraunschweigEnzyme Database (BRENDA) (Chang et al., 2015) and verifiedthe values described against those in the primary literature. In

many instances, enzymes were associated with multiplekinetic parameter sets. We selected kinetic parameters basedupon the following criteria: (i) specificity to the wild-typeenzyme in one of the three main mammalian model species:human, mouse or rat; (ii) sourced from a primary literaturereference describing in vivo or in vitro experimental data asopposed to computationally derived structural modellingdata; and (iii) sourced from a reference that could be accessedand therefore verified. For many enzymes, this yielded arange of values for each parameter, and where this was thecase, we used the mean of the values obtained.

Inhibitor listInhibitor compounds not already indexed in GtoPdb wereidentified for each reaction from ChEMBL and BRENDA,databases that we took to be representative of the commu-nity of target databases. We included a compound in ourset if it met three criteria: (i) the enzyme used in the assaywas wild-type from one of the three main mammalianmodel species: human, mouse or rat; (ii) an experimentallydetermined reaction-specific inhibition constant (Ki) wasreported; and (iii) the assay conditions were reported.Crucially, all data were checked against the primary litera-ture references. Where this yielded a range of inhibitionconstants for nominally identical compounds, the mostpotent Ki values were used.

We verified the correct chemical structures of the inhibi-tors by cross-referencing the original references against theonline chemical databases PubChem (Kim et al., 2016) andChemSpider (Pence andWilliams, 2010). The actual chemicalstructures of the marketed statin drugs were established bychecking the FDA labels and the international non-proprietary name-assigned structures on the World HealthOrganization MedNet site (https://mednet-communities.net/inn). Comparison of unique structural identifiers allowedus to identify duplicates within the ChEMBL, BRENDA andliterature-derived dataset, and to establish whether the chem-ical structure reported in a given reference matched themarketed drug or research compound structures.

Curated content describing the enzymes in this pathway,their substrates and small molecule inhibitors was used toconsolidate and expand GtoPdb using the same approachand guidelines as described elsewhere (Pawson et al., 2014).The enzymes, list of inhibitors and kinetic parameters arenow all updated in the July 2016.3 release of GtoPdb.

Hypothesis generationWe combined ODE kinetic models, the pathway parametersand the inhibitor parameters to create a model describingthe dynamics of the mevalonate pathway. We sought toidentify the drug combination that would best suppressthe production of squalene as a precursor for cholesterol,but would also maintain production of geranylgeranyl-diphosphate at the same levels as in the absence of any in-hibitors, thereby eliminating a significant side-effect oftreatment. Firstly, we identified the steady-state activity ofthe pathway in the absence of any inhibitors. Then we usedcomputational optimization to identify the drug combina-tion that, at steady state, minimized squalene production,but left geranylgeranyl diphosphate production the sameas in the absence of inhibitors.

BJP H Benson et al.

4364 British Journal of Pharmacology (2017) 174 4362–4382

http://www.guidetopharmacology.org/GRAC/LigandDisplayForward?ligandId=3038http://www.guidetopharmacology.org/GRAC/LigandDisplayForward?ligandId=3054http://www.guidetopharmacology.org/GRAC/LigandDisplayForward?ligandId=3052http://www.guidetopharmacology.org/GRAC/LigandDisplayForward?ligandId=3052http://www.guidetopharmacology.org/GRAC/FamilyDisplayForward?familyId=104http://www.guidetopharmacology.org/GRAC/FamilyDisplayForward?familyId=104http://www.yworks.com/products/yedhttp://www.yworks.com/products/yedhttp://www.guidetopharmacology.org/GRAC/LigandDisplayForward?ligandId=3053http://www.guidetopharmacology.org/GRAC/LigandDisplayForward?ligandId=3053https://mednet-communities.net/innhttps://mednet-communities.net/inn

This was implemented using the Genetic Algorithm func-tion available on Matlab (MathWorks, http://www.mathworks.com) in parallel with a population size of 200and a function tolerance of 10�6. Matlab was chosen as themodelling platform for its flexibility, stability and compre-hensive libraries. The genetic algorithm started with one in-stance of a set of drug concentrations where each drug wasassigned a concentration equal to its Ki. A 199 further in-stances of sets of drug concentrations were automaticallygenerated from this instance by adding Gaussian noise tothe concentration of each drug (with standard deviation 1,the default setting). These 200 instances comprised the firstgeneration of candidate interventions. All instances of setsof concentrations were evaluated for their efficacy at sup-pressing squalene synthesis whilst maintaining geranylger-anyl diphosphate production. Two hundred new instanceswere created as a second generation of candidate interven-tions from the two most effective instances of the first gener-ation and with the addition of Gaussian noise. The 200 newinstances were then themselves evaluated with the two mosteffective instances used to generate a further 200 new in-stances, the third generation. This process was repeated untilwe arrived at instances from which no improvement in effi-cacy could be found for 20 consecutive generations, at whichpoint we interpreted the best performing instance identifiedthus far as optimal.

Nomenclature of targets and ligandsKey protein targets and ligands in this article are hyperlinkedto corresponding entries in http://www.guidetopharma-cology.org, the common portal for data from theIUPHAR/BPS Guide to PHARMACOLOGY (Southan et al.,2016), and are permanently archived in the Concise Guideto PHARMACOLOGY 2015/16 (Alexander et al., 2015).

Results

Pathway productionWe produced the model of the mevalonate arm of the choles-terol biosynthesis pathway shown in Figure 1 in SBGN nota-tion, describing the sequence of metabolic steps that leadfrom acetyl-CoA and acetoacetyl-CoA consumption tosqualene and geranylgeranyl diphosphate production. Thispathway comprises 12 steps (see Table 1), involving 10 en-zymes and 14 metabolites.

The parameters required for the resulting ODE model areshown in Table 1. After pooling results across mouse, humanand rat models, we were able to obtain experimental valuesfor only 12 out of the 24 required parameters. Across the stud-ies reported, pH values ranged from 7.0 to 8.0 and tempera-tures ranged from 25°C to 37°C, although in some studies,neither pH nor temperature values were given. When verifiedagainst the primary references, we found that one parametervalue obtained from BRENDA was missing from the literaturereference provided, suggesting that it had been misattributed[Kcat = 0.023/s for hydroxymethylglutaryl-Coa reduc-tase (HMGCR)]. A second parameter had been transcribed(for diphosphomevalonate decarboxylase) where theliterature source contradicted itself, specifying Km = 10 μM

in the abstract and Km = 10 mM in the manuscript. Becausecomputational hypothesis generation is highly sensitive tothe values of the parameters, ambiguous or inaccuratereporting can have a significant impact on any predictionsmade.

Substrates were reported in varying levels of structural de-tail. Common names were used that could refer to multipleexplicit forms of a chemical structure. However, variationsin the chirality and chemical structure can significantly affectsubstrate affinity. The relative enzyme concentrations hadbeen inferred previously (Watterson et al., 2013) and arelisted in Table 2.

Figure 1The mevalonate arm of the cholesterol biosynthesis pathway.



http://www.mathworks.comhttp://www.mathworks.comhttp://www.guidetopharmacology.orghttp://www.guidetopharmacology.orghttp://www.guidetopharmacology.org/GRAC/LigandDisplayForward?ligandId=3039http://www.guidetopharmacology.org/GRAC/ObjectDisplayForward?objectId=639http://www.guidetopharmacology.org/GRAC/ObjectDisplayForward?objectId=639http://www.guidetopharmacology.org/GRAC/ObjectDisplayForward?objectId=642

Supporting Information Table S1 compares representa-tions of the cholesterol biosynthesis pathway across themainpublicly available pathway and chemical databases. Itincludes a summary of cross-referencing between databaseswith standard identifiers for unambiguous representation,which will be essential for future cross-platforminteroperability.

InhibitorsThe inhibitors obtained from GtoPdb, BRENDA and theliterature, together with their inhibition constants (Ki), arelisted in Table 3. Six of the 10 enzyme targets had quantifiedparameters in humans. It was necessary to include two inhib-itors that had been only reported for rat enzymes [L-659,699for hydroxymethylglutaryl-CoA synthase (HMGCS1)and 3-hydroxy-3-methyl-6-phosphohexanoic acid forphosphomevalonate kinase] in order to maximizecoverage of the pathway. Two enzyme paralogues(isopentenyl diphosphate Δ-isomerases 1 and 2) hadno reported inhibitors with available Ki values, representinga region of the pathway that cannot currently be modulatedin our modelling process. This can be contrasted with theenzymes HMGCR and farnesyl diphosphate synthase,each of which had an extensive list of inhibitors. Inhibitionconstants could be obtained for 8 of the 10 enzymes in thepathway. Where reported, these values came from studiesconducted across a range of pH levels from 6.8 to 7.5 andtemperatures from 25°C to 37°C.

Both explicit structure and name-to-structure (n2s)ambiguities existed around the reporting of inhibitor enti-ties. In some cases, the common or trade name of a com-pound was used, without specification of the exactchemical structure and stereochemistry. In other cases, wefound a different n2s assignment across different databaseresources or indeed within the same resource. For example,under the HMGCS1 entry of BRENDA, the same inhibitoris listed twice as L-659,699 and (E,E)-11-[3-(hydroxy-methyl)-4-oxo-2-oxytanyl]-3,5,7-trimethyl-2,4-undecadienenoic acid.

Several results recorded in ChEMBL were transcribedagainst the incorrect drug target. Three inhibitors listedagainst the enzyme HMGCS1 describe results obtained fromexperiments with HMGCR (Balasubramanian et al., 1989).There were also cases where the incorrect species had been re-corded. For example, the compound with ChEMBL IDCHEMBL88601 cited in one study (Procopiou et al., 1994)(ChEMBL document ID CHEMBL1151052) is listed againstthe human squalene synthase (FDFT1) enzyme, whilst infact, the paper describes results for the yeast Candida albicansand rats.

Hypothesis generationIn order to complete the gaps in the available parameter sets,we proceeded by assuming that where parameters were takenfrom separate studies, the same metabolite chemicalstructures were referenced. For all the unknown parameters,we substituted a single representative value, obtained byaveraging across all known corresponding parameters.

Calculating the steady-state behaviour of the pathway inthe absence of inhibitors yielded the profile of flux shownon the left of Figure 2A, which we take to be wild-type

behaviour. Using computational optimization, we identifiedthe following drug combination that produced the steady-state profile of flux shown in the middle of Figure 2A and inFigure 2B: L-659,699 = 0.0294 nM, rosuvastatin = 2.60 nM,farnesyl thiodiphosphate = 0.0340 nM, cinnamicacid = 0.00104 nM, 6-fluoromevalonate 5-diphosphate = 0.0213 nM, zoledronic acid = 9.97 nM,BPH-628 = 5.86 nM; zaragozic acid A = 0.755 nM (seeTable 3 and Supporting Information Tables S2 and S3). Here,the production of squalene, a precursor of cholesterol, isheavily suppressed and the production of geranylgeranyldiphosphate is maintained at wild-type levels. In Figure 2B,we see specifically the flux at endpoints of the two pathwaybranches. With this drug combination, the flux fromgeranylgeranyl diphosphate → protein prenylation is thesame between the wild-type (inhibitor free) case and theoptimal multi-drug intervention case. However, the flux fromsqualene → cholesterol synthesis has been significantlysuppressed.

In Figure 2A, B, we compare the flux profiles for wild-type and optimal multi-drug interventions to the casewhere rosuvastatin, a type of statin, is applied alone. Thisinhibitor targets the interaction catalysed by HMGCR, andwe chose a concentration sufficient to suppress the rate ofsqualene formation and consumption to the same extentas the multi-drug intervention. As can be seen in Figure 2B, rosuvastatin intervention impacts upon both branchesof the pathway, suppressing geranylgeranyl diphosphateformation and protein prenylation as an off-target effectof treatment.

Interestingly, a concentration of 362 nM rosuvastatin wasrequired to achieve the same level of squalene suppression asthe multi-drug intervention. The greatest individual drugconcentration required in the optimal multi-drug interven-tion was 9.97 nM, and the total combined concentrationwas 19.3 nM, a dramatically lower dosage.

The value of drug combinations can also be seen inSupporting Information Figure S1 where we consider the im-pact of pairs of drugs (Lehár et al., 2007). Here, we see thatdrug pairs with targets above the fork inhibit the flux throughboth pathway endpoints (Supporting Information FigureS1A, B). Drug pairs with targets above and below the forktogether inhibit the flux through the cholesterol synthesiz-ing branch (Supporting Information Figure S1C, D). How-ever, drug pairs with targets above and below the fork athigh doses can have a low impact on the flux through theprotein prenylation branch (Supporting Information FigureS1E, F). Critically, Supporting Information Figure S1B, Eshows that concentrations can be selected that significantlysuppress the cholesterol synthesizing branch but that donot suppress the protein prenylation branch. The resultsdemonstrate comparable impact to the multi-drug interven-tion described above, but at higher individual and combinedconcentrations.

In order to identify the optimal multi-drug combination,it was necessary to use a high-performance computing(HPC) platform. However, the HPC demands were modest.Using an eight-node desktop computer running MATLAB inparallel, we can see that the score (a dimensionless value,greater than or equal to zero, that quantifies how effectivelythe best performing multi-drug intervention identified

BJP H Benson et al.


http://www.guidetopharmacology.org/GRAC/LigandDisplayForward?ligandId=5886http://www.guidetopharmacology.org/GRAC/ObjectDisplayForward?objectId=638http://www.guidetopharmacology.org/GRAC/LigandDisplayForward?ligandId=3202http://www.guidetopharmacology.org/GRAC/ObjectDisplayForward?objectId=641http://www.guidetopharmacology.org/GRAC/ObjectDisplayForward?objectId=646http://www.guidetopharmacology.org/GRAC/ObjectDisplayForward?objectId=646http://www.guidetopharmacology.org/GRAC/ObjectDisplayForward?objectId=647http://www.guidetopharmacology.org/GRAC/ObjectDisplayForward?objectId=644http://www.guidetopharmacology.org/GRAC/ObjectDisplayForward?objectId=645http://www.guidetopharmacology.org/GRAC/LigandDisplayForward?ligandId=2954http://www.guidetopharmacology.org/GRAC/LigandDisplayForward?ligandId=3216http://www.guidetopharmacology.org/GRAC/LigandDisplayForward?ligandId=3203http://www.guidetopharmacology.org/GRAC/LigandDisplayForward?ligandId=3203http://www.guidetopharmacology.org/GRAC/LigandDisplayForward?ligandId=3205http://www.guidetopharmacology.org/GRAC/LigandDisplayForward?ligandId=3205http://www.guidetopharmacology.org/GRAC/LigandDisplayForward?ligandId=3177http://www.guidetopharmacology.org/GRAC/LigandDisplayForward?ligandId=3188http://www.guidetopharmacology.org/GRAC/LigandDisplayForward?ligandId=3057

Table 1A list of the enzymes of the mevalonate branch of the cholesterol synthesis pathway, with gene and protein identifiers and EC numbers

E.Cnumber

Enzyme/GtoPdbtarget ID

UniProtID

HGNCID

IUBMBenzymeapproved name

Reactioncatalysed

Km(mM)/PMID

Reportedsubstrate/GtoPdb ligand ID

Kcat(s-1)/PMID

2.3.3.10 HMGCS1/638

Q01581 5007 Hydroxymethylglutaryl-CoA synthase

Acetyl CoA + H2O+ acetoacetylCoA = (S)-3-hydroxy-3-methylglutaryl-CoA +coenzyme A

0.009/6118268

Acetyl-CoA/3038

–

– – – – – 0.2/6118268 Acetyl-CoA/3038

–

– – – – – 0.073/19706283

Acetyl-CoA/3038

–

– – – – – 0.076/19706283

Acetyl-CoA/3038

–

– – – – – 0.084/19706283

Acetyl-CoA/3038

–

– – – – – 0.029/7913309

Acetyl-CoA/3038

–

1.1.1.34 HMGCR/639

P04035 5006 Hydroxymethylglutaryl-CoA reductase(NAPDH)

(S)-3-Hydroxy-3-methylglutaryl-CoA + 2 NADPH =mevaldyl CoA+ 2NADP+

0.006/4985697

3-Hydroxy-3-methylglutaryl-CoA/3040

–

– – – – – 0.012/4985697


–

– – – – Mevaldyl CoA+ 2NADP+ = (R)-mevalonate +coenzymeA + 2 NADP+

0.01/10392455


–

– – – – – 0.014/10392455


–

– – – – – 0.015/10392455


–

– – – – – 0.019/10392455


–

– – – – – 0.024/10392455


–

– – – – – 0.07/16128575


–

– – – – – 0.6/�� 3-Hydroxy-3-methylglutaryl-CoA/3040

–

– – – – – 0.068/18446881

Hydroxymethylglutaryl-CoA

0.023/18446881

– – – – – 0.004/7077140

S-3-Hydroxy-3-methylglutaryl-CoA/3040

–

continues



Table 1(Continued)

E.Cnumber

Reactioncatalysed Organism

Reportedconditions

MeanKm(mM)

SubstitutedmeanKm

SubstitutedmeanKcat

2.3.3.10 Acetyl CoA + H2O+ acetoacetylCoA = (S)-3-hydroxy-3-methylglutaryl-CoA +coenzyme A

Rattusnorvegicus

absence ofacetoacetyl-CoA, hydrolysisreaction

0.0785 – 6.651575

– Rattusnorvegicus

0.01 Macetoacetyl-CoA

– – –

– Homosapiens

– – – –

– Homosapiens

– – – –

– Homosapiens

– – – –

– Homo sapiens – – – –

1.1.1.34 (S)-3-Hydroxy-3-methylglutaryl-CoA + 2 NADPH =mevaldyl CoA+ 2NADP+

Rattusnorvegicus

Only oneenantiomer

0.0765 – 0.0023

– Rattus norvegicus – – – –

Mevaldyl CoA+ 2NADP+ = (R)-mevalonate +coenzymeA + 2 NADP+

Mus musculus Enzyme fromtumour

– – –

– Mus musculus Enzyme from liverand tumour

– – –

– Mus musculus Enzyme from liver,implanted tumour

– – –

– Mus musculus Enzyme from liver,implanted tumour

– – –

– Mus musculus Enzymefrom liver

– – –


– Homo sapiens pH 7.5/Tempnot specified

– – –



2.7.1.36 ATP + (R)-mevalonate= ADP +(R)-5-phosphomevalonate

Rattusnorvegicus

pH 7.5/25C 0.0337 – –

– Rattusnorvegicus

pH 7.5/34C – – –

– Homo sapiens pH 7.5/30C – – –


2.7.4.2 ATP + (R)-5-phosphomevalonate =

Homosapiens

pH 7.0/30C 0.034 – 6.651575

continues

BJP H Benson et al.


Table 1 (Continued)

E.Cnumber


UniProtID

HGNCID


Reactioncatalysed

Km(mM)/PMID


Kcat(s-1)/PMID

2.7.1.36 MVK/640 Q03426 7530 Mevalonatekinase

ATP + (R)-mevalonate= ADP +(R)-5-phosphomevalonate

0.035/14680942

(RS)-mevalonate/3056

–

– – – – – 0.035/17964869


21.9/18302342

– – – – – 0.0408/18302342


–

– – – – – 0.024/9325256

Mevalonate/3056

–

2.7.4.2 PMVK/641 Q15126 9141 Phosphomevalonatekinase

ATP + (R)-5-phosphomevalonate =ADP + (R)-5-diphosphomevalonate

0.034/17902708

(R)-5-Phosphomevalonate/3046

–

4.1.1.33 MVD/642 P53602 7529 Diphosphomevalonatedecarboxylase

ATP + (R)-5-diphosphomevalonate =ADP + phosphate+ isopentenyldiphosphate + CO2

0.02/8744421

5-Diphosphomevalonate/3055

–

– – – – – 0.0289/18823933


4.5/18823933

– – – – – 0.036/16626865


–

– – – – – 0.036/17888661


–

– – – – – 0.01/11913522

Mevalonatediphosphate/3055

–

5.3.3.2 IDI1 andIDI2*/646& 647

Q13907/Q9BXS1

5387/23487

Isopentenyl-diphosphatedelta isomerase

Isopentenyldiphosphate =dimethylallyldiphosphate

0.0228/17202134

Isopentenyldiphosphate/3048

–

– – – – – 0.033/8806705


–

2.5.1.1 FDPS/644 P14324 3631 Farnesyldiphosphatesynthase

Dimethylallyldiphosphate +isopentenyldiphosphate= diphosphate +geranyl diphosphate

– – –

2.5.1.10 – – – – Geranyl diphosphate+ isopentenyldiphosphate =diphosphate +trans,trans-farnesyldiphosphate

– – –

2.5.1.1 GGPS1/643 O95749 4249 Farnesyltranstransferase Dimethylallyldiphosphate +isopentenyldiphosphate= diphosphate+ geranyldiphosphate

– – –



Table 1 (Continued)

E.Cnumber

Reactioncatalysed Organism

Reportedconditions

MeanKm(mM)

SubstitutedmeanKm

SubstitutedmeanKcat

ADP + (R)-5-diphosphomevalonate

4.1.1.33 ATP + (R)-5-diphosphomevalonate =ADP + phosphate+ isopentenyldiphosphate + CO2

Rattus norvegicus – 0.0262 – –

– Homo sapiens 30C – – –



– Mus musculus pH 7.2 – – –

5.3.3.2 Isopentenyldiphosphate =dimethylallyldiphosphate

Homo sapiens pH 8.0 0.0279 – 6.651575


2.5.1.1 Dimethylallyldiphosphate +isopentenyldiphosphate= diphosphate +geranyl diphosphate

– – – 0.0351375 6.651575

2.5.1.10 Geranyl diphosphate+ isopentenyldiphosphate =diphosphate +trans,trans-farnesyldiphosphate

– – – 0.0351375 6.651575

2.5.1.29 Trans,trans-farnesyldiphosphate +isopentenyldiphosphate =diphosphate +geranylgeranyldiphosphate

Rattus norvegicus pH 7.0/37C 0.0027 – –


– Rattus norvegicus pH 7.0/37C – – –


2.5.1.21 2 Trans,trans-farnesyldiphosphate =diphosphate +presqualenediphosphate

Homo sapiens – 0.0016 – 6.651575

Presqualenediphosphate+ NAD(P)H + H+

= trans-squalene +diphosphate+ NAD(P)+

Rattus norvegicus – – – –

BJP H Benson et al.


achieves our objective, with zero indicating success)converges rapidly on an effective drug combination. Itsuccessfully identified an optimal combination in 46 minand achieved an approximately optimal solution in less than10 min.

The results of our curation of the pathway and the inhib-itors that target it are available in GtoPdb at http://www.guidetopharmacology.org/GRAC/FamilyDisplayForward?familyId=104, an example of which is shown in SupportingInformation Figure S2.

The model produced is available from http://biomodels.org (Chelliah et al., 2013) (ID: MODEL1506220000).

Discussion

The importance of systems pharmacologyMulti-drug interventions. Multi-drug approaches are alreadyemployed in areas including HIV and oncology (Petrelli andGiordano, 2008; Thakur and Marchand, 2012). However, theexisting interventions have typically been developedheuristically, rather than through systematic studies of thepathways involved, requiring significant domain expertise andsubjective judgement. Systems pharmacology introducesobjective metrics that have the potential to transform therapy

Table 1 (Continued)

E.Cnumber


UniProtID

HGNCID


Reactioncatalysed

Km(mM)/PMID


Kcat(s-1)/PMID

2.5.1.10 – – – – Geranyl diphosphate+ isopentenyldiphosphate =diphosphate +trans,trans-farnesyldiphosphate

– – –

2.5.1.29 – – – – Trans,trans-farnesyldiphosphate +isopentenyldiphosphate =diphosphate +geranylgeranyldiphosphate

0.0029/17846065


–

– – – – – 0.003/16698791


–

– – – – – 0.00071/17846065

Trans,trans-farnesyldiphosphate/3050

–

– – – – – 0.0042/16698791

Trans,trans-farnesyldiphosphate/3050

0.204/16698791

2.5.1.21 FDFT1/645 P37268 3629 Farnesyl-diphosphatefarnesyltransferase 1

2 Trans,trans-farnesyldiphosphate =diphosphate +presqualenediphosphate

0.0023/9473303

Farnesyldiphosphate/2910

–

– – – – Presqualenediphosphate+ NAD(P)H + H+

= trans-squalene +diphosphate+ NAD(P)+

0.001/1569107

Trans-farnesyldiphosphate/3050

–

Reported substrates, kinetic values and details of the experimental studies from which they were obtained, along with references are recorded.Please note that ligands outlined in the table are listed using the nomenclature from the original literature. Where the reference did not specify theisomer used experimentally, it was assumed the racemate was used.FDPS, farnesyl diphosphate synthase; IDI1, isopentenyl diphosphate delta isomerase 1; IDI2, isopentenyl diphosphate delta isomerase 2; MVD,diphosphomevalonate decarboxylase; MVK, mevalonate kinase; PMVK, phosphomevalonate kinase.



http://www.guidetopharmacology.org/GRAC/FamilyDisplayForward?familyId=104http://www.guidetopharmacology.org/GRAC/FamilyDisplayForward?familyId=104http://www.guidetopharmacology.org/GRAC/FamilyDisplayForward?familyId=104http://biomodels.orghttp://biomodels.org

development, yielding therapeutichypothesesmore rapidly andcost-effectively.

Many diseases are multifactorial in nature, involvingmultiple pathways in their pathology. Effective futuretherapies will likely employ multi-drug approaches thattargetmultiple points in the network of pathways responsible(i.e. polypharmacology). Promiscuous drugs can beincorporated advantageously into the generation of thesehypothetical interventions, provided that their interactionsare known and parameterized.

Multi-drug approaches can minimize the pleotropiceffects of an intervention. As we demonstrated for statins,where a single drug intervention suppressed the output of apathway to the same extent as multiple drugs targeting thesame pathway, not only was the dose of each of the multipledrugs significantly lower than the dose of the single drugbut also the combined dose of all of the multiple drugs wassignificantly lower than the dose of the single drug. Thisintrinsically reduces the risk from off-target or pleotropiceffects for each drug.

The systems pharmacology approach allows us to pre-dict multi-drug strategies that may be optimal to treat adisease and can be used as a prioritization triage for futuredrug development. It can support personalized and strati-fied medicine, where we adapt the parameter sets of theunderlying models of pathway activity to represent an in-dividual (for personalized medicine) or a subpopulation(for stratified medicine) and we develop interventions thatare customized to be optimal for the patient or patientgroup. A challenge lies in developing optimized therapiesso that they preferentially target key tissues. Pathwaymodels and pharmacological interactions can be madetissue specific by generating a new parameter set for eachtissue. Hypothesis generation would then use optimizationto determine an intervention that impacted upon a keypathway in a key tissue, leaving other pathways unchangedacross all tissues and with a minimal impact on the keypathway in non-targeted tissues.

Drug development. Fewmulti-drug treatmentsmake it throughthe development process. The number of combinationaltherapies listed in the Therapeutic Target Database at the timeof writing is 115 (Qin et al., 2014). A combination therapy,LCZ696, with the brand name Entresto, was approved in 2015and is in Phase III of clinical trials for the treatment ofcardiovascular disorders. Establishing drug combinations usinga conventional drug development pipeline creates significantchallenges as development essentially replicates the single drugdevelopment process multiple times. Systems pharmacology istherefore critical to expanding the range of multi-druginterventions available in a cost-effective manner. Although itmay add extra steps to the preclinical stages of the drugdevelopment process, it could have a significant positiveimpact on the cost-efficiency associated with each success byreducing the attrition rate in the later stages of the pipeline(Bowes et al., 2012).

Integrating our understanding of pharmacology and sys-tems biology will also enable us to make better predictionsof the behaviour of individual drugs. For example, squalenesynthase (FDFT1) has been investigated as a potential drugtarget that lies downstream of HMGCR, the target for statins,in the cholesterol biosynthesis pathway (see Figure 1). FDFT1catalyses an interaction after the fork to geranylgeranyl-diphosphate production, and it has been speculated thatsqualene synthase inhibitors might suppress cholesterolproduction without impacting on the geranylgeranyl-diphosphate producing branch, in contrast to statin treat-ment. However, squalene synthase inhibitors typically haveKi values orders of magnitude greater than the typical Ki forstatins (See Table 3). As a result, squalene synthase inhibitorconcentrations are required to be orders of magnitude greaterthan statin concentrations to suppress the correspondingenzyme activity comparably. Such high concentrations riskunforeseen off-target effects, making squalene synthaseinhibitors a higher risk drug to develop.

Systems level analysis. At the heart of systems pharmacologyis the growing recognition that we will only be able to trulyunderstand the best ways to therapeutically intervene inphysiological function by considering biology at a systemslevel. The network of interactions that mediatephysiological function is a dynamical system, and just ashealth and disease are different dynamical states of cells,tissues and organs, they also describe different dynamicalstates of the networks (Ahn et al., 2006). In a networkcontext, dynamical states can comprise a single stableconfiguration of the whole network or a sequence ofconfigurations that repeat cyclically and stably. However, itis the configuration (species concentrations, distributionsand structural conformations) of the network as a whole, orat least of critical subnetworks, that relate to phenotype,rather than any single component of the network (Lewisand Glass, 1991).

Small networks often yield dynamics that are intuitiveand predictable. However, as networks become larger andricher in structure, novel and often counter-intuitivedynamics can emerge and it will only be once we are ableto build high-confidence models at this scale that the fullpotential of systems level studies will be realized (Aderem,2005). Building high confidence networks at this scale is

Table 2Normalized enzyme levels

Enzyme Level

HMGCS1 1441

HMGCR 258

MVK 76

PMVK 874

MVD 111

IDI1 2707

IDI2 –

FDPS 7029

GGPS1 86

FDFT1 3425

FDPS, farnesyl diphosphate synthase; IDI1, isopentenyl diphosphatedelta isomerase 1; IDI2, isopentenyl diphosphate delta isomerase 2;MVD, diphosphomevalonate decarboxylase; MVK, mevalonate ki-nase; PMVK, phosphomevalonate kinase.

BJP H Benson et al.


Table

3List

ofinhibitors

forea

chof

theen

zymes

inthemev

alon

atebran

chof

thech

olesterolsyn

thesispa

thway

withK i

values

andreferenc

es

E.C

num

ber

Enzy

me

Inhib

itor

name/

Gto

Pdb

ligandID

InChi

key

Appro

ved

dru

g?

Org

anism

KiorIC

50

(nM)/

PMID

Rep

orted

conditions

Multidru

gco

nce

ntration

(nM)

2.3.3.10

HMGCS1

L-65

9699/58

86

ODCZJZWSX

PVLA

W-

KXCGKLM

DSA

-NNo

Rattus

norveg

icus

Ki=53

.7/

7913

309

–0.215

1.1.1.34

HMGCR

Rosuva

statin/2

954

BPRH

UIZQVS

MCRT

-YX

WZHEE

RSA-N

Yes

Homosapien

sKi

=2.3/

1612

857

5pH6.8,3

7C

5.48

Rosuva

statin/2

954

BPRH

UIZQVS

MCRT

-YX

WZHEE

RSA-N

Yes

Homosapien

sKi

=3.5/

1277

315

0–

–

Rosuva

statin/2

954

BPRH

UIZQVS

MCRT

-YX

WZHEE

RSA-N

Yes

Homosapien

sKi

=0.9/

1568

689

8–

–

Cerivastatin/29

50SE

ERZIQ

QUAZTO

L-ANMDKA

QQSA

-NYe

sHomosapien

sKi

=5.7/

1612

857

5–

–

Cerivastatin/29

50SE

ERZIQ

QUAZTO

L-ANMDKA

QQSA

-NYe

sHomosapien

sKi

=10

/12

77315

0–

–

Atorvastatin/

2949

XUKU

URH

RXDUEB

C-

KAYW

LYCHSA

-NYe

sHomosapien

sKi

=8/

1277

315

0–

–

Atorvastatin/

2949

XUKU

URH

RXDUEB

C-

KAYW

LYCHSA

-NYe

sHomosapien

sKi

=14

/16

12857

5–

–

Lova

statin/2739

PCZOHLX

UXFIOCF-

BXMDZJJM

SA-N

Yes

Homosapien

sKi

=0.6/

doi:1

0.102

1/np

5006

1a020

–

Lova

statin/2739

PCZOHLX

UXFIOCF-

BXMDZJJM

SA-N

Yes

Homosapien

sKi

=0.6/

6933

445

––

Simva

statin/2

955

RYMZZMVNJRMUDD-

HGQWONQES

A-N

Yes

Homosapien

sKi

=2.6/

1568

689

8–

–

CHEM

BL393

12/7991

VWKZ

OIO

UHUHQKZ-

HZPD

HXFC

SA-N

No

Homosapien

sKi

=3/

doi:1

0.101

6/S0

960-894X

(01)

8078

8-5

–

CHEM

BL391

02/7993

XKZCNQAYF

RBCKR-

HNNXBM

FYSA

-NNo

Homosapien

sKi

=16

/doi:1

0.101

6/S0

960-894X

(01)

8078

8-5

–

Fluv

astatin/29

51FJLG

EFLZ

QAZZCD-

MCBH

FWOFS

A-N

Yes

Homosapien

sKi

=27

5/

1612

857

5–

–

2.7.1.36

MVK

Farnesyl

thiodiph

ospha

te/321

6DRA

DWUUFB

CYM

DM-

UHFFFA

OYS

A-L

No

Homosapien

sKi

=29

/14

67922

5pH7.5,3

0C

0.050

0

2.7.4.2

PMVK

Cinnam

icac

id/3

203

WBYW

AXJH

AXSJNI-

VOTS

OKGWSA

-NNo

Rattus

norveg

icus

Ki=24

800

00/

2260

78

pH7.4,3

7C

0.024

0

Isoferulic

acid/798

0**

QURC

VMIEKC

OAJU-

HWKA

NZRO

SA-N

No

Rattus

norveg

icus

Ki=38

500

00/

2260

78

pH7.4,3

7C

–

XRC

IRZGXKW

CWNQ-

UHFFFA

OYS

A-N

No

Rattus

norveg

icus

Ki=14

500

0/

doi:1

0.102

1/ja004

93a0

44–

continue

s



Table

3(C

ontinue

d)

E.C

num

ber

Enzy

me

Inhib

itor

name/

Gto

Pdb

ligandID

InChi

key

Appro

ved

dru

g?

Org

anism

KiorIC

50

(nM)/

PMID

Rep

orted

conditions

Multidru

gco

nce

ntration

(nM)

3-hyd

roxy

-3-m

ethyl-

6-pho

spho

hexa

noic

acid/3

202

P-co

umaric

acid/5

787

NGSW

KAQJJW

ESNS-

ZZXKWVIFSA

-NNo

Rattus

norveg

icus

Ki=23

900

00/

2260

78

pH7.4,3

7C

–

4.1.1.33

MVD

6-fluo

romev

alon

ate5-

diph

ospha

te/3205

GLN

COGHKIHKS

A-

UHFFFA

OYS

A-N

No

Homo

sapien

sKi

=62

/18

82393

3pH7.0,3

0C

0.050

0

2-fluo

romev

alon

ate5-

diph

ospha

te/3204

WPX

HWHACORB

SDS-

UHFFFA

OYS

A-N

No

Rattus

norveg

icus

Ki=30

20/

1662

686

5pH7.5,2

5C

–

Diphospho

glycolyl

proline/320

6CDFD

GXYB

ANXCPC

-UHFFFA

OYS

A-N

No

Homosapien

sKi

=23

00/

1882

393

3–

–

CHEM

BL116

0330

/79

94

YERU

UUBB

RAPJND-

UHFFFA

OYS

A-N

No

Homosapien

sKi

=75

0/

doi:1

0.101

6/096

0-89

4X(96)00

374-5

–

CHEM

BL116

0328

/79

96

YIGLD

WRZ

XXHIG

Z-

ZCFIWIBFS

A-N

No

Homo

sapien

sKi

=37

/doi:1

0.101

6/096

0-89

4X(96)00

374-5

–

P0-geran

yl2-

fluo

romev

alon

ate

5-dipho

sphate/32

07

ACYP

MTK

DKJZHBJ-

MDWZMJQ

ESA-N

No

Homo

sapien

sKi

=41

69/

1882

393

3–

–

P0-geran

yl3,5,

9-trihyd

roxy

-3-methy

lnon

anate9-

diph

ospha

te/5621

PMUQIJK

CGIYWGT-

GZTJUZNOSA

-NNo

Rattus

norveg

icus

Ki=64

57/

1652

425

6–

–

5.3.3.2

IDI1

––

––

––

–

5.3.3.2

IDI2

––

––

––

–

2.5.1.1,2

.5.1.10

FDPS

Zoledron

icacid/3177

XRA

SPMIURG

NCCH-

UHFFFA

OYS

A-N

Yes

Homosapien

sKi

=0.07/

1832

789

9–

12.7

Zoledron

icacid/3177

VWKZ

OIO

UHUHQKZ-

HZPD

HXFC

SA-N

Yes

Homosapien

sKi

=85

/18

32789

9–

–

Risedronate/317

6IID

JRNMFW

XDHID

-UHFFFA

OYS

A-N

Yes

Homosapien

sKi

=0.36/

1832

789

9–

–

Risedronate/317

6IID

JRNMFW

XDHID

-UHFFFA

OYS

A-N

Yes

Homosapien

sKi

=81

/18

32789

9–

–

NE5

8062/316

6XUCBN

FJYK

WKAMN-

UHFFFA

OYS

A-N

No

Homosapien

sKi

=1/

1832

789

9–

–

NE9

7220/317

1NAIJO

BGUXRH

QJW

-UHFFFA

OYS

A-N

No

Homosapien

sKi

=1.09/

1832

789

9–

–

NE9

7220/317

1NAIJO

BGUXRH

QJW

-UHFFFA

OYS

A-N

No

Homosapien

sKi

=12

/18

32789

9–

–

continue

s

BJP H Benson et al.


Table

3(C

ontinue

d)

E.C

num

ber

Enzy

me

Inhib

itor

name/

Gto

Pdb

ligandID

InChi

key

Appro

ved

dru

g?

Org

anism

KiorIC

50

(nM)/

PMID

Rep

orted

conditions

Multidru

gco

nce

ntration

(nM)

NE5

8018/316

8XXNASZ

AYA

NFLID

-UHFFFA

OYS

A-N

No

Homosapien

sKi

=0.74

/1832

789

9–

–

NE5

8018/316

8XXNASZ

AYA

NFLID

-UHFFFA

OYS

A-N

No

Homosapien

sKi

=59

/18

32789

9–

–

2.5.1.1,

2.5.1.10,

2.5.1.29

GGPS

1BP

H-628/318

8MPB

UFK

ZCEB

TBSK

-UHFFFA

OYS

A-N

No

Homosapien

sKi

=20

/17

53589

5–

13.5

BPH-608/797

7YX

QQNSY

ZOQHKHD-

UHFFFA

OYS

A-N

No

Homosapien

sKi

=60

/17

53589

5–

–

BPH-675/797

5MZVW

VRVN

MXTD

AK-

UHFFFA

OYS

A-N

No

Homosapien

sKi

=70

/17

53589

5–

–

BPH-629/797

6BY

VXAUZOTG

ITQZ-

UHFFFA

OYS

A-N

No

Homosapien

sKi

=11

0/

1753

589

5–

–

BPH-676/797

8NWIARQ

RYIRVY

CM-

UHFFFA

OYS

A-N

No

Homosapien

sKi

=11

0/

1753

589

5–

–

2.5.1.21

FDFT

1Zarag

ozicacid

A/3

057

DFK

DOZMCHOGOBR

-NCSQ

YGPN

SA-N

No

Homosapien

sKi

=0.25/

7864

626

pH7.4,3

7C

4.79

CHEM

BL243

62/

3105

FBPJEW

KDFU

WVKV

-UHFFFA

OYS

A-N

No

Homosapien

sKi

=43

/doi:1

0.101

6/S0

960-

894X

(97)00

053-X

–

CHEM

BL120

8103

/31

20

HGDWHTA

SNMRJMP-

UHFFFA

OYS

A-N

No

Homosapien

sKi

=30

0/

1945

609

9Re

combinan

ten

zyme

expressed

inEsch

erichiacoli

–

CHEM

BL120

7858

/31

27

AGJZDRX

KAQZWEP

-UHFFFA

OYS

A-N

No

Homosapien

sKi

=52

0/

1945

609

9Re

combinan

ten

zyme

expressed

inE.

coli

–

BPH-830/312

1GNET

VUVZ

FYJATO

-UHFFFA

OYS

A-N

No

Homosapien

sKi

=53

0/

1945

609

9Re

combinan

ten

zyme

expressed

inE.

coli

–

SQ-109

/7997

JFIBVD

BTCDTB

RH-

REZTV

BANSA

-NNo

Homosapien

sKi

=74

0/

2248

671

0Re

combinan

ten

zyme

expressed

inE.

coli

–

[1-(Hyd

roxy

carbam

oyl)-

4-(3-phen

oxy

phe

nyl)

butyl]p

hospho

nate/3120

HGDWHTA

SNMRJMP-

UHFFFA

OYS

A-N

No

Homosapien

sKi

=30

2/

1945

609

9–

–

Compou

nd13

[PMID

:19

45609

9]/312

7AGJZDRX

KAQZWEP

-UHFFFA

OYS

A-N

No

Homosapien

sKi

=52

5/

1945

609

9–

–

(1-M

ethyl-1-{[3-

(3-phen

oxyp

hen

yl)

prop

yl]carbam

oyl}

ethyl)pho

sphon

ate/312

7

AGJZDRX

KAQZWEP

-UHFFFA

OYS

A-N

No

Homosapien

sKi

=52

5/

1945

609

9–

–

**Den

otesinteractionno

tlistedon

GtoPd

b.Th

esereac

tions

wereselected

from

either

BREN

DAor

ChEM

BLto

complete

theda

tasetrequiredforthemodellin

gprocess.

FDPS

,farne

syld

iphospha

tesynthase;

IDI1,isope

nten

yldipho

spha

tede

ltaisom

erase1;

IDI2,isopen

teny

ldipho

sphatedelta

isom

erase2;

MVD,d

ipho

sphom

evalona

tede

carboxy

lase;M

VK,

mev

alon

ate

kina

se;P

MVK

,phosph

omev

alon

atekina

se.



inherently challenging as we see here. Coherently andunambiguously parameterizing all the interactions of anetwork is a significant logistical challenge. However, wehave also seen that doing so enables us to identify andaddress the side-effects of treatment whilst the therapy isbeing designed, rather than retroactively. Hence, systems-level approaches are well suited to pharmacologicalapplications.

Current impediments to systems pharmacology

Problem 1: lack of systematic recordingThe absence of systematic and rigorous descriptions of me-tabolites and pharmaceutical compounds poses a significantchallenge. Example 1, fluvastatin consists of two

enantiomers, represented by PubChem compound identifiers(CIDs) 1548972 and 446155, with the 3R, 5S enantiomer(CID 446155) being significantly more pharmacologically ac-tive than the other (Di Pietro et al., 2006; Boralli et al., 2009).Commercial preparations used in vitro often vary in theirstereochemical composition, with both enantiomersavailable individually, as well as in a racemic mixture. How-ever, the authors did not always specify the stereochemicalcomposition used despite this necessarily impacting uponthe inhibition constant, Ki, reported. Example 2, mevalonateis a metabolite that occurs naturally in mammals as the (R)-isomer form. Sigma-Aldrich currently refers to its marketedversion as ‘(RS)-mevalonic acid’. However, in one study(Potter and Miziorko, 1997), the metabolite is obtained fromthe supplier Sigma-Aldrich, and it is recorded on BRENDA

Figure 2(A) The profile of flux through the pathway shown in Figure 1 described as a cone plot for the three scenarios: wild-type (treatment free), opti-mized multi-drug intervention and single-drug statin-like intervention. Cone size and colour both represent flux level. We show only the flux lead-ing to cholesterol synthesis [the flux to protein prenylation is presented in (B)]. Interactions are numbered by their product: (1: 3-hydroxy-3-methylglutaryl-CoA; 2: melvaldyl-CoA, 3: mevalonate, 4: mevalonate-P, 5: mevalonate diphosphate, 6: isopentenyl diphosphate, 7: dimethylallyldiphosphate, 8: geranyl diphosphate, 9: farnesyl diphosphate, 10: presqualene diphosphate, 11: squalene, 12: cholesterol synthesis). (B) The fluxthrough the endpoints of the two branches for the three scenarios: wild-type, optimized multi-drug intervention and single-drug statin-likeintervention. Flux through the squalene/cholesterol synthesis branch is shown in blue. Flux through the geranylgeranyl-PP/protein prenylationbranch is shown in red. The statin concentration has been selected to ensure that the flux through the cholesterol synthesis branch is the sameas in the multi-drug intervention. (C) Convergence on the optimal multi-drug intervention that suppresses cholesterol synthesis whilst minimizingoff target effects, shown against time and against generations of the genetic algorithm.

BJP H Benson et al.


http://www.guidetopharmacology.org/GRAC/LigandDisplayForward?ligandId=2951http://www.guidetopharmacology.org/GRAC/LigandDisplayForward?ligandId=3042http://www.guidetopharmacology.org/GRAC/LigandDisplayForward?ligandId=3042

under the general name ‘mevalonic acid’ without unambigu-ous chemical identifiers such as the Simplified Molecular-Input Line-Entry System or the International Chemical Iden-tifier. The isomer form affects the parameterization of the me-tabolite. Hence, this ambiguity creates potential inaccuracyin any resulting modelling.

Problem 2: reporting of the wrong dataWe found cases of incorrect or incomplete kinetic data reportedin the primary literature that undermined the ability to modelinteractions. Vmax values were regularly reported instead of Kcatvalues where Vmax is related to Kcat by Vmax = Kcat × (enzymeconcentration). For a Vmax value to be reusable in subsequentstudies, the enzyme concentration must also be reportedalongside it. However, we regularly found this not to be thecase, making most reported Vmax values unusable.

Similarly, inhibitors were frequently parameterized byIC50 values instead of Ki values, where Ki and IC50 are relatedby Ki = IC50/(1 + S/Km) and S is the substrate concentration.For IC50 values to be reusable in future studies, the substrateconcentrations must also be reported. Here, again, we foundregular omissions that rendered most reported IC50 valuesunusable.

Solution (1 and 2): introduce data capture standards that facilitateunambiguous reconstruction of the results withoutoptimization. Reporting must include clear and thoroughdescriptions of experimental configurations andunambiguous identification of chemical structures throughthe use of comprehensive and standard nomenclature. Pastexperience has shown that effective standards can bedeveloped through open community exercises (e.g. SBMLand SBGN). The necessity for appropriate standards hasbeen recognized previously by the chemical biology andpharmacometric communities (Oprea et al., 2011; Swatet al., 2015).

Standards are already employed widely across the life sci-ences, frequently building upon ontologies (controlled vo-cabularies of biological/chemical entities and concepts). TheInternational Union of Pure and Applied Chemistry, the In-ternational Union of Biochemistry and Molecular Biology(IUBMB) Joint Commission on Biochemical Nomenclatureand the Nomenclature Committee of IUBMB have providedguidelines on biochemical descriptions and enzyme classifi-cation. A library of ontologies for the life sciences has alsobeen proposed by the Open Biomedical Ontologies Foundry(Smith et al., 2007). Standards and guidelines also exist forreporting biomedical studies, including the minimum infor-mation (MI) standards overseen by the Minimum Informa-tion for Biological and Biomedical Investigations Foundrywho include the Standards for Reporting Enzymology DataCommission (Gardossi et al., 2010). The MI standards of di-rect relevance include the ‘minimum information about abioactive entity’ (Orchard et al., 2011), the ‘minimum infor-mation about a proteomic experiment’ (Taylor et al., 2007)and the ‘minimum information about a molecular interac-tion experiment’ (Orchard et al., 2007).

Problem 3: curation errorsOnline databases can contain errors. We have identified caseswhere the incorrect structures, enzyme targets, species and

parameter values had been recorded. Errors were at low fre-quency, but some would undermine systems pharmacologyapproaches, and these fell into two groups: errors that derivedfrom mistakes in the literature itself, such as from mis-interpretation of data, and errors that derived from the incor-rect transcribing from the literature to the database. The for-mer derive from verbatim acceptance of results frommanuscripts following author error. The latter errors can beintroduced by databases themselves, either from semi-automated triage tools or inadvertent curator mistakes, andthis can be associated with a lack of clarity in the originaldocument. In the present study and for the GtoPdb, wereviewed the primary literature when expanding our datasetsand re-curated existing database coverage.

Solution 3: quality control in curation of results. Using teams ofcurators to validate each other’s work can reduce errors. Thiscan be arranged systematically into error-identifying or error-correcting curation quality control programmes. In an error-identifying programme, each result is independentlycurated twice and where disagreements are identified, thedata is reviewed. Such approaches have been discussedwithin the International Society for Biocuration (Bateman,2010). However, the funding limitations of most publicdatabases preclude this degree of validation. In an error-correcting programme, each result would be independentlycurated three times and where a disagreement is found, theconsensus would be accepted automatically as correct.

Systems pharmacology for the futureA workflow for future studies and hypothesis generation. Withan adequate set of standards and a well-characterizedexperimental system, it should be possible to developintervention hypotheses that can be tested to inform futuretherapy development and to contribute to iterativerefinement of databases. To make this a consistent, highconfidence process, it would be advantageous to work inone experimental system. Such an experimental systemcould be in vivo or in vitro. However, an in vitro model wouldoffer more control and consistency. Such an in vitro systemwould serve as a first approximation to in vivo physiologyand would contribute to determining how in vitroparameters are mapped to in vivo parameters in order tomaximize their value. An advantage of using an in vitrosystem is that it would lend itself to automated hypothesisgeneration and testing and it could be used tosystematically search for new protein–protein anddrug–target interactions. It has been suggested that artificialintelligence methods would be suitable for this purpose inthe laboratory (King et al., 2004). Automation would bothminimize the time required for study and reduce the risk ofmisreporting or mis-curating the results.

Our current systems-level understanding has grown to ascale where manual manipulation is no longer feasible.Standards such as SBML, SBGN and SBGN-ML and reposito-ries such as BioModels have been developed partially to ad-dress this and automated model development allows the fullvalue of databases to be realized (Swainston et al., 2011).Open Pharmacological Concept Triple Store (Williamset al., 2012) is a consortium responsible for a number ofpharmacological and life science databases whose aims



include the improvement of data availability through theuse of data standards, the incorporation of contextual datathrough semantic web standards and the cross-platformlinkage of datasets through an identity mapping service. De-veloping multi-drug hypotheses is a challenge that growsexponentially with the number of drugs and interactionsconsidered. HPC resources are likely to be essential for thisdevelopment.

The following workflow would enable the process to beautomated (see Figure 3).

(I) Pharmacological literature seeds databases of pharma-cological interactions.

(II) Pharmacological and chemical databases containingsufficient information for experimental results to bereproduced accurately. Database Application ProgramInterfaces (APIs) facilitate extraction of results for hy-pothesis generation.

(III) Interaction literature seeds databases of biologicalpathways.

(IV) Pathway databases containing sufficient informationfor experimental results to be reproduced accurately.Database APIs facilitate extraction of results for hypoth-esis generation.

(V) Hypothesis generation for single drug and multi-druginterventions using data obtained through APIs fromthe pharmacological and pathway databases.

(VI) Hypothesis testing. Success yields a candidate therapyand provides validation of the database. Failure initiatesfurther exploration of the underlying interactions thatin turn refine the databases.

(VII) Candidate Intervention. Following success, the groupof compounds enters an optimization pipeline thatreduces them to a minimal set of lead compounds forpreclinical testing to establish their efficacy and safety.

ConclusionThe growth in our understanding of pharmacologicalinteractions and the continuing development of our abilityto computationally model pathway biology will increasinglyenable us to explore drug combinations that target multiplepoints on multiple pathways to reprogram system levelbehaviour. In this way, systems pharmacology may lead tomore effective therapies with fewer side effects. Here, weexplored this approach for the mevalonate arm of thecholesterol biosynthesis pathway, and in doing so, we iden-tify many of the current barriers to progress.

We attempted to build a systems pharmacology model ofthe mevalonate arm of the cholesterol biosynthesis pathway,but gaps and inconsistencies in the data prevented us fromachieving this to a high level of confidence. In particular, we

Figure 3The proposed systems pharmacology workflow.

BJP H Benson et al.


found the lack of comprehensive and systematic parameteriza-tions, experimental variation, ambiguity in structural detailand inappropriate and inaccurate reporting from the primaryliterature to be obstacles. That this should be the case for a path-way of such high biomedical and commercial significance wasunexpected. For this reason, our best current parameterizationrepresents a patchwork of values taken from multiple speciesand experimental configurations. Nonetheless, by completinggaps in our knowledge with representative values, we were ableto demonstrate subtle reprogramming of pathway dynamicsthat may contribute significantly to drug development. Wepropose that these obstacles can be removed through theadoption of standards and quality control.

Although we focused on the mevalonate arm of choles-terol biosynthesis, this approach could be applied to anypathway of interest for which targets and ligands are known.However, before this can happen at a general level, both thecomputational biology and the pharmacology communitiesmust collaborate to remove the current barriers to progress.

Acknowledgements

Initial calculations of optimal multi-drug interventions werecompleted using the supercomputing cluster made availableby the Intelligent Systems Research Centre at the Universityof Ulster. We gratefully acknowledge the teams and fundersthat support the range of external database resources used,without which this work would not have been possible. Anyobservations of error are meant to form part of a constructivediscussion rather than criticism. We are indebted to the lateProf Anthony Harmar (dedication below) for his engagementand enthusiasm during the early phases of this project.

The IUPHAR/BPS Guide to PHARMACOLOGY databaseis funded by the International Union of Basic and ClinicalPharmacology, the British Pharmacological Society andWellcome Trust Biomedical Resources grant 099156/Z/12/Z(H.B., J.S. and C.S.). This work was in part supported by agrant awarded to Professor Tony Bjourson from EuropeanUnion Regional Development Fund (ERDF) EU SustainableCompetitiveness Programme for N. Ireland; NorthernIreland Public Health Agency (HSC R&D); and UlsterUniversity.

This paper is dedicated to the memory of Prof Anthony(Tony) J. Harmar, Emeritus Professor of Pharmacology, Univer-sity of Edinburgh.

Author contributionsThis work was conceived by H.B., S.W., P.G. and A.H. Theanalysis and data compilation was undertaken by H.B., S.W., J.S. and A.P. The manuscript was written by H.B., S.W.,J.S., C.M., A.P., C.S. and P.G.

Conflict of interestH.B., J.S., C.M., C.S. and A.H. have served as curators for theIUPHAR/BPS GtoPdb.

Declaration of transparency andscientific rigourThis Declaration acknowledges that this paper adheres to theprinciples for transparent reporting and scientific rigour ofpreclinical research recommended by funding agencies,publishers and other organisations engaged with supportingresearch.

References

Abaan OD, Polley EC, Davis SR, Zhu YJ, Bilke S, Walker RL et al.(2013). The exomes of the NCI-60 panel: a genomic resource forcancer biology and systems pharmacology. Cancer Res 73:4372–4382.

Aderem A (2005). Systems biology: its practice and challenges. Cell121: 511–513.

Ahn AC, Tewari M, Poon CS, Phillips RS (2006). The clinicalapplications of a systems approach. PLoS Med 3: e209.

Alexander SPH, Fabbro D, Kelly E, Marrion N, Peters JA, Benson HEet al. (2015). The Concise Guide to PHARMACOLOGY 2015/16:Enzymes. Br J Pharmacol 172: 6024–6109.

Balasubramanian N, Brown P, Catt J, Han W, Parker R, Sit S et al.(1989). A potent, tissue-selective, synthetic inhibitor of HMG-CoAreductase. J Med Chem 32: 2038–2041.

Bateman A (2010). Curators of the world unite: the InternationalSociety of Biocuration. Bioinformatics (Oxford, England) 26: 991.

Benson N, Matsuura T, Smirnov S, Demin O, Jones H, Dua P et al.(2013). Systems pharmacology of the nerve growth factor pathway:use of a systems biology model for the identification of key drugtargets using sensitivity analysis and the integration of physiologyand pharmacology. Interface Focus 3, p. 20120071.

Berger S, Iyengar R (2009). Network analyses in systemspharmacology. Bioinformatics (Oxford, England) 25: 2466–2472.

Bhattacharya B, Sweby P, Minihane A, Jackson K, Tindall M (2014). Amathematical model of the sterol regulatory element binding protein2 cholesterol biosynthesis pathway. J Theor Biol 349: 150–162.

BlancM, HsiehWY, Robertson KA,Watterson S, Shui G, Lacaze P et al.(2011). Host defense against viral infection involves interferonmediated down-regulation of sterol biosynthesis. PLoS Biol 9e1000598.

Boralli V, Coelho E, Sampaio S, Marques M, Lanchote V (2009).Enantioselectivity in the pharmacokinetic interaction betweenfluvastatin and lercanidipine in healthy volunteers. J Clin Pharmacol49: 205–211.

Boran ADW, Iyengar R (2010a). Systems approaches topolypharmacology and drug discovery. Curr Opin Drug Discov Devel13: 297–309.

Boran ADW, Iyengar R (2010b). Systems pharmacology. Mount SinaiJournal of Medicine 77: 333–344.

Bowes J, Brown AJ, Hamon J, JarolimekW, Sridhar A, Waldron G et al.(2012). Reducing safety-related drug attrition: the use of in vitropharmacological profiling. Nat Rev Drug Discov 11: 909–922.

Caspi R, Billington R, Ferrer L, Foerster H, Fulcher CA, Keseler IM et al.(2016). The MetaCyc database of metabolic pathways and enzymes



http://onlinelibrary.wiley.com/doi/10.1111/bph.13405/abstract

and the BioCyc collection of pathway/genome databases. NucleicAcids Res 44 (D1): D471–D480.

Chang A, Schomburg I, Placzek S, Jeske L, Ulbrich M, Xiao M et al.(2015). BRENDA in 2015: exciting developments in its 25th year ofexistence. Nucleic Acids Res 43 (Database issue): D439–D446.

Chelliah V, Laibe C, Le Novère N (2013). BioModels Database: arepository of mathematical models of biological processes. Methodsin Molecular Biology (Clifton, NJ) 1021: 189–199.

Croft D, Mundo AF, Haw R, Milacic M, Weiser J, Wu G et al. (2014).The Reactome pathway knowledgebase. Nucleic Acids Res 42(Database issue): D472–D477.

Czauderna T, Klukas C, Schreiber F (2010). Editing, validating andtranslating of SBGN maps. Bioinformatics (Oxford, England) 26:2340–2341.

Darwich AS, Ogungbenro K, Vinks AA, Powell JR, Reny JL, Marsousi Net al. (2017). Why has model-informed precision dosing not yetbecome common clinical reality? Lessons from the past and aroadmap for the future. Clin Pharma & Ther 101: 646–656.

Di Pietro G, Coelho EB, Geleilete TM, Marques MP, Lanchote VL(2006). Chiral evaluation of fluvastatin in human plasma by high-performance liquid chromatography electrospray mass spectrometry.J Chromatogr B Analyt Technol Biomed Life Sci 832: 256–261.

Fabregat A, Sidiropoulos K, Garapati P, Gillespie M, Hausmann K,Haw R et al. (2016). The Reactome pathway Knowledgebase. NucleicAcids Res 44 (D1): D481–D487.

Gardossi L, Poulsen PB, Ballesteros A, Hult K, Svedas VK, Vasić-RackiD et al. (2010). Guidelines for reporting of biocatalytic reactions.Trends Biotechnol 28: 171–180.

Gaulton A, Hersey A, Nowotka M, Bento AP, Chambers J, Mendez Det al. (2016). The ChEMBL database in 2017. Nucleic Acids Res 45(D1): D945–D954.

van der Graaf PH, Benson N (2011). Systems pharmacology: bridgingsystems biology and pharmacokinetics-pharmacodynamics (PKPD)in drug discovery and development. Pharm Res 28: 1460–1464.

van der Greef J, McBurney R (2005). Innovation: rescuing drugdiscovery: in vivo systems pathology and systems pharmacology. NatRev Drug Discov 4: 961–967.

van Hasselt J, van der Graaf P (2015). Towards integrative systemspharmacology models in oncology drug development. DrugDiscovery Today: Technologies 15: 1–8.

Henderson R, O’Kane M, McGilligan V, Watterson S (2016). Thegenetics and screening of familial hypercholesterolaemia. J BiomedSci 23: 39.

Hucka M, Finney A, Sauro HM, Bolouri H, Doyle JC, Kitano H et al.(2003). The systems biology markup language (SBML): a medium forrepresentation and exchange of biochemical network models.Bioinformatics (Oxford, England) 19: 524–531.

Kanehisa M, Goto S, Sato Y, Kawashima M, Furumichi M, Tanabe M(2014). Data, information, knowledge and principle: back tometabolism in KEGG. Nucleic Acids Res 42 (D1): D199–D205.

Kanehisa M, Furumichi M, Tanabe M, Sato Y, Morishima K (2017).KEGG: new perspectives on genomes, pathways, diseases and drugs.Nucleic Acids Res 45 (D1): D353–D361.

Kell DB, Goodacre R (2014). Metabolomics and systemspharmacology: why and how to model the human metabolicnetwork for drug discovery. Drug Discov Today 19: 171–182.

Kim S, Thiessen PA, Bolton EE, Chen J, Fu G, Gindulyte A et al. (2016).PubChem Substance and Compound databases. Nucleic Acids Res 44(D1): D1202–D1213.

King R, Whelan K, Jones F, Reiser P, Bryant C, Muggleton S et al.(2004). Functional genomic hypothesis generation andexperimentation by a robot scientist. Nature 427: 247–252.

Kutmon M, Riutta A, Nunes N, Hanspers K, Willighagen EL, Bohler Aet al. (2016). WikiPathways: capturing the full diversity of pathwayknowledge. Nucleic Acids Res 44 (D1): D488–D494.

Law V, Knox C, Djoumbou Y, Jewison T, Guo AC, Liu Yet al. (2013).DrugBank 4.0: shedding new light on drugmetabolism. Nucleic AcidsRes 42 (D1): D1091–D1097.

Le Novère N, Hucka M, Mi H, Moodie S, Schreiber F, Sorokin A et al.(2009). The systems biology graphical notation. Nat Biotechnol 27:735.

Lehár J, Zimmermann GR, Krueger AS, Molnar RA, Ledell JT, HeilbutAM et al. (2007). Chemical combination effects predict connectivityin biological systems. Mol Syst Biol 3: 80.

Lewington S, Whitlock G, Clarke R, Sherliker P, Emberson J, Halsey Jet al. (2007). Blood cholesterol and vascular mortality by age, sex, andblood pressure: a meta-analysis of individual data from 61prospective studies with 55,000 vascular deaths. Lancet 370:1829–1839.

Lewis JE, Glass L (1991). Steady states, limit cycles and chaos inmodels of complex biological networks. Int J Bifurcation Chaos01: 477.

Lu H, Talbot S, Robertson K, Watterson S, Forster T, Roy D et al.(2015). Rapid proteasomal elimination of 3-hydroxy-3-methylglutaryl-CoA reductase by interferon-γ in primarymacrophages requires endogenous 25-hydroxycholesterol synthesis.Steroids 99 (Pt B): 219–229.

Madrasi K, Burns RN, Hendrix CW, Fossler MJ, Chaturvedula A(2014). Linking the population pharmacokinetics of tenofovir and itsmetabolites with its cellular uptake and metabolism. CPT:Pharmacometrics & Systems Pharmacology 3: 1–6.

Mazein A, Watterson S, Gibbs HC, Hsieh W, Abdullah MB, Blanc Met al. (2011). Regulation and feedback of cholesterol metabolism.Nature Precedings https://doi.org/10.1038/npre.2011.6315.1.

Mazein A, Watterson S, Hsieh W, Griffiths W, Ghazal P (2013). Acomprehensive machine-readable view of the mammaliancholesterol biosynthesis pathway. Biochem Pharmacol 86: 56–66.

Oprea T, May E, Leitão A, Tropsha A (2011). Computational systemschemical biology. Methods in Molecular Biology (Clifton, N.J.) 672:459–488.

Orchard S, Salwinski L, Kerrien S, Montecchi-Palazzi L, Oesterheld M,Stümpflen Vet al. (2007). The minimum information required forreporting a molecular interaction experiment (MIMIx). NatBiotechnol 25: 894–898.

Orchard S, Al-Lazikani B, Bryant S, Clark D, Calder E, Dix I et al.(2011). Minimum information about a bioactive entity (MIABE). NatRev Drug Discov 10: 661–669.

Parton A, McGilligan V, O’Kane M, Baldrick F, Watterson S (2016).Computational modelling of atherosclerosis. Brief Bioinform 17:562–575.

Pawson A, Sharman J, Benson H, Faccenda E, Alexander S, BunemanO et al. (2014). The IUPHAR/BPS guide to PHARMACOLOGY: anexpert-driven knowledgebase of drug targets and their ligands.Nucleic Acids Res 42 (Database issue): D1098–D1106.

BJP H Benson et al.


https://doi.org/10.1038/npre.2011.6315.1

Pence H, Williams A (2010). ChemSpider: an online chemicalinformation resource. J Chem Educ 87: 1123–1124.

Peterson MC, Riggs MM (2015). FDA advisory meeting clinicalpharmacology review utilizes a quantitative systems pharmacology(QSP) model: a watershed moment? CPT: Pharmacometrics &Systems Pharmacology 4: 189–192.

Petrelli A, Giordano S (2008). From single- to multi-target drugs incancer therapy: when aspecificity becomes an advantage. Curr MedChem 15: 422–432.

Pichardo–Almarza C, Metcalf L, Finkelstein A, Diaz-Zuccarini V(2015). Using a systems pharmacology approach to study the effect ofstatins on the early stage of atherosclerosis in humans. CPT:Pharmacometrics & Systems Pharmacology 4 (1): 41–50.

Potter D, Miziorko H (1997). Identification of catalytic residues inhuman mevalonate kinase. J Biol Chem 272: 25449–25454.

Procopiou P, Bailey E, Bamford M, Craven A, Dymock B, Houston Jet al. (1994). The squalestatins: novel inhibitors of squalene synthase.Enzyme inhibitory activities and in vivo evaluation of C1-modifiedanalogues. J Med Chem 37: 3274–3281.

Qin C, Zhang C, Zhu F, Xu F, Chen S, Zhang P et al. (2014).Therapeutic target database update 2014: a resource for targetedtherapeutics. Nucleic Acids Res 42 (Database issue): D1118–D1123.

Robertson K, Hsieh W, Forster T, Blanc M, Lu H, Crick P et al. (2016).An interferon regulated microRNA provides broad cell-intrinsicantiviral immunity through multihit host-directed targeting of thesterol pathway. PLoS Biol 14, p.e1002364.

Rostami-Hodjegan A (2012). Physiologically based pharmacokineticsjoined with in vitro–in vivo extrapolation of ADME: a marriage underthe arch of systems pharmacology. Clinical Pharmacology &Therapeutics 9

edinburgh research explorer · biosynthesis pathway' british journal of pharmacology, vol. 174, no....

Documents