Msc project report

Published on: **Mar 4, 2016**

- 1. Populationpharmacokinetic harmacokineticmodelling of aninvestigational prodrug. drug.Erasmus 2010 – 2011, Pharmacometrics smusresearch group, Uppsala University.Crunenberg Robin, second master ,Pharmaceutical sciences.
- 2. AcknowledgementsI would like to thank the following persons who have all contributed to this work.My supervisor MSc. Sebastian Ueckert for not only his time and quality of management, butalso for his encouragement and guidance these past four months.My project provider MSc. Martin Bergstrand for his availability and advice from thebeginning to the final level of this research project.My university supervisor Professor Vincent Seutin for supporting me in this internshipproject.Professor Mats Karlsson for his cordial welcome into his department.And finally all the members of the pharmacometrics research group for their help, advice andsympathy. 2
- 3. Table of contents1. List of abbreviations ............................................................................................................ 42. Abstract ............................................................................................................................... 53. Introduction ......................................................................................................................... 6 3.1. Overview of the field ....................................................................................................... 6 3.2. Introduction to the prodrug .............................................................................................. 74. Materials and methods ........................................................................................................ 8 4.1. Mathematical and statistical methods .............................................................................. 8 4.1.1. Theory of nonlinear mixed effects models ................................................................ 8 4.1.2. General NMLE model formulation ........................................................................... 8 4.2. Model selection and validation ........................................................................................ 9 4.2.1. Goodness of fit criteria .............................................................................................. 9 4.2.2. Likelihood ratio test ................................................................................................ 10 4.2.3. Visual predictive check ........................................................................................... 10 4.3. Study data....................................................................................................................... 11 4.4. Software and computing tools. ...................................................................................... 135. Results ............................................................................................................................... 14 5.1. First model: active compound ........................................................................................ 14 5.2. Second model: Pro-drug ................................................................................................ 18 5.2.1. Prodrug data results ................................................................................................. 21 5.2.2. Metabolite data results............................................................................................. 24 5.2.3. Active compound data results, prodrug treatment .................................................. 27 5.2.4. Active compound data results, active compound treatment .................................... 306. Discussion ......................................................................................................................... 33 6.1. First model: active compound ........................................................................................ 33 6.2. Second model: Prodrug .................................................................................................. 34 6.2.1. Prodrug data results ................................................................................................. 34 6.2.2. Metabolite data results............................................................................................. 34 6.2.3. Active compound data results, prodrug treatment .................................................. 34 6.2.4. Active compound data results, active compound treatment .................................... 357. Conclusion ......................................................................................................................... 368. References ......................................................................................................................... 379. Appendix ........................................................................................................................... 38 3
- 4. 1. List of abbreviationsAdd AdditiveBQL Below quantification limitBGOF Basic goodness of fitCL ClearanceCV Coefficient of variationCWRES Conditional weighted residualsDV Dependent variableEH Hepatic extractionFM Metabolised fractionFOCE First order conditional estimateIPRED Individual predictionIRES Individual residualsIWRES Individual weighted residualsLLOQ Lower limit of quantificationLRT Likelihood ratio testNLMEM Non-linear Mixed Effects ModellingNONMEM Software of Non-linear Mixed Effects ModellingOFV Objective function valuePD PharmacodynamicPK PharmacokineticPRED Population predictionProp ProportionalPsN Pearl speaks NONMEMQ Inter-compartmental clearanceRuv Residual errorSD Standard deviationSE Standard errorV Volume of distributionVPC Visual predictive check 4
- 5. 2. AbstractThe aim of this study was to develop a population pharmacokinetic model for aninvestigational prodrug with the intention of linking this model to previous works on amodified release form of this prodrug. The data used to create this model were provided froma phase I study. Concentration-time measurements were available for three compounds; theprodrug, the metabolite and the active compound. The model developed is able to describe thepharmacokinetics of the prodrug and its metabolite. Perspectives for future investigations arepresented. 5
- 6. 3. Introduction3.1. Overview of the fieldPharmacometrics is the science of developing and applying mathematical and statisticalmethods to characterize, understand and predict the pharmacokinetic, pharmacodynamic, andbiomarker-outcome behaviour of drugs[1]. The goal of this emerging science is to influencedrug development, regulatory and therapeutic decisions [2].The Journal of Clinical Pharmacology gives a structural representation of this emergingscience in constant evolution. The structure of this evolution focuses on three general inter-connected themes: integration, innovation and impact [2]. “The quantitative integration ofmultisource data and knowledge (a)”, as this field is not solely focused on thepharmacological, statistical, mathematical, engineering or biological concepts, but insteadtakes them all together. This management will lead to the “continuous methodological andtechnological innovation enhancing scientific understanding and knowledge (b)”, which inturn has an “impact on discovery, research, development, approval and utilization of newmedicine (c)”. Fig. 1: Evolution structure of the pharmacometrics.Pharmacometric models describe the relationship between dose, concentrations and time, i.e.the pharmacokinetics (PK) of the drug, and/or the change in the effects due to drug treatmentover time as a function of drug exposure (dose, concentration or other summary measure), i.e.the pharmacodynamics (PD) of the drug.This work concerns the modelling of pharmacokinetic data. Pharmacokinetics describes thedynamics of drug absorption, distribution, metabolism and elimination. PK models aredefined with pharmacologically meaningful coefficients, i.e. clearance, volume of distributionand rate constant. A well characterized PK model can be used to predict, for example, theconcentration variations when altering the doses or the time of dosing. PK model in 6
- 7. association with PD guides recommendations for optimal dosage. A clear dose-concentration-effect relationship should prevent marketing a drug at a dose later recognized to beunnecessarily high [3]. Fig. 2: Overview of the pharmacometric field and its impacts.The aim of this work is to develop a PK model describing the pharmacokinetics of aninvestigational prodrug. In order to link this modelling with previous studies, the model willcontain a hepatic compartment describing the inter-conversion process between the differentcompounds.3.2. Introduction to the prodrugThe investigational prodrug is under development for the treatment/prevention of thrombosis.This project is part of a larger modelling initiative towards describing the populationpharmacokinetics of an investigational drug following the administration of modified releaseformulations [4]. 7
- 8. 4. Materials and methodsThis section is divided in four parts. Firstly a short introduction to the main statistic andmathematics of the population modelling is presented. Secondly the different tools used toevaluate and select the model are discussed. The third part constitutes a presentation of thedata available to produce the model. Finally, the different software packages used to developthe model are listed.4.1. Mathematical and statistical methodsPharmacometric research focuses on population data. Population modelling involvesanalyzing data from all individuals simultaneously instead of data from each individualseparately. To account for the different levels of variability in the population, nonlinear mixedeffects models are used.4.1.1. Theory of nonlinear mixed effects modelsThe nonlinear mixed effects (NMLE) modelling approach involves the simultaneousestimation of the typical and variance parameters using data from all patients, i.e. thepopulation parameters. Many statistical processes can be described by models that incorporatefixed and random effects.The term mixed refers to the combination of these fixed and random effects for thedescription of the data. The fixed effects are those not occurring at random or associated withan entire population e.g. the dose. They are described in the model as fixed effects parametersand they give a model prediction for the typical individual.The random effects are those occurring at random in the population or associated withindividual experimental units. They are described with a distribution function in the modeland these random effects parameters are usually an estimate of the variability [5, 6].Mixed effects models allow the analysis of different levels of variability. For pharmacometricmodels, the two most important levels are inter- and intra-individual variability. Inter-individual or between-subject variability is a result of considering multiple individuals withdifferent physiological parameters. Intra-individual variability is associated with themeasurement error and the limited ability of the model to describe the response; because ofthat it is sometimes called residual variability [7].4.1.2. General NMLE model formulationThe NLME models are used in different areas and can be formulated by many mathematicallyidentical ways. The following pharmaceutical terminologies reflect one of its applications in apharmacological sense.The observed response (e.g. concentration) in an individual within the framework ofpopulation in NLME can be described as [1] = (ф , ) + 8
- 9. where is the observed data (e.g. concentration). This value depends on the system outputthrough a function on the individual parameter ф and on all the experimental/designvariables (e.g. time) but also on the random vector or within-subject error terms,normally distributed with mean 0 and covariance matrix Σ .At a second stage the variability of the individual parameters ф is modelled through afunction by ф = ( , z i, i)where is a fixed effect population parameter vector, zi is a vector of possibly time varyingcovariates, and i is a vector of subject specific random effect parameters. The i are restrictedto be normally distributed with mean 0 and covariance matrix Ω, i.e. ∼ (0,Ω). Examplesof covariates (zi) are body mass, age and sex [7, 8].Given a model function of the form described above and a vector of observed values , themathematical problem is to estimate the fixed effect parameters and the covariance matrixof the random effects Ω and error terms Σ. Different estimation methods have been proposedto solve this. Two of these methods have been investigated in the present work, the first one isthe first order conditional estimate (FOCE) and the second one is the Laplace method.The FOCE method [9] makes the linearization around the current conditional estimate of therandom effect [10].A higher order approximation method is the LAPLACE estimation method [9]. It usessecond-order Taylor series linearization around the current conditional estimate of the randomeffect.In this work the LAPLACE method was used to handle data below the limit of quantification.More specifically the M3 method as described by Ahn [11].All the step wise model building was done firstly with the FOCE and secondly with the M3method to decrease the risk of model misspecification.Both methods belong to the class of maximum likelihood estimators, which allow drawinginference on the parameters of a distribution given a set of observed data. The generalapproach of a maximum likelihood estimator is to find an estimate for a parameter such thatthe likelihood of actually observing the data is maximal [7, 12]. Applied to the specificproblem of a NLME model, the maximum likelihood approach is to maximize the likelihoodfunction H over the set of possible values for , Ω and Σ.4.2. Model selection and validationThe models selection is performed by comparison between new model and previous one.Various tools are used to evaluate if the change made led to an improvement. With theevaluation, the validation is investigated as presented below. In this work the followingdiagnostics were used:4.2.1. Goodness of fit criteriaGraphical evaluations [10], where used to explore the model fits to the data. In this work, fourkinds of plots were used (Fig. 8). 9
- 10. DV versus PRED – Dependent variable versus population predictions:The plot is generated by plotting the population predictions of the model (i.e. withoutconsidering random effects) versus the observed data. The satisfactory model is expected toproduce data points that scatter evenly around the line of identity. Since, random effects areignored in this diagnostic, deviations from the expected appearance are usually due tomisspecifications in the structural model.DV versus IPRED – Dependent variable versus individual predictions:The plot is generated by plotting the individual predictions (i.e with random effects) versusthe observed data. It used to diagnose misspecifications in the random structure of the model.|IWRES| versus IPRED – Absolute value of the individual weighted residual versus theindividual predictions:This plot is used for assessing the stochastic model, in particular the residual error model.Ideally there should be no trend in magnitude of |IWRES|.CWRES versus TIME – conditional weighted residual [13] versus the independent variable:This is a plot is used to diagnose the structural model where ideally the residuals should bescattered evenly around the zero line.4.2.2. Likelihood ratio testThe likelihood ratio test (LRT) [14] can be used to test which one of two competing modelsfits the data best. Usually this involves one full model and one reduced model.The LRT is an approximate test of adding or deleting parts of a model and utilises theminimum objective function value (OFV). The OFV is a goodness-of-fit statistic, calculated byNONMEM (see “4.4. Software and computation tools”). This value is proportional to the likelihoodof the data, when the value decreases the fit to the data is improved. The critical differencebetween the OFVs for a reduced and full model (∆OFV) values for certain significance levels(α) and degrees of freedom (df = number of differing parameters between models) are shownbelow:Table 1: If we take a full model and delete a parameter (df=1), then for a significance level of 0.05, the increment of OFV should be less than 3.84 to keep the reduced model. α = 0.05 α = 0.01 α = 0.001 df = 1 3.84 6.63 10.83 df = 2 5.99 9.21 13.824.2.3. Visual predictive checkThe visual predictive check (VPC) [15, 16] can be used to evaluate how the model candescribe the data used for model development (Fig. 10). The main principle of this method isthe simulation of a high number of data sets from the model, calculation of summary statistics 10
- 11. over all replicates and the comparison of those statistics with the corresponding statistics ofthe original data. In this work, the 95% confidence interval of the simulated median responseis compared to the median of the original data. For a model that adequately describes the data,the observed median is expected to be entirely contained in the confidence interval of thesimulation.[10]4.3. Study dataData from a phase I clinical study was available. This included nine different treatmentapproaches (pro-drug or drug, different routes of administration, doses and formulations), fouranalytes (pro-drug, intermediate, drug and non-release pro-drug in tablet) on a single occasionand covariates (weight, height, sex, age and fed status).The initial modelling will use only a part of the data available, i.e. patients treated with one 10mg i.v. dose of the investigational active compound and twenty concentration-timemeasurements per patient. In total, 200 active compound concentration-time measurements,with 12% data below the limit of quantification (BQL) were available from ten patients.Fig. 3: Blood concentration versus time profiles for the investigational active compound, after 10 mg i.v. dose of the investigational active compound. A different color is used for each individual.To build a prodrug model the data from the active compound treatment and the same patientstreated with 30 mg i.v. dose of the prodrug (forty-eight concentration-time measurements perpatient) were used. For the prodrug treatment, 480 concentration-time measurements wereavailable: 151 prodrug measurements, with 11.25% BQL data, 140 metabolite compoundmeasurements, with 20.71% of BQL data and 189 active compound measurements, with11.64% BQL data. 11
- 12. Fig. 4: Blood concentration versus time profiles for the investigational prodrug, after 30 mg i.v. dose of the investigational prodrug. A different color is used for each idividual. Fig. 5: Blood concentration versus time profiles for the investigational intermediatecompound, after 30 mg i.v. dose of the investigational prodrug. A different color is used for each idividual. 12
- 13. Fig. 6: Blood concentration versus time profiles for the investigational active compound, after 30 mg i.v. dose of the investigational prodrug. A different color is used for each idividual.4.4. Software and computing tools.Data were analysed using the nonlinear mixed effects modelling (NLMEM) software [5],NONMEM (version VII). NLMEMs approach has become increasingly common inpopulation PK/PD analysis.During this work a toolbox for population PK/PD model building, Perl-speaks-NONMEM(PsN) [17, 18] was used with NONMEM. It has a broad functionality ranging from parameterestimate extraction from output files, data file sub setting and resampling, to advancedcomputer-intensive statistical methods and NONMEM job handling in large distributedcomputing systems [19]. The PsN functions used were mainly execute, sumo, update intits,runrecord and vpc.Numerical and graphical diagnostics will be generated using Xpose4 [20]. Xpose 4 is anopen-source population PK/PD model building aid for NONMEM. Xpose tries to make iteasier for a modeler to use diagnostics in an intelligent way, providing a toolkit for datasetcheckout, exploration and visualization, model diagnostics, candidate covariate identificationand model comparison. 13
- 14. 5. ResultsThe different models developed are firstly presented as schematic drawing of the differentcompartments and parameters estimated. Secondly for each model the results of theevaluations tools presented earlier in the materiel and methods section (goodness of fit plotsand predictive check) will be presented as final result.5.1. First model: active compoundThe disposition of drug was described by a three compartment model (Fig. 7) with a centraland two peripheral compartments. To match the constraint of this project a hepaticcompartment where the clearance process occurs is added. The parameters estimated in thismodel are a central (CVC) and two peripheral (CVP) volumes, two inter-compartmentalclearances (CQ), the hepatic extraction ratio (CEH) and a proportional residual error (Prop.ruv).An inter-individual variability is estimated for all the parameters except CQ1, which is fixedto zero. The intra-individual or residual variability is fixed to 1. The event dose andobservation are done in the central compartment and the extraction ratio underwent a logittransformation. The logit transformation is used to allow any value from negative infinity topositive infinity as input, whereas the output is confined to values between 0 and 1. Thiscontainment gives a meaning to the ratio (and the fraction, see above). Fig. 7: Schematic illustration of the PK model of the active compound. The observation and the treatment events are done in the central compartment. The constraint of hepaticcompartment and the relation between central and hepatic is described by physiological value to minimize the impact of this compartment on the model. 14
- 15. Table 2: Estimated parameter values for the model of the active compound model. Parameter Mean Standard Coefficient of Standard Eta shrinkage error (%) variation (%) error (%) (%) CVC 11.22 L 1.44 0.46 168.25 8.54 CVP1 66.30 L 0.15 0.007 219.06 27.51 CVP2 40.29 L 0.61 0.012 228.68 3.42 CEH 0.188 50.07 2.96 120.18 3.95 CQ1 8.56 L/h 3.66 CQ2 36.62 L/h 0.18 0.012 319.04 12.5 Prop. ruv 0.125 177.95Fig. 8: Basic goodness of fit-plots: the red line is a non-parametric smoothing spline of thedata points; the blue dots are the data points; the solid black line in each plot is the line of identity. Data points from the same individual are linked by lines. 15
- 16. Fig. 9: Individual plots for the first four individuals. Each dot represents a data point; the red line is the individual prediction given by the model and the blue line is the population prediction of the model. 16
- 17. Fig. 10: Visual predictive check from 1000 simulated data sets. The 95% confidence intervalof the median of the simulated data is represented by the pink square in the upper plot and theblue area in the plot down below. The real data are the blue dots. The median of the observed data is the red line in the upper plot and the blue line in the plot down below. 17
- 18. 5.2. Second model: Pro-drugAs stated previously, the aim of this work is to develop a model which describes thepharmacokinetics of the prodrug. It means developing a linked model with the prodrug-metabolite-active compound data.The disposition of the pro-drug appears to be adequately described by a six compartmentsmodel (Fig. 11) with the same active compound compartment as described before and the newmetabolite and pro-drug compartments. This includes a central volume (BVC), a metabolitemetabolised fraction (BFM), a proportional (Prop. ruv), and an additive residual error (Add.ruv) for the metabolite. For the prodrug, the new parameters are a central (AVC) andperipheral (AVP) volume, an inter-compartmental clearance (AQ), a hepatic extraction ratio(AEH), a prodrug metabolised fraction (AFM), a proportional (Prop. ruv) and an additiveresidual error (Add. ruv).A slope-intercept (additive and proportional) residual error model is used to describe themetabolite and prodrug data. The active compound model part was adequately described by aproportional residual error. An inter-individual variability is estimated for all the parametersexcepting AQ1, CVP2, CQ1 and CQ2 fixed to zero. The intra-individual or residualvariability is fixed to 1. The event doses are done in the central compartment of the prodrugand active compound (Fig. 11). The observation or concentration-time measurement eventhappens in the central compartment of each compound. The metabolised fraction and theextraction ratio underwent a logit transformation. 18
- 19. Fig. 11: Schematic illustration of the pro-drug PK model. The data from the two treatments were used to develop the model. The constraint of hepatic compartment and the relation between central and hepatic is described by physiological value to minimize the impact of this compartment on the model.
- 20. Table 3: Estimated parameter values of the prodrug model.Parameter Mean Standard Coefficient Standard Eta error of variation error (%) shrinkage (%) (%) (%)AEH, pro-drug 0.172 25.79 2.35 166.95 4.48AVC, pro-drug 5.94 L 1.11 0.07 315.45 3.66AQ1, pro-drug 19 L/h 0.234AVP1, pro-drug 9.52 L 0.388 0.03 635.03 7.75AFM, pro-drug 0.462 12.08 1.01 175.88 8.27BVC, metabolite 7.88 L 0.88 0.05 330 6.53BFM, metabolite 0.282 20.87 1.45 173.93 1.68CVC, drug 16.7 L 0.58 0.03 228.41 18.61CVP1, drug 51.3 L 0.19 0.01 149.31 4.02CVP2, drug 81.7 L 0.16CQ1, drug 13.7 L/h 2.94CQ2, drug 2.33 L/h 2.95CEH, drug 0.162 48.33 3.39 137.51 3.14Addi ruv, pro-drug 5.51 2.39Prop ruv , pro-drug 0.0557 172.35Addi ruv, metabolite 8.92 152.33Prop ruv, metabolite 0.0663 153.32Prop ruv, drug 0.206 38.83
- 21. 5.2.1. Prodrug data results Fig. 12: Basic goodness of fit-plots for prodrug i.v. bolus treatment and prodrug data. For description see Fig. 8. 21
- 22. Fig. 13: Individual plots prodrug i.v. bolus treatment and prodrug data. For description see Fig. 9. 22
- 23. Fig. 14: Visual predictive check for prodrug i.v. bolus treatment and prodrug data. For description see Fig. 10. 23
- 24. 5.2.2. Metabolite data results Fig. 15: Basic goodness of fit-plots for prodrug i.v. bolus treatment and metabolite data. For description see Fig. 8. 24
- 25. Fig. 16: Individual plots for prodrug i.v. bolus treatment and metabolite data. For description see Fig. 9. 25
- 26. Fig. 17: Visual predictive check for prodrug i.v. bolus treatment and metabolite data. For description see Fig. 10. 26
- 27. 5.2.3. Active compound data results, prodrug treatment Fig. 18: Basic goodness of fit-plots for prodrug i.v. bolus treatment and active compound data. For description see Fig. 8. 27
- 28. Fig. 19: Individual plots for prodrug i.v. bolus treatment and active compound data. For description see Fig. 9. 28
- 29. Fig. 20: Visual predictive check for prodrug i.v. bolus treatment and active compound data. For description see Fig. 10. 29
- 30. 5.2.4. Active compound data results, active compound treatment Fig. 21: Basic goodness of fit-plots for active compound i.v. bolus treatment and active compound data. For description see Fig. 8. 30
- 31. Fig. 22: Individual plots for active compound i.v. bolus treatment and active compound data. For description see Fig. 9. 31
- 32. Fig.23: Visual predictive check for active compound i.v. bolus treatment and active compound data. For description see Fig. 10. 32
- 33. 6. DiscussionThe following section is structured according to the results part, with each section of theresults being addressed in a separate section.6.1. First model: active compoundAll four basic goodness of fit plots (Fig. 8) illustrate that the model is describing the dataadequately. There is no major trend in magnitude of |IWRES| and the CWRES vs timeregression line is close to the zero line. The short rise of the curve after fifteen minutes isdriven by tree data points and considered insignificant. For both the DV vs PRED and DV vsIPRED plots, the data points scatter evenly around the line of identity. For higherconcentrations a slightly higher deviation is observed. One possible explanation is perhapsinaccuracies in the application of the infusion that can’t be described by the model. Notably,one of the individuals displays a bigger discrepancy between the individual predicted and theobserved value. However, since the deviations are on both sides of the line of identity (i.e.representing under and over prediction) this is less worrying.The individual plots (Fig. 9) represent a very easy to interpret and natural representation ofthe model predictions and underline the very good performance of the model to describe theobservations.In contrast, VPCs might be harder to interpret, but constitutes a very powerful tool to evaluatea model. In Fig. 10 it can be seen that the median of the observed data is always contained inthe 95% confidence interval for the simulated medians. This further illustrates that the modelis describing the data adequately. In addition, this plot allows diagnosing how BLQ data ispredicted by the model. Since the observed fraction is entirely contained in the predictedinterval, this can also be considered as satisfactory.The parameter estimates (Table 2) for the model have a low standard error for all fixed effectsexcept for the hepatic extraction (CEH). The latter can be explained by its logittransformation. Due to the low number of individuals in the study, the random effectparameters (including the RUV) have a very high standard error. This might limit the usage ofthe model for simulations of large populations. However, the inclusion of each random effectwas tested using the likelihood ratio test and only significant random effects were included inthe final model.All parameter estimates seem physiologically plausible with a low standard error for the fixedeffects. This together with the excellent performance of this model in the graphicaldiagnostics, justifies the use of model 1 to describe the PK of the active compound in the fullmodel. 33
- 34. 6.2. Second model: Prodrug6.2.1. Prodrug data resultsThe BGOF plots (Fig. 12) illustrate that the model is describing the data adequately except forthe CWRES vs time plot. There are irregularities in the trend of the curve but it remainsbetween -1 and 1 and it is not deemed significant. The individual plots (Fig. 13) show thegood performance of the model to describe the observations. For the VPC (Fig. 14), it can beseen that the median of the observed data is always contained in the 95% interval for thesimulated median. However, it also shows a few data points not covered by the prediction atthe high concentration. The prediction (blue line) for the data below limit of quantification(LLOQ) is close to the center of the simulated data (blue area). The analysis of these plotssupports this model to describe the first part of this complex model.6.2.2. Metabolite data resultsThe population prediction (Fig. 15) is centerred on the mean but the data are scattered aroundthis prediction. With the links between the dots that show the individual, we can clearly seesome individuals are not well described. Overall, the individual prediction looks good. Thereis a trend in the |IWRES| vs IPRED but it is a small one and deemed acceptable. The CWRESvs time is problematic but we can go further and check the result of the other diagnostic tools.The individual plots (Fig. 16) show difficulty in matching the highest concentrations but theIPRED looks globally good. The VPC (Fig. 17) shows problems in describing the highconcentrations and few data points at the beginning of the observation. After this analysis wecan see few model misspecifications. We think these are mostly due to the active compoundmisspecification as described below.6.2.3. Active compound data results, prodrug treatmentThe BGOF (Fig. 18) shows a population prediction close to the mean but few individuals aremismatched and the data points are too sparse around the mean. The data points of theindividual prediction are also scattered but the general shape of the curve is satisfactory. The|IWRES| vs IPRED show a clear trend, that could indicate a bad choice of residual errormodel but the slop-intercept or additive models were investigated without success. TheCWRES vs time is not satisfactory and further work will be necessary. The individual plots(Fig. 19) present a good shape but some data points are not matched. In the VPC (Fig. 20) theprediction is not in the center of the simulated data and data points are not matched especiallyat the high concentrations. To do the modelling of this part, data from two different treatmentswere available. The model selected to describe this part of the prodrug model is the onedeveloped for the I.V. bolus active compound treatment data. However, with a simpleparameter comparison we can already expect problems in the model specification. Onepossible explanation of these problems is a saturation process in the degradation of the activecompound in the prodrug treatment. The molecular weight of the different compounds is veryclose and in the prodrug treatment the degradation of the three compounds could use the sameenzyme. This enzyme saturation could diminish the degradation rate of the active compoundand produce the model misspecification. 34
- 35. 6.2.4. Active compound data results, active compound treatmentThe predictions vs observation (Fig. 21) plots show problems in matching the highconcentrations. The |IWRES| vs IPRED and CWRES vs time show like the previous BGOFplot(Fig. 18) a real model misspecification. The individual plots (Fig. 22) present a badprediction at the beginning of the observation. The VPC for this observation is produced onthe log scale to allow the comparison with the first model (Fig. 10). We can see that theprediction is no longer in the center of the simulated data. The first model developed is arelevant model to describe this data. However, due to the bad capability of the model todescribe the active compound of the prodrug, the estimation methods used try to find newparameter estimates. These parameters are a compromise between the two models and thusgive a worse result for each of them. 35
- 36. 7. ConclusionFor the first time a full population pharmacometric model, describing the complete PK of thisinvestigational prodrug, including the active compound and an intermediate has beendeveloped. In general, the model developed in this work describes the observed data to asatisfactory degree. Data from the intra-venous application of the pro-drug and the activecompound are especially well described. The description of the intermediate form and theactive compound after an i.v. dose of the prodrug may need further improvement. Asdiscussed, one possibility would be the inclusion of a saturation phenomenon to theconversion process between the different compounds. 36
- 37. 8. References1. Ette Ene I. , W.P.J., Pharmacometrics : the science of quantitative pharmacology.Hoboken ed. 2007, Hoboken, N.J.: John Wiley & Sons. xix, 1205 p.2. Gobburu, J.V., Pharmacometrics 2020. J Clin Pharmacol, 2010. 50(9 Suppl): p. 151S-157S.3. Machado, S.G., R. Miller, and C. Hu, A regulatory perspective onpharmacokinetic/pharmacodynamic modelling. Stat Methods Med Res, 1999. 8(3): p. 217-45.4. Bergstrand M, S.E., Eriksson U, Weitschies W, Karlsson MO. Semi-mechanisticmodeling of absorption from extended release formulations - linking in vitro to in vivo. inPAGE. 2010. Berlin.5. Beal, S., Sheiner, L.B., Boeckmann, A., & Bauer, R.J., NONMEM Users Guides.(1989-2009). 2009, Icon Development Solutions: Ellicott City, MD, USA.6. Grasela, T. and L. Sheiner, Pharmacostatistical modeling for observational data.Journal of Pharmacokinetics and Pharmacodynamics, 1991. 19(0): p. 25S-36S.7. Ueckert, S., A Population Disease Progression Model for Osteoporosis, inDepartment of Mathematics and Computer Science. 2009, Freie Universität Berlin: Berlin. p.91.8. Foracchia, M., et al., POPED, a software for optimal experiment design in populationkinetics. Comput Methods Programs Biomed, 2004. 74(1): p. 29-46.9. Wang, Y., Derivation of various NONMEM estimation methods. J PharmacokinetPharmacodyn, 2007. 34(5): p. 575-93.10. Pharmacometrics, U.u.g., Introduction to the PHARMACOMETRICS GROUP atUppsala University. Vol. 1. 2010, Uppsala. 143.11. Ahn, J.E., et al., Likelihood based approaches to handling data below thequantification limit using NONMEM VI. J Pharmacokinet Pharmacodyn, 2008. 35(4): p. 401-21.12. Bonate, P.L., Pharmacokinetic-Pharmacodynamic Modeling and Simulation. 2006:Springer US.13. Hooker, A.C., C.E. Staatz, and M.O. Karlsson, Conditional weighted residuals(CWRES): a model diagnostic for the FOCE method. Pharm Res, 2007. 24(12): p. 2187-97.14. Neyman, J. and E.S. Pearson, Joint statistical papers. 1967, Berkeley,: University ofCalifornia Press. 299 p.15. Holford, N., The Visual Predictive Check – Superiority to Standard Diagnostic(Rorschach) Plots, in PAGE. 2005: Pamplona, Spain.16. Karlsson MO , H.N., Tutorial on Visual Predictive Checks. , in PAGE. 2008:Marseille, France.17. Lindbom L, R.J., Jonsson EN. , Perl-speaks-NONMEM (PsN)--a Perl module forNONMEM related programming. Comput Methods Programs Biomed, 2004. 75(2): p. 85-94. 37
- 38. 18. Lindbom L, P.P., Jonsson EN. , PsN-Toolkit--a collection of computer intensivestatistical methods for non-linear mixed effect modeling using NON MEM. Comput MethodsPrograms Biomed., 2005. 79(3): p. 241-57.19. K. Harling, S.U., A. C. Hooker, E. N. Jonsson and M. O. Karlsson, Xpose and Perlspeaks NONMEM (PsN), in PAGE. 2010: Berlin, Germany.20. Jonsson, E.N.K., M.O., Xpose--an S-PLUS based populationpharmacokinetic/pharmacodynamic model building aid for NONMEM. Computer Methodsand Programs in Biomedicine, 1999. 58(1): p. 51-64. 9. AppendixModel file developed for NONMEM. 38