Prediction of intestinal absorption: comparative assessment of gastroplus™ and idea™

European Journal of Pharmaceutical Sciences 17 (2002) 51–61www.elsevier.nl / locate/ejps

P rediction of intestinal absorption: comparative assessment ofGASTROPLUS�and IDEA�

* ´Neil Parrott , Thierry LaveF. Hoffmann-La Roche AG, Pharmaceuticals Division, CH-4070 Basel, Switzerland

Received 17 April 2002; received in revised form 9 July 2002; accepted 17 July 2002

Abstract

We have assessed two commercial software tools employing physiologically based models for prediction of intestinal absorption inhuman.IDEA� 2.0 andGASTROPLUS� 3.1.0 were compared both in their ability to predict fraction absorbed for a set of 28 drugs and interms of the functionality offered. The emphasis was placed on the practical usefulness to pharmaceutical drug discovery. Predictionswere assessed for three levels of input data (i) pure in silico input, (ii) thermodynamic solubility and in silico permeability, (iii)thermodynamic solubility and human colon carcinoma cell line (CACO-2) permeability. We found the pure in silico prediction ability ofthe tools to be comparable with 70% correct classification rate. With measured input data theIDEA� prediction rate improved to 79% whileGASTROPLUS� stayed at 70%. In terms of functionalityGASTROPLUS� is a powerful system for the trained user. Open access to modelparameters, diagnostic tools and the ability to integrate data make it particularly suitable for the later stages of discovery and development.IDEA� is web based and presents a simple interface suitable for widespread use with minimal training. However the limited functionalityand inconvenient handling of multiple compound batches currently restrict the usefulness of version 2.0 for drug discovery. 2002 Elsevier Science B.V. All rights reserved.

Keywords: ADME; Oral absorption; Physiologically based pharmacokinetics; Simulation; Modeling

1 . Introduction challenges. In particular a need for reliable models of oralabsorption exists.

The ability to be administered by the oral route is a The prediction of in vivo absorption is complex and thehighly desirable property for new pharmaceutical drugs number of factors to be considered large. Absorption ofbecause it is the safest, most convenient and economical drugs from the gastrointestinal tract can be influenced bymethod (Goodman et al., 1999). For this reason good oral physicochemical, physiological and formulation factors.availability is a required property for drug candidate The physicochemical factors include pK , solubility,a

molecules in a large percentage of pharmaceutical discov- stability, lipophilicity, and salt forms. The physiologicalery projects. It has been observed that drug development factors include gastrointestinal pH, gastric emptying, smalloften failed for reasons of poor pharmacokinetics (Prentis and large bowel transit times, active transport and efflux,et al., 1988). To avoid the high costs associated with such and gut wall metabolism. The formulation factors arefailures the current practice is to consider metabolism and related to drug particle size, crystal form and dosage formspharmacokinetic properties in parallel with pharmacologi- such as solution, tablet, capsule or suspension.cal tests during the discovery phase. High throughput in Considering the complexity of absorption and the num-vitro technology now allows properties of importance for ber of processes involved, an integrated approach takingoral absorption like solubility, permeability, lipophilicity most of the data into account is highly desirable. Physio-and pK to be measured early and for many compounds. logically based models provide a rational basis for integra-a

Consequently the discovery scientist is presented with tion of data and can predict both extent and rate oflarge volumes of multivariate data and is faced with absorption.considerable data integration and information generation A further step forward is the pure in silico approach

where measured in vitro data is replaced with propertiespredicted from chemical structure alone. Input of in silico*Corresponding author. Tel.:141-61-688-0813.

E-mail address: neil [email protected](N. Parrott). predictions into validated predictive models then gives the]

0928-0987/02/$ – see front matter 2002 Elsevier Science B.V. All rights reserved.PI I : S0928-0987( 02 )00132-X

mailto:[email protected]



´52 N. Parrott, T. Lave / European Journal of Pharmaceutical Sciences 17 (2002) 51–61

possibility for absorption potential to be assessed before compartment the logP and the pK values of the drug. Thea

compounds are synthesized. With access to such tools the exact function for adjusting compartmental permeabilitiesmedicinal chemist is in the position to synthesize only was optimized bySIMULATIONS PLUS to explain the observedcompound collections with a good chance of being well rate and extent of absorption for a proprietary training setabsorbed. of drugs. The logD model and other default settings are

Two physiologically based models to predict oral ab- recommended when sufficient data needed for constructionsorption recently became commercially available. These of drug-specific models are not available (sufficient datamodels areIDEA� (in vitro determination for the estimation might constitute measured regional permeabilities or inof ADME) From Lion Bioscience (LION Bioscience) and vivo concentration versus time data from multiple doses).the ACAT (advanced compartmental absorption and trans- For pure in silico predictions it is necessary to use ait) model available inGASTROPLUS� from SIMULATIONS PLUS companion product,QMPRPLUS�, that takes an input file of

1(Simulations Plus ). multiple structures and computes properties includingIn this study we describe a comparative evaluation solubility, permeability and logP. Models based upon 2D

aimed at assessing the predictiveness of these tools when or 3D structures are available (in this evaluation we usedused with the typical data available to drug discovery only 2D). The model for human permeability is based on ascientists. Our evaluation also discusses usability and combination of in vivo human values and in situ rat wallfunctionality. permeabilities converted to human values based upon a

correlation. The model was constructed using partial leastsquare regression on a training set of 47 drugs (N547,

22 . Methods R 50.76, RMSE50.29). Several solubility models areavailable. We used the model where the only input is 2D

Brief overviews of GASTROPLUS� and IDEA� are pre- chemical structure (an artificial neural network trained on2sented here. For more detailed information we refer the 1204 examples (N51204, R 50.943, RMSE50.47 log

reader to the web sites of the respective companies. units)).GASTROPLUS� 3.1.0 andQMPRPLUS� 2.2 were the versions

2 .1. GASTROPLUS� 3.1.0 evaluated. The latest versionsGASTROPLUS� 3.1.1a andQMPRPLUS� 3.0 were released in February and April 2002

GASTROPLUS� simulates gastrointestinal absorption and and provide enhanced functionality but without majorpharmacokinetics for drugs dosed orally or intravenously changes to the models.in humans and animals. The simulation model underlyingGASTROPLUS� is known as the advanced compartmental 2 .2. IDEA� 2.0absorption and transit model (ACAT) (Agoram et al.,2001). The ACAT model ofSIMULATIONS PLUS is based IDEA� simulates human physiology and accounts forupon an original CAT model described by Yu et al. (1996) regional variations in intestinal permeability, solubility,and is semiphysiological with nine compartments corre- surface area and fluid movement. TheIDEA� absorptionsponding to different segments of the digestive tract module runs as a web application on the corporate Intranet(stomach, seven small intestinal compartments and colon). and is aimed specifically at facilitating lead selection earlyIn addition to human physiology, models for other species in drug discovery.(rat, dog, rabbit or cat) are provided. The model accounts The simulation model underlyingIDEA� is based uponfor controlled-release profiles, pH dependence of dissolu- published work by Grass (1997). Subsequently work wastion and permeability, transport of drug material through done within Lion to develop and train the model with inputthe gastrointestinal tract and absorption of drug material from a consortium of pharmaceutical companies. Consor-through the intestinal wall into the portal vein. tium members supplied oral and intravenous clinical dataGASTROPLUS� provides the ability for customization by while Lion generated a database of in vitro data for aallowing absorption constants to be set individually for training set of around 70 nonmetabolized drugs (Norris eteach intestinal compartment. In addition an optimization al., 2000). Many model details are kept proprietary andmodule permits fitting of model parameters such as only briefly described in theIDEA� reference manual.regional permeabilities, physiological variables and formu- In addition to the physiologically based absorptionlation factors to observed data. For early drug discovery a modelIDEA� 2.0 includes a structure-based model for ingeneric logD model for regional permeability is based silico prediction of absorption class. This model usesupon the premise that as the ionized fraction of drug statistical pattern recognition techniques trained on in vitroincreases permeability decreases. Therefore permeability in and clinical pharmacokinetic data for 121 drug com-each compartment is scaled according to the pH of that pounds. A separate model for prediction of Caco-2 per-

meability was trained on Lion data for 250 drugs.1 Detailed assay protocols for generation of in vitro dataSimulations Plus, I., 1220 W. Avenue J, Lancaster, California 93534-2902, USA.http: / /www.simulations-plus.com/. for use with the absorption model are described in the

http://www.simulations-plus.com/





´N. Parrott, T. Lave / European Journal of Pharmaceutical Sciences 17 (2002) 51–61 53

Table 1IDEA� reference manual. In addition, a set of data for tenClinical fraction absorbed dataexample drugs is provided so that results of companyDrug Dose Fraction absorbedassays may be compared to the data of Lion.

(mg) (%)In a previous evaluation ofIDEA� (Leesman et al., 2000)Roche provided eight drug samples as an external valida-Aciclovir 350 23

Amiloride 10 50tion set for a blinded study and the predictions based uponAntipyrine 600 97the CACO-2 permeability model were within previouslyAtenolol 50 50

agreed acceptance criteria for seven of the eight com- Carbamazepine 200 70pounds. Chloramphenicol 250 90

Version 2.0 ofIDEA� was evaluated. Version 2.2 is due Desipramine 150 100Diazepam 5 100for release in Q4 of 2002 [Lion, personal communication].Diltiazem 90 90Etoposide 350 50Furosemide 80 612 .3. Input data usedGanciclovir 75 3Hydrochlorothiazide 50 69

Three different combinations of input data were assessedKetoprofen 75 92Metoprolol 100 95Naproxen 500 99(i) Pure in silico input for solubility and permeability (i.e.Penicillin V 200 38chemical structure alone as input)Pirenzepine 50 27

(ii) thermodynamic solubility and in silico permeability Piroxicam 20 100and Progesterone 2.5 100

Propranolol 240 99(iii) thermodynamic solubility and measured CACO-2 per-Ranitidine 60 63meability.Saquinavir 600 30Sulpiride 200 44Terbutaline 10 62For a set for 28 drugs, data as above was generated inTheophylline 200 100the standard screening assays used in our drug discoveryVerapamil 120 100projects. Additional common input was the clinical doseWarfarin 5 98

levels corresponding to the fraction absorbed reported inWhere not otherwise indicated fraction absorbed is from Zhao et al.the literature (Table 1).

(2001). Diltiazem, etopside are from Dollery (1998), carbamazepine isThe drugs were selected for diversity in physicochemi-from Goodman et al. (1999), pirenzepine is from Vergin (1986),

cal properties and to cover the full range of fraction saquinavir is from Hoffmann LaRoche (data on file).absorbed in man (Fig. 1). Practical considerations relatedto the availability of compound samples necessitated a bias

the input of solubility values measured at several pHtowards well-absorbed drugs. For the 28 drugs the solu-values, the ability ofGASTROPLUS� to generate solubilitybility at pH 6.5 covered the range from 0.005 to overversus pH profiles from a single measured value and a.1000 mg/ml. The CACO-2 permeability covered a range

26 26 table of pK was not used. Rather solubility values at 5 pHfrom 0.2310 to .60310 cm/s. a

values ranging from 1.5 to 7.5 were used for bothAt this point some explanation of the input requirementssoftwares.of the two software as well as the differences in the modes

Concerning dose,GASTROPLUS� takes a number of inputsof operation is necessary. Table 2 compares the input datadescribing dosage form and formulation whereasIDEA�requirements while Figs. 2 and 3 illustrate the operationaccepts just a single value (Table 2). In this evaluation wemodes for pure in silico prediction and prediction wheretook theGASTROPLUS� defaults namelymeasured data for permeability and solubility are used.

Regarding the pure in silico predictions, one required1. Dosage form: immediate release tabletinput where estimates are not currently available from2. Dose volume: 250 mlQMPRPLUS� is pK . Therefore to complete the in silico dataa

3. Drug particle density: 1.2 g/mlset for GASTROPLUS� we used a separate tool pKalc2 4. Effective particle radius: 25mm(CompuDrug ). ForIDEA� the pure in silico capability,

consisting of a classification into low, medium or highFor predictions based upon measured permeability, the(#33%, 33–66%,>66%), requires only chemical struc-

GASTROPLUS� model has been trained to accept values forture as input.human jejunal permeability as input. Thus a preliminaryTo facilitate a direct comparison it was necessary tostep when using CACO-2 data is transformation basedensure identical input to both tools. SinceIDEA� requiresupon a correlation. In this study we built a correlation oflog human permeability against log CACO-2 permeability

2CompuDrug International, I., P.O.B. 160, Budapest, 1255, Hungary. that included 20 drugs. The correlation coefficient was


Fig. 2. Operation modes for pure in silico prediction.

Fig. 3. Operation modes for prediction when measured data for per-Fig. 1. Distribution of fraction absorbed in human, calculated logP and

meability and solubility are used.measured pK for the 28 drugs studied.a

Table 2Input data forGASTROPLUS� and IDEA�

GASTROPLUS� IDEA�

Chemical ISIS structure SMILES structurestructure (or SMILES) (or ChemDraw structure)

Dose and Initial dose (mg) Dose (mg)formulation Subsequent doses (mg)

Dosing interval (h)Dose volume (ml)Drug particle density (g/ml)Effective particle radius (microns)Dosage form (selection from

a list of options)

Solubility Solubility at different pH values Solubility at different pH valuesin the range 1.5–7.5 in range 1.5–7.5or solubility at one known pHplus a table of pK valuesa

Permeability Permeability measure that is Caco-2 (may need to be transformedtransformed based on a correlation using a correlation to Lion data)to human permeability or in or rabbit intestinal tissuesilico estimate of humanP (four segments) plus permeabilityeff

efflux ratio (optional)pK Table of pK values (used in thea a

log D model of regional permeability)

Lipophilicity log D at known pH or logP (Lipophilicity is predicted internally—no input is possible)


0.88 and the standard error of estimate was 0.28 log units. of these in the test set can lead to a misleadingly lowThe CACO-2 assay and the correlation will be published RMSE.separately (development of a new 7-day, 96-well Caco-2 Therefore all three measures taken together were used topermeability assay combined with HT compound analysis characterize the predictions.[J. Alsenz et al. in preparation].

The IDEA� model has been trained with CACO-2 valuesgenerated in Lions’ assay. It is well known that CACO-2 2 .5. Variabilitydata generated in different laboratories can vary con-siderably (Artursson et al., 1996) and theIDEA� manual In considering the possible relevance of differences intherefore recommends comparison to data from the Lion predictiveness it is obviously important to note the vari-assay for a set of example drugs. In our case, our data were ability in the input data used as well as that of the clinicalnot within 2-fold of Lion data for three out of twelve fraction absorbed data that we are attempting to predict. Incompounds. As this seemed to be a borderline situation we the paper where we obtained the majority of our fractioninvestigated the results using both unchanged CACO-2 absorbed values, Zhao et al. (2001) estimate that their

2data as well as data transformed via the correlation (R 5 RMSE of prediction of 14% is close to the variability in0.77, SE50.5 log units, N 5 12). In fact there was no the observed clinical fraction absorbed. Regarding thesignificant difference in the results obtained and so we calculated properties,SIMULATIONS PLUS gives the RMSE ofreport here the predictions where our CACO-2 data was predicted human jejunal permeability as 0.3 log units andinput unchanged. of predicted solubility as 0.5 log units. For measured

permeability and solubility the experimental error in ourin-house assays is of the order of 0.2–0.3 log units. The

2 .4. Assessment of the predictions models have not been trained with our data and error isobviously introduced when we transform input data via

In addition to graphical plots of observed versus pre- correlation. Our evaluation also ignores the effects of otherdicted fraction absorbed the following numerical measures processes that might be involved for certain drugs such ashave been applied to facilitate assessment of the predic-active transport, efflux, gut metabolism and formulationtions. effects. For the purposes of comparative assessment we

The ability to correctly categorize compounds. have assumed that neglecting such factors will affect theCategories used are: predictions of both models similarly.

We have found that generation of surface plots can be a• High: Fraction absorbed>66% useful aid to understanding the sensitivity of prediction to• Medium: 33%,Fraction absorbed,66% input data. Such plots may be generated by systematically• Low: Fraction absorbed#33% varying the input to the predictive absorption model. In

other words we simulate for a large number of vital drugsThe root mean squared error (RMSE) for predictions of each with slightly different properties. Fig. 4 shows one

fraction absorbed. example of such a plot generated withGASTROPLUS�. Inthis case we held the dose and logP fixed at 100 mg and]]]]]]]]N0, respectively and varied permeability and solubility. Onto2O(observed2 predicted) the plot we have added an indication of the range covered1

]]]]]]]]RMSE5 when the two key input parameters vary by 0.5 logœ Nunits—a variability that we believe is a reasonable esti-mate. As can be seen, given such variability in input, theEach of these measures has weaknesses when usederror associated with fraction absorbed is of the order ofalone. Graphical plots provide an overall impression of20%.prediction quality but do not allow quantitative com-

To gain an impression of the significance of differencesparisons. Categorization reflects well the likely practicalobtained in the RMSE for our test set we carried out someapplication of the predictions in drug discovery projectssimulations. A normal distribution with standard deviationwhere one tends to rank and classify compounds. How-of 14% (an estimate of the variability in observed absorp-ever, classification can be misleading when the predictionstion) was applied to the observed fraction absorbed valuesfall close to the boundary regions e.g. 35% predicted vs.for our set of 28 drugs (with truncation at 0% and 100%).30% observed would be in the wrong class but is actually aWe then ran a series of trials where we took randomgood prediction.samples from these distributions and computed the RMSERMSE is a commonly used measure for such predictionsfor each trial. Based upon 300 trials the RMSE showed aand thus allows comparison with previous publishednormal distribution with mean of 7% and standard devia-results but this too can be misleading when the test sets aretion of 0.75%. Bearing in mind this variability of RMSEdifferent. For example 100% absorbed compounds are notthat arises from the variability in the observed data alonethe most challenging to predict and a very large proportion


Table 3Prediction of absorption class with pure in silico input

Observed IDEA� GASTROPLUS�

absorption predicted predictedclass class class

Aciclovir Low Med MedAmiloride Med Med HighAntipyrine High High HighAtenolol Med Med MedCarbamazepine High High HighChloramphenicol High High HighDesipramine High Med HighDiazepam High High HighDiltiazem High High HighEtoposide Med High MedFurosemide Med Med MedGanciclovir Low Low LowHydrochlorothiazide High High MedKetoprofen High High HighMetoprolol High High HighNaproxen High High HighPenicillin V Med High MedPirenzepine Low High HighPiroxicam High High HighProgesterone High Med MedPropranolol High High HighRanitidine Med High HighSaquinavir Low Low HighSulpiride Med Med HighFig. 4. Surface plot generated withGASTROPLUS� showing the dependencyTerbutaline Med High Highof fraction absorbed on solubility and permeability for a logP of 0 and aTheophylline High High Highdose of 100 mg. The ellipse superimposed onto the plot shows how aVerapamil High High Highvariability of 0.5 log units in the two key input parameters translates toWarfarin High High Highapproximately 20% error in predicted fraction absorbed.

20 correct 19 correctpredictions predictionsfor 28 drugs for 28 drugswe should certainly not assign significance to differences

of ,1.5% in RMSE of the predictions. Fraction absorbed classes are defined as: low#33%, medium 33–66%,high $66%.

3 . Results and discussion to conclude that our results must be considered as optimis-tic for the predictions likely to be achieved for novel

3 .1. Prediction based upon chemical structure alone structures.Overall the pure in silico performance of the two tools is

Table 3 compares the results after conversion of the similar at around 70% correct classification and it isGASTROPLUS� prediction of fraction absorbed to a classifi- interesting to compare this with other in silico approaches.cation identical to that used byIDEA�. For example the ‘Pfizer rules’ or ‘rule of 5’ (Lipinski et

For IDEA� 20 out of 28 drugs (71%) are assigned to the al., 1997) is often used as an alert for compounds likely tocorrect absorption class. Two of the Low class compounds be poorly absorbed. In our implementation, the ‘Pfizerare correctly predicted (ganciclovir, saquinavir) while two rules’ correctly provide an alert for saquinavir but not forare missed (aciclovir, pirenzepine) and in the case of the three other low absorption drugs and a false alert ispirenzepine by more than one class. triggered for etoposide, a medium absorption drug. Using a

For GASTROPLUS� 19 out of 28 drugs (68%) are assigned more sophisticated approach (Wessel et al., 1998) trained ato the correct absorption class. In addition to the two low six-descriptor neural network model with a data set of 76category drugs missed byIDEA� saquinavir (fabs%530) is compounds to predict for a test set of 10 drugs andalso misclassified byGASTROPLUS�. achieved a RMSE of 16%. Zhao et al. (2001) used

In considering the performance of the pure in silico multiple regression based upon Abraham structural de-approach it must be noted that the set of 28 drugs used are scriptors and achieved an RMSE of 14% when predictingwell studied and might belong to the training sets that have for a test set of 131 drugs. The RMSE obtained frombeen used for the development of the models. Neither GASTROPLUS� was 24%, however, direct comparison ofcompany discloses their internal training sets and so we RMSE is not valid because the data sets are different.can only speculate on this. However, we feel that it is fair Table 4 enables comparison of the drugs from our data set


Table 4Summary of pure in silico predictions achieved by Zhao and Wessel

Obs. Zhao Correct Wessel CorrectAbs. Pred. class? Pred. class?

Aciclovir 23 63 NAmiloride 50 54 YAntipyrine 97 95 Y 92 YAtenolol 50 79 N 80 NCarbamazepine 70 – – – –Chloramphenicol 90 79 Y 96 YDesipramine 100 106 YDiazepam 100 104 YDiltiazem 90 – – – –Etoposide 50 68 N 52 YFurosemide 61 69 N 89 NGanciclovir 3 55 N 6 YHydrochlorothiazide 69 55 Y 62 YKetoprofen 92 95 Y 100 YMetoprolol 95 94 Y 91 YNaproxen 99 94 Y 100 YPenicillin V 59 76 NPirenzepine 27 – – – –Piroxicam 100 78 YProgesterone 100 111 Y 94 YPropranolol 99 98 Y 96 YRanitidine 64 78 N 77 NSaquinavir 30 – – – –Sulpiride 44 72 NTerbutaline 62 58 YTheophylline 100 76 YVerapamil 100 108 YWarfarin 98 94 Y 100 Y

Classification 16/24 11/14rate (66%) (78%)RMSE 19% 12%

in common with the Zhao and Wessel data sets in terms of that is consistently well predicted with aciclovir, piren-the rate of correct classification and the RMSE. zepine and saquinavir being over predicted by bothIDEA�

and GASTROPLUS�.3 .2. Prediction based upon measured solubility and Based on pure in silico as well as on measured solu-predicted permeability bilities there was no significant difference in performance

of GASTROPLUS� and IDEA�. However,IDEA� did show anFig. 5 shows the results obtained withIDEA�. There is improvement in predictive accuracy when measured

an apparent tendency for underprediction of the high CACO-2 permeability was used as input. This improve-absorption compounds while for medium category drugs ment was apparent when measured by either the correctthe fraction absorbed was overpredicted. classification rate or the RMSE. But even here only one of

Fig. 6 shows the results obtained withGASTROPLUS�. The the four poorly absorbed compounds was correctly iden-predictions show a different pattern to those fromIDEA� tified and pirenzepine, a low absorption compound wouldbut overall the accuracy of the predictions as measured by have been misclassified as high absorption. Interestingly,the rate of correct categorization or the RMSE is the same GASTROPLUS� did not show a comparable improvement infor IDEA� andGASTROPLUS� being 19 out of 28 and 22% in predictiveness. Thus independently of the permeabilityboth cases. used (in silico or experimental), the number of correct

classifications was 19. Although two of the poorly ab-3 .3. Prediction based upon measured solubility and sorbed compounds were correctly recognized there was ameasured CACO-2 permeability tendency to underpredict absorption for medium range

compounds (e.g. hydrochlorothiazide, theophylline). ForFig. 7 shows the results obtained withIDEA� and Fig. 8 early drug discovery such a tendency for false negatives

those withGASTROPLUS�. would be more problematic than overpredictions (falseTable 5 summarizes all predictions. Of the four poorly positives) because it could lead to rejection of potentially

absorbed drugs in our data set, ganciclovir is the only drug interesting compounds before in vivo tests are performed.


Fig. 6. GASTROPLUS� predicted fraction absorbed based on measuredFig. 5. IDEA� predicted fraction absorbed based on measured solubility solubility and predicted human jejunal permeability. The line of unity isand predicted CACO-2 permeability. The line of unity is shown on the shown on the graph. Categorization of fraction absorbed into high,graph. Categorization of fraction absorbed into high, medium or low medium or low absorption is correct for 19 out of 28 drugs. The rootabsorption is correct for 19 out of 28 drugs. The root mean squared error mean squared error of prediction is 22%.of prediction is 22%.

On the other hand, false positives, although clearly unde- For hydrochlorothiazide the fraction absorbed predictedsirable, would probably be found out at the stage of the in withGASTROPLUS� based on CACO-2 data and measuredvivo assay. Part of the reason for the poorer performance solubility (24%) was significantly lower than the 69%of GASTROPLUS� might be related to the need to transform observed in vivo. Hydrochlorothiazide is consistentlyin vitro permeability data to in vivo human permeability. under predicted byGASTROPLUS� based upon both in silicoThe human jejunal permeability data is associated with and in vitro permeability. In fact we found that even whenlarge variability and this certainly introduces considerable using the measured human jejunal permeability this com-additional error into the predictions. Conversely,IDEA� is pound is considerably under predicted. It seems thatcalibrated directly for CACO-2 data and so the transforma- hydrochlorothiazide could not be well modeled bySIMULA-

tion step is avoided. TIONS PLUS and both hydrochlorothiazide and furosemideIndependently of the approach and the software used, were left out of theQMPRPLUS� models of human per-

several drugs were consistently badly predicted with an meability most likely because the low in vivo permeabilityerror of more than 20%. Some comments and possible seems inconsistent with the observed intestinal absorptionexplanations for these incorrect predictions are considered and bioavailability in man. Amiloride is also under pre-here. dicted byGASTROPLUS� when using in vitro permeability

A number of these compounds like pirenzepine, but is better predicted byIDEA�. In common with hydro-ranitidine and sulpiride are absorbed passively through the chlorothiazide this drug lies in the region of lower per-paracellular route. Pirenzepine is overpredicted by both meability where theIDEA� model seems to be betterGASTROPLUS� and IDEA� based upon both in silico and in predictive.vitro permeability. Furosemide, ranitidine and sulpiride are Saquinavir is overpredicted byGASTROPLUS� based upondrugs with low CACO-2 permeability which are better both in silico and in vitro permeability. Interestingly,predicted byIDEA� than GASTROPLUS�. IDEA� predicts the fraction absorbed for this drug better


Fig. 7. IDEA� predicted fraction absorbed for measured solubility andmeasured CACO-2 permeability. The line of unity is shown on the graph. Fig. 8.GASTROPLUS� predicted fraction absorbed for measured solubilityCategorization of fraction absorbed into high, medium or low absorption and measured CACO-2 permeability. The line of unity is shown on theis correct for 22 out of 28 drugs. The root mean squared error of graph. Categorization of fraction absorbed into high, medium or lowprediction is 19%. absorption is correct for 19 out of 28 drugs. The root mean squared error

of prediction is 24%.

when using the in silico prediction of CACO-2 permeabili-ty that is 60 times lower then the value measured in our Although bothGASTROPLUS� and IDEA� aim to predictassay. Based upon the measured permeabilityIDEA� also oral absorption in man there are considerable differencesover predicts. A likely explanation for this over prediction between them in usability and functionality.IDEA� pre-is that saquinavir is known to be rapidly metabolized by sents a simple interface as a web page allowing verycytochrome P450 3A4 [Roche in house data]. CYP3A4 is limited access to model parameters. Access is via anexpressed at high levels in the mucosa of the gastrointesti- Internet browser and so corporate-wide deployment isnal tract and it is likely that a part of the dose is simple and because of the simplicity of the interfacemetabolized before being absorbed into the portal vein, a training requirements are minimal. In this respectIDEA�process that was not accounted for by the models as they could be aimed at use by non-DMPK specialists in drugwere used in the present evaluation. discovery (e.g. medicinal chemists) who wish to assess the

Progesterone is under predicted byGASTROPLUS� based absorption potential of compounds with a quick and easyupon in silico and in vitro permeability. It is also under tool. HoweverIDEA� version 2.0 falls short in this respectpredicted byIDEA� but to a lesser extent. This drug shows because the capabilities for import and export are poor andlow solubility (0.005 mg/ml) and it seems that the the guidance provided is too limited. The batch importdiscrepancy between the high observed absorption (100%) capability is cumbersome and handles onlySMILES

3and the lower prediction is related to the dissolution rate (Daylight ) and not multiple structures inMDL format from4and formulation. Thus, it is possible to achieve the correct ISIS (MDL ) the desktop tool most familiar to our chemists.

fraction absorbed inGASTROPLUS� by reducing the particle 3Daylight (www.daylight.com), Suite 360, 27401 Los Altos, Missionsize from 25 to 5mm. This is a factor not taken into Viejo, CA, USA.

4account in our current evaluation where a default of 25mm MDL, M.I. S., 14600 Catalina Street, San Leandro, CA 94577, USA.was used for all drugs. www.mdli.com.

www.daylight.com

www.mdli.com


Table 5Summary of predictions

Correct RMSE Drugs overpredicted Drugs underpredictedcategorization (%) by.20% by .20%

GASTROPLUS� pure in silico 19 24

IDEA� pure in silico 20

GASTROPLUS� predicted permeability 19 22 ac, am, pe, hy, th1measured solubility pi, sa, su

IDEA� predicted permeability 19 22 ac, et, fu,1measured solubility pe, pi, su

IDEA� measured permeability 22 19 ac, pe, pi, sa1measured solubility

GASTROPLUS� measured permeability 19 24 pi, sa am, fu, hy,1measured solubility pg, ra, su

ac5aciclovir, am5amiloride, et5etoposide, fu5furosemide, hy5hydrochlorothiazide, pe5penicillin v, pi5pirenzepine, pg5progesterone, ra5ranitidine, sa5saquinavir, su5sulpiride, te5terbutaline, th5theophylline.

Also there is no capability for batch export of results and In closing we would like to express our belief thatthis would be essential in the early stages when numerous physiologically based models do have great potential forcompounds are being assessed. In this respect we feel that use within both early and later stages of drug discoveryversion 2.0 ofIDEA� needs some enhancements before it [see the recent review provided by Grass and Sinkowould be acceptable for widespread use in drug discovery. (2002)]. Typically in early drug discovery, in silico and inIn addition, to better understand the sensitivity of absorp- vitro data become available for numerous drug candidatestion to changes in key parameters could be very useful and and the discovery scientist requires a tool to integrate thesethis type of information is currently not easily obtained. parameters and weight their overall impact in vivo. Physio-Provision of surface plots (Fig. 4) would be one suggestion logically based approaches provide this capability and canthat could help in this respect. Version 2.2 ofIDEA� is due go beyond the prediction of a single parameter like fractionto be released in Q4 of 2002 and we are assured that many absorbed to predict absorption versus time profiles andof the interface deficiencies will be addressed [Lion, even plasma concentration levels if an estimate of disposi-personal communication]. tion kinetics is available. We foresee the prospect for

GASTROPLUS� is a native NT application presenting a improved prediction by means of optimization on companysophisticated user interface. Very many aspects of the data to allow direct input without the need to transform viamodel are available to be changed and although the basic correlation. In addition, within drug discovery projects,input required is not too much different toIDEA� there is models could be optimized for use with a specific com-the possibility to specify many additional data. Server pound class.based deployment within a single site is easily possible, As the discovery process moves on, open and flexiblehowever, because of the complexity of the application, physiologically based models provide a framework forthorough training is required. This complexity arises integration of further complexities when data related tobecauseGASTROPLUS� provides broad functionality includ- these processes becomes available (e.g. active transporting pharmacokinetic model fitting, optimization for drug- and efflux). At this stage an additional benefit is the abilityspecific models, sensitivity analysis, and variability estima- to estimate the impact of formulation, physiological vari-tion via stochastic mode operation. For pure in silico ability and food effects and to provide the ability toprediction one uses theQMPRPLUS� module and the combi- extrapolate across species and doses.nation of QMPRPLUS� and GASTROPLUS� provides the capa-bility for batch analysis of large numbers of compounds ina convenient and efficient way. AlsoGASTROPLUS� showsstrengths in its ability to integrate additional data (dosage A cknowledgementsform and formulation, transport /efflux processes, intestinalmetabolism). Such data often becomes available later in We gratefully acknowledge the work of Roche col-the drug discovery process and we feel that at this stage leagues Manfred Kansy and Jochem Alsenz who providedGASTROPLUS� would be the tool of choice. In our opinion thermodynamic solubility and CACO-2 permeability dataGASTROPLUS� is a tool for specialist users who have been and gave advice on the evaluation. We also thank Liontrained and who make use of the software relatively Bioscience and Simulations Plus for reviewing a draft offrequently. the manuscript and providing helpful comments.


Lipinski, C., Lombardo, F., Dominy, B., Feeney, P., 1997. ExperimentalR eferencesand computational approaches to estimate solubility and permeabilityin drug discovery and development settings. Adv. Drug Deliv. Rev.

Agoram, B., Woltosz, W.S., Bolger, M.B., 2001. Predicting the impact of 23, 3–25.physiological and biochemical processes on oral drug bioavailability. Norris, D.A., Leesman, G.D., Sinko, P.J., Grass, G.M., 2000. Develop-Adv. Drug Deliv. Rev. 50 (Suppl. 1), S41–S67. ment of predictive pharmacokinetic simulation models for drug

Artursson, P., Palm, K., Luthman, K., 1996. Caco-2 monolayers in discovery. J. Controlled Release 65, 55–62.experimental and theoretical predictions of drug transport. Adv. Drug Prentis, R.A., Lis, Y., Walker, S.R., 1988. Pharmaceutical innovation byDeliv. Rev. 22, 67–84. the seven UK-owned pharmaceutical companies (1964–1985). Br. J.

Dollery, C., 1998. Therapeutic Drugs, 2nd Edition. Churchill Livingstone. Clin. Pharmacol. 25, 387–396.Goodman, L.S.O.G., A (Orig); Hardman, J.G. Limbird, L.E. Molinoft, Vergin, H.M.H., Strobel, K., Nitsche, V., 1986. Pharmacokinetics and

P.B. Ruddon, R.W., 1999. Goodman and Gilman’s The Pharmaco- bioequivalence of different formulations of pirenzepine Arzneimittel-logical Basis of Therapeutics. Forschung 36 (9), 1409–1412.

Grass, G.M., 1997. Simulation models to predict oral drug absorption Wessel, M.D., Jurs, P.C., Tolan, J.W., Muskal, S.M., 1998. Prediction offrom in vitro data. Adv. Drug Deliv. Rev. 23, 190–219. human intestinal absorption of drug compounds from molecular

Grass, G.M., Sinko, P.J., 2002. Physiologically-based pharmacokinetic structure. J. Chem. Inf. Comp. Sci. 38, 726–735.simulation modelling. Adv. Drug Deliv. Rev. 54 (3), 433–451. Yu, L.X., Crison, J.R., Amidon, G.L., 1996. Compartmental transit and

Leesman, G., Norris, D., Timony, G., Chen, Y., Pirson, W., Theil, F.P., dispersion model analysis of small intestinal transit flow in humans.Schmitt-Hoffman, A.H., Lee, Y.H., Doerr-Stevens, J., Gordon, P., Int. J. Pharmaceut. 140, 111–118.Retajczyk, R., Tran, S., Lee, K.-J., Johnson, N., Castelo, J., Christ- Zhao, Y.H., Le, J., Abraham, M.H., Hersey, A., Eddershaw, P.J., Lus-opher, R., Grass, G., Sinko, P., 2000. Assessment of the physiologi- combe, C.N., Boutina, D., Beck, G., Sherborne, B., Cooper, I., 2001.cally-basedIDEA� predictive model using an external (blinded) data Evaluation of human intestinal absorption data and subsequent deriva-set. In: Poster presented at the AAPS 2000. tion of a quantitative structure activity relationship (QSAR) with the

LION Bioscience, W.S., 69123 Heidelberg, Germany.http: / /www.lion- Abraham descriptors. J. Pharm. Sci. 90 (6), 749–784.bioscience.com/.

http://www.lionbioscience.com/






Prediction of intestinal absorption: comparative assessment of gastroplus™ and idea™

Documents

Transcript of Prediction of intestinal absorption: comparative assessment of gastroplus™ and idea™