Provisional Classification and in Silico Study of Biopharmaceutical

Apr 26, 2013 - Comparison of concordances between BCS permeability classifications using calculated physicochemical properties (partition coefficient,...
6 downloads 10 Views 2MB Size
Download PDF
Article pubs.acs.org/molecularpharmaceutics

Provisional Classification and in Silico Study of Biopharmaceutical System Based on Caco‑2 Cell Permeability and Dose Number Hai Pham-The,† Teresa Garrigues,‡ Marival Bermejo,§ Isabel González-Á lvarez,§ Maikel Cruz Monteagudo,∥,⊥,# and Miguel Á ngel Cabrera-Pérez*,†,‡,§ †

Molecular Simulation & Drug Design Group, Centre of Chemical Bioactive, Central University of Las Villas, Santa Clara 54830, Villa Clara, Cuba ‡ Department of Pharmacy and Pharmaceutical Technology, University of Valencia, Burjassot 46100, Valencia, Spain § Department of Engineering, Area of Pharmacy and Pharmaceutical Technology, Miguel Hernández University, 03550 Sant Joan d’Alacant, Alicante, Spain ∥ CIQ, Department of Chemistry and Biochemistry, Faculty of Sciences, University of Porto, 4169-007 Porto, Portugal ⊥ REQUIMTE, Department of Chemistry and Biochemistry, Faculty of Sciences, University of Porto, 4169-007 Porto, Portugal # Applied Chemistry Research Center (CEQA), Faculty of Chemistry and Pharmacy, Central University of Las Villas, Santa Clara, 54830, Cuba S Supporting Information *

ABSTRACT: Today, early characterization of drug properties by the Biopharmaceutics Classification System (BCS) has attracted significant attention in pharmaceutical discovery and development. In this direction, the present report provides a systematic study of the development of a BCS-based provisional classification (PBC) for a set of 322 oral drugs. This classification, based on the revised aqueous solubility and the apparent permeability across Caco-2 cell monolayers, displays a high correlation (overall 76%) with the provisional BCS classification published by World Health Organization (WHO). Current database contains 91 (28.3%) PBC class I drugs, 76 (23.6%) class II drugs, 97 (31.1%) class III drugs, and 58 (18.0%) class IV drugs. Other approaches for provisional classification of drugs have been surveyed. The use of a calculated polar surface area with a labetalol value as a high permeable cutoff limit and aqueous solubility higher than 0.1 mg/mL could be used as alternative criteria for provisionally classifying BCS permeability and solubility in early drug discovery. To develop QSPR models that allow screening PBC and BCS classes of new molecular entities (NMEs), 18 statistical linear and nonlinear models have been constructed based on 803 0-2D Dragon and 126 Volsurf+ molecular descriptors to classify the PBC solubility and permeability. The voting consensus model of solubility (VoteS) showed a high accuracy of 88.7% in training and 92.3% in the test set. Likewise, for the permeability model (VoteP), accuracy was 85.3% in training and 96.9% in the test set. A combination of VoteS and VoteP appropriately predicts the PBC class of drugs (overall 73% with class I precision of 77.2%). This consensus system predicts an external set of 57 WHO BCS classified drugs with 87.5% of accuracy. Interestingly, computational assignments of the PBC class reasonably correspond to the Biopharmaceutics Drug Disposition Classification System (BDDCS) allocations of drugs (accuracy of 63.3−69.8%). A screening assay has been simulated using a large data set of compounds in different drug development phases (1, 2, 3, and launched) and NMEs. Distributions of PBC forecasts illustrate the current status in drug discovery and development. It is anticipated that a combination of the QSPR approach and well-validated in vitro experimentations could offer the best estimation of BCS for NMEs in the early stages of drug discovery. KEYWORDS: Biopharmaceutics Classification System (BCS), Biopharmaceutics Drug Disposition Classification System (BDDCS), Provisional Biopharmaceutical Classification (PBC), dose number, Caco-2 cell permeability, Quantitative Structure Activity/Property Relationship (QSAR/QSPR)

■

INTRODUCTION

solid oral dosage forms are categorized as having either rapid or slow in vitro dissolution and then classified based on aqueous

After almost 20 years of the introduction and exploration of the Biopharmaceutics Classification System (BCS), it has gained a major impact on the regulation and development of immediate release (IR) solid oral drug products.1,2 Based on the principal factors that determine the rate and extent of drug absorption, the BCS provides a scientific framework for classifying drug substances into one of four categories. According to BCS, IR © XXXX American Chemical Society

solubility and intestinal permeability of the active pharmaceutical Received: February 3, 2013 Revised: March 30, 2013 Accepted: April 26, 2013

A

dx.doi.org/10.1021/mp4000585 | Mol. Pharmaceutics XXXX, XXX, XXX−XXX

Molecular Pharmaceutics

Article

ingredient (API).1 This system has been formally adopted by the US Food and Drug Administration (FDA),3 the European Medicines Agency (EMA),4 and the World Health Organization (WHO)5 as a technical standard for waiving bioequivalence (BE) test requirements for oral drugs. A recent study of the economic impact of granting biowaivers for class I and III BCS demonstrated an impressive saving annual expenditure on running BE studies, being more than 120 million dollars between the two classes.6 Because it avoids unnecessary drug exposures to healthy subjects, while maintaining the high public health standard for therapeutic equivalence, the BCS is, without doubt, a potential tool for speeding up and reducing the cost of drug development. Initially, BCS-based in vivo BE study waivers were granted for Scale-Up and Post-Approval Changes (SUPAC).7 Current biowaiver approval is also applied for generic IR oral drug products of BCS class I API. Therefore, pharmaceutical scientists in the early drug development always attempt to utilize pharmaceutical profiling data to establish the preliminary BCS classification for the lead compound.8 The most popular form of such classification is based on the secondary aqueous solubility references and permeability estimations (e.g., clogP).9 The so-called provisional BCS classification gives a round prediction of solubility and permeability class membership that can be revised as more experimental data become available.10 A precise BCS classification based on more extensive solubility, dissolution, and permeability experiments to officially support a biowaiver application is further needed. Several works have been carried out to provisionally classify large groups of drugs.9,11−15 In these studies, aqueous solubility criterion was obtained from commonly available references such as US Pharmacopeia16 and Merck Index,17 considering the US FDA3 or the WHO guidance.5 Following these rules, a drug substance can be seen as a highly soluble drug when the highest dose strength is soluble in 250 mL or less of aqueous media over the pH range of 1 to 7.5 (US FDA) or 1.2 to 6.8 (WHO) at 37 °C. However, in most of the cases the authors could neither fill out multi-pH profiles nor clarify the appropriate method of solubility determination.14 In addition, reported maximum dose strengths vary depending on the references. On the other hand, the permeability classification is based on diverse estimations, which remain somewhat uncertain. First, as the principal purpose of regulatory authorities when utilizing BCS is to support BE study, no benchmarking exists for the low-permeability term which would define BCS class III or IV. Up to now, it is commonly accepted that, if a drug does not meet the high permeability criterion, it can be seen as a low permeability drug. From the FDA and WHO guidance, the permeability class boundary can be determined directly by measuring the rate of mass transfer across human intestinal membrane. It can also be determined indirectly based on the extent of absorption (specifically, fraction of dose absorbed, Fa) of a drug substance in humans.2,3,5 However, the definitions of highly permeable drugs between guidances are different. According to FDA, a drug substance is considered highly permeable when Fa is 90% or more of the oral dose. This absorption-based permeability criterion was relaxed to 85% in the WHO acceptance.5 This shifting causes some changes in the allocation of class I and III drugs.5 There are five main approaches to demonstrate high permeability in BCS: (i) absolute bioavailability or mass balance studies in humans, (ii) urinary recovery of unchanged drug in humans, (iii) in vivo intestinal perfusion studies in humans, (iv) in vitro permeation studies across a monolayer of cultured

epithelial cells, and/or (v) high metabolism as defined under BDDCS (Biopharmaceutical Drug Disposition Classification System).2,18 Among these methods, in vitro assays are widely used as high throughput screening (HTS) methods for measuring permeability and appropriately estimate the oral absorption of drug-like molecules in early preclinical phase. It should be noted that the Caco-2 (adenocarcinoma cells derived from colon) monolayers are recommended by both the US FDA and EMA3,4 as the most suitable model for estimating intestinal permeability. This cell culture exhibits morphological and functional similarities to human enterocytes that makes it a better surrogate for in vivo drug absorption potential than other epithelial cell cultures.19 The expression of multiple transporters in Caco-2 cells, such as P-glycoprotein (P-gp), the protoncoupled oligopeptide transporter (PEPT1) among others,20 offers great advantages over other simplified models in investigating the interplay among different transport systems and to establish the relative contributions of passive and active transport mechanisms to the overall permeability.19 In this report, we attempt to provide a Provisional Biopharmaceutical Classification (PBC) for a group of drug and drug-like uncharged molecules, using consensus-revised references for BCS solubility and Caco-2 cell permeability measurements for identifying high permeability class in the context of BCS. This classification was compared to current available allocations according to BCS and BDDCS. Based on computed physicochemical properties of this data set some computational models were developed to assign a provisional biopharmaceutical class for new molecular entities (NMEs). These models were rigorously validated on various published BCS class drug sets,5,9,11,12,18 and the feasibility of performing PBC prediction in early drug discovery is discussed.

■

MATERIALS AND METHODS Data Set. BCS based-provisional classification requires both solubility and permeability measurements. In this work, a set of 322 drugs was obtained from published works. A provisional classification was executed by means of an extensive literature revision of experimental values and assigned classes, as follows. Solubility Data. The drug solubility data (in mg/mL) can be obtained from standard references,9 such as the Pharmacopeia16 or the Merck Index.17 Since the FDA criteria for solubility specifies water as a solvent, simulated intestinal fluid containing a surfactant cannot be used.21 Furthermore, the cutoff value between high and low solubility requires knowledge of the lowest solubility over the wide pH range (as stated in the FDA or WHO guidance). The temperature condition of the assay is also specified to be 37 °C. However, all of these requirements are seldom fulfilled. Hence, we propose the use of other reliable sources apart from current standard references. There are many free databases which provide a vast number of physicochemical data, including aqueous solubility, in user-friendly web servers. For instance, it can be mentioned the Physprop database with more than 25 000 compounds (available at http://www.srcinc.com/what-we-do/databaseforms.aspx?id=386) or ChemIDplus, a database of about 350 000 chemical compounds (available at http://chem.sis.nlm.nih.gov/chemidplus/ chemidheavy.jsp), and so on. As the number of experimental data increases, the final consensus choice is more reliable. A valuable reference is a recent application of BDDCS to 927 drugs.21 Since both BCS and BDDCS use the same criteria for solubility, this work has been taken as reliable source to standardize our data set. Due to the extensive survey, herein we only report the lowest solubility under the conditions listed above. In addition, scale-up guidelines B

dx.doi.org/10.1021/mp4000585 | Mol. Pharmaceutics XXXX, XXX, XXX−XXX

Molecular Pharmaceutics

Article

were taken from Kasim et al.9 whenever solubility data was not available or was undefined. Maximum Dose Strength. Two reference sources were mainly used for searching values of maximum dose strength (mg): (i) the WHO Model List of Essential Medicines22 and (ii) Orange Book.23 For drugs that are not included in these documents or exist in different market presentations, the first introduced strengths were revised and used as highest dosages. Doses in mg/kg were transformed into mg assuming 70 kg as adult body weight. Dose Number Calculations. The dose number (D0) was calculated using the following equation D0 =

(M 0 /V0) S

ceutics classification very useful for identifying good drug candidates (or NMEs) as it provides a round approximation of their actual BCS allocations. However, since the clinical dose and the absorbed fraction are unknown at the initial stage of drug discovery, such a classification remains a challenge. Computational or in silico methods made important advancements over the past decade, and they are the first choice for facing this challenge. Previous attempts for BDDCS modeling were carried out and demonstrated that consensus models based on twoseparated properties (dose number and % total metabolism) predictions can achieve reasonable performance.28,29 That is to say, a robust prediction of dose number, despite being a property of drug in its formulation that can only be defined for NMEs in a clinical context,29 can be effectively conducted using appropriate physicochemical or molecular descriptors such as those implemented in Volsurf+ package.30 On the other hand, several researchers have explored Quantitative Structure−Property Relationship (QSPR) classification approaches involving Caco-2 cell permeability for predicting highly absorbed NMEs. All of these models presented appropriate performances.25 Up to now, a threshold of Papp ≥ 8 × 10−6 cm/s has been used for identifying compounds with Fa ≥ 80%.25,31 Rank-order relationships between high Caco-2 cell permeability and high absorption level showed wide variability. For example, Papp ≥ 1 × 10−6 cm/s has been associated to complete absorption,32 Papp ≥ 10 × 10−6 cm/s with Fa ≥ 70%,33 Papp ≥ 14 × 10−6 cm/s related to Fa ≥ 90%,34 and so on. By adopting a new cutoff value derived from an internal standard drug (Metoprolol, Papp ≥ 20 × 10−6) to define the low/high permeability class boundary,24 Caco-2 measurement based classification models could suitably classify the BCS permeability. Finally, a cutoff value of 16 × 10−6 cm/s was used for classification. Taking all of the above together, in this work computational efforts have been made to construct potential predictors of BCS classes for NMEs based on two separate model series of dose number and Caco-2 cell permeability. To attain this purpose, the following computational procedures were considered: (i) suitably computing physicochemical and molecular descriptors, (ii) rational selection of training and test sets, (iii) establishment of modeling strategy and appropriated variable selection, (iv) rigorous validation with BCS classified known drugs, and (v) prediction of BCS classes for NMEs according to chemical space covered by obtained models. Molecular Descriptor Calculations. Chemical structures of 322 drugs from current data set were expressed as 2D SMILE strings. Using this encoded inputs, 803 simple (0-2D) descriptors belonging to 29 families implemented in Dragon software version 6.0,35 and 126 molecular descriptors in VolSurf+ version 1.0.430 were calculated. A preliminary variable selection was carried out removing missing and nonvariance variables (standard deviation, SD < 10−4). Detailed descriptions of molecular descriptors (MDs) calculated by the Dragon and VolSurf packages can be found in the literature.36,37 Rational Design of the Training and Test Sets. This procedure was carried out by means of k-mean cluster analysis (k-MCA), a multivariate technique implemented in the Statistica software version 8.0.38 To guarantee acceptable statistical quality of data cluster, the number of members in each cluster and the SD of the variables in the cluster (as low as possible) were considered. Additionally, the SD between and within cluster, the respective Fisher ratio, and p-level of significance (p < 0.05) were inspected.

(1)

where M0 is the highest dose strength (mg), S is the aqueous solubility (mg/mL) determined under conditions mentioned above, and water volume V0 is assumed to be 250 mL.1,9 Drugs with D0 ≤ 1 were classified as high-solubility drugs. Conversely, drugs with D0 > 1 were assigned as low solubility drugs.9 Permeability Estimations. There are different forms to estimate the BCS permeability class.9,13,24 As mentioned above, the scientific community currently accepts the application of five methods to prove the high permeability class (BCS class I or II). However, as the number of available data increases, it is more reasonable to use these references in combination for permeability assignments. In this work, in vitro Caco-2 cell permeability is used to classify drug according to BCS. For this purpose, we take advantage of our previous research where an extensive literature survey of this kind of data was processed.25 Besides, we have adopted the same method proposed by Kim et al.,24 taking the average permeability value of Metoprolol (average apparent permeability Papp = 20 × 10−6 cm/s) for benchmarking the high permeability class boundary. All experimental data was revised taking into account the main factors contributing to data variability, such as passage range (28−65), confluence (>85%), transport assay in range of 18−21 days postseeding in pH medium of 6.5−7.4 at 37 °C, and so forth.26 There is not an official standard protocol for Caco-2 available to predict drug absorption in the context of the BCS. Behrens and Kissel have published four important points that should be considered: filter inserts (polycarbonate, PC), collagen coating, average seeding density of 4 × 104 to 6 × 104 cells/cm2 and invariable medium supplement.27 The apparent coefficient of permeation (Papp) must be calculated (cm/s) at sink condition: Papp =

dQ 1 1 × × × Vdonor dt mo A

(2)

where dQ/dt is the permeability rate (steady state transport rate) obtained from the mass transport-time profile of the substrate, e.g. counts/s, A (cm2) the surface area of the monolayer, m0 the original mass of the marker substance in the donor compartment, for e.g. counts, and Vdonor (cm3) is the volume of donor compartment. Due to the large revised literature, the mean values were listed, excluding those laid outside of the mean ±2 SD (standard deviation) ranges. Additionally, available data obtained on both directions apical to basolateral (Papp,A‑B) and vice versa (Papp,B‑A) were taken into account. Computational Methods. BCS is very important for drug authorization. It also predicts drug probability for achieving good oral absorption. This fact makes early provisional biopharmaC

dx.doi.org/10.1021/mp4000585 | Mol. Pharmaceutics XXXX, XXX, XXX−XXX

Molecular Pharmaceutics

Article

(D2), F, and corresponding p were also considered. In the case of BLR, Forward Selection (Conditional) was used. This is a stepwise selection method with entry testing based on the significance of the score statistic and removal based on the probability of a likelihood-ratio statistic (p < 0.05) based on conditional parameter estimates. As a default, entry scores lower than 0.05 and removal scores higher than 0.1 were applied. Additionally, to compare the potency of Dragon and Volsurf+ approaches for BCS classification, discriminant models (LDA, QDA, and BLR) were analyzed according to descriptor families. In silico models were constructed under different combinations: 0-2D Dragon, Volsurf+ and 0-2D Dragon plus Volsurf+. The same procedure was carried out for both solubility and permeability. Model Validation. From the three models obtained with different descriptor groups, the best one was chosen for each classification algorithm (LDA, QDA, and BLR). The final consensus system (for solubility and permeability) was constructed with the predictions of these three selected models. The validation procedure was carried out using results of consensus models. Performances of models were evaluated using false positive rate (FPr), true negative rate (TN, for specificity), true positive rate (TP, for sensitivity), Matthews Correlation Coefficient (MCC), and predictive accuracy, as defined below:

Compounds of the training and test sets were randomly collected from the previous clusters. This procedure allows selecting, in a representative way and in all level of the linking distance (Y-axis), compounds for both sets. Compounds belonging to the test set (around 20% of the complete database) were not used in the development of the discriminant functions and were set aside to validate the obtained models. The splitting procedures according to solubility and permeability response variables were performed independently. Model Building and Feature Selection. Three statistical classification algorithms were applied in order to detect all possible (linear or nonlinear) relationships between solubility/ permeability and computed parameters: LDA (linear discriminant analysis), QDA (quadratic discriminant analysis), and BLR (binary logistic regression). LDA and QDA are well-known to be parametric statistical algorithms based on Fisher’s discriminant analysis,39 while BLR is a parametric system derived from the method of maximum likelihood.40 Briefly, LDA searches for a hyperplane which divides the p-dimensional space of attributes to distinguish the two classes (±) and is represented by a linear discriminant function f LDA(±) fLDA ( +/−) = a0 + a1x1 + a 2x 2 + ··· + apxp

(3)

where a0 is a constant, a1 to ap are the regression coefficients for p variables (x1 to xp). This method makes the parametrical assumptions such as noncolinearity, normality, and homocedasticity that were further ascertained. The quadratic discrimination is similar to the linear one, but in this case the hypersurfaces, which divide the classes, are quadratic. Then the QDA function is f QDA(±) fQDA ( +/−) = a0 + a11x1 + ··· + a1ppxp + a 21x1x 2 + ··· + a 2nxixj +

a31x12

+ ··· +

a3pxp2

(6)

sensitivity = TP/(TP + FN)

(7)

precision = TP/(TP + FP)

(8)

MCC = [(TP × TN) × (FP × FN)] /[(TP + FP)(TP + FN)(TN + FP) − (TN + FN)]1/2

(4)

(9)

where a0 is a constant, a11 to a1p, a21 to a2n, and a31 to a3p are quadratic coefficients, and xi and xj are the i/jth predictor variables. The BLR method produces a mathematic formula that output the probability P of each case to belong to one class, for example, positive (P+), in function (typically a sigmoidal dependency) of predictive variables. That is ln(P+/P −) = a0 + a1x1 + a 2x 2 + ··· + apxp

specificity = TN/(TN + FP)

accuracy = (TN + TP)/(TN + TP + FN + FP)

(10)

The TP (true positive), FP (false positive), TN (true negative), and FN (false negative) were extracted from confusion matrix (Table SI3 in the Supporting Information). The high solubility/permeability class was seen as the positive class. Additionally, the receiver operating characteristic curve (ROC) was used to evaluate the accuracy of the discriminant functions. The ROC curve is a representation of sensitivity versus 1-specificity. The closer the curve follows the left-hand border and the top border of the ROC space, the more accurate the test. Meanwhile the closer the curve comes to the 45-degree diagonal of the ROC space, the less accurate the test is. Accuracy is measured by the area under the ROC curve (AUC). An area of 1 represents a perfect test; an area of 0.5 represents a worthless test.42 Application for BCS Prediction Purposes. The developed models need to be validated with the published BCS drug classifications. As many of them are based on estimations of other parameters, such as logP/CLogP or surface area for permeability, they only serve as a reference for comparing the predictions. Even in WHO guidelines,5 many drugs display inconclusive classification, for example, ciprofloxacin (I/III), cyclosporine A (III/IV), etoposide (II/IV), haloperidol (III/IV), saquinavir (II/IV), and much more. Besides, some commonly occurring deficiencies are recognized, related to incomplete information of current reported studies.14 Here, three data sets were utilized for testing computational models: (i) 57 BCS-classified drugs

(5)

where (P+/P−) is called the odds ratio and ln(P+/P−) the logit transform of P+; a1 to ap are the regression coefficients for p variables (x1 to xp). A probability (P+) of 0.5 is considered as default cutoff for BLR classification rules. As expected, QDA and BLR are useful for capturing nonlinear relationships between dependent variable and predictors. The model building was performed using SPSS and Statistica software.38,41 Stepwise discriminant analysis was mainly used to identify the appropriate subset of variables and to build discriminant models. Forward stepwise methods were first employed for each descriptor family. Specifically, in case of LDA and QDA, at each step all variables were reviewed and evaluated, considering the p-level (p) or Fisher ratio (F), to determine the most influential variable on the discrimination between classes. Subsequently, the “best subset” technique was applied taking Wilks’ λ as subset significance evaluator. Hence the fundamental hypothesis of being equal variance could be refused, leading to develop model that properly separates two groups. Standard statistical parameters such as the square of Mahalanobis distance D

dx.doi.org/10.1021/mp4000585 | Mol. Pharmaceutics XXXX, XXX, XXX−XXX

Molecular Pharmaceutics

Article

extracted from reliable references,5,9,12 (ii) 679 BDDCSclassified oral drugs obtained from a recent in-depth study of Benet et al.,21 and (iii) 37 377 compounds belonging to different categories that have been previously employed for proving computational BDDCS forecast in early drug discovery.29 For reliable predictions of these three external data sets, it is important to consider all applicability domains (ADs) defined by the chemical spaces of the training set. There are many approaches for AD estimation.43 Here, the leverage approach, a geometric method commonly used for QSAR problems, was employed. The leverage of a compound in the original variable space is defined as hi = [X(X′X)−1X′], where X is the descriptor matrix derived from the training set descriptor values.43 The warning leverage (h*) is defined as h* = 3(p + 1)/n, where n is the number of training compounds and p is the number of predictor variables. Compounds with hi > h* were observed to reveal their influence on classification performance. It is not necessary to exclude them from predictions although they appear to be outside AD. However, compounds are considered to be outliers if they lay outside the ±3 standardized residual (δ) range.43 Since the final PBC classification is performed by a consensus system obtained from six models (three solubility and three permeability models) and outlier predictions are excluded, some cases become nonconclusive-classified ones.

correspond to high BCS solubility compounds. Herein, 23 of 25 compounds (92%) with Dmax < 5 mg are classified as high solubility substances. This concordance is 81.69% for drugs with Dmax < 20 mg and dropped quickly to 68.49% for drugs with Dmax < 100 mg. On the contrary, when Dmax increases, drugs are more likely to fall into the low solubility class. However, this rule was found to be misleading for this data. Less than 54% of drugs with Dmax ≥ 203, 300, or 500 belong to low solubility class. To investigate the correlation and impact of solubility and dose on the classification of BCS solubility, the maximum reported aqueous solubility (Smax) and the drug’s lowest dose strength (Dmin) were used for Do calculation (Figure SI1 in the Supporting Information). A high opposite correlation (r2 > 0.76) between solubility and Do was observed, while dose and Do exhibit a low proportional correlation (r2 < 0.24). In addition, 28 drugs changed their solubility classification (low to high class) when the Smax was used to calculate Do. Meanwhile, the use of Dmin changed the classification of 23 drugs from low to high class. Correlation with Calculated Solubility. pH dependent solubility (mg/mL) was calculated using Volsurf + v.1.0.4,30 ALOGPS v.2.1,46 and ACD/Laboratories v.3.0 (Advanced Chemistry Development: Toronto, Canada, http://www. acdlabs.com/products/pc_admet/physchem/physchemsuite/). For Volsurf and ACD/Laboratories, the lowest predicted solubility between pH ranges of 3−8 and 2−8, respectively, were employed for the calculations of Do. For highly soluble drugs classified by reported data, solubility predictions of ACD/ Laboratories package correlate 63.1% on classification. The true positive rates were 79.1% with Volsurf and 57.2% with ALOGPS. Conversely, the calculated solubility of ACD/Laboratories accurately predicted 88.1% of low solubility class; meanwhile Volsurf achieved 80% accuracy, and the prediction accuracy of ALOGPS was 85.9%. Classification of in Vitro Permeability. The mean Caco-2 permeability values ranged from 0.049 (raffinose) to 378.33 × 10−6 cm/s (ethinyl estradiol). The mean and median values of data set were 25.09 and 17.33 × 10−6 cm/s, respectively. The internal standard (metoprolol) was applied as class boundary for high permeability.9,24 However, the Fa value of metoprolol (Fa ≥ 95%)47 is considerably more conservative than permeability criteria of the FDA (Fa ≥ 90%) and WHO (Fa ≥ 85%) guidelines.3,5,48 Based on Kim et al.’s method, the lower limit of 90% confidence interval of the mean permeability test/reference ratio to be greater than 0.8 is an acceptable criteria of highly permeable drug.24 The 90% confidence interval can be calculated from the mean permeability ratio of the test drug and the internal permeability standard. Applying this scheme for classifying current data, 167 compounds were found in the high permeability category. Interestingly, 51 drugs fall into the [0.8−1.25] confidence interval of metoprolol, and 26 compounds display a lower permeability than metoprolol. Among them, only three drugs: etodolac, ketoconazol, and vitamin B2 have a lower GI absorption than 85%. Suitability of Permeability Classification. To further verify the suitability of the permeability class assignments based on Caco-2 permeability, correlation of in vitro permeability/ in vivo intestinal absorption was obtained for 282 drugs. All qualitatively reported drug absorptions (poor, complete, etc.) were not included in the analysis. The scatter plots with error bars of apparent permeability versus fraction of dose absorbed Fa(%) are shown in Figure 1. A relative good correlation between Papp and Fa was achieved. The fitted sigmoidal curve with mean values had a proportional correlation and a high standard error (StdE ≈ 18.5%). The rank-order

■

RESULTS A number of 123 orally administered drugs on the World Health Organization (WHO) Essential Medicine List (EML) were initially classified into BCS.9,12 Later, 200 oral drug products in the United States, Great Britain, Spain, and Japan were classified based on published solubility data and permeability data estimated by calculated log P.11 Recently, increasing attention has been turned out for determining the Provisional Biopharmaceutical location of orally administered immediaterelease (IR) drug products using different estimated gastrointestinal permeability, such as partition coefficients (log D and log P), molecular surface area (PSA), or other in vitro permeability.13−15,44 It is emphasized that the distributions of BCS class I, II, III, and IV drugs in each classification are quite different. In this report, taking advantage of the availability of experimental in vitro Caco-2 cell data, a Provisional Biopharmaceutical Classification (PBC) of 322 oral drug products gathered from literature was performed. To our knowledge, it is the largest data set for such classification. Classification of Drug Solubility. It was carried out by the direct comparison between dimensionless Do parameter and unity. On the basis of this criterion, 188 of the 322 drugs were classified as high-solubility. Compounds with aqueous solubility lower than 0.1 mg/mL could negatively affect the oral absorption.45 This is a useful thumb rule for classifying the BCS solubility. The application of this rule to the current data reveals that 218 compounds had a minimum solubility value (Smin) ≥ 0.1 mg/mL, and 171 of them have Do ≤ 1, corresponding to 78.44% of true positive (TP). Conversely, from 104 compounds with Smin < 0.1 mg/mL, 90 are low solubility entities, corresponding to 86.54% of true negative (TN). With respect to the maximum dose (Dmax), values that correspond to the 10th percentile, the 25th percentile, the median, the average, the 75th percentile and 90th percentile of the dose distribution for this data were determined to be 5, 20, 100, 203, 300, and 500 mg. Based on these k-th percentile, the rank-order relationship between Dmax and Do could be examined. BCS background leads to expect that low Dmax would usually E

dx.doi.org/10.1021/mp4000585 | Mol. Pharmaceutics XXXX, XXX, XXX−XXX

Molecular Pharmaceutics

Article

The first one is based on correlations of human intestinal permeability with of pH-dependent and pH-independent n-octanol/ water partition coefficients (log P) and n-octanol/water distribution coefficients (log D).9,11 This method also takes metoprolol as the reference compound for directly comparing the estimated log D or log P. The second work applies criteria of polar surface area (PSA) for permeability class assignments.15 It involves the use of labetalol (Fa ≈ 90%) as a high permeable internal standard.55 These methods can be considered as in silico approaches since they provide a simple, fast, and conservative classification. Herein we considered interesting to evaluate such criteria for current data. The Mlog P, log D7.5 (at pH = 7.5), and TPSA(Tot) descriptors of Dragon and Volsurf softwares were used.35−37 Figure 2 reveals the suitability of this classification and correlation of in vitro Caco-2 permeability with computed properties. As can be seen from the scatter plots, there is no clear trend between Fa or in vitro permeability and three parameters. Likewise Caco-2 permeability, compounds with Fa < 85% (or 90%) could not be identified. Mlog P and log D7.5 of Metoprolol are 1.65 and 0.09, respectively. It was evidenced that Mlog P provides a better correlation with high absorption than log D7.5. Among 155 drugs with Mlog P ≥ 1.65, 111 compounds have Fa ≥ 85%, and 101 have Fa ≥ 90%, corresponding to 71.6% and 65% of TP according to WHO and FDA criteria, respectively. Interestingly, TPSA(Tot) yields slightly better prediction of high absorption than Mlog P. Calculated TPSA(Tot) for labetalol by Dragon software receives a value of 95.85 Å2.35 From data set, 172 drugs have TPSA(Tot) ≤ 95.85 Å2. As the results, 124 of them show a Fa ≥ 85% (accuracy of 72.1%), and 116 have Fa ≥ 90% (correctly 67.4%). This descriptor also predicts the high in vitro permeability class better than the others with 67.7% TP rate. It should be noted that TPSA(Tot) shows an inverse relationship with Fa and Caco-2 Papp. In the literature, polar molecular surface is recognized as a relevant physicochemical property governing drug transfer across cell membranes and oral absorption.56,57 Provisional Biopharmaceutical Classification (PBC) of Drugs. Previous solubility and permeability classifications enable to develop a PBC for current data. According to this, 91 drugs have been classified as PBC class I (28.3% of data), 76 drugs as PBC class II (23.6%), 97 drugs as PBC class III (31.1%), and 58 drugs as PBC class IV (18.0%). Table SI1 details all of the classifications and other related absorption properties. Additionally, 171 drugs are encountered with BCS classification in reliable references,5,9,11,12,18,44 and 264 drugs are published recently with BDDCS.21 They were analyzed to ascertain the concordance degree between PBC with literature. Please note that, for cases which could not be found a specific BCS class (also, namely, inconclusive data),12 the drug classification of PBC is considered to be accurate if it agrees with one of the assignments of existing BCS. Consequently, of the 79 orally administered drugs on the BCS classification of WHO list,5 60 were correctly classified by PBC (76% of global precision). The accuracies were 80.8% for Class I, 78.6% for Class II, 62.5% for Class III, and 86.67% for Class IV. The translocation was adopted to BCS class I of some drugs that were previously considered to be in class III in WHO report,5 such as paracetamol, acetylsalicylic acid, allopurinol, lamivudine, and promethazine. In comparison with other studies that used the permeability criteria of log P or log D,9,11 172 drugs have been analyzed. In the report of Takagi et al., together with the use of metoprolol as a reference drug for permeability classification, atenolol, cimetidine, and ranitidine were also used for comparing the correlation between the BCS (based on estimated log P) and BDDCS. Therefore, some drugs appeared with two or more classifications;

Figure 1. Scatter plots with error bar illustrating the sigmoidal and rankorder relationships between Caco-2 cell permeability (Papp) and human intestinal absorption (Fa). The fitting constants (A; B; C) of Fa function f(x) = A(1 + B exp(−C·Papp) were determined as (103.837; 0.001; 1.124). The pink and blue dotted lines express the high BCS permeability cutoff as defined by US FDA (Fa ≥ 90%) and WHO (Fa ≥ 85%) guidelines. The continuous red line indicates the high PBC permeability criterion derived from Caco-2 permeability of metoprolol.

relationship between high permeability and high absorption can be clearly appreciated.24,34 141 compounds from this set are categorized as high permeability class: 131 of them have Fa ≥ 85% (92.9% of true positive with WHO criteria),5 and 123 had Fa ≥ 90% (being equivalent to 87.2% of TP with US FDA guideline).3 Only for 10 compounds, the Caco-2 model overestimates their absorption extent in human: bromazepam (PBC class I), chloropheniramine (class I), danazol (class II), efavirenz (class II), etodolac (class II), flecainide (class I), griseofulvin (class II), guanabenz (class I), ketoconazole (class II), and riboflavin (class I). From the Table SI1 in the Supporting Information, bromazepam displays a Fa value of 84%. This value is slightly lower than absorption boundary (85%) used by the WHO for BCS permeability classification.5 Because 85% is an arbitrary value and because we can tolerate the threshold by applying the [0.8, 1.25] confidence interval rule,24 bromazepam can be considered as unclassified permeability drug. Similarly, riboflavin (active transport),49 guanabenz and flecainide (passive diffusion) show a Fa ≈ 80%. For compounds such as chlorpheniramine and danazol, widely varied absorption data have been reported in the literature. Other water-insoluble drugs are griseofulvin,50 efavirenz (higher Fa at dose ≤ 100 mg),51 and ketoconazole (Fa decreases at gastric pH 4.0−6.5).52 Furthermore, the efflux effects limit the absorption of efavirenz and ketoconazol (saturable).53 In case of etodolac, which is a weak acid, different absorption (Fa from bioavailability) data were found in the literature. This compound has pH-dependent absorption with a higher value (bioavailability >80%) at basic medium. Previous report has classified this compound as BCS class I at pH 7.4 alone.54 On the other hand, weak correlation was found between low permeability and low Fa. The scatter plot reveals this limitation since 30.5% of low permeability compounds display Fa ≥ 85%. Other Criteria for Permeability Classification. There are two common ways to conduct a BCS permeability classification. F

dx.doi.org/10.1021/mp4000585 | Mol. Pharmaceutics XXXX, XXX, XXX−XXX

Molecular Pharmaceutics

Article

Figure 2. Comparison of concordances between BCS permeability classifications using calculated physicochemical properties (partition coefficient, distribution coefficient, and polar surface area) and intestinal absorption criteria (A1, B1, C1). The same procedure was carried out for comparing with the Caco-2 permeability classification (A2, B2, C2). The pink, blue, and red dotted lines express the high BCS permeability cutoff as defined by the US FDA (Fa ≥ 90%), WHO (Fa ≥ 85%), and Caco-2 permeability. The continuous red lines in A1, A2, B1, and B2 indicate the calculated log P/log D values of metoprolol. The continuous red lines in C1 and C2 are calculated PSA of labetalol.

exhibit Fa ≤ 70%. Regarding these compounds, it is important to emphasize the difficulty in classifying amoxicillin, chlorpromazine, doxycycline, haloperidol, lamivudine, levodopa, nitrofurantoin, salbutamol, and stavudine (failure for either PBC or log P based BCS). This analysis sought that the combination of PBC and log P based BCS could enhance the prediction of the in vivo BCS. The work of Lindenberg et al. presented a different method for classifying drugs according to BCS. This study utilized bioavailability data to permeability BCS class assignments (with 90% as a conservative cutoff).12 70 drugs of present data were previously classified by Lindenberg et al.12 Interestingly, from 23 compounds assigned as class I PBC, 15 (over 65.2%) agree with Lindenberg’s classification, 100% coincide in class II

for example, antipyrine, caffeine, cetirizine, codeine, ephedrine, and lomefloxacin were BCS class I according to Atenolol and meanwhile class III with respect to metoprolol, whereas for morphine, the situation was the opposite. The differences between log P and CLogP values also contribute to the inconclusive classification. In detail, the agreement between previous BCS classifications with PBC for class I were 69.6% and 89.8% for class II, 74.5% for class III, and 14.8% for class IV. To investigate the lack of correspondence of class IV assignments, the classification of WHO was taken as a reference. Interestingly, nearly half (43.5%) of class IV PBC are inconclusive drugs (BCS class II or IV according to WHO), while the log P based BCS of these drugs corresponds to class I or II. Besides, the inspection of intestinal absorption reveals that 81.8% of class IV PBC drugs G

dx.doi.org/10.1021/mp4000585 | Mol. Pharmaceutics XXXX, XXX, XXX−XXX

Molecular Pharmaceutics

Article

Figure 3. Relationship among intestinal absorption (A1), Caco-2 permeability (A2), and efflux ratios. The legend of pink and blue dotted lines and continuous red line are found in Figure 2 capture. The dotted gray line and continuous gray line show two criteria for defining substrate status (NES vs ES) according to in vitro data.

substrate of the efflux transporter, the efflux ratio (EfR) is calculated as follows:

BCS assignment, 67% coincide in class III BCS, and 73% is in agreement with previous class IV BCS. More than 41% of drugs classified as III PBC are of class I and I/III in the Lindenberg’s study,12 suggesting the risk of high false negative when Caco-2 data is used. However, this in vivo permeability criterion (human oral bioavailability) seems to be too exigent with respect to the BCS application. Furthermore, current PBC was compared with the BCS classification scheme proposed by Wu and Benet.18,44 Although their method was not specified, this data has been widely applied for comparing many classification results according to BCS and BDDCS.10,13,21 Current data includes 98 drugs of that study. Likewise analysis above, the PBC agrees with the other classification: 87.1% for class I BCS, 82.6% for class II BCS, 69.0% for class III BCS, and 13.3% for class IV BCS. Indeed, 81.2% of PBC class IV drugs are classified as class II by Wu and Benet.18,44 Most of them (92.9%) have Fa ≤ 70%, and the BCS classification of WHO shows that they are rather inconclusive data. Strong correlation (about 90−95%) between extent of metabolism and intestinal permeability of drugs is recognized.18,21 This correlation was also inspected with PBC assignments. Recently, over 900 marketed drugs were classified for BDDCS20 and 264 of them appeared in current PBC data. Based on BDDCS, 90 drugs are class I, 89 are class II, 68 are class III, and 17 are class IV. The concordance between the BDDCS and our PBC was quite good, being a 73.3% for class I, 63.0% for class II, 78.1% for class III, and 64.7% for class IV. Besides, from 179 extensively metabolized drugs, only 126 display high Caco-2 permeability for a correlation of 70.4%. In contrast, according to the criterion of WHO, only 71.4% of high metabolism class of BDDCS matches with the high permeability of BCS (regardless of inconclusive data). PBC and Efflux Transport Effect. ABC (ATP-binding cassette) efflux transporters such as P-glycoprotein (P-gp), the multidrug resistance-associated protein 2 (MRP2), and the breast cancer-resistant protein (BCRP) are expressed on the mucosal membrane of Caco-2.58,59 These proteins located at the apical membrane are considered as the rate-limiting barrier to monolayer permeation and intestinal drug absorption. Therefore the BCS and/or PBC may be strongly influenced by effluxmediated transport drugs. To ascertain whether a drug is

EfR =

Papp,B → A Papp,A → B

(11)

Figure 3 represents the Caco-2 permeability (172 drugs) and human intestinal absorption (157 drugs) of different PBC classes versus efflux substrate status. As can be appreciated from Figure 3, drugs of the PBC class III and IV are mostly affected by efflux transport. Generally, compounds with EfR ≥ 2 are possible efflux substrate (ES), and EfR ≤ 1.2 is a sign of nonefflux substrate (NES). The range 1.2 < EfR < 2 is borderline and would need further experiments to confirm the substrate status.60 On Figure 3 (A1), 94% of PBC class I drugs present Fa ≥ 85%, and 84% of them are NES. Only one compound (fluvastatin) displays an EfR > 2. Subsequently, 91.5% of class II drugs have Fa ≥ 85%, and 82.9% of them are NES. Bilutamide and telmisartan are two class II drugs out of range EfR > 2. On the opposite, 73.7% of class III drugs does not present high absorption (Fa < 85%). Compounds from this category are the most susceptible drugs with respect to the substrate specificity. Only 67.6% of PBC class IV drugs have absorption characteristic (Fa) similar to class III. More than 71% of drugs in this category are possible ES. As being included more scatters on Figure 3 (A2), the proportions of NES in class I and II diminished to 84% and 85%, respectively; meanwhile, 70.3% of class IV drugs are ES. Unfortunately, identification of the substrate status of PBC class III compounds is quite limited. The impact of efflux substrate on PBC class III drugs may be reduced due to the possible saturation effect. Among ES drugs, it is emphasized that the absorption profiles of nine compounds (cetirizine, clarithromycin, domperidone, ketorolac, lamivudine, nitrofurantoin, prednisolone, sulfinpyrazone, and sulindac) present in vitro characteristics of substrates with relatively low Caco-2 permeability (Papp ≤ 5 × 10−6 cm/s) but still achieve high absorption (Fa ≥ 85%). All of them are classified as class III and IV by PBC. These drugs are substrates and/or inhibitors of P-gp, multidrug resistance protein 2 and 4 (ABCC4, ABCC2), and breast cancer resistance protein (BCRP/ABCG2).61 Especially, nitrofurantoin (EfR > 23) is classified as class II (according to bioavailability data) by WHO H

dx.doi.org/10.1021/mp4000585 | Mol. Pharmaceutics XXXX, XXX, XXX−XXX

Molecular Pharmaceutics

Article

and Lindenberg et al.;5,12 meanwhile all log P based BCS, BDDCS as well as current PBC identify it as class IV. The active uptake involved nitrofurantoin in vivo transport62 may be the reason for this apparent contradiction. Distribution of Drug Therapeutic Category. The PBC class distribution based on therapeutic category of drugs was examined. The Anatomical Therapeutic Chemical (ATC) Classification System, proposed by WHO (available at http:// www.whocc.no/atc/structure_and_principles/), was used for the classification of drugs. Herein, only immediate-release oral formulations were considered, and drugs with two or more codes were taken into account. Due to a limited number, some ATC classes were combined based on the mechanism and target allocation. Figure 4 displays the distributions of PBC classified drugs within each therapeutic category.

hormone/antihormone drug groups represent no more than 5% within this PBC class. Physicochemical Profiling of PBC. It is very useful to analyze the similarity between physicochemical spaces characterized by PBC classes, especially for developing computational predictions of current PBC and further BCS. Thus, six commonly used physicochemical parameters were calculated by Dragon and Volsurf+ for this analysis:30,35 molecular weight (MW), polar surface area (PSA), Mlog P, log D6, log D7.5, total number of hydrogen bond donors and acceptors (nHA+B), number of free rotatable bonds (RBN), and estimated ionization states. The average and median values of maximum dose strength (Dmax) as well as Caco-2 Papp were also analyzed for each class. Table 1 displays the distributions of these properties within each class. As can be observed in Table 1, class II drugs display the highest lipophilicity, while class III and IV are more hydrophilic. Class I drugs represent a balanced physicochemical profile even though they tend to be more lipophilic. In general, only the hydrogen bonding term is fairly different from one class to another. There is certain physicochemical similarity between class I and II (Mlog P, log D at basic medium), class III and IV (nHA+B, PSA), or class II and III (MW), etc. Values of Dmax do not display any trend. It is demonstrated that poor absorption is more likely when the compounds violate two or more of the Lipinski’s rules (Ro5):63 (i) logP < 5, (ii) MW < 500, (iii) HBD (hydrogen bond donors) < 5, and (iv) HBA (hydrogen bond acceptors) < 10. Current data was recollected mostly from successful drugs. Then, it is easy to understand that many of them (>95%) passed the Ro5. Computational Models To Predict PBC Class from Chemical Structures. Solubility and Caco-2 permeability were modeled independently. The final computational PBC classification was achieved using two voting consensus (permeability and solubility) systems. QSPR models obtained by different statistical techniques for each property are described below. Solubility Modeling. Three model series were obtained using LDA, QDA, and BLR. Different molecular descriptors (MDs) were used for building QSPR models. From every constructed model groups, the best one was selected (detailed comparisons are described in Supporting Information). Table 2 summarizes the mathematical equations and performances of the three best models for PBC solubility classification. Performances of the consensus model (VoteS) have been calculated and compared with those of other models. The ROC curves are shown in the Supporting Information. Permeability Modeling. As with solubility, the same procedure was carried out to select the best models for predicting the PBC permeability classes. Table 3 displays the relevant information of these models. Classifications of Four PBC Classes. The two obtained voting models were finally combined to estimate the four PBC classes of the data (322 compounds). Table 4 displays the confusion matrix (for four classes) of this consensus system. A good overall accuracy of 73.0% was obtained. Analysis of Molecular Descriptors (MDs). Interestingly, the PBC solubility and permeability properties are well-described based on a small set of MDs. Many of them are simple and interpretable descriptors, while the others are calculated physicochemical properties, such as CACO2, LgD5, and so forth (Table SI10 in the Supporting Information). Although the Volsurf solubility calculations and their derived descriptors have been shown to be better correlated with Do,29 actual selected descriptors do not agree with that findings. It is important to note that there are some MDs directly related to polarizability and dispersion forces within molecules

Figure 4. Distribution of therapeutic classes of drugs. The first level encoding letters used in the Anatomical Therapeutic Chemical (ATC) Classification System were employed. NS: anesthetics, analgesics, antipyretics, nonsteroidal anti-inflammatory, medicines used to treat gout and disease-modifying agents used in rheumatoid disorders; RS: antiallergics, medicines used in anaphylaxis and antidotes; N: anticonvulsants/antiepileptics, antimigraine, antiparkinsonism, and other psychotherapeutic medicines; J1: antibacterial and antifungals; J2: antiviral; JP: antiprotozoal; L: antineoplastics, immunosuppressives; B: medicines affecting the blood; BH: diuretics and cardiovascular medicines; AV: gastrointestinal medicines; GH: hormones, other endocrine medicines and contraceptives; A: vitamins, minerals, and other nutritional supplements.

Current data includes a high number of cardiovascular medicines and diuretics (79 drugs) and anti-infective medicines (a total of 80 antibacterial, antiviral, and antiprotozoal drugs). The second group is psychotherapeutic medicines (64 drugs). The following group is analgesics, antipyretics, nonsteroidal antiinflammatory medicines (49 drugs), and so on. According to PBC, the class I drugs are mainly present in psychotic, vitamin, and nutritional drug products (>50%); meanwhile they are absent from antineoplastic and antibiotic groups (