Article pubs.acs.org/jmc
Clearance Mechanism Assignment and Total Clearance Prediction in Human Based upon in Silico Models Franco Lombardo,*,† R. Scott Obach,§ Manthena V. Varma,§ Rowan Stringer,‡ and Giuliano Berellini*,† †
Metabolism and Pharmacokinetics, Novartis Institutes for Biomedical Research, 250 Massachusetts Avenue, Cambridge, Massachusetts 02139, United States ‡ Metabolism and Pharmacokinetics, Novartis Institutes for Biomedical Research, Wimblehurst Road Horsham, West Sussex, RH12 5AB, United Kingdom § Pharmacokinetics, Dynamics and Metabolism, Pfizer Global Research and Development, Groton, Connecticut 06340, United States S Supporting Information *
ABSTRACT: We introduce a two-tier model based on an exhaustive data set, where discriminant models based on principal component analysis (PCA) and partial least squares (PLS) are used separately and in conjunction, and we show that PCA is highly discriminant approaching 95% accuracy in the assignment of the primary clearance mechanism. Furthermore, the PLS model achieved a quantitative predictive performance comparable to methods based on scaling of animal data while not requiring the use of either in vivo or in vitro data, thus sparing the use of animal. This is likely the highest performance that can be expected from a computational approach, and further improvements may be difficult to reach. We further offer the medicinal scientist a PCA model to guide in vitro and/or in vivo studies to help limit the use of resources via very rapid computations.
■
INTRODUCTION Human clearance is a very important pharmacokinetic (PK) parameter, as it influences both half-life (together with the volume of distribution) and bioavailability (together with oral absorption), and thus impacts the dose regimen (how often) and dose size (how much) of a drug. Pharmaceutical researchers are mostly concerned with the fate of potential new medicines in human and rely on in vitro and in vivo animal studies. However, the difficulty in predicting clearance is enhanced by the many pathways in which the drug can be cleared. For any given drug the mechanism of clearance is generally conserved across mammalian species, although it is likely that the rate of clearance will differ and/or exceptions related to mechanism are possible. Drugs that are cleared by hepatic metabolism in rats generally tend to be cleared by hepatic metabolism in humans; drugs that are secreted unchanged in urine in rats will be likewise cleared renally in humans and so forth. This is generally due to the fact that clearance mechanisms are largely governed by physicochemical properties such as lipophilicity and charge state of the compound. Lipophilic drugs require metabolism to form more hydrophilic metabolites in order to become membraneimpermeable enough to permit their removal in excreta without reabsorption from excretory tissues (i.e., kidney and bowel). For compounds that are eliminated by metabolism potential interspecies differences in both the expression and activity of drug metabolizing enzymes may make clearance prediction difficult. The situation is complicated further by compounds © 2014 American Chemical Society
that either show limited hepatocyte permeability (e.g., HIV protease inhibitors) or undergo accumulation in hepatocytes by active processes (e.g., statins). In vivo scaling methods have been proposed to predict human clearance, and they have traditionally relied on data from in vivo studies in preclinical species, mainly rat, dog, and monkey.1−14 These methods, however, are generally resourceintensive and time-consuming with significant cost. By use of these methods, 60−65% of compounds are predicted within 2fold error of observed values, the latter being a seemingly accepted threshold in the field. This was shown by a recent report where in vivo methods have been compared on the basis of a large set of experimental human and animal PK data.15 The difficulty in predicting the human clearance within 2-fold was exhibited, and the comparison highlights the need of nonrodent species, in particular monkey, the latter being of crucial importance for neutral compounds. Thus, considering the variability in quantitative aspects of clearance, its prediction may not go as far as one would expect in terms of accuracy. Hence, the use of several methods may improve confidence in prediction, especially when there is agreement between in vitro and in vivo methods. In vitro clearance predictions utilizing human liver microsomes or hepatocytes have been described. While these methods are less expensive compared to in vivo studies, they Received: March 19, 2014 Published: April 28, 2014 4397
dx.doi.org/10.1021/jm500436v | J. Med. Chem. 2014, 57, 4397−4405
Journal of Medicinal Chemistry
Article
Figure 1. (A) Two-component PCA score plot of discriminant model based on computed physicochemical descriptors. Compounds primarily (≥70%) cleared by metabolic pathways are shown in red. Those cleared by renal pathways and by biliary pathways are in blue and green, respectively. The dotted line represents the equation PC2 = 3PC1 − 1 used to discriminate renal from metabolic pathways, while the shaded area represents the uncertain region with either renally or metabolically cleared compounds. The few red dots located within the largely renal (blue) region represent some phosphate prodrugs or compounds with a formamide group recognized by the identification of those specific fragments. The compounds outside the outer ellipsoid were omitted for illustration clarity. (B) PCA loading plot representing the descriptors underlying the relative position of molecules in the PCA score plot in (A). The descriptors are colored according to their type.
Gombar and Hall30 recently reported in silico methods for the prediction of volume of distribution, based on a sizable set of human data originally published by Obach et al.32 This study reported a low percentage ( 0, for the assignment of a compound as metabolically cleared, and PC1 < 0 for the assignment of a compound as renally cleared. A somewhat more balanced assignment could be achieved by using the equation PC2 = 3PC1 − 1, as shown in Table 2, between renal and metabolic classes considering that the renal compounds (blue in Figure 1) are largely located at the right of the line and thus in the region of space identified by PC2 > 3PC1 − 1. Conversely, metabolic compounds (shown in red) occupied a region of space identified by PC2 < 3PC1 − 1. However, the changes are fairly small and the two assignment approaches could be considered equivalent with some slight preference for the M3/R3 classes which improve by a few points in going from using PC1 = 0 to using PC2 = 3PC1 − 1. We further tested the quality of the assignment and its ruggedness using only the M1 and R1 compounds. To perform the test, we listed each class in ascending order of clearance and selected only 20% of the compounds choosing the first, sixth, eleventh, and so on for group 1 (i.e., M1_1 for the metabolic compounds), then the second, the seventh and the twelfth (M1_2) and repeating the process in order to generate five groups. The same was done for the renal compounds, and groups R1_1 to R1_5 were generated. We excluded each group, in turn, cast a new PC model using the other four groups combined and then used the excluded group as a test set. It is possible that analogues of a compound present in a given group (test set) may be represented in the other four groups (training set), but this approach still serves as a rugged test of predictive accuracy and stability. From Table 3, it can be
on On the our
■
RESULTS AND DISCUSSION In this section we discuss the findings from the application of a two-tiered approach and the statistical and predictive quality of the models: the principal component analysis (PCA), on which the mechanism discriminant analysis is based, and the partial least squares (PLS), on which the quantitative clearance prediction is based. Mechanism Discriminant Analysis. We sought to build a qualitative model, based on principal component statistics, that would be capable of assigning the mechanism of elimination based on computed descriptors only. We also found that the utilization of formamide and phosphate (prodrugs) fragments improved the results, and we maintained these fragments in the data set throughout the testing performed and discussed below. The results are graphically shown in Figure 1A together with the discriminant equation (PC2 = 3·PC1 − 1) utilized for the assignment of compounds projected onto the plot. The principal component (PC) model was built using classes M1 (metabolic, red dots) and R1 (renal, blue dots) together with biliary compounds (termed “B1” and depicted by green dots), although the latter were too few to allow further (quantitative) application of those data. Nevertheless the model is able to segregate well the region of space mostly occupied by the latter compounds as well as showing a good discrimination between renal and metabolic compounds. The corresponding statistics are reported in Tables 1 and 2, and they are based on the model shown and built using only
Table 3. 5-Fold Cross-Validation of Mechanism Prediction Based on PC2 = 3PC1 − 1 Line as Discriminant
Table 1. Accuracy of Predicted vs Experimental Mechanism Assignments Based on PC1 = 0 Line as Discriminant M1−3/R1−3 739 metabolic renal total
88 77 84
M1/R1
M2/R2
M3/R3
Number of Compounds in Class 469 115 155 % of Accurate Assignments 91 79 83 85 57 64 88 67 81
% accuracy for metabolic compounds
M2−3/R2−3 % accuracy for renal compounds overall total
270 84 58 77
M1/R1
M2/R2
M3/R3
M2−3/R2−3
Number of Compounds in Class 739 metabolic renal total
87 80 84
469 115 155 % of Accurate Assignments 90 73 87 87 62 64 88 67 85
M1_1 83
M1_2 84
M1_3 96
M1_4 93
M1_5 91
all 86
R1_1 86
R1_2 89
R1_3 81
R1_4 86
R1_5 89
88
seen that the method is very stable, as there are very small variations across subgroups and the model performance does not fall below 81% of accurate assignments. Indeed, some groups (M1_3, M1_4, M1_5 and R1_2 and R1_5) closely approach or exceed 90%. This cross-validation indicated that the performance of the model for the mechanism discrimination is good and further confirms that the assignment is reliable and while approaching an overall accuracy of 90%. We further show that the 469 compounds used so far (M1/ R1 classes) well represent the total space comprised by the total 1003 compounds and the 42 physicochemical descriptors used to generate the model (Figure 2). Thus, the PCA model may be considered rugged and represents a very wide variety of structural and therapeutic classes, physicochemical properties, and clearance mechanisms. These findings prompted some consideration related to the application of this approach to address (i) whether further improvement in the computational assignment of the mechanism would be possible within the data set available and (ii) whether the computational mechanistic assignment, if
Table 2. Accuracy of Predicted vs Experimental Mechanism Assignments Based on PC2 = 3PC1 − 1 Line as Discriminant M1−3/R1−3
all 90
270 83 62 77
M1 and R1 compounds, i.e., the compounds with the highest degree of clearance by a single mechanism (≥70%). Therefore, the mechanism assignment statistics for M1/R1 compounds should be considered as coming from the training set while M2/R2, M3/R3, and M2−3/R2−3 are a bona fide test of accuracy. The data reported in Table 1 allow distinction 4399
dx.doi.org/10.1021/jm500436v | J. Med. Chem. 2014, 57, 4397−4405
Journal of Medicinal Chemistry
Article
Figure 2. Superposition of the 469 compounds (black dots) used for the development of the PCA discriminant method in the two-component plot of all 1003 compounds present in the data set. The 534 compounds not used are shown as white dots.
Table 4. 5-Fold Cross-Validation of Quantitative Clearance Prediction Using Known Experimental Mechanism no. compds GMFE %
