Subscriber access provided by GUILFORD COLLEGE
Brief Article
Prediction of fraction unbound in microsomal and hepatocyte incubations – a comparison of methods across industry datasets Susanne Winiwarter, George Chang, Prashant Desai, Karsten Menzel, Bernard Faller, Rieko Arimoto, Christopher Keefer, and Fabio Broccatelli Mol. Pharmaceutics, Just Accepted Manuscript • DOI: 10.1021/acs.molpharmaceut.9b00525 • Publication Date (Web): 26 Jul 2019 Downloaded from pubs.acs.org on July 31, 2019
Just Accepted “Just Accepted” manuscripts have been peer-reviewed and accepted for publication. They are posted online prior to technical editing, formatting for publication and author proofing. The American Chemical Society provides “Just Accepted” as a service to the research community to expedite the dissemination of scientific material as soon as possible after acceptance. “Just Accepted” manuscripts appear in full in PDF format accompanied by an HTML abstract. “Just Accepted” manuscripts have been fully peer reviewed, but should not be considered the official version of record. They are citable by the Digital Object Identifier (DOI®). “Just Accepted” is an optional service offered to authors. Therefore, the “Just Accepted” Web site may not include all articles that will be published in the journal. After a manuscript is technically edited and formatted, it will be removed from the “Just Accepted” Web site and published as an ASAP article. Note that technical editing may introduce minor changes to the manuscript text and/or graphics which could affect content, and all legal disclaimers and ethical guidelines that apply to the journal pertain. ACS cannot be held responsible for errors or consequences arising from the use of information contained in these “Just Accepted” manuscripts.
is published by the American Chemical Society. 1155 Sixteenth Street N.W., Washington, DC 20036 Published by American Chemical Society. Copyright © American Chemical Society. However, no copyright claim is made to original U.S. Government works, or works produced by employees of any Commonwealth realm Crown government in the course of their duties.
Page 1 of 28 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60
Molecular Pharmaceutics
Prediction of fraction unbound in microsomal and hepatocyte incubations – a comparison of methods across industry datasets Susanne Winiwarter*†, George Chang‡, Prashant Desai║, Karsten Menzel§, Bernard Faller#, Rieko Arimoto¤, Christopher Keefer‡, Fabio Broccatell+. †DMPK,
Research and Early Development, Cardiovascular, Renal and Metabolism (CVRM),
BioPharmaceuticals R&D, AstraZeneca, Gothenburg, Sweden; ‡Pfizer Inc., Groton, CT 06340; ║Eli
Lilly and Company, Indianapolis, IN 46285; §MSD, West Point, PA 19486; #Novartis,
Switzerland; ¤Vertex Pharmaceuticals Inc., Boston, MA 02210; +Genentech Inc., South San Francisco, CA 94080
KEYWORDS: IVIVc, fraction unbound in the incubation, hepatocytes, microsomes, predictive models.
ACS Paragon Plus Environment
1
Molecular Pharmaceutics 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60
Page 2 of 28
ABSTRACT
The fraction unbound in the incubation, fu,inc, is an important parameter to consider in the evaluation of intrinsic clearance measurements performed in vitro in hepatocytes or microsomes. Reliable estimates of fu,inc based on a compound’s structure have the potential to positively impact the screening timelines in drug discovery. Previous works suggest that fu,inc is primarily driven by passive processes and can be described using physico-chemical properties such as lipophilicity and the protonation state of the molecule. While models based on these principles proved predictive in relatively small datasets that included marketed drugs, their applicability domain has not been extensively explored. The work presented here from the in silico ADME discussion group (part of the International Consortium for Innovation through Quality in Pharmaceutical Development, the IQ consortium) describes the accuracy of these models in large proprietary datasets that include several thousand compounds across chemical space. Overall, the models do well for compounds with low lipophilicity. In other words, the equations correctly predict that fu,inc is in general above 0.5 for compounds with a calculated logP of less than 3. When applied to lipophilic compounds, the models failed to produce quantitatively accurate predictions of fu,inc with a high risk to underestimate binding properties. These models can, therefore, be used quantitatively for less lipophilic compounds. On the other hand, internal machine-learning models using the company’s own proprietary data-set also predict compounds with higher lipophilicity reasonably well. Additionally, data shown indicates that microsomal binding is in general a good proxy for hepatocyte binding.
ACS Paragon Plus Environment
2
Page 3 of 28 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60
Molecular Pharmaceutics
INTRODUCTION The fraction unbound in the incubation, fu,inc, is an important parameter to be included in the estimation of a compound’s in vivo clearance from in vitro metabolic stability measurements.1-7 This parameter is now routinely used in drug discovery: to predict human pharmacokinetics (PK) from in vitro data, to understand the in vitro to in vivo correlation of PK in preclinical species, and to better judge the significance of the in vitro CLint value. Various methods have been described to measure the fraction unbound in both microsome and hepatocyte incubations.3, 8-10 Lipophilicity and ion class of the compound have been identified as the main determinants of fu,inc.3,
8, 11-13
Other authors have suggested to use plasma protein binding to account for
incubational binding during scaling4 or to estimate binding in microsomes by mechanistic considerations.1 Machine-learning models have also been presented.9,
14
The simple equations
based on lipophilicity or plasma protein binding are attractive since these parameters are easily available or predictable for any compound. Unlike machine-learning models, which rely on the availability of extensive proprietary datasets, these simple models can be used and reproduced across different organizations. However, the published empirical correlations are based on rather limited datasets with less than 100 marketed drugs; only the method proposed by Turner et al12 is based on several hundreds of acidic, basic and neutral compounds. Earlier evaluations indicated some usage limitations, especially with lipophilic compounds and specific ion classes.10, 15 The aim of the present work was to investigate the performance of the models on a large and diverse set of compounds, with focus on evaluating the impact of using predicted fu,inc values in absence of measured data on decision making in a real-world setting. As such, fold-error in the estimation of the absolute fu,inc values provides the direct assessment of its impact on estimating the intrinsic clearance. Members of the in silico ADME discussion group of the International
ACS Paragon Plus Environment
3
Molecular Pharmaceutics 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60
Page 4 of 28
Consortium for Innovation through Quality in Pharmaceutical Development (the IQ consortium) used compounds in their respective company’s in-house datasets to evaluate these models. In order to reduce the complications associated with sharing and publishing proprietary data, results for each company were summarized and categorized internally. Only the summarized data were then combined into a master dataset for overall analysis. In-house machine-learning models were utilized for the respective datasets when available at a company for additional comparison. We are, thus, able to show an unbiased comparison on several thousands of compounds across chemical spaces from multiple pharmaceutical companies, with a special focus on the effect of lipophilicity and ionization state on the performance of the predictive approaches.
ACS Paragon Plus Environment
4
Page 5 of 28 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60
Molecular Pharmaceutics
EXPERIMENTAL SECTION Datasets and experimental details for determination of fu,inc The values of fu,inc were taken from each company’s in-house database. Both microsomal and hepatocyte data were considered as available. The values were measured according to standard procedures utilized at each company (see Table 1).
Table 1: Measurement details and number of data points in each company’s dataset Company
n
Matrix
Method reference
AstraZeneca
1122
human liver microsomes
Austin et al., 20023
(1 mg protein/mL) AstraZeneca
3639
rat hepatocytes (106 cells/mL),
Austin et al., 20058
metabolism precluded by inhibitor Genentech
166
human liver microsomes
See Supporting Information
(0.5 mg protein/mL) Lilly
8375
human liver microsomes
See Supporting Information
(0.5 mg protein/mL) MSD
2602
human liver microsomes
Austin et al., 20023
(0.25 mg protein/mL) Novartis
102
human liver microsomes
See Supporting Information
(0.5 mg protein/mL) Pfizer
3438
human liver microsomes
Zhang et al., 20106
(0.76 mg protein/mL) Vertex
649
rat hepatocytes
Austin et al., 20058
(0.5*106 cells/mL), dead
ACS Paragon Plus Environment
5
Molecular Pharmaceutics 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60
Page 6 of 28
Methods for predicting fu,inc Both established methods to predict fu,inc and in-house models based on a company’s proprietary data were used in this comparison.
Methods for predicting fu,inc in microsomes (fu,mic) Three different models were used in the prediction of fu,mic. The first model was developed by Austin et al.3 and is shown in Equation 1. fu,mic =
1
eq.1
0.56 × 𝑙𝑜𝑔𝑃/𝐷 ― 1.41
1 + Cp × 10
The second model by Hallifax and Houston13 is shown in Equation 2. fu,mic =
1
eq.2
0.072 × (𝑙𝑜𝑔𝑃/𝐷)2 + 0.067 × 𝑙𝑜𝑔𝑃/𝐷 ― 1.126
1 + Cp × 10
where Cp is the protein concentration in the incubation, 0.5 or 1mg/mL and logP/D a lipophilicity descriptor depending on ion class. For bases, and neutral compound, the descriptor refers to logP, the logarithm of the compound’s partition coefficient in octanol-water, whereas for acids it refers to logD, the logarithm of the compound’s distribution coefficient in octanolwater at pH 7.4. The third model for predicting fu,inc in microsomes, developed by Turner et al.12, 16 involves separate equations for bases, acids and neutral compounds as shown in Equations 3a-3c. 1
predominantly ionized bases: fu,mic =
1 + Cp × 100.58 × logP ― 2.02
predominantly ionized acids: fu,mic =
1 + Cp × 100.20 × logP ― 1.54
neutral compounds: fu,mic =
1
1 1 + Cp × 100.46 × logP ― 1.51
eq.3b eq.3b eq.3c
ACS Paragon Plus Environment
6
Page 7 of 28 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60
Molecular Pharmaceutics
Compounds were classified as predominately ionized acids and bases or neutrals based on the calculated pKa as described below.
Methods for predicting fu,inc in hepatocytes (fu,hep) Two lipophilicity-based models have been utilized for prediction of fu,hep. Equation 4 shows Austin et al.’s model for hepatocyte binding.8 1
fu,hep = 1 + 100.4 × logD/P ― 1.38
eq.4
The second model by Kilford et al.11 is described in Equation 5. fu,hep =
1 1 + 125 × V𝑅 × 100.072
× (logP/D)2 + 0.067 × logP/D – 1.126
eq.5
where VR is the ratio of the cell and incubation volumes; standard value is 0.005 for 106 cells/mL incubation.17 Estimating fu,hep simply from plasma protein binding, fu, and ion class was suggested by Page,4 see Equation 6a for bases and Equation 6b for other ion classes. bases:
fu,hep = 𝑓𝑢
eq.6a
non-bases: fu,hep = 𝑓𝑢0.5
eq.6b
In addition, microsomal binding data (fu,mic) were used to estimate fu,hep using three different methods. Austin et al.8 suggested Equation 7. 1
fu,hep = log
(
1 ― f𝑢,𝑚𝑖𝑐 f𝑢,𝑚𝑖𝑐
)
eq.7 ― 0.06
1.52
1 + 10
The second method by Kilford et al.11 is described by Equation 8. fu,hep =
1 Kp
1 + Ka ∗
V𝑅 Cp
∗
(
1 ― f𝑢,𝑚𝑖𝑐 f𝑢,𝑚𝑖𝑐
eq.8
)
ACS Paragon Plus Environment
7
Molecular Pharmaceutics 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60
Page 8 of 28
where Kp is the hepatocyte/medium concentration ratio and Ka is the microsomal protein binding affinity (Kp/Ka is considered to be 125). The simplest model assumes that fu,hep is equal to fu,mic, as is shown in Equation 9. fu,hep=fu,mic
eq.9
In-house models used to calculate fu,inc Internal machine-learning models were used when available at the specific company. 80% of the fu,inc data defined by a temporal split were used as training set, and the remaining 20% served as test set and were used in the comparison. Modeling details for each company can be found in the Supporting Information (Table S1).
Determination of calculated or experimental logP, logD, pKa, and ion class LogP, logD and pKa were calculated for all compounds (see Supporting Information, Table S2, for methods used at each company). Calculated pKa was used to define a compound’s ion class: a compound where the most basic functional group had a pKa above 7.4 was considered to be a base, a compound where the most acidic functional group had a pKa below 7.4 was considered to be an acid, compounds with both conditions were labeled as zwitterions, and the remaining compounds were considered to be neutral. No guidance exists regarding zwitterions in the lipophilicity-based models. We used logD as input for Equations 1, 2, 4 and 5, and utilized the equation for neutral compounds (eq. 3c) defined by Turner et al.12, 16 For a considerable number of compounds also experimental logD data was available (see Supplemental Table 2, for experimental details at each company). Experimental logD was
ACS Paragon Plus Environment
8
Page 9 of 28 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60
Molecular Pharmaceutics
converted to logP using calculated pKa values according to Equations 10 and 11, disregarding any potential ion pair extraction component. eq.10
𝑙𝑜𝑔𝑃 = 𝑙𝑜𝑔𝐷 ― log(𝑓𝑟_𝑛𝑒𝑢𝑡𝑟𝑎𝑙)
where fr_neutral is the fraction of neutral species at pH 7.4, calculated as shown in Equation 11. 𝑓𝑟_𝑛𝑒𝑢𝑡𝑟𝑎𝑙 =
(
1 1+
1 10 ―7.4 + 𝑝𝐾𝑎_𝑎𝑐𝑖𝑑
+
)
1 107.4 ― 𝑝𝐾𝑎_𝑏𝑎𝑠𝑒
eq.11
where pKa_acid describes the pKa for the most acidic functional group in a molecule and pKa_base describes the pKa for the most basic functional group. If no acidic group was present in a molecule, pKa_acid was set to 14. Similarily, pKa_base was set to 0, if no basic group was present in a molecule. For neutral compounds no conversion was necessary and logD was used directly. Only compounds with logD (logP) values between -6 and 8 were considered in the evaluation since predicted fu,inc values are extremely sensitive to high or low logD (logP) values. The numbers of compounds evaluated could, thus, be slightly different depending on the model. Experimental procedures for determination of plasma protein binding followed standard procedures at each company (see Supporting Information, Table S2, for details).
Evaluation procedure An Excel (Microsoft, WA) template was created to gather the data at each company and perform the calculations. An initial analysis was conducted at each company to identify the information needed for the overall analysis. For example, predictivity differences due to lipophilicity could be well described by distinguishing compounds with ClogP above or below 3 (see Supporting, Figure S1). The compounds were categorized according to their ion class, their
ACS Paragon Plus Environment
9
Molecular Pharmaceutics 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60
Page 10 of 28
calculated logP (≤3, >3), their experimental fu,inc value, normalized to 1 mg/mL protein content or 106 cells/mL, (≤0.2, >0.2 to 0.4, >0.4 to 0.6, >0.6 to 0.8, >0.8), and how well they were predicted by each model (≤2-fold, >2- to 3-fold, >3 to 5-fold, >5-fold). The categorized data from all companies were combined and the total set analyzed in Spotfire (Tibco Software, MA; http://spotfire.tibco.com/).
RESULTS Evaluation of prediction of fu,mic values Altogether, fu,mic data for more than 15,000 compounds were analyzed, with about 2,500 (16%) acids, 3,400 (22%) bases, more than 9,000 (58%) neutral compounds and about 700 (4%) zwitterions. For close to 30% of the dataset fu,mic normalized to 1 mg protein/mL was above 0.6. When considering only compounds with ClogP ≤ 3 (about 50% of the dataset) the amount of compounds with fu,mic above 0.6 increased to almost 50%, whereas less than 12% of the compounds with ClogP > 3 showed normalized fu,mic values above 0.6. More than 70% of the acids had a ClogP > 3, whereas the other ion classes had about 50% of the compounds with ClogP > 3. Detailed information about the physico-chemical properties of the individual company’s datasets can be found in supporting information (Table S3). All three of the lipophilicity-based models that were investigated predicted more than 70% of the compounds within 2-fold of the experimental fu,mic value (see Figure 1a). Use of the experimental lipophilicity values, available for about 50% of the compounds, enhanced the predictivity of fu,mic in all three methods only slightly. Predicting fu,mic from experimental plasma protein binding4 gave the least accurate results (44% of the compounds were predicted within 2-
ACS Paragon Plus Environment
10
Page 11 of 28 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60
Molecular Pharmaceutics
fold). The proprietary machine-learning models based on big internal data sets predicted 81% of the compounds within 2-fold.
ACS Paragon Plus Environment
11
Molecular Pharmaceutics 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60
Page 12 of 28
Figure 1: (a) Overall results for evaluation of fu,mic predictions and (b) by lipophilicity class (upper panel: ClogP ≤ 3; lower panel: ClogP > 3); color indicates fold-error: green: ≤2-fold, yellow: >2- to ≤3-fold, orange: >3 to ≤5-fold, red: >5-fold; the number of compounds evaluated is indicated for each bar. These trends were generally also found within the respective company datasets, although some differences as to which lipophilicity-based model predicted best could be noted (see Supporting Information, Figure S1 for details, and Table S4 for 6 example structures). Stratifying the analysis by calculated logP shows that approximately 90% of the low-lipophilic compounds (ClogP≤3) were predicted within 2-fold by almost all the models that were investigated (see Figure 1b). For the lipophilicity-based models, this value decreases to below 60% for compounds with ClogP greater than 3, and around 20% of the lipophilic compounds were predicted outside 5-fold. Internal models, on the other hand, predicted about 70% of the lipophilic compounds within 2-fold, with about 5% outside of 5-fold. Page’s approximation predicted the least percentage of compounds within 2-fold, both for lipophilic and less lipophilic compounds. Still, lipophilic compounds were less well predicted. More than 90% of the compounds with experimental, normalized fu,mic greater than 0.8 were predicted within 2-fold by the lipophilicity-based methods, whereas only about 20% (40% in Austin model) of the compounds with an fu,mic less than 0.2 were predicted with the same accuracy (Figure 2a and Supporting Information, Figure S3). While the same trend was observed for the internal machine-learning models, the proportion of compounds with an fu,mic less than 0.2 that were predicted within 2-fold was significantly higher, almost 60%. In contrast, Page’s estimation is least accurate for compounds with an experimental fu,mic between 0.2 and 0.6.
ACS Paragon Plus Environment
12
Page 13 of 28 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60
Molecular Pharmaceutics
Figure 2: (a) Evaluation of fu,mic predictions by fu,mic class, normalized for protein content, and (b) by ion class (upper left panel: Hallifax equation; upper right panel: Turner equation; lower left panel: correlation to protein binding as suggested by Page; lower right panel: internal
ACS Paragon Plus Environment
13
Molecular Pharmaceutics 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60
Page 14 of 28
model); color indicates fold-error: green: ≤2-fold, yellow: >2- to ≤3-fold, orange: >3 to ≤5-fold, red: >5-fold; the number of compounds evaluated is indicated for each bar. Stratifying by ion class did not give significant differences for most prediction methods, with usually more than 60% of the predictions within 2-fold for acids, bases, and zwitterions, and about 80% for neutral compounds (see Figure 2b and Supporting Information, Figure S4). However, while estimation of fu,mic from protein binding for neutral compounds and zwitterions shows about 55% predictions within 2-fold, only 35% of the bases and less than 20% of the acids are predicted as well.
Evaluation of prediction of fu,hep values Data on fu,hep were available from two companies, AstraZeneca and Vertex, and included values for more than 4,000 compounds (about 1,000 acids, 1,200 bases, 1,800 neutral compounds and 200 zwitterions). In this dataset, most acids were notably lipophilic (almost 90% of the acids had ClogP > 3 in both the AstraZeneca and the Vertex datasets), whereas for all other ion classes, the percentage of lipophilic compounds was approximately 50%. The Vertex dataset contained 44% acids, and 70% of the compounds had ClogP above 3, whereas the AstraZeneca dataset contained mostly neutral compounds (43%) and only 20% acids, with about 50% of the compounds having a ClogP of greater than 3 (see Supporting Information, Table S3).
ACS Paragon Plus Environment
14
Page 15 of 28 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60
Molecular Pharmaceutics
ACS Paragon Plus Environment
15
Molecular Pharmaceutics 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60
Page 16 of 28
Figure 3: (a) Overall results for evaluation of fu,hep predictions, (b) by ion class and (c) by fu,hep class, normalized for cell concentration, (b and c: upper left panel: Austin equation; upper right panel: correlation to protein binding as suggested by Page; lower left panel: direct estimation from fu,mic; lower right panel: internal model); color indicates fold-error: green: ≤2-fold, yellow: >2- to ≤3-fold, orange: >3 to ≤5-fold, red: >5-fold; the number of compounds evaluated is indicated for each bar. The lipophilicity-based models (Austin and Kilford) predicted around 60% and 65% of the experimental values within 2-fold when using calculated and measured lipophilicity, respectively (see Figure 3a). The use of experimental plasma protein binding data gave lower predictivity, with only 45% of the compounds being predicted within 2-fold. AstraZeneca’s internal machinelearning model predicted 80% of the test compounds in the company’s data set within 2-fold. Considering experimental fu,mic data more than 85% of the compounds were predicted within 2fold, both when corrections suggested by Austin and Kilford were applied, and when the fu,mic value was taken directly. Lipophilicity-based models showed less ability to predict the Vertex data set, whereas estimation from plasma protein binding gave similar results as for the AstraZeneca data set (see Supporting Information, Figure S5). Overall, the predictivity was lower for compounds with ClogP above 3, similar to what was seen for fu,mic predictions (see Supporting Information, Figure S6). Considering ion classes (see Figure 3b and Supporting Information, Figure S7), a lower accuracy for acids was noted, especially for predictions from lipophilicity-based models. When plasma protein binding was used, both acids and bases clearly had a lower predictivity, whereas when estimating fu,hep from microsomal binding, only a slight predictivity decrease for acids was
ACS Paragon Plus Environment
16
Page 17 of 28 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60
Molecular Pharmaceutics
perceived. The internal machine-learning model did not show predictivity differences across ion classes. Figure 3c (and Supporting Information, Figure S8) shows the results stratified by fu,hep class normalized to 1*106 cells/mL: lipophilicity-based models show poor predictivity for highly bound compounds, whereas the differences for other models are less pronounced. Basing fu,hep predictions on plasma protein binding again has the lowest predictivity for medium bound compounds, whereas about 55% of the compounds are predicted within 2-fold when fu,hep is above 0.8 or below 0.2.
ACS Paragon Plus Environment
17
Molecular Pharmaceutics 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60
Page 18 of 28
DISCUSSION Datasets from six companies were available for the evaluation of fu,mic predictions. The size of these datasets varied. Nevertheless, the overall properties of the datasets with regard to ion class and lipophilicity were fairly similar, even though they likely covered different regions in chemical space. Overall trends were that the lipophilicity-based models3,
12-13
could predict most of the
compounds with a ClogP less than 3 within 2-fold of the experimental value, whereas lipophilic compounds were quite often predicted with an error above 5-fold. Additionally, the use of experimental logD values enhanced the overall predictivity only slightly. The method that relies on experimental plasma protein binding,4 originally suggested to account for binding in scaling from hepatocyte CLint data, was the least accurate of those evaluated. Still, compounds with ClogP below 3 were predicted better than those with higher lipophilicity. For this method, the availability of data was limited compared to the other equations analyzed. Actually, simply assuming fu,mic to be 0.7 will give a reasonable estimate for compounds with ClogP3) the literature models are less accurate. Correcting measured CLint using these calculations will, in most cases, provide a “best case scenario” for the unbound intrinsic clearance, which is valuable information during compound triaging.
ACS Paragon Plus Environment
22
Page 23 of 28 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60
Molecular Pharmaceutics
ASSOCIATED CONTENT Supporting Information. Specific methods for fu,inc measurements, model details for the internal fu,inc models, calculation methods for logP, logD and pKa calculations and experimental details for measurements of logD and plasma protein binding, and physico-chemical characterization of the internal datasets, are given as supporting information. Additionally, the structure and data for six example compounds as well as detailed evaluation plots are given. This material is available free of charge at http://pubs.acs.org.
AUTHOR INFORMATION Corresponding Author *e-mail:
[email protected] Author Contributions The manuscript was written through contributions of all authors. All authors have given approval to the final version of the manuscript.
ACS Paragon Plus Environment
23
Molecular Pharmaceutics 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60
Page 24 of 28
ACKNOWLEDGMENT We would like to acknowledge the experimental teams in each company for generating the data for our analysis. Special thanks to Michael Mohutsky (Eli Lilly), Zhengyin Yan and Ivy Chen (Genentech), Kajsa Kanebratt (AstraZeneca), Gaelle Chenal and Barun Bhhatarai (Novartis) for assisting this work with data retrieval and data interpretation at the respective company. This article was developed with the support of the International Consortium for Innovation and Quality in Pharmaceutical Development (IQ, www.iqconsortium.org). IQ is a not-for-profit organization of pharmaceutical and biotechnology companies with a mission of advancing science and technology to augment the capability of member companies to develop transformational solutions that benefit patients, regulators and the broader research and development community.
ABBREVIATIONS ADME, absorption, distribution, metabolism and excretion; CLint, intrinsic clearance; fu,inc, fu,mic, fu,hep, unbound fraction in the incubation, in microsomal incubation, and in hepatocyte incubation, respectively; fu, fraction unbound in plasma; hep, hepatocytes; mic, microsomes; PK, pharmacokinetic.
ACS Paragon Plus Environment
24
Page 25 of 28 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60
Molecular Pharmaceutics
REFERENCES 1.
Poulin, P.; Haddad, S., Microsome composition-based model as a mechanistic tool to
predict nonspecific binding of drugs in liver microsomes. J Pharm Sci 2011, 100 (10), 4501-17. 2.
Sohlenius-Sternbeck, A.-K.; Afzelius, L.; Prusis, P.; Neelissen, J.; Hogstraate, J.;
Johansson, J.; Floby, E.; Bengtsson, A.; Gissberg, O.; Sternbeck, J.; Petersson, C., Evaluation of the human prediction of clearance from hepatocytes and microsome intrinsic clearance for 52 drug compounds. Xenobiotica 2010, 40 (9), 637-649. 3.
Austin, R. P.; Barton, P.; Cockroft, S. L.; Wenlock, M. C.; Riley, R. J., The Influence of
Nonspecific Microsomal Binding on Apparent Intrinsic Clearance, and its Prediction from Physicochemical Properties. Drug Metab Dispos 2002, 30 (12), 1497–1503. 4.
Page, K. M., Validation of Early Human Dose Prediction: A Key Metric for Compound
Progression in Drug Discovery. Mol Pharm 2016, 13 (2), 609-620. 5.
Wan, H.; Bold, P.; Larsson, L.-O.; Ulander, J.; Peters, S.; Löfberg, B.; Ungell, A.-L.;
Någård, M.; Llinàs, A., Impact of Input Parameters on the Prediction of Hepatic Plasma Clearance Using the Well-Stirred Model. Curr Drug Metab 2010, 11 (7), 583-594. 6.
Zhang, Y.; Yao, L.; Lin, J.; Gao, H.; Wilson, T. C.; Giragossian, C., Lack of appreciable
species differences in nonspecific microsomal binding. J Pharm Sci 2010, 99, 3620−3627. 7.
Obach, R. S., The Importance of Nonspecific Binding in In Vitro Matrices, Its Impact on
Enzyme Kinetic Studies of Drug Metabolism Reactions, and Implications for In Vitro-In Vivo Correlations. Drug Metab Dispos 1996, 24 (10), 1047-1049.
ACS Paragon Plus Environment
25
Molecular Pharmaceutics 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60
8.
Page 26 of 28
Austin, R. P.; Barton, P.; Mohmed, S.; Riley, R. J., The binding of drugs to hepatocytes
and its relationship to physicochemical properties. Drug Metab Dispos 2005, 33 (3), 419-25. 9.
Gao, H.; Yao, L.; Mathieu, H. W.; Zhang, Y.; Maurer, T. S.; Troutman, M. D.; Scott, D.
O.; Ruggeri, R. B.; Lin, J., In silico modeling of nonspecific binding to human liver microsomes. Drug Metab Dispos 2008, 36 (10), 2130-5. 10. Chen, S.; Prieto Garcia, L.; Bergstrom, F.; Nordell, P.; Grime, K., Intrinsic Clearance Assay Incubational Binding: A Method Comparison. Drug Metab Dispos 2017, 45 (4), 342-345. 11. Kilford, P. J.; Gertz, M.; Houston, J. B.; Galetin, A., Hepatocellular binding of drugs: correction for unbound fraction in hepatocyte incubations using microsomal binding or drug lipophilicity data. Drug Metab Dispos 2008, 36 (7), 1194-7. 12. Turner, D. B.; Rostami-Hodjegan, A.; Tucker, G. T.; Yeo, K. R., Prediction of NonSpecific Hepatic Microsomal Binding from Readily Available Physicochemical Properties. Drug Metab Rev 2006, 38 (S1), 162. 13. Hallifax, D.; Houston, J. B., Binding of drugs to hepatic microsomes: comment and assessment of current prediction methodology with recommendation for improvement. Drug Metab Dispos 2006, 34 (4), 724-6. 14. Nair, P. C.; McKinnon, R. A.; Miners, J. O., A Fragment-Based Approach for the Computational Prediction of the Nonspecific Binding of Drugs to Hepatic Microsomes. Drug Metab Dispos 2016, 44 (11), 1794-1798.
ACS Paragon Plus Environment
26
Page 27 of 28 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60
Molecular Pharmaceutics
15. Gertz, M.; Kilford, P. J.; Houston, J. B.; Galetin, A., Drug lipophilicity and microsomal protein concentration as determinants in the prediction of the fraction unbound in microsomal incubations. Drug Metab Dispos 2008, 36 (3), 535-42. 16. Turner, D. B.; Rostami-Hodjegan, A.; Tucker, G. T.; Yeo, K. R. Prediction of NonSpecific Hepatic Microsomal Binding from Readily Available Physicochemical Properties Available
Physicochemical
Properties.
https://www.certara.com/wp-
content/uploads/Resources/Posters/DavidISSX2006.pdf (accessed 2019-04-24). 17. Brown, H. S.; Griffin, M.; Houston, J. B., Evaluation of Cryopreserved Human Hepatocytes as an Alternative in Vitro System to Microsomes for the Prediction of Metabolic Clearance. Drug Metab Dispos 2006, 35 (2), 293-301.
ACS Paragon Plus Environment
27
Molecular Pharmaceutics 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60
275x174mm (150 x 150 DPI)
ACS Paragon Plus Environment
Page 28 of 28