Perspective Cite This: J. Med. Chem. 2017, 60, 9097-9113
pubs.acs.org/jmc
In Silico Absorption, Distribution, Metabolism, Excretion, and Pharmacokinetics (ADME-PK): Utility and Best Practices. An Industry Perspective from the International Consortium for Innovation through Quality in Pharmaceutical Development Miniperspective Franco Lombardo,† Prashant V. Desai,‡ Rieko Arimoto,§ Kelly E. Desino,∥ Holger Fischer,⊥ Christopher E. Keefer,# Carl Petersson,∇ Susanne Winiwarter,○ and Fabio Broccatelli*,◆ †
Alkermes Inc., 852 Winter Street, Waltham, Massachusetts 02451, United States Computational ADME, Drug Disposition, Eli Lilly and Company, Indianapolis, Indiana 46285, United States § Vertex Pharmaceuticals Inc., 50 Northern Avenue, Boston, Massachusetts 02210, United States ∥ AbbVie, Inc., North Chicago, Illinois 60064, United States ⊥ Roche Pharmaceutical Research and Early Development, Pharmaceutical Sciences, Innovation Center Basel, F. Hoffmann-La Roche Ltd., 4070 Basel, Switzerland # Pfizer Inc., Groton, Connecticut 06340, United States ∇ Discovery Drug Disposition, Biopharma, R&D Global Early Development, EMD Serono, Frankfurter Strasse 250 I Postcode D39/001, 64293 Darmstadt, Germany ○ Drug Safety and Metabolism, AstraZeneca R&D Gothenburg, 431 83 Mölndal, Sweden ◆ Genentech Inc., South San Francisco, California 94080, United States ‡
ABSTRACT: In silico tools to investigate absorption, distribution, metabolism, excretion, and pharmacokinetics (ADME-PK) properties of new chemical entities are an integral part of the current industrial drug discovery paradigm. While many companies are active in the field, scientists engaged in this area do not necessarily share the same background and have limited resources when seeking guidance on how to initiate and maintain an in silico ADME-PK infrastructure in an industrial setting. This work summarizes the views of a group of industrial in silico and experimental ADME scientists, participating in the In Silico ADME Working Group, a subgroup of the International Consortium for Innovation through Quality in Pharmaceutical Development (IQ) Drug Metabolism Leadership Group. This overview on the benefits, caveats, and impact of in silico ADME-PK should serve as a resource for medicinal chemists, computational chemists, and DMPK scientists working in drug design to increase their knowledge in the area.
1. INTRODUCTION
acronym ADME-PK implies, in vivo properties can also be the target of such efforts. Appropriately applied during the discovery phase, in silico ADME-PK has the potential to enhance the probability of identifying compounds with favorable ADME properties. Importantly, these in silico tools are not always intended to be in lieu of experimental measurements but are intricately connected with the preclinical assay they attempt to predict and are ideally applied by chemists and DMPK scientists alike, presynthetically to guide the next round of synthesis. The intended result of such effort is to accelerate the drug optimization process by reducing the number of cycles of
In silico ADME-PK (absorption, distribution, metabolism, excretion, and pharmacokinetics) is the use of computer modeling to understand structure−property relationships and to predict DMPK (drug metabolism and pharmacokinetics) properties from compound structure. This is related but distinct from physiologically based pharmacokinetic (PBPK) modeling, which strives to provide accurate predictions of the PK profile of drug candidates. The focus of in silico ADME-PK is to guide the design of novel compounds with superior ADME properties. Most often a QSPR (quantitative structure− property relationship) approach is used to relate a compound’s structure to the property in question (e.g., cell permeability or metabolic clearance) measured in an in vitro assay. As the © 2017 American Chemical Society
Received: March 28, 2017 Published: June 13, 2017 9097
DOI: 10.1021/acs.jmedchem.7b00487 J. Med. Chem. 2017, 60, 9097−9113
Journal of Medicinal Chemistry
Perspective
Figure 1. Changes observed across all projects in microsomal stability at Genentech (GNE), solubility at AstraZeneca (AZ), and CYP3A4 timedependent inhibition at Eli Lilly (Lilly) after adoption of in silico models.
implications in model building, evaluation, and use. The review further suggests individual components necessary to implement a successful in silico ADME-PK modeling culture and infrastructure in industry. Complementary discussions focusing on the technical aspects of building an ADME QSAR infrastructure can be found elsewhere.7 While the work described focuses mainly on QSPR modeling, many of the learning points are transferrable to other technologies such as molecular matched pairs (MMPs) and chemical data mining, which will be discussed briefly before the conclusion of the manuscript. The technologies discussed in this work require availability of hundreds of consistently measured data points for each property target of the modeling effort; this could be a practical limitation for scientists working in small organizations approaching this field.
synthesis necessary to identify a quality drug candidate. Additionally in silico ADME tools can be used to prioritize in vitro testing after synthesis. Computational ADME modeling is quite a mature but still evolving field.1−4 In silico ADME tools are routinely applied to drug design. However, scientists engaged in this arena often witness how the use of the in silico ADME tools is largely dependent on the perception of the individuals within the project teams. This skepticism is often the result of lack of clarity on the scope and limitations of the model and of the assay being modeled. For example, one cannot expect a reliable quantitative model of thermodynamic aqueous solubility when we are not aware of the solid form of the material used in generating the solubility data.5 Models, by definition, will provide numbers, and numbers can be misinterpreted if used out of context. This is also true of experimental values, which have associated variability and experimental error and need to be interpreted with caution, especially for some high-throughput screens. In a recent perspective article, Kuhn and coauthors6 highlight the value of qualitative assessments in their closing notes. These authors invite us to not use absolute values but rather consider a more qualitative approach within the context of the project. They also suggest to “stay close to the experiments”. This should be interpreted, in our view, as a coupling of the use of computed physicochemical, in vitro, and in vivo readouts with a solid understanding of the experimental assay and its limitations. For example, cLogP can easily be calculated by one of several commercially available software packages. The validity of this calculated property for a given chemotype of interest should be validated by submitting a sufficient number of compounds for experimental log P/log D determinations. In other words, “trust but check” to ensure the models are applicable to the structural space of interest; this is a critical consideration that will be further discussed later. This review discusses key lessons learned by scientists actively working in the in silico ADME-PK field within drug discovery. The present impact will be discussed as well as the importance of paying attention to the purpose and quality of the experimental assay being modeled and the resulting
2. IMPACT OF IN SILICO ADME-PK MODELING The aim of in silico ADME-PK modeling is to improve the design of candidate drugs. Experimental end points that are most commonly targeted include metabolic stability, permeability, solubility, CYP inhibition, and transporter-mediated efflux. For these properties most members of our group reported using a “global” model, i.e., including all the available internal data in the training set, as opposed to a “local” model focusing on a specific chemotype. While good models are needed, the mere existence of high quality models does not guarantee success. Models can only be truly impactful if the organization embraces and fosters an in silico culture; this requires a diversified approach involving in silico scientists, DMPK scientists (in vitro and in vivo), chemists, and to some extent management. Figure 1 reports real life examples where models influenced the quality of compound design organization-wide.8,9 An in silico solubility model at AstraZeneca and an in silico microsomal stability model at Genentech had been available for years. Nevertheless, true impact could only be seen once usage was enforced through departmental goals in 2004 at AstraZeneca and in 2010 at Genentech. This can be done by creating the expectation that projects facing a ADME challenges should factor in silico ADME models into decision making or 9098
DOI: 10.1021/acs.jmedchem.7b00487 J. Med. Chem. 2017, 60, 9097−9113
Journal of Medicinal Chemistry
Perspective
Figure 2. Results of the survey among industrial scientists involved in in silico ADME.
understanding decision drivers during design. In companies adopting a tiered screening strategy, where only compounds that are stable in microsomes (tier 1) will be tested in hepatocytes (tier 2), a good in silico filter for bad actors in the liver microsomes assay can effectively act as a tier 0 assay. In contrast, a model for stability in hepatocytes may be less useful as guidance for the tier 2 assay, since in vitro metabolic stability would already be available during candidate selection for the hepatocytes stability assay. The opposite could be true in a different paradigm where metabolic stability in hepatocytes is used as a tier 1 assay, for example, in a project where phase II metabolism is recognized as important for the metabolic clearance. In essence, the success of the in silico tools is inherently linked to the local company environment. Scientists at AstraZeneca reported that their in silico ADMETox modeling infrastructure was utilized to predict at least one property during the design stage for ∼55% of the compounds synthesized between 2011 and 2012.8 While this does not inform on the quality of the predictions, it is a clear indication of their importance in decision-making. Another example of the impact of in silico ADME models was described by scientists at Eli Lilly and Company, based on an integrated approach involving an iterative learning cycle through in silico−in vitro−in vivo assessments.10 In an effort to provide an estimate of the impact of in silico ADME-PK modeling in the industry today, 11 members of the In Silico ADME IQ Working Group answered a survey (Figure 2). The scientists who responded to this survey are industry experts working in some of the major pharmaceutical companies in U.S., Europe, and Asia. The findings of the survey are summarized as follows:
provide explanations if they do not. At AstraZeneca this resulted in a 7-fold increase in the number of compounds with good solubility, whereas at Genentech the number of stable compounds increased by 2-fold. The process was slightly different for the third example in Figure 1: a high-throughput assay for assessing time-dependent inhibition (TDI) of CYP3A4 was made available at Eli Lilly in mid-2009. After proactive enrichment of the TDI data to capture a wide range of structural diversity, a global in silico model was introduced in 2011. Subsequently the ratio of compounds with low vs high potential for CYP3A4 TDI improved by 3-fold (1.4 before 2011 and 4.3 after 2011). While high quality of a model is an absolute prerequisite and management may help a model gain momentum, an equally important aspect is continued education and support promoted by a core team composed of the following: • In silico ADME-PK experts with the capability to develop robust models (usually computational chemists or cheminformaticians with a keen interest in understanding ADME properties) • In silico and in vitro (or in vivo) experts monitoring model performance over time as well as experimental data quality • IT and infrastructure experts ensuring model access and ease of use • Early adopters (or “champions”) of the models positioned within the project team that can lead by example • Educators creating awareness by presenting in different forums • Management encouragement of the use of the models Furthermore, usage and thus the impact of a model depend on the quality of the software infrastructure as well as the ability to promote the model’s interplay and synergy with available in vitro tools. The latter point is very important and is tied to
• In silico ADME is perceived as an impactful reality in industry; yet, approximately 3 out of 4 of the members think that there is still room for growth. 9099
DOI: 10.1021/acs.jmedchem.7b00487 J. Med. Chem. 2017, 60, 9097−9113
Journal of Medicinal Chemistry
Perspective
• QSPR predictors are the most established tools. Complementary approaches focused on interpretation and idea generation, such as MMPs and data trend analysis, are also widely adopted. • Structure-based modeling of ADME targets (e.g., CYPs, P-gp, PXR) can be useful but should be considered as a tool to study specific cases. It is important to point out that in silico predictions of metabolic soft spots and computer aided metabolite identification may be regarded as components of the in silico ADME toolbox but will not be reviewed here. This is an extremely active and impactful area at the interface between in silico and in vitro science but is beyond the scope of this work.
(MDCK-MDR1) efflux ratio models are of high interest in industry, since it was shown that the experimental value is a good predictor for poor in vivo brain penetration.12 A project targeting the CNS needs to understand the efflux risk at the stage of design. Table 1 shows the example of an MDCKMDR1 efflux ratio model used for categorical predictions at Genentech. Another example is presented in Figure 3 depicting prospective predictions for the AbbVie MDCK-MDR1 efflux
3. MODEL BUILDING CONSIDERATIONS The necessary starting point for all good and useful models is an in-depth understanding of the property that is being modeled, how that property factors into decision making, and how it is related to the in vivo situation. In addition, model usage often depends on how the experimental data are used, which will also influence the approach. A categorical filter for bad actors will not necessarily need to capture 80% of the experimental data within 2-fold from the prediction, while a continuous model with increased accuracy may be preferred for uses such as establishing an in silico to in vitro correlation around clearance. Some experimental end points lend themselves to being easily modeled as in the case of fraction unbound in human plasma (h-fup) (Table 1). This parameter is experimentally Figure 3. Prospective predictions for the AbbVie MDCK-MDR1 efflux ratio model. The compounds are color coded by prediction confidence (purple, high; blue, medium; green, low confidence).
Table 1. Summary of in Vitro Variability and in Silico Error in Prediction Based on Genentech Data property
R2
rmse
within 2-fold
count
h-fup, exptla h-fup, calcdb efflux ratioc in MDCK-MDR1, exptla efflux ratioc in MDCK-MDR1, calcdb
0.92 0.64 0.59 0.22
0.05 0.08 17 26
0.91 0.52 0.77 0.46
1240 239 677 138
ratio model, which can be confidently used as a filter for compounds with high efflux ratios and thus poor chances of permeating the blood−brain barrier. The model is actually based on two different models using orthogonal descriptor types and is thereby able to estimate the prediction confidence for individual compounds; medium to high confidence predictions (blue and purple dots) rarely appear in the topleft quadrant of the plot; that is, false positives are rarely encountered for predictions associated with good confidence. In another example authors from Eli Lilly have successfully incorporated a categorical model for P-gp efflux to prioritize compounds for design, synthesis, and testing.10 In this case, given the potential limitations in interpreting numerical values of efflux ratio (see section 3.1 for details), the authors chose to build a categorical model wherein the P-gp efflux category determined in a similar assay (substrates or nonsubstrates) was utilized to train the model. The model was applied prospectively to predict P-gp efflux category along with “confidence” scores associated with each prediction. It was demonstrated that the use of confidence scores improved the accuracy of prediction from 73% to 81%. 3.1. Experimental Data. All modeling efforts require the availability of experimental values, usually measured at one or several occasions under specific, mostly defined conditions. Such data show variability and inherent errors. For example efflux ratio measurements in the MDCK-MDR1 cell line are aimed at understanding whether a compound is substrate of the key efflux transporter P-gp. However, interaction with P-gp may actually be underestimated for compounds with low permeability, which some define as
