Progress in predicting human ADME parameters in silico Sean Ekinsa,*, Chris L. Wallerb, Peter W. Swaanc, Gabriele Crucianid, Steven A. Wrightona, James H. Wikela aLilly Research Laboratories, Eli Lilly and Company, Lilly Corporate Center, Drop Code 0730, Indianapolis, IN 46285, USA bSphinx Pharmaceuticals, A Division of Eli Lilly and Company, 20 T.W. Alexander Drive, Research Triangle Park, NC 27709, USA cDivision of Pharmaceutics and Ohio State Biophysics Program, The Ohio State University, 500 West 12th Avenue, Columbus, OH 43210-1291, USA dLaboratory for Chemometrics, Via Elce di Sotto, 10-1-06123, Perugia, Italy Received 22 September 2000 accepted 25 September 2000 Abstract Understanding the development of a scientific approach is a valuable exercise in gauging the potential directions the process could take in the future. The relatively short history of applying computational methods to absorption, distribution, metabolism and excretion (ADME) can be split into defined periods. The first began in the 1960s and continued through the 1970s with the work of Corwin Hansch et al. Their models utilized small sets of in vivo ADME data. The second era from the 1980s through 1990s witnessed the widespread incorporation of in vitro approaches as surrogates of in vivo ADME studies. These approaches fostered the initiation and increase in interpretable computational ADME models available in the literature. The third era is the present were there are many literature data sets derived from in vitro data for absorption, drug���drug interactions (DDI), drug transporters and efflux pumps [P-glycoprotein (P-gp), MRP], intrinsic clearance and brain penetration, which can theoretically be used to predict the situation in vivo in humans. Combinatorial synthesis, high throughput screening and computational approaches have emerged as a result of continual pressure on pharmaceutical companies to accelerate drug discovery while decreasing drug development costs. The goal has become to reduce the drop-out rate of drug candidates in the latter, most expensive stages of drug development. This is accomplished by increasing the failure rate of candidate compounds in the preclinical stages and increasing the speed of nomination of likely clinical candidates. The industry now understands the reasons for clinical failure other than efficacy are mainly related to pharmacokinetics and toxicity. The late 1990s saw significant company investment in ADME and drug safety departments to assess properties such as metabolic stability, cytochrome P-450 inhibition, absorption and genotoxicity earlier in the drug discovery paradigm. The next logical step in this process is the evaluation of higher throughput data to determine if computational (in silico) models can be constructed and validated from it. Such models would allow an exponential increase in the number of compounds screened virtually for ADME parameters. A number of researchers have started to utilize in silico, in vitro and in vivo approaches in parallel to address intestinal permeability and cytochrome P-450-mediated DDI. This review will assess how computational approaches for ADME parameters have evolved and how they are likely to progress. D 2001 Elsevier Science Inc. All rights reserved. Keywords: QSAR Catalyst ADME 1. Introduction The past two decades have witnessed the increasing application of in vitro models in both industry and acade- mia by groups involved in the study of absorption, distribution, metabolism and excretion (ADME) (see in vitro metabolism review in this journal by Ekins et al.). One reason for this includes the recognition that undesir- able ADME properties of a new drug entity uncovered in the clinical phase of drug development represent a sig- nificant proportion of failures in drug development. One such ADME property historically discovered late in deve- lopment is clinically significant drug���drug interactions (DDI). If the DDI resulted in severe adverse effects, it may prevent the continuation of development and thus be costly in terms of the financial investment in a particular project. Therefore, it is important to screen for potential interactions (Wrighton & Silber, 1996) early on in the discovery/development process. Since the number of mole- cules synthesized by pharmaceutical companies has dra- matically increased with the utilization of combinatorial chemistry, there is now a shift in emphasis towards earlier * Corresponding author. Tel.: +1-317-433-5387 fax: +1-317-433- 0311. E-mail address: ekins_sean@lilly.com (S. Ekins). Journal of Pharmacological and Toxicological Methods 44 (2000) 251���272 1056-8719/00/$ ��� see front matter D 2001 Elsevier Science Inc. All rights reserved. PII: S1056-8719(00)00109-X

implementation of higher throughput in vitro metabolism and toxicology studies (Wrighton & Silber, 1996). Many of the in vitro, automation and bioanalytical technologies are revolutionizing sample throughput for drug metabolism and toxicology measurements. The pressure is now to screen not only earlier and faster but also smarter to avoid failure (or to fail as quickly as possible in the process). Such an approach will ultimately lead to more rational approaches to screening ADME properties in vitro and in vivo, and it is likely this will be facilitated by computa- tional filters. Validated and generally applicable computa- tional filters will avoid the need to screen or even synthesize every molecule for a given program and will allow simultaneous optimization of efficacy with other properties such as those related to ADME. By incorporat- ing these filters into combinatorial library design/screening, molecules with optimal ADME properties and structural diversity may be readily selected (Darvas, Dorman, & Papp, 2000 Ekins, Ring, Bravi, Wikel, & Wrighton, 2000). Thus, we appear to be at a turning point where computational ADME approaches can prospectively assist in the selection of more desirable drug candidates and enable drug discovery to reach its goals in terms of cost, speed and efficiency. In order to anticipate how this field may advance, it is worthwhile to understand how it has been shaped. Compu- tational technologies were initially successfully applied to modeling potency parameters and are beginning to be incorporated into ADME. Only recently have ADME departments demonstrated influence across discovery groups and it may be suggested that the development of in vitro approaches such as their ability to predict DDI in vitro has assisted in this development. By bringing ADME and computational chemistry techniques together, numerous ADME models are now in the data collection and model building stages. Computational ADME models can be thought of as having evolved through distinct phases that will be described in the following review. The appreciation of this development may ultimately influence the direction taken in their application in the future and their impact on the drug discovery process. 2. The first phase: computational models for ADME (1960s���1980) The first phase of ADME computational models began in the 1960s with classical QSAR methods developed by Hansch (1972). The simplicity of Hansch analysis was perhaps initially overlooked by ADME departments to some extent. This is not surprising as the ADME discipline within the pharmaceutical arena is generally not thought of as an early adopter of this technology and now is beginning to appreciate the potential of this technology after 30 years of progress. In contrast, toxicologists recognized the impor- tance of the partition coefficient (log P) as a potential indicator of toxicity and its (limited) relationship with absorption very early on. The work of Hansch et al. introduced the use of the octanol���water log P to a wider audience. Eventually, this parameter was later found to be useful in generating QSAR models for CYP data. As early as 1972, Hansch already suggested that quantitative rela- tionships could be generated for the lipophilic character of drugs and metabolic parameters such as microsomal hydro- xylation, demethylation, NADPH disappearance, CYP450��� CYP420 conversion and duration of drug action (Hansch, 1972). Many of the data sets used in this study were of limited size with less than 20 molecules that were conge- neric series. This immediately emphasizes the limitation of Hansch analysis applied to CYPs in that it works best with structurally homologous series whereas most individual CYPs are now known to metabolize many different classes of structurally diverse molecules (Gao & Hansch, 1996). This was acknowledged by the authors in a follow up to the review from 21 years earlier (Hansch, 1972), in which they discussed the ��� up to that point ��� limited application and acceptance of QSAR in understanding CYP binding, induc- tion and metabolism (Gao & Hansch, 1996). Furthermore, it was suggested that the lack of a unified approach to structure���activity relationships resulted as the research to date with multiple in vitro and in vivo systems had only provided ������glimpses of structure activity relationships������ (Hansch & Zhang, 1993). Undoubtedly, Hansch was ahead of his time if not a visionary when he stated, ������... the modification of organic compounds by the microsomal enzymes can be understood in terms of physiochemical constants in a quantitative fashion������ (Hansch, 1972). The success of ADME QSAR in the future will owe much to the initial efforts in influencing this discipline by Hansch and colleagues. 3. The intermediate phase: computational models for ADME (1980���2000) The fact that ADME is now embracing computational approaches could be due to a combination of improve- ments in technology as well as a change in mindset. It could also represent a shift from direct Hansch-type analysis to more graphical 3D (pharmacophore/toxico- phore-oriented) approaches. This change began during the 1980s and early 1990s as increased development of computational hardware enabled rapid computation and graphical representation (Evans and Sutherland, Cray, Silicon Graphics). Consequently, molecular modeling soft- ware was developed to take advantage of this new hard- ware [comparative molecular field analysis (CoMFA), Catalyst, as described later], such that 3D visualization became an important direction for QSAR. In parallel, the crystallization of nonmammalian soluble CYP enzymes, starting with P450cam, facilitated homology modeling of CYP active sites. These crystallized enzymes were then S. Ekins et al. / Journal of Pharmacological and Toxicological Methods 44 (2000) 251���272 252

used as templates for modeling the membrane bound human CYPs, the applications and limitations of which have been reviewed in detail (Szklarz & Halpert, 1998). It was not until the arrival of CoMFA and similar techniques that scientists had additional exciting tools to interpret and understand enzyme active sites and receptors in cases where no crystal structure was available. Although CoMFA and related techniques were initially not applied to CYPs, at the same time CYP DDI data prompted modeling the active sites of these enzymes (Fuhr et al., 1993 Strobl, von Kruedener, Stockigt, Guengerich, & Wolff, 1993). This period is perhaps critical in bridging the work of Hansch et al. and the increasing availability of in vitro data generated by studying DDI and the increased acceptance of other in vitro models in general. This shift towards computational ADME approaches stimulated additional classical QSAR studies (Lewis, 1997), substrate template models (DeGroot, Ackland, Horne, Alex, & Jones, 1999 DeGroot, Bijloo, Martens, Van Acker, & Vermeulen, 1997 DeGroot, Bijloo, Van Acker, et al., 1997 Jones, Hawks- worth, et al., 1996) as well as CYP 3D-QSAR and 4D- QSAR modeling later in the 1990s (Ekins et al., 1999a, 1999b Ekins, Bravi, Ring, et al., 1999 Ekins, Bravi, Wikel, & Wrighton, 1999 Ekins, Bravi, et al., 2000 Ekins, Ring, et al., 2000 He, Korzekwa, Jones, Rettie, & Trager, 1999 Jones, He, Trager, & Rettie, 1996 Poso, Juvonen, & Gynther, 1995 Rao et al., 2000) to define the ligand features necessary for interaction with these enzymes (Table 1). During this period, a large number of membrane perme- ability and intestinal absorption models were generated based on Caco-2 data (Table 2) and Lipinski���s ���rule of 5��� was defined as a method to identify molecules with advan- tageous solubility and permeability properties that would lead to high bioavailability (Lipinski, Lombardo, Dominy, & Feeney, 1997). Such models and general rules have had considerable impact as they have helped identify ADME problems solvable early in discovery. Ultimately, ���the rule of 5��� may be too simplistic, however, it has stimulated a great deal of interest in fast, generally applicable filters for ADME. As with any rule, it is meant to be broken, as clearly exemplified by its overt rejection of substrates for solute transporters that generally exert high oral bioavail- ability, and Lipinski et al. (1997) would likely suggest it as a ������rule of thumb������ or guide rather than a definitive cutoff. This rule has been judged by some as having a high false- positive level (George, 1999), yet it is still apparently widely used in the pharmaceutical industry. 4. In silico techniques developed in the 1980s���2000. The following sections describe several representative computational concepts, approaches and applications that have been developed in the 1980s through 1990s and have been of particular interest to ADME scientists. 4.1. Physiologically based pharmacokinetic (PB-PK) models There is a large body of work in the literature on the theory, development and utilization of PB-PK models (mul- ticompartment simulation models) for the prediction and characterization of xenobiotic absorption, distribution, metabolism and elimination (Ploemen et al., 1997). Dynamic models of this type when coupled with more traditional quantitative structure���property relationships (QSPR) and QSAR models for the estimation of tissue:- blood partition coefficients, metabolic rate constants and putative biological outcomes are being developed and implemented primarily in the environmental contaminant/ protection arena. The resource intensive nature of these types of models has precluded their widespread acceptance in the pharmaceutical industry and has been reviewed else- where (Clewell & Andersen, 1996). 4.2. 3D-QSAR Structure���activity methods that identify atomic spatial arrangements related to a measured endpoint are collectively termed three-dimensional QSAR (3D-QSAR) methods. These methods attempt to identify spatial regions for com- plementary molecular properties that are indicative of given biological measurements (Green & Marshall, 1995 Mar- shall & Cramer, 1988) and many reviews and monographs exist, which cover this approach in some detail (Kubinyi, 1997 and references therein). These procedures extend the QSAR approach in three dimensions by selecting either manually (Cramer, Patterson, & Bunce, 1988) or in an automated fashion (Jain, Koile, & Chapman, 1994), a particular geometry for each modeled compound. The molecular scaffold (Cramer et al., 1988), the pharmacophore (Van Drie, Weininger, & Martin, 1989) and/or the molecular field (Kearsley & Smith, 1990) is then used for super- imposition of the molecules and the basis for comparison for each compound. The underlying assumptions of 3D-QSAR methods are as follows: (i) the modeled compound, and not its metabo- lites, produce the effect (ii) the proposed or modeled conformation is the bioactive one (iii) the binding site and/or mode is the same for all modeled compounds (iv) the biological activity is largely explained by enthalpic processes (v) entropic terms are similar for all compounds (vi) the system is considered to be at equilibrium, and kinetic aspects are usually not considered (vii) solvent effects, diffusion, transport, etc., are not included. For any QSAR paradigm to be applied, at least 20 compounds (depending on the software and various references, preferably more) have to be individually characterized (i.e., synthesized and tested for biological response). These compounds constitute the ������training set������ from which the model is derived, which is extremely important for the predictive value of the model as a whole (Oprea & Waller, 1997). S. Ekins et al. / Journal of Pharmacological and Toxicological Methods 44 (2000) 251���272 253