Ebook: Virtual ADMET Assessment in Target Selection and Maturation

“The Solvay Pharmaceutical Conferences: where industry meets academia in a search for novel therapies”

At the Crossroad of Drug Discovery

Drug discovery has undergone a revolution over the last two decades. The times are over when absorption, distribution, metabolism, excretion and toxicity (ADMET) of candidate molecules was an add-on to discovery aimed to prepare for development or to comply with regulatory requirement. The current targeted drug design process requires that ADMET profiling is initiated early in the discovery process. Today, biologists and medicinal chemists realize that there is a strong relationship between pharmacodynamic (what the drug does to the organism) and pharmacokinetic (what the organism does to the drug) effects.

A significant contributing factor to the evolution in drug discovery was the methodological and technological revolution with the advent of combinatorial chemistry, high-throughput screening and profiling, and in silico prediction of target-based activity and ADMET properties. High-throughput screening and in silico methods have accelerated the process towards drugability of new chemical structures.

Another component of the revolution in drug discovery is the replacement of the disease (indication)-based approach by a target-based approach. A better understanding of pathophysiology of diseases and the underlying biological processes of diseases combined with explosive development of genomics and proteomics have been instrumental in the birth of this new paradigm. With target ligands now being designed in silico by molecular modeling and obtained later by means of chemical synthesis, the road is open to a selective in silico screening of ADMET properties.

This volume summarizes discussions of these three aspects of modern drug discovery, i.e. priority for targets, early ADMET assessment, and in silico screening, held during the Solvay Pharmaceuticals Conference entitled “Virtual ADMET Assessment in Target Selection and Maturation” organized May 11–13th, 2005 in Lucerne (Switzerland). The volume offers a selection of the lectures delivered at this conference. We trust that readers from academia as well as from industry will benefit from these studies.

W. Cautreels, C. Steinborn, L. Turski

Although access to novel technologies has increased and insights in biology and underlying disease mechanisms have improved immensely over the last decades, less innovative molecules are currently being approved and successfully launched by pharmaceutical industry. Some of the reasons are that development costs increased dramatically over the last two decades while productivity stagnated or declined, competition for in-licensing increased and marketing costs rose significantly, and regulatory scrutiny has mounted. As a result of these challenges growth and profitability of pharmaceutical industry declines. Therefore, new business models for the industry are required in order to respond to challenges imposed by economical conditions and by increased pressure of aging global population, eager to secure and to finance the new era of medicine leading to additional prolongation of life expectancy and improvement of quality of life at advanced age. The response to this new realm is the focus to therapeutic areas of strength and migration towards life-saving medicines. Therefore, continuation of prosperity of pharmaceutical industry requires profound changes in the research and development and sales organizations; the old paradigm of decision-making and existing power structures dictate if and how breakthrough scientific innovation, new technologies and insights are applied. Instead, the next pharmaceutical industry needs to radically reshape its business model to enable sustained growth and high profitability coupled with leadership in medical progress and innovation.

There has been exponential growth in the interest and availability within the last five years for the developments in predictive ADME modeling. The ambitious goal of those virtual assessments in target selection and maturation is to provide predictive tools to characterize and possibly drop compounds that are most likely to exhibit ADME or toxicity problems sooner. The impact of simulation and prediction in the success of drug discovery and drug development is here briefly presented. Starting from general consideration of what are a model and a prediction, arguments are presented that show that quality of models not only depends on development of computers and algorithms but also rely on the quality of the input data. In this sense, experimental and in silico aspects are very complementary. Evolution of the techniques, some benefits but also limits will be underlined. Among the topics that are covered, progresses in methods and technologies coupled with technical advances (e.g. power of computers) are summarized; evolution of predictive parameters used in pharmacophore elaboration and current limits of structure-based drug design or virtual screening approaches are presented; more recent challenging fields (e.g. polymorphism) are also presented.

In this overview, we first examine Structure-Activity Relations (SARs) and their components from a general point of view. Four types of interpretation emerging from statistically valid relations are considered, namely causal (mechanistic), contextual (empirical), fortuitous and tautological correlations. Implications for ADME predictions arise when discussing the diversity of interactions between active compounds (e.g. drugs) and biological systems.

In a second part, we share our views on the differences between pharmacodynamic targets (namely the sites of action of bio-active compounds, e.g. receptors and ion channels) and pharmacokinetic agents (namely the biological components that act on drugs to transport, metabolize, retain and excrete xenobiotics). While the former are usually characterized by a high (i.e. narrow) specificity towards their ligands, the latter have evolved to recognize chemically diverse compounds and thus to display a low (i.e. broad) specificity.

In a last part, we discuss the concept of molecular structure and focus on the fluctuations undergone by molecular form and functions. As a result, a molecule can exist in a large number of distinct microstates, the ensemble of which constitutes the property space of the molecule.

The first aim of this chapter is to describe the main representatives of substructure approaches and whole molecule approaches for logP calculation, to highlight their characteristics and to discuss their advantages and disadvantages. A further goal is a classification of currently available logP programs according to their methodological background. The limited reliability of logP programs urgently favours the use of a diverse set of programs instead of a single procedure. The majority of drugs is ionized at pH 7.4. Distribution coefficients (logD) reflect the lipophilicity of ionizable compounds; the implication of logD in drug development and available logD software are discussed. LogP and logD are of prime importance in ADMET predictions; a few examples for protein binding, BBB permeability, metabolism, and toxicity are given.

An overview of the structure-information representation (SIR) method of molecular structure description is presented with an emphasis on structure descriptors found useful in QSAR modeling of ADME-Tox endpoints. An artificial neural network (ANN) model of human plasma protein binding will be presented as an example of the utility of this method. A novel non-linear trend analysis used to interpret the relationship between changes in the input descriptor values and the corresponding change in the output prediction will be discussed as a means of interpretation for the input descriptors. Molconn structure-information representation descriptors were used in conjunction with artificial neural network descriptor selection and data-fitting techniques to produce a human plasma protein binding model based on 1000 drugs. The model demonstrated an external validation MAE of 12.7% on 200 drugs, with 90% of predictions within 30% of the experimental value. Discussion will include the role of the predicted properties that are used as inputs for this model.

Insufficient knowledge on absorption, distribution, metabolism and excretion limits drug discovery. Chemical synthesis of molecules and testing in experimental animals preceeds human use of drug candidates. However, this approach has low throughput and significantly decreases speed of drug finding. Therefore, in order to predict behavior of novel structures prior to the first administration in humans molecular descriptors are employed. Molecular descriptors can be applied in modeling relevant properties of molecules such as passive transcellular permeability, solubility, unspecific binding to proteins or volume of distribution. In addition, sets of molecular descriptors can be used to predict interaction of a molecule with biological membranes or proteins. They can also be used to predict regions around the molecule relevant for efficacy. Molecular descriptors have successfully been applied in many problem solving approaches in drug discovery and the future use of them is expanding.

Since the emergence of combinatorial chemistry and chemical libraries, great attention is being paid to the concepts of chemical diversity and chemical space. This approach is based on the assumption that molecular properties are invariant ones. But a growing computational power shows that a molecule cannot be considered as a static object but as an animated subject whose conformational changes may significantly affect the profile of any of its computable property. The ensemble of all conformers of a given compound is often taken as defining a conformational space. In a similar manner, many molecular properties can be shown to vary with the 3D-geometry of the molecule.

In particular, powerful computational methods based on molecular fields now allow some physicochemical properties to be computed for each conformer, as discussed in the first part of the chapter. Such methods include MEPs (Molecular Electrostatic Potentials), MLPs (Molecular Lipophilicity Potentials), which allows to back-calculate a partition coefficient of a given conformer, and the recent MHBPs (Molecular Hydrogen-Bonding Potentials). A range of property values corresponding to all realistic conformations must be examined and taken into account. The range of these values defines a property space whose form and extent will depend on both the solute and the relevant environment.

In a second part, the chapter focuses on the property spaces of the acetylcholine in a variety of polar and hydrophobic solvents. The effect of the solvent on the conformational behaviour of acetylcholine is analyzed together with the corresponding effects on the property spaces. Moreover, attention is being paid to the cross-correlation among the profiles of these spaces (both physicochemical and structural). These interrelations lead to a definition of the concept of molecular sensitivity which describes the ability of a molecule to modify its physicochemical properties as its geometry changes, as presented in the third part of the chapter. The receptor selectivity of alpha-adrenergic ligands offers an illustration of the interest and limits of molecular sensitivity and property space range in dynamic QSAR analyses.

From the historically grown archive of protein-ligand complexes in the Protein Data Bank small organic ligands are extracted and interpreted in terms of their chemical characteristics and features. Subsequently, pharmacophores representing ligand-receptor interactions are derived from each of these small molecules and its surrounding amino acids. Based on a defined set of only six types of chemical features and volume constraints, three-dimensional pharmacophore models are constructed, which are sufficiently selective to identify the described binding mode and are thus a useful tool for in silico screening of large compound databases. The algorithms for ligand extraction and interpretation as well as the pharmacophore creation technique from the automatically interpreted data are presented and applied to complexes derived from inhibitors of HRV coat proteins and of ABL tyrosine kinase.

For convenience to the patients and to increase compliance most drugs are given orally. Therefore high bioavailability is a key quest in most drug discovery projects. Low bioavailability usually results in undesired variability due to population differences. Early estimates of oral bioavailability can help to focus on most promising lead series and clinical candidates. This paper reviews some of the in silico attempts to predict oral bioavailability. However, bioavailability is a complex property and various pros and cons of current approaches will be discussed. Physiologically-based pharmacokinetic (PBPK) modelling is discussed as a promising approach to predict and simulate pharmacokinetics, including estimating bioavailability.

The chapter starts by describing what is meant by ‘brain uptake’ and will then focus on two quantitative measurements, (a) blood-brain distribution, and (b) brain perfusion. It is shown that there are no satisfactory ‘experimental models’ for either of these processes, and hence the need for in silico models. For blood-brain distribution, numerous in silico models have been proposed. The chapter will concentrate on those models for which a predictive assessment through independent test sets has been made, and which are capable of being implemented reasonably easily. There are several models in this category all of which can predict blood-brain distribution, as logBB, to about 0.4 log units, probably not far from the experimental error. Of course, such predictions are for passive transport only, and do not take into account active transport or efflux mechanisms.

The second quantitative measurement is brain perfusion that has replaced methods such as the Brain Uptake Index. Brain perfusion can be carried out using an injection of a bolus of a drug in blood, plasma, or saline, all at pH 7.4. Most recent work has used saline as the vehicle, but the number of published observations is relatively small, as compared to blood-brain distribution. Two recent papers that use perfusion from saline in order to construct in silico models will be discussed.

The effects of ionization of proton acids and bases will then be considered. It appears that for blood-brain distribution, no correction for ionization at pH 7.4 is needed (one or two models have an indicator variable for proton acids, but this is not very significant). On the other hand, in perfusion from saline at pH 7.4 it seems that some correction is needed, although the magnitude of the correction is not clear. Unless the problem of strong proton acids and strong proton bases can be solved, it will be difficult to construct general in silico models for brain perfusion from saline. In addition, the usual caveats over active transport and efflux mechanisms apply.

Finally, brain uptake as classified as CNS+ and CNS- is briefly considered.

The significance of transport proteins for the pharmacokinetics and pharmacodynamics of drug molecules has been recognized during the past decade. The most important and best characterized is P-glycoprotein (P-gp), a typical multidrug transporter that belongs to the large class of the ABC-transport protein superfamily. P-gp holds an important physiological role as a natural detoxification system. It is actively involved in various processes such as lowering oral drug bioavailability, preventing drugs from penetrating across the blood-brain barrier, decreasing intracellular accumulation of anti-cancer and other cytotoxic agents, HIV-protease inhibitors and many more. P-gp is involved in resistance mechanisms of tumor and other cells to a broad spectrum of agents thus ranking it as a major multidrug resistance (MDR) protein. Hence pharmaceutical companies have a strong interest in identifying potential P-gp substrates at an early stage of the drug development process.

Together with experimental assays computational or in silico methods are very useful in achieving this purpose. In silico tools do not require compound synthesis and biological testing and can be applied to hypothetical compounds permitting the rapid exclusion of the likely failures and contributing simultaneously to a better understanding of drug-protein interactions.

In silico approaches to model drug-P-gp interactions have undergone several stages depending on particular knowledge of the structure and the structure-function relationship of the protein. Due to absence of 3D structural data of P-gp with sufficient resolution the ligand-based drug design approaches are mostly applied.

A number of QSAR and 3D-QSAR models have been developed based on the experimental MDR reversing activity data of different classes of compounds proven to interact with P-gp. These QSAR models use structural fragments and various physicochemical parameters. Mostly, lipophilicity (logP) and size (molar refractivity) were identified as the main determinants for P-gp recognition. 3D-QSAR models using Comparative Molecular Field (CoMFA) and Similarity Indices (CoMSIA) Analyses provide more detailed information pointing to the role of hydrophobicity as a space distributed molecular property and involve additionally hydrogen-bond (HB) acceptor molecular fields.

Accumulation of appropriate data about interactions of drugs with particular sites on P-gp stimulated in silico pharmacophore modeling for a more detailed elucidation of the structural features of the drugs responsible for activity. Several pharmacophore models have been developed. Almost all pharmacophore patterns involve at least two hydrophobic and one HB-acceptor points. Some of these models also consider HB-donor interactions of P-gp related drugs.

The recently published X-ray structures of the bacterial ABC-transporter MsbA from Escherichia coli and Vibrio cholerae gave rise to homology models of P-gp. Several 3D-models have been developed that could contribute to elucidation of the structure-function relationships of the transporter. Very recent results on in silico identification of the protein binding pockets confirm the hypothesis of the existence of multiple binding sites within the protein.

Failures due to bad metabolic properties are one of the major reasons of problems in drug companies. Biotransformation reactions due to human CYP-enzymes must be taken into account as early as possible in the lead optimization process during the development of new therapeutic agents. To this end, it would be extremely valuable for the drug industry to have a computational predictive method to predict the isoform selectivity, the potential inhibitor effect and the site of metabolism (i.e. the place in a molecule where the metabolic reaction occurs) for xenobiotics. The experimental elucidation of these properties is usually a high resource-demanding task, which requires several experimental techniques and consumes a considerable amount of compound. The recognition of the metabolic site in silico could be of great help to design new compounds with better pharmacokinetic profile as well as to avoid the presence of toxic metabolites. Moreover, the methodology can be used either to suggest new positions that should be protected in order to avoid certain metabolic profile or to check the suitability of a pro-drug. The aim of the present paper is to report a new method, fast, easy and computationally inexpensive for predicting CYP-inhibition, substrate selectivity and site of metabolism, using human CYP X-ray structures and ad hoc developed 3D-homology models. The methodology works for the most important human cytochromes, but can be applied automatically to all the cytochromes about which the 3D-structure is known by X-ray or homology models. The methods thus appear as a valuable new tool in virtual screening and in early ADME-Tox field where potential drug-drug interaction and metabolic stability information must be evaluated to enhance drug design efforts.

The prediction of the structure of metabolites based solely on consideration of the chemical structure of the parent can provide value during drug discovery, including lead optimisation and development. Since metabolism frequently has a direct influence on pharmacological activity profiles, toxicity and idiosyncratic drug reactions, consideration of metabolism has become increasingly recognised as vital in the optimisation of lead compounds selected for development. By whatever means prediction is achieved it will always be limited by the extent of existing knowledge. The challenge is to find the means of capturing and accessing knowledge and complementing this with tools that can that can take account of specific structural and physicochemical properties for compounds of interest. Two types of expert systems can be defined, the first being human experts with access to knowledge databases and the second in silico prediction systems. The tools available to a human expert and the approaches to their use are discussed and illustrated. The development of in silico systems and the existing products are reviewed. In silico systems are still at an early stage of development and there is considerable scope for their further enhancement. However it is also the case that some human expertise should be included to modulate outputs. One element of metabolism is the ability to predict potential for formation of reactive intermediates, which may contribute to toxicity or other adverse effects. This issue is more directed to providing alerts rather than predicting the structures of metabolites and in silico systems with the capability of recognising structural elements according to defined rules could provide the ideal vehicle to make predictive assessments. Examples are described of how structural alerts can be identified from existing knowledge. Input from the above initiatives could be at any stage of the drug discovery process since it is a purely theoretical exercise. There could be value in early implementation since this will expand the chemical space of candidates based on metabolic considerations and will therefore provide the opportunity to include compounds with a broader range of metabolic stability. The strategy adopted will clearly depend on what is most suitable to each specific company based on its organisation.

Expert systems for the prediction of toxicity use a knowledge base of rules, collected from human experts, as the basis of their predictions. Publicly available examples include DEREK for Windows, HazardExpert, OncoLogic and the German Federal Institute for Risk Assessment's Decision Support System. Advantages of the expert system approach include the transparency of predictions, the ease with which new knowledge can be added and the potential for knowledge sharing, particularly between organisations. Practical applications within the pharmaceutical industry include their use in drug discovery, occupational hazard assessment and as supporting evidence in regulatory submissions.

High-throughput screening technologies in biological sciences of large libraries of compounds obtained via combinatorial or parallel chemistry approaches, as well as the application of design rules for drug-likeness, have resulted in more hits to be evaluated with respect to their ADME or DMPK properties. The traditional in vivo methods using pre-clinical species such as rat, dog, monkey, are no longer sufficient to cope with this demand. This contribution discusses the changes towards medium to high-throughput in vitro and in silico ADME screening [1]. In addition much more attention is now put on early safety and risk assessment of promising lead series and potential clinical candidates.