Advances in the Study of Drug Metabolism – Symposium Report of the 12^th Meeting of the International Society for the Study of Xenobiotics (ISSX)

Abstract

The 12^th International Society for the Study of Xenobiotics (ISSX) meeting, held in Portland, OR, USA from July 28–31, 2019, was attended by diverse members of the pharmaceutical sciences community. The ISSX New Investigators Group provides learning and professional growth opportunities for student and early career members of ISSX. To share meeting content with those who were unable to attend, the ISSX New Investigators herein elected to highlight the “Advances in the Study of Drug Metabolism” symposium, as it engaged attendees with diverse backgrounds. This session covered a wide range of current topics in drug metabolism research including predicting sites and routes of metabolism, metabolite identification, ligand docking, and medicinal and natural products chemistry, and highlighted approaches complemented by computational modeling. In silico tools have been increasingly applied in both academic and industrial settings, alongside traditional and evolving in vitro techniques, to strengthen and streamline pharmaceutical research. Approaches such as quantum mechanics simulations facilitate understanding of reaction energetics towards prediction of routes and sites of drug metabolism. Furthermore, in tandem with crystallographic and orthogonal wet lab techniques for structural validation of drug metabolizing enzymes, in silico models can aid understanding of substrate recognition by particular enzymes, identify metabolic soft spots and predict toxic metabolites for improved molecular design. Of note, integration of chemical synthesis and biosynthesis using natural products remains an important approach for identifying new chemical scaffolds in drug discovery. These subjects, compiled by the symposium organizers, presenters, and the ISSX New Investigators Group, are discussed in this review.

Introduction:

The International Society for the Study of Xenobiotics (ISSX) Symposium “Advances in the Study of Drug Metabolism” held in Portland, OR, USA on July 30, 2019 highlighted techniques and approaches within the field of drug metabolism that further our scientific understanding and improve the efficiency of drug discovery. The co-chairs, Iain Martin (Relay Therapeutics, USA) and Mitchell Taub (Boehringer Ingelheim, USA) recognized that comprehensive coverage of this topic in one symposium would be impossible. Nonetheless, they attempted to pique the interests of the attendees with emerging efforts of particular importance to the drug metabolism community.

The ISSX New Investigator Group works together to ensure that student and early career stage members of ISSX are actively engaged in the Society and find value in their membership by providing learning opportunities, networking and mentorship events, and career development. In this spirit, the ISSX New Investigator Group elected to highlight a symposium from the ISSX Annual Meeting to share the insights of the speakers and the questions from the audience; in particular, to provide those who were not in attendance the opportunity to be enriched by the symposium. Though each of the symposia organized for the annual ISSX meeting provided great scientific insights, the ISSX New Investigator Group selected the “Advances in the Study of Drug Metabolism” symposium as it was engaging to attendees with diverse backgrounds. The topics discussed were wide-reaching and spanned medicinal chemistry, drug metabolism, biotransformation, metabolite identification, molecular docking, crystallography, biosynthetic approaches, and in silico predictive techniques. The multidisciplinary nature of these topics highlights the diverse expertise of each speaker, and reinforces the synergistic effect the combination of approaches has on improving drug discovery and development programs. We believe that each of the approaches discussed can have near-immediate positive impacts on drug discovery and development. Uniquely, the already successful computational approaches mentioned will improve over time as further experimental data is gathered to improve data and training sets, resulting in an iterative process likely to positively affect the field of drug metabolism.

From visualizing substrates in the active sites of drug-metabolizing enzymes, to predicting sites and routes of metabolism, pharmacokinetic parameters, and doses in man, in silico techniques are an integral part of the drug metabolism scientist’s toolbox. These techniques are increasingly used to screen virtual compounds and help direct synthetic efforts in drug discovery. Towards the synthesis of these optimized molecules, creative biosynthetic methods are often employed to complement traditional synthetic techniques. In addition to being a tool used to broaden and explore chemical diversity, biosynthesis is also being leveraged to generate metabolites in support of drug discovery and development efforts. Computational approaches to predicting drug metabolism and pharmacokinetic properties have long been of interest, especially to pharmaceutical industry scientists. Indeed, one of the co-chairs recalls attending a demonstration of metabolism prediction software at the 1990 ISSX annual meeting!

The speakers for the session represented diverse research expertise from both academic and industry settings. Matt Segall (Optibrium, UK) provided an update on current strategies to predict routes and sites of drug metabolism, integrating how this knowledge can be harnessed to identify possible drug-drug interactions, as well as potentially active and reactive metabolites. Emily Scott (University of Michigan, USA) described recent advances in understanding cytochrome P450 (P450) substrate recognition. Current research on CYP1A1 illustrated the necessity of a broad range of P450 structures in order to increase the reliability of docking for structurally diverse substrates. Fabio Broccatelli (Genentech, USA) provided an overview of how in silico drug metabolism and pharmacokinetic (DMPK) approaches complement experimental paradigms in drug discovery and development. Such modeling techniques can be used, for example, to predict molecular structures with the highest likelihood of desirable absorption, distribution, metabolism, and excretion (ADME) properties, thereby helping to identify which molecules to synthesize and test (or not test) in preclinical assays. In silico endpoints are also increasingly incorporated into early predictions of human therapeutic dose. In the last talk of the session, Griff Humphreys (Aranmore Consultants, USA) addressed the use of biosynthetic methods alongside traditional chemical synthesis to explore chemical diversity and to provide new starting points for drug discovery. Metabolites accessed via such strategies may be required for assessment of pharmacological activity, as analytical standards in clinical programs, or for determination of their safety.

The co-chairs and meeting organizers were very pleased with the high level of attendance at this symposium, the engaging question and answer session following each speaker, and the positive feedback following its conclusion. Clearly, technological advancement, with computational technologies being a fundamental pillar, provides a powerful approach for expediting and economizing drug discovery and development programs, and thus will continue to be of high interest to ISSX meeting attendees.

Predicting Routes, Sites, and Products of Metabolism (Matthew Segall, Optibrium Limited)

Understanding the metabolic fate of a compound is essential throughout drug discovery and development. Rapid metabolism limits the systemic exposure to a compound, active, reactive, or toxic metabolites may be formed, and metabolic clearance by a single predominant enzyme isoform increases safety risks related to drug-drug interactions or genetic variation in the patient population. Many approaches have been applied to the prediction of drug metabolism, from empirical data-driven approaches, such as machine learning which frequently provides rapid predictions with lower precision due to inherent assumptions and approximations, to mechanistic models of the action of an enzyme based on molecular dynamics or quantum mechanics (QM) which are significantly more precise but extremely time- and effort-intensive (Table 1). The ideal blend of speed, performance (precision and accuracy), and transferability comes from a combination of methods across this spectrum.

Table 1.

The Pros and Cons of Approaches to Modeling of Drug Metabolism

Category	Empirical Models	Mechanistic Models
Speed	Fast	(Very) slow
Ease of Use	Easy to assemble	Requires detailed understanding
Data Requirements	Requires significant data	Can be built on smaller high-quality data sets
Transferability	Non-transferable	Transferable – based on physical principles
Data Output	Qualitative	Semi-quantitative

Predicting Cytochrome P450 Metabolism:

Cytochromes P450 (P450s) have been studied for many years and there are several well-established computational models that predict the sites of metabolism (SOMs), and hence products, by individual P450 isoforms (Kirchmair et al. 2012). There are two factors that determine the SOM: (1) the electronic properties of the substrate, which determine the reactivity of each potential SOM to hydrogen-abstraction or direct oxidation by the active oxy-haem compound I and (2) the accessibility of each potential SOM in the binding pocket of a specific isoform. The former can be calculated using quantum mechanics simulations to estimate the activation energy of the rate-limiting step of the reaction, while the latter can be included as corrections to the activation energies using statistical models trained to high-quality experimental regioselectivity data for individual P450 isoforms (Tyzack et al. 2017). This combination enables accurate prediction of SOM and can capture subtle long-range effects of small substitutions on a substrate that can be observed during lead optimization. These calculations take 1–2 minutes for a typical drug-like molecule, making them appropriate for routine application.

Models have been developed that predict SOM for seven P450 isoforms (CYP3A4, CYP2D6, CYP2C9, CYP1A2, CYP2C8, CYP2C19, and CYP2E1). Therefore, the next important question is “Which isoforms will be responsible for metabolism of a novel compound of interest?” This can be addressed by utilizing a machine learning model trained on a data set of the P450 isoforms experimentally observed to metabolize a diverse range of substrates. Such a ‘WhichP450’ model can accurately rank the isoforms that are most likely to metabolize a particular compound (Hunt et al., 2018). The output of the WhichP450 model can be combined with SOM predictions to identify the most likely metabolites to be formed by P450-mediated metabolism.

Beyond P450 Metabolism:

In order to apply similar principles to modeling the contributions by non-P450 enzymes to xenobiotic metabolism, it is important to first thoroughly understand their catalytic reaction mechanisms. One example is the flavin-containing monooxygenases (FMOs). This class of enzymes predominantly catalyze nitrogen and sulfur oxidation reactions by transfer of an oxygen from a reactive flavin-adenine dinucleotide peroxide (FAD-OOH). Detailed QM studies of the FAD-OOH reaction with several known FMO substrates indicate that the reaction proceeds by a concerted, SN2 mechanism, as illustrated in Figure 1 for trimethylamine (Walton et al., 2019). A second example is in phase II metabolism with UDP-glucuronosyltransferases (UGTs), which are attributed with catalyzing approximately 40% of all conjugation reactions in humans (Guillemette et al., 2014). UGTs catalyze the conjugation of a substrate with glucuronic acid, transferred from a UDP-glucuronic acid cofactor. In this case, detailed QM calculations have revealed that the reaction mechanism is more complex than the previous examples, involving the transfer of protons from two highly conserved histidine residues in the active site (Öeren et al., 2019).

Figure 1.

The reaction mechanism for Flavin-containing Monooxygenase (FMO) metabolism, showing the energetics for oxidation of trimethyl amine, calculated with density functional theory.

As with prediction of P450 metabolism, the activation energies calculated using QM simulations can be combined with steric and orientation descriptors using a machine learning model. In the case of the FMO and UGT models, classification models were built using the Gaussian processes method (Obrezanova and Segall 2010) and resulted in an excellent ability to identify SOMs for FMO and UGT metabolism, with accuracies of 92% and 83% for the FMO3 and UGT1A1 models, respectively (Walton et al., 2019) (Öeren et al., 2019).

Predictions of SOMs can be enhanced with an understanding of the extent of metabolism to contextualize the potential impact on overall drug disposition. As alluded to above, characterizing the most likely SOMs can guide optimization of molecules with desirable pharmacologically active metabolites or eliminate compounds with the potential for bioactivation-related toxicity via reactive metabolite formation. However, understanding of the extent of such formation allows for meaningful predictions of clearance and/or risk assessment in such scenarios. Further, SOM prediction can implicate the specific metabolic enzyme(s) involved in the biotransformation of a molecule of interest, thus providing insight into the potential for drug-drug interactions or pharmacogenomic variance to impact drug clearance prior to any in vitro experimentation. In this scenario, knowledge of the extent of metabolism for each isoform can help understand the potential risks.

In conclusion, combining models of different steps in the catalytic cycle enables us to predict routes, sites, and products of metabolism. Detailed QM simulations facilitate understanding of the reaction mechanisms responsible for metabolism. This improves metabolism prediction performance and enhances transferability by calculating the reaction energetics and supplementing this with chemical descriptors capturing the steric and orientation effects of the protein environment using machine learning models.

The integration of deep learning into predicting sites of metabolism may enhance prediction performance relative to conventional approaches through using increasingly large sets of data. With this in mind, revealing mechanisms of less well-characterized drug metabolizing enzymes will be required to complete these datasets in order to develop robust machine learning models. One such example is better understanding the plasticity of metabolic enzyme active sites under varying conditions, as described for CYP1A1 below, which could ultimately guide improved drug design.

Substrate Recognition by Cytochromes P450:Applying Structure to Understand Cytochrome P450 1A1 Metabolism (Emily E. Scott and Aaron G. Bart, University of Michigan)

Using cytochrome P450 structural information to predict drug metabolites is of significant value in drug design but is often made difficult by the promiscuity and flexible active site of most human drug-metabolizing P450 enzymes. While structures are available for many of the human xenobiotic-metabolizing enzymes, the structure of a single P450/ligand complex does not often provide a comprehensive perspective on its ligand binding capacity and orientation. For example, human CYP1A1 is well-known for oxidizing planar polyaromatic hydrocarbons (Guengerich 2015) and, until recently, the only structure available was with such a compound: alpha-naphthoflavone (Walsh et al., 2013). It was obvious from this initial CYP1A1 structure demonstrating a planar, enclosed active site that conformational changes would be essential to accommodate larger and/or nonplanar validated CYP1A1 substrates.

The first was a complex with bergamottin, a furanocoumarin found in grapefruit juice, well-known for its ability to inactivate CYP3A4 (He et al., 1998) resulting in clinically significant alterations of the pharmacokinetics of CYP3A substrates. Bergamottin bound to the CYP1A1 active site with its geranyl side chain over the heme, with C6’ and C7’ closest to the heme iron (Bart and Scott 2018), consistent with production of the non-toxic 6’-hydroxy, 7’-hydroxy, and 6’,7’-dihydroxy metabolites observed. A second new complex was with the lung cancer drug and tyrosine kinase EGFR inhibitor erlotinib. Erlotinib bound to CYP1A1 with its ethynylbenzene ring directed toward the heme (Bart and Scott 2018). The benzene para position and the terminal carbon of the triple bond were closest to the iron. This orientation is consistent with CYP1A1 oxidation known to occur at these two positions, resulting in reactive metabolites that have been linked to adverse pharmacological effects (Lu et al., 2006; Liu et al., 2007; Li et al., 2010; Zhao et al., 2018). The planar portions of both substrates generally occupied a similar region of the active site, stacking with the side chain of an F helix Phe224 residue, similar to the arrangement observed when CYP1A1 binds alpha-naphthoflavone. However, both new compounds perturbed residues composing the active site roof to somewhat expand the CYP1A1 active site volume in the case of bergamottin, but creating a channel to the protein surface in the case of erlotinib (Bart and Scott 2018) (Figure 2A).

Figure 2.

CYP1A1 experimental X-ray structures and docking results. A. Comparison of CYP1A1 active site with bound to alpha-naphthoflavone (ANF) bound (light grey ribbons) vs. erlotinib (green ribbons), showing expansions in the active site cavity volume and shape (grey transparent surface for CYP1A1/ANF and green mesh for CYP1A1/erlotinib). Green arrows indicate secondary structure elements or residues specifically perturbed. B. Docking results using the three currently available CYP1A1 crystal structures with various ligands metabolized by CYP1A1. The experimental structures and corresponding PDB codes are listed across the top and the ligand docked into each structure is listed on the left side. The percentage of docking poses that successfully recapitulated orientations consistent with the experimental structure and/or known metabolites are color coded by the legend below. C. Docking into the CYP1A1/erlotinib structure was most successful in orienting known sites of metabolism (red dashes) adjacent to the heme iron.

While these new structures reveal the capacity to significantly enlarge the CYP1A1 active site and conformational landscape, the question remains whether this expanded structural dataset is sufficient to understand and predict the orientations of other compounds using readily accessible and computationally inexpensive docking methods. Towards addressing this question, initial docking studies were used to first validate the approach with compounds whose orientations are known from solved crystal structures. Docking was then initiated for a series of CYP1A1 substrates that are clinical tyrosine kinase inhibitors. Initial validation consisted of (1) taking the three CYP1A1 structures determined with alpha-naphthoflavone, bergamottin, and erlotinib, (2) deleting the respective ligand, (3) docking each of the three ligands into each of the three structures, and (4) comparing the dominant orientation(s) to those observed in the crystallographic structure. Minimization and docking procedures were performed with a rigid receptor and flexible ligand sampling using the Schrödinger Maestro suite (Schrödinger 2018). If the known site(s) of binding or metabolism were within 6Å from the heme iron, the docking prediction was categorized as successful. Aggregated docking outcomes were classified by the percentage of these correct orientations (Figure 2B).

The experimental CYP1A1 structure determined with bergamottin successfully oriented bergamottin and erlotinib most of the time but was less successful in yielding the correct orientation of alpha-naphthoflavone. Alternatively, all three ligands docked with the dominant solution being the correct orientation using the experimental CYP1A1 structure observed when erlotinib is bound, suggesting that this structure might be most useful for docking an even wider range of compounds.

Subsequent tests focused on predicting the binding orientations of three structurally diverse tyrosine kinase inhibitors (TKIs) metabolized by CYP1A1. The first TKI, sunitinib, is approved for renal cell carcinoma and gastrointestinal tumors (Goodman et al., 2007) and is primarily N-de-ethylated to a nontoxic metabolite (Zhao et al., 2019). The second TKI, gefitinib, is used in the treatment of non-small cell lung cancer and is defluorinated by CYP1A1 to a para-hydroxyaniline moiety. This moiety may be subject to further oxidation to a reactive quinone-imine with potential to form adducts with glutathione or proteins, which have been linked to adverse pulmonary and hepatic effects (Li et al., 2009). Ponatinib, the third TKI, is a BCR-ABL inhibitor used in the treatment of chronic myeloid leukemia that is oxidized by CYP1A1 on the purine, ethynyl linkage and methylphenyl group (Lin et al., 2017), proposed to generate an epoxide which forms undesirable protein adducts. Understanding why some TKIs result in toxic metabolites is especially important in smokers, where CYP1A1 expression is induced. The original CYP1A1/alpha-naphthoflavone structure was unable to accommodate the larger compounds sunitinib and ponatinib, while incorrectly predicting the pose for the gefitinib site of metabolism. Docking into the CYP1A1/bergamottin structure successfully positioned sunitinib in an orientation consistent with the known metabolite but produced many fewer correct poses for gefitinib and none for ponatinib. In contrast, the experimental CYP1A1 structure determined with erlotinib bound was able to correctly position all three drugs in orientations consistent with the known metabolites (Figure 2B and 2C).

Overall, this study highlights several critically underappreciated concepts. First, docking outcomes are fundamentally dependent on the initial protein structure used, thus significant structural validation is important prior to utilization, particularly with flexible drug metabolizing P450 enzymes. Second, in many cases we do not have enough structures of drug metabolizing P450 enzymes and thus fail to reliably explore the breadth of active site conformational flexibility. Third, docking results must be more frequently validated with orthogonal wet lab experiments. In the case of CYP1A1, the expanded active site cavity with a channel performed best in correctly orienting test compounds with a range of sizes and structures. A newer human CYP1A1 structure with the nonplanar kinase inhibitor GDC-0339 demonstrates an even more expanded active site that may be more appropriate for docking bulkier ligands (Bart et al., 2019). However, a current limitation revealed by conformation-dependent binding of different substrates to CYP1A1 is that simulated ligand-enzyme interactions may only apply to one conformation. This leaves space for improving prediction models, where manipulation of the model to favor a particular protein conformation may be appropriate depending on the drug or drug class of interest. Nonetheless, t his new structural information for CYP1A1 may be useful in the redesign of clinical agents like gefitinib and ponatinib to remove metabolic liabilities and to better anticipate toxic metabolites for new pharmaceutical agents.

Application of in silico DMPK Drug Discovery and Development (Fabio Broccatelli, Genentech, Inc.)

In addition to their utility in predicting sites of metabolism and metabolite production, in silico approaches are central to discovering novel and effective therapeutics. Drug discovery is an iterative process involving the synthesis and experimental assessment of thousands of compounds in order to identify a single clinical candidate. Consequently, technology that expedites the identification of clinical candidates during lead optimization has the potential to positively impact the efficiency and cost-effectiveness of research and development. The scientific and bioinformatic advancements in lead molecule finding and optimization over recent decades have increased the availability of high-quality in silico, in vitro, and in vivo models (Waring et al., 2015). These models allow investigators to tease out the mechanistic components of new molecular entity disposition. As a result, modern scientists focusing on lead optimization are tasked with interpretation of highly dimensional data matrices spanning across multiple disciplines. In addition to challenges related to supporting a diverse array of customized assays in a weekly schedule, pharmaceutical scientists are now attempting to enhance the value of their experimental data by investigating its translatability through bioinformatic analysis and in vitro to in vivo extrapolation.

Experimental and computational data collected during drug discovery campaigns has an immediate data value that rapidly drives project decisions following compound testing and a latent data value which can be harnessed with the use of cheminformatics, machine learning, and data mining. At Genentech, an array of in silico absorption, distribution, metabolism and excretion (ADME) tools are used to exploit the latent data value by focusing on five key areas: (1) generation and scoring compound design ideas, (2) prediction of in vitro assay results and assisting in vitro data analysis, (3) rationalization of the assay cascade, (4) in silico exploration to produce in vitro to in vivo correlations, and (5) utilization of mechanistic simulations to inform project strategy. This segment seeks to highlight how computational modeling is broadly incorporated throughout the preclinical and clinical testing of potential therapeutic agents.

Artificial intelligence (AI) is routinely used to estimate in vitro DMPK properties from molecular structures (Lombardo et al., 2017), however, prediction accuracy is limited by inherent prediction errors produced by AI models and by the inherent variability in experimental assays. Nonetheless, before a compound is synthesized, admittedly biased human expertise is the only alternative tool available to guide compound design. In this space, categorical and semi-quantitative AI models built on large in vitro datasets have consistently shown value over the years and have transformed the ADME landscape of compounds selected for synthesis (Aliagas et al., 2015). In particular, liver microsomal and hepatocyte stability, liver microsomal binding, solubility, permeability, transporter-mediated efflux, and lipophilicity are routinely estimated and considered in the selection of synthesis targets. AI models are easily scalable and can be used to score millions of compounds in a matter of hours; these technological improvements are gradually shifting the bottleneck in small molecule drug design from using AI to score synthetically viable design ideas, to use AI to generate synthetically viable ideas . Chemical design strategies that are driven not only by the judgment of medicinal chemists but also augmented by AI- and cheminformatics-driven strategies are currently active areas of research (Ford et al., 2015)(Rifaioglu, 2019). Overall, the increase in computer processing power seen in recent years, combined with increasing availability of large datasets (both proprietary and open source) allowed for development of more sophisticated machine learning models that are to a large extent being open-sourced. These tool are progressively being tested and incorporated in the background of a growing number of drug discovery scientists, and represent a substantial promise to positively streamline drug discovery at present.

Matched Molecular Pairs (MMPs) have been employed as tools to investigate the impact of a well-defined structural transformation on properties of interest, with resulting information used to guide decision-making (Lombardo et al., 2017). Specifically, an MMP is a pair of compounds that differ only by a single structural fragment, with the chemical change referred to as a transformation (e.g. phenyl to pyridine). The same transformation may occur across dozens of different chemical scaffolds. Whereas quantitative structure activity/property relationship (QSAR or QSPR) models aim to predict absolute ADME values for entire molecules, MMP models focus on subtle differences between a pair of molecules to predict the effects of a particular transformation. The average changes in ADME properties from MMPs can subsequently be summarized and used to develop future design approaches. For example, mutating a phenyl into a pyridine tends to decrease lipophilicity, increase solubility, and shows little to no improvement in liver microsomal metabolic stability. Databases of MMPs can therefore be used to identify transformations that may result in improved ADME profiles for a given chemical structure. At Genentech, the exploitation of MMPs is enabled by a web-based tool that allows compound enumeration and multi-parameter optimization.

An MMP-based approach was recently applied to the optimization of in vivo drug half-life in rats by investigating the impact of changes in lipophilicity versus changes in in vitro clearance in rat hepatocytes, based on in-house Genentech data. Examination of MMPs with a transformational change that results in similarities of one property (i.e. lipophilicity) but differences in another property (i.e. in vitro metabolic stability) revealed that the best approach to extend in vivo drug half-life was to decrease clearance in rat hepatocytes without altering lipophilicity (Broccatelli et al., 2018). These findings could prove useful in guiding future drug design efforts, as they contradict conventional medicinal chemistry approaches of decreasing lipophilicity to prolong in vivo half-life. Further, it should be noted that half-life depends on both clearance and volume of distribution (V_ss), therefore attempts to decrease clearance by utilizing the commonly-employed strategy of increasing polarity can be misleading, due to the counter effect on V_ss (Broccatelli et al., 2018). Therefore, it may be more appropriate to identify structural modifications that will improve half-life by a thorough analysis of metabolite identification (MetID) experimental data. This reinforces the importance of embedding MetID experiments early in research. While the expertise of MetID scientists is far from replaceable, computer assistance has decreased the experimental burden while increasing experimental throughput (Bonn et al., 2010) (Zamora et al., 2013).

Recognizing progress at the end of a ‘design-make-test’ cycle is not as straightforward as one might anticipate. One must investigate the in vitro to in vivo correlations of pharmacokinetic properties in preclinical species with the understanding that species differences are inherent, and therefore a high clearance compound in mice may not necessarily display high clearance in humans. In silico tools can be used to incorporate in vitro to in vivo correlation information into the design of new molecules, however, this warrants examination of the in silico-to in vitro- to in vivo correlations. This requires tracking prospective in silico predictions, correlating in vitro predictions to in vivo data for each preclinical species, and examining degree of prediction success for each pharmacokinetic property in each species, for various projects, and for various scaffolds within each project. Though benefits of implementing in silico predictions into a chemical optimization paradigm can be demonstrated, making these tools accessible and easy-to-use is necessary to ensure they are routinely applied. At Genentech, an investment was made to enable seamless exploration of this high dimensional data matrix, allowing insights into in silico- to in vitro- to in vivo- correlations. Additionally, user-friendly web modeling tools based on a simple one compartment model were made available to enable non-PK experts to explore chemical property trade-offs in the context of dose projection. This facilitates estimating cost benefits of improving one property against another (e.g. clearance vs. potency), as well as exploring the difference in dose estimate resulting between exposure- vs. half-life- driven PKPD relationships. This provides a platform for scientists with different backgrounds to communicate more efficiently in shaping project strategy.

Biosynthesis of Drug Candidates and Metabolites (W. Griffith Humphreys, Aranmore Pharma Consulting)

While modern synthetic chemistry provides the ability to synthesize an incredible variety of new structures, the chemical diversity of the natural world remains unmatched. This diversity has historically been employed in the drug discovery process through isolation of an organic molecule from a biological source as a drug candidate, usually referred to as natural product chemistry. Recently, an open access database named The Natural Products Atlas (www.npatlas.org) has compiled over 24,000 compounds related to natural products chemistry, including structural data, source organisms, and isolation references (van Santen et al., 2019). The ability to deposit information as well as search by physical properties, structure, and substructure are only a few capabilities offered by this database. This collaborative resource presents an attractive means to expedite microbial natural products chemistry research in drug discovery and development. Another application of biological processes in drug discovery is by providing enzymes that perform chemistry not easily accessed through synthetic chemistry approaches. This overview of the use of biosynthesis in drug discovery and development focuses on this latter aspect, the use of biosynthesis to augment overall medicinal chemistry efforts.

Natural or engineered enzymes can be used in candidate discovery and development during multiple stages, such as: (1) to generate chemical diversity in conjunction with synthetic efforts, (2) to provide metabolite standards during candidate characterization, or (3) to scale-up synthesis of a drug candidate molecule. Because they perform chemistry on an incredibly wide range of substrates and generate products with interesting pharmacology, the predominant enzymes appropriate for these activities are from the cytochrome P450 family.

The same diversity of enzymatic reactions that produces natural products can be tapped for the diversification of lead compounds or libraries. There has been considerable literature on lead diversification through enzymatic transformation with an excellent example being a study by Arnold and co-workers that describes a small library being modified by a panel of mutated bacterial CYP BM-3 enzymes (Sawayama et al., 2009). This approach can be used to produce a modified library covering a significantly wider range of chemical space than synthetic libraries alone. It does, however, necessitate a more complex screening protocol with a deconvolution step to discriminate the activity of mixtures. Another approach is to use lead compounds as the starting point for enzymatic diversification (Masood et al., 2012; Obach et al., 2018; Fessner 2019).

The most common use of biosynthetic tools by ADME scientists is likely for targeted metabolite synthesis. This process begins with the identification of a metabolite of interest during an in vitro or in vivo experiment with a drug candidate. The initial system used for identification may provide sufficient quantities of the metabolite for structure identification. However, it is often the case that larger quantities of the metabolites of interest are required for characterization purposes. This is especially true for major metabolites, which are present in human plasma at quantities greater than 10% of the total drug-related material. There are many options for the process of metabolite scale-up (Cusack et al., 2013), several of which include: (1) chemical synthesis, (2) isolation from human or animal in vitro systems or excreta samples, or (3) isolation from incubations with microbes. The important factors to consider when selecting a method for metabolite synthesis are overall resource cost, timing of delivery that meets project needs, and quality of final product.

Many metabolites can be synthesized with relatively small changes to the synthetic scheme used for the parent. This is often the case with metabolites formed from alterations of side chains that have simple points of attachment, e.g., N- or O-dealkylations or side-chain oxidation reactions. In these cases, chemical synthesis of the metabolite will likely be the most cost-effective method. For more complex metabolite structures, biosynthetic methods will be very competitive in total resource costs. Examples of in vitro systems for human metabolite biosynthesis include (1) human hepatocytes, (2) subcellular fractions containing drug-metabolizing enzymes (microsomes, cytosol, and S9 fractions), or (3) drug-metabolizing enzymes expressed in systems derived from mammalian cells, inset cells, yeast, or bacteria such as E. coli. There have been several interesting new developments in engineered mammalian CYP proteins that may be useful for these applications (Gumulya and Gillam 2017; Zhang, 2019). The engineered proteins described in these new studies have properties such as improved thermal stability or new functional capabilities.

Another potential source of enzymatic activity useful for metabolite synthesis is microbial cultures. Bacteria and fungi synthesize many natural compounds that have drug-like properties. Indeed, the majority of antibiotics have been isolated from these sources. There is limited literature on the use of microbial systems in the context of metabolite synthesis. This is likely the result of the inability of ADME-groups to readily access the systems for microbial culture. Also, it is cumbersome to screen multiple microbial strains for targeted activity if conducted with traditional shake flask cultures. One solution to this problem, as described by Li et al. (2008), is to group microbes that can be cultured under the same conditions on 24-well plates to streamline the process. Machine learning models can also be harnessed to improve antibacterial drug discovery, where conventional drug discovery approaches have been unable to meet demands for novel antibiotics. Although the application of AI to natural product drug discovery is in its early days, it has the potential to identify patterns within nature and to improve high throughput rapid screening of biosynthetic drug candidates (Durrant and Amaro, 2014).

BMS-986142, a reversible inhibitor of the Bruton’s Tyrosine Kinase (BTK) enzyme, was an early development candidate that had progressed into first-in-human studies (Lee et al., 2017). Human plasma samples were profiled for circulating metabolites and two hydroxylated species were identified (Zhao, 2017). Based on LC-UV quantitation, one of these species was judged to be a major metabolite. The compound displayed very slow metabolism in microsomes and hepatocytes, and isolation of a sufficient quantity of the key metabolites for structure identification or further characterization proved to be challenging. The metabolites were also in low quantities in human or animal in vivo samples; thus, these were not useful sources to isolate the metabolites. The in-house microbial screening system described above was employed (Zhao et al., 2017). One particular strain, Streptomyces violascens, produced a metabolite that matched the mass spectra and retention time of the human metabolite and was thus deemed to be the same species (Figure 4). The nuclear magnetic resonance (NMR) of this metabolite demonstrated that the hydroxyl group was on the dehydrocyclohexane ring, a structure not readily accessed via chemical synthesis. The microbial system could be scaled to provide sufficient quantities of the human metabolite for it to be isolated and characterized. The metabolite was found to be approximately 10-fold less potent against the pharmacological target and did not demonstrate any new off-target pharmacology. In particular, no activities were observed against related kinase enzymes or in drug-drug interaction screens.

Figure 4.

Biosynthesis of the major circulating human metabolites of BMS-986142. The parent drug was incubated with multiple bacteria strains and the culture containing Streptomyces violascens was found to be the most active towards formation of the metabolite that matched the target human metabolite. Scale-up incubations with Streptomyces violascens allowed purification of mg quantities of the target metabolite (Zhao et al., 2017).

Metabolite biosynthesis and characterization is an important step in drug development. Much of the attention regarding metabolites is devoted to meeting the requirements of the metabolites in safety testing guidance (U.S. FDA Guidance, 2016). While this is a very important goal, there are other potential benefits of metabolite characterization. In select cases, the metabolite may retain the potent pharmacology of the parent compound, and if found early enough in the discovery and development process, can become the lead compound (Fura et al., 2004). Furthermore, the process of complete characterization of metabolite structure and properties can be important regarding intellectual property. Metabolite characterization will continue to be an important part of drug development and can be accomplished by a multi-pronged and ever-evolving approach, including both synthetic and biosynthetic techniques.

Conclusions:

In summary, the “Advances in the Study of Drug Metabolism” symposium provided ISSX attendees insight into the diverse approaches currently being pursued in academia and industry to advance the field of drug metabolism. A theme throughout this session was the application of predictive in silico tools to improve the efficiency and effectiveness of drug discovery and development. Although machine learning presents unique challenges including the need for robust datasets to avoid biases, more so than ever before the advancement in computational technology is at its peak. Therefore, as we accumulate more data and enhance model structures and performance, predictions can only improve. For the approaches discussed within this manuscript, computational predictions are employed to aid in identification of the routes (enzymes) and products (metabolites) of drug metabolism, as well as in prediction of ADME properties in support of lead optimization efforts. Furthermore, SOM predictions employing P450 structural information emphasized the necessity for a broad range of initial structural information for the specific P450 and integrated validation using both wet lab and in silico experiments. In addition, the biosynthetic exploitation of metabolic sources such as enzymes, microbes, and fungi to supplement medicinal chemistry efforts in the generation of novel or hard-to-synthesize chemical matter demonstrates the power of harnessing natural product chemistry to facilitate metabolite synthesis. Together, these efforts highlight significant advancements within the field to address specific drug discovery and development challenges. Though difficult to directly quantify the potential impacts of the novel approaches discussed in this manuscript on the future of drug development and discovery, we are confident that these methodologies, along with the current collaborative efforts facilitating integration of biologists, chemists, bioinformaticians, computational scientists, physicians, among others, will provide tangible immediate and future improvements, creating a more streamlined and efficient approach to discover effective therapeutics. It should be noted that these valuable tools address just a portion of the many challenges faced by the field, and we look forward to updates on drug metabolism research at future ISSX symposia. As advancements towards understanding the impact of the microbiome on drug metabolism, metabolism of new drug modalities (RNAs, cyclic peptides, etc.) and new enzyme/cell models to study metabolism for instance, can provide an opportunity for further research. Additionally, the New Investigator Group looks forward to outlining additional sessions from future annual ISSX meetings, to provide members of the pharmaceutical science community who were unable to attend the opportunity to learn from the informative discussions held by the attendees.

Figure 3.

Top-down approach to drug discovery guided by in silico approaches

Abbreviations:

ADME

Absorption, Distribution, Metabolism, Excretion

AI

Artificial Intelligence

DMPK

Drug Metabolism and Pharmacokinetics

FMO

Flavin-containing Monooxygenase

MetID

Metabolite Identification

MMP

Matched Molecular Pairs

NMR

Nuclear Magnetic Resonance

P450

Cytochrome P450

PK

Pharmacokinetics

PKPD

Pharmacokinetics and Pharmacodynamics

QM

Quantum Mechanics

QSAR

Quantitative Structure Activity Relationship

SOM

Site of Metabolism

TKI

Tyrosine Kinase Inhibitor

UGT

Uridine Diphosphate-Glucuronosyltransferase

V_ss

Volume of Distribution at Steady State