Prediction of In Vivo Pharmacokinetic Parameters and Time-Exposure Curves in Rats Using Machine Learning from the Chemical Structure.

Animal pharmacokinetic (PK) data as well as human and animal in vitro systems are utilized in drug discovery to define the rate and route of drug elimination. Accurate prediction and mechanistic understanding of drug clearance and disposition in animals provide a degree of confidence for extrapolation to humans. In addition, prediction of in vivo properties can be used to improve design during drug discovery, help select compounds with better properties, and reduce the number of in vivo experiments. In this study, we generated machine learning models able to predict rat in vivo PK parameters and concentration-time PK profiles based on the molecular chemical structure and either measured or predicted in vitro parameters. The models were trained on internal in vivo rat PK data for over 3000 diverse compounds from multiple projects and therapeutic areas, and the predicted endpoints include clearance and oral bioavailability. We compared the performance of various traditional machine learning algorithms and deep learning approaches, including graph convolutional neural networks. The best models for PK parameters achieved R2 = 0.63 [root mean squared error (RMSE) = 0.26] for clearance and R2 = 0.55 (RMSE = 0.46) for bioavailability. The models provide a fast and cost-efficient way to guide the design of molecules with optimal PK profiles, to enable the prediction of virtual compounds at the point of design, and to drive prioritization of compounds for in vivo assays.

Improving the Accuracy of Predicted Human Pharmacokinetics: Lessons Learned from the AstraZeneca Drug Pipeline Over Two Decades.

During drug discovery and prior to the first human dose of a novel candidate drug, the pharmacokinetic (PK) behavior of the drug in humans is predicted from preclinical data. This helps to inform the likelihood of achieving therapeutic exposures in early clinical development. Once clinical data are available, the observed human PK are compared with predictions, providing an opportunity to assess and refine prediction methods. Application of best practice in experimental data generation and predictive methodologies, and a focus on robust mechanistic understanding of the candidate drug disposition properties before nomination to clinical development, have led to maximizing the probability of successful PK predictions so that 83% of AstraZeneca drug development projects progress in the clinic with no PK issues; and 71% of key PK parameter predictions [64% of area under the curve (AUC) predictions; 78% of maximum concentration (Cmax) predictions; and 70% of half-life predictions] are accurate to within twofold. Here, we discuss methods to predict human PK used by AstraZeneca, how these predictions are assessed and what can be learned from evaluating the predictions for 116 candidate drugs.

Prediction of Fraction Unbound in Microsomal and Hepatocyte Incubations: A Comparison of Methods across Industry Datasets.

The fraction unbound in the incubation, fu,inc, is an important parameter to consider in the evaluation of intrinsic clearance measurements performed in vitro in hepatocytes or microsomes. Reliable estimates of fu,inc based on a compound's structure have the potential to positively impact the screening timelines in drug discovery. Previous works suggested that fu,inc is primarily driven by passive processes and can be described using physicochemical properties such as lipophilicity and the protonation state of the molecule. While models based on these principles proved predictive in relatively small datasets that included marketed drugs, their applicability domain has not been extensively explored. The work presented here from the in silico ADME discussion group (part of the International Consortium for Innovation through Quality in Pharmaceutical Development, the IQ consortium) describes the accuracy of these models in large proprietary datasets that include several thousand of compounds across chemical space. Overall, the models do well for compounds with low lipophilicity. In other words, the equations correctly predict that fu,inc is, in general, above 0.5 for compounds with a calculated logP of less than 3. When applied to lipophilic compounds, the models failed to produce quantitatively accurate predictions of fu,inc, with a high risk of underestimating binding properties. These models can, therefore, be used quantitatively for less lipophilic compounds. On the other hand, internal machine-learning models using a company's own proprietary dataset also predict compounds with higher lipophilicity reasonably well. Additionally, the data shown indicate that microsomal binding is, in general, a good proxy for hepatocyte binding.

Novel Chemical Series of 5-Lipoxygenase-Activating Protein Inhibitors for Treatment of Coronary Artery Disease.

5-Lipoxygenase (5-LO)-activating protein (FLAP) inhibitors have proven to attenuate 5-LO pathway activity and leukotriene production in human clinical trials. However, previous clinical candidates have been discontinued and the link between FLAP inhibition and outcome in inflammatory diseases remains to be established. We here describe a novel series of FLAP inhibitors identified from a screen of 10k compounds and the medicinal chemistry strategies undertaken to progress this series. Compound 4i showed good overall properties and a pIC50 hWBfree of 8.1 and an lipophilic ligand efficiency of 5.2. Target engagement for 4i was established in dogs using ex vivo measurement of leukotriene B4 (LTB4) levels in blood with good correlation to in vitro potency. A predicted human dose of 280 mg b.i.d. suggests a wide margin to any identified in vitro off-target effects and sufficient exposure to achieve an 80% reduction of LTB4 levels in humans. Compound 4i is progressed to preclinical in vivo safety studies.

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.

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.

In Vitro Intrinsic Permeability: A Transporter-Independent Measure of Caco-2 Cell Permeability in Drug Design and Development.

In vitro permeability data have a central place in absorption risk assessments in drug discovery and development. For compounds where active efflux impacts permeability in vitro, the inherent passive membrane permeability ("intrinsic permeability") gives a concentration-independent measure of the compound's permeability. This work describes the validation of an in vitro intrinsic permeability assay and application of the data in a predictive in silico model. Apparent intrinsic permeability (Papp) across Caco-2 cell monolayers is determined in the presence of an optimized cocktail of chemical inhibitors toward the three major efflux transporters ABCB1, ABCC2, and ABCG2. The intrinsic Papp value gives an estimate of passive permeability, which is independent of transporter expression levels and not limited by solubility or cell toxicity. An in silico model has been established to predict the Caco-2 intrinsic permeability and shown to consistently identify highly permeable compounds. The new intrinsic permeability assay is useful for early absorption estimates and suitable for absorption risk assessment in DMPK and pharmaceutical development.

Modeling Organic Anion-Transporting Polypeptide 1B1 Inhibition to Elucidate Interaction Risks in Early Drug Design.

The importance of transporter proteins for the disposition of drugs has become increasingly apparent during the past decade. A noted drug-drug interaction risk is the inhibition of organic anion-transporting polypeptides (OATPs), key transporters for the liver uptake of the widely used statins. We show here the development of a ligand-based in silico model for interaction with OATP1B1, an important representative of the OATP family. The model is based on a structural overlay of 6 known OATP1B1 inhibitors. A data set of about 150 compounds with published OATP1B1 inhibition data was compared to the resulting "transportophor," and a similarity threshold was defined to distinguish between active and inactive molecules. In addition, using a statistical model based on physicochemical properties of the compounds as prefilter was found to enhance the overall predictivity of the model (final accuracy 0.73, specificity 074, and sensitivity 0.71, based on 126 compounds). The combined model was validated using an in-house data set (accuracy, specificity, and sensitivity were 0.63, 0.59, and 0.78, respectively; 62 compounds). The model gives also a structural overlay to the most similar template enabling visualization of where a change in a given structure might reduce the interaction with the transporter.

Use of HµREL Human Coculture System for Prediction of Intrinsic Clearance and Metabolite Formation for Slowly Metabolized Compounds.

Design of slowly metabolized compounds is an important goal in many drug discovery projects. Standard hepatocyte suspension intrinsic clearance (CLint) methods can only provide reliable CLint values above 2.5 µL/min/million cells. A method that permits extended incubation time with maintained performance and metabolic activity of the in vitro system is warranted to allow in vivo clearance predictions and metabolite identification of slowly metabolized drugs. The aim of this study was to evaluate the static HµREL coculture of human hepatocytes with stromal cells to be set up in-house as a standard method for in vivo clearance prediction and metabolite identification of slowly metabolized drugs. Fourteen low CLint compounds were incubated for 3 days, and seven intermediate to high CLint compounds and a cocktail of cytochrome P450 (P450) marker substrates were incubated for 3 h. In vivo clearance was predicted for 20 compounds applying the regression line approach, and HµREL coculture predicted the human intrinsic clearance for 45% of the drugs within 2-fold and 70% of the drugs within 3-fold of the clinical values. CLint values as low as 0.3 µL/min/million hepatocytes were robustly produced, giving 8-fold improved sensitivity of robust low CLint determination, over the cutoff in hepatocyte suspension CLint methods. The CLint values of intermediate to high CLint compounds were at similar levels both in HµREL coculture and in freshly thawed hepatocytes. In the HµREL coculture formation rates for five P450-isoform marker reactions, paracetamol (CYP1A2), 1-OH-bupropion (CYP2B6), 4-OH-diclofenac (CYP2C9), and 1-OH-midazolam (3A4) were within the range of literature values for freshly thawed hepatocytes, whereas 1-OH-bufuralol (CYP2D6) formation rate was lower. Further, both phase I and phase II metabolites were detected and an increased number of metabolites were observed in the HµREL coculture compared to hepatocyte suspension. In conclusion, HµREL coculture can be applied to accurately estimate intrinsic clearance of slowly metabolized drugs and is now utilized as a standard method for in vivo clearance prediction of such compounds in-house.

Time dependent analysis of assay comparability: a novel approach to understand intra- and inter-site variability over time.

We demonstrate here a novel use of statistical tools to study intra- and inter-site assay variability of five early drug metabolism and pharmacokinetics in vitro assays over time. Firstly, a tool for process control is presented. It shows the overall assay variability but allows also the following of changes due to assay adjustments and can additionally highlight other, potentially unexpected variations. Secondly, we define the minimum discriminatory difference/ratio to support projects to understand how experimental values measured at different sites at a given time can be compared. Such discriminatory values are calculated for 3 month periods and followed over time for each assay. Again assay modifications, especially assay harmonization efforts, can be noted. Both the process control tool and the variability estimates are based on the results of control compounds tested every time an assay is run. Variability estimates for a limited set of project compounds were computed as well and found to be comparable. This analysis reinforces the need to consider assay variability in decision making, compound ranking and in silico modeling.

Exploring in silico prediction of the unbound brain-to-plasma drug concentration ratio: model validation, renewal, and interpretation.

Recently, we built an in silico model to predict the unbound brain-to-plasma concentration ratio (Kp,uu,brain), a measure of the distribution of a compound between the blood plasma and the brain. Here, we validate the previous model with new additional data points expanding the chemical space and use that data also to renew the model. The model building process was similar to our previous approach; however, a new set of descriptors, molecular signatures, was included to facilitate the model interpretation from a structure perspective. The best consensus model shows better predictive power than the previous model (R(2) = 0.6 vs. R(2) = 0.53, when the same 99 compounds were used as test set). The two-class classification accuracy increased from 76% using the previous model to 81%. Furthermore, the atom-summarized gradient based on molecular signature descriptors was proposed as an interesting new approach to interpret the Kp,uu,brain machine learning model and scrutinize structure Kp,uu,brain relationships for investigated compounds.

Discovery of AZD6642, an inhibitor of 5-lipoxygenase activating protein (FLAP) for the treatment of inflammatory diseases.

A drug discovery program in search of novel 5-lipoxygenase activating protein (FLAP) inhibitors focused on driving a reduction in lipophilicity with maintained or increased ligand lipophilic efficiency (LLE) compared to previously reported compounds led to the discovery of AZD6642 (15b). Introduction of a hydrophilic tetrahydrofuran (THF) ring at the stereogenic central carbon atom led to a significant shift in physicochemical property space. The structure-activity relationship exploration and optimization of DMPK properties leading to this compound are described in addition to pharmacokinetic analysis and an investigation of the pharmacokinetic (PK)-pharmacodynamic (PD) relationship based on ex vivo leukotriene B4 (LTB4) levels in dog. AZD6642 shows high specific potency and low lipophilicity, resulting in a selective and metabolically stable profile. On the basis of initial PK/PD relation measured, a low dose to human was predicted.

Resolving the distribution-metabolism interplay of eight OATP substrates in the standard clearance assay with suspended human cryopreserved hepatocytes.

Uptake transporters may act to elevate the intrahepatic exposure of drugs, impacting the route and rate of elimination, as well as the drug-drug interaction potential. We have here extended the assessment of metabolic drug stability in a standard human hepatocyte incubation to allow for elucidation of the distribution-metabolism interplay established for substrates of drug transporters. Cellular concentration-time profiles were obtained from incubations of eight known OATP substrates at 1 µM, each for two different 10-donor batches of suspended cryopreserved human hepatocytes. The kinetic data sets were analyzed using a mechanistic mathematical model that allowed for separate estimation of active uptake, bidirectional diffusion, metabolism and nonspecific extracellular and intracellular binding. The range of intrinsic clearances attributed to active uptake, diffusion and metabolism of the test set spanned more than 2 orders of magnitude each, with median values of 18, 5.3, and 0.5 µL/min/10(6) cells, respectively. This is to be compared with the values for the apparent clearance from the incubations, which only spanned 1 order of magnitude with a median of 2.6 µL/min/10(6) cells. The parameter estimates of the two pooled 10-donor hepatocyte batches investigated displayed only small differences in contrast to the variability associated with use of cells from individual donors reported in the literature. The active contribution to the total cellular uptake ranged from 55% (glyburide) to 96% (rosuvastatin), with an unbound intra-to-extracellular concentration ratio at steady state of 2.1 and 17, respectively. Principal component analysis showed that the parameter estimates of the investigated compounds were largely influenced by lipophilicity. Active cellular uptake in hepatocytes was furthermore correlated to pure OATP1B1-mediated uptake as measured in a transfected cell system. The presented approach enables the assessment of the key pathways regulating hepatic disposition of transporter and enzyme substrates from one single, reproducible and generally accessible human in vitro system.

In silico prediction of unbound brain-to-plasma concentration ratio using machine learning algorithms.

Distribution over the blood-brain barrier (BBB) is an important parameter to consider for compounds that will be synthesized in a drug discovery project. Drugs that aim at targets in the central nervous system (CNS) must pass the BBB. In contrast, drugs that act peripherally are often optimised to minimize the risk of CNS side effects by restricting their potential to reach the brain. Historically, most prediction methods have focused on the total compound distribution between the blood plasma and the brain. However, recently it has been proposed that the unbound brain-to-plasma concentration ratio (K(p,uu,brain)) is more relevant. In the current study, quantitative K(p,uu,brain) prediction models have been built on a set of 173 in-house compounds by using various machine learning algorithms. The best model was shown to be reasonably predictive for the test set of 73 compounds (R(2)=0.58). When used for qualitative prediction the model shows an accuracy of 0.85 (Kappa=0.68). An additional external test set containing 111 marketed CNS active drugs was also classified with the model and 89% of these drugs were correctly predicted as having high brain exposure.

Structure-brain exposure relationships in rat and human using a novel data set of unbound drug concentrations in brain interstitial and cerebrospinal fluids.

New experimental methodologies were applied to measure the unbound brain-to-plasma concentration ratio (K(p,uu,brain)) and the unbound CSF-to-plasma concentration ratio (K(p,uu,CSF)) in rats for 43 structurally diverse drugs. The relationship between chemical structure and K(p,uu,brain) was dominated by hydrogen bonding. Contrary to popular understanding based on the total brain-to-plasma concentration ratio (logBB), lipophilicity was not a determinant of unbound brain exposure. Although changing the number of hydrogen bond acceptors is a useful design strategy for optimizing K(p,uu,brain), future improvement of in silico prediction models is dependent on the accommodation of active drug transport. The structure-brain exposure relationships found in the rat also hold for humans, since the rank order of the drugs was similar for human and rat K(p,uu,CSF). This cross-species comparison was supported by K(p,uu,CSF) being within 3-fold of K(p,uu,brain) in the rat for 33 of 39 drugs. It was, however, also observed that K(p,uu,CSF) overpredicts K(p,uu,brain) for highly effluxed drugs, indicating lower efflux capacity of the blood-cerebrospinal fluid barrier compared to the blood-brain barrier.

Modeling of drug-transporter interactions using structural information.

Drug-transporter interactions have recently been demonstrated to play an important part in multidrug resistance, drug-drug interactions and drug disposition. Such interactions can occur as inhibition of transport, efflux out of cellular systems or enhanced transport into cellular systems. Modeling efforts are currently being undertaken using both ligand- and transporter-based methods, such as (3D) quantitative structure-activity relationship studies, pharmacophore modeling, homology modeling and molecular dynamics studies. The aim of these efforts is to explain how differences in chemical structures either enhance or weaken interactions with the transporter, or to elucidate how the transporter functions in general. This review summarizes recent modeling efforts in the light of difficulties such as the lack of X-ray crystal structures and the complexity and inconsistency of available experimental data on drug-transporter interactions.

Presentation of a structurally diverse and commercially available drug data set for correlation and benchmarking studies.

A multivariate analysis of drugs on the Swedish market was the basis for the selection of a small, physicochemically diverse set of 24 drug compounds. Factors such as structural diversity, commercial availability, price, and a suitable analytical technique for quantification were considered in the selection. Lipophilicity, pKa, solubility, and permeability across human Caco-2 cell monolayers were measured for the compiled data set. The results show that, by use of a physicochemically diverse data set, experimental responses over a wide range were obtained. The paper also shows how experimental difficulties due to the diversity of the data set can be overcome. We anticipate that this data set can serve as a benchmark set for validation of new experimental techniques or in silico models. It can also be used as a diverse starting data set for the development of new computational models.

Acyl sulfonamides as potent protease inhibitors of the hepatitis C virus full-Length NS3 (protease-helicase/NTPase): a comparative study of different C-terminals.

Synthesis and inhibitory potencies of three types of protease inhibitors of the hepatitis C virus (HCV) full-length NS3 (protease-helicase/NTPase) are reported: (i) inhibitors comprising electrophilic serine traps (pentafluoroethyl ketones, alpha-keto acids, and alpha-ketotetrazoles), (ii) product-based inhibitors comprising a C-terminal carboxylate group, and (iii) previously unexplored inhibitors comprising C-terminal carboxylic acid bioisosteres (tetrazoles and acyl sulfonamides). Bioisosteric replacement with the tetrazole group provided inhibitors equally potent to the corresponding carboxylates, and substitution with the phenyl acyl sulfonamide group yielded more potent inhibitors. The hexapeptide inhibitors Suc-Asp-D-Glu-Leu-Ile-Cha-Nva-NHSO(2)Ph and Suc-Asp-D-Glu-Leu-Ile-Cha-ACPC-NHSO(2)Ph with K(i) values of 13.6 and 3.8 nM, respectively, were approximately 20 times more potent than the corresponding inhibitors with a C-terminal carboxylate and were comparable to the carboxylate-based inhibitor containing the native cysteine, Suc-Asp-D-Glu-Leu-Ile-Cha-Cys-OH (K(i)=28 nM). The acyl sulfonamide group constitutes a very promising C-terminal functionality that allows for prime site optimization.

Hydrogen bonding descriptors in the prediction of human in vivo intestinal permeability.

Hydrogen bonding has been identified as an important parameter for describing drug permeability. Recently, we derived models for predicting intestinal permeability using the hydrogen bonding descriptors polar surface area (PSA) and number of hydrogen bond donors (HBD), and a lipophilicity descriptor [J. Med. Chem. 41 (1998) 4939]. We have now explored other types of hydrogen bonding descriptors to see if these improve the models. Both an experimental hydrogen bonding descriptor, deltalogP, and calculated descriptors, based either on semiempirical calculations or on experimentally derived hydrogen bond strength values of small molecules, were used. Principal component analyses (PCA) were performed in order to characterize the different parameters, using both a drug data set and a data set of small non-drug-like molecules for which deltalogP-values had been published. For a set of diverse drug molecules, for which human intestinal permeability data was available, a PLS-analysis was performed to study the correlation of permeability to the different hydrogen bonding parameters. No correlation could be identified between deltalogP and human intestinal permeability in this data set. However, the combination of a hydrogen bond donor descriptor, a general hydrogen bonding descriptor and a lipophilicity descriptor enabled the prediction of human intestinal permeability, whereas hydrogen bond acceptor descriptors were found to be less important. The obtained models successfully predicted the intestinal permeability values of two external data sets.

Tetrapeptides as potent protease inhibitors of Hepatitis C Virus full-length NS3 (protease-helicase/NTPase).

A library of tetrapeptides was evaluated for Hepatitis C Virus NS3 protease inhibitor activity in an in vitro assay system comprising the native bifunctional full-length NS3 (protease-helicase/NTPase) protein. Tetrapeptides with K(i) values in the high nanomolar range were identified, for example Suc-Chg-Glu-2-Nal-Cys (K(i)=0.27+/-0.03 microM) and Suc-Dif-Glu-Glu-Cys (K(i)=0.40+/-0.10 microM). Furthermore, it was shown that the inhibitory potencies are not affected significantly by assay ionic strength. As suggested by molecular modelling, potential binding interactions of the tetrapeptide inhibitors with the helicase domain might explain the data and structure-activity relationships thus obtained. Hence, we postulate that the full-length NS3 assay is a relevant system for inhibitor identification, offering new opportunities for inhibitor design.