Article pubs.acs.org/jmc
Cite This: J. Med. Chem. 2018, 61, 10106−10115
Understanding Molecular Drivers of Melanin Binding To Support Rational Design of Small Molecule Ophthalmic Drugs Paulina Jakubiak,†,‡ Michael Reutlinger,† Patrizio Mattei,† Franz Schuler,† Arto Urtti,‡,§ and Rubeń Alvarez-Sań chez*,†
Downloaded via DURHAM UNIV on November 25, 2018 at 05:59:22 (UTC). See https://pubs.acs.org/sharingguidelines for options on how to legitimately share published articles.
†
Roche Pharmaceutical Research and Early Development, Roche Innovation Center Basel, F. Hoffmann-La Roche Ltd., Grenzacherstrasse 124, 4070 Basel, Switzerland ‡ School of Pharmacy, University of Eastern Finland, 70211 Kuopio, Finland § Centre for Drug Research, Division of Pharmaceutical Biosciences, Faculty of Pharmacy, University of Helsinki, 00014 Helsinki, Finland S Supporting Information *
ABSTRACT: Binding of drugs to ocular melanin is a prominent biological phenomenon that affects the local pharmacokinetics and pharmacodynamics in the eye. In this work, we report on the development of in vitro and in silico tools for an early assessment and prediction of melanin binding properties of small molecules. A robust high-throughput assay has been established to study the binding of large sets of compounds to melanin. The extremely randomized trees approach was used to develop an in silico model able to predict the extent of melanin binding from the molecular properties of the compounds. After the last iteration of the model, strong melanin binders could prospectively be identified with 91% accuracy. On the basis of in vitro data generated for approximately 3400 chemically diverse drug-like small molecules, pronounced correlations were observed between the extent of melanin binding and the basicity, lipophilicity, and aromaticity of the compounds.
■
INTRODUCTION With the recent development of potent therapeutic agents for the treatment of neovascular ocular diseases, such as agerelated macular degeneration, diabetic retinopathy, or central retinal artery and vein occlusion, the challenging task of efficiently delivering drugs to the posterior eye segment has drawn considerable attention among pharmaceutical scientists.1 Limitations in delivering effective doses of therapeutics to the target tissues in the back of the eye are posed by various inherent ocular barriers that are specific depending upon the route of administration and the molecule type.2 The most common modes of administration to the posterior regions are via the topical, systemic, periocular, and intravitreal routes.1 Topical drug application is noninvasive with good patient convenience, but it results in low bioavailability in the posterior tissues due to the tear dilution and drainage of the eye drop, limited corneal permeability, and systemic absorption across the conjunctiva.2 On the other hand, systemic drug administration requires a large excess of dose to allow a pharmacologically relevant ocular concentration, © 2018 American Chemical Society
whereas in fact, most of the dose exposes other tissues where undesired side effects may occur. Furthermore, the entry of drugs from blood into the retina is restricted by the blood− retinal barrier.3 Direct delivery to the vitreous and retinal compartments is possible with intravitreal injections, but this approach is invasive and may result in potential adverse effects such as retinal detachment or hemorrhage. In addition, small drug molecules have vitreal half-lives of a few hours, necessitating frequently repeated injections not feasible in the clinical setting.4 Before reaching the desired site of action, the administered drug may encounter ocular membrane transporters and metabolizing enzymes that can alter its biodistribution and disposition profile.5 Despite the work done to identify and quantify drug transporters in the eye and ocular cell lines,6 there is still limited knowledge on the relevance of these transporters in the overall drug ocular tissue distribution and Received: August 13, 2018 Published: November 6, 2018 10106
DOI: 10.1021/acs.jmedchem.8b01281 J. Med. Chem. 2018, 61, 10106−10115
Journal of Medicinal Chemistry
Article
bioavailability.7 In terms of drug metabolism, there is evidence for the activity of phase 1 and phase 2 metabolic enzyme systems mainly in the ciliary body and the retinal pigment epithelium (RPE).8 Also, a high activity of the heme oxygenase and NADPH-cytochrome P450 reductase has been identified in the corneal epithelium.9 As different drugs can induce various enzyme systems in the eye, the role of ocular metabolism in the disposition of individual ophthalmic drugs should be evaluated.5b,8 The access of small molecules to the posterior part of the eye can be improved by developing strategies for sustained delivery of intravitreal drugs.7 Current advances in the fields of drug formulation, biomaterials, and nanotechnology highlight the potential of a wide variety of controlled-release systems and technologies, including liposomes, nanoparticles, biodegradable and nonbiodegradable implants, to prolong the duration of action of intravitreally administered drugs.10 Targeting biological components of the posterior eye segment is another approach that can be adopted to prolong the ocular residence time of therapeutics and achieve effective concentrations in the retinal tissues.7 It has been shown that binding of macromolecules to ocular albumin extends their vitreal half-life by 3-fold.11 In this context, melanin is a polyanionic polymer found in the pigmented cell layers of several tissues in the body. It is biosynthesized in specialized vesicles called melanosomes by enzymatic oxidation of tyrosine, followed by polymerization.12 Pigmented tissues in the eye include the uveal tract (i.e., iris, ciliary body, and choroid) and the retinal pigment epithelium. The main function of ocular melanin is the protection of the eye against light and UV radiation as well as the maintenance of the cellular homeostasis with regard to metal ions and oxidative stress.13 Binding of compounds to ocular melanin is a well-known phenomenon that affects the local pharmacokinetic and pharmacodynamic profiles of ophthalmic drugs.14 As described by Potts, chlorpromazine administered to rabbits as a single dose was selectively and persistently stored in their uveal tract during a 30 day long experiment and the uveal pigmentation turned out to be a prime factor in this accumulation.14d Similarly, Salminen et al. demonstrated that after ocular application, pilocarpine accumulated in the pigmented iris, ciliary body, choroid, and retina more than in the corresponding albino tissues of the rabbits.14g Moreover, the concept of utilizing melanin binding as a sustained-release drug depot at the back of the eye was successfully applied in animal models for the treatment of ocular neovascular disease using VEGF/PDGF receptor tyrosine kinase inhibitors (RTKI).15 In this work, the two investigated strong melanin binders showed significant retention in the ocular tissues of pigmented rats compared to albino animals due to melanin binding. The long residence time in the eye was linked to chronic efficacy in a choroidal neovascularization model highlighting the potential of drug retention in the pigmented layers of the eye to treat neovascular ocular diseases. Despite melanin binding being a known and prominent mechanism for drug distribution, the molecular properties driving melanin binding are not comprehensively understood.16 It has been reported that melanin binding is related to the basicity and lipophilicity of compounds.14a,17 Also, the number of rigid bonds and rings as well as charge-transfer interactions are associated with strong melanin binding.18 However, the data sets used in these studies were relatively
small, limited to narrow chemical spaces, and may not allow drawing of general or quantitative conclusions about the correlations between the molecular properties of the ligands and melanin binding. More systematic research on melanin binding mechanisms is needed to broadly apply melanin binding-related retention as a means to extend pharmacological action of ophthalmic drugs.16 The purpose of this work was to establish the foundations for rational drug design of strong melanin binders in ophthalmic drug discovery. We developed in vitro and in silico tools to (1) help in gaining a comprehensive understanding of main molecular drivers of melanin binding and (2) offer support to discovery research programs to identify and optimize new chemical entities (NCE).
■
RESULTS Melanin Binding Assay Development. The melanin binding assay determines the unbound fraction of a substance in melanin suspension at equilibrium. We aimed for a highthroughput (HT) setting allowing a rapid screening of large compound libraries. An initial set of 72 chemically diverse small molecules was selected for method development. These compounds were first measured in a small-scale binding assay as reported elsewhere.19 Subsequently, the same compound set was used as a reference set during the optimization of the HT method as described in the Experimental Section. Melanin binding was initially studied at compound concentrations of 1 and 10 μM. Figure 1 shows the correlation between the
Figure 1. Correlation between the experimental fraction unbound values obtained for 72 compounds at concentrations of 1 and 10 μM.
fraction unbound values obtained for the investigated 72 compounds at these two concentrations. The data show a trend of the fraction unbound (fu) being higher at 10 μM than at 1 μM, indicating that some of the compounds start to saturate the binding capacity. As a result of a compromise between the potential risk for compound precipitation and satisfactory sensitivity of the analytical system, the compound concentration of 2 μM was selected for subsequent investigations. On the basis of data from a kinetic pilot study, incubation of compounds for 2 h was adequate for the binding to reach equilibrium. Model-Based Compound Screening. Approximately 3400 compounds were tested in the HT melanin binding assay in four separate experimental rounds. Four melanin 10107
DOI: 10.1021/acs.jmedchem.8b01281 J. Med. Chem. 2018, 61, 10106−10115
Journal of Medicinal Chemistry
Article
The goal of the fourth campaign was to test the gained knowledge and to identify hit compounds, i.e., very high melanin binders. Predictions were made on the molecules from three libraries: two selected Roche ophthalmology research programs (1312 compounds), the Roche internal library of readily available screening compounds (1.8 million compounds), a database of publicly known chemical compounds20 (14 400), and a database of known ophthalmology drugs (89 compounds; Supporting Information). Experimental investigations were performed with 1055 molecules that were selected from the above-mentioned databases based on a variety of criteria aiming to explore a druggable chemical space: M between 200 and 400 g/mol, predicted medium and high kinetic solubility (≥5 μg/mL), predicted log D ≤ 3.5, and predicted high/medium permeability in artificial lipid bilayers (defined by Pe ≥ 2.0 (10−6 × cm/s) or Pe < 0.2 (10−6 × cm/s) and retention in the membrane of ≥20%). Further, strong bases (pKa > 10) were excluded due to known potential concerns (e.g., polypharmacology, phospholipidosis) and the compounds were clustered to focus only on a set of structurally diverse molecules.21 We only selected compounds predicted to be very strong melanin binders from the internal screening compound library but did not restrict the selection for the other sources mentioned above. Figure 3 indicates the distribution of all the tested molecules with regard to the represented physicochemical properties. The percentage distribution of the studied compounds with respect to their melanin binding category is depicted in Figure 4. Experimental fraction unbound values obtained for the publicly known compounds are appended in Supporting Information. Performance of the in Silico Model. To assess the machine learning model, we evaluated the performance at each iteration using the predictions generated by the model from the previous iteration. This ensures that strictly prospective predictions are analyzed and the performance is not artificially enhanced by including training data predictions in the analysis. During the exploration phase (iterations 1−3) the majority of compounds were selected to be outside of the applicability domain of the model. Figure 5A represents the experimental versus predicted output for the fourth iteration of the model. This version of the model was successful in correctly categorizing the compounds within the defined thresholds for the proposed melanin binding classes. 91% of the predicted high and very high melanin binders were confirmed (precision), and 97% of the measured high and very high melanin binders were predicted as such (recall) which translates to a Matthews correlation coefficient of 0.51 for the four category model. The validation related to the fraction unbound values obtained in vitro is depicted in more detail in Figure 5B. The 91% precision considers the performance for classification of high and very high melanin binders. The model was considered nonprecise if the experimental value fell in the medium or low categories. In fact, among the 16% outliers, 6% were still very high or high binders and thus considered as well classified. We performed a virtual trial on the druggable space of the internal compound library. The percentage distribution of the compounds with respect to their predicted melanin binding is depicted in Figure 6. As indicated, 28% of the compounds are classified as high melanin binders. Important Molecular Descriptors. A wide variety of molecular descriptors were considered in the development of the model. These were divided into three distinct sets, namely,
binding categories were defined based on the fraction unbound values at equilibrium: very high binder (0−1% free fraction), high binder (1−5% free fraction), medium binder (5−25% free fraction), and low binder (>25% free fraction). The extremely randomized trees machine learning algorithm was used to build an in silico model. The initial training data set consisted of the experimental data generated for 72 compounds in the HT melanin binding assay. However, the first version of the model showed lack of predictive power and was not able to accurately categorize the compounds. During the course of the study, new data were generated for a significant number of structurally diverse molecules that were deliberately not covered by the previous model version leading to an iterative refinement of the prediction tool as illustrated in Figure 2. In that setting, the chemical space was expanded each
Figure 2. Illustrative description of the iterative model development strategy adopted in our in vitro−in silico melanin binding studies.
time a new experiment was performed. At the same time, specific compound selection criteria were applied in the process of compound selection for each experimental campaign in order to reduce observed uncertainties and further expand the applicability domain of the trained model. The compound selection for the first experimental screening campaign performed with 250 compounds was made out of the Roche internal compound library including drug-like small molecules (reflecting the Lipinski’s rule of five) which were diversity-filtered based on ECFP fingerprint similarity. Additional filters were applied for molar mass (M ≥ 200 g/mol), calculated lipophilicity (1 ≤ log D ≤ 4), and number of aromatic rings (≥2) without distinction of fused and nonfused systems. The second round of experimental testing was performed with 1030 compounds selected based on the following criteria: in silico model predictions with high uncertainty or predicted high binding, Tanimoto structural similarity to measured compounds of 25%). Number of compounds, n = 3391.
MoKa descriptors22 (predicted pKa-derived descriptors), RDKit descriptors,23 and ErG fingerprints.24 MoKa descriptors are ionization related compound properties, e.g., maximum basic pKa, minimum acidic pKa, and cumulative and total charge. RDKit descriptors cover a broad variety of calculated molecular properties including atom counts and van der Waals surface area descriptors. ErG fingerprints are based on the reduced graph and encode a 2D pharmacophore representation of the molecule. Table 1 indicates the most important physicochemical and structural descriptors recognized by the in silico model.
■
DISCUSSION AND CONCLUSIONS Short half-life of small molecules in the ocular compartments poses the need for developing sustained drug delivery approaches. Extension of the retention time of molecules in the posterior eye tissues leading to prolonged drug responses is a highly sought-after strategy.1 Melanin binding is an important feature affecting the local pharmacokinetic and pharmacody10109
DOI: 10.1021/acs.jmedchem.8b01281 J. Med. Chem. 2018, 61, 10106−10115
Journal of Medicinal Chemistry
Article
Figure 6. Percentage distribution of the compounds with respect to their melanin binding category obtained after performing a virtual trial on the internal druggable library. The dark green color represents the very high melanin binders (fu < 1%). The light green represents the high binders (fu = 1−5%). Yellow depicts medium binders (fu = 5− 25%), and orange indicates the poorly binding compounds (fu > 25%).
Table 1. The Most Important Physicochemical and Structural Descriptors Recognized by the in Silico Melanin Binding Model
Figure 5. (A) Experimental fraction unbound values related to the prospective melanin binding predictions obtained at the fourth iteration of the developed in silico model. The dark green chart represents the predicted very high binders (f u < 1%). The light green chart indicates the high binders (fu = 1−5%). Yellow chart depicts medium binders (fu = 5−25%), and orange chart represents the low binding compounds (fu > 25%). Number of compounds, n = 1055. (B) Validation of the in silico model for the prediction of very high melanin binders ( fu < 1%) related to the fraction unbound values obtained in vitro. The dark green color indicates the percentage of compounds that were classified as very high binders prospectively by the model and in the experiment. The light green color represents the percentage of the compounds from the predicted very high category which turned out to be high binders (fu = 1−5%) in the experiment. The yellow and orange colors show the percentage of compounds predicted as very high binders which were experimentally classified as medium (fu = 5−25%) and low (fu < 25%), respectively, in the experiment.
namic profiles of ophthalmic therapeutic agents.14g,h,15,25 Binding to ocular melanin leads to significant accumulation of compounds in the pigmented tissues as previously reported for phenothiazines, tricyclic antidepressants, antimalarial quinolines, and aminoglycoside antibiotics, when comparing compound concentration in the eyes of experimental pigmented animals with their albino counterparts.14d,f,26 This prominent biological mechanism, if harnessed appropriately, may offer an attractive option for ocular drug delivery. As recently stated by Rimpelä et al., even though melanin binding has been studied for decades, the understanding of the mechanisms and the molecular drivers behind it is still very poor.16 With the aim of gaining a more comprehensive understanding of the molecular properties driving the binding of compounds to melanin, we developed in vitro and in silico tools to serve as a foundation of a framework supporting discovery programs in designing and optimizing the new chemical entities with potential to become ophthalmic drugs. We first established a high-throughput melanin binding assay
Descriptor
Importance
MOKA:isCharged RDKIT:NumAromaticRings RDKIT:SlogP_VSA8 ErG:202 MOKA:MAX_BPKA ErG:201 MOKA:TotalCharge RDKIT:SMR_VSA10 ErG:203 RDKIT:BertzCT RDKIT:RingCount ErG:281 ErG:199 RDKIT:Estate_VSA2 RDKIT:Estate_VSA4 RDKIT:BalabanJ ErG:282 ErG:200 RDKIT:FractionCSP3 ErG:280
0.023 0.020 0.016 0.013 0.0093 0.0091 0.0075 0.0069 0.0067 0.0065 0.0065 0.0063 0.0060 0.0054 0.0053 0.0053 0.0052 0.0051 0.0051 0.0050
that allowed us to profile a large number of molecules from our internal libraries in a relatively short time span. The in vitro binding studies were performed with melanin isolated from Sepia of f icinalis which we considered the most convenient for our screening purposes due to its animal origin and ready availability, which enables its use in HT screening assays and its ease of handling in the described assay setup. Nevertheless, there are known chemical and morphological differences between Sepia, synthetic and biological melanin isolated from bovine or porcine eyes,27 which could influence the binding parameters and consequently the pharmacokinetic translation from in vitro to in vivo.28 Yet the described approach allows rapid identification of strong melanin binders for further characterization in mammalian melanin binding test systems, which may be advisible in late lead optimization stages of drug discovery. The experimental testing was performed at a single compound concentration of 2 μM which was found to be a 10110
DOI: 10.1021/acs.jmedchem.8b01281 J. Med. Chem. 2018, 61, 10106−10115
Journal of Medicinal Chemistry
Article
binding capacity of melanin, serves as a reliable surrogate for binding affinity. Nevertheless, binding capacity, which can be derived from isotherm studies, is another important parameter that can drive the expected drug accumulation in the pigmented tissue. These more laborious studies are often precluded because of the solubility limitations and would be advisible to be performed as follow-up assessments with selected compounds of interest. Parallel to melanin binding testing an in silico model was developed where model performance informed compound selection for testing in an iterative way. As the data set evolved during the course of the study, a new model was trained every time new data became available. These iterative experimental and modeling efforts resulted ultimately in a robust prediction tool able to accurately categorize the compounds with regard to their extent of melanin binding. The model showed very good predictivity for high binders as 75% of compounds predicted to have fu < 1% were confirmed experimentally and another 16% of this set showed fu of 1−5%, resulting in 91% precision for identification of hit compounds. Thus, an iterative combination of focused compound testing and machine learning can be a powerful and cost-effective way to screen molecules. Our extensive data set of approximately 3400 tested compounds allowed us to address the influence of molecular properties on melanin binding more comprehensively compared to previous, more qualitative QSPR approaches,14a,18,29 yet confirmed their qualitative correlation tendencies. Strong correlations between the binding extent and the ionization category (at pH 7.4), lipophilicity, and number of aromatic rings were obtained. As expected, basic compounds turned out to be the highest binders due to their protonation enabling electrostatic interactions with anionic melanin. The molecules with basic pKa values of 5.7−9.4 displayed the lowest values of the fraction unbound. This pKa range covers the physiological pH as well as the pH at which the melanin binding assay is performed. As the basicity increases beyond pKa > 9.4, no gain in fu is observed as no more ionization effect is expected. Moreover, as the basicity increases, the compounds become more hydrophilic. As shown in our study, very hydrophilic molecules display poor melanin binding. Nonetheless, the maximal effect with regard to binding extent was reached at 2 < log D < 3, indicating no need for highly lipophilic compounds, which are usually not desirable because of other properties, i.e., poor solubility. The highest extent of melanin binding was observed for molecules with the number of aromatic rings being ≥3, pointing at the involvement of π−π interactions with the melanin structure.30 As high number of condensed heteroaromatic rings and planarity are related to solubility limitations, it is important to find a compromising solution taking into account all the contributing physicochemical properties. Inverse dependency was observed when relating these physicochemical properties to each other, which strengthened the conclusions about their direct influence on the melanin binding as presented in the Supporting Information. However, the log D and the number of aromatic rings were directly correlated with each other. Melanin is a polyanionic polymer composed of covalently bound oligomers including dihydroxyindole (DHI) and dihydroxyindolecarboxylic acid (DHICA) monomer units. These oligomers appear to have supramolecular organization forming planar sheets stacked with spacings of 3.7−4.0 Å that
Figure 7. Correlations between the extent of melanin binding and the compound’s physiochemical properties: (A) ionization category at pH 7.4; (B) basic pKa; (C) log D; (D) number of aromatic rings.
good compromise between the linearity of the binding, solubility limitations, and the sensitivity of the analytical method. This test concentration, expected to be below the 10111
DOI: 10.1021/acs.jmedchem.8b01281 J. Med. Chem. 2018, 61, 10106−10115
Journal of Medicinal Chemistry
Article
characterize π−π aromatic interactions and acidic groups, which could be involved in hydrogen bonding and electrostatic interactions with other molecules.30 Structurally similar compounds can have very different melanin binding properties as indicated in Table 2. For example, the marketed drug
melanin correlates with the concentrations in pigmented cell lines and the enrichment in pigmented ocular tissues of animals.15,18,25b,31 However, in addition to melanin binding, a lead molecule needs to show pharmacological activity on the intended therapeutic target as well as other properties supporting its drugability and developability. Foremost, the molecules need to be delivered via the desired route of administration. In this regard, solubility and permeability are important factors to be taken into account, especially if the compound is intended for topical or intravenous use. In the latter case, a high free plasma concentration is essential to enable a favorable mass gradient, and thus low unbound clearance is highly desirable to enable efficient uptake. Additionally, melanin is deposited in mildly acidic melanosomes; therefore, chemical stability of the compounds needs to be ensured over the range of weeks. Besides chemical stability, the desired compounds should rather not be substrates of ocular metabolic enzymes, e.g. hydrolases, mainly active in the cornea, ciliary body, and the retinal pigment epithelium.32 Yet little is known about the transporters in the eye and the state of knowledge is limited to allow for an efficient and relevant optimization.7 Since the melanin binders do extensively accumulate in the pigmented tissues over long periods of time, the safety profile of the molecules must be considered. It has been demonstrated that after a single oral dose, 14C-labeled chloroquine was retained in ocular tissues of pigmented rats for 6 months, which was the last measured time point included in the study.33 Similar observations were made 1 year after an intravenous bolus administered to pigmented mice, where the radioactivity was still found in the melanin-containing eye tissues.14f Concerns have been raised regarding the ocular toxicity caused by chloroquine and other melanin-binding drugs due to their long-term accumulation in pigmented tissues.34 However, in most of the cases the toxic effects displayed by these drugs could be recapitulated in both albino and pigmented animals, which indicates that the toxicity is not a direct consequence of melanin binding but an inherent property of the molecule. Also, many clinical drugs that are known to strongly bind to melanin do not cause toxicity. The occurrence of adverse effects in the eye of a given drug can depend on its physicochemical properties, pharmacokinetics, and the type of the pharmacological target and off-target mechanisms on which the drug acts.17a Besides being prone to bind to melanin and having desired physicochemical and ADME properties, a compound must show a favorable release profile from melanin to be considered as a promising candidate for sustained ocular delivery. Since only the free, unbound drug can reach the pharmacological target and exert activity, it is crucial for the melanin bound compound to display a slow and continuous release from the pigmented tissues. The elimination rate of the compound from melanosomes governed by its dissociation rate from melanin and membrane permeability is expected to be a key determinant for the ocular half-life and the steady state concentration.16 Whereas the release kinetics from the pigmented tissue is an important parameter defining the concentration−time profile, there have been very limited attempts to address this release process experimentally.19,31 In this context, there is a need to better understand the kinetic aspects of melanin binding of small molecules to warrant the opportunity to access novel treatments for the back of the eye. In summary, we established the foundation of a technology platform combining both (1) an experimental in vitro assay,
Table 2. Compound Pairs with Distinct Melanin Binding Properties
riluzole (1a) is a low melanin binder (fu = 48%), whereas the related benzothiazole (1b) is a very high binder with a fu of 0.2%. Similarly, the benzofuro[3,2-d]pyrimidine derivative (2a) is a low binder (fu = 34%), whereas replacement of the 2-methylphenyl substituent by a 4-pyridyl substituent dramatically increases melanin binding (2b, fu = 0.1%). Despite the substantial discrepancies in fu for each pair, these four compounds share characteristics often found in melanin binders (weakly basic condensed heteroaromatic rings enabling π−π aromatic interactions with melanin,30 combined with log D > 2), yet our in silico model correctly predicts these differences in melanin binding. The in silico model is also able to correctly classify compounds that might not be readily recognizable as melanin binders. For instance, the 5-HT4 agonist RS-67333 (3b) is a very high binder (fu = 0.8%) whereas the analogue lacking the butyl chain (3a) is a low binder (fu = 36%). It has been demonstrated that melanin binding can be used as a strategy to extend the retention time of a drug in the eye,15 even though medicinal chemists may object to the physicochemical properties of the drugs used in that study (pazopanib, GW-771806). Compounds like 3b show that a wide variety of pharmacophores is compatible with very strong melanin binding. Melanin binding is an easily accessible property, and it is not particularly difficult to find a drug candidate with melanin binding affinity. After performing a virtual trial on the druggable space of our internal compound library, we have seen that 28% of the molecules are recognized by the model as melanin binders. It has been reported in the literature that the extent of melanin binding measured in vitro with isolated 10112
DOI: 10.1021/acs.jmedchem.8b01281 J. Med. Chem. 2018, 61, 10106−10115
Journal of Medicinal Chemistry
Article
final compound concentration of 0.2 μM. Analytes were infused into the MS detector, and the optimized parameters such as collision energy (CE) and declustering potential (DP) as well as the traces for the parent and fragment ions were recorded. The chromatographic separation was carried out in a dual mode on a YMC-Triart C18 Narrowbore HPLC column (3 μm particle size, 50 mm × 2.0 mm; YMC, Kyoto, Japan) with the temperature set at 70 °C. Sample volumes of 1 μL were injected, and the analytes were eluted using an aqueous mobile phase A (0.5% formic acid in water) and an organic mobile phase B (acetonitrile) with a flow rate of 500 μL min−1. The following LC gradient was applied: 0.0−0.26 min, 95% solvent A, 0.26−0.90 min, from 95% solvent A to 95% solvent B, 0.90−1.23 min, 95% solvent B, 1.23−1.30 min, from 95% solvent B to 95% solvent A. Analyst 1.6.2 software (AB Sciex, Toronto, Canada) was used for LC−MS data acquisition and analysis. The fraction of unbound drug was calculated as percentage of peak area ratio of the supernatant isolated from the melanin samples compared to the corresponding PBS buffer solutions used as reference samples corrected by ISTD. In addition, the mean, standard deviation, and coefficient of variation for the replicates were calculated. In Silico Melanin Binding Predictive Modeling. The extremely randomized tree machine learning algorithm35 as implemented within the scikit-learn Python package version 0.17.136 was used to generate predictive models for melanin binding prediction. The training data set consisted of the experimental data generated in-house during this study from the HT melanin binding assay. During the course of the study the data set evolved and a new model was iteratively trained as new data became available. Physicochemical descriptors and molecular fingerprints used for model building and evaluation were calculated using RDKit37 (extended reduced graph (ErG) fingerprints,24 physicochemical descriptors, MORGAN circular-fingerprints23), and MoKa22 (predicted pKa-derived descriptors). The descriptors used for model building were selected at each iteration based on retrospective 10-fold stratified cross-validation. The following model parameters were applied: n_estimators = 300; min_samples_split = 3; max_features = 0.7; min_samples_leaf = 2; other parameters with model default. Performance was evaluated using Matthews correlation coefficient38 as implemented in scikitlearn.
allowing for an early compound screening and assessment of the melanin binding profile of small molecule drug candidates, and (2) an in silico model for predicting the extent of melanin binding based on molecular descriptors. We conducted the compound testing needed to understand the main chemical and structural factors influencing the binding to melanin. On the basis of the in vitro binding data generated for a set of approximately 3400 chemically diverse drug-like small molecules, very strong correlations were observed between the high extent of melanin binding and the compound’s physicochemical properties. Binding of compounds to pigmented ocular tissues provides a promising strategy for prolonged retention and sustained delivery of drugs for back of the eye therapy. We believe that this knowledge can be integrated in early drug discovery to support rational design of ocular therapeutics.
■
EXPERIMENTAL SECTION
Chemicals and Materials. All test compounds used in this study were provided by F. Hoffmann-La Roche Ltd. Melanin from Sepia of ficinalis, albumin from bovine serum, and dimethyl sulfoxide were purchased from Sigma-Aldrich (St. Louis, MO, USA). Phosphate buffered saline was purchased from Life Technologies (Paisley, U.K.). 96-well polypropylene microplates were obtained from Thermo Fisher Scientific Nunc (Roskilde, Denmark). 384-well polypropylene microplates were purchased from Greiner Bio-One GmbH (Frickenhausen, Germany). YMC-Triart C18 Narrowbore HPLC columns were purchased from YMC Co. (Kyoto, Japan). All other chemicals were of high purity or reagent grade. Quality control HPLC-UV data were available for 75% of the test compounds included in the study. Among them, 81% showed purity above 90% and the average purity measured was 95%. High-Throughput Melanin Binding Assay. A 1 mg/mL suspension of melanin from Sepia of f icinalis was prepared in phosphate buffered saline (PBS) at pH 6.5. As a cleaning step of the matrix, the suspended granules were spun down at 3000 rpm for 5 min. The resultant pellet was resuspended in a fresh aliquot of PBS buffer and warmed up to 37 °C for 1 h under continuous shaking. Incubations of the test compounds with melanin were performed in 96-well polypropylene microplates at a final compound concentration of 2 μM. All dispensing steps were performed on TECAN Freedom EVO liquid handling platform (TECAN Group, Männedorf, Switzerland). In brief, an amount of 98 μL of the melanin suspension was transferred into the wells and spiked with 2 μL of the test compound dissolved in DMSO at 0.1 mM. Similarly, an amount of 98 μL of PBS buffer solution containing 0.1% bovine serum albumin (BSA) was mixed with the same amount of the respective test compound. The experimental plates were generated in triplicate and incubated for 2 h at 37 °C under shaking conditions at 900 rpm. After the incubation, the samples were centrifuged at 4000 rpm for 10 min at room temperature to pellet down the bound melanin-drug complexes. Lastly, an amount of 50 μL of the resultant supernatants was transferred into 384-well polypropylene microplates and quenched with acetonitrile solution containing oxazepam and an in-house compound as internal standards (ISTD) at 100 ng/mL. Samples were stored at −20 °C until analysis. Analytical Conditions. Samples from four individual wells were pooled for analysis by liquid chromatography coupled with tandem mass spectrometry (LC−MS/MS). The analytical LC system consisted of a Shimadzu HPLC controller with two Shimadzu LC pumps equipped with a CTC PAL autosampler coupled to a hybrid triple quadrupole Sciex 4000 QTrap mass spectrometer with an ESI Turbo source operated in positive ion mode. Prior to analysis, the MS/MS methods were developed and optimized for each compound in an automated manner using DiscoveryQuant 2.1.3 optimize software (AB Sciex, Toronto, Canada). In brief, the DMSO compound stock solutions of 0.1 mM were diluted with a tuning solution containing ethanol/water (75:25) and 0.5% formic acid to a
■
ASSOCIATED CONTENT
S Supporting Information *
The Supporting Information is available free of charge on the ACS Publications website at DOI: 10.1021/acs.jmedchem.8b01281.
■
Figure S1 showing relationships between the physicochemical properties directly influencing melanin binding; Table S1 listing structural and melanin binding properties of publicly known compounds tested in the high-throughput melanin binding studies (PDF) Molecular formula strings and some data (XLSX)
AUTHOR INFORMATION
Corresponding Author
*Phone: +41 61 688 1941. Fax: +41 61 688 29 08. E-mail:
[email protected]. ORCID
Arto Urtti: 0000-0001-6064-3102 Rubén Alvarez-Sánchez: 0000-0001-7035-8022 Author Contributions
P.J., A.U., and R.A.-S. conceived the idea and directed the project. P.J. carried out the experimental investigations and supported M.R. with the iterative model development. P.J., M.R., and P.M. performed the chemical analysis of the data. R.A.-S., A.U., and F.S. discussed the results. P.J. wrote the 10113
DOI: 10.1021/acs.jmedchem.8b01281 J. Med. Chem. 2018, 61, 10106−10115
Journal of Medicinal Chemistry
Article
(5), 961−964. (b) Thrimawithana, T. R.; Young, S.; Bunt, C. R.; Green, C.; Alany, R. G. Drug delivery to the posterior segment of the eye. Drug Discovery Today 2011, 16 (5−6), 270−277. (11) Fuchs, H.; Igney, F. Binding to ocular albumin as a half-life extension principle for intravitreally injected drugs: evidence from mechanistic rat and rabbit studies. J. Ocul. Pharmacol. Ther. 2017, 33 (2), 115−122. (12) (a) Wasmeier, C.; Hume, A. N.; Bolasco, G.; Seabra, M. C. Melanosomes at a glance. J. Cell Sci. 2008, 121 (24), 3995−3999. (b) Wakamatsu, K.; Ito, S. Advanced chemical methods in melanin determination. Pigm. Cell Res. 2002, 15 (3), 174−183. (13) Hu, D. N.; Simon, J. D.; Sarna, T. Role of ocular melanin in ophthalmic physiology and pathology. Photochem. Photobiol. 2008, 84 (3), 639−644. (14) (a) Zane, P. A.; Brindle, S. D.; Gause, D. O.; O’Buck, A. J.; Raghavan, P. R.; Tripp, S. L. Physicochemical factors associated with binding and retention of compounds in ocular melanin of rats correlations using data from whole-body autoradiography and molecular modeling for multiple linear-regression analyses. Pharm. Res. 1990, 7 (9), 935−941. (b) Salazar-Bookaman, M. M.; Wainer, I.; Patil, P. N. Relevance of drug-melanin interactions to ocular pharmacology and toxicology. J. Ocul. Pharmacol. Ther. 1994, 10 (1), 217−239. (c) Ings, R. M. J. The melanin binding of drugs and its implications. Drug Metab. Rev. 1984, 15 (5−6), 1183−1212. (d) Potts, A. M. The concentration of phenothiazines in the eye of experimental animals. Invest. Ophthalmol. 1962, 1, 522−530. (e) Potts, A. M. The reaction of uveal pigment in vitro with polycyclic compounds. Invest. Ophthalmol. 1964, 3, 405−416. (f) Lindquist, N. G.; Ullberg, S. The melanin affinity of chloroquine and chlorpromazine studied by whole body autoradiography. Acta Pharmacol. Toxicol. 1972, 31 (Suppl. 2), 1−32. (g) Salminen, L.; Urtti, A.; Periviita, L. Effect of ocular pigmentation on pilocarpine pharmacology in the rabbit eye 0.1. Drug distribution and metabolism. Int. J. Pharm. 1984, 18 (1−2), 17−24. (h) Urtti, A.; Salminen, L.; Kujari, H.; Jantti, V. Effect of ocular pigmentation on pilocarpine pharmacology in the rabbit eye 0.2. Drug response. Int. J. Pharm. 1984, 19 (1), 53−61. (i) Katz, I. M.; Berger, E. T. Effects of iris pigmentation on response of ocular pressure to timolol. Surv. Ophthalmol. 1979, 23 (6), 395−398. (15) Robbie, S. J.; von Leithner, P. L.; Ju, M. H.; Lange, C. A.; King, A. G.; Adamson, P.; Lee, D.; Sychterz, C.; Coffey, P.; Ng, Y. S.; Bainbridge, J. W.; Shima, D. T. Assessing a novel depot delivery strategy for noninvasive administration of VEGF/PDGF RTK inhibitors for ocular neovascular disease. Invest. Ophthalmol. Visual Sci. 2013, 54 (2), 1490−1500. (16) Rimpela, A. K.; Reinisalo, M.; Hellinen, L.; Grazhdankin, E.; Kidron, H.; Urtti, A.; Del Amo, E. M. Implications of melanin binding in ocular drug delivery. Adv. Drug Delivery Rev. 2018, 126, 23−43. (17) (a) Leblanc, B.; Jezequel, S.; Davies, T.; Hanton, G.; Taradach, C. Binding of drugs to eye melanin is not predictive of ocular toxicity. Regul. Toxicol. Pharmacol. 1998, 28 (2), 124−132. (b) Aubry, A. F. Applications of affinity chromatography to the study of drug-melanin binding interactions. J. Chromatogr. B: Anal. Technol. Biomed. Life Sci. 2002, 768 (1), 67−74. (18) Reilly, J.; Williams, S. L.; Forster, C. J.; Kansara, V.; End, P.; Serrano-Wu, M. H. High-throughput melanin-binding affinity and in silico methods to aid in the prediction of drug exposure in ocular tissue. J. Pharm. Sci. 2015, 104 (12), 3997−4001. (19) Du, W. N.; Sun, S. M.; Xu, Y.; Li, J.; Zhao, C. H.; Lan, B. F.; Chen, H.; Cheng, L. Y. The effect of ocular pigmentation on transscleral delivery of triamcinolone acetonide. J. Ocul. Pharmacol. Ther. 2013, 29 (7), 633−638. (20) Ertl, P.; Patiny, L.; Sander, T.; Rufener, C.; Zasso, M. Wikipedia chemical structure explorer: substructure and similarity searching of molecules from Wikipedia. J. Cheminf. 2015, 7, 10. (21) Stahl, M.; Mauser, H.; Tsui, M.; Taylor, N. R. A robust clustering method for chemical structures. J. Med. Chem. 2005, 48 (13), 4358−4366.
manuscript with input from all the authors who commented on it. Notes
The authors declare no competing financial interest.
■
ACKNOWLEDGMENTS We thank Sandrine Simon, Pascal Schenk, Charles Tournillac, Isabelle Walter, and Valerie Chabrier at Roche Innovation Center Basel for technical assistance.
■ ■
ABBREVIATIONS USED f u, unbound drug fraction; HT, high throughput; QSPR, quantitative structure−property relationship REFERENCES
(1) Marsh, D. Selection of drug delivery approaches for the back of the eye: opportunities and unmet needs. In Drug Product Development for the Back of the Eye; Kompella, U. B., Edelhauser, H. F., Eds.; AAPS Advances in the Pharmaceutical Sciences Series 2; Springer: New York, NY, 2011; pp 1−20, DOI: 10.1007/978-1-4419-9920-7_1. (2) Yavuz, B.; Kompella, U. B. Ocular drug delivery. In Pharmacologic Therapy of Ocular Disease; Whitcup, S. M., Azar, D. T., Eds.; Handbook of Experimental Pharmacology Series 242; Springer International Publishing: Cham, Switzerland, 2017; pp 57− 93, DOI: 10.1007/164_2016_84. (3) Gaudana, R.; Ananthula, H. K.; Parenky, A.; Mitra, A. K. Ocular drug delivery. AAPS J. 2010, 12 (3), 348−360. (4) Del Amo, E. M.; Urtti, A. Current and future ophthalmic drug delivery systems. A shift to the posterior segment. Drug Discovery Today 2008, 13 (3−4), 135−143. (5) (a) Zhang, T.; Xiang, C. D.; Gale, D.; Carreiro, S.; Wu, E. Y.; Zhang, E. Y. Drug transporter and cytochrome P450 mRNA expression in human ocular barriers: implications for ocular drug disposition. Drug Metab. Dispos. 2008, 36 (7), 1300−1307. (b) Vadlapatla, R. K.; Vadlapudi, A. D.; Pal, D.; Mitra, A. K. Role of membrane transporters and metabolizing enzymes in ocular drug delivery. Curr. Drug Metab. 2014, 15 (7), 680−693. (c) Mannermaa, E.; Vellonen, K. S.; Urtti, A. Drug transport in corneal epithelium and blood-retina barrier: emerging role of transporters in ocular pharmacokinetics. Adv. Drug Delivery Rev. 2006, 58 (11), 1136−1163. (6) (a) Chen, P.; Chen, H.; Zang, X.; Chen, M.; Jiang, H.; Han, S.; Wu, X. Expression of efflux transporters in human ocular tissues. Drug Metab. Dispos. 2013, 41 (11), 1934−1948. (b) Mannermaa, E.; Vellonen, K. S.; Ryhanen, T.; Kokkonen, K.; Ranta, V. P.; Kaarniranta, K.; Urtti, A. Efflux protein expression in human retinal pigment epithelium cell lines. Pharm. Res. 2009, 26 (7), 1785−1791. (c) Dahlin, A.; Geier, E.; Stocker, S. L.; Cropp, C. D.; Grigorenko, E.; Bloomer, M.; Siegenthaler, J.; Xu, L.; Basile, A. S.; Tang-Liu, D. D.; Giacomini, K. M. Gene expression profiling of transporters in the solute carrier and ATP-binding cassette superfamilies in human eye substructures. Mol. Pharmaceutics 2013, 10 (2), 650−663. (7) Del Amo, E. M.; Rimpela, A. K.; Heikkinen, E.; Kari, O. K.; Ramsay, E.; Lajunen, T.; Schmitt, M.; Pelkonen, L.; Bhattacharya, M.; Richardson, D.; Subrizi, A.; Turunen, T.; Reinisalo, M.; Itkonen, J.; Toropainen, E.; Casteleijn, M.; Kidron, H.; Antopolsky, M.; Vellonen, K. S.; Ruponen, M.; Urtti, A. Pharmacokinetic aspects of retinal drug delivery. Prog. Retinal Eye Res. 2017, 57, 134−185. (8) Reddy, I. K.; Ganesan, M. G. Ocular therapeutics and drug delivery: an overview. In Ocular Therapeutics and Drug Delivery: A Multi-Disciplinary Approach; Reddy, I. K., Ed.; Technomic Publishing Company: Lancaster, PA, 1996; pp 3−29. (9) Abraham, N. G.; Lin, J. H. C.; Dunn, M. W.; Schwartzman, M. L. Presence of heme oxygenase and NADPH cytochrome-P-450 (c) reductase in human corneal epithelium. Invest. Ophthalmol. Visual Sci. 1987, 28 (9), 1464−1472. (10) (a) Geroski, D. H.; Edelhauser, H. F. Drug delivery for posterior segment eye disease. Invest. Ophthalmol. Visual Sci. 2000, 41 10114
DOI: 10.1021/acs.jmedchem.8b01281 J. Med. Chem. 2018, 61, 10106−10115
Journal of Medicinal Chemistry
Article
(22) Milletti, F.; Storchi, L.; Sforna, G.; Cruciani, G. New and original pKa prediction method using grid molecular interaction fields. J. Chem. Inf. Model. 2007, 47 (6), 2172−2181. (23) Rogers, D.; Hahn, M. Extended-connectivity fingerprints. J. Chem. Inf. Model. 2010, 50 (5), 742−754. (24) Stiefl, N.; Watson, I. A.; Baumann, K.; Zaliani, A. ErG: 2D pharmacophore descriptions for scaffold hopping. J. Chem. Inf. Model. 2006, 46 (1), 208−220. (25) (a) Nagata, A.; Mishima, H. K.; Kiuchi, Y.; Hirota, A.; Kurokawa, T.; Ishibashi, S. Binding of antiglaucomatous drugs to synthetic melanin and their hypotensive effects on pigmented and nonpigmented rabbit eyes. Jpn. J. Ophthalmol. 1993, 37 (1), 32−38. (b) Shinno, K.; Kurokawa, K.; Kozai, S.; Kawamura, A.; Inada, K.; Tokushige, H. The relationship of brimonidine concentration in vitreous body to the free concentration in retina/choroid following topical administration in pigmented rabbits. Curr. Eye Res. 2017, 42 (5), 748−753. (26) Larsson, B. S. Interaction between chemicals and melanin. Pigm. Cell Res. 1993, 6 (3), 127−133. (27) (a) Prota, G. The chemistry of melanins and melanogenesis. Fortschr. Chem. Org. Naturst. 1995, 64, 93−148. (b) Nofsinger, J. B.; Forest, S. E.; Eibest, L. M.; Gold, K. A.; Simon, J. D. Probing the building blocks of eumelanins using scanning electron microscopy. Pigm. Cell Res. 2000, 13 (3), 179−184. (28) Koeberle, M. J.; Hughes, P. M.; Skellern, G. G.; Wilson, C. G. Binding of memantine to melanin: influence of type of melanin and characteristics. Pharm. Res. 2003, 20 (10), 1702−1709. (29) (a) Kaliszan, R.; Kaliszan, A.; Wainer, I. W. Prediction of drugbinding to melanin using a melanin-based high-performance liquidchromatographic stationary-phase and chemometric analysis of the chromatographic data. J. Chromatogr., Biomed. Appl. 1993, 615 (2), 281−288. (b) Radwanska, A.; Frackowiak, T.; Ibrahim, H.; Aubry, A. F.; Kaliszan, R. Chromatographic modelling of interactions between melanin and phenothiazine and dibenzazepine drugs. Biomed. Chromatogr. 1995, 9 (5), 233−237. (30) (a) Watt, A. A. R.; Bothma, J. P.; Meredith, P. The supramolecular structure of melanin. Soft Matter 2009, 5 (19), 3754−3760. (b) Kim, Y. J.; Khetan, A.; Wu, W.; Chun, S. E.; Viswanathan, V.; Whitacre, J. F.; Bettinger, C. J. Evidence of porphyrin-like structures in natural melanin pigments using electrochemical fingerprinting. Adv. Mater. 2016, 28 (16), 3173−3180. (31) Rimpela, A. K.; Schmitt, M.; Latonen, S.; Hagstrom, M.; Antopolsky, M.; Manzanares, J. A.; Kidron, H.; Urtti, A. Drug distribution to retinal pigment epithelium: studies on melanin binding, cellular kinetics, and single photon emission computed tomography/computed tomography imaging. Mol. Pharmaceutics 2016, 13 (9), 2977−2986. (32) (a) Coupland, S. E.; Penfold, P. L.; Billson, F. A. Histochemical survey of the anterior segment of the normal human foetal and adult eye. Graefe's Arch. Clin. Exp. Ophthalmol. 1993, 231 (9), 533−540. (b) Hayasaka, S. Lysosomal enzymes in ocular tissues and diseases. Surv. Ophthalmol. 1983, 27 (4), 245−258. (33) Ono, C.; Yamada, M.; Tanaka, M. Absorption, distribution and excretion of 14C-chloroquine after single oral administration in albino and pigmented rats: binding characteristics of chloroquine-related radioactivity to melanin in-vivo. J. Pharm. Pharmacol. 2003, 55 (12), 1647−1654. (34) (a) Hobbs, H. E.; Sorsby, A.; Freedman, A. Retinopathy following chloroquine therapy. Lancet 1959, 274 (7101), 478−480. (b) Bernstein, H.; Zvaifler, N.; Rubin, M.; Mansour, A. M. The ocular deposition of chloroquine. Invest. Ophthalmol. 1963, 2, 384−392. (c) Luellman-Rauch, R. Lipidosis of the retina due to cationic amphiphilic drugs, rat. In Eye and Ear, Monographs on Pathology of Laboratory Animals; Jones, T. C., Mohr, U., Hunt, R. D., Eds.; Springer: Berlin, 1991; pp 87−92, DOI: 10.1007/978-3-642-766404_16. (35) Geurts, P.; Ernst, D.; Wehenkel, L. Extremely randomized trees. Mach. Learn. 2006, 63 (1), 3−42.
(36) Pedregosa, F.; Varoquaux, G.; Gramfort, A.; Michel, V.; Thirion, B.; Grisel, O.; Blondel, M.; Prettenhofer, P.; Weiss, R.; Dubourg, V.; Vanderplas, J.; Passos, A.; Cournapeau, D.; Brucher, M.; Perrot, M.; Duchesnay, E. Scikit-learn: machine learning in Python. J. Mach. Learn. Res. 2011, 12, 2825−2830. (37) RDKit: Open-Source Cheminformatics and Machine Learning. http://www.rdkit.org. (38) Matthews, B. W. Comparison of the predicted and observed secondary structure of T4 phage lysozyme. Biochim. Biophys. Acta, Protein Struct. 1975, 405 (2), 442−451.
10115
DOI: 10.1021/acs.jmedchem.8b01281 J. Med. Chem. 2018, 61, 10106−10115