A high content screening assay to predict human drug-induced liver injury during drug discovery

Journal of Pharmacological and Toxicological Methods 68 (2013) 302–313

Contents lists available at ScienceDirect

Journal of Pharmacological and Toxicological Methods journal homepage: www.elsevier.com/locate/jpharmtox

Original article

A high content screening assay to predict human drug-induced liver injury during drug discovery Mikael Persson, Anni F. Løye, Tomas Mow, Jorrit J. Hornberg ⁎ Department of Exploratory Toxicology, Non-Clinical Safety Research, H. Lundbeck A/S, Ottiliavej 9, 2500 Valby, Denmark

a r t i c l e

i n f o

Article history: Received 19 April 2013 Accepted 1 August 2013 Keywords: Cytotoxicity Drug-induced liver injury Hepatotoxicity High content screening Methods Mitochondrial toxicity Predictive toxicology Risk assessment

a b s t r a c t Introduction: Adverse drug reactions are a major cause for failures of drug development programs, drug withdrawals and use restrictions. Early hazard identification and diligent risk avoidance strategies are therefore essential. For drug-induced liver injury (DILI), this is difficult using conventional safety testing. To reduce the risk for DILI, drug candidates with a high risk need to be identified and deselected. And, to produce drug candidates without that risk associated, risk factors need to be assessed early during drug discovery, such that lead series can be optimized on safety parameters. This requires methods that allow for medium-to-high throughput compound profiling and that generate quantitative results suitable to establish structure–activity-relationships during lead optimization programs. Methods: We present the validation of such a method, a novel high content screening assay based on six parameters (nuclei counts, nuclear area, plasma membrane integrity, lysosomal activity, mitochondrial membrane potential (MMP), and mitochondrial area) using ~100 drugs of which the clinical hepatotoxicity profile is known. Results discussion: We find that a 100-fold TI between the lowest toxic concentration and the therapeutic Cmax is optimal to classify compounds as hepatotoxic or non-hepatotoxic, based on the individual parameters. Most parameters have ~50% sensitivity and ~90% specificity. Drugs hitting ≥2 parameters at a concentration below 100-fold their Cmax are typically hepatotoxic, whereas non-hepatotoxic drugs typically hit b 2 parameters within that 100-fold TI. In a zone classification model, based on nuclei count, MMP and human Cmax, we identified an area without a single false positive, while maintaining 45% sensitivity. Hierarchical clustering using the multi-parametric dataset roughly separates toxic from non-toxic compounds. We employ the assay in discovery projects to prioritize novel compound series during hit-to-lead, to steer away from a DILI risk during lead optimization, for risk assessment towards candidate selection and to provide guidance of safe human exposure levels. © 2013 Elsevier Inc. All rights reserved.

1. Introduction Clinical safety and non-clinical toxicity are major causes for failure of drug development programs in the pharmaceutical industry. It is estimated that toxicology and clinical safety accounted for N30% of all attrition after first-in-human dosing in the year 2000 (Kola & Landis, 2004), and ~20% of late development programs in 2007–2010 (Arrowsmith, 2011a, 2011b). Though serious adverse drug reactions that impact drug development occur in different ways and organs, drug-induced liver injury (DILI) is one of the most common reasons for development failures (Olson et al., 2000). It has also been one of the most common Abbreviations: AC50, activity concentration 50%; ALT, alanine transferase; Cmax, maximum total concentration in plasma; CCCP, carbonyl cyanide m-chlorophenyl hydrazine; DILI, drug-induced liver injury; DMSO, dimethyl sulfoxide; FBS, fetal bovine serum; HCS, high content screening; IC50, inhibitory concentration 50%; LEC, lowest efficacious concentration; MMP, mitochondrial membrane potential; PBS, phosphate buffered saline; SD, standard deviation; SEM, standard error of the mean; TI, therapeutic interval. ⁎ Corresponding author. Tel.: +45 36433783. E-mail addresses: [email protected] (M. Persson), [email protected] (A.F. Løye), [email protected] (T. Mow), [email protected] (J.J. Hornberg). 1056-8719/$ – see front matter © 2013 Elsevier Inc. All rights reserved. http://dx.doi.org/10.1016/j.vascn.2013.08.001

reasons for safety-related withdrawals of marketed drugs (Fung et al., 2001; Stevens & Baker, 2009) and has led to black box warnings and use restrictions. It is estimated that drug-induced liver injury accounts for more than half of all acute liver failures, and though the majority of those cases can be attributed to an overdose of paracetamol, more than 10% is attributed to other drugs (Holt & Ju, 2006). The causes for DILI are only partially understood. A limited number of drugs on the market (e.g. paracetamol) produce a so-called predictable hepatotoxicity (Kaplowitz, 2004), which is dose dependent and can often be identified in preclinical animal toxicology studies. A substantial number of marketed drugs produce a so-called idiosyncratic DILI which occurs infrequently (typically 1 in 10,000 patients or less) and is thus not identified until a large number of patients have been treated, often after regulatory approval (Kaplowitz, 2004, 2005; Lee, 2003). Though some patient-related risk factors for idiosyncratic DILI have been identified, e.g. age, gender, lifestyle, general health and genetic polymorphisms (Chalasani & Bjornsson, 2010), it is currently not possible to quantify the risk for the individual patient. In order to reduce the risk for DILI, it is therefore essential to identify and deselect those drug candidates that have the propensity to cause

M. Persson et al. / Journal of Pharmacological and Toxicological Methods 68 (2013) 302–313

hepatotoxicity. In fact, in order to produce drug candidates that do not have that risk associated, we argue that drug-related risk factors need to be assessed early in drug discovery programs, such that lead series can be optimized on safety parameters. This requires methods that allow for medium to high throughput compound profiling and quantitative results that are suitable to generate structure–activity relationships during lead optimization programs. Idiosyncratic drug reactions are per definition hard to predict. Animal models and conventional cytotoxicity assays have modest sensitivity (50% and 25%, respectively) to identify hepatotoxic drugs (Olson et al., 2000). Image-based high content screening (HCS) using organspecific cellular models has emerged as a potential solution to the problem. Several drug-related risk factors, such as mitochondrial toxicity, oxidative stress, intracellular glutathione, as well as cell loss and nuclear factors, are sensitive predictors of DILI (Lin & Will, 2011; O'Brien et al., 2006; Xu et al., 2008), and can be readily assessed by high content imaging on a cell by cell basis. Here we present a novel HCS assay for the prediction of human DILI, which can be employed during drug discovery to identify risk factors early, aid optimization and ultimately provide guidance on safe human dose levels. In the assay, called the Quadprobe assay, we employ four fluorescent probes to analyze six parameters related to cytotoxicity, mitochondrial toxicity, and changes in lysosomal activity. The assay was validated for its ability to predict DILI by screening ~100 drugs of which the clinical hepatotoxicity profile is known. It is compatible with formaldehyde fixation of cells to allow relatively high throughput. Compounds are screened in dose response such that quantitative data can be fed back to drug discovery teams. The multi-parametric analysis applied provides an integrated view on cellular toxicity for standardized predictions of the propensity of new drugs to cause hepatotoxicity and give insight into possible toxic modes of action. 2. Methods

303

assessed by visual examination of stock solutions, again after dilution series had been transferred into media, and then finally in the microscope after incubation with cells for 24 or 72 h. No visible precipitation was noted for any of the 102 test compounds. Eight wells with 1% DMSO were kept as negative controls, and 30 μM chlorpromazine and 100 μM troglitazone were used as positive controls in four wells each. After incubation for 24 or 72 h with compound, the culture medium was aspirated and carefully replaced with 100 μl new culture medium with a cocktail of 50 nM MitoTracker Orange, 50 nM LysoTracker Green, 1 μM TOTO-3 and 1 μg/ml Hoechst 33342 (all from Invitrogen). The plates were incubated for 60 min at 37 °C and then gently washed thrice with pre-warmed PBS, leaving 100 μl after the last wash step. For fixation, 100 μl/well 4% paraformaldehyde (Sigma) was added followed by 30 min incubation at room temperature. The cells were then washed three times with PBS and were then ready for imaging. Six image fields per well were acquired using the ThermoFisher/Cellomics ArrayScan VTI HCS Reader with an LED light engine, and using an LD Plan Neofluar Apochromat 20× objective and a BGFR filter cube (Cellomics, Pittsburgh, PA, USA). The nuclei fluorescence of Hoechst 33342 was used for automatic focusing. Fig. 1 shows typical images from the immunocytochemistry. Image analysis was performed using the Compartmental Analysis V4 module of the ThermoFisher/Cellomics BioApplications Analysis Software (Cellomics). Background removal algorithms were used on all acquired channels. The parameters nuclei counts and nuclear area were acquired from the 386/23 nm channel as “Selected object counts per valid field” and “Mean object area”. Lysosomal activity was analyzed from the 486/20 nm channel as “Mean ring average intensity”, and mitochondrial parameters were analyzed from the 549/15 nm channel as “Mean ring average intensity” for MMP and “Mean ring spot total area” for mitochondrial area. Plasma membrane integrity was analyzed from the 650/13 nm channel as “Mean circ average intensity”. Reported results are mean values from six image fields per well.

2.1. Cell culturing 2.3. Cellular toxicity profiling HepG2 cells were grown in 75 cm2 cell culture flasks (NuncA/S, Roskilde, Denmark) in Dulbecco's Modified Eagle Medium (DMEM) without glucose (Invitrogen, Carlsbad, CA) and supplemented with the following to reach a final concentration of 10% FBS (Invitrogen), 1 mM Sodium Pyruvate (Invitrogen), 2 mM L-glutamine (Invitrogen), 10 mM galactose (Sigma Aldrich), 5 mM Hepes (Invitrogen) and 100 IU/ml Penicillin/Streptomycin (Invitrogen). A single stock of HepG2 cells (passage 3–29) was used for the experiments. Cells were passaged every 3–4 days by briefly rinsing the cell mono-layer twice with prewarmed (37°C) phosphate buffered saline (PBS; Invitrogen), with subsequent addition of 0.05% Trypsin-EDTA solution (Biochrom AG, Berlin, Germany) and incubation for 5–7 min at 37 °C. Once the cell-layer was dispersed, the Trypsin-EDTA solution was deactivated by adding 5 ml of complete growth medium, and cells were aspirated using a syringe to assure single cell formation, and dispersed into 75 cm2 cell culture flasks or 96-well plates. The cells were kept in an incubator at 37 °C with a humidified atmosphere with 5% CO2 and 95% air. 2.2. High content screening Quadprobe assay For toxicity screening, HepG2 cells were plated out to a final density of 15,000 cells/well (for 24 hour treatment) or 6,000 cells/well (for 72 hour treatment) in Poly-L-lysine coated (Sigma) thin bottom imaging plates (Nunc). The cells were incubated overnight in an incubator at 37 °C with a humidified atmosphere with 5% CO2 and 95% air. Test compounds were then added the following day to a final dose range of 0.01–100 μM at a final concentration of 1% (v/v) DMSO in supplemented DMEM in each well. All drugs and chemicals were acquired from the Lundbeck compound repository or from Sigma–Aldrich and were of the highest quality possible. Compound precipitation was

For each parameter, the normal population was defined as the mean ± 3 standard deviations (SDs) for the negative controls. The lowest concentration of a test compound generating a value outside the normal population, considered as a toxic cellular response, was defined as a lowest efficacious concentration (LEC) for toxicity. Compounds were evaluated for cellular toxicity in the assay according to these LEC values, which were calculated for all parameters, and additionally, IC50 values for nuclei counts, and AC50 values (encompassing either IC50 or EC50, dependent on if there was a loss or gain of signal, respectively) for MMP, were calculated using a 4-parameter logistic model. Curves were only fitted if there was at least a 50% change from base-line (1% DMSO) within the test concentrations. In order to focus on the initial toxic response, values were knocked out from calculations in cases of hormesis effects or bi-phasic toxic responses (e.g. an initial druginduced hyperpolarization of the MMP turning into a hypopolarization as a result of secondarily induced cell death). Compounds which did not generate data points outside the normal population were considered as non-toxic in the assay. Fig. 2 provides an example of cellular toxicity profiling for a single parameter, in this case MMP. Compounds were clustered based on LEC and IC50 values. Hierarchical clustering was performed in Spotfire using unweighted averages and Euclidean distance similarity measures. In order to assess reproducibility, a pilot experiment with seven drugs was performed in three independent replicates. The AC50 values never varied more than ±10 μM, and LEC values were always within a factor of two, highlighting robustness of the assay (results not shown). Additionally, throughout the course of screening, reference compounds were spiked to screening plates regularly, and no assay drift has been observed.

304

M. Persson et al. / Journal of Pharmacological and Toxicological Methods 68 (2013) 302–313

Fig. 1. Typical images from the Quadprobe assay. Four different probes were used for assessing cell health. Hoechst 33342 (blue) was used for nuclei counts and nuclear size. Lysotracker Green (A, green) stains the lysosomes and was used for investigating lysosomal activation. MitoTracker Orange (B, red) shows typical mitochondrial morphology and was used for assessing the mitochondrial membrane potential (MMP). A loss of MMP is seen as a decrease of incorporated MitoTracker Orange as can be seen for the apoptotic cell with fragmented DNA indicated by an arrow in Fig. 1B. TOTO-3 (C, red) stains DNA of cells with compromised membrane integrity, signifying necrotic or late apoptotic cells as indicated by the arrow in Fig. 1C. Images were acquired using a 20× objective.

Toxicity Database (http://www.livertox.nih.gov), and FDA drug labels. Classification was only performed according to hepatotoxicity profile, and consequently, signs of other toxicities were not taken into account. The marketed drugs were selected to cover a diverse set of therapeutic areas. A selection of toxic chemicals was included as assay controls for cellular and/or mitochondrial toxicity.

2.4. Hepatotoxicity classification of drugs Drugs were classified into three groups for their potential to cause clinical drug-induced liver injury in humans as described before (O'Brien et al., 2006). Severe human hepatotoxic compounds were defined as producing N1% increase (defined as N3-fold above normal controls) in frequency of serum alanine transferase (ALT) plus two of either jaundice, N 3 reports of liver failure, or containing a warning for adverse liver effects on the FDA label. Moderate human hepatotoxic compounds were defined as producing 0.1–1% frequency of ALT plus either jaundice, or a warning for adverse liver effects on the FDA label. Non-hepatotoxic drugs were defined as those causing b 1% frequency of increases in ALT and showing no clinical signs of hepatotoxicity. The classification of the drugs used in the validation can be seen in Table 1. Information on hepatotoxicity and therapeutic total Cmax (free plus bound concentration) was extracted from the paper of O'Brien and colleagues (O'Brien et al., 2006), PharmaPendium (http://www. pharmapendium.com), the U.S. National Library of Medicine's Liver

3. Results 3.1. HCS Quadprobe assay for cellular toxicity profiling In order to analyze compound-induced cellular toxicity, we developed an HCS assay, called the Quadprobe assay since it employs four fluorescent probes (Fig. 1), using the HepG2 hepatocarcinoma cell line grown in galactose medium as a model system. The fluorescent probes were chosen based on their characteristics and compatibility with paraformaldehyde fixation, which enables sufficiently high throughput to use the assay in early drug discovery. The assay assesses the parameters

Mitochondrial Membrane Potential 140

Relative mitochondrial membrane potential [%]

Normal population 120

100

DMSO Troglitazone Chlorpromazine

80

Disulfiram LEC for Disulfiram

Valproate

60

Caffeine Calculated AC50 for Disulfiram

Amisulpride

40

Paracetamol Mean + 3SD Mean - 3SD

20

0 0,01

0,1

1

10

100

-20

µM compound Fig. 2. Typical results from a single parameter (mitochondrial membrane potential). Compounds/drugs were tested from 0.01 μM to a top concentration of 100 μM at a fixed concentration of 1% DMSO. The lowest efficacious concentration (LEC) for each compound is calculated as the lowest concentration to cause a deviation from the normal population defined as the mean and three times the standard deviation from eight wells of DMSO controls, indicated by the red dashed lines. As an example, the MMP LEC value for Disulfiram (#11) is 13 μM. Data is expressed as a percentage of DMSO controls and as mean value ± SEM per cell based on six images per well. Numbers associated with the drugs correspond to the drug numbers in Table 1.

M. Persson et al. / Journal of Pharmacological and Toxicological Methods 68 (2013) 302–313

nuclear counts, nuclear area, membrane integrity, lysosomal activity, mitochondrial membrane potential (MMP), and mitochondrial area. Data were derived in terms of IC50 (or AC50) values, and LEC values generated from deviations from the normal population, as described in Methods and seen in Fig. 2. The extent, to which the applied Quadprobe assay can predict human liver toxicity, was determined with a validation set of ~100 drugs with a known clinical hepatotoxicity profile (classified as severely, moderately or non-hepatotoxic, see Methods for details). It is recommended to screen drugs at 30–100× the human therapeutic Cmax (Lin & Will, 2011; O'Brien et al., 2006; Xu et al., 2008), however the Cmax values of compounds are typically not known in the early phase of pharmaceutical research. Therefore, all compounds in our study were assayed at a fixed concentration range of 0.01–100 μM at a fixed DMSO concentration of 1%, to mimic the in vitro toxicology screening practice during early drug discovery. A full overview of all results from the validation set is depicted in Table 1. The assay accurately identifies drugs known to be cytotoxic, such as amodiaquine and chlorpromazine (Greer, Barber, Eakins, & Kenna, 2010; Zhang, Solomon, Hu, Ulibarri, & Lee, 2008), and drugs affecting the MMP such as efavirenz (Blas-Garcia et al., 2010), and discriminates those from non-toxic compounds such as carbidopa, within the tested concentration range (Fig. 3). The data generated for the drugs in the validation set, can be used to effectively predict whether novel chemical series, lead compounds or drug candidates have the propensity to induce hepatotoxicity in humans. We will now describe three models for data analysis to illustrate this: i) estimation of a safe drug exposure (Cmax) limit based on the individual parameters, ii) a zone classification model based on MMP, nuclei count and exposure (Cmax) and, iii) clustering of compounds with similar toxicity profiles based on all parameters. 3.2. Estimation of a safe exposure (Cmax) limit based on individual parameters Others have shown that a threshold therapeutic interval (TI) of 100 between the minimum toxic concentration (LEC in this study) and the therapeutic Cmax can be used to classify a compound as hepatotoxic or non-toxic (Lin & Will, 2011; O'Brien et al., 2006; Xu et al., 2008). In other words, hepatotoxic drugs typically produce LEC values for cellular toxicity lower than 100-fold their Cmax, whereas non-hepatotoxic drugs typically produce LEC values higher than 100-fold their Cmax. Reversely, a safe drug exposure (Cmax) threshold can be deduced from the LEC values: a drug should not induce hepatotoxicity if the Cmax is more than 100-fold lower than the LEC value. In order to determine an optimal threshold TI in our Quadprobe assay, we calculated sensitivity and false positive frequency for a wide range of TI values (1–1000), for each parameter in the assay for 24 h or 72 h incubation (Fig. 4). At higher TI threshold values more hepatotoxic compounds are classified correctly (high sensitivity), and at lower TI threshold values more non-hepatotoxic compounds are classified correctly (low number of false positives). A low false positive rate is important in order to avoid discarding promising compounds too early in drug discovery projects. Based on the curves in Fig. 4, we concluded that a 100-fold TI from the LEC for the given in vitro toxicity endpoint to the human therapeutic Cmax, provides an acceptable sensitivity, i.e. hepatotoxic drugs are correctly identified, while maintaining a low false positive rate, i.e. non-toxic compounds are correctly identified for the individual parameters (Table 2). Incubation with compound for 72 h increases the sensitivity for the parameters nuclei counts and membrane integrity, compared to 24 h incubation. Most drugs that hit two or more parameters at a concentration below 100-fold their therapeutic Cmax are hepatotoxic, whereas most non-hepatotoxic drugs hit less parameters within that 100-fold TI (Fig. 4I). If drug exposure is not taken into account (i.e. compounds are scored solely on the basis of whether they hit a parameter within the test concentration range or

305

not), the sensitivity of the assay increases in general. However, it also generates higher false positive rates (Table 2). 3.3. A zone classification model based on MMP, nuclei count and Cmax Compound-induced changes in nuclei counts and MMP have previously been identified as sensitive predictors of human drug-induced liver injury (Dykens, Marroquin, & Will, 2007; O'Brien et al., 2006; Russmann, Kullak-Ublick, & Grattagliano, 2009; Xu et al., 2008). Interestingly, these parameters also produce very few false positives when using IC50 or AC50 values (Fig. 4A and D). We therefore hypothesized that factoring in both parameters simultaneously would add strength to safety predictions. Indeed, in a three dimensional graph depicting nuclei count IC50, MMP AC50 and the human therapeutic Cmax, many hepatotoxic drugs clearly separate from the others (Fig. 5A). Many hepatotoxic drugs affect either one or both of these parameters. Based on this, we constructed a simple model accounting for both parameters by multiplying both factors (nuclei count IC50 × MMP AC50) and plotting the resulting value against the therapeutic Cmax of the drugs (Fig. 5B). The model identifies a zone that only harbors hepatotoxic drugs. This simple zone classification model, which combines nuclei counts and MMP data from the Quadprobe assay with therapeutic Cmax values, predicts human DILI with a sensitivity of 45% and a no false positives using 24 h incubation with compound, and a sensitivity of 51% and a false positive rate of 6% using 72 h incubation with compound. 3.4. Clustering of compounds with similar toxicity profiles based on ally parameters Next, we were interested to see whether the data on all parameters in the Quadprobe assay would allow for clustering drugs with similar toxicity. We performed hierarchical clustering of compounds in the validation set based on the LEC, IC50 and AC50 values (Fig. 6). The two main resulting clusters roughly separate hepatotoxic drugs from the non-hepatotoxic drugs. In fact, one of the two main clusters contains hepatotoxic drugs only (withdrawn, severe and moderate). The other main cluster contains all non-toxic drugs as well as those toxic drugs the assay does not pick up. Although the clustering approach does not take drug exposure into account, it can accurately group compounds with similar toxicity profiles. 4. Discussion In this study we present a novel high content screening assay, which is highly predictive for human drug-induced liver injury (DILI). It can be employed in drug discovery projects to optimize compound series on their safety profile, and provide a risk assessment tool towards candidate selection. This “Quadprobe assay” assesses six parameters based on general cytotoxicity, mitochondrial toxicity and lysosomal dysfunction. In addition to cytotoxicity, mitochondrial toxicity is known to be an important contributor to DILI (Begriche, Massart, Robin, BorgneSanchez, & Fromenty, 2011; Boelsterli & Lim, 2007; Chan, Truong, Shangari, & O'Brien, 2005; Labbe, Pessayre, & Fromenty, 2008; Pessayre et al., 2012; Russmann et al., 2009) and it has been recommended to screen for this as part of early safety assessments (Dykens & Will, 2007; Dykens et al., 2007). Lysosomal dysfunction is less well-known as a predictive parameter for DILI, but it has been used in an HCS assay to identify lysosomotropic compounds (Nadanaciva et al., 2011). Such compounds reduce lysosomal activity and thereby disturb the regular breakdown of organelles and aggregated proteins, which can lead to an increase in liver size (Czaja, 2011; Schneider, Korolenko, & Busch, 1997). An increase in lysosomal activity (i.e. increased breakdown of e.g. organelles) on the other hand, may indicate that a compound induces damage to e.g. organelles. The sensitivity of the lysosomal activity parameter to identify hepatotoxic drugs is similar to the

306 Table 1 Multi-parametric data for 24 h and 72 h compound incubation. The compounds are divided into hepatotoxicity categories as described under Methods. The given Cmax values are total Cmax (free plus bound concentrations) at human therapeutic doses. Information on hepatotoxicity and therapeutic Cmax was extracted from the paper of O'Brien et al. (2006), PharmaPendium (http://www.pharmapendium.com), the U.S. National Library of Medicine's Liver Toxicity Database (http://www.livertox.nih.gov/), and FDA drug labels. #

Compound

Main indication

Nuclei counts LEC [μM]

Nuclei counts IC50 [μM]

Nuclei area LEC [μM]

Membrane integ. LEC [μM]

Lysosomal activity LEC [μM]

MMP LEC [μM]

MMP AC50 [μM]

Mito. Area LEC [μM]

Cmax [μM]

Antituberculosis Antiarrythmia Antidepressant Cancer Antipsychotic Immunosuppressant Testosterone derivative Muscle relaxant Pain, fever, NSAID Antiretroviral, HIV Drug abuse Antiretroviral, HIV Anticancer Anticonvulsant Antiviral, Hepatitis B Antipsychotic Antidepressant Pain, fever, NSAID Antifungal Antifungal Pain, fever, NSAID Blood pressure Antiviral, Hepatitis B Immunosuppressant Antimetabolite, antifolate Sympatholytic, antihypertensive Antibiotic Antibiotic Antibacterial Pain, fever Pain, fever, NSAID Hyperthyroidism Antipsychotic Antiretroviral, HIV Pain, fever, NSAID Anticonvulsant Asthma

24 h N100 30 67 N100 100 20 N100 N100 N100 N100 13 44 N100 N100 100 N100 100 N100 N100 20 N100 N100 N100 N100 N100 N100 N100 N100 N100 N100 N100 N100 44 N100 N100 10 N100

72 h N100 10 10 30 13 10 N100 67 N100 N100 20 20 1 N100 13 100 30 100 1 13 N100 44 N100 10 N100 N100 N100 10 N100 N100 N100 N100 100 100 N100 N100 N100

24 h N100 29 54 N100 90 19 N100 N100 N100 N100 47 40 N100 N100 N100 N100 90 N100 N100 24 N100 N100 N100 N100 N100 N100 N100 100 N100 N100 N100 N100 74 N100 N100 N100 N100

72 h N100 11 14 30 22 10 N100 94 N100 N100 17 15 1 N100 29 N100 22 N100 1 12 N100 67 N100 5 N100 N100 N100 5 N100 N100 N100 N100 16 N100 N100 N100 100

24 h N100 44 100 N100 44 20 N100 N100 100 N100 20 44 1 N100 13 N100 44 N100 10 44 N100 N100 20 20 N100 N100 N100 N100 N100 13 N100 N100 10 30 N100 N100 N100

72 h N100 10 20 20 100 10 N100 N100 N100 N100 20 13 1 N100 13 67 30 100 10 10 N100 100 N100 10 N100 100 N100 10 N100 N100 N100 N100 44 100 N100 N100 67

24 h N100 10 67 44 67 10 N100 N100 N100 N100 20 67 44 N100 N100 30 44 N100 20 30 N100 N100 N100 N100 N100 N100 N100 N100 N100 20 N100 100 13 N100 N100 10 N100

72 h N100 13 13 13 10 13 N100 N100 100 N100 44 0.1 1 N100 20 10 0.1 N100 0.1 20 N100 44 N100 10 100 N100 N100 10 N100 N100 N100 N100 10 100 N100 N100 N100

24 h 20 10 10 30 0.01 20 N100 N100 10 N100 30 20 10 30 N100 10 67 N100 10 30 20 20 N100 13 10 44 N100 20 N100 N100 N100 67 100 N100 N100 N100 67

72 h N100 20 30 44 10 10 N100 N100 N100 N100 20 30 20 N100 N100 N100 100 N100 10 20 N100 67 N100 100 N100 N100 N100 10 N100 67 N100 N100 44 N100 N100 N100 N100

24 h N100 0.1 67 30 44 10 N100 100 100 N100 30 10 10 N100 20 10 67 N100 1 13 100 N100 44 1 N100 10 N100 13 N100 N100 N100 100 100 N100 N100 N100 44

72 h N100 1 10 N100 100 13 67 N100 N100 N100 44 20 1 N100 10 10 67 N100 10 1 N100 44 N100 N100 N100 44 100 13 N100 N100 N100 N100 44 N100 N100 N100 N100

24 h N100 33 44 80 47 N100 N100 N100 N100 N100 25 17 33 N100 N100 N100 50 N100 8 28 N100 N100 N100 11 N100 N100 N100 12 N100 N100 N100 N100 73 N100 N100 N100 100

72 h N100 2 3 48 50 26 N100 N100 N100 N100 31 21 N100 N100 N100 3 64 N100 4 24 N100 40 N100 N100 N100 100 N100 13 N100 N100 N100 N100 N100 N100 N100 N100 N100

24 h N100 1 67 67 100 10 N100 100 44 N100 30 20 10 N100 30 13 30 N100 1 20 100 N100 20 1 N100 100 N100 20 N100 20 N100 N100 44 N100 N100 N100 100

72 h N100 10 1 20 100 10 N100 N100 N100 N100 30 30 1 N100 10 10 10 N100 10 10 N100 30 N100 44 N100 100 N100 10 N100 N100 N100 N100 10 N100 N100 N100 N100

49 0.81 0.3 1.9 0.69 0.2 0.16 7.9 4.2 12 5.4 13 17 42 1 0.053 0.6 6 0.4 7 7 0.4 17 1 0.02 11 8 6 1 130 5 42 0.039 4 19 540 17

Moderate hepatotoxicants 38 Aspirin 39 Chlorpromazine 40 Clofibrate 41 Estradiol 42 Fenofibrate 43 Fluoxetine 44 Furosemide 45 Ibuprofen 46 Ketoprofen 47 Naproxen 48 Nimodipine 49 Paclitaxel

Pain, fever Antipsychotic Cholesterol Sex hormone Cholesterol Antidepressant Diuretic Pain, fever Pain, fever, NSAID Pain, fever, NSAID Blood pressure, brain ischemia Cancer, anti-mitotic

N100 10 N100 44 100 30 N100 N100 N100 N100 67 0.1

N100 13 N100 30 100 10 N100 N100 N100 N100 20 0.1

N100 26 N100 43 100 22 N100 N100 N100 N100 30 33

N100 12 N100 17 72 12 N100 N100 N100 N100 19 0.05

N100 30 N100 44 67 44 N100 N100 N100 N100 67 1

N100 10 N100 30 100 20 N100 100 N100 N100 13 0.1

N100 13 N100 44 N100 20 N100 N100 N100 N100 20 13

N100 44 N100 30 67 13 N100 100 N100 N100 13 0.1

1 30 N100 44 N100 10 N100 N100 10 1 44 1

N100 20 N100 30 N100 10 N100 N100 N100 N100 67 1

30 30 N100 44 0.1 10 N100 N100 N100 100 20 0.1

N100 44 N100 67 30 13 N100 N100 N100 N100 44 10

N100 25 N100 31 N100 21 N100 N100 N100 N100 24 1

N100 30 N100 29 N100 17 N100 N100 N100 N100 28 2

N100 10 N100 30 20 10 N100 N100 N100 N100 30 1

N100 20 N100 30 100 10 N100 44 N100 N100 44 10

1650 0.5 470 0.0006 25 0.048 16 250 15 0.2 0.024 0.02

M. Persson et al. / Journal of Pharmacological and Toxicological Methods 68 (2013) 302–313

Severe hepatotoxicants 1 Aminosalicylic acid 2 Amiodarone 3 Amitriptyline 4 Bicalutamide 5 Clozapine 6 Cyclosporin 7 Danazol 8 Dantrolene 9 Diclofenac 10 Didanosine 11 Disulfiram 12 Efavirenz 13 Etoposide 14 Felbamate 15 Fialuridine 16 Haloperidol 17 Imipramine 18 Indomethacin 19 Itraconazole 20 Ketoconazole 21 Ketorolac 22 Labetalol 23 Lamivudine 24 Mercaptopurine 25 Methotrexate 26 Methyldopa 27 Minocycline 28 Nitrofurantoin 29 Novobiocin 30 Paracetamol 31 Piroxicam 32 Propylthiouracil 33 Risperidone 34 Stavudine 35 Sulindac 36 Valproate 37 Zileuton

50 51

67 30

10 30

33 30

3 12

67 13

10 30

10 13

10 10

13 30

13 30

67 13

1 1

50 2

28 3

67 10

10 1

Withdrawn hepatotoxicants 52 Alpidem 53 Amineptine 54 Amodiaquine 55 Bromfenac 56 Flutamide 57 Iproniazid 58 Lumiracoxib 59 Menadione 60 Nefazodone 61 Nomifensine 62 Pemoline 63 Tolcapone 64 Troglitazone

Anxiolytic Antidepressant Antimalaria, antiinflammatory Pain, fever, NSAID Cancer Antidepressant Pain, fever, NSAID Nutritional supplement Antidepressant Antidepressant ADHD Parkinson's disease Diabetes

24 h 30 N100 N100 N100 N100 N100 N100 20 13 N100 N100 N100 100

72 h 20 N100 13 N100 N100 N100 N100 1 10 100 N100 44 44

24 h 29 N100 100 N100 N100 N100 N100 16 12 N100 N100 N100 80

72 h 15 N100 14 N100 N100 N100 N100 12 6 100 N100 20 46

24 h N100 N100 67 N100 N100 N100 N100 13 10 N100 N100 N100 100

72 h 30 N100 13 N100 N100 N100 N100 13 10 N100 N100 44 20

24 h 44 N100 44 N100 N100 N100 N100 13 20 44 N100 N100 100

72 h 30 N100 30 N100 N100 N100 100 13 0.1 100 N100 30 100

24 h 13 N100 44 N100 N100 N100 N100 44 10 30 N100 67 44

72 h 44 N100 13 N100 N100 N100 N100 20 20 44 N100 1 30

24 h 20 N100 30 N100 N100 N100 N100 20 10 67 N100 13 44

72 h N100 N100 30 N100 100 30 N100 20 10 N100 N100 30 67

24 h 22 N100 49 N100 N100 N100 N100 13 12 N100 N100 50 48

72 h N100 N100 28 N100 N100 100 N100 25 10 N100 N100 44 59

24 h 20 N100 13 N100 N100 N100 N100 20 20 67 N100 67 N100

72 h 30 N100 10 N100 N100 N100 N100 10 13 N100 N100 30 67

Non-hepatotoxicants 65 Amisulpride 66 Bambuterol 67 Biotin 68 Bisacodyl 69 Buspirone 70 Caffeine 71 Carbidopa 72 Cisapride 73 Citicoline 74 Cromolyn 75 Dexamethasone 76 Diphenhydramin 77 DMSOb 78 Eserine 79 Fexofenadine 80 Flumazenil 81 Folinic Acid 82 Glimepiride 83 Isoproterenol 84 Isosorbide dinitrate 85 Lidocaine 86 Moxisylyte 87 Nialamide 88 Nicotine 89 Oxyphenonium bromide 90 Pargyline 91 Pergolide 92 Pinacidil 93 Pioglitazone 94 Propranolol 95 Pyridoxine 96 Streptomycin 97 Thiamine 98 Topiramate 99 Vancomycin

Antipsychotic Asthma Vitamin B7 Laxative Anxiolytic Stimulant Parkinson's disease Gastrointestinal Psychostimulant Antihistamine Antiinflammatory Antihistamine Solvent, negative control Choline esterase inhibitor Allergy Benzodiazepine antagonist Adjuvant Antidiabetic Adrenergic agonist Vasodilator Analgesic Erectile dysfunction Antidepressant Stimulant Antimuscarinic MAO inhibitor Parkinson's disease Vasodilator Diabetes Beta blocker Vitamin B6 Antibiotic Vitamin B1 Anticonvulsant Antibiotic

13 0.1 N100 N100 10 N100 44 20 N100 N100 N100 N100 N1% N100 67 N100 N100 N100 N100 N100 N100 N100 N100 N100 N100 N100 1 N100 N100 N100 N100 N100 N100 N100 N100

N100 N100 N100 30 10 N100 N100 N100 N100 N100 N100 13 N1% N100 N100 N100 N100 100 67 N100 N100 100 100 N100 N100 N100 N100 20 N100 30 N100 N100 N100 30 N100

N100 N100 N100 N100 N100 N100 N100 N100 N100 N100 N100 N100 N1% N100 N100 N100 N100 N100 N100 N100 N100 N100 N100 N100 N100 N100 N100 N100 N100 N100 N100 N100 N100 N100 N100

N100 N100 N100 90 10 N100 N100 N100 N100 N100 N100 N100 N1% N100 N100 N100 N100 N100 55 N100 N100 N100 N100 N100 N100 N100 N100 90 N100 32 N100 N100 N100 N100 N100

N100 N100 N100 N100 100 100 N100 0.1 N100 N100 N100 N100 N1% N100 67 N100 N100 N100 30 N100 N100 N100 30 N100 N100 N100 N100 N100 N100 N100 N100 N100 N100 N100 N100

N100 N100 N100 67 10 N100 N100 N100 N100 N100 100 13 N1% N100 N100 N100 N100 100 67 N100 N100 N100 N100 N100 N100 N100 67 100 100 67 N100 N100 N100 N100 N100

67 30 N100 N100 10 0.1 N100 30 N100 N100 N100 N100 N1% N100 0.1 N100 N100 N100 N100 N100 N100 N100 N100 N100 N100 N100 44 N100 0.01 N100 N100 N100 N100 N100 N100

N100 N100 N100 30 30 N100 N100 13 N100 N100 N100 20 N1% N100 N100 N100 N100 N100 100 N100 N100 N100 N100 N100 N100 N100 10 N100 100 20 N100 N100 N100 N100 N100

N100 N100 N100 N100 N100 N100 N100 N100 N100 N100 44 30 N1% 100 N100 N100 N100 20 N100 N100 44 1 N100 N100 20 N100 10 10 0.01 30 N100 N100 30 N100 100

N100 N100 N100 N100 N100 13 N100 N100 N100 N100 10 44 N1% 30 N100 N100 N100 30 100 N100 44 N100 N100 N100 N100 N100 1 N100 N100 20 N100 N100 N100 N100 N100

N100 N100 N100 100 44 N100 N100 N100 N100 N100 N100 N100 N1% N100 N100 N100 N100 N100 N100 N100 100 N100 13 N100 N100 N100 30 N100 0.01 30 N100 N100 N100 N100 N100

N100 N100 N100 N100 N100 N100 N100 20 N100 N100 N100 N100 N1% N100 N100 N100 N100 N100 100 N100 N100 N100 N100 N100 N100 N100 N100 N100 N100 44 N100 N100 N100 N100 N100

N100 N100 N100 N100 69 N100 N100 N100 N100 N100 N100 N100 N1% N100 N100 N100 N100 N100 N100 N100 N100 N100 N100 N100 N100 N100 N100 N100 N100 100 N100 N100 N100 N100 N100

N100 N100 N100 N100 N100 N100 N100 N100 N100 N100 N100 N100 N1% N100 N100 N100 N100 N100 90 N100 N100 N100 N100 N100 N100 N100 N100 N100 N100 12 N100 N100 N100 N100 N100

N100 N100 N100 N100 20 N100 N100 N100 N100 N100 N100 N100 N1% N100 N100 N100 100 N100 N100 N100 N100 N100 10 N100 N100 N100 N100 N100 N100 30 N100 N100 0.1 N100 N100

N100 N100 N100 44 67 N100 N100 13 N100 N100 0.1 30 N1% 10 N100 N100 N100 N100 100 N100 N100 30 N100 N100 N100 N100 44 N100 67 10 N100 N100 N100 N100 N100

Toxins/positive controls 100 Antimycin 101 CCCP 102 Rotenone

Toxin, Complex III inhibitor Toxin, uncoupler Toxin, complex I inhibitor

0.1 10 0.01

0.01 10 1

0.01 8 0.02

0.01 1 0.01

0.1 1 0.01

0.01 10 0.01

0.1 10 0.01

0.01 1 0.01

0.01 13 0.01

0.01 10 0.01

0.01 13 0.01

0.1 10 0.1

0.01 11 0.01

0.01 1 0.1

0.01 13 0.01

0.01 1 0.01

b

0.02 4 0.217a 0.0025 0.42 26.36 6 0.248 5 4.3 0.035 5.7 14.64 6.4

3.4 0.0015 1 0.15 0.01 42 2 0.12 700 0.016 0.23 0.3 0.22 0.57 0.1 3 0.73 0.006 0.0008 36 1.65 14 0.09 0.3 0.3 0.082 0.17 2.6 0.1 1.1 52 6.8 20 19

M. Persson et al. / Journal of Pharmacological and Toxicological Methods 68 (2013) 302–313

Cholesterol Cancer

a

Simvastatin Tamoxifen

Presuming a therapeutic dose of 100 mg. DMSO was added as a negative control (compound solvent) and was not included in calculating false positive rates.

307

308

M. Persson et al. / Journal of Pharmacological and Toxicological Methods 68 (2013) 302–313

A

B

Amodiaquine 300

Chlorpromazine 300

Nuclei counts

250

250

Nuclear area

Relative value [%]

Relative value [%]

Membrane Integrity Lysosomal activity

200

Mitochondrial membrane potential Mitochodrial area

150

100

50

200

150

100

50

0 0.01

0.1

1

10

0 0.01

100

0.1

D

Efavirenz

300

250

250

200

150

100

50

0 0.01

10

100

10

100

Carbidopa

300

Relative value [%]

Relative value [%]

C

1

µM compound

µM compound

200

150

100

50

0.1

1

10

100

µM compound

0 0.01

0.1

1

µM compound

Fig. 3. Typical results from multi-parametric analysis. Shown are results for all parameters for four compounds, as examples for the dataset. (A) Amodiaquine (#54), a drug which has been withdrawn from the market due to drug-induced liver injury, causes deviations from the normal population, starting with a loss of MMP and lysosomal activity at 30 μM, followed by a decrease in cell counts and loss of membrane integrity. (B) Chlorpromazine (#39) causes prominent cytotoxicity by loss of cells and plasma membrane integrity occurring at 20 μM with complete cell loss at 44 μM. (C) Efavirenz (#12) induces a decrease in the MMP at 20 μM and loss of cells and plasma membrane integrity at 30 μM. (D) Carbidopa (#71), which is nonhepatotoxic, does not cause any changes in cellular health parameters at tested concentrations. Data is expressed as in percentage of DMSO controls and as mean value ± SEM per cell based on six images per well. Numbers associated with the drugs correspond to the drug numbers in Table 1.

sensitivity of the mitochondrial toxicity parameters, but it does produce more false positives. Based on the data for ~100 drugs, we present three data analysis models, which can be used to predict whether a compound carries the risk to cause DILI: i) estimation of a safe drug exposure (Cmax) limit based on the individual parameters, ii) a zone classification model based on MMP, nuclei count and exposure (Cmax) and, iii) clustering of compounds with similar toxicity profiles based on all parameters. We show that a 100-fold TI between the lowest concentration at which toxicity is observed and the human therapeutic Cmax is optimal to classify compounds as hepatotoxic or non-hepatotoxic, based on the individual parameters. In other words, compounds that are picked up in the assay may cause human DILI if their exposure is higher than 1/100th of the toxic concentration in vitro. The parameters assessed in our assay typically have sensitivities of ~50% with ~10% false positives, except for the parameter nuclear counts which gives a sensitivity ranging from 20% (24 h compound incubation) to 40% (72 h incubation). This 100fold TI threshold is in line with what others report for their high content screening-based assays (O'Brien et al., 2006; Xu et al., 2008). In those studies, the TI values were calculated using the most sensitive in vitro parameter. It has also been shown by others that hepatotoxic compounds in general hit multiple in vitro parameters. In a recent study, for instance, an aggregated in vitro panel score was used to classify compounds according to their risk to induce idiosyncratic adverse drug

reactions (Thompson et al., 2012). The score indicated the number of parameters hit by a compound within a particular concentration range, but did not include therapeutic exposure data. We compared this approach (does the compound hit our assay parameters within the tested concentration range up to 100 μM) with our TI approach (does the compound hit our assay parameters within 100-fold below their therapeutic total Cmax), and we found that not taking the TI into account in general increases sensitivity at the expense of generating a higher false positive rate (Table 2). In screening efforts for early drug discovery projects, one may favor the TI approach, since it is essential to maintain a low false positive rate. One example of the benefit of taking TI into account is exemplified with buspirone. It is the only compound that comes out as a false positive in the hierarchical clustering model, which does not take TI into account. When analyzed in the zone classification system or using 100-fold TI, buspirone is identified as a non-hepatotoxicant which is in line with its clinical profile. We furthermore found that drugs that hit two or more parameters in the Quadprobe assay at a concentration below 100-fold their therapeutic Cmax are typically hepatotoxic, whereas non-hepatotoxic drugs typically hit less parameters within that 100-fold TI (Fig. 4I). The approach of estimating a safe exposure level by applying a TI is supported by the fact that a low daily dose is correlated to a low DILI risk. Even though idiosyncratic toxicity is not always dose-dependent in individual patients, high risk drugs are typically dosed at 100 mg/day or higher, only a few

M. Persson et al. / Journal of Pharmacological and Toxicological Methods 68 (2013) 302–313

A

Nuclei counts IC50

B

70

309

C Nuclear area LEC

Nuclei counts LEC 70

70

60

60

50

50

Sensitivity @ 24h

60

False positives @ 24h

20

20

10

10

10

0

1x

100x 300x 1000x

F

MMP LEC

60

60

60

50

50

50

30

Percentage

70

Percentage

70

40 30

20

10

10

10

0 30x

50x

Membrane integrity LEC 60

60

50

50

Percentage

70

30

5x

10x

30x

50x

40 30 20

10

10

I

5x

10x

2 1 0

100x 300x 1000x

1x

5x

10x

30x

50x

100x 300x 1000x

pa he e at er

50x

od

30x

M

10x

100x 300x 1000x

3

-h

5x

50x

4

on

0 1x

30x

5

N

0

100x 300x 1000x

Number of parameters hit

ep

20

1x

100x 300x 1000x

H Lysosomal activity LEC

70

40

0 1x

100x 300x 1000x

at o

10x

Number of parameters hit at 100x Cmax

G

5x

50x

30

20

1x

30x

40

20

0

10x

Mitochondrial area LEC

70

40

5x

aw n

50x

dr

30x

ith

10x

W

5x

to xi ci ty

E

MMP AC50

0 1x

100x 300x 1000x

pa to

50x

ity

30x

he

10x

to xi c

D

5x

re

1x

Percentage

30

20

0

Percentage

30

40

to

30

40

Se ve

40

to xi c

False positives @ 72h

Percentage

Percentage

50

Percentage

Sensitivity @ 72h

Fig. 4. Predictivity of individual parameters. Sensitivity and false positive calculations at different fold therapeutic indices using LEC values or IC50 or AC50 values for the individual parameters at 24 or 72 h incubation with compounds. Sensitivity towards human drug-induced liver injury is increased at 72 h compared to 24 h incubation with drugs for cytotoxicity related parameters (nuclei counts (B) and membrane integrity (G)) while other parameters remain largely unaffected by the longer incubation time. Drugs that hit two or more parameters at a concentration below 100-fold their therapeutic Cmax are typically hepatotoxic, whereas non-hepatotoxic drugs typically hit less parameters within that 100-fold TI (I). Lysosomal activity was excluded from this analysis, since we judged it to produce too many false positives.

compounds dosed below 50 mg/day have a DILI-risk, while no compounds dosed below 10 mg/day have been identified to cause DILI so far (Lammert et al., 2008; Stepan et al., 2011). In addition to estimating what exposure level may be safe, the data on the individual parameters can also be used to rank-order compounds according to cell-based TIs, which can be calculated by dividing an LEC or IC50 value from the Quadprobe assay by relevant efficacy data. Depending on the stage of the project, this could be data from an enzymatic assay, a functional cellular assay, an animal model, or predicted human therapeutic Cmax values (Muller & Milton, 2012). It has been shown that compounds

with a high cellular TI are more likely to provide safe treatments than compounds with a low cellular TI (Abraham, Towne, Waring, Warrior, & Burns, 2008). In order to avoid deselecting potentially promising compounds in drug discovery programs, it is essential that an in vitro toxicology assay does not produce too many false positives. In our zone classification model, based on the parameters nuclei count IC50 and MMP AC50, and the human therapeutic Cmax values (Fig. 5B), we identified an area without a single false positive, while maintaining a sensitivity of 45%. This confirms that nuclei counts and MMP are highly predictive

310

M. Persson et al. / Journal of Pharmacological and Toxicological Methods 68 (2013) 302–313

Table 2 Sensitivity and false positive rates for individual parameters. Shown are the sensitivity (% of hepatotoxic drugs correctly scored as hepatotoxic) and false positive rate (% of non-hepatotoxic drugs incorrectly scored as hepatotoxic) for each parameter in the Quadprobe assay, both 24 h and 72 h incubation with drug. Compared are two approaches to score compounds as hepatotoxic: with TI (does the compound hit the assay parameter within 100-fold below their therapeutic total Cmax) versus without TI (does the compound hit the assay parameter within the tested concentration range up to 100 μM). Parameter

Sensitivity [%]

False positive rate [%]

24 h

Nuclei counts LEC Nuclei counts IC50 Nuclear area LEC Membrane integrity LEC Lysosomal activity LEC MMP LEC MMP AC50 Mito Area LEC

72 h

24 h

72 h

100-fold TI

Without TI

100-fold TI

Without TI

100-fold TI

Without TI

100-fold TI

Without TI

19 19 33 28 47 47 30 44

36 36 45 45 64 63 42 58

42 41 42 41 30 36 28 38

55 41 56 58 47 52 42 52

12 0 9 12 24 9 0 9

21 0 18 24 38 21 6 15

12 0 6 6 15 0 0 12

29 15 29 24 27 9 6 32

parameters for human drug induced liver injury (O'Brien et al., 2006). Although based on different parameters, the model shares similarity with the zone classification system presented by Nakayama et al. (2009) which effectively classifies drugs for their risk to cause idiosyncratic toxicity based on covalent binding data in human hepatocytes and daily dose. Hierarchical clustering of the drugs according to the data in the Quadprobe assay groups compounds with a similar toxicity profile. If a novel compound from a drug discovery program is included in this analysis, one can identify which known drugs have a similar toxicity profile and subsequently extrapolate that the compound may have a similar clinical profile, although we did not take drug exposure into account for this clustering analysis. Several of the drugs in our validation set have been extensively studied and their toxic modes of action are often known. The clustering analysis can therefore help to elucidate the toxic mode of action of new compounds. Although our Quadprobe assay identifies ~50% of hepatotoxic drugs, and typically correctly classifies non-hepatotoxic drugs, a subset of hepatotoxic drugs are not identified as toxic. These false negatives can be explained by limitations of the model system and the choice of drugs to screen. DILI can be caused by a variety of drug-related risk factors, such as bioactivation, bile salt transporter inhibition, immune activation, and oxidative stress (Dawson, Stahl, Paul, Barber, & Kenna, 2012; Stepan et al., 2011; Thompson et al., 2012; Xu et al., 2008), which are not scored in the present method. Still, in order to obtain a fair estimate of the assay performance, we have intentionally included drugs with different mechanisms of toxicity. For instance, compounds which require metabolism beyond the capacity of HepG2 cells in order to become toxic, such as bromfenac (Otto et al., 2008), or which cause immune-mediated DILI, such as iproniazid (Homberg et al., 1982; Pons et al., 1996), are not identified. We found that the sensitivity can be further increased by including additional parameters, such as oxidative stress, glutathione depletion and transporter inhibition (our unpublished data). Similarly, it has been shown that combining in vitro toxicity data with covalent binding burden data further increases predictivity (Thompson et al., 2012), although it can be challenging to generate those data for large compound sets in drug discovery programs. A further limitation may be the fixed concentration range at which we tested, set to a maximum of 100 μM, which hampers identification of hepatotoxic drugs that are cytotoxic at higher concentrations. We chose this fixed top concentration in order to mimic the daily screening practice in drug discovery programs, in which the test concentration range cannot be adjusted for individual compounds and where a clinically relevant Cmax is often not known. A further complication in the translation of in vitro screening data to human organ toxicity could arise from the fact that the ratio between in vitro drug concentration and in vivo drug exposure is not equal for each drug, e.g. due to differences in binding of the drugs to proteins, lipids and tissue-specific matrices. The presence of 10% serum in our assay

allows for protein binding, and may compensate for this, at least in part. Using ‘free Cmax’ or ‘daily dose’ instead of total Cmax has in our experience not improved the predictive power of the parameters in this assay. All compounds were screened at a fixed final concentration of 1% DMSO, which is needed to achieve compound solubility (especially in early lead series). We have not observed any DMSO effects (compared to control wells without DMSO) on the parameters used in this assay. Although it cannot be excluded that the DMSO sensitizes cells towards drug-induced toxicity, this is not considered to have much impact, as the assay produces very few false positives. It can be argued that the investigated parameters in the Quadprobe assay are rather descriptors of general organ toxicity than specific hepatotoxicity. It has indeed been shown that high content screening cytotoxicity screening in HepG2 cells can identify compounds that are toxic to other organs as well (O'Brien et al., 2006), and that model systems using cells originating from liver, kidney or heart are equally sensitive to detecting general toxicities instead of being organ specific (Lin & Will, 2011). In vitro cytotoxicity values correlate with the severity of in vivo toxicity findings when assessing systemic toleration, clinical chemistry, and multi-organ pathology in rodents (Benbow, Aubrecht, Banker, Nettleton, & Aleo, 2010), which further indicates that in vitro cytotoxicity screening data do not only predict liver toxicity. Since DILI is difficult to identify in animal toxicology studies, and is one of the leading causes for safety-related failures in drug development as well as market withdrawals and other regulatory actions, early hazard identification, risk assessment and risk avoidance strategies are essential. Our Quadprobe assay is well suited as a first-line safety screening assay in drug discovery projects: the assay is cost-effective and allows for relative high throughput testing, since it is plate-based and does not require live cell imaging; and it allows for decision making by project teams, since the data for novel compounds can be placed in relation to the data for the validation set with known drugs, as presented here. We positioned the assay at various places in our drug discovery process: to gather information on novel compound series, which aids in prioritization between hit series during the hit-to-lead process; to optimize on safety endpoints by generating structure–activity relationships and thereby steer away from a potential DILI risk during the lead optimization process; to assess the DILI risk towards candidate selection and to provide guidance of safe exposure levels in humans. In this way, the assay provides early risk assessment as well as risk avoidance and, as such, should contribute to reduce safety-related late stage failures.

Acknowledgments We kindly thank Dr. Annemette Thougaard for scientific discussions, and Bente Rotbøl and Tina Christensen for technical assistance.

M. Persson et al. / Journal of Pharmacological and Toxicological Methods 68 (2013) 302–313

311

A 13

63

37 4

24

94

19

69 Nuclei counts IC50

42

28

54 5

17 64 33

3 41

11 49 51

50

48

52

2

12 39

6

59

MMP AC50

20

60 101

100 102

Cmax [µM]

B Nuclei counts * Mitochondrial membrane potential

Clean

20000

17500 94

37

15000

42

12500

4 69

10000 63

7500 33

5000

54 17

41

50 48

52

19

39

43

0

49

0.001

0.01

100

0.1

64

24

3

2500

13

5

6

2 101 102

1

28

11

60

51 59 20

10

12

100

1000

Cmax [µM] Fig. 5. DILI risk assessment based on nuclei counts, mitochondrial membrane potential and exposure. (A): Most non-hepatotoxic drugs (green) do not affect either parameter, while some moderately hepatotoxic compounds (orange), severely hepatotoxic compounds (red), and compounds withdrawn due to hepatotoxicity (dark red) affect either nuclei counts or MMP to various degrees, or both, similar to the positive control chemicals (black). (B): The multiplied product of the values for nuclei counts IC50 and MMP AC50 is plotted against the human therapeutic total Cmax. A simple divider (red dashed line) identifies a zone that only harbors hepatotoxic drugs. This zone classification model has a sensitivity of 45%, and produces no false positives for predicting DILI. The black dotted line indicates the maximum concentration tested in the assay. Compounds that do not produce an IC50 or AC50 value in the tested concentration range were given the value of 150 μM, in order to be able to depict them in the plot. Positive control chemicals are arbitrarily set to a Cmax of 1 μM, but are excluded from the sensitivity and false positive analysis. Numbers associated with the symbols correspond to the drug numbers in Table 1.

312

M. Persson et al. / Journal of Pharmacological and Toxicological Methods 68 (2013) 302–313

Fig. 6. Dendrogram visualization of hierarchical clustering results. Hierarchical clustering on all cytotoxicity parameters from the Quadprobe assay was performed in order to identify compounds with similar toxicity profiles. Of the two main resulting clusters, a “toxic” cluster (highlighted in red) only contains hepatotoxic drugs (withdrawn, severe and moderate). The other main cluster (highlighted in green) contains all non-toxic drugs as well as those toxic drugs the assay does not pick up. Compounds tend to cluster together according to their mechanisms of action, such as the mitochondrial toxins used as positive control chemicals (antimycin A (#100), CCCP (#101), and rotenone (#102)). Since most of the compounds in the training set have known mechanisms of toxicity, the method can be used for elucidating toxicity profiles for new drug candidates. Numbers associated with the chemicals correspond to the numbers in Table 1.

References Abraham, V. C., Towne, D. L., Waring, J. F., Warrior, U., & Burns, D. J. (2008). Application of a high-content multiparameter cytotoxicity assay to prioritize compounds based on toxicity potential in humans. Journal of Biomolecular Screening, 13, 527–537. Arrowsmith, J. (2011a). Trial watch: Phase II failures: 2008–2010. Nature Reviews. Drug Discovery, 10, 328–329. Arrowsmith, J. (2011b). Trial watch: phase III and submission failures: 2007–2010. Nature Reviews. Drug Discovery, 10, 87. Begriche, K., Massart, J., Robin, M.A., Borgne-Sanchez, A., & Fromenty, B. (2011). Drug-induced toxicity on mitochondria and lipid metabolism: mechanistic diversity and deleterious consequences for the liver. Journal of Hepatology, 54, 773–794. Benbow, J. W., Aubrecht, J., Banker, M. J., Nettleton, D., & Aleo, M.D. (2010). Predicting safety toleration of pharmaceutical chemical leads: Cytotoxicity correlations to exploratory toxicity studies. Toxicology Letters, 197, 175–182.

Blas-Garcia, A., Apostolova, N., Ballesteros, D., Monleon, D., Morales, J. M., Rocha, M., et al. (2010). Inhibition of mitochondrial function by efavirenz increases lipid content in hepatic cells. Hepatology, 52, 115–125. Boelsterli, U. A., & Lim, P. L. (2007). Mitochondrial abnormalities — A link to idiosyncratic drug hepatotoxicity? Toxicology and Applied Pharmacology, 220, 92–107. Chalasani, N., & Bjornsson, E. (2010). Risk factors for idiosyncratic drug-induced liver injury. Gastroenterology, 138, 2246–2259. Chan, K., Truong, D., Shangari, N., & O'Brien, P. J. (2005). Drug-induced mitochondrial toxicity. Expert Opinion on Drug Metabolism & Toxicology, 1, 655–669. Czaja, M. J. (2011). Functions of autophagy in hepatic and pancreatic physiology and disease. Gastroenterology, 140, 1895–1908. Dawson, S., Stahl, S., Paul, N., Barber, J., & Kenna, J. G. (2012). In vitro inhibition of the bile salt export pump correlates with risk of cholestatic drug-induced liver injury in humans. Drug Metabolism and Disposition, 40, 130–138. Dykens, J. A., Marroquin, L. D., & Will, Y. (2007). Strategies to reduce late-stage drug attrition due to mitochondrial toxicity. Expert Review of Molecular Diagnostics, 7, 161–175.

M. Persson et al. / Journal of Pharmacological and Toxicological Methods 68 (2013) 302–313 Dykens, J. A., & Will, Y. (2007). The significance of mitochondrial toxicity testing in drug development. Drug Discovery Today, 12, 777–785. Fung, M., Thornton, A., Mybeck, K., Wu, J. H., Hornbuckle, K., & Muniz, E. (2001). Evaluation of the characteristics of safety withdrawal of prescription drugs from worldwide pharmaceutical markets — 1960 to 1999. Drug Information Journal, 35, 293–317. Greer, M. L., Barber, J., Eakins, J., & Kenna, J. G. (2010). Cell based approaches for evaluation of drug-induced liver injury. Toxicology, 268, 125–131. Holt, M. P., & Ju, C. (2006). Mechanisms of drug-induced liver injury. The AAPS Journal, 8, E48–E54. Homberg, J. C., Stelly, N., Andreis, I., Abuaf, N., Saadoun, F., & Andre, J. (1982). A new antimitochondria antibody (anti-M6) in iproniazid-induced hepatitis. Clinical and Experimental Immunology, 47, 93–102. Kaplowitz, N. (2004). Drug-induced liver injury. Clinical Infectious Diseases, 38(Suppl. 2), S44–S48. Kaplowitz, N. (2005). Idiosyncratic drug hepatotoxicity. Nature Reviews. Drug Discovery, 4, 489–499. Kola, I., & Landis, J. (2004). Can the pharmaceutical industry reduce attrition rates? Nature Reviews. Drug Discovery, 3, 711–715. Labbe, G., Pessayre, D., & Fromenty, B. (2008). Drug-induced liver injury through mitochondrial dysfunction: mechanisms and detection during preclinical safety studies. Fundamental and Clinical Pharmacology, 22, 335–353. Lammert, C., Einarsson, S., Saha, C., Niklasson, A., Bjornsson, E., & Chalasani, N. (2008). Relationship between daily dose of oral medications and idiosyncratic drug-induced liver injury: search for signals. Hepatology, 47, 2003–2009. Lee, W. M. (2003). Drug-induced hepatotoxicity. The New England Journal of Medicine, 349, 474–485. Lin, Z., & Will, Y. (2011). Evaluation of drugs with specific organ toxicities in organ specific cell lines. Toxicological Sciences, 126, 114–127. Muller, P. Y., & Milton, M. N. (2012). The determination and interpretation of the therapeutic index in drug development. Nature Reviews. Drug Discovery, 11, 751–761. Nadanaciva, S., Lu, S., Gebhard, D. F., Jessen, B.A., Pennie, W. D., & Will, Y. (2011). A high content screening assay for identifying lysosomotropic compounds. Toxicology In Vitro, 25, 715–723. Nakayama, S., Atsumi, R., Takakusa, H., Kobayashi, Y., Kurihara, A., Nagai, Y., et al. (2009). A zone classification system for risk assessment of idiosyncratic drug toxicity using daily dose and covalent binding. Drug Metabolism and Disposition, 37, 1970–1977. O'Brien, P. J., Irwin, W., Diaz, D., Howard-Cofield, E., Krejsa, C. M., Slaughter, M. R., et al. (2006). High concordance of drug-induced human hepatotoxicity with in vitro

313

cytotoxicity measured in a novel cell-based model using high content screening. Archives of Toxicology, 80, 580–604. Olson, H., Betton, G., Robinson, D., Thomas, K., Monro, A., Kolaja, G., et al. (2000). Concordance of the toxicity of pharmaceuticals in humans and in animals. Regulatory Toxicology and Pharmacology, 32, 56–67. Otto, M., Hansen, S. H., Dalgaard, L., Dubois, J., & Badolo, L. (2008). Development of an in vitro assay for the investigation of metabolism-induced drug hepatotoxicity. Cell Biology and Toxicology, 24, 87–99. Pessayre, D., Fromenty, B., Berson, A., Robin, M. A., Lettéron, P., Moreau, R., et al. (2012). Central role of mitochondria in drug-induced liver injury. Drug Metabolism Reviews, 44, 34–87. Pons, C., Dansette, P. M., Gregeois, J., Homberg, J. C., Billett, E. E., & Mansuy, D. (1996). Human anti-mitochondria autoantibodies appearing in iproniazid-induced immunoallergic hepatitis recognize human liver monoamine oxidase B. Biochemical and Biophysical Research Communications, 218, 118–124. Russmann, S., Kullak-Ublick, G. A., & Grattagliano, I. (2009). Current concepts of mechanisms in drug-induced hepatotoxicity. Current Medicinal Chemistry, 16, 3041–3053. Schneider, P., Korolenko, T. A., & Busch, U. (1997). A review of drug-induced lysosomal disorders of the liver in man and laboratory animals. Microscopy Research and Technique, 36, 253–275. Stepan, A. F., Walker, D. P., Bauman, J., Price, D. A., Baillie, T. A., Kalgutkar, A. S., et al. (2011). Structural alert/reactive metabolite concept as applied in medicinal chemistry to mitigate the risk of idiosyncratic drug toxicity: A perspective based on the critical examination of trends in the top 200 drugs marketed in the United States. Chemical Research in Toxicology, 24, 1345–1410. Stevens, J. L., & Baker, T. K. (2009). The future of drug safety testing: Expanding the view and narrowing the focus. Drug Discovery Today, 14, 162–167. Thompson, R. A., Isin, E. M., Li, Y., Weidolf, L., Page, K., Wilson, I., et al. (2012). In vitro approach to assess the potential for risk of idiosyncratic adverse reactions caused by candidate drugs. Chemical Research in Toxicology, 25, 1616–1632. Xu, J. J., Henstock, P. V., Dunn, M. C., Smith, A.R., Chabot, J. R., & de, G. D. (2008). Cellular imaging predictions of clinical drug-induced liver injury. Toxicological Sciences, 105, 97–105. Zhang, H., Solomon, V. R., Hu, C., Ulibarri, G., & Lee, H. (2008). Synthesis and in vitro cytotoxicity evaluation of 4-aminoquinoline derivatives. Biomedicine and Pharmacotherapy, 62, 65–69.