The integrated stress response‐related expression of CHOP due to mitochondrial toxicity is a warning sign for DILI liability

Drug‐induced liver injury (DILI) is one of the most frequent reasons for failure of drugs in clinical trials or market withdrawal. Early assessment of DILI risk remains a major challenge during drug development. Here, we present a mechanism‐based weight‐of‐evidence approach able to identify certain candidate compounds with DILI liabilities due to mitochondrial toxicity.


| INTRODUC TI ON
Drug-induced liver injury (DILI) is a common reason for failure in drug development or market withdrawal and cause of acute liver failure. 1,2Due to its function in drug metabolism, the liver is considered one of the most important target organs in the context of adverse drug reactions.Although knowledge about mechanisms involved in DILI has increased over the past, 3 early assessment of DILI risk during drug development and chemical safety assessment remains a major challenge.Thus, clinical DILI hazard has mostly been missed in pre-clinical animal models. 4We anticipate that mechanism-based testing using human-relevant test methods will be pivotal to advance the assessment of DILI risk for novel drug candidates.
Mechanisms of DILI are diverse and involve specific drug on-target and off-target effects on transporters or nuclear hormone receptors leading to cholestasis or steatosis respectively. 5,6ternatively, drugs may cause cell injury through reactive metabolite formation or disturbances of normal cell function including mitochondrial toxicity. 7These cell disturbances may lead to the activation of cellular stress response activation involving NRF2mediated antioxidant stress response, p53-mediated DNA damage response, or the ATF4-and ATF6-mediated unfolded protein response. 8Recent gene co-expression network analysis (WGCNA) allowed the association of stress response activation with hepatotoxicity histopathology and marking the ATF4-mediated CHOP induction in direct relationship with liver cell apoptosis and the NRF2-mediated SRXN1 induction with liver cell hyperplasia. 91][12] However, the underlying mechanism for the stress response activation remained unclear.Secondly, the translation of these findings to observations in primary human hepatocytes as well as the observations of postmarketing information was lacking.Furthermore, a wider analysis on a larger set of drugs that is not biased towards DILI is missing.This has limited a direct application for a mechanism-based hazard characterization to deprioritize compounds with a likelihood for DILI liability.
In the current study, we have integrated a large-scale mechanism-based high-throughput screen with further mechanistic studies.We performed a high-throughput imaging-based screening for cellular stress response activation in HepG2 CHOP-GFP, P21-GFP and SRXN1-GFP reporter cell lines.We screened a total of 1587 FDA-approved molecules and 378 kinase inhibitors, an important novel drug class for which post-marketing information is largely missing.Using primary human hepatocytes, the cellular response to a selection of compounds that were cross-referenced with existing DILI datasets for their possible DILI involvement and sharing chemical or pharmacological similar features, was further uncovered using concentration response targeted RNAsequencing and detailed mitochondrial functionality assays.Our combined results indicate that CHOP activation in the absence of the canonical activation of the unfolded protein response is an important marker for mitochondrial toxicity.We suggest a weightof-evidence tiered testing strategy to guide hazard identification and characterization to deprioritize compounds with possible DILI concerns.

| Primary and secondary screen
Compounds from FDA-approved drugs and kinase inhibitor libraries were aliquoted using a Labcyte Echo® 550 Liquid Handler (Labcyte Inc., United Kingdom).To prevent hit identification due to compound volatility and plate location, compounds were grouped and per biological replicate randomly allocated on a plate.At the start of the experiment, HepG2 BAC-GFP reporter cell lines were seeded at a density of .25 million cells/mL in Greiner μclear plates (Greiner Bio-One, 781 091).Prior to exposure, nuclei were stained with 100 ng/ mL Hoechst 33342 (Sigma-Aldrich, H3570) for 24 h.Thereafter, the medium was removed, and cells were exposed to all compounds at 50 μM (primary screen) or eight concentrations between 0.31 and 50 μM (secondary validation screen) for 48 h.For the control compounds, indicated concentrations were used to generate a concentration-response curve and determine reporter activation levels.To track cell death, in-house manufactured Annexin V-Alexa633 and 100-nM propidium iodide (PI, Fisher Scientific, P1304MP) were added to the exposure medium. 13Seeding, nuclear staining and exposure procedures were optimized for laboratory automation by a liquid-handling robot (BioMek FX, Beckman Coulter).After incubation, images were taken using the Molecular Devices ImageXpress Micro Confocal microscope (P21-GFP and SRXN1-GFP) using a 20× Plan Apo objective and DAPI, FITC, TRITC and CY5 filter sets or Nikon TE-2000-E eclipse microscope (CHOP-GFP) using a 20× Plan Apo objective, 408, 488, 561 and 633 nm lasers.Information on the analysis of the screen can be found in the Supplementary Materials and Methods section.
After 6 hours, culture medium was changed to InvitroGro maintenance medium.Twenty-four hours after plating, PHH were treated for 24 h with selected compounds at eight different concentrations between .1×Cmax and 316× Cmax.After treatment, medium was collected, and cytotoxicity was assessed using Roche's LDH Cytotoxicity kit (Merck/Roche, 11 644 793 001).For each compound, a maximum of six concentrations was chosen for transcriptomic analysis based on the cytotoxic response.Directly after medium collection, cells were washed once with PBS and lysed with 1× TempO-Seq lysis buffer (BioClavis, Glasgow, Scotland) for 15 min at RT. Next, samples were frozen and shipped at −80°C.
All samples were analysed using TempO-Seq technology deploying a whole transcriptome probe set.Probe alignment was performed by BioClavis as described previously. 14Read counts were analysed using an in-house developed bioinformatics pipeline using R. First, samples with less than one million counts were discarded.
Thereafter, PCA analysis was performed to identify possible outliers.All samples passed these quality controls.Next, DEGs per sample were calculated based on a negative binominal regression using the R DeSeq2 package. 15All samples were compared towards DMSO control to calculate log2 fold change gene expression.These values were uploaded to a previously established WGCNA-based PHH TXG-MAPr bioinformatics tool (http:// www.txg-mapr.eu) to calculate and extract gene network module eigengene scores (EGS). 16To gain mechanistic and translational insights, the previously reported module annotation was used. 16Only modules with a preserved status towards rat in vivo data or with a clear mechanistic annotation based on module enrichment were considered for analysis.

| Mitochondrial seahorse assay
The extracellular flux assay was used to assess mitochondrial toxicity in HepG2 and LiverPool 10 Donor PHH cells by determining the oxygen consumption rate (OCR), reserve capacity and extracellular acidification rate (ECAR) utilizing the XF e 96 flux analyser (Agilent, Cheadle, UK) as previously described. 17More details can be found in the Supplementary Materials and Methods.

| FDA-approved molecules induce cellular stress response pathways throughout the drug-like chemical space
First, the compound libraries, FDA-approved drugs and KIs, were characterized for their indications, primary target and marketing status using the manufacturer's annotation and ChEMBL database. 18The FDA-approved drug library consists of 1587 drugs, targeting a diverse panel of cellular pathways for various indications (Figure 1A,B).The 378 drugs from the KI library are mainly involved in cancer treatment (Figure 1A).Note that only KIs that are in advanced clinical development stages have established indications for which these drugs are used.As expected, most of the KIs target well-known cancer-related pathways like the cell cycle regulators, EGFR and MAPK pathway (Figure 1B).After mining the ChEMBL database, most FDA-approved drugs were confirmed to have a marketed status, whereas most kinase inhibitors were in preclinical development (Figure 1C).Furthermore, 42 molecules were identified that have been withdrawn for various reasons, including hepatotoxicity.
To test the ability of FDA-approved drugs and KIs to induce DILIassociated cellular stress response pathways, we evaluated these compounds in three critical HepG2 BAC-GFP reporters: CHOP-GFP for the unfolded protein response (UPR), P21-GFP for the DNA damage response (DDR) and SRXN1-GFP for the oxidative stress response (OSR) (Figure 2A). 10,11Activation of these pathways does not directly indicate DILI liabilities in the clinic, but ultimately could be associated with both intrinsic and idiosyncratic DILI based on previous research. 11,19,20Cell death was determined using Annexin V-Alexa633 and PI staining.For the primary screen, all cells were exposed to a concentration of 50 μM for 48 h.After quality assessment, point-of-departures (PoD) were calculated for each reporter for the positive controls to identify the lowest significant reporter onset intensity levels (Figure 2B,C).Compounds exceeding these PoD values were considered a positive hit.Here, we identified 389 compounds for CHOP-GFP induction, 219 compounds for P21-GFP and 104 compounds for SRXN1-GFP that were further evaluated in a secondary screen (Figure 2D).Compounds that activated these reporters covered the entire drug chemical space and some compounds triggered several stress responses.In addition, 157 compounds caused significant levels of cell death in multiple reporter systems and/or the HepG2 wild-type cell line (Supporting Information Figure S1); these compounds were taken along into the secondary screen to assess the stress pathway activation at lower concentrations.
In a secondary screen, 742 compounds were evaluated at a dose range between 0.31 and 50 μM for 48 h.For each compound, a PoD was calculated (Supporting Information Table S1).
Compounds for which no PoD could be calculated were not considered for follow-up experiments.To further strengthen the validation, a Williams trend test was applied to the doseresponse curve of each compound.PoD values were determined for 154, 119 and 51 FDA-approved drug candidates for CHOP-GFP, P21-GFP and SRXN1-GFP induction respectively (Figure 2E, Supporting Information Figure S2).For the validated KIs, the majority only showed induction of cellular stress response pathway activation at 25 μM or higher (Supporting Information Figure S3).This was not unexpected, since kinase inhibitors are known for their high target affinity and are typically on target at higher nM concentrations. 21Since the stress pathway activation was often observed at a concentration more than 100 times the IC50, we did not further consider KIs for our further studies and focused on FDA-approved drugs.

| Some antimycotic and central nervous system FDA-approved drugs that induce CHOP-GFP expression are associated with DILI liabilities
Next, drug properties, the molecular fingerprint and known hepatotoxicity claims of the identified hits from the secondary screen

F I G U R E 1 Overview of the screened drug libraries (FDA-approved drugs and Kinase Inhibitors). (A) Number of drugs per indication.
Colour indicates whether the drug is allocated in the FDA-approved drug library (blue) or in both the FDA-approved drug and kinase inhibitor library (Green).(B) Overview of pathways and primary molecular targets of the screened drugs.The surface area corresponds to the number of drugs targeting a specific pathway (middle layer) or protein (outside layer).(C) Maximum clinical stage of all drugs for any indication.
were evaluated using various DILI datasets.Again, the library manufacturer's annotation was used to identify compound classes and their molecular targets.Among the CHOP-GFP reporter positive CNS agents, chemotherapeutics and anti-infection drugs were identified as the major drug classes (Figure 3A).Closer examination of the intended targets revealed that the anti-infection drugs mainly contained antimycotic drugs and 11 out of 46 tested antimycotic drugs were positive for CHOP-GFP activation (Figure 3B, Supporting Information Figure S4).Furthermore, 31 out of 254 tested CNS drugs were positive and these did mainly target classical neuronal receptors.The majority of P21-GFP inducing compounds were drugs involved in cancer therapeutics and for SRXN1-GFP positive FDA compounds we observed a highly diverse set of drugs targeting various pathways (Supporting Information Figure S5).
The CHOP-GFP positive hits showed the largest group of compounds sharing similar targets, including antimycotic and CNS drugs (Figure 3A,B).Of note, due to the screening set-up, for most of these drugs, the Cmax was exceeded by more than 100× (Supporting Information Table S2).In the past decade, various in silico studies were performed to establish prediction models for hepatotoxicity. 22ese studies often suggest that compounds with a similar molecular fingerprint lead to a similar biological outcome.To test this hypothesis for our compounds, nine antimycotic drugs with structural similarities and 13 dissimilar CNS agents were selected for further interrogation (Figure 3C,D).To better hypothesize possible outcomes for follow-up experiments for the selected hits, molecular fingerprints of the antimycotic drugs and CNS agents were compared.The comparison of the molecular fingerprints revealed that seven out of nine antimycotic drugs shared a highly similar scaffold, whereas all the 13 selected CNS agents were structurally different (Figure 3C,D).This suggests that induction of CHOP can likely be explained by specific drug-target interactions and similar biological mechanisms for antimycotic drugs can be expected due to similar biological interactions (e.g.similar (off-)targets).Interestingly, although all compounds induced CHOP-GFP, cross-referencing our selected compounds with dilirank, medline, sider and pharmapendium for hepatic adverse events showed that three out of nine antimycotic (clotrimazole, miconazole and sulconazole) compounds and 10 out of 13 CNS agents (asenapine, doxazosin, fluoxetine, duloxetine, chlorpromazine, desloratadine, azelastine, indacaterol, escitalopram and vilazodone) were linked to a DILI liability in the clinic (Figure 3E).A few compounds (isavuconazole, domperidone, loperamide, vilazodone) were found to be DILI negative, although some data for these compounds usually are missing in the mined datasets.For sertaconazole, econazole, isoconazole, tioconazole, butoconazole and ifenprodil, there were no data available at all.For these compounds, no conclusions can be drawn on their DILI status in the clinic based on these datasets.

| Selected antimycotic and CNS drugs do not activate the canonical UPR
We further aimed to uncover a common underlying mechanism that determines CHOP-GFP activation by both the antimycotic and CNS drugs.We performed transcriptome analysis to uncover the overall biological perturbations.To facilitate translation to a clinical setting, we used plated cryopreserved 10 donor-pooled primary human hepatocytes (PHH).PHH were exposed to a concentration range of the selected drugs, based on the maximum blood-plasma concentration (Cmax), where the Cmax was obtained via literature data or in silico prediction (Table 1 and Supporting Information Table S2).PHH were treated with increasing Cmax values up to where cytotoxicity was observed or 316x Cmax.First, the induction of DDIT3 (CHOP) and the upstream ER stress sensor HSPA5 (BIP) in PHH were evaluated (Figure 4A).A reference control compound that induces UPR, tunicamycin, enhanced expression of both genes.For most of the selected antimycotic and CNS drugs, a clear induction of DDIT3 was observed at levels above the predicted Cmax levels.However, except for duloxetine, none of the tested concentrations induced HSPA5 expression.The activation of gene network modules associated with ER stress was further explored using our PHH TXG-MAPr platform. 16Tunicamycin showed activation of ER stress-associated modules (Figure 4B).
For the compounds of interest, only at the highest concentrations, F I G U R E 2 FDA-approved molecules induce CHOP-GFP expression in HepG2 cells throughout whole drug-like chemical space.(A) Overview of the screening set-up.(B) Established point-of-departure (PoD) values for all reporters (CHOP-GFP, P21-GFP, SRXN1-GFP) in the primary screen.GFP intensity was min/max normalized, using the negative control (.5% DMSO) as the minimum value and the intensity value after treatment with the highest concentration of the positive control as the maximum response.Thereafter, a LOESS curvefit was applied.The vertical line indicates the concentration of the control compound from where a significant induction of the respective reporter was identified.The horizontal line indicates the intensity value at which the induction becomes significant and was set as the point-ofdeparture (PoD) for the respective reporter.(C) Representative images of data shown at B. (D) tSNE plot representing all drugs that were screened throughout the chemical space.Red dots indicate a hit, meaning that these compounds induced a response in the reporter that was exceeding the PoD established at B. (E) Results of the secondary screen.In the foreground, possible trends for the curve are shown.In the background, the concentration-response curve for each validated hit is shown.The colour of the line represents the reporter cell line it was validated for.Compounds were screened at indicated doses.Thereafter, a PoD was calculated for each compound using a similar approach as described at B. If a PoD was found, the umbrella-protected Williams trend test using a contrast matrix was applied to these compounds to identify significant induction (p-adjusted value <.05) at any concentration.If a significant value was found, the compound was allocated to the trend with the most significant (lowest) p-value.activation of gene networks associated with ATF4 (PHH:15, PHH:205, PHH:367), ATF6 (PHH:13) and DDIT3 (PHH:280) was observed (Figure 4B).When closely examining the individual gene activation within module PHH:280, no significant differences between tunicamycin and the tested compounds were seen (Figure 4C).However, in contrast to tunicamycin, the expression of genes in module 62 remained unchanged for both antimycotic and CNS drugs (Figure 4D), suggesting that rather than the canonical UPR, different mechanisms drive the activation of CHOP by the antimycotics and CNS agents.

| Antimycotic and CNS drugs affect mitochondrial gene networks in primary human hepatocytes
We further systematically explored the transcriptomic data using the PHH TXG-MAPr.First, we evaluated similarities in the transcriptomic response of the selected compounds using Pearson correlation scores of module's EGS (Supporting Information Figure S6).PHH treated with structural similar antimycotic compounds did cluster together (cluster 3).Interestingly, the samples with the highest concentrations of four TA B L E 1 Predicted maximum blood-plasma concentrations (Cmax) for selected drugs.CNS agents were also present in cluster 3 (Figure 5A,B), indicative that these compounds have a comparable biological perturbation of PHH as the antimycotic compounds.The antimycotic drugs, isavuconazole and clotrimazole, did cluster with a lower Pearson correlation score (−0.2-0.4).We anticipated that the similarity in mode-of-action that defines the grouping in cluster 3 might also be underlying to the mechanisms leading to CHOP activation.We selected the modules based on their enrichment terms as well as preservation status towards rat-based testing systems. 16Since the observed response at the highest Cmax levels has a high Pearson correlation score, for this analysis all antimycotic compounds were included.Since module EGS are Z-scores, we only considered modules with a mean absolute EGS of at least 2 at the highest concentration for all antimycotic drugs and modules associated with ER stress. 16dose-dependent increase of NRF2 activity (PHH:144, PHH:337) and Xenobiotic stress (PHH:134, PHH:358) was observed for the antimycotic drugs (Figure 5D).These modules were not affected by the CNS agents (Supporting Information Figure S7).Furthermore, modules associated with mitochondrial activity (PHH:113, PHH:256, PHH:97) were deactivated for most compounds; the antimycotic compounds were more potent compared to the CNS agents (Figure 5D, Supporting Information Figure S7).For example, mitochondrial associated module PHH:113 showed a decrease in gene expression for most genes, involving MRPS15, NDUFB10 and PDZD11 that are part of the mitochondrial transcriptional regulation and oxidative phosphorylation and showed a downregulation of up to 2.8-log2 fold (Figure 5E).
In module PHH:256, a similar pattern was visible for MRPL41, MTIF3 and NDUFA8 (Figure 5F).For these modules a more in-depth time course study was further evaluated using butoconazole and isavuconazole (Supporting Information Figure S8).Here, samples were exposed for up to 72 h, followed by a recovery period and sample lysis.
Sustained (72 h) exposure resulted in significant induction of cytotoxicity (Supporting Information Figure S8A).For butoconazole, but not isavuconazole, DDIT3 showed a decreasing trend of induction at 48 h after wash-out, but overall stable activation of modules (Supporting Information Figure S8B,C).Mitochondrial modules showed reducing modulation over time, suggesting that mitochondrial gene expression did recover (Supporting Information Figure S8C).
To test whether the response isavuconazole and butoconazole were specific towards PHH, RPTEC/TERT1 cells were treated with these antimycotics.Genes from the PHH modules were used as marker for mitochondrial functionality.A similar trend was observed for the gene expression in RPTEC-TERT1 cells, albeit with lower potency (Supporting Information Figure S9).The observed decrease in gene expression was expected based on previous observations for mitochondrial toxicants in both HepG2 and RPTEC/TERT1 cells. 23Taken together, the data suggest that the antimycotic drugs and a select group of CNS agents impacted mitochondrial function as an important mode of action.Also for the other CNS agents, at the highest doses, the mitochondrialassociated WGCNA modules were deactivated (Supporting Information Figure S7).Since the Pearson correlation between these CNS agents was relatively low and this mitochondrial effect was mainly visible at the highest Cmax, we further focused on the antimycotic compounds and the CNS agents with a similar transcriptomic fingerprint to validate onset of mitochondrial toxicity.

| Antimycotic and CNS drugs that induce DDIT3 expression impair mitochondrial oxygen consumption
Since the transcriptome data suggested mitochondrial perturbation underlying the observed toxicity, we further evaluated the effect of selected antimycotic drugs and CNS agents on mitochondrial function and measured the mitochondrial oxygen consumption rate (OCR) and reserve capacity (Figure 6A,B). 24For almost all selected compounds, a decrease in mitochondrial OCR was observed at relatively low concentrations in both PHH and HepG2 cells (Figure 6C, Table 2).Moreover, all antimycotic drugs showed a major decrease in the reserve capacity of mitochondria respiration after treatment (Table 2, Supporting Information Figure S10A).We observed  improvements can be made by using translational approaches including knowledge from pathophysiological programs involved in human liver disease and/or animal data.We have used large toxicogenomics datasets from rat 28-day repeat dose toxicity studies to establish a rat liver TXG-MAPr and identified activation of gene co-expression networks that are associated with, e.g., in vivo onset of liver single cell necrosis (=apoptosis).HepG2 BAC-GFP reporters for TRIB3 and MTHFD2 that represent this gene network contribute to screening for liabilities that have direct in vivo toxicological relevance. 26Other improvements are required to be able to directly relate in vitro reporter outcomes towards DILI risk in the clinic.The current screening set-up had several limitations that need to be overcome in the future: (1) some drugs have a Cmax that is higher than the tested concentration of 50 μM, and therefore toxic effects may have been missed; (2) HepG2 cells have limited metabolic capacity prohibiting bioactivation of candidate drugs; (3) DILI often involves complex mechanisms including different (immune) cell types, e.g., parenchymal cells, Kupffer cells, liver natural killer cells and infiltrating immune cells that are critical in the pathogenesis of specific drug-induced liver disease phenotypes.Related to this, HLA-haplotypes can also determine whether an individual is susceptible to DILI for specific drugs. 27Currently, further exploration of these complex adaptive immune-mediated mechanisms can only be achieved in in vivo models. 28Technological advancements in perfusion culture models are required to be able to measure these interactions in vitro.
In our mechanistic studies using transcriptomics analysis on a select set of antimycotic and CNS drugs, we confirmed the ATF4-CHOP/DDIT3 cellular stress response pathway activation in PHH in conjunction with perturbation of mitochondrial function, further validated by mitochondrial oxygen consumption assays in both HepG2 and PHH.The combined data indicate that CHOP can be used in high-throughput screening assay as a molecular biomarker for mitochondrial perturbation, and, thereby, potential DILI liability.Many studies report the involvement of CHOP in the regulation of pro-apoptotic processes in multiple cell types. 29,30r many FDA-approved drugs, we did not observe significant initiation of cell death upon CHOP-GFP induction in the experimental time frame.CHOP-mediated apoptosis is mainly observed upon induction of the UPR, which in the present study was not the driver of CHOP induction since the canonical UPR responses such as XBP1 and BIP expression were not observed.Selective mitochondrial toxicants can also activate CHOP through induction of ATF4 without activation of the canonical UPR program. 31Recently, in mice cardiomyocytes, CHOP was shown to be a regulator that

F I G U R E 3 F I G U R E 4 F I G U R E 5
Characterization of validated hits.(A) Chord diagram of the validated hits.The upper half of the graph shows the intended target pathway of the identified compounds.The lower half shows the indications for which the compounds are used.Numbers indicate the number of compounds belonging to the mentioned class.(B) Overview of the intended molecular targets of the antimycotic (purple) and neuronal signalling (green) compounds.(C) Chemical similarity of the selected drugs.The similarity score of the selected drugs was calculated using the Tanimoto similarity.(D) Molecular structures of a selected subset of molecules shown at C. (E) DILI association of identified CHOP-GFP positive compounds in dilirank, medline, sider and pharmapendium.Dili_rank column indicates DILI status according to DILIrank (Chen et al. 2016, 21(4):648-653).Colours in the heatmap indicate a positive (red) or negative (blue) association with various DILI pathologies according to medline, sider or pharmapendium.Grey indicates that no data were available for this compound and pathology.Information about these datasets can be found in the supplemental documents.Induction of the unfolded protein response is not the primary mechanism of toxicity.(A) Log2 fold change induction levels of DDIT3 (CHOP) and the upstream regulator HSPA5 (BIP).(B) Eigengene scores of ER stress-associated modules after weighted correlation gene network analysis.A score of ≥2 or ≤ −2 indicated either activation or deactivation respectively.(C,D) Gene networks of modules 62 and 280.Colours indicate log2 fold change levels of genes compared to the negative control (.2% DMSO).Genes in a square indicate the hub gene of a module.Size of the circle/square indicates the correlation eigengene score of a gene for that module.Modules associated with mitochondrial processes are deactivated.(A) Table of compounds belonging to a cluster.A compound was allocated to a cluster if the majority of the samples with these compounds were present in that cluster.If a compound was equally represented in multiple clusters, the allocation of the highest concentration decided the allocation to the cluster.(B) Tanimoto similarity score across clusters 1, 2 and 3. Colours indicate which compound classes were compared; green CNS agent vs. CNS agents, purple antimycotic drug vs. antimycotic drug and blue antimycotic drug vs. CNS agent.(C) Identified modules that contribute to the mechanism of toxicity of cluster 3. Based on the high Pearson correlation score of the module eigengene scores, shown in A, for the antimycotic compounds and selected CNS agents, modules were included if a mean absolute eigengene score of 2 across all shown treatments at the highest concentration was observed and if they were preserved towards rat in vivo or had a clear mechanistic annotation according to Callegaro et al. 16 (D,E) Gene networks of modules 113 and 256.Colours indicate log-2 fold change levels of genes compared to the negative control.Genes in a square indicate the hub gene of a module.Size of the circle/square indicates the correlation eigengene score of a gene for that module.

a
strong inverse relationship between OCR and ATF4 (Pearson correlation −0.74) and CHOP (Pearson correlation −0.6) activity (Figure 6E,F), further indicative that mitochondrial toxicity would induce the expression of ATF4 and CHOP.In concordance with this hypothesis, a CHOP-GFP response could be observed upon treatment of HepG2 cells with mitochondrial toxicants Rotenone and CCCP, without induction of BIP-GFP and XBP1-GFP (Supporting Information Figure S10B).

F I G U R E 6
Treatment with selected antimycotic drugs and CNS agents results in a drop in mitochondrial oxygen consumption and reserve capacity.HepG2 cells were treated with compounds for 24 h with indicated concentrations.Thereafter, the Seahorse assay was performed.(A) Schematic overview of principle of the Seahorse assay.Known mitochondrial toxicants can inhibit different complexes of the oxidative phosphorylation chain.(B) Schematic overview of measurable parameters of the Seahorse assay after treatment with compounds indicated at A. The mitochondrial oxygen consumption rate (OCR) is first measured at basal conditions (black line) or with compounds of interest (orange line).A decrease in OCR upon treatment with the test compound indicates impaired mitochondrial function.Upon the addition of oligomycin, the amount of oxygen used in ATP production can be measured.Furthermore, it also indicates the basal respiration that is not linked to ATP production due to proton leakage.To measure maximum mitochondrial respiration, FCCP is added.The difference between the basal OCR and maximum OCR indicates the reserve capacity of the mitochondria.The addition of rotenone and antimycin A indicates non-mitochondrial oxygen consumption.(C) Mitochondrial OCR upon treatment with indicated compounds.Values were fold change normalized towards the control (.2% DMSO).(D) Mitochondrial reserve capacity upon treatment with indicated compounds.Values were fold change normalized towards the control (.2% DMSO).(E) Correlation plots between mitochondrial OCR and eigengene scores of modules 15 (ATF4) and 280 (DDIT3/CHOP).(F) Schematic overview of the weight-of-evidence approach to elucidate mechanisms of toxicity for antimycotic and central nervous system drugs.Identification of mechanistic hallmarks that drive DILI at early stages during drug development is essential to make informed decisions on liabilities for undesired adversities of drug candidates.Here, we present a high-throughput mechanism-based weight-of-evidence test strategy that can be used to uncover DILI liabilities in relation to mitochondrial toxicity with subsequent activation of the integrated stress response ATF4-DDIT3 pathway.The approach involves an early screening for cellular stress response activation (imagingbased HepG2 BAC-GFP reporter assay), high-throughput transcriptomics for MoT support and tailored mitochondrial functional assays in human-relevant hepatocellular systems (PHH).We have applied this strategy to a large group of FDA-approved drugs, and linked antimycotic and CNS drug profiles to DILI liabilities derived from diverse pre-clinical and clinical datasets on drug adversities.The presented screening assay can be used to rapidly test small molecules for their potency to induce various stress response pathways associated with DILI, but not directly indicative of risks upon clinical use.Therefore, data-driven follow-up experiments should always be considered to gain more information about the risk of DILI and the cellular mechanisms involved.To further improve this strategy, additional reporters indicative of other pathways that are known to be involved in DILI (e.g.oxidative stress, inflammatory response, heat shock response) could be included within the reporter panel to reduce the amount of missed DILI-positive compounds. 25Additional Overview of DILI compound status in the DILI dataset collection and oxygen consumption rates in HepG2 and PHH cell lines.