A metabolomics cell-based approach for anticipating and investigating drug-induced liver injury

In preclinical stages of drug development, anticipating potential adverse drug effects such as toxicity is an important issue for both saving resources and preventing public health risks. Current in vitro cytotoxicity tests are restricted by their predictive potential and their ability to provide mechanistic information. This study aimed to develop a metabolomic mass spectrometry-based approach for the detection and classification of drug-induced hepatotoxicity. To this end, the metabolite profiles of human derived hepatic cells (i.e., HepG2) exposed to different well-known hepatotoxic compounds acting through different mechanisms (i.e., oxidative stress, steatosis, phospholipidosis, and controls) were compared by multivariate data analysis, thus allowing us to decipher both common and mechanism-specific altered biochemical pathways. Briefly, oxidative stress damage markers were found in the three mechanisms, mainly showing altered levels of metabolites associated with glutathione and γ-glutamyl cycle. Phospholipidosis was characterized by a decreased lysophospholipids to phospholipids ratio, suggestive of phospholipid degradation inhibition. Whereas, steatosis led to impaired fatty acids β-oxidation and a subsequent increase in triacylglycerides synthesis. The characteristic metabolomic profiles were used to develop a predictive model aimed not only to discriminate between non-toxic and hepatotoxic drugs, but also to propose potential drug toxicity mechanism(s).

administered vehicle 4 . Treatment was repeated during 4 consecutive days and sample collection and euthanasia were carried out 24 h after the last administration.
Rats were anesthetized with sodium thiobarbital (0.1 g/kg). Livers were removed, rinsed in PBS, divided into small portions, flash-frozen in liquid N2, and stored at -80 ºC until further processing. All the experimental protocols were approved by the Institutional Animal Ethics Committee.

Liver tissue processing for LC-MS untargeted metabolomic studies
Each frozen tissue sample (around 100 mg) was placed in a 2 mL tube containing CK14 ceramic beads and weighted. For each 100 mg of tissue, 650 µL of methanol:water (3:1) containing the IS reserpine (0.375 µg/mL) and sulfadimethoxine (0.075 µg/mL) were added. Then, tissue was homogenized twice for 25 s at 6,000 rpm at 4 ºC in a Precellys 24 Dual system. After centrifugation (3000 g, 5 min, 4°C), the supernatant was transferred to a clean tube. A second extraction was performed with 350 µL per 100 mg of tissue of the solvent. Finally, the two extraction supernatants were pooled and stored at -80 ºC until further processing.
A 100 µL aliquot was transferred to a clean tube and evaporated to dryness using a speedvac. The dry residue was stored at -80 ºC until analysis. The residue was resuspended in 100 µL of water:methanol (1:1) containing 4 µg/mL LCA-D4 as IS. After centrifugation (10 min, 10000g, 4 ºC), the clarified supernatant was transferred to a 96-well HPLC plate and analyzed using the generic-RP analysis conditions in ESI (-) mode.
A 200 µL aliquot was transferred to a clean tube and 100 µL of chloroform containing 0.01 µg/mL terfenadine as IS were added. After vortexing (3 x 10 s), samples were allowed to rest at -20 ºC for 20 min and centrifuged (10 min, 10000 g, 4 ºC). Each phase (the upper aqueous and the lower organic) was separately transferred to a clean tube and evaporated to dryness in a speedvac. The organic phase was resuspended in 100 µL of methanol:chloroform (3:1) containing 0.5 µg/mL verapamil as IS and analyzed using the lipidomic-RP approximation in ESI (+) mode. The aqueous phase was resuspended in 100 µL of acetonitrile:water (70:30) containing 40 µg/mL Phe-D5 and 10 µg/mL Val-Tyr-Val as IS and analyzed using the HILIC approximation in both ESI (+) and ESI (-) modes.
respectively; cone voltage was set at 40 V; desolvation and source temperatures were set at 380ºC and 120ºC, respectively; the flow rates of the cone and nebulization gases were set at 50 L/h and 800 L/h, respectively. The same parameters were applied for the simultaneous MS and MS/MS analyses, with a collision energy ramp from 5 to 60 eV in the MS/MS channel.
Eluent solutions were 0.1% formic acid in water (solvent A) and 0.1% formic acid in acetonitrile (solvent B). A 26-minute elution gradient was run as follows: for the first 2 min, eluent composition was set at 99.9% A and 0.1% B, which was linearly changed to 75% A and 25% B in 4 min; then the proportion of B was increased to 80% over the next 4 min, followed by a further increase to 90% B reached at min 12 and 100% B at min 17, and was maintained for 5.5 min. Finally, the initial conditions were recovered and maintained for 2 min for column conditioning. Mass detection was carried out in the MS scan mode from 50 to 1000 Da in ESI(-).
Eluent solutions were 0.1% formic acid ammonium acetate 10 mM in water (solvent A), and 0.1% formic acid and ammonium acetate 10 mM in acetonitrile:isopropanol (5:2) (solvent B). An 18-minute elution gradient was performed as follows: the initial eluent composition was set at 65% A and 35% B, which was linearly changed to 20% A and 80% B in 2 min; then the proportion of B was increased to reach 100% at min 9 and was maintained for 7 min. Finally, the initial conditions were recovered and maintained for 2 min for column conditioning. Mass detection was run in the MS scan mode from 200 to 1200 Da in ESI (+).
Eluent solutions were acetonitrile (solvent A) and ammonium acetate pH 3 20 mM in water (solvent B).
An 18-minute elution gradient was performed as follows: for the first 3 minutes, eluent composition was set at 95% A and 5% B, which was linearly changed to 75% A and 25% B in 6 min; then the proportion of B was increased to reach 65% at min 13 and was kept for 2 min. Finally, the initial conditions were recovered and maintained for 2.5 min for column conditioning. Mass detection was carried out in the MS scan mode from 50 to 1000 Da in both ESI (+) and ESI (-).

Targeted analysis of oxidative stress markers
Targeted analysis of oxidative stress markers was performed in a Waters Acquity UPLC chromatograph hyphenated to a Waters Xevo TQS mass spectrometer (Waters, UK) by following a previously described LC-MS/MS method 1 . HepG2 cells (70-80% confluence) were treated for 24 h with either control compounds ( Table 1) or hepatotoxins (i.e. tert-butyl hydroperoxide, amiodarone and tetracycline) and processed following the protocol described by Carretero et al 1 .   and in the external validation data set.

Class Model Development External Validation
Control Til5 Til20 Am5 Clo20

Dox250
Tet50 Tet100 Tet400 Val2000 Val4000 Val8000 Dox500 Tet200 See Table 1 for detailed information regarding the characteristics and abbreviation correspondence of each condition. PLS-DA model aimed at the discrimination between HepG2 cells treated with either non-toxic (control) or toxic compounds belonging to either of the mechanisms of hepatotoxicity (i.e., oxidative stress, phospholipidosis, steatosis).   Table S4). The PLS-DA model is developed (and the parameters optimized) using only the model developent data set. Optimization of PLS-DA parameters is performed via cross validation (CV). First the number of latent variables (LVs) is set to that providing the best performance (3LVs).

ANOVA C-OS C-P C-S OS-P OS-S P-S
Then, the variables are ranked according to their VIP value and PLS-DA models with an increasing number of variables are built and their performance evaluated using cross validation. The optimum number of variables in the PLS-DA model is set to that providing the highest figures of merit (n=26).
Thus based on the optimized paramenters a PLS-DA model using 3LVs and 26 retained variables is built.
Model validation is performed using three different strategies: i) Cross validation, that allows to calculate