A targeted lipidomics workflow for high-throughput phenotyping and mechanism-based in vitro hepatotoxicity studies Authors: Tzu-Hsin Yen, Chander K. Negi, Courtney Sakolish, Erin S. Baker, Ivan Rusyn Published: 2026 ABSTRACT Animal-based toxicity testing is limited in throughput and mechanistic characterization, underscoring the need for cell-based approaches. Incorporating molecular readouts, such as lipidomics, can enhance the translational relevance of cell-based assays. Because untargeted lipidomic analyses are time-consuming and can require large sample volumes, a targeted workflow was used in this study to characterize drug-induced liver injury-related lipid responses in primary hepatocytes from multiple species. Multiple reaction monitoring was applied in a targeted lipidomics workflow by coupling liquid-chromatography with triple quadrupole mass spectrometry (LC-MS/MS) for 148 lipid targets across 3 categories and 11 classes. We used primary hepatocytes from human, rat, monkey, and dog cultured in 96-well plates and exposed to chlorpromazine (1–30 μM), bosentan (1–200 μM), and fialuridine (1–100 μM) from culture day 4 to 8. Chlorpromazine induced the most pronounced lipid alterations in human, monkey, and rat hepatocytes, while responses in dog hepatocytes were minimal, with human cells showing a dose-dependent trend up to 10 μM. Bosentan induced dose-dependent lipid changes across species, strongest in human hepatocytes, whereas fialuridine caused limited lipid alterations. Overall, lipidomic effects closely aligned with conventional toxicity markers, reflecting compound and species-specific hepatotoxic sensitivity. The targeted lipidomics workflow provided mechanism-informed phenotyping of chemical-induced lipid responses in cultured hepatocytes. Moreover, it provided a framework for mechanistic evaluation, supported cross-species comparison, and complemented in vitro hepatotoxicity assessments. Introduction New approach methodologies (NAMs) have gained increasing adoption in recent years, driven by the need to enhance the efficiency and predictive performance of preclinical research and to improve the translational relevance of preclinical findings to humans (Wu et al., 2026). Traditional animal-based toxicity testing has long served as the backbone of safety assessments. However, animal-based testing faces limitations including ethical concerns, species differences, high costs, and relatively low throughput. Furthermore, drug candidates that appear safe in animal studies may fail in clinical development due to lack of efficacy or human-specific toxicities, particularly hepatotoxicity (US FDA, 2025). Consequently, there is a growing shift toward human-relevant, mechanistically informed approaches capable of predicting adverse outcomes earlier in the drug development pipeline. NAMs encompass a wide spectrum of technologies, including advanced in vitro models such as spheroids and liver-on-a-chip microphysiological systems (MPS), in silico tools, and integrative testing strategies supported by multi-omics readouts (Marx et al., 2020). Together, these approaches aim to transition safety testing toward mechanistically based, human-relevant, and scientifically rigorous methods (Tieghi et al., 2025; US FDA, 2026). While the number and complexity of cell-based NAMs is increasing, new high-throughput phenotyping strategies are also needed to assess their ability to capture chemical effects at the molecular level. High-throughput transcriptomics has become a mature technology that is being widely used for both mechanistic interpretation and dose-response assessment in vitro (Harrill et al., 2019; Tsai et al., 2023). In addition, lipidomics is increasingly recognized as uniquely informative for studies of liver safety. Dysregulated lipid homeostasis, mitochondrial impairment, phospholipid remodeling, bile acid metabolism, and lipotoxicity are well-established mechanisms of drug-induced liver injury (DILI) and metabolic dysfunction in the liver (Ooi et al., 2021; Rusyn et al., 2021; Weaver et al., 2020). Lipids play a myriad of roles in biology from serving as structural components of cellular membranes (Harayama and Riezman, 2018) to being critical regulators of signaling, energy metabolism, inflammation, oxidative stress, and cell death (Cockcroft, 2021; Morgan et al., 2024; Park et al., 2021; Yoon et al., 2021). Therefore, evaluation of lipid alterations provides a direct window into early toxicological mechanisms rather than downstream consequences such as cell death or inflammation (Rai et al., 2025; Vinken, 2015). For example, changes in phosphatidylcholine and phosphatidylethanolamine (PC/PE) molar ratios have been associated with liver injury severity and may indicate membrane destabilization and cholestatic processes (van der Veen et al., 2017; Wan et al., 2019). Similarly, increased abundances in some ceramide (Cer) species have been linked to activation of apoptotic pathways, whereas changes in certain lysophosphatidyl cholines (LPCs) act as pro-inflammatory signaling molecules (Chitkara and Atilla-Gokcumen, 2025; Li et al., 2020; Sevastou et al., 2013). Because these molecular alterations often precede overt cell death, they provide mechanistic insights and early detection of hepatotoxic potential. Thus, integrating targeted lipidomic analyses with in vitro hepatocyte experiments during the preclinical phase of drug development may further strengthen predictive capability of cell-based assays. Primary human hepatocytes, spheroids, and liver MPS preserve metabolic competence and transporter activity more effectively than traditional immortalized cell lines (Godoy et al., 2013). When combined with these physiologically relevant models, lipidomic readouts can identify pathway-specific toxicity signatures rather than relying solely on phenotypic outcomes. This approach moves beyond simply determining whether a compound is toxic to elucidating mechanism-informed biomarkers. Such understanding will improve safety assessment, support compound prioritization, and guide medicinal chemistry efforts. Despite their promise, metabolomics studies, including lipidomics, are no longer universally considered as “game-changing” technologies by the end-users (Pognan et al., 2023). While valuable for mechanistic research, their integration into regulatory submissions presents challenges. Untargeted lipidomics studies typically generate large datasets requiring complex data processing and structural annotation (Rakusanova and Cajka, 2024). These complex workflows can also introduce uncertainty and decrease reproducibility across laboratories and experiments (Bowden et al., 2017). Additionally, untargeted approaches often detect thousands of potential lipid “features,” many of which lack clear toxicological and mechanistic interpretation, further confounding data interpretation. By contrast, high-throughput targeted lipidomics approaches offer a more practical alternative (Huynh et al., 2023; Meikle et al., 2026; Migni et al., 2026). By focusing on a defined panel of toxicologically relevant lipid targets across multiple classes (e.g., glycerolipids, phospholipids, and sphingolipids) known to be important for liver physiology and/or involved in established mechanisms of liver injury, data interpretation can be streamlined while preserving mechanistic relevance. Furthermore, targeting specific lipid classes enables the use of simpler, faster sample processing methods. This strategy reduces sample volume requirements and accelerates processing, enabling lipidomic characterization using small aliquots from in vitro experiments done in multi-well plates without compromising throughput. The approach proposed herein builds on the advances in targeted lipidomics approaches that are being applied in clinical settings (Huynh et al., 2023; Meikle et al., 2026). Specifically, the targeted lipid workflow reduces sample volume requirements and accelerates data processing, enabling lipidomic characterization from small aliquots. Our specific objective was to develop a high-throughput lipidomics method for phenotyping small volumes of in vitro samples and characterization of drug-induced liver injury (DILI)-related lipid responses in pre-clinical studies. By integrating this method with a NAM-based model, we generated mechanistically informative safety data that can be incorporated into early drug discovery workflow. Thus, the combined human-relevant biological models and targeted molecular readouts provide a pathway toward more predictive, ethical, and efficient safety assessment strategies. Materials and methods Experimental design This study was a cross-species comparison of drug-specific effects on intracellular lipids from selected lipid sub-classes (Table 1) in primary hepatocytes from human, monkey, rat, and dog (Fig. 1A). We used samples from a previously published study (Negi et al., 2025b). Specifically, cells were cultured in 96-well plates (Fig. 1B) and on day 4 were exposed to three known human hepatotoxicants in concentration response: chlorpromazine (CPZ; 1, 10, or 30 μM), bosentan (BOS; 1, 10, 30, or 200 μM), and fialuridine (FIAU: 1, 10, 30, or 100 μM). Lipidomic profiles were analyzed in cell lysates on day 8 of the experiment after two 48-h exposures to test compounds using targeted liquid chromatography coupled with tandem mass spectrometry (LC-MS/MS) using a triple quadrupole mass spectrometer. We focused our targeted analysis workflow on major lipid classes of glycerolipids (DG, TG), phospholipid (PC, PE, lysoPC, lysoPE and ether linked PCs), and sphingolipid classes (SM, Cer, dhCer and dhSM). The lipid classes and individual molecules included in our targeted lipidomics analysis were selected based on their roles in hepatic lipid homeostasis and evidence linking them to hepatotoxicity-associated lipid remodeling. Specifically, this targeted panel was designed to cover several major and complementary lipid classes in liver lipid biology, including membrane phospholipids, lysophospholipids, sphingolipids, and glycerolipds. Phosphatidylcholine (PC) and phosphatidylethanolamine (PE) lipids are the principal membrane lipids and play an essential roles in maintaining membrane integrity and phospholipid balance (Harayama and Riezman, 2018), while disruption of the PC/PE molar ratio has been widely associated with the progression of liver disease (van der Veen et al., 2017). In contrast, lysophospholipids, particularly lysophosphatidylcholines (LPCs) and lysophosphatidyle-thanolamines (LPEs), are bioactive lipid species and commonly regarded as being involved in the regulation of inflammatory responses and stress signaling (Papadopoulos et al., 2022). In addition, ceramides (Cers) and diacylglycerols (DGs) are also highly bioactive lipid classes and are considered major contributors to lipotoxicity (Engin, 2017). Cers have been extensively studied for their role in regulating apoptosis, whereas DGs have been implicated in multiple signaling pathways (Kolczynska et al., 2020; Yu et al., 2025). Triacylglycerides (TGs), on the other hand, are the major form of neutral lipid storage in the liver, and its abnormal accumulation is one of the most common characteristics of steatosis. Previous studies suggest that upregulation of TG synthesis may represent an early buffering response that esterifies fatty acids into TGs to prevent lipotoxicity, as disruption of TG synthesis has been shown to promote lipotoxic injury (Engin, 2017; Mantzaris et al., 2011). When fatty acids are primarily derived from de novo synthesis, the accumulation of TGs with lower degrees of unsaturation may represent a major product of this process (Roumans et al., 2020). Based on these considerations, potential hepatotoxicity-related mechanistic categories were chosen as shown in Table 2, and the selected lipids were further used to derive biologically-informed indicators, including class-level relative abundance changes, lipid class ratios derived from relative abundances, and BioPAN-inferred changes at the level of specific lipid reactions. A full summary of the lipids and their corresponding potential biological indicators is provided in Table 2. Not all monitored lipids, however, were assigned to the predefined mechanistic categories. First, ether lipids are a class of peroxisome-related lipids whose ether bond structure is considered relatively sensitive to oxidative stress, and they may therefore serve as potential indicators of oxidative stress-related alterations (Dean and Lodhi, 2018). However, because these lipids are present at relatively low abundance in the liver and their biological functions remain less clearly defined than those of other major lipid classes, their signals were retained in this study but were not assigned to any specific mechanistic category. Cell culture methods and drug treatment The following is a summary of the methods for studies where the samples originate, those studies were reported previously (Negi et al., 2025b). Primary human hepatocytes (HMCPIS, lot #HU8373) were purchased from Gibco (Thermo Fisher Scientific, Waltham, MA, USA). Cells were thawed in cryopreserved hepatocyte recovery medium (CM7000, Thermo Fisher, USA), collected by centrifugation (100 ×g, 10 min), and resuspended in William’s E media (A1217601, ThermoFisher) supplemented with 3.6% cocktail A, 5% FBS, 10 nM dexamethasone (CM3000, ThermoFisher). Primary rat (Sprague–Dawley; M00005-P, lot #FIM), monkey (Cynomolgus; M003055-P, lot #JCW), and dog (Beagle Dog; M002055-P, lot #BTQ) hepatocytes were purchased from BioIVT (Baltimore, USA). These cells were thawed directly into the INVITROGRO CP medium (BioIVT, USA) types Z990003, Z990028, and Z990025, respectively. All cells were seeded at a density of 70,000 cells per well in collagen I–precoated 96-well plates (200 μg/mL; CB354249, Corning). Plates were incubated at 37 °C with 5% CO2, and culture media were refreshed every 1–2 days. The day of cell seeding was designated as day 0 (D0). After overnight attachment, culture media were replaced with maintenance media consisting of either William’s E medium supplemented with 4% cocktail B and 1 nM dexamethasone (CM4000, ThermoFisher) for human hepatocytes, INVITROGRO CP medium supplemented with TORPEDO antibiotic mix (Z990008, 1 mL per 45 mL medium) for monkey and dog hepatocytes, or TORPEDO Rodent antibiotic mix (Z990027) for rat hepatocytes. Drug treatments with different concentrations of CPZ, BOS, and FIAU were initiated on day 4 and repeated on day 6 of culture. Stock solutions of each compound were prepared in cell culture grade dimethyl sulfoxide (DMSO; BP231–100 Fisher Scientific, Waltham, MA), and working concentrations were obtained by serial dilution while maintaining a final DMSO concentration at 0.1%. Vehicle control wells received only 0.1% DMSO. At the end of the culture period (day 8), media was removed and the plates with cells were stored at −80 ° C prior to lipid extraction. Lipid extraction In-plate lipid extraction of cell lysates was performed using a single-phase extraction approach as described previously (Negi et al., 2025a). In brief, the extraction solvent consisted of isopropanol/methanol (50:50, v/v), supplemented with 6% (v/v) SPLASH LIPIDOMIX Mass Spec Standard (330,707, Avanti Polar Lipids, Alabaster, AL) and 6% (v/v) Cer/Sph Mixture I (LM6002, Avanti Polar Lipids) (Table S1). Internal standards (ISTDs) were added to each sample prior to extraction. Because total area normalization was applied in the primary analysis, the ISTDs were used only for quality assurance to monitor chromatographic performance and instrument stability throughout sample processing and LC–MS analysis. As this study used relative quantitation rather than absolute quantification, and because class-matched calibration standards were not available for all monitored lipids, lipid levels are reported as relative abundances rather than absolute concentrations. Ice-cold extraction solvent (200 μL) was added to each well and plates were sealed with aluminum foil prior to sonication for 2 min and vortexing for 8 min. The plates were then centrifuged at 4000 ×g for 10 min at 4 °C. Subsequently, 160 μL of the supernatant was collected from each well, transferred into 0.3 mL amber glass vials (6PSV9-03FIVAP, Thermo Scientific, Waltham, MA, USA) and evaporated to dryness under a gentle stream of nitrogen (N2) at room temperature in chemical safety cabinet. Dried extracts were reconstituted in 60 μL of isopropanol/methanol (50:50, v/v), followed by vortexing for 5 min and brief centrifugation. A pooled quality control (QC) sample was prepared by aliquoting 10 μL from each reconstituted sample to monitor analytical stability throughout the analytical run. Extracted samples were stored at −20 °C and analyzed within one week (Want et al., 2013). Targeted lipidomic profiling by LC-MS/MS Lipidomic profiling was performed using an Agilent 1290 UPLC Infinity II system coupled to an Agilent 6470 triple quadrupole mass spectrometer (Agilent Technologies, Santa Clara, CA). The LC-MS/MS method was adapted from previously published protocols (Negi et al., 2025a; Negi et al., 2023) and optimized for the current study. Specifically, chromatographic separation was achieved on a Waters ACQUITY Premier CSH C18 column (2.1 × 100 mm, 1.7 μm; Waters Corporation, Milford, MA) that was maintained at 55 °C using an optimized LC–TQMS method. The mobile phase consisted of (A) 60:40 (v/v) acetonitrile/water and (B) 90:10 (v/v) isopropanol/acetonitrile, both supplemented with 10 mM ammonium formate. The flow rate was set at 0.35 mL/min, and the injection volume was 10 μL. The gradient program was as follows: 0–1.86 min, 15–30% B; 1.86–2.32 min, 30–48% B; 2.32–5.0 min, 48–60% B; 5.0–6.0 min, 60–64% B; 6.0–6.8 min, 64% B; 6.8–7.2 min, 64–68% B; 7.2–9.5 min, 68–82% B; 9.5–12.5 min, 82–99% B; 12.5–13.5 min, 99–100% B; 13.5–14.0 min, 100–15% B; and 14.0–16.0 min, 15% B for column re-equilibration. The total run time was 16 min per sample, including column re-equilibration. Mass spectrometric analysis was performed under the following conditions. The source parameters were optimized as follows: gas temperature, 200 °C; gas flow, 12 L/min; nebulizer pressure, 45 psi; sheath gas temperature, 400 °C; and sheath gas flow, 11 L/min. The capillary and nozzle voltages were set at 3500 V and 1000 V, respectively. Data were acquired in positive electrospray ionization (ESI) mode using multiple reaction monitoring (MRM), targeting 148 MRM transitions. Fragmentor voltage and collision energy for each lipid targets were optimized using Agilent MassHunter Optimizer software to maximize sensitivity. MRM transitions were selected based on previously published study (Negi et al., 2025a). For lipid classes that share common fragment ions and contain isomeric or isobaric species (e.g., ether-linked PCs and SMs), RT was also used to further confirm lipid identity (Lee et al., 2023). Detailed information on MRM transitions, including precursor/product ions and optimized parameters, is provided in Table S2. Prior to sample analysis, a pooled QC sample was injected in triplicate to equilibrate the system and evaluate initial performance. Sequence-randomized analysis proceeded only after predefined acceptance criteria were met, including a coefficient of variation (CV) ≤ 10% for internal standard peak areas and retention time deviation within ±0.1 min. Throughout the analytical run, pooled QC samples (see “Lipid Extraction” section) were interspersed at regular intervals to monitor instrument stability and ensure data reproducibility across the entire batch (Want et al., 2013). Spectrum and data preprocessing Raw data acquired from LC-MS/MS were directly imported into Skyline software (version 25.1, MacCoss Lab Software; https://skyline.ms/) for signal extraction and processing (Adams et al., 2020). The Skyline transition list (Table S3) included information for all targeted lipids, including precursor name, precursor m/z, precursor adduct, precursor charge, product m/z, product charge, and retention time (RT), enabling extraction of signal intensities for each target lipid. Integration boundaries for all lipid features were manually inspected and confirmed. Repeated injections of pooled QC samples were used for quality assurance. Specifically, the stability of peak areas for each lipid feature was first evaluated using three consecutive QC injections prior to sample analysis, and features with a coefficient of variation (CV) greater than 10% were excluded. In addition, variability of repeatedly injected QC samples across the entire analytical sequence was assessed, and lipid signals with CVs exceeding 20% were considered unstable and removed. In this study, absolute lipid quantification was limited by the availability of appropriate internal standards across all lipid classes and targets, therefore, total peak area normalization was applied, and lipid levels are reported as relative abundances. However, this limited class-based comparisons. For this study, abundances of the individual lipids are reported in a semi-quantitative manner as relative abundances rather than normalized concentrations. Previous in vitro studies have used one or a limited number of internal standards to represent a lipid class for normalization, together with total protein normalization as an estimate of cell number, to report estimated concentrations (Negi et al., 2025a; Nemergut et al., 2023). However, previous studies have shown that total protein normalization does not effectively improve the reproducibility of lipidomics data for cell number correction. In contrast, data-based normalization strategies, such as sum normalization, showed more stable performance in maintaining data reproducibility (Chao et al., 2017; Leegwater et al., 2025) and we reason that using total peak area normalization and directly reporting relative lipid abundance provides a more sensible strategy for data analysis (Want et al., 2013). The lipidomics workflow used in this study is also reported according to the Lipidomics Standards Initiative minimal reporting checklist (Kopczynski et al., 2024) and included as Supplemental Text. Statistical analyses Statistical analyses were performed using R software (version 4.3.3). Lipid intensities were log-transformed prior to analysis. Statistical analyses were conducted separately for each species and compound. For individual lipid-target analyses, Benjamini–Hochberg correction was applied across all retained lipid targets within each species–compound dataset for the omnibus Welch’s ANOVA. For post hoc analyses, Benjamini–Hochberg correction was applied across all retained lipid targets separately for each dose-versus-vehicle comparison within each species and compound. A lipid target was considered significantly altered only when both the adjusted omnibus Welch’s ANOVA p value and the corresponding adjusted pairwise p value were < 0.05. The same statistical analysis strategy was applied to lipid class-level analyses and relative class abundance ratio analyses. For lipid ontology–level analysis, relative abundance of lipid targets within the same ontology were summed at the sample level, log-transformed, and analyzed using the same Welch’s ANOVA framework. Reaction-level lipid metabolism analysis was conducted using BioPAN on the relative abundance data to infer reaction responses, with reactions showing an absolute Z-score greater than 1.65 considered significant (Gaud et al., 2021). Data visualization was performed using R and GraphPad Prism (version 10.2.2; Prism Software, Laguna Hills, CA). Results To investigate cross-species hepatotoxic response patterns following drug treatment, we evaluated the 148 lipid targets to characterize molecular effects of tested drugs in hepatocytes from different species (Fig. 1A). Fig. 2 and Tables S4–S15 show the overall patterns in lipid alterations in primary human, monkey, rat, and dog hepatocytes following exposure to different concentrations of CPZ, BOS, and FIAU. Almost all selected lipid targets were observed including 140 for human, 142 for monkey, 140 for rat, and 141 for dog. CPZ induced the most pronounced lipid alterations based on relative abundance changes between control and treated samples in human, monkey, and rat hepatocytes; in dog hepatocytes the responses were more muted. In human hepatocytes, a monotonic dose-response relationship was observed up to 10 μM, but this trend was lost at the highest dose, in accord with significant cytotoxicity as reported previously (Negi et al., 2025b) (elevations in alanine aminotransferase (ALT) and lactate dehydrogenase (LDH) at 10 μM, a biphasic albumin response, and severe cytotoxicity at 30 μM) across all measured toxicity markers. In monkey hepatocytes, marked lipid alterations were observed even at low doses, but a clear monotonic dose-response relationship was not evident. These observations closely align with the drug effects on biomarkers of cytotoxicity as reported previously (Negi et al., 2025b). Specifically, ALT was significantly elevated at all concentrations tested, as low as 1 μM while no significant effects were observed for other biomarkers, even though they trended in the direction of hepatotoxicity. In contrast, rat hepatocytes exhibited a monotonic dose-response relationship across all doses consistent with the relatively limited toxicity effects wherein the only significant effect was LDH elevation at the highest concentration tested (Negi et al., 2025b). In dog hepatocytes lot used in this study, lipid alterations were minimal in agreement with the absence of any significant effects on toxicity markers (ALT, LDH, or albumin) in these cells (Negi et al., 2025b). BOS exposure induced notable lipid alterations primarily at the highest dose in human, monkey, and rat hepatocytes. These observations are also consistent with the toxicity marker data (Negi et al., 2025b), wherein LDH was the most sensitive biomarker in human hepatocytes at 1 μM, while albumin, AST, and ALT reached statistical significance only at the highest concentration (200 μM). In dog hepatocytes lot used in this study, lipid changes were largely absent, consistent with the lack of significant toxicity marker responses in these cells. FIAU induced limited changes in lipid profiles across hepatocytes from all tested species. This is consistent with the toxicological observations wherein FIAU exhibited the fewest effects among the three compounds. In human hepatocytes, traditional liver injury markers (ALT, LDH, AST) were largely unaffected, while albumin secretion was significantly suppressed at >30 μM (Negi et al., 2025b), suggesting a decrease in the metabolic function associated with its known human-specific mitochondrial toxicity (Hendriks et al., 2019). In monkey hepatocytes, only ALT showed a significant increase at the lowest concentration tested (1 μM), with no significant effects at higher concentrations (Negi et al., 2025b). No effects were observed in rat or dog hepatocytes for any toxicity markers, in alignment with the limited lipid perturbations observed in these species. Collectively, these observations demonstrate that the lipidomics data align well with the conventional toxicity markers (LDH, ALT, AST, and albumin) with the degree and pattern of lipid alterations reflecting species-specific hepatotoxic sensitivities. Fig. 3A shows the number of significantly altered lipids for each drug treatment and concentration. CPZ exposure resulted in the highest number of lipid alterations with 101 lipids at 10 μM in human hepatocytes; monkey hepatocytes showed a similar number of altered lipids at 1 and 10 μM (89 lipids), whereas rat hepatocytes exhibited the greatest changes at 30 μM (95 lipids). Under BOS exposure, all species showed more pronounced lipid alterations at the highest dose, with the strongest response observed in human hepatocytes, followed by monkey and rat hepatocytes. In contrast, FIAU did not induce significant lipid alterations in human, monkey, or rat hepatocytes. Additionally, dog hepatocytes lot used in this study exhibited only limited lipid alterations under all drug exposure conditions. Since non-monotonic dose–response relationships may be associated with overt toxicity, and previous studies have demonstrated that CPZ at 30 μM exhibits high toxicity in human hepatocytes, 10 μM was selected as the representative dose for CPZ, whereas the highest dose was selected as the representative dose for BOS and FIAU. Fig. 3B presents the magnitude and direction of lipid alterations at the subclass level. Based on the toxicologically relevant events defined in Table 2, we focused on the effects on DG, TG, LPC, Cer, and PC. When treated with CPZ, human hepatocytes showed increases in DG, TG, and LPC, along with a decrease in PC; monkey hepatocytes showed increases in TG and a decrease in PC; rat hepatocytes showed an increase in TG; and dog hepatocytes had no appreciable effects on these lipids. When treated with BOS, human hepatocytes showed increases in DG, TG, and Cer, along with a decrease in PC; monkey hepatocytes showed an increase in TG and a decrease in PC; rat hepatocytes did not exhibit lipid alterations consistent with predefined toxicological events; and dog hepatocytes showed increases in DG and Cer. In contrast, FIAU did not induce lipid alterations at the subclass level in any species. In addition, TGs exhibited chain length- and double bond-dependent alterations, and this pattern was specific to human hepatocytes. Fig. 4 shows that in human hepatocytes, CPZ and BOS exposure led to the accumulation of more saturated TG species with shorter total carbon numbers, whereas TG species with longer carbon chains or a higher degree of unsaturation showed a decreasing trend. In contrast, no such pattern was observed in monkey, rat, or dog hepatocytes. To provide mechanistic insights, we conducted several analyses. First, we used BioPAN approach to perform reaction-level analysis based on lipid relative abundance changes. The resulting lipid reaction responses across species at representative doses are shown in Fig. 5. When treated with CPZ, inhibition of the PE → PC reaction was inferred in both human and monkey hepatocytes; additionally, activation of PC → LPC and inhibition of DG → TG were inferred in human hepatocytes. Furthermore, the dhCer→Cer conversion was consistently inferred across all species under CPZ exposure. When treated with BOS, inhibition of the PE → PC reaction was inferred in human and rat hepatocytes; inhibition of DG → TG was inferred in human, monkey, and dog hepatocytes; and activation of TG was inferred in rat hepatocytes. Under FIAU exposure, inhibition of the PE → PC reaction was inferred in monkey, rat, and dog hepatocytes, as was inhibition of DG → TG in dog hepatocytes. Second, given the observed effects on the lipid targets and classes, we also explored whether intensity-derived lipid class ratios can provide additional information as empirical indicators of the drug effects. Fig. S1 shows changes in PC/PE, TG/DG, LPC/PC, and Cer/SM. When treated with CPZ, both human and monkey hepatocytes showed decreased PC/PE; in addition, human hepatocytes showed decreased TG/DG and increased LPC/PC. Notably, at 30 μM CPZ exposure, a decrease in TG/DG was also observed in rat hepatocytes. When treated with BOS, human hepatocytes showed decreased PC/PE, decreased TG/DG, and increased Cer/SM. Rat hepatocytes showed decreased PC/PE and increased TG/DG; however, the decrease in PC/PE was modest, and neither change exhibited a clear dose–response trend. Dog hepatocytes showed an increase in Cer/SM. In contrast, FIAU exposure resulted in minor decreases in PC/PE only in monkey hepatocytes. To integrate the lipid alterations identified across different analyses, Table 2 and Fig. 6 provide a comprehensive summary of lipid-derived signatures suggestive of hepatotoxicity inferred from targeted lipidomics data in hepatocytes from four species exposed to representative doses of CPZ, BOS, and FIAU (10, 200, and 100 μM, respectively). Overall, Fig. 6 shows that CPZ exposure induced more evident toxicologically-relevant lipid features in human and monkey primary hepatocytes. In human hepatocytes, the lipid alterations were consistent with features related to membrane damage, inflammation, and lipotoxicity, whereas in monkey hepatocytes, the observed changes were mainly more consistent with membrane damage. In contrast, rat hepatocytes did not show obvious lipid alterations corresponding to the predefined toxicological events at 10 μM CPZ. However, it should be noted that the most pronounced lipid response in rats occurred at 30 μM. No obvious lipid alterations corresponding to these toxicological events were observed in dog hepatocytes lot used in this study. In response to BOS exposure, lipid alterations observed in human hepatocytes showed a greater degree of correspondence with the predefined toxicological features than those in other species, mainly manifesting as changes associated with membrane damage, as well as some apoptosis-related features. The other species did not exhibit lipid alterations that fully corresponded to any single toxicological event; only a few similar trends were observed. Notably, despite showing the fewest overall lipid changes, dog hepatocytes displayed the second-highest overall consistency with the predefined toxicological mechanisms after human hepatocytes, exceeding that of monkey and rat hepatocytes. In contrast, FIAU exposure did not induce fully corresponding lipid responses to the predefined toxicological events in any species. For our final comparison, we explored potential concordance between lipid alterations and commonly used liver toxicity biomarkers (ALT, LDH, AST, and albumin) across species (Fig. 7). The correlation patterns were largely consistent with both the lipidomics response observations and the toxicity biomarkers. For example, in human hepatocytes, albumin showed a significant positive correlation with PE, suggesting that alterations in PE may be associated with hepatic functional decline. This is consistent with the observed suppression of albumin secretion, particularly at higher concentrations of CPZ and FIAU, where albumin was identified as a sensitive functional marker of hepatotoxicity (Negi et al., 2025b). Additionally, LDH and AST both showed negative correlations with PC indicating that hepatocellular injury progression, which is reflected by elevated LDH and AST, is also associated with membrane disruption and loss of membrane integrity as indicated by a decrease in PC. In monkey hepatocytes, ALT showed a significant positive correlation with P-PC suggesting that PC remodeling may be closely associated with hepatocellular injury in this species. This aligns well with the observation that ALT was the most sensitive toxicity biomarker in monkey hepatocytes, showing significant elevations at concentrations as low as 1 μM for CPZ. The strong ALT P-PC relationship may also indicate that oxidative stress-related membrane phospholipid remodeling can be interpreted as an early event in drug-induced hepatotoxicity in monkeys. For rat hepatocytes, albumin showed a significant positive correlation with Cer, suggesting that Cer accumulation is associated with hepatic functional changes in this species. Additionally, ALT showed a negative correlation with PC indicating that as hepatocellular injury developed, PC levels decreased, consistent with membrane remodeling under cytotoxic stress. These observations align with the toxicological response in rat hepatocytes, where LDH was the only significantly affected toxicity marker, and only at the highest concentration tested suggesting that lipid-biomarker relationships in rats reflect a more hepatotoxic response. In dog hepatocytes lot used in this study, no significant correlations were observed, consistent with the absence of significant toxicity marker responses in this species. This further supports the notion that dog hepatocytes are the least sensitive species to the hepatotoxic effects of the tested compounds. Beyond lipid-toxicity marker relationships, several significant lipid-lipid correlations were also observed across species. Notably, sphingolipid subclasses (dhSM, Cer, dhCer) showed strong positive associations in human and rat hepatocytes, suggesting potential coordinated regulation of sphingolipid metabolism under drug-induced stress. Similarly, phospholipid subclasses (PE, PC, P-PC, O-PC) showed significant positive associations particularly in rat hepatocytes, potentially reflecting coordinated membrane phospholipid remodeling in response to hepatotoxic insult. These lipid-lipid relationships suggest that drug-induced hepatotoxicity may trigger broader and interconnected lipid metabolic reprogramming. Overall, the species-specific correlation patterns between lipid subclasses and toxicity biomarker are well aligned with the hepatotoxic sensitivity profiles observed across species with the strongest and most significant lipid-toxicity marker correlations observed in human and monkey hepatocytes the most toxicologically sensitive species and an absence of significant correlations in dog hepatocytes, consistent with its lack of hepatotoxic response. Discussion This study presents a targeted lipidomics workflow that is applicable for molecular phenotyping of hepatotoxic responses in primary hepatocytes from human, monkey, rat, and dog. The targeted lipid workflow is based on recent clinical advances in lipidomics, providing a fast and mechanistically informative approach to the analysis of complex biological samples (Huynh et al., 2023; Meikle et al., 2026). Additionally, this study extends targeted lipidomics to small sample in vitro research and screening applications. In previous in vitro hepatocyte lipidomics studies, untargeted high-resolution lipidomics provided broader lipid coverage and more extensive structural evidence for lipid annotation but required longer acquisition times and more complex sample preparation and data-processing workflows (Migni et al., 2025). By focusing on a predefined panel of 148 lipid targets spanning phospholipids, lysophospholipids, sphingolipids, ether-linked phospholipids, and glycerolipids, the workflow balances analytical throughput with biological interpretability, specifically for DILI-related contexts of use. Furthermore, the drugs of interest were selected to represent different mechanisms of DILI and also species-specific sensitivity. Therefore, lipid perturbations were considered parallel with conventional indicators of hepatocellular injury (Negi et al., 2025b), supporting the value of targeted lipidomics as a complementary readout in in vitro liver safety studies. One key finding of this work is that CPZ, a drug known to induce liver injury via both cytotoxic and cholestatic mechanisms in all tested species (Greer et al., 2010), elicited the most pronounced lipidomic responses across species. This effect was most pronounced in human hepatocytes, where increases in DGs, TGs, and LPCs were observed while PCs decreased. These perturbations were broadly consistent with known effects of hepatotoxic stress on membrane lipid remodeling, neutral lipid handling, and bioactive lipid signaling (Buratta et al., 2021). Specifically, decreases in PC relative abundance and in the PC/PE relative abundance ratio are relevant because PC and PE are major membrane phospholipids, and disruption of their balance has been associated with impaired membrane integrity and liver disease progression (van der Veen et al., 2017). Likewise, increases in lysophospholipids may reflect phospholipid remodeling and inflammatory signaling processes, as LPCs are widely recognized as bioactive lipids involved in stress and inflammatory responses (Knuplez and Marsche, 2020). The CPZ response also highlighted the potential utility of glycerolipid patterns as mechanistically informative indicators (Scott et al., 2014). In human hepatocytes, CPZ exposure was associated with increased DG and TG relative abundance, but also with a decreased TG/DG relative abundance ratio and inferred suppression of the DG➔TG reaction at the representative dose, suggesting that interpretation depends not simply on whether neutral lipids accumulate, but on the balance between potentially lipotoxic intermediates and storage lipids. This distinction is biologically important because DGs are bioactive signaling lipids linked to lipotoxic stress, whereas TG accumulation can in some contexts represent an adaptive buffering response that sequesters excess fatty acids and reduces cellular injury (Listenberger et al., 2003). The species-specific TG behavior was also observed in this study, including chain length and unsaturation changes to the lipids in the human hepatocytes. This further suggests that class-level summaries alone may obscure biologically relevant remodeling patterns. CPZ is also a known sphingomyelinase inhibitor, and suppression of sphingomyelin-to-ceramide conversion has been reported as a mechanistically expected response to this drug (Albouz et al., 1986). However, this effect is unlikely to be directly associated with CPZ-related hepatotoxicity, as sphingomyelinase inhibition is specific to the parent compound, whereas CPZ-induced hepatotoxicity is thought to be largely driven by CYP-generated reactive metabolites that lack sphingomyelinase inhibitory activity (Albouz et al., 1986; MacAllister et al., 2013). Consistent with this mechanism, CPZ exposure across species resulted in a pronounced decrease in the Cers/SM relative abundance ratio. Notably, this response was not included in our predefined toxicity-mechanism categories and was therefore not interpreted as a toxicity-related event. This indicates that our workflow did not misclassify this large-effect-size lipid response as toxicity-related simply because it showed a pronounced change. BOS is a drug known to cause DILI primarily by cholestatic mechanism (Fattinger et al., 2001) and it produced a more selective lipid response that was most pronounced in human hepatocytes. This response included ceramide accumulation, increased Cer/SM relative abundance ratio, decreased PCs, and decreased PC/PE relative abundance ratio. These findings are notable considering the known human sensitivity to BOS-associated liver injury and are consistent with the view that lipidomic profiling can capture species-dependent toxicological liabilities. Cers are of particular interest because they are sphingolipid mediators implicated in stress signaling and apoptosis (Hajduch et al., 2021), and increased Cer relative abundance has been associated with BOS-induced hepatotoxicity (Cuykx et al., 2019). At the same time, the Cer-related changes may be interpreted as evidence of sphingolipid remodeling, but our data demonstrating altered relative abundance of selected Cers and lipid relative abundance ratios should not be interpreted as evidence for increased de novo ceramide synthesis or causal involvement in cell death. In contrast, FIAU which is a mitochondrial toxicant that was missed in pre-clinical animal studies but caused severe hepatotoxicity in a clinical trial (McKenzie et al., 1995), induced minimal lipidomic perturbation in this in vitro model. This muted response likely reflects an important limitation of the present study design and choice of targeted lipids, rather than a lack of toxic potential of FIAU. FIAU hepatotoxicity is strongly linked to mitochondrial dysfunction (Lewis et al., 1996) and is more readily captured in cell-based models that can enable prolonged exposure and have sustained metabolic competence (Foster et al., 2019; Negi et al., 2025b). In these studies, the primary indicators of potential hepatotoxicity were decreases in albumin synthesis, a stress response that is a leading indicator of mitochondria stress. Mitochondrial dysfunction is known to result in profound changes in lipid catabolism resulting in lipid accumulation (Moschandrea et al., 2024); however, our data shows that longer study duration may be needed to observe this effect in vitro. The lack of response observed here may also result from the limitations of the targeted lipid panel itself, which may not sufficiently cover lipid signatures indicative of mitochondrial injury. This point is important for the future use of the proposed workflow as the targeted lipidomics approach can be highly informative for some mechanisms of liver injury, but no single lipid panel is likely to capture all forms of hepatotoxicity equally well. An important strength of this study is demonstration of the utility of the proposed workflow to a study of primary hepatocytes from four species under equivalent experimental conditions. Comparative analysis across species revealed that lipid perturbations were well aligned with known species differences in sensitivity to tested model DILI compounds, with stronger and more numerous responses in human hepatocytes for BOS and in human/monkey hepatocytes for CPZ, while dog hepatocytes were comparatively less responsive. Such cross-species data on in vitro models are valuable because species differences remain a major challenge in preclinical liver safety assessments yet most of the NAM development focuses on human cell-based models. Likewise, most of the comparative cross-species lipidomics data is on plasma levels, showing that while major species differences exist (Ishikawa et al., 2015), the plasma profiles of major lipid classes between human and rodents also have great similarities (Kaabia et al., 2018). In this context, our proposed targeted lipidomics workflow may be useful not only as a descriptive phenotyping tool but also as a means of improving mechanistic interpretation of translational studies. While there are practical advantages of targeted lipidomics (Huynh et al., 2023; Meikle et al., 2026) because it can simplify data processing and support higher-throughput analysis from small sample volumes, including material derived from 96-well plate experiments in early discovery workflows, there are a number of limitations. The targeted approach used herein does not amount to comprehensive lipidome coverage which may produce incomplete picture of the role that the lipids may play in the hepatotoxicity phenotype. We also used positive electrospray ionization that preferentially captures glycerolipids, choline-containing phospholipids, selected ethanolamine lipids, and sphingolipids, but misses important negative-mode lipid classes. Furthermore, even though the use of lipid class relative abundance ratios is a helpful approach to molecular phenotyping, it is an approach that can mask molecular detail, may lack mechanistic specificity, and may be prone to methodological bias (Burla et al., 2018). Indeed, summing lipid relative abundances into classes and interpreting class ratios can mask significant shifts at the level of individual molecular species, especially for low relative abundance species or those that do not ionize well. Some pathological or signaling mechanisms are reflected in specific lipid species, not in bulk class levels. Therefore, we propose to use the workflow detailed herein as a complement to species-level or pathway-level analysis, and as part of the overall weight-of-evidence with other biochemical and molecular data that will be measured in the in vitro studies of DILI, not as an independent biological readout. The scope of our study was limited to the evaluation of one donor lot from each species and specific experimental designs; therefore, the broad utility of the proposed method cannot be assumed. Additional limitations to the proposed approach are listed below. First, the current workflow relied on normalized signal intensities rather than absolute quantification for each lipid targets. While our approach is suitable for relative comparisons within the study, ionization efficiency differs substantially across lipid classes and species (Yang and Han, 2011), which limits the quantitative interpretation of summed class relative abundances and derived ratios. As a result, lipid class relative abundance ratios such as PC/PE, TG/DG, LPL/PL, and Cers/SM should be interpreted here as empirical indicators within this assay rather than direct molar estimates. To evaluate the sensitivity of the findings to the normalization strategy, the data from monkey hepatocytes were additionally analyzed using class-matched internal-standard normalization (Table S16). The magnitude and direction of treatment-associated changes were consistent between the two normalization approaches, although the number of statistically significant lipid targets differed for some treatments. These differences coincided with greater within-group variability after internal-standard normalization, which may be partly because the cell numbers were not quantified after drug treatments. Future research may need to introduce class-matched calibration and more explicit normalization to internal standards, combined with correction for between-well differences in cell number, to support semi-quantitative or quantitative reporting. Second, mechanistic assignments for many lipid in the publicly available databases remain inferential. The predefined categories in Table 2 are useful for organizing overall mechanistic interpretation; yet we acknowledge that changes in lipids are dynamic, multifunctional and context-dependent (Hornburg et al., 2023), and the same directional change may reflect adaptation in one setting and toxicity in another. For example, TG accumulation may indicate steatotic remodeling but may also represent a transient protective sink for excess fatty acids (Listenberger et al., 2003). Similarly, changes in LPC, Cers, or PC/PE relative abundance ratio are compatible with inflammation, lipoapoptosis, or membrane damage (Law et al., 2019), but lipidomics alone cannot establish pathway activation without orthogonal evidence. For this reason, the mechanistic interpretation offered by targeted lipidomics is best viewed as hypothesis-generating and most powerful when integrated with other injury markers, transcriptomics, imaging, or functional readouts. For example, the lipidomics data collected in this study were broadly concordant with traditional hepatocellular injury markers, supporting the biological relevance of the targeted lipid panel. In general, compounds and species that showed stronger effects on conventional endpoints such as ALT, AST, LDH, and albumin also exhibited more numerous and pronounced lipid alterations, whereas minimal lipidomic perturbation was seen in conditions with weak or absent injury-marker responses. This agreement was most evident for CPZ and BOS, which produced both marked lipid remodeling and corresponding changes in biochemical injury markers, particularly in human hepatocytes, while FIAU caused comparatively limited effects in both datasets. In addition, several lipid classes were significantly correlated with injury markers across suggesting that lipidomic endpoints may complement traditional biomarkers by providing mechanistic context to hepatotoxic responses. Indeed, our previous study that explored biochemical and transcriptional effects of these three drugs in the same samples showed value of such combination in terms of both biomarker discovery and mechanistic interpretation (Negi et al., 2025b). Thus, addition of high-throughput lipidomics to other methods and its ability to be informative in small sample volumes from the same experiment would greatly enhance the value of the comprehensive model-omics approach desired by the pharmaceutical industry as a prerequisite for the adoption of NAMs (Homan, 2023). Third, the biological scope of the lipid panel used in this study was intentionally narrow. Although selected lipids cover several major pathways relevant to liver physiology and hepatotoxicity, not all monitored lipids were assigned to mechanistic categories, and some potentially informative lipid classes remain underrepresented. We note that the biological interpretation and annotation challenges of ether lipids in this context introduce the greatest uncertainty. Further expansion of the targeted lipid panel could improve sensitivity for oxidative stress, mitochondrial dysfunction, and bile-related or transporter-mediated mechanisms, depending on the intended contexts of use. Overall, the findings presented in this study support the conclusion that targeted lipidomics can provide rapid, mechanistically informative phenotyping of compound-induced liver responses in vitro. The approach appears especially useful for detecting perturbations linked to phospholipid remodeling, sphingolipid changes, and glycerolipid consumption, and for determining species-dependent differences in response to known hepatotoxicants using an in vitro model, rather than studies in animals. Targeted lipidomics also shows promise when combined with advanced hepatocyte-based models and both conventional and transcriptomic in vitro biomarkers of function and toxicity. Specifically, targeted lipidomics may strengthen mechanistic interpretation in preclinical screening and help both prioritize compounds for follow-up investigation and de-risk compounds that show effects in vivo. Thus, future studies should focus on improving quantitative rigor, expanding mechanistic coverage of the lipid panel, and benchmarking lipid signatures against orthogonal endpoints and longer-term more complex tissue chip culture systems to better define where this approach adds the most predictive value in drug-induced liver injury assessment. Supplementary Material