In previous studies, HepaRG cells have been successfully used to analyze chemically induced changes in hepatic lipid metabolism, mainly focusing on neutral fatty acid changes (Karaca et al. 2023; Lichtenstein et al. 2021). So far, only few studies have investigated PLD induction in HepaRG cells, and if so, a focus has been placed on few selected substances (Anthérieu et al. 2011; Tomida et al. 2017; Zhang et al. 2022). To our knowledge, this is the first study to investigate phospholipid accumulation in HepaRG cells using the LipidTOX assay as an in vitro cell-based high-throughput method for predicting PLD for a large number of compounds.

In the present study, the assay performance identifying PLD-inducing compounds was evaluated. A panel of 35 substances was examined for their ability to induce phospholipid accumulation. The tested compounds comprise a selection of PLD-inducing substances, with the majority being classified as CADs, as well as a set of non-PLD-inducing substances. Compared to in vivo data, results revealed a sensitivity of 95% and a specificity of 73% with an overall predictivity of 86% (Table 2). Although the results accord very well with in vivo findings on PLD induction (Table 1), some discrepancies were observed between in vitro and in vivo data. As an example, disopyramide and methapyrilene exhibited PLD induction in our assay, as well as in HepG2 cells (Tilmant et al. 2011; van de Water et al. 2011), but no induction in rats (Sahini et al. 2014; Yudate et al. 2012). Those discrepancies are not unexpected as the majority of in vivo data originates from rodent experiments which may exhibit limited comparability to human cells. One reason could be that oral absorption rates in vivo limit compound bioavailability, which would impair the identification of substance-specific effects. Moreover, the detection of PLD in vivo is mainly done by electron microscopy. This method provides high resolution, but has only a low throughput without quantification, and bears risks of subjective interpretation sampling bias, which could cause false negative categorization. Another factor might be that the ability of certain compounds to induce PLD might differ between humans and other species. It is well established that the activity and expression of metabolic enzymes involved in the biotransformation of substances and the expression and regulation of hepatic drug transporters mediating drug uptake and efflux differ across species, which impacts substance uptake, metabolism, and clearance (Martignoni et al. 2006; Wang et al. 2015). Interspecies differences become also apparent in the context of PLD considering in vitro studies comparing PLD induction in the same cell type of different species. For example, human and murine macrophages exhibited in vitro different phospholipid and lysosome accumulation (Öhlinger et al. 2020). Bhandari et al. could also demonstrate differences in PLD induction in vitro between the human HepG2 cell line and hepatocytes from rats and rhesus monkeys, confirming inter-species differences in PLD induction (Bhandari et al. 2008). Therefore, comparisons of in vitro data on PLD induction in this work have been limited to in vitro studies utilizing human-derived hepatic-like cells, while in vitro studies using other cell types have not been included in the comparative analyses.

To overcome the limitations of in vivo assessment of PLD, attempts have made to predict the PLD-inducing potential of chemicals with in silico NAMs, as fast, cost-efficient, and animal-free alternatives. Published in silico models focus mainly on physicochemical ClogP and pKa calculations of substances, as those seem to be the shared characteristics among CADs (Fischer et al. 2000). Here, in silico methodologies discussed are the Ploemen model utilizing basic physicochemical calculations (pKa and ClogP) (Pelletier et al. 2007; Ploemen et al. 2004), the Bayesian model which incorporates pKa, ClogP, amphiphilic moment, number of basic and acidic centers, structural and atom classes (Pelletier et al. 2007), and the model by Hanumegowda focusing pKa and ClogP as well as tissue distribution factors (Hanumegowda et al. 2010). As shown by the good concordance of those models compared to in vivo data (75–88% predictivity), in silico NAMs offer a powerful and efficient tool for PLD prediction. The advantage of their high throughput and cost-effectiveness enables rapid screening of large compound libraries for early risk identification and prioritization. However, looking at the here investigated substances, in silico predictions show several false negatives when compared to in vivo data (Table 1), which negatively impacts method sensitivity. As all models are based on physicochemical characteristics, they probably can not fully capture all biological mechanisms leading to PLD (esp. for non-CAD compounds); this is why relatively low ClogP and tissue distribution values lead to false negative predictions as for several antibiotics, such as amikacin, erythromycin, and gentamicin (Hanumegowda et al. 2010). The inclusion of amphiphilic moments, number of basic and acidic centers, structural and atom classes seem to improve the prediction for this substance class (Pelletier et al. 2007). Interestingly, the negative PLD induction for amikacin aligns with results of the here established in vitro assay in HepaRG cells. Moreover, also for disopyramide and methapyrilene, which displayed PLD induction in HepaRG cells, were positively predicted by one of the in silico models. Moreover, factors like bioavailability, substance accumulation, and xenobiotic metabolism are neglected in most prediction tools which might result in false classification (Hanumegowda et al. 2010). For example, ketoconazole, haloperidol, and erythromycin are metabolized by various cytochrome P450 enzymes. Their metabolites can differ in their PLD-inducing properties from the parent compound, as for ketoconazole, which is extensively metabolized to de-N-acetyl ketoconazole, a potent PLD inducer (Whitehouse et al. 1994). In silico methods also lack information on the potency of substances to induce PLD, and the models are not suitable for assessing mixture effects. While in silico NAMs are highly useful for early-stage screening, their limitations advocate their use in combination with experimental approaches to enhance overall predictive accuracy for PLD.

In vitro assays represent a good alternative to in vivo and in silico methods. Previous to the here established assay, several studies have been published using the human liver carcinoma cell line HepG2 with different PLD detection methods. For the comparative analysis of PLD induction in vitro, only those studies which tested a larger set of substances and not only individual compounds were considered in this work. Across these studies, the reported accuracies were ranging from approximately 80% to 100%. Data obtained from the different HepG2 studies discussed in this work revealed differences in PLD induction potency among substances, similar to the results observed in HepaRG cells. Comparing results per substance, the induction strength varied between the different HepG2 assays leading partly to a categorization of three different PLD induction classes for the same substance (Table 1). The selection of seeded cell numbers, used dye, and incubation conditions seems to influence assay performance and measured PLD induction potency in HepG2 cells. Moreover, the HepG2 cell line is deficient in many liver-specific functions essential for mediating cellular responses to toxicological stressors, such as sufficiently high levels of xenobiotic-metabolizing enzymes (Guguen-Guillouzo & Guillouzo 2010). This limitation is particularly relevant, as metabolites can exhibit different toxicological effects compared to the parent compound. For instance, haloperidol and erythromycin do not exhibit PLD induction in the studies conducted by Sawada et al. and van de Water et al., although they are classified as PLD inducers according to in vivo data. In this study, haloperidol and erythromycin displayed PLD induction in HepaRG cells indicating the importance to use a model that provides xenobiotic metabolism functionality for PLD prediction. Comparing our findings with PLD induction in HepaRG cells with the LYSO-ID Red dye revealed a high degree of concordance for the majority of compounds tested in both assays, including induction potencies (Tomida et al. 2017). In contrast to this study, Tomida et al. reported no PLD induction for dexamethasone, but the tested concentrations in their study were considerably lower. The strong concordance between Tomida’s and our study emphasizes the reproducibility and utility of HepaRG cells for predicting PLD.

A key advantage offered by in vitro methods in comparison to other methodologies is the possibility to systematically investigate PLD induction potency and its concentration-dependent characteristics. As observed in HepaRG cells, curve shapes of concentration-dependent PLD induction in HepG2 cells seem to differ amongst substances. For the majority of the substances, a continuous increase in phospholipid content with rising concentration was observed, while for others, like chloroquine and clomipramine, the phospholipid content plateaued at a certain level (Nioi et al. 2007; Park et al. 2012). The appearance of plateaus may be due to a saturation threshold within cells beyond which no additional phospholipid accumulation is possible. However, these plateaus are primarily observed at phospholipid levels substantially lower than those associated with maximal PLD induction for other compounds. For example, chloroquine treatment reveals a plateau at 300% phospholipid content, whereas tilorone induces a maximum of 600% (Fig. S2). Therefore, the differences in PLD curve progressions could be an indicator for different modes of action instead of an indicator for saturation with phospholipids. Even though, no studies have been published correlating this observation to underlying mechanisms of PLD or specific chemical groups. In this study, PLD induction by some substances displayed a maximum accumulation followed by a decline in phospholipid content at higher concentrations. A drop in the phospholipid content was previously associated with cytotoxic effects (Bhandari et al. 2008). In contrast, viability measurement in HepaRG cells in our study revealed no cytotoxic effects (Fig. S1), suggesting the involvement of alternative mechanisms. Moreover, PLD induction in HepaRG cells using the LYSO-ID Red dye exhibited curve shapes similar to what has been observed in our study (Tomida et al. 2017). This reinforces the hypothesis that the appearance of plateaus or a declining content of phospholipids correlates with different underlying modes of action instead of saturation and cytotoxicity and is highlighting the complexity of PLD induction and the utility of in vitro HepaRG assays.

While applied concentrations of test compounds in vitro might exceed their typical serum levels in vivo, the objective of this study was to confirm the utility of the assay as a NAM, thereby providing framework for future refinement and broader applicability, including its potential use with non-pharmaceutical chemicals. This assay holds significant promise for future applications, such as the pre-screening of substances prior to animal testing, the evaluation of testing substance batteries to assess metabolic disruption of hepatic lipid metabolism, the investigation of effects of compound mixtures, and mechanistic studies aimed at elucidating the modes of action of PLD. The reliability of the here developed in vitro assay assessing PLD accumulation in HepaRG cells strongly supports the utility of in vitro NAMs. Furthermore, these findings emphasize the importance of integrating human-relevant in vitro models, such as the HepaRG cell line, alongside with in silico methods within a framework of NAM-based testing strategies.