Biomarkers of drug-induced liver injury: a mechanistic perspective through acetaminophen hepatotoxicity.

Abstract

Introduction:

Liver injury induced by drugs is a serious clinical problem. Many circulating biomarkers for identifying and predicting drug-induced liver injury (DILI) have been proposed.

Areas covered:

Biomarkers are mainly predicated on the mechanistic understanding of the underlying DILI, often in the context of acetaminophen overdose. New panels of biomarkers have emerged that are related to recovery/regeneration rather than injury following DILI. We explore the clinical relevance and limitations of these new biomarkers including recent controversies. Extracellular vesicles have also emerged as a promising vector of biomarkers, although the biological role for EVs may limit their clinical usefulness. New technological approaches for biomarker discovery are also explored.

Expert Opinion:

Recent clinical studies have validated the efficacy of some of these new biomarkers, cytokeratin-18, macrophage colony stimulating factor receptor, and osteopontin for DILI prognosis. Low prevalence of DILI is an inherent limitation to DILI biomarker development. Furthering mechanistic understanding of DILI and leveraging technological advances (e.g. machine learning and omics approaches) is necessary to improve upon the newest generation of biomarkers. The integration of omics approaches with machine learning has led to novel insights in cancer research and DILI research is poised to leverage these technologies for biomarker discovery and development.

1. Introduction

Drug-induced liver injury (DILI) is initiated by the parent xenobiotic or its associated metabolites causing unanticipated damage either to hepatocytes or the non-parenchymal cells of the liver [1]. The liver is particularly sensitive to injury due to the magnitude of xenobiotic exposure and its critical role in the metabolism of drugs following oral administration. There are two primary classifications for DILI: intrinsic and idiosyncratic. Intrinsic DILI is typified by its dose dependency with a predictable response across a large proportion of individuals whereby the latency is relatively short ranging from hours to days. Idiosyncratic DILI (iDILI) is generally not dose-dependent and occurs only in a select population of individuals exposed to the drug with a latency period that is difficult to estimate. For these reasons, clinical trials designed to establish safety are often underpowered to capture these rare events of idiosyncratic hepatoxicity [2].

DILI is the most common cause of acute liver failure (ALF) [3,4] with APAP overdose being responsible for 46% of all ALF cases in the US and iDILI causes approximately 11% of the cases [5]. The challenge in clinical practice is to identify if this is a DILI case and which drug is the culprit. In the pharmaceutical industry, DILI is a frequent reason for stopping the development of a drug and is a major cause for withdrawal of approved drugs from the market throughout the world [6]. Due to the financial cost of bringing new drugs to the market in conjunction with the propensity for DILI to curtail the success of the new drug, there has been an increasing incentive for novel biomarker discovery. However, many of these biomarkers remain firmly in the realm of preclinical models with limited understanding of their practical utility in the clinic. Mechanistic insights into the hepatotoxicity of various drugs such as acetaminophen (APAP) has spurred interest in using damage associated molecular patterns (DAMPs) as biomarkers such as, high mobility box group box 1 (HMGB1), mitochondrial DNA (mtDNA), nuclear DNA fragments (nDNA) and other mitochondrial proteins [7]. Many approved drugs have known drug toxicities that target critical cellular functions such as cellular respiration, lending potential credence to the validity for using mtDNA or mitochondrial proteins in circulation as biomarkers. Some of these mitochondria associated biomarkers have been validated clinically, although others have not. There also is a surge in interest for using circulating microRNAs, such as miR-122 and 192 [8–10]. More recently, significant interest has been targeting the capacity for plasma extracellular vesicles (EVs) to serve as biomarkers [11] because EVs can encapsulate molecules such as miR-122 [12] shielding them from the extracellular environment and improve stability of these molecules [13]. A new paradigm focusing on biomarkers at the intersection of inflammation and liver regeneration, such as macrophage colony stimulating factor receptor (MCSFR) and osteopontin (OPN), are finding significant traction. We will discuss the rationale for these preclinical biomarkers, their potential for translational value, and the recent controversies surrounding them. Finally, we will discuss how the integration of machine learning with omics is an exciting and unexplored approach to biomarker discovery in DILI.

2. Acetaminophen: The Classic Example of Intrinsic DILI

Acetaminophen (APAP) is the primary ingredient in many over-the-counter analgesic/antipyretic medications. Further, APAP is often combined with more powerful analgesics such as hydrocodone, leading to an even greater number of individuals exposed to APAP. Thus, overdoses of APAP account for nearly 50% of all ALF in the United States [14]. Correspondingly, while half the cases involving APAP-induced ALF are due to large single overdoses, nearly 50% are the result of unintentional overdoses from individuals consuming moderate supratherapeutic doses over several days [15]. Increased susceptibility to moderate APAP overdoses may be caused by concomitant use of other drugs, fasting or chronic alcohol abuse. Often times, these various factors are targeting two central prerequisites to APAP toxicity: CYP2E1 activity and glutathione (GSH) status. APAP is rapidly absorbed in the gut and is transferred to the liver where it is metabolized mainly by phase II conjugation reactions and only a minor fraction is oxidized by CYPs, especially CYP2E1 [16]. However, an overdose can overwhelm the capacity of the conjugation pathways resulting in a greater percentage of APAP being metabolized by CYP2E1 which forms the highly reactive electrophile, N-acetyl-para-benzoquinone imine (NAPQI) [17]. NAPQI has a short half-life and will readily conjugate with sulfhydryl groups. Conjugation with GSH can be a safe detoxification, however, at significantly high doses of APAP the increased levels of NAPQI will lead to a depletion of hepatic GSH, and result in increased covalent binding to sulfhydryl groups on cellular proteins [16]. However, the binding to mitochondrial proteins is of particular importance because it leads to disturbance of the mitochondrial electron transport chain and subsequent reactive oxygen species (ROS) production [18]. This triggers a cascade of stress signaling mediated through the MAPK pathway and activation (phosphorylation) of c-jun N-terminal kinase (JNK) [19,20]. Sustained JNK activation and the mitochondrial translocation of phospho-JNK amplifies the mitochondrial oxidant stress, which leads to the mitochondrial membrane permeability transition pore (MPTP) opening, cessation of ATP synthesis, matrix swelling, rupture of the outer mitochondrial membrane, endonuclease release from the intermembrane space and translocation to the nucleus with DNA fragmentation, ultimately driving oncotic necrosis [21]. Due to these extensive, mechanistic insights into APAP-induced liver injury along with the greatest proportion of ALF being due to APAP misuse, this has encouraged the development of APAP-induced liver injury biomarkers [7]. Many drugs that cause intrinsic DILI share similar features with the pathogenesis of APAP toxicity [22]. There is a parent drug or formation of a reactive metabolite that interferes with essential cell function that effects the redox status of the cell, drives mitochondria or endoplasmic reticulum (ER) stress and causes irreparable DNA damage [1,22].

3. Idiosyncratic Drug-Induced Liver Injury

It is difficult to diagnose or predict iDILI because it occurs in less than 1 out of 10,000 individuals taking the drug, is associated with a prolonged latency period, and has a range of clinical manifestations [23]. The lack of defining features are presumably due to the wide array of drugs that have established idiosyncratic reactions. Over a 30 year period, iDILI was responsible for 32% of drugs being withdrawn from the market [6] and in 2013 11% of ALF cases were a result of idiosyncratic drug reactions [14]. The majority of evidence indicates that iDILI is dependent on activation of the adaptive immune system which occurs due to an individual’s specific human leukocyte antigens (HLA) polymorphisms and propensity for synthesis of immunogenic haptens [24–26]. This idea is congruent with the observation that drugs with a higher daily dosing tend to be involved in iDILI, as it has been suggested that there is probably a threshold for activation of the immune system [1]. The ‘hapten hypothesis’ is the idea that either the parent drug or its metabolites covalently bind to endogenous proteins giving rise to a neoantigen. The prototypical neoantigen-forming drug is halothane, which is predominately eliminated safely though the lungs, but a proportion is metabolized by CYP2E1 in the liver to trifluoroacetic acid [27]. This leads to trifluoroacetylated hepatic proteins, which cause an aggressive immune response [27]. However, many drugs causing idiosyncratic reactions do so in the absence of eosinophilia [2]. Although metabolic activation of drugs is frequently associated with iDILI, drug hepatotoxicity can also occur in the absence of drug metabolism and formation of reactive metabolites [28]. Neither HLA polymorphisms nor the formation of haptens fully captures the complexity of idiosyncratic reactions. The majority of populations with HLA polymorphisms associated with iDILI do not have any adverse effects to the drug [29] and two common drugs, ibuprofen and APAP, form drug-protein adducts [30–32] without triggering an adaptive immune response. The reasoning behind this dichotomy remains largely unknown. There is evidence that drugs or associated metabolites can directly bind to HLA molecules or the T-cell receptors themselves [33,34], although if this were the case consistently in patients one would anticipate a relatively short latency period. Perhaps more convincing is the idea that chronic, minor stress of the mitochondria and ER while initially safely resolved becomes problematic in aggregate. This emerging idea fits well with the hypothesis that iDILI may result from a failure of immune tolerance [1]. Recent work inhibiting regulators of immune tolerance led to a more severe DILI phenotype [35]. Essentially, there is a dose threshold that is being slightly exceeded causing mild liver injury, but which can be resolved through adaptive mechanisms. However, as a function of time, the aberrant dysfunction exceeds the immune tolerance threshold resulting in activation of the immune system and increased risk for iDILI. Although the exact etiology for iDILI remains unclear, there is increasing evidence that it is multifactorial, time-dependent, and requires immune activation [36].

Ultimately, iDILI is commonly assessed by means of exclusion of alternative causes. Most frequently this is done using the Roussel Uclaf Causality Assessment Method (RUCAM) established in 1993 [37,38] and updated in 2016 [39]. The advent of the RUCAM was to help address the incorrect diagnosis of DILI, which has been reviewed in the literature [40–42], and to provide a structured methodology that could be generally adopted. RUCAM uses a continuous scoring system with seven major elements that can be further subdivided. A summation of both positive and negative scores yields a final number which can be interpreted for relative likelihood of DILI (≤ 0 excluded, 1–2 unlikely, 3–5 possible, 6–8 probable, and ≥ 9 highly probable). While the original RUCAM performed well when validated [37] and RUCAM is frequently used, other causality assessment methods do exist. For example, the Drug-Induced Liver Injury Network (DILIN) in the United States uses a panel of expert hepatologists that evaluate the clinical, blood chemistry and histology data that then make a final assessment using a percent-based score (definite > 95%, highly likely 75–95%, probable 50–74%, possible 25–49%, unlikely <25%) [43]. For exploration into the strengths and weaknesses of the causality assessment methods we refer the interested reader to [39,43,44].

4. Traditional Approaches to Assessing Drug-Induced Liver Injury

Traditional measurements used to detect liver injury include blood measurements of alanine aminotransferase (ALT), aspartate aminotransferase (AST), alkaline phosphatase (ALP), bilirubin, and albumin. Clinically, DILI is defined when one of the three following thresholds are met, 1) ≥5 x ULN elevation in ALT or R ≥ 5, 2) ≥ 2 x ULN in ALP or R ≤ 2, 3) ≥3 x ULN in ALT and concurrent elevation of total bilirubin 2 x ULN or R > 2 and < 5 [38,39] (where R represents the ratio of ALT and ALP expressed as multiples of ULN). These various requisites allow for a classification of injury as either hepatocellular, cholestatic, or mixed [42]. Functionally, this can be useful as certain drugs are associated with certain patterns of injury. As an example, APAP and isoniazid are associated with a hepatocellular phenotype, while flucloxacillin and amoxicillin-clavulanate (drugs highly linked to hapten formation) are linked to a cholestatic pattern of injury [45]. When using the updated RUCAM, the classification of injury as either hepatocellular or cholestatic/mixed results in a different weight in the scoring to delineate the likelihood of DILI [39]. A prospective study of DILI patients did not show a difference in mortality and rate of liver transplantation between groups of hepatocellular and cholestatic DILI [46]. Importantly, blood chemistries and their relative distributions can change dramatically during the time course of injury [47], so the pattern of DILI is classified based on the initial measurements taken during the patients admission [48]. However, these traditional measurements all lack specificity as they are general parameters of cell injury and overall liver dysfunction, with limited specificity to any single drug or the mechanism by which the drug is inducing liver injury.

In the case of APAP overdose, the Rumack-Matthew nomogram is routinely employed to determine if the standard of care, N-acetylcysteine (NAC) is needed. By measuring APAP levels in the blood, the nomogram provides a template to predict the likelihood for developing severe liver injury, e.g. an ALT value above 1,000 IU/L after a single overdose. Originally published in 1975 [49], it still remains the gold standard for prognostication. While the nomogram initially was developed for patients presenting to the hospital in the first 18 hours after APAP exposure, it has been extended to 24 hours following ingestion of a toxic dose (>10g for adults). However, there are limitations to its use as it relies upon self-reported time of ingestion from the presenting patient as well as blood measurements of APAP which is rapidly metabolized. Therefore, in late presenting patients with low plasma APAP levels there is often uncertainty as to whether this is a case of APAP overdose. ALT values typically increase 36–48h following APAP overdose while the nomogram only covers the first 24 hours leaving a significant window for uncertainty. While NAC is generally regarded as safe, there have been reported adverse effects [50] and approximately 15% of early-presenting patients who receive NAC treatment go on to develop liver toxicity [51,52].

Due to the ‘orphan’ status of DILI it is unlikely that drug-specific biomarkers will be developed, with one notable exception, APAP. However, a significant body of work has focused on identifying new biomarkers that are related to the mechanism of injury or some cell-related processes. These types of biomarker hold significant promise as they may potentially be a direct representation of the pattern of damage in parenchymal or non-parenchymal cells in DILI. Lastly, the most promising diagnostic marker in APAP-induced liver injury is APAP-protein adducts, which can be measured in circulation and have a much longer half-life than the parent compound [53].

5. Biomarkers of DILI Diagnosis – the example of APAP

The relatively high incidence of APAP overdose and the efficacy of the standard of care, NAC, has encouraged the development of a diagnostic marker for APAP overdose patients in the clinic. The toxic APAP metabolite, NAPQI binds to cysteine residues on hepatic proteins as 3-(cystein-S-yl) APAP, commonly referred to as APAP-CYS [54]. Measurement of protein-derived APAP-CYS can be employed clinically to confirm a diagnosis for APAP overdose especially when a possible combination of drugs have been consumed [55]. Three general approaches have been used to measure APAP-CYS: the first developed was an antibody-based test which also evolved to encompass an ELISA format [56], a high-pressure liquid chromatography and electrochemical detection (HPLC-ECD) method [57] and most recently HPLC with mass spectrometry [58]. Early use of the antibody-based testing revealed an enrichment of APAP-protein adducts in the plasma membrane and mitochondria, which led the authors to suggest that the mitochondria may be important for APAP hepatotoxicity [56]. Additionally, antibody-based testing (Western blotting) is still employed in some preclinical, mechanistic studies to validate APAP metabolism [59]. The ELISA format allowed for a slightly higher throughput approach and was successfully used in both mice [60] and in the serum of APAP overdose patients [61]. Nevertheless, a major drawback with antibody-based methods is the low sensitivity as it failed to detect adducts in patients with ALT values < 5000 IU/L [61]. The HPLC-ECD method is significantly more sensitive, with a lower detection limit of 3 pmol/mg protein, allowing for the detection of protein-derived APAP-CYS in the blood of patients with mild liver injury [57]. However, for patients who ingest therapeutic or moderate supratherapeutic doses, the mass spectrometric detection is more sensitive and more accurate at these lower levels [62]. APAP is rapidly absorbed in the gut reaching peak levels in the liver and plasma within 30 minutes and has a median half-life of 5.4 hours depending on the initial amount of drug consumed [63]. As many patients present late to the clinic when plasma levels of APAP may be low or even undetectable, it can be difficult to determine whether the patients took an overdose of APAP. However, the elimination half-life of APAP-protein adducts in the blood is approximately 42 hours [53]. This not only greatly extends the window of detection for confirming APAP-mediated liver injury, but it also provides a high degree of specificity. Further, peak levels of protein-derived APAP-CYS were found to correlate with peak aminotransferase levels in APAP overdose patients [53]. In patients with ALT > 1000 U/L it has been shown that protein-derived APAP-CYS concentrations in plasma > 1.1 nmol/ml are adequately sensitive and specific to establish toxicity [53]. Recent work has extended these findings by suggesting that APAP-CYS concentrations of 0.58 nmol/ml could identify patients who would later develop hepatotoxicity [64]. Notably, APAP-CYS does not improve risk prediction models when the time of ingestion and plasma APAP levels are known [64]. However, because the time of APAP ingestion cannot be reliably established in many cases, a benefit of using APAP-CYS is that the time of APAP ingestion does not have to be known [62]. A commercial immunoassay is in development to detect APAP protein adducts allowing for rapid measurement of these adducts in a clinical setting and initial results are promising [65]. Taken together, this work provides a strong rationale for measuring APAP-protein adducts as a diagnostic and possibly even predictive biomarker for APAP hepatotoxicity in the clinic, particularly in patients where the dose and time of APAP ingestion is unknown. Similar approaches to developing improved diagnostic measurements for other forms of DILI involving the measurement of metabolic products of the parent compound remain relatively unexplored, mainly due to the low prevalence of a single compound resulting in liver injury and/or liver failure.

6. Biomarkers of Injury and Recovery

The majority of biomarkers being investigated can be classified as prognostic and are derived from a mechanistic understanding of the pathology of interest. The aim of these biomarkers is to determine if hepatotoxicity will occur and if the injury will progress to liver failure. Most commonly these evolving biomarkers are studied in the context of APAP hepatotoxicity, due to the abundance of DILI cases associated with APAP and the greater mechanistic understanding. However, biomarkers which have been studied in incidences of acute DILI outside of APAP hepatotoxicity will also be addressed. DILI causes a cascade of molecular signaling which can broadly be bifurcated into an injury response and a recovery/regeneration response. While certainly the injury program begins before the recovery program, it is unlikely they are mutually exclusive (e.g. recovery/regeneration response will begin while injury is still ongoing). Therefore, while we will group these evolving biomarkers by clustering them as either relating to the cellular injury response or the cellular recovery/regeneration response, there is a recognized potential overlap (Table 1).

Table 1:

Prominent novel biomarkers in DILI.

Biomarker	Phase	Efficacy	Limitation	Ref #
HMGB1	Injury	↑ HMGB1 before ↑ ALT, potential use in AH	Modest prognostic value, useful when integrated with miR-122 and K18	10, 51, 73
K18/ccK18	Injury	Cell death mode, predict early stage mortality in AH patients, diagnostic and prognostic for DILI	Lacks liver specificity	10, 74, 76, 77
GLDH	Injury	Liver specific, may be specific to mitochondrial targeted injury, ↑ GLDH before ↑ ALT	↑ levels of GLDH independent of hepatotoxicity	51, 76, 80, 81, 82, 83
AFP	Regeneration	Moderate prognostic and diagnostic value for DILI	Moderate inter- subject variability	76, 94
LECT2	Inflammation/ Regeneration	↑ during liver regeneration in humans	Ambiguity: ↑ in survivors following non-DILI ALF, ↓ in survivors following DILI	108, 109, 110
MCSFR	Inflammation	May be sensitive to severe IDILI, diagnostic and prognostic for DILI	Limited evidence for efficacy in IDILI, lacks liver specificity	76
ITGB3	Inflammation/Regeneration	Prognostic and diagnostic potential for Diclofenac-DILI	Limited evidence, incorrect RUCAM classification	128
OPN	Inflammation/ Regeneration	Promotes liver regeneration, diagnostic and prognostic for DILI	Counterintuitive: ↑ OPN in patients with poor outcome	76, 111, 112, 113
miR-122#	Injury	Liver specific, early prognostic biomarker, useful as biomarker isoniazid hepatotoxicity	High inter- and intra- variability, challenging to measure in a clinic	10, 12, 76, 118, 119

^#EV-associated miR-122 has been explored as a biomarker due to its greater stability, however, preclinical evidence demonstrates significant variability limiting its clinical benefit.

6.1. Liver Injury: Cell Death and Mitochondrial Toxicity

The primary form of cell death during APAP overdose is hepatocyte necrosis [21] which is established from morphological characteristics such as cell swelling, karyolysis, and ultimately the release of intracellular contents [66]. This rupturing of the cell and release of biomolecules normally sequestered in intracellular compartments (e.g. mitochondria/nucleus) activates an immune response and has prompted interest in establishing these molecules as biomarkers. HMGB1 is a nuclear protein which canonically functions as a chromatin-binding factor but when released by necrotic cells acts as a DAMP targeting TLRs and the receptor for advanced glycation end products (RAGE). Further, hyperacetylated HMGB1 can be actively secreted from immune cells as part of an inflammatory response [67]. Acetylated HMGB1 (acHMBG1) generated significant interest as a predictive biomarker for outcome in APAP-induced acute liver failure patients when first reported [68]. However, an internal investigation at the University of Liverpool revealed that the acHMGB1 data published in this manuscript and others are not reproducible and hence the manuscript was recently retracted [69]. In addition, a number of other journals have published notes of concerns regarding the acHMGB1 data. This has led to the retraction in the letter of support by the European Medicines Agency (EMA) (which is no longer found on their website) for further development and exploratory use of not only acHMGB1, but also K18, OPN, and MCSFR [70], which were part of the initially approved package by EMA and FDA [71]. Nevertheless, while the acHMGB1 data are clearly questionable, the data of the other biomarkers used in this manuscript (total HMGB1, K18 and ccK18) could be verified using some of the original samples [69]. In addition, total HMGB1 as a biomarker has also been vindicated in other studies with APAP overdose patients [10,72], as well as the other aforementioned biomarkers, which are explored throughout this text. Total HMGB1 is elevated in patients after APAP overdose [10] and notably is increased in early-presenting APAP overdose patients prior to any elevation in ALT [51]. As a standalone diagnostic biomarker, HMGB1 produces only modest results, however, when utilized in a combined model using miR-122 and K18 it functions as a better predictor for APAP-induced liver injury than ALT alone [10]. Additionally, patients with acute alcoholic hepatitis (AH) exhibit elevated levels of HMGB1 which positively correlated with markers of injury and inflammation. However, HMGB1 failed to act as an independent predictor for metrics of AH disease progression limiting its clinical potential [73].

A more promising biomarker is cytokeratin 18 (K18), which can be measured in the plasma as either total K18 or as a caspase-cleaved product (ccK18). K18, a cytoskeletal protein, is cleaved by caspases resulting in a truncated form that is detectable by the monoclonal antibody M30. This earned ccK18 the moniker M30 while total K18 (full-length K18 + ccK18) was denoted as M65 (after the detection antibody). Correspondingly, the ratio of M30/M65 (‘apoptotic index’ or AI) has been used as a method to characterize the mode of cell death as predominately necrotic or apoptotic. In APAP overdose patients, M65 was greatly increased compared to M30 indicating that hepatic necrosis is the primary modality of cell death [74] which is congruent with human hepatocytes [75] and animal studies [66]. Both total K18 and ccK18 have successfully predicted elevations in ALT in early presenting APAP overdose patients [10] and have predicted poor outcome in patients with DILI [76]. These findings have spurred letters of support from the FDA for further development of K18 as a diagnostic/prognostic DILI biomarker. Lastly, both the M30/M65 ratio or total K18 (M65) can predict early stage mortality in AH patients [77].

Mitochondrial injury is central to APAP toxicity [18] as well as numerous other DILI associated drugs such as fialuridine, troglitazone, nefazodone and benzbromarone [2,22]. Glutamate dehydrogenase (GLDH) is a mitochondrial matrix enzyme that is highly expressed throughout the liver [78] particularly in the centrilobular region [79]. GLDH has been shown to be released into circulation during APAP-induce liver injury in patients and in mice [80]. In a larger cohort, GLDH plasma levels successfully predicted outcome in APAP-induced injury [81] as well as in early presenting patients before any elevations in ALT [10,51]. In rodent models, GLDH was increased in the plasma of mice treated with APAP but not furosemide (FS), an agent that causes centrilobular necrosis without localized mitochondrial toxicity, suggesting that GLDH is specific to mitochondria-targeted toxicants [80]. However, this idea has recently been challenged as a new investigation found that both APAP and FS caused centrilobular necrosis corresponding with increases in ALT and GLDH [82]. However, consistent with the previous data [74], there was also a substantially higher release of ALT than GLDH in FS hepatotoxicity [76] suggesting some correlation with mitochondrial injury. Nevertheless, GLDH as a prognostic biomarker has low sensitivity [10] and GLDH levels were elevated in healthy patients who had received heparin treatments without hepatotoxicity [83]. One of the limitations of traditional liver injury parameters such as ALT activity in the plasma is the lack of liver specificity as muscle damage also causes increases in blood ALT levels [84]. In this context, GLDH has gained traction due to its liver specificity and it has been suggested as a potential alternative to ALT in certain clinical circumstances where greater discrimination is needed [76]. This is being actively pursued by the Critical Path Institute (C-Path) Predictive Safety Testing Consortium (PSTC) and is supported by the FDA [71].

Other potential biomarkers that have been explored include mitochondrial DNA (mtDNA) [80,81], nuclear DNA (nDNA) [72,80,81], long chain acyl-carnitines [85], sorbitol dehydrogenase (SDH) [76] and GSTα [76,86,87]. MtDNA and nDNA proved to be useful for distinguishing survivors from non-survivors in APAP-induced ALF patients, but produced underwhelming metrics of sensitivity/specificity [81]. Long chain acyl-carnitines appeared promising as an early biomarker that was specific to mitochondrial toxicants based on mouse data [85]. However, acyl-carnitines were not significantly increased in APAP overdose patients, which may reflect the improved mitochondrial bioenergetics when treated with NAC [85]. SDH is a cytosolic enzyme that has long been used as a biomarker for liver diseases [88]. It has been suggested to be an early detector for DILI [83] and as a viable biomarker for isoniazid-induced liver injury [89]. In a recent investigation of patient data, SDH performed moderately producing an AUC of 0.819 for DILI prediction with relatively minimal inter- and intra-variability in healthy patients, but simply was outperformed by other burgeoning biomarkers such as K18 and ccK18 [76]. GST is frequently examined due to its integral role in phase II metabolism, including detoxification of reactive metabolites such as NAPQI, and because genetic polymorphisms of GST correlates with increased risk of hepatotoxicity from multiple drugs [90]. Increases in GSTα preceded changes in ALT in human APAP overdose patients [91] and performed better than ALT in rats exposed to a variety of hepatotoxic drugs [87]. Ultimately, GSTα is a useful indicator for DILI, but fails to have value as a prognostic biomarker [76].

6.2. Liver Regeneration: Cell Repair and Inflammation

While traditional biomarkers used in the clinic are associated with cell injury or associated loss of function (e.g. bilirubin), there has been a drive towards developing biomarkers related to cellular regeneration/recovery. In part, this stems from the recognition that liver regeneration is dependent on the severity of the injury and that beyond a certain threshold regeneration can be impaired [92,93]. The idea is to use one of these harbingers of regeneration to predict patient survival. The underlying assumption is that patients with greater activation of the regenerative response have a greater chance for survival. As an example, rising levels of alpha-fetoprotein (AFP), a liver regeneration marker, over the first 3 days at the hospital was associated with survival in ALF patients [94]. However, while key mediators of liver regeneration are highly activated after severe APAP exposure, there can be a failure to initiate a regenerative response because there is an even greater induction in proteins responsible for restricting regeneration [92,95]. Therefore, while this is an exciting avenue for biomarker discovery, further understanding of the mechanisms driving liver regeneration in DILI will improve our ability to use these regeneration biomarkers in the clinic.

In the context of APAP hepatotoxicity, inflammation sits at the nexus of injury and repair. The role of the inflammatory response during APAP hepatoxicity has been an area of contention. Both sides have presented arguments to the detrimental impact of invading inflammatory cells and to the pro-recovery/regenerative function of immune cells [96, 106]. Furthermore, recent work in mice has highlighted neutrophil mediated hepatic recovery following APAP [97,98]; consistent with these preclinical studies is the observation that neutrophil activation in APAP overdose patients occurs only during the recovery phase but not during the injury [99]. In addition, alternatively activated bone-derived macrophages were used as a therapeutic to abrogate APAP-induced liver injury [100]. Seminal work comparing APAP overdose patient liver samples to pathological controls (e.g. liver disease) found increased anti-inflammatory/hepatoprotective mediators (IL-6, IL-10, TGF-β1) without any differences in proinflammatory mediators and a significant increase in hepatic macrophage numbers (due to recruitment of bone-marrow derived circulating monocytes) [101]. Notably, while this sterile inflammation is predominately beneficial in APAP-DILI [96, 106], the role of inflammation in other forms of DILI is unknown. Generally, if the inflammatory response is proportional to the amount of injury then it is beneficial, however, if the inflammatory response is chronic it may promote fibrosis and carcinogenesis [102] and if disproportionate (to the initial insult) it may aggravate the injury [103]. This integral role for the immune system in both arms of DILI has encouraged the exploration of inflammatory biomarkers as prognostic indicators.

Chemokine ligand 2/monocyte chemoattractant protein 1 (CCL2/MCP1) is a signaling molecule responsible for initial recruitment of monocytes and other immune cells in response to inflammatory conditions [104]. The severity of APAP-induced liver injury in patient serum samples positively correlated with CCL2/MCP1 levels and negatively correlated with the number of circulating monocytes. The authors postulated that the increased CCL2/MCP1 was a compensatory mechanism to enhance recruitment of bone marrow-derived monocytes into circulation, however, the capacity for generation of these cells in the bone marrow was less than the rate of recruitment to the liver, ultimately leading to reduced monocytes in the blood [101]. Further, they believed that since hepatic macrophage expansion occurred in an ‘anti-inflammatory/regenerative’ environment it supported the notion of a pro-recovery role for these cells. Interestingly, CCL2/MCP1 has never been evaluated for its prognostic potential but it strongly indicates the need for greater study into this class of proteins. Two candidates that fit in this mold are leukocyte cell-derived chemotaxin 2 (LECT2) and MCSFR. LECT2 sits at this intersection of cell repair and inflammation as it can recruit neutrophils [105], which growing evidence supports as a requisite for optimal hepatic recovery following APAP [106]. In addition, LECT2 is a target gene of Wnt/β-catenin signaling, which is critical for liver regeneration [107]. LECT2 is increased during liver regeneration following hepatectomy in humans [108] and one study found serum LECT2 was significantly lower in non-survivor patients than in survivors following non-DILI ALF [109]. Nevertheless, more recent data with a larger patient population found that lower LECT2 is associated with survival in ALF patients [110], contrasting with the previous findings [109]. Additionally, a comprehensive investigation of 14 biomarkers in the largest DILI biomarker study found that LECT2 performed the worst for DILI identification (AUC = 0.519), the only one that could not accurately predict DILI [76]. However, in this same study, MCSFR performed very favorably as a biomarker for DILI prediction and as a prognostic for ALF. Interestingly, MCSFR was higher in patients with flupirtine-induced hepatotoxicity compared to APAP overdose patients, suggesting that MCSFR may be useful for distinguishing iDILI from DILI [76].

OPN is another molecule bridging inflammation and liver recovery/regeneration. OPN is secreted by a variety of cells, such as macrophages and endothelial cells, serving as both an extracellular matrix protein and pro-inflammatory cytokine. In rodent models of partial hepatectomy, OPN promotes liver regeneration [111–113], although somewhat paradoxically, OPN is higher in DILI patients with poor outcome [76]. This may simply be due to the direct relationship between injury and regeneration: if there is greater injury, there is need for greater regeneration. Alternatively, it may represent that the regenerative signal is failing to elicit the desired effect (e.g. initiate a recovery program), therefore, OPN continues to be released into circulation. Nevertheless, in the non-surviving patient this is a futile attempt, but results in dramatically elevated OPN levels in the blood [76]. In support of this hypothesis for the elevated OPN levels in patients with poor outcome, in our preliminary data, we observed a prolonged elevation of MCP-1 in non-surviving compared to surviving APAP overdose patients [114].

Importantly, OPN had the greatest prognostic value for predicting liver transplantation in DILI patients when compared to other promising biomarkers such as ccK18/K18 and MCSFR [76]. Other notable molecules associated with liver regeneration that have been assessed as potential biomarkers are AFP and the lipid phosphatidic acid (PA). Of the two, AFP has been more rigorously investigated and has performed moderately with utility as a predictor for DILI [76]. The interest in PA is derived from the finding that it is elevated in plasma from APAP overdose patients and may serve as a facilitator of liver regeneration [115]. Further effort is needed to determine the value of PA in a clinical environment.

7. MicroRNA and EV-associated Biomarkers

MicroRNAs are short, approximately 15–25 nucleotide molecules that are post-transcriptional regulators of gene expression [116]. They have garnered significant interest as potential therapeutics for their capacity as biomarkers across many fields such as cancer and DILI. Early work in APAP-treated rodents identified miR-122 and miR-192 as viable biomarkers as they increased prior to detectable ALT alterations and had a dose-dependent response [117]. Further excitement for miR-122/192 developed when these miRNAs were increased in the serum of APAP overdose patients, but not in chronic kidney disease patients, suggesting liver specificity [8]. miR-122 also has performed very well as a prognostic biomarker for ALF in APAP overdose patients who present with normal ALT [10,76]. Additionally, miR-122 has efficacy as a biomarker in other forms of DILI, such as isoniazid-induced hepatotoxicity [118]. Utilizing miR-122 and albumin together, one group was able to accurately predict 6-month survival in patients with DILI [119]. Various other miRNAs have been associated or increased in other models of injury, such as trovafloxacin-induced liver injury [120], cholestasis, or hepatic steatosis [121]. Nevertheless, the only miRNAs that has been seriously investigated is miR-122. Overall, the excitement for miR-122 has begun to wane, as it is not easily assayed in a clinical environment and had high inter- and intra-variability in healthy volunteers, which makes it a challenge to establish a reliable reference interval for this parameter [76]. Also, the initial appeal for miR-122 was its superiority over ALT for predicting liver injury, however, GLDH is approximately equivalent in its liver specificity and can be measured using standard clinical assays [7]. Despite these concerns, miR-122 still remains a viable biomarker, particularly in APAP-induced liver injury, where the sheer magnitude of change in APAP overdose patients compared to healthy volunteers overcomes the limitation of the baseline variability in miR-122 [8].

There also has been increasing interest in studying extracellular vesicles (EVs) and exploring their potential as biomarkers. EVs are nanosized particles, enclosed in a lipid bilayer that can house the diverse armamentarium of molecules found in a cell [122]. EVs can be classified into three categories as a result of their size and origin: exosomes, microvesicles, and apoptotic bodies [123]. The most widely studied are exosomes although many studies fail to use the appropriate nomenclature making it unclear what exactly is being measured. Many of the molecules that can be assayed in the blood can also be found inside EVs obtained from blood. In most cases of DILI, there is a massive increase in the amount of circulating EVs highlighting their potential as a novel biomarker target [11]. EV-associated miR-122 and miR-192 were increased in a mouse model of APAP-induced liver injury and demonstrated similar liver specificity as ‘free’ miR-122 [124]. The excitement for EV-associated miRNA stemmed from the fact that EVs enclose the miRNA preventing degradation from RNAses in circulation [13]. However, the limitation of ‘free’ miR-122, namely the large variability in patient populations, seems to also be true of EV-associated miR-122 [125]. This idea of greater stability of molecules housed in EVs prompted the search for APAP-CYS in small EVs (e.g. exosomes) collected from mice treated and patients overdosed on APAP, unfortunately, there was none detected [126]. It would be interesting to measure APAP-CYS in larger EV populations (such as microvesicles), as APAP-CYS is predominately in the mitochondria [56] and it has been shown that mitochondria can be transported within larger EVs [127]. Importantly, EVs play a fundamental role in maintaining homeostasis and mounting evidence supports an integral role for EVs in DILI [11]. Therefore, when considering EVs for biomarker development, it is necessary to recognize that they are part of a coordinated biological response by a system (e.g. liver) that are released with the intention of facilitating an effect, not the leaking of cell content from dead or dying cells.

8. Machine Learning and Omics in Biomarker Development

For all the excitement behind this new wave of mechanistic and prognostic biomarkers, they do have inherent limitations. While some of these new biomarkers are specific to liver injury and can offer insight into the mode of cell death, they can exhibit high variability, are rarely specific for the offending hepatotoxicant, and require further validation in clinical settings. We currently sit at the precipice of the explosion of ‘omics’ data. These omics studies in DILI often are proteomics or metabolomics that analyze patient blood or urinary samples. A proteomic analysis of cells from patients with Diclofenac-DILI, which was assessed by RUCAM according to the authors, identified integrin beta 3 (ITGB3) as being upregulated and then went on to validate this as a potential biomarker by staining patient liver biopsy samples (interestingly, it is decreased in the blood of Diclofenac-DILI patients) [128]. Although a promising biomarker, the limited number of patients (16) used in this study [128], the current lack of independent validation, and the unanswered question whether this could be a general iDILI biomarker or if it is specific for diclofenac, limits the current conclusions regarding this biomarker [70]. Nevertheless, this study demonstrates how a big data approach can be leveraged to a potentially meaningful impact in the clinic.

In APAP-derived human samples, patterns of metabolites have been used to predict potential for liver toxicity with greater efficacy than ALT [129]. Other approaches have assessed the transcriptome and metabolome of blood samples collected from rats and humans who had consumed APAP. This work identified a host of changes in gene expression that were associated with impaired oxidative phosphorylation [130]. The wealth of information generated in studies using omics approaches requires a massive computational analysis to find patterns that are biologically meaningful and that can be leveraged for translational application. Taking a step outside of the DILI field and looking towards the domain of cancer research there is great promise in the use of machine learning approaches to identify biomarkers or integrate known biomarkers into a singular, composite feature that has clinical value. Machine learning (ML) is the ability for a system to ‘learn’ from experience without being directly told (programmed) to do so [131]. This is an area of highly active research, as there has been twice as many studies published in the past two years as there were in the past two decades. As an example, recent work investigating patients with advanced melanoma used a ML approach to identify a panel of 16 cytokines that are associated with clearance of immune checkpoint inhibitors [132]. Previous, empirical evidence has demonstrated that lower clearance of immune checkpoint inhibitors is a predictor of overall survival in this patient population [133]. This group then successfully used this novel cytokine panel to determine clearance of the immune checkpoint inhibitor, which was then linked to overall survival in three independent clinical trials [132]. Importantly, the use of an individual cytokine as a biomarker has been insufficient as a prognostic biomarker in advanced melanoma and instead required a combinatorial approach, whereby each cytokine provided useful information. The R value to determine DILI or Kings College Criteria [134] used in the clinic as a predictor for ALF represents this underlying idea: that integration of multivariable factors into a unifying feature can be more powerful than one single factor. While human-powered integration of these features may work for 2–4 variables as it increases further it necessitates a computational approach. Additionally, the query and identification of these biological signatures are well-suited to ML approaches. As ML pipelines become widely accessible there will be a whirlwind of biomarker discovery in DILI, although their utility will have to be thoroughly vetted. ML approaches such as support vector machine algorithms (SVM) have already seen use in preclinical models to predict human risk for DILI as a function of oral dose and blood concentrations [135].

APAP biomarker research is particularly poised to take advantage of this surge in computational approaches due to the vast amount of information that has been collected. A critical feature for the successful identification of the cytokine panel was due to the prior knowledge that lower drug treatment clearance is a predictor of improved overall survival, therefore the target was to identify cytokines associated with clearance. Due to extensive mechanistic research into APAP key elements of APAP overdose pathology is well understood such as, how toxicity is initiated, where toxicity manifests, duration of injury, initiation of the immune response, and onset of recovery. Taking advantage of precipitously declining sequencing costs or even plate-based arrays, a novel cytokine or miR signature could be derived and related to the initiation of a hepatic regenerative program in preclinical models of APAP-induced liver injury. The next step being the measurement and confirmation in patient data tied to a clinical outcome (such as ALF in DILI patients). While not a short-term or easy project to tackle, it is the future: integrating big data using the drastic increases in computational power to capture novel insights in biology.

9. Expert Opinion

The trifecta for DILI prognosis (K18, OPN, MCSFR) was poised to undergo validation in tightly controlled trials where regulatory agencies could then incorporate them into existing guidelines. However, the retraction of a notable paper due to falsified data regarding acHMGB1 [68,69], followed by a series of subsequent retractions or investigations of related works, has derailed support from EMA for further development of K18, HMGB1, OPN, and MCSFR. While this has hampered excitement around these biomarkers at a regulatory level, the data supporting K18, total HMGB1, OPN and MCSFR has been validated [10,72,76], and these biomarkers are tainted due to their association with acHMGB1. Further mechanistic studies exploring the exact role for OPN and MCSFR, as these sort of intermediaries between recovery and inflammation, will help uncover the ‘why’ behind their efficacy and support their further development.

Collectively, the novel biomarkers that have been elucidated and explored over the past decade hold significant promise. However, as we discussed thoroughly in our previous review [7], there is ultimately the underlying problem of low prevalence of DILI. While it is common to use sensitivity (True Positive Rate) and specificity (False Positive Rate) for validating models in DILI biomarker research, positive predictive value (PPV) and negative predictive value (NPV) are often used instead (Table 2). This is because predictive values integrate sensitivity and specificity while also considering the prevalence of the condition in the given population. Therefore, predictive values represent the probability that a given test result (e.g. disease or no disease) is actually true. Taking prevalence into consideration is necessary for biomarker development in DILI because even a biomarker with very high sensitivity and specificity will generate a large number of false positives. In the case of iDILI, assuming a 1:10000 incidence rate, this means that a biomarker with 95% sensitivity/specificity will result in PPV less than 20%. This means that if you are diagnosed with DILI, there is only a 20% probability that this diagnosis is correct (e.g. there is still a much greater likelihood that you do not have DILI). Clinical decisions cannot effectively be made with this level of confidence in the information. However, this 1:10000 metric applies to the entire United States population (e.g. out of 10,000 people who take this drug, 1 will have an adverse reaction), therefore, if you are assessing people who have already presented to the clinic, reported some sort of GI/liver pain, it is reasonable to assume a much greater proportion of those individuals are experiencing DILI. For example, there are approximately 30,000 hospital admissions associated with APAP overdose and 500 deaths annually [136]. This means the risk of death is 1:60. Correspondingly, a biomarker with 95% sensitivity/specificity would result in a positive predictive value of ~80%. These sorts of metrics have already been broached with these new biomarkers [10,76], so the question is what salient information is not captured by PPV/NPV metrics and how can we improve from here? What is not robustly accounted for in PPV/NPV metrics is the large amount of inter- and intra- variability inherent to biomarkers [76]. Significant effort needs to be spent in establishing a reference range for these biomarkers. Many studies do not adequately explore the standard range of values, and when it is completed there is often very high variability between individuals [76]. This creates very noisy data, which coupled with low prevalence, can lead to false conclusions.

Table 2:

Confusion matrix relating the terminologies used in statistics, biomarker discovery and machine learning (ML).

		Actual
		ALF	No ALF
Prediction	ALF	True Positive	False Positive (Type I error)		Precision (Positive Predictive Value) $\frac{\mathit{T}\mathit{N}}{\mathit{T}\mathit{P}+\mathit{F}\mathit{P}}$
	No ALF	False Negative (Type II error)	True Negative		False Negative Rate $\frac{\mathit{F}\mathit{N}}{\mathit{T}\mathit{N}+\mathit{F}\mathit{N}}$	Negative Predictive Value $\frac{\mathit{T}\mathit{N}}{\mathit{T}\mathit{N}+\mathit{F}\mathit{N}}$
		True Positive Rate* (Sensitivity) $\frac{\mathit{T}\mathit{P}}{\mathit{T}\mathit{P}+\mathit{F}\mathit{N}}$	True Negative Rate (Specificity) $\frac{\mathit{T}\mathit{N}}{\mathit{T}\mathit{N}+\mathit{F}\mathit{P}}$	False Positive Rate $\frac{\mathit{F}\mathit{P}}{\mathit{F}\mathit{P}+\mathit{T}\mathit{N}}$

Acute liver failure (ALF) is being predicted. Sensitivity and specificity are useful indices for validating the efficacy for biomarkers, however, they do not account for the prevalence of the disease in the population. Due to the relatively low prevalence of DILI, PPV/NPV is often used in DILI biomarker discovery. PPV attempts to estimate the probability of ALF when a patient has a positive test result.

^*True positive rate/sensitivity is also referred to as ‘recall’ in ML.

Biomarker discovery or improvement of existing biomarkers can occur either by further refinement of our mechanistic understandings of DILI etiology and/or by using untargeted omics approaches in conjunction with ML to ‘see’ patterns in the data that us humans have missed. Studies have reported contrasting findings in regard to OPN, with one report finding OPN is higher in DILI patients with poor outcome [76], and the other finding reduced OPN in patients who died or received liver transplants [137]. This discrepancy may result from the evidence that liver regeneration is proportional to injury up to a certain threshold, at which point liver regeneration is strongly impaired [92]. Therefore, probabilistic sampling of a similar patient population may lead to divergent outcomes. By having the contextual understanding that there is not a simple, direct relationship between severity of injury and regeneration, we can begin to make sense of these contrasting reports. This highlights why further delineation of both the positive and negative mediators of liver regeneration will be critical to the successful development of biomarkers indicative of repair/regeneration.

10. Conclusion

As previously alluded, the canonical biomarkers and those that are in development are reaching the maximum limit of their effectiveness. In order to make significant improvements, a greater understanding of the biology, or the use of technological advances (e.g. ML/omics) need to be employed. In the case of APAP-DILI, this means an enhanced understanding of the mechanisms driving recovery/regeneration and the cellular control mechanisms behind the ‘switch’ from injury to repair. ML can be used on two fronts: it can be used as tool to assess the wealth of data that has already been collected from DILI patients to identify signatures that were previously missed and it can be used by preclinical researchers to gain insights into the key processes that may be behind patient survival (e.g. induction of a recovery program). The explosion in omics technologies allows for an untargeted approach behind data acquisition and ML allows for an untargeted approach to deriving biological insights. When these two technologies are leveraged judiciously, new candidate biomarkers or a composition of markers can be discovered.

As proof-of-principle, just published work conducted a proteomic profiling of EVs collected from human patients and used a ML approach to distinguish a protein profile that could accurately classify tumors of unknown primary origin [138]. This publication links three growing fields of interest: EVs, omics, and ML. Similar methodologies used in DILI biomarker development may usher in a new wave of markers as well as improve the current combinatorial models used for DILI prediction and prognosis.

Article Highlights.

• Mechanistic insights into DILI, especially acetaminophen-induced liver injury, has helped identify novel biomarkers that can be broadly classified as relating to the injury phase or to the recovery phase.

• The recent retraction in the letter of support by EMA from further development of acHMGB1/HMGB1, K18, MCSFR, and osteopontin as biomarkers was due to falsified data regarding acHMGB1. However, the other biomarkers have been validated by other groups.

• K18, MCSFR, and osteopontin have emerged as novel biomarkers for DILI prognosis. GLDH and miR-122 are both liver-specific, but the high variability in miR-122 and difficulty in measuring it in a clinical setting has somewhat dampened excitement for miR-122.

• A common acetaminophen metabolite, APAP-CYS, is effective at identifying APAP overdose patients due to a longer half-life than APAP. An immunoassay that is feasible in a clinical setting has been developed and tested.

• Extracellular vesicle (EV)-associated biomarkers may have enhanced stability in the blood compared to their ‘free’ counterparts, however, evidence suggests that EVs are released in DILI to facilitate a biological effect and the kinetics of endogenous EVs remains unexplored.

• The declining costs in gene sequencing and other omics approaches correspond with the exponential increase in machine-learning (ML) approaches that can be used in an untargeted fashion to identify novel biological signatures. ML, as a tool, is becoming much more accessible which will foster interest in leveraging ML for DILI biomarker discovery/development.