1. Introduction
More than 30,000 drugs have been developed for diverse diseases, of which 1100 drugs could potentially cause liver injury. In the United Kingdom, the incidence of drug-induced liver injury (DILI) was reported as 13.9 per 100,000 inhabitants [
1], while contemporary studies in China suggested a higher incidence of 23.8 per 100,000 persons with a different etiology from that of Western countries [
2]. As one of the most severe adverse drug reactions (ADRs), DILI can damage the liver, causing acute liver failure (ALF), and fulminant hepatic failure that eventually requires a liver transplant or causes death [
3,
4,
5]; however, due to the assorted clinical features and complex mechanisms of DILI, clinicians often fail to detect the condition early and miss the critical window to treat the patient effectively [
6]. Incidents of DILI have been the major reason for regulatory bodies to decline new drug applications, or for pharmaceutical companies to modify dosing and regimens, declare prescription warnings, or withdraw the drug entirely from the market [
7].
Currently, there are emerging preclinical human-relevant in vitro models used to evaluate the toxic injury of drug candidates to the liver. In these models, either single-cell type or multi-cell type assays can be performed [
8]. The main difference between these two kinds of assays is the number of cell types used in the experiments. Only one of the primary human hepatocytes, immortalized liver-derived cell lines (e.g., HepG2, HuH7) or hepatocyte-like cells derived from stem cells, are generally used in single cell-type in vitro models, while multiple cell lines or multicellular co-culture systems are used as the representative of in vivo cellular behavior in multicell type assays [
9]. Three-dimensional in vitro liver co-culture systems were also developed for the investigation of DILI, where cytochrome P450 (CYP450) inducibility and bile canaliculi-like structures are imitated [
10]. Besides cytotoxicity assays, other in vitro assays are conducted to research the specific mechanism of potential hepatotoxicity. For example, the glucose-galactose assay and oxygen uptake assay can be conducted for the determination of mitochondrial injury [
11]. Investigation for BSEP inhibition by drugs can be also a manner to evaluate DILI [
12]. In addition, covalent binding assays and reactive metabolite trap** are used to detect the formation of reactive metabolites that can commonly cause liver injury [
13].
Animal models also play an important role in the pharmacokinetics and toxicity researches of drug metabolism in vivo relevant to DILI. Several models of chimeric mice with humanized hepatocytes have been developed over the years, most of which require the damage of endogenous mouse hepatocytes followed by a transplant of human liver cells. Highly immunodeficient NOG mice (TK-NOG) are a humanized liver model expressing a herpes simplex virus type 1 thymidine kinase (HSVtk) transgene and mouse liver cells that were ablated after exposure to ganciclovir [
14]. Humanized liver Fah−/−/Rag2−/−/Il2rg−/− (FRG) mice were developed by Azuma et al. by transplanting human hepatocytes into FRG mice whose endogenous hepatocytes were damaged due to the genetic block of the tyrosine catabolic pathway [
15]. These humanized liver models exhibit comparable liver enzyme expression levels and activity to the donor livers [
16], as an alternative tool to study the potential damage to the human liver of drug candidates. In addition to mouse models with humanized livers, several human CYP-transgenic mouse models have been generated. CYP450 humanization in mice can be achieved through the cross-breeding of a human CYP-transgenic mouse with a mouse-CYP knockout mouse or directly knocking in the human genes to replace the mouse genes [
17]. However, CYP450 humanization in mice can only investigate the action of a single human CYP transgene on drugs, hence the limited significance of these models to human drug hepatotoxicity assessment [
18].
Commonly used tests to assess DILI, including in vitro experiments or in vivo animal models mentioned above, are sometimes inaccurate when the data were extrapolated to humans. In addition, it is difficult to detect DILI during prospective clinical trials as the relatively low DILI occurrence rate compared to other serious adverse events (SAEs) requires a disproportionately large sample size for detection. Ethical considerations also forbid re-challenge tests for DILI [
19,
20]. Although an interactive software “Evaluation of Drug-Induced Serious Hepatotoxicity” (eDISH), released by the FDA in 2004, aims at monitoring and evaluating drug liver toxicity in clinical trials, it mainly depends on the acquired clinical data (e.g., serum level of alanine aminotransferase) obtained from each trial subject and it cannot predict drug hepatotoxicity before drugs are applied in humans [
21]. Therefore, early diagnosis of DILI or the timely referral of patients is of great importance for a drug throughout its whole lifespan. In silico methods provide an effective approach to screening drugs that may cause hepatotoxicity. We aim to review the major underlying mechanisms of DILI and the in silico approaches to predict DILI risks, including computer algorithm models and the DILIsym model.
4. Discussion
According to the General Practice Research Database (GPRD), the crude incidence of drug hepatotoxicity in the UK is 2.4 cases per 100,000 people per year [
1]. Delayed diagnosis and treatment of DILI have resulted in an increased burden on the healthcare system, as well as causing preventable morbidity and mortality among patients who rely on these drugs. In addition to clinical consequences, DILI is also a major reason for drug withdrawal and a cost driver for pharmaceutical companies. As such, there is an urgent need to improve the capability to predict DILI risks at the early stage of drug development. The pharmaceutical industry has increasingly turned to in silico modeling, and the technique could hold the key to better risk prediction.
Knowledge-based predictive models use the computational algorithm and specific descriptors to develop the relationship between the drug properties and DILI outcome and further predict the DILI risk. The most challenging problem of knowledge-based predictive models is the lack of DILI annotation datasets, which can be defined as a comprehensive collation of drug information, including the hepatotoxicity descriptions from available data, dosing regimen, basic properties, related study results, and the risk classification schemes of DILI into a single database. There are various public DILI annotation datasets with different drugs and content, for example, the LiverTox Dataset [
93], the LTKB dataset [
94], and DILIst [
95]. However, these datasets are still several orders of magnitude smaller than the benchmark datasets used in drug discovery [
96]. This is the bottleneck that restricts widespread validation and application of knowledge-based predictive models to the development of new chemical entities. Ma et al. established a property augmentation approach to include massive training data and significantly improved the predictive accuracy to 81.4% using cross-validation [
97], indicating the importance of the data size for the predictive results. In this case, more effort should be made with the extension of DILI annotation datasets. For example, some drugs withdrawn from the world market or only listed in one country can be embraced into the research scope. In addition, the risk judgment of drugs in datasets can be more concrete and precise by incorporating information both from the literature and clinical outcomes.
In addition, DILI is often simplified to a qualitative classification problem in the case of computational predictions. This approach, however, does not provide sufficient information on essential factors such as dose dependency or affected patient population during model development. Consequently, the practical applicability of such models is limited. Mechanism-based prediction algorithms such as DILIsym can combine cell assays, PBPK modeling, and interindividual variation into a single model to predict the time course of drug hepatotoxicity. The ability to simulate different dosing regimens and to predict the corresponding DILI risk of each regimen can help to clarify some fundamental clinical questions, including the dose-dependency of a particular drug and determination of safety margin can be undertaken through dose escalations in the virtual subjects within DILIsym [
98]. Additionally, hepatotoxicity prediction using DILIsym can circumvent the problem of species differences in preclinical DILI research, especially bile acid-associated liver injury [
99,
100]. However, the DILIsym model does not include all mechanisms of hepatotoxicity, primarily the immune-related liver injury as previously described. This could potentially lead to a certain degree of underestimation of the risk of DILI. In addition, the DILIsym model cannot mimic various special populations, especially the diseased population that is most vulnerable to DILI. Lastly, most in silico predictive models of DILI, exclude information on macromolecules, metallic, and inorganic compounds resulting in few predictions of DILI of these drugs. These problems can turn into the main research interests of DILI prediction in the future.
DILI in China accounts for about 20% of the hospitalization rate of acute liver injury [
101], which could be attributed to both western medicine and traditional Chinese medicine (TCM) [
24]. This highlights the importance and urgent need to predict TCM-induced liver injury in China and other regional countries that utilize TCM as part of the healthcare system. Huang et al. used the validated QSAR model that was based on the Liver Toxicity Knowledge Base to investigate the hepatotoxic potential of identified ingredients in the Traditional Chinese Medicine Database of Taiwan. The result showed 74.8% of 9160 unique chemicals possessed hepatotoxic potential, with high hepatotoxic potential for 100 chemical ingredients [
102]. The results called for immediate attention to conducting comprehensive assessments of TCM-induced liver injury, especially using mixed model approaches that have been proposed to improve the predictive accuracy of computational models of TCM-induced toxicity. Typically these models would integrate physicochemical data, observational data on drug toxicity, and biological information into a single database for model development using different machine-learning methods [
103]. For example, Li et al. successfully used an SVM classifier to delineate the relationship between in vitro hepatotoxic benchmark concentrations and in vivo AUC of training sets, followed by estimating the in vivo plasma profile based on the cytotoxicity data of natural products derived from traditional Chinese medicines (NP-TCMs) of interest by in vitro–in vivo extrapolation (IVIVE) relationship. Then, the oral dosing schedules of NP-TCMs were predicted using PBPK modeling reversely to help safety assessment of TCM-induced liver injury [
104].
Theoretically, all subjects who received a drug share the same risks of experiencing hepatotoxic drugs. However, certain populations will be more vulnerable to hepatotoxic drugs. Applying in silico models mentioned above in the analyses of DILI risk factors among vulnerable patient groups is of great significance for DILI research. Related genes have always been studied to understand the pathogenesis of specific DILIs. Moore et al. used machine learning approaches, including multivariate adaptive regression splines (MARS), multifactor dimensionality reduction (MDR), and logistic regression to investigate single-nucleotide polymorphism (SNP)–SNP interaction as a potential DILI risk factor [
105]. Moreover, gender has also played an important role in IDILI. Women are more susceptible to hepatotoxicity from certain drugs, such as minocycline and methyldopa [
106]. In addition, environmental factors leading to drug-induced liver injury cannot be ignored. Excessive drinking may increase the risk of DILI caused by duloxetine, APAP, methotrexate, and isoniazid [
24]. However, few DILI predictions focus on vulnerable populations, which can be one of the important issues for further study.
In conclusion, we have introduced two main types of in silico models for DILI prediction: knowledge-based models using the computational algorithm and mechanism-based models using DILIsym, which could be used to evaluate the risk of drug-associated liver adverse events in both clinical settings and drug development processes.